Adding upstream version 1:115.7.0.upstream/1%115.7.0upstream

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

10 files changed, 6071 insertions, 0 deletions

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..6ae776c7c4
--- /dev/null
+++ b/media/libaom/config/win/ia32/config/aom_config.asm
@@ -0,0 +1,76 @@
+;
+; Copyright (c) 2018, 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.
+;
+
+ARCH_ARM equ 0
+ARCH_MIPS equ 0
+ARCH_PPC equ 0
+ARCH_X86 equ 1
+ARCH_X86_64 equ 0
+CONFIG_2PASS_PARTITION_SEARCH_LVL equ 1
+CONFIG_ACCOUNTING equ 0
+CONFIG_ANALYZER equ 0
+CONFIG_AV1_DECODER equ 1
+CONFIG_AV1_ENCODER equ 0
+CONFIG_BIG_ENDIAN equ 0
+CONFIG_BITSTREAM_DEBUG equ 0
+CONFIG_COEFFICIENT_RANGE_CHECKING equ 0
+CONFIG_COLLECT_INTER_MODE_RD_STATS equ 0
+CONFIG_COLLECT_RD_STATS equ 0
+CONFIG_DEBUG equ 0
+CONFIG_DENOISE equ 1
+CONFIG_DIST_8X8 equ 0
+CONFIG_ENTROPY_STATS equ 0
+CONFIG_FILEOPTIONS equ 1
+CONFIG_FIX_GF_LENGTH equ 1
+CONFIG_FP_MB_STATS equ 0
+CONFIG_GCC equ 1
+CONFIG_GCOV equ 0
+CONFIG_GLOBAL_MOTION_SEARCH equ 1
+CONFIG_GPROF equ 0
+CONFIG_INSPECTION equ 0
+CONFIG_INTERNAL_STATS equ 0
+CONFIG_INTER_STATS_ONLY equ 0
+CONFIG_LIBYUV equ 0
+CONFIG_LOWBITDEPTH equ 1
+CONFIG_MAX_DECODE_PROFILE equ 2
+CONFIG_MISMATCH_DEBUG equ 0
+CONFIG_MULTITHREAD equ 1
+CONFIG_NORMAL_TILE_MODE equ 0
+CONFIG_OS_SUPPORT equ 1
+CONFIG_PIC equ 0
+CONFIG_RD_DEBUG equ 0
+CONFIG_REDUCED_ENCODER_BORDER equ 0
+CONFIG_RUNTIME_CPU_DETECT equ 1
+CONFIG_SHARED equ 0
+CONFIG_SHARP_SETTINGS equ 0
+CONFIG_SIZE_LIMIT equ 0
+CONFIG_SPATIAL_RESAMPLING equ 1
+CONFIG_STATIC equ 1
+CONFIG_WEBM_IO equ 0
+DECODE_HEIGHT_LIMIT equ 0
+DECODE_WIDTH_LIMIT equ 0
+HAVE_AVX equ 1
+HAVE_AVX2 equ 1
+HAVE_DSPR2 equ 0
+HAVE_FEXCEPT equ 1
+HAVE_MIPS32 equ 0
+HAVE_MIPS64 equ 0
+HAVE_MMX equ 1
+HAVE_MSA equ 0
+HAVE_NEON 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_VSX equ 0
+HAVE_WXWIDGETS 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..090a455778
--- /dev/null
+++ b/media/libaom/config/win/ia32/config/aom_config.h
@@ -0,0 +1,82 @@
+/*
+ * Copyright (c) 2018, 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 ARCH_ARM 0
+#define ARCH_MIPS 0
+#define ARCH_PPC 0
+#define ARCH_X86 1
+#define ARCH_X86_64 0
+#define CONFIG_2PASS_PARTITION_SEARCH_LVL 1
+#define CONFIG_ACCOUNTING 0
+#define CONFIG_ANALYZER 0
+#define CONFIG_AV1_DECODER 1
+#define CONFIG_AV1_ENCODER 0
+#define CONFIG_BIG_ENDIAN 0
+#define CONFIG_BITSTREAM_DEBUG 0
+#define CONFIG_COEFFICIENT_RANGE_CHECKING 0
+#define CONFIG_COLLECT_INTER_MODE_RD_STATS 0
+#define CONFIG_COLLECT_RD_STATS 0
+#define CONFIG_DEBUG 0
+#define CONFIG_DENOISE 1
+#define CONFIG_DIST_8X8 0
+#define CONFIG_ENTROPY_STATS 0
+#define CONFIG_FILEOPTIONS 1
+#define CONFIG_FIX_GF_LENGTH 1
+#define CONFIG_FP_MB_STATS 0
+#define CONFIG_GCC 1
+#define CONFIG_GCOV 0
+#define CONFIG_GLOBAL_MOTION_SEARCH 1
+#define CONFIG_GPROF 0
+#define CONFIG_INSPECTION 0
+#define CONFIG_INTERNAL_STATS 0
+#define CONFIG_INTER_STATS_ONLY 0
+#define CONFIG_LIBYUV 0
+#define CONFIG_LOWBITDEPTH 1
+#define CONFIG_MAX_DECODE_PROFILE 2
+#define CONFIG_MISMATCH_DEBUG 0
+#define CONFIG_MULTITHREAD 1
+#define CONFIG_NORMAL_TILE_MODE 0
+#define CONFIG_OS_SUPPORT 1
+#define CONFIG_PIC 0
+#define CONFIG_RD_DEBUG 0
+#define CONFIG_REDUCED_ENCODER_BORDER 0
+#define CONFIG_RUNTIME_CPU_DETECT 1
+#define CONFIG_SHARED 0
+#define CONFIG_SHARP_SETTINGS 0
+#define CONFIG_SIZE_LIMIT 0
+#define CONFIG_SPATIAL_RESAMPLING 1
+#define CONFIG_STATIC 1
+#define CONFIG_WEBM_IO 0
+#define DECODE_HEIGHT_LIMIT 0
+#define DECODE_WIDTH_LIMIT 0
+#define HAVE_AVX 1
+#define HAVE_AVX2 1
+#define HAVE_DSPR2 0
+#define HAVE_FEXCEPT 1
+#define HAVE_MIPS32 0
+#define HAVE_MIPS64 0
+#define HAVE_MMX 1
+#define HAVE_MSA 0
+#define HAVE_NEON 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_VSX 0
+#define HAVE_WXWIDGETS 0
+#define INCLUDE_INSTALL_DIR INSTALLDIR/include
+#define INLINE inline
+#define LIB_INSTALL_DIR INSTALLDIR/lib
+#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..5b3762f109
--- /dev/null
+++ b/media/libaom/config/win/ia32/config/aom_dsp_rtcd.h
@@ -0,0 +1,2379 @@
+// 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/enums.h"
+#include "av1/common/blockd.h"
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+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 subx, int suby);
+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 subx, int suby);
+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 subx, int suby);
+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 subx, int suby);
+
+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_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, 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_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);
+RTCD_EXTERN void (*aom_convolve_copy)(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_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_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_2x2 aom_dc_128_predictor_2x2_c
+
+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_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_2x2 aom_dc_left_predictor_2x2_c
+
+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_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_2x2 aom_dc_predictor_2x2_c
+
+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_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_2x2 aom_dc_top_predictor_2x2_c
+
+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_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_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_2x2 aom_h_predictor_2x2_c
+
+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_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 subx, int suby, 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);
+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 subx, int suby, 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 subx, int suby, 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 subx, int suby, 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_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 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 bps);
+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 bps);
+
+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 bps);
+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 bps);
+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 bps);
+
+void aom_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);
+void aom_highbd_convolve_copy_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);
+void aom_highbd_convolve_copy_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 bps);
+RTCD_EXTERN void (*aom_highbd_convolve_copy)(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 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_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_2x2 aom_highbd_dc_128_predictor_2x2_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);
+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_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_2x2 aom_highbd_dc_left_predictor_2x2_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);
+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_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_2x2 aom_highbd_dc_predictor_2x2_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);
+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_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_2x2 aom_highbd_dc_top_predictor_2x2_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);
+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_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_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_2x2 aom_highbd_h_predictor_2x2_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);
+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_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 *limt1, 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 *limt1, 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 *limt1, 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 *limt1, 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);
+
+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_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_2x2 aom_highbd_paeth_predictor_2x2_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_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_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_2x2 aom_highbd_smooth_h_predictor_2x2_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_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_2x2 aom_highbd_smooth_predictor_2x2_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_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_2x2 aom_highbd_smooth_v_predictor_2x2_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
+
+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_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_2x2 aom_highbd_v_predictor_2x2_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);
+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_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 subx, int suby, 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 subx, int suby, 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 subx, int suby, 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 subx, int suby, 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_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_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_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_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_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_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_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_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_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_paeth_predictor_2x2 aom_paeth_predictor_2x2_c
+
+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_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_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_2x2 aom_smooth_h_predictor_2x2_c
+
+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_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_predictor_2x2 aom_smooth_predictor_2x2_c
+
+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_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_2x2 aom_smooth_v_predictor_2x2_c
+
+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);
+
+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_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_2x2 aom_v_predictor_2x2_c
+
+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);
+
+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);
+
+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_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_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;
+    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_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_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_convolve8_horiz = aom_highbd_convolve8_horiz_c;
+    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_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_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_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_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_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_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_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_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_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_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_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_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_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_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_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;
+    av1_round_shift_array = av1_round_shift_array_c;
+    if (flags & HAS_SSE4_1) av1_round_shift_array = av1_round_shift_array_sse4_1;
+}
+#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..6b56795cd4
--- /dev/null
+++ b/media/libaom/config/win/ia32/config/aom_scale_rtcd.h
@@ -0,0 +1,88 @@
+// 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_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_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_copy_u aom_yv12_partial_copy_u_c
+
+void aom_yv12_partial_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_copy_v aom_yv12_partial_copy_v_c
+
+void aom_yv12_partial_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_copy_y aom_yv12_partial_copy_y_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..a2a6f14ff5
--- /dev/null
+++ b/media/libaom/config/win/ia32/config/av1_rtcd.h
@@ -0,0 +1,605 @@
+// 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/txfm_common.h"
+#include "av1/common/common.h"
+#include "av1/common/enums.h"
+#include "av1/common/quant_common.h"
+#include "av1/common/filter.h"
+#include "av1/common/convolve.h"
+#include "av1/common/av1_txfm.h"
+#include "av1/common/odintrin.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;
+
+/* Function pointers return by CfL functions */
+typedef void (*cfl_subsample_lbd_fn)(const uint8_t *input, int input_stride,
+                                     uint16_t *output_q3);
+
+typedef void (*cfl_subsample_hbd_fn)(const uint16_t *input, int input_stride,
+                                     uint16_t *output_q3);
+
+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);
+
+typedef void (*cfl_predict_hbd_fn)(const int16_t *src, uint16_t *dst,
+                                   int dst_stride, int alpha_q3, int bd);
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void 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);
+void 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);
+void 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 void (*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_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_convolve_2d_copy_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_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+void av1_convolve_2d_copy_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_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+void av1_convolve_2d_copy_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_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_convolve_2d_copy_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_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+
+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_q4, 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_q4, 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_q4, 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_q4, const int subpel_y_q4, 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_q4, const int subpel_y_q4, 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_q4, const int subpel_y_q4, 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_q4, const int subpel_y_q4, 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);
+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 InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, 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 InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, 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 InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, 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 InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, 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_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+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_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+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_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+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_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, 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);
+#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
+
+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_edge_high_c(uint16_t *p, int sz, int strength);
+void av1_filter_intra_edge_high_sse4_1(uint16_t *p, int sz, int strength);
+RTCD_EXTERN void (*av1_filter_intra_edge_high)(uint16_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_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_copy_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_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
+void av1_highbd_convolve_2d_copy_sr_sse2(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_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
+void av1_highbd_convolve_2d_copy_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_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_convolve_2d_copy_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_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
+
+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_q4, const int x_step_qn, const int subpel_y_q4, 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_q4, const int x_step_qn, const int subpel_y_q4, 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_q4, const int x_step_qn, const int subpel_y_q4, 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_q4, const int subpel_y_q4, 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_q4, const int subpel_y_q4, 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_q4, const int subpel_y_q4, 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_q4, const int subpel_y_q4, 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_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 InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, 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 InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, 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 InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, 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 InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, 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_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, 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_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, 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_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, 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_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, 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);
+#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
+
+void av1_highbd_inv_txfm_add_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_sse4_1(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_avx2(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_16x16_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_16x16_sse4_1(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_16x16)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_16x8_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_16x8_sse4_1(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_16x8)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_32x32_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_32x32_sse4_1(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_32x32_avx2(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_32x32)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_4x4_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_4x4_sse4_1(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_4x4)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_8x16_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_8x16_sse4_1(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_8x16)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_8x8_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_8x8_sse4_1(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_8x8)(const tran_low_t *dqcoeff, uint8_t *dst, 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
+
+void av1_highbd_jnt_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_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
+void av1_highbd_jnt_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_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
+void av1_highbd_jnt_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_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_jnt_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_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_jnt_convolve_2d_copy_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_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
+void av1_highbd_jnt_convolve_2d_copy_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_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
+void av1_highbd_jnt_convolve_2d_copy_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_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_jnt_convolve_2d_copy)(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_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_jnt_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 InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
+void av1_highbd_jnt_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 InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
+void av1_highbd_jnt_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 InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_jnt_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 InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_jnt_convolve_y_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_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
+void av1_highbd_jnt_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_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
+void av1_highbd_jnt_convolve_y_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_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_jnt_convolve_y)(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_q4, const int subpel_y_q4, ConvolveParams *conv_params, 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);
+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 ConvolveParams *conv_params, int bps);
+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 ConvolveParams *conv_params, int bps);
+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 ConvolveParams *conv_params, int bps);
+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 ConvolveParams *conv_params, int bps);
+
+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_jnt_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_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+void av1_jnt_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_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+void av1_jnt_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_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_jnt_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_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+
+void av1_jnt_convolve_2d_copy_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_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+void av1_jnt_convolve_2d_copy_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_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+void av1_jnt_convolve_2d_copy_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_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_jnt_convolve_2d_copy)(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_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+
+void av1_jnt_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 InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+void av1_jnt_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 InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+void av1_jnt_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 InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_jnt_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 InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+
+void av1_jnt_convolve_y_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_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+void av1_jnt_convolve_y_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_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+void av1_jnt_convolve_y_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_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_jnt_convolve_y)(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_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+
+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_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_upsample_intra_edge_high_c(uint16_t *p, int sz, int bd);
+void av1_upsample_intra_edge_high_sse4_1(uint16_t *p, int sz, int bd);
+RTCD_EXTERN void (*av1_upsample_intra_edge_high)(uint16_t *p, int sz, int bd);
+
+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);
+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_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 ConvolveParams *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 ConvolveParams *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 ConvolveParams *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 ConvolveParams *conv_params);
+
+void cdef_filter_block_c(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int max, int coeff_shift);
+void cdef_filter_block_sse2(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int max, int coeff_shift);
+void cdef_filter_block_ssse3(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int max, int coeff_shift);
+void cdef_filter_block_sse4_1(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int max, int coeff_shift);
+void cdef_filter_block_avx2(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int max, int coeff_shift);
+RTCD_EXTERN void (*cdef_filter_block)(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int max, int coeff_shift);
+
+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);
+
+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);
+
+void copy_rect8_16bit_to_16bit_c(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int v, int h);
+void copy_rect8_16bit_to_16bit_sse2(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int v, int h);
+void copy_rect8_16bit_to_16bit_ssse3(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int v, int h);
+void copy_rect8_16bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int v, int h);
+void copy_rect8_16bit_to_16bit_avx2(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int v, int h);
+RTCD_EXTERN void (*copy_rect8_16bit_to_16bit)(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int v, int h);
+
+void copy_rect8_8bit_to_16bit_c(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h);
+void copy_rect8_8bit_to_16bit_sse2(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h);
+void copy_rect8_8bit_to_16bit_ssse3(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h);
+void copy_rect8_8bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h);
+void copy_rect8_8bit_to_16bit_avx2(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h);
+RTCD_EXTERN void (*copy_rect8_8bit_to_16bit)(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h);
+
+cfl_predict_hbd_fn get_predict_hbd_fn_c(TX_SIZE tx_size);
+cfl_predict_hbd_fn get_predict_hbd_fn_ssse3(TX_SIZE tx_size);
+cfl_predict_hbd_fn get_predict_hbd_fn_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_predict_hbd_fn (*get_predict_hbd_fn)(TX_SIZE tx_size);
+
+cfl_predict_lbd_fn get_predict_lbd_fn_c(TX_SIZE tx_size);
+cfl_predict_lbd_fn get_predict_lbd_fn_ssse3(TX_SIZE tx_size);
+cfl_predict_lbd_fn get_predict_lbd_fn_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_predict_lbd_fn (*get_predict_lbd_fn)(TX_SIZE tx_size);
+
+cfl_subtract_average_fn get_subtract_average_fn_c(TX_SIZE tx_size);
+cfl_subtract_average_fn get_subtract_average_fn_sse2(TX_SIZE tx_size);
+cfl_subtract_average_fn get_subtract_average_fn_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subtract_average_fn (*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;
+
+    apply_selfguided_restoration = apply_selfguided_restoration_c;
+    if (flags & HAS_SSE4_1) apply_selfguided_restoration = apply_selfguided_restoration_sse4_1;
+    if (flags & HAS_AVX2) apply_selfguided_restoration = apply_selfguided_restoration_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_convolve_2d_copy_sr = av1_convolve_2d_copy_sr_c;
+    if (flags & HAS_SSE2) av1_convolve_2d_copy_sr = av1_convolve_2d_copy_sr_sse2;
+    if (flags & HAS_AVX2) av1_convolve_2d_copy_sr = av1_convolve_2d_copy_sr_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_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_edge_high = av1_filter_intra_edge_high_c;
+    if (flags & HAS_SSE4_1) av1_filter_intra_edge_high = av1_filter_intra_edge_high_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_highbd_convolve_2d_copy_sr = av1_highbd_convolve_2d_copy_sr_c;
+    if (flags & HAS_SSE2) av1_highbd_convolve_2d_copy_sr = av1_highbd_convolve_2d_copy_sr_sse2;
+    if (flags & HAS_AVX2) av1_highbd_convolve_2d_copy_sr = av1_highbd_convolve_2d_copy_sr_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_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_16x16 = av1_highbd_inv_txfm_add_16x16_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_16x16 = av1_highbd_inv_txfm_add_16x16_sse4_1;
+    av1_highbd_inv_txfm_add_16x8 = av1_highbd_inv_txfm_add_16x8_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_16x8 = av1_highbd_inv_txfm_add_16x8_sse4_1;
+    av1_highbd_inv_txfm_add_32x32 = av1_highbd_inv_txfm_add_32x32_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_32x32 = av1_highbd_inv_txfm_add_32x32_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_inv_txfm_add_32x32 = av1_highbd_inv_txfm_add_32x32_avx2;
+    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_8x16 = av1_highbd_inv_txfm_add_8x16_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_8x16 = av1_highbd_inv_txfm_add_8x16_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_jnt_convolve_2d = av1_highbd_jnt_convolve_2d_c;
+    if (flags & HAS_SSE4_1) av1_highbd_jnt_convolve_2d = av1_highbd_jnt_convolve_2d_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_jnt_convolve_2d = av1_highbd_jnt_convolve_2d_avx2;
+    av1_highbd_jnt_convolve_2d_copy = av1_highbd_jnt_convolve_2d_copy_c;
+    if (flags & HAS_SSE4_1) av1_highbd_jnt_convolve_2d_copy = av1_highbd_jnt_convolve_2d_copy_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_jnt_convolve_2d_copy = av1_highbd_jnt_convolve_2d_copy_avx2;
+    av1_highbd_jnt_convolve_x = av1_highbd_jnt_convolve_x_c;
+    if (flags & HAS_SSE4_1) av1_highbd_jnt_convolve_x = av1_highbd_jnt_convolve_x_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_jnt_convolve_x = av1_highbd_jnt_convolve_x_avx2;
+    av1_highbd_jnt_convolve_y = av1_highbd_jnt_convolve_y_c;
+    if (flags & HAS_SSE4_1) av1_highbd_jnt_convolve_y = av1_highbd_jnt_convolve_y_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_jnt_convolve_y = av1_highbd_jnt_convolve_y_avx2;
+    av1_highbd_warp_affine = av1_highbd_warp_affine_c;
+    if (flags & HAS_SSE4_1) av1_highbd_warp_affine = av1_highbd_warp_affine_sse4_1;
+    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_jnt_convolve_2d = av1_jnt_convolve_2d_c;
+    if (flags & HAS_SSSE3) av1_jnt_convolve_2d = av1_jnt_convolve_2d_ssse3;
+    if (flags & HAS_AVX2) av1_jnt_convolve_2d = av1_jnt_convolve_2d_avx2;
+    av1_jnt_convolve_2d_copy = av1_jnt_convolve_2d_copy_c;
+    if (flags & HAS_SSE2) av1_jnt_convolve_2d_copy = av1_jnt_convolve_2d_copy_sse2;
+    if (flags & HAS_AVX2) av1_jnt_convolve_2d_copy = av1_jnt_convolve_2d_copy_avx2;
+    av1_jnt_convolve_x = av1_jnt_convolve_x_c;
+    if (flags & HAS_SSE2) av1_jnt_convolve_x = av1_jnt_convolve_x_sse2;
+    if (flags & HAS_AVX2) av1_jnt_convolve_x = av1_jnt_convolve_x_avx2;
+    av1_jnt_convolve_y = av1_jnt_convolve_y_c;
+    if (flags & HAS_SSE2) av1_jnt_convolve_y = av1_jnt_convolve_y_sse2;
+    if (flags & HAS_AVX2) av1_jnt_convolve_y = av1_jnt_convolve_y_avx2;
+    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_upsample_intra_edge = av1_upsample_intra_edge_c;
+    if (flags & HAS_SSE4_1) av1_upsample_intra_edge = av1_upsample_intra_edge_sse4_1;
+    av1_upsample_intra_edge_high = av1_upsample_intra_edge_high_c;
+    if (flags & HAS_SSE4_1) av1_upsample_intra_edge_high = av1_upsample_intra_edge_high_sse4_1;
+    av1_warp_affine = av1_warp_affine_c;
+    if (flags & HAS_SSE4_1) av1_warp_affine = av1_warp_affine_sse4_1;
+    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_filter_block = cdef_filter_block_c;
+    if (flags & HAS_SSE2) cdef_filter_block = cdef_filter_block_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_block = cdef_filter_block_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_block = cdef_filter_block_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_block = cdef_filter_block_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;
+    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;
+    copy_rect8_16bit_to_16bit = copy_rect8_16bit_to_16bit_c;
+    if (flags & HAS_SSE2) copy_rect8_16bit_to_16bit = copy_rect8_16bit_to_16bit_sse2;
+    if (flags & HAS_SSSE3) copy_rect8_16bit_to_16bit = copy_rect8_16bit_to_16bit_ssse3;
+    if (flags & HAS_SSE4_1) copy_rect8_16bit_to_16bit = copy_rect8_16bit_to_16bit_sse4_1;
+    if (flags & HAS_AVX2) copy_rect8_16bit_to_16bit = copy_rect8_16bit_to_16bit_avx2;
+    copy_rect8_8bit_to_16bit = copy_rect8_8bit_to_16bit_c;
+    if (flags & HAS_SSE2) copy_rect8_8bit_to_16bit = copy_rect8_8bit_to_16bit_sse2;
+    if (flags & HAS_SSSE3) copy_rect8_8bit_to_16bit = copy_rect8_8bit_to_16bit_ssse3;
+    if (flags & HAS_SSE4_1) copy_rect8_8bit_to_16bit = copy_rect8_8bit_to_16bit_sse4_1;
+    if (flags & HAS_AVX2) copy_rect8_8bit_to_16bit = copy_rect8_8bit_to_16bit_avx2;
+    get_predict_hbd_fn = get_predict_hbd_fn_c;
+    if (flags & HAS_SSSE3) get_predict_hbd_fn = get_predict_hbd_fn_ssse3;
+    if (flags & HAS_AVX2) get_predict_hbd_fn = get_predict_hbd_fn_avx2;
+    get_predict_lbd_fn = get_predict_lbd_fn_c;
+    if (flags & HAS_SSSE3) get_predict_lbd_fn = get_predict_lbd_fn_ssse3;
+    if (flags & HAS_AVX2) get_predict_lbd_fn = get_predict_lbd_fn_avx2;
+    get_subtract_average_fn = get_subtract_average_fn_c;
+    if (flags & HAS_SSE2) get_subtract_average_fn = get_subtract_average_fn_sse2;
+    if (flags & HAS_AVX2) get_subtract_average_fn = 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..4ab2dedb42
--- /dev/null
+++ b/media/libaom/config/win/x64/config/aom_config.asm
@@ -0,0 +1,76 @@
+;
+; Copyright (c) 2018, 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.
+;
+
+ARCH_ARM equ 0
+ARCH_MIPS equ 0
+ARCH_PPC equ 0
+ARCH_X86 equ 0
+ARCH_X86_64 equ 1
+CONFIG_2PASS_PARTITION_SEARCH_LVL equ 1
+CONFIG_ACCOUNTING equ 0
+CONFIG_ANALYZER equ 0
+CONFIG_AV1_DECODER equ 1
+CONFIG_AV1_ENCODER equ 0
+CONFIG_BIG_ENDIAN equ 0
+CONFIG_BITSTREAM_DEBUG equ 0
+CONFIG_COEFFICIENT_RANGE_CHECKING equ 0
+CONFIG_COLLECT_INTER_MODE_RD_STATS equ 0
+CONFIG_COLLECT_RD_STATS equ 0
+CONFIG_DEBUG equ 0
+CONFIG_DENOISE equ 1
+CONFIG_DIST_8X8 equ 0
+CONFIG_ENTROPY_STATS equ 0
+CONFIG_FILEOPTIONS equ 1
+CONFIG_FIX_GF_LENGTH equ 1
+CONFIG_FP_MB_STATS equ 0
+CONFIG_GCC equ 1
+CONFIG_GCOV equ 0
+CONFIG_GLOBAL_MOTION_SEARCH equ 1
+CONFIG_GPROF equ 0
+CONFIG_INSPECTION equ 0
+CONFIG_INTERNAL_STATS equ 0
+CONFIG_INTER_STATS_ONLY equ 0
+CONFIG_LIBYUV equ 0
+CONFIG_LOWBITDEPTH equ 1
+CONFIG_MAX_DECODE_PROFILE equ 2
+CONFIG_MISMATCH_DEBUG equ 0
+CONFIG_MULTITHREAD equ 1
+CONFIG_NORMAL_TILE_MODE equ 0
+CONFIG_OS_SUPPORT equ 1
+CONFIG_PIC equ 0
+CONFIG_RD_DEBUG equ 0
+CONFIG_REDUCED_ENCODER_BORDER equ 0
+CONFIG_RUNTIME_CPU_DETECT equ 1
+CONFIG_SHARED equ 0
+CONFIG_SHARP_SETTINGS equ 0
+CONFIG_SIZE_LIMIT equ 0
+CONFIG_SPATIAL_RESAMPLING equ 1
+CONFIG_STATIC equ 1
+CONFIG_WEBM_IO equ 0
+DECODE_HEIGHT_LIMIT equ 0
+DECODE_WIDTH_LIMIT equ 0
+HAVE_AVX equ 1
+HAVE_AVX2 equ 1
+HAVE_DSPR2 equ 0
+HAVE_FEXCEPT equ 1
+HAVE_MIPS32 equ 0
+HAVE_MIPS64 equ 0
+HAVE_MMX equ 1
+HAVE_MSA equ 0
+HAVE_NEON 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_VSX equ 0
+HAVE_WXWIDGETS 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..9a0be7cd11
--- /dev/null
+++ b/media/libaom/config/win/x64/config/aom_config.h
@@ -0,0 +1,82 @@
+/*
+ * Copyright (c) 2018, 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 ARCH_ARM 0
+#define ARCH_MIPS 0
+#define ARCH_PPC 0
+#define ARCH_X86 0
+#define ARCH_X86_64 1
+#define CONFIG_2PASS_PARTITION_SEARCH_LVL 1
+#define CONFIG_ACCOUNTING 0
+#define CONFIG_ANALYZER 0
+#define CONFIG_AV1_DECODER 1
+#define CONFIG_AV1_ENCODER 0
+#define CONFIG_BIG_ENDIAN 0
+#define CONFIG_BITSTREAM_DEBUG 0
+#define CONFIG_COEFFICIENT_RANGE_CHECKING 0
+#define CONFIG_COLLECT_INTER_MODE_RD_STATS 0
+#define CONFIG_COLLECT_RD_STATS 0
+#define CONFIG_DEBUG 0
+#define CONFIG_DENOISE 1
+#define CONFIG_DIST_8X8 0
+#define CONFIG_ENTROPY_STATS 0
+#define CONFIG_FILEOPTIONS 1
+#define CONFIG_FIX_GF_LENGTH 1
+#define CONFIG_FP_MB_STATS 0
+#define CONFIG_GCC 1
+#define CONFIG_GCOV 0
+#define CONFIG_GLOBAL_MOTION_SEARCH 1
+#define CONFIG_GPROF 0
+#define CONFIG_INSPECTION 0
+#define CONFIG_INTERNAL_STATS 0
+#define CONFIG_INTER_STATS_ONLY 0
+#define CONFIG_LIBYUV 0
+#define CONFIG_LOWBITDEPTH 1
+#define CONFIG_MAX_DECODE_PROFILE 2
+#define CONFIG_MISMATCH_DEBUG 0
+#define CONFIG_MULTITHREAD 1
+#define CONFIG_NORMAL_TILE_MODE 0
+#define CONFIG_OS_SUPPORT 1
+#define CONFIG_PIC 0
+#define CONFIG_RD_DEBUG 0
+#define CONFIG_REDUCED_ENCODER_BORDER 0
+#define CONFIG_RUNTIME_CPU_DETECT 1
+#define CONFIG_SHARED 0
+#define CONFIG_SHARP_SETTINGS 0
+#define CONFIG_SIZE_LIMIT 0
+#define CONFIG_SPATIAL_RESAMPLING 1
+#define CONFIG_STATIC 1
+#define CONFIG_WEBM_IO 0
+#define DECODE_HEIGHT_LIMIT 0
+#define DECODE_WIDTH_LIMIT 0
+#define HAVE_AVX 1
+#define HAVE_AVX2 1
+#define HAVE_DSPR2 0
+#define HAVE_FEXCEPT 1
+#define HAVE_MIPS32 0
+#define HAVE_MIPS64 0
+#define HAVE_MMX 1
+#define HAVE_MSA 0
+#define HAVE_NEON 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_VSX 0
+#define HAVE_WXWIDGETS 0
+#define INCLUDE_INSTALL_DIR INSTALLDIR/include
+#define INLINE inline
+#define LIB_INSTALL_DIR INSTALLDIR/lib
+#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..2856d4ede7
--- /dev/null
+++ b/media/libaom/config/win/x64/config/aom_dsp_rtcd.h
@@ -0,0 +1,2001 @@
+// 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/enums.h"
+#include "av1/common/blockd.h"
+
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+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 subx, int suby);
+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 subx, int suby);
+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 subx, int suby);
+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 subx, int suby);
+
+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_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, 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_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);
+#define aom_convolve_copy aom_convolve_copy_sse2
+
+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_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_2x2 aom_dc_128_predictor_2x2_c
+
+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_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_2x2 aom_dc_left_predictor_2x2_c
+
+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_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_2x2 aom_dc_predictor_2x2_c
+
+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_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_2x2 aom_dc_top_predictor_2x2_c
+
+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_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_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_2x2 aom_h_predictor_2x2_c
+
+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_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 subx, int suby, 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);
+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 subx, int suby, 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 subx, int suby, 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 subx, int suby, 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_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 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 bps);
+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 bps);
+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 bps);
+
+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 bps);
+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 bps);
+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 bps);
+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 bps);
+
+void aom_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);
+void aom_highbd_convolve_copy_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);
+void aom_highbd_convolve_copy_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 bps);
+RTCD_EXTERN void (*aom_highbd_convolve_copy)(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 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_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_2x2 aom_highbd_dc_128_predictor_2x2_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);
+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_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_2x2 aom_highbd_dc_left_predictor_2x2_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);
+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_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_2x2 aom_highbd_dc_predictor_2x2_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);
+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_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_2x2 aom_highbd_dc_top_predictor_2x2_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);
+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_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_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_2x2 aom_highbd_h_predictor_2x2_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);
+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_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 *limt1, 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 *limt1, 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 *limt1, 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 *limt1, 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);
+
+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_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_paeth_predictor_2x2 aom_highbd_paeth_predictor_2x2_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_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_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_h_predictor_2x2 aom_highbd_smooth_h_predictor_2x2_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_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_predictor_2x2 aom_highbd_smooth_predictor_2x2_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_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_smooth_v_predictor_2x2 aom_highbd_smooth_v_predictor_2x2_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
+
+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_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_2x2 aom_highbd_v_predictor_2x2_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);
+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_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 subx, int suby, 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 subx, int suby, 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 subx, int suby, 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 subx, int suby, 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_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_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_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_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_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_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_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_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_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_paeth_predictor_2x2 aom_paeth_predictor_2x2_c
+
+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_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_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_h_predictor_2x2 aom_smooth_h_predictor_2x2_c
+
+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_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_predictor_2x2 aom_smooth_predictor_2x2_c
+
+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_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_smooth_v_predictor_2x2 aom_smooth_v_predictor_2x2_c
+
+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);
+
+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_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_2x2 aom_v_predictor_2x2_c
+
+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
+
+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);
+
+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_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_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_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_h_predictor_32x32 = aom_h_predictor_32x32_sse2;
+    if (flags & HAS_AVX2) aom_h_predictor_32x32 = aom_h_predictor_32x32_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_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_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_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_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_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_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;
+    av1_round_shift_array = av1_round_shift_array_c;
+    if (flags & HAS_SSE4_1) av1_round_shift_array = av1_round_shift_array_sse4_1;
+}
+#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..6b56795cd4
--- /dev/null
+++ b/media/libaom/config/win/x64/config/aom_scale_rtcd.h
@@ -0,0 +1,88 @@
+// 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_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_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_copy_u aom_yv12_partial_copy_u_c
+
+void aom_yv12_partial_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_copy_v aom_yv12_partial_copy_v_c
+
+void aom_yv12_partial_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_copy_y aom_yv12_partial_copy_y_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..d273182086
--- /dev/null
+++ b/media/libaom/config/win/x64/config/av1_rtcd.h
@@ -0,0 +1,594 @@
+// 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/txfm_common.h"
+#include "av1/common/common.h"
+#include "av1/common/enums.h"
+#include "av1/common/quant_common.h"
+#include "av1/common/filter.h"
+#include "av1/common/convolve.h"
+#include "av1/common/av1_txfm.h"
+#include "av1/common/odintrin.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;
+
+/* Function pointers return by CfL functions */
+typedef void (*cfl_subsample_lbd_fn)(const uint8_t *input, int input_stride,
+                                     uint16_t *output_q3);
+
+typedef void (*cfl_subsample_hbd_fn)(const uint16_t *input, int input_stride,
+                                     uint16_t *output_q3);
+
+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);
+
+typedef void (*cfl_predict_hbd_fn)(const int16_t *src, uint16_t *dst,
+                                   int dst_stride, int alpha_q3, int bd);
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+void 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);
+void 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);
+void 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 void (*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_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_convolve_2d_copy_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_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+void av1_convolve_2d_copy_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_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+void av1_convolve_2d_copy_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_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_convolve_2d_copy_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_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+
+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_q4, 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_q4, 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_q4, 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_q4, const int subpel_y_q4, 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_q4, const int subpel_y_q4, 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_q4, const int subpel_y_q4, 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_q4, const int subpel_y_q4, 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);
+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 InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, 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 InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, 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 InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, 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 InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, 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_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+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_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+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_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+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_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, 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);
+#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
+
+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_edge_high_c(uint16_t *p, int sz, int strength);
+void av1_filter_intra_edge_high_sse4_1(uint16_t *p, int sz, int strength);
+RTCD_EXTERN void (*av1_filter_intra_edge_high)(uint16_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_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_copy_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_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
+void av1_highbd_convolve_2d_copy_sr_sse2(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_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
+void av1_highbd_convolve_2d_copy_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_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_convolve_2d_copy_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_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
+
+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_q4, const int x_step_qn, const int subpel_y_q4, 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_q4, const int x_step_qn, const int subpel_y_q4, 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_q4, const int x_step_qn, const int subpel_y_q4, 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_q4, const int subpel_y_q4, 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_q4, const int subpel_y_q4, 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_q4, const int subpel_y_q4, 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_q4, const int subpel_y_q4, 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_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 InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, 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 InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, 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 InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, 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 InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, 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_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, 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_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, 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_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, 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_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, 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);
+#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
+
+void av1_highbd_inv_txfm_add_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_sse4_1(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_avx2(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_16x16_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_16x16_sse4_1(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_16x16)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_16x8_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_16x8_sse4_1(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_16x8)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_32x32_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_32x32_sse4_1(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_32x32_avx2(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_32x32)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_4x4_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_4x4_sse4_1(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_4x4)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_8x16_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_8x16_sse4_1(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_8x16)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_8x8_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_8x8_sse4_1(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_8x8)(const tran_low_t *dqcoeff, uint8_t *dst, 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
+
+void av1_highbd_jnt_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_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
+void av1_highbd_jnt_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_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
+void av1_highbd_jnt_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_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_jnt_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_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_jnt_convolve_2d_copy_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_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
+void av1_highbd_jnt_convolve_2d_copy_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_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
+void av1_highbd_jnt_convolve_2d_copy_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_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_jnt_convolve_2d_copy)(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_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_jnt_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 InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
+void av1_highbd_jnt_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 InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
+void av1_highbd_jnt_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 InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_jnt_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 InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_jnt_convolve_y_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_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
+void av1_highbd_jnt_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_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
+void av1_highbd_jnt_convolve_y_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_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_jnt_convolve_y)(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_q4, const int subpel_y_q4, ConvolveParams *conv_params, 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);
+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 ConvolveParams *conv_params, int bps);
+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 ConvolveParams *conv_params, int bps);
+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 ConvolveParams *conv_params, int bps);
+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 ConvolveParams *conv_params, int bps);
+
+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_jnt_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_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+void av1_jnt_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_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+void av1_jnt_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_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_jnt_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_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+
+void av1_jnt_convolve_2d_copy_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_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+void av1_jnt_convolve_2d_copy_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_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+void av1_jnt_convolve_2d_copy_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_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_jnt_convolve_2d_copy)(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_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+
+void av1_jnt_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 InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+void av1_jnt_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 InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+void av1_jnt_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 InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_jnt_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 InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+
+void av1_jnt_convolve_y_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_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+void av1_jnt_convolve_y_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_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+void av1_jnt_convolve_y_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_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_jnt_convolve_y)(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_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+
+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_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_upsample_intra_edge_high_c(uint16_t *p, int sz, int bd);
+void av1_upsample_intra_edge_high_sse4_1(uint16_t *p, int sz, int bd);
+RTCD_EXTERN void (*av1_upsample_intra_edge_high)(uint16_t *p, int sz, int bd);
+
+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);
+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_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 ConvolveParams *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 ConvolveParams *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 ConvolveParams *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 ConvolveParams *conv_params);
+
+void cdef_filter_block_c(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int max, int coeff_shift);
+void cdef_filter_block_sse2(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int max, int coeff_shift);
+void cdef_filter_block_ssse3(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int max, int coeff_shift);
+void cdef_filter_block_sse4_1(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int max, int coeff_shift);
+void cdef_filter_block_avx2(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int max, int coeff_shift);
+RTCD_EXTERN void (*cdef_filter_block)(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int max, int coeff_shift);
+
+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);
+
+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);
+
+void copy_rect8_16bit_to_16bit_c(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int v, int h);
+void copy_rect8_16bit_to_16bit_sse2(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int v, int h);
+void copy_rect8_16bit_to_16bit_ssse3(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int v, int h);
+void copy_rect8_16bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int v, int h);
+void copy_rect8_16bit_to_16bit_avx2(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int v, int h);
+RTCD_EXTERN void (*copy_rect8_16bit_to_16bit)(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int v, int h);
+
+void copy_rect8_8bit_to_16bit_c(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h);
+void copy_rect8_8bit_to_16bit_sse2(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h);
+void copy_rect8_8bit_to_16bit_ssse3(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h);
+void copy_rect8_8bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h);
+void copy_rect8_8bit_to_16bit_avx2(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h);
+RTCD_EXTERN void (*copy_rect8_8bit_to_16bit)(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h);
+
+cfl_predict_hbd_fn get_predict_hbd_fn_c(TX_SIZE tx_size);
+cfl_predict_hbd_fn get_predict_hbd_fn_ssse3(TX_SIZE tx_size);
+cfl_predict_hbd_fn get_predict_hbd_fn_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_predict_hbd_fn (*get_predict_hbd_fn)(TX_SIZE tx_size);
+
+cfl_predict_lbd_fn get_predict_lbd_fn_c(TX_SIZE tx_size);
+cfl_predict_lbd_fn get_predict_lbd_fn_ssse3(TX_SIZE tx_size);
+cfl_predict_lbd_fn get_predict_lbd_fn_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_predict_lbd_fn (*get_predict_lbd_fn)(TX_SIZE tx_size);
+
+cfl_subtract_average_fn get_subtract_average_fn_c(TX_SIZE tx_size);
+cfl_subtract_average_fn get_subtract_average_fn_sse2(TX_SIZE tx_size);
+cfl_subtract_average_fn get_subtract_average_fn_avx2(TX_SIZE tx_size);
+RTCD_EXTERN cfl_subtract_average_fn (*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;
+
+    apply_selfguided_restoration = apply_selfguided_restoration_c;
+    if (flags & HAS_SSE4_1) apply_selfguided_restoration = apply_selfguided_restoration_sse4_1;
+    if (flags & HAS_AVX2) apply_selfguided_restoration = apply_selfguided_restoration_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_convolve_2d_copy_sr = av1_convolve_2d_copy_sr_sse2;
+    if (flags & HAS_AVX2) av1_convolve_2d_copy_sr = av1_convolve_2d_copy_sr_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_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_edge_high = av1_filter_intra_edge_high_c;
+    if (flags & HAS_SSE4_1) av1_filter_intra_edge_high = av1_filter_intra_edge_high_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_highbd_convolve_2d_copy_sr = av1_highbd_convolve_2d_copy_sr_sse2;
+    if (flags & HAS_AVX2) av1_highbd_convolve_2d_copy_sr = av1_highbd_convolve_2d_copy_sr_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_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_16x16 = av1_highbd_inv_txfm_add_16x16_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_16x16 = av1_highbd_inv_txfm_add_16x16_sse4_1;
+    av1_highbd_inv_txfm_add_16x8 = av1_highbd_inv_txfm_add_16x8_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_16x8 = av1_highbd_inv_txfm_add_16x8_sse4_1;
+    av1_highbd_inv_txfm_add_32x32 = av1_highbd_inv_txfm_add_32x32_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_32x32 = av1_highbd_inv_txfm_add_32x32_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_inv_txfm_add_32x32 = av1_highbd_inv_txfm_add_32x32_avx2;
+    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_8x16 = av1_highbd_inv_txfm_add_8x16_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_8x16 = av1_highbd_inv_txfm_add_8x16_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_jnt_convolve_2d = av1_highbd_jnt_convolve_2d_c;
+    if (flags & HAS_SSE4_1) av1_highbd_jnt_convolve_2d = av1_highbd_jnt_convolve_2d_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_jnt_convolve_2d = av1_highbd_jnt_convolve_2d_avx2;
+    av1_highbd_jnt_convolve_2d_copy = av1_highbd_jnt_convolve_2d_copy_c;
+    if (flags & HAS_SSE4_1) av1_highbd_jnt_convolve_2d_copy = av1_highbd_jnt_convolve_2d_copy_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_jnt_convolve_2d_copy = av1_highbd_jnt_convolve_2d_copy_avx2;
+    av1_highbd_jnt_convolve_x = av1_highbd_jnt_convolve_x_c;
+    if (flags & HAS_SSE4_1) av1_highbd_jnt_convolve_x = av1_highbd_jnt_convolve_x_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_jnt_convolve_x = av1_highbd_jnt_convolve_x_avx2;
+    av1_highbd_jnt_convolve_y = av1_highbd_jnt_convolve_y_c;
+    if (flags & HAS_SSE4_1) av1_highbd_jnt_convolve_y = av1_highbd_jnt_convolve_y_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_jnt_convolve_y = av1_highbd_jnt_convolve_y_avx2;
+    av1_highbd_warp_affine = av1_highbd_warp_affine_c;
+    if (flags & HAS_SSE4_1) av1_highbd_warp_affine = av1_highbd_warp_affine_sse4_1;
+    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_jnt_convolve_2d = av1_jnt_convolve_2d_c;
+    if (flags & HAS_SSSE3) av1_jnt_convolve_2d = av1_jnt_convolve_2d_ssse3;
+    if (flags & HAS_AVX2) av1_jnt_convolve_2d = av1_jnt_convolve_2d_avx2;
+    av1_jnt_convolve_2d_copy = av1_jnt_convolve_2d_copy_sse2;
+    if (flags & HAS_AVX2) av1_jnt_convolve_2d_copy = av1_jnt_convolve_2d_copy_avx2;
+    av1_jnt_convolve_x = av1_jnt_convolve_x_sse2;
+    if (flags & HAS_AVX2) av1_jnt_convolve_x = av1_jnt_convolve_x_avx2;
+    av1_jnt_convolve_y = av1_jnt_convolve_y_sse2;
+    if (flags & HAS_AVX2) av1_jnt_convolve_y = av1_jnt_convolve_y_avx2;
+    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_upsample_intra_edge = av1_upsample_intra_edge_c;
+    if (flags & HAS_SSE4_1) av1_upsample_intra_edge = av1_upsample_intra_edge_sse4_1;
+    av1_upsample_intra_edge_high = av1_upsample_intra_edge_high_c;
+    if (flags & HAS_SSE4_1) av1_upsample_intra_edge_high = av1_upsample_intra_edge_high_sse4_1;
+    av1_warp_affine = av1_warp_affine_c;
+    if (flags & HAS_SSE4_1) av1_warp_affine = av1_warp_affine_sse4_1;
+    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_filter_block = cdef_filter_block_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_block = cdef_filter_block_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_block = cdef_filter_block_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_block = cdef_filter_block_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;
+    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;
+    copy_rect8_16bit_to_16bit = copy_rect8_16bit_to_16bit_sse2;
+    if (flags & HAS_SSSE3) copy_rect8_16bit_to_16bit = copy_rect8_16bit_to_16bit_ssse3;
+    if (flags & HAS_SSE4_1) copy_rect8_16bit_to_16bit = copy_rect8_16bit_to_16bit_sse4_1;
+    if (flags & HAS_AVX2) copy_rect8_16bit_to_16bit = copy_rect8_16bit_to_16bit_avx2;
+    copy_rect8_8bit_to_16bit = copy_rect8_8bit_to_16bit_sse2;
+    if (flags & HAS_SSSE3) copy_rect8_8bit_to_16bit = copy_rect8_8bit_to_16bit_ssse3;
+    if (flags & HAS_SSE4_1) copy_rect8_8bit_to_16bit = copy_rect8_8bit_to_16bit_sse4_1;
+    if (flags & HAS_AVX2) copy_rect8_8bit_to_16bit = copy_rect8_8bit_to_16bit_avx2;
+    get_predict_hbd_fn = get_predict_hbd_fn_c;
+    if (flags & HAS_SSSE3) get_predict_hbd_fn = get_predict_hbd_fn_ssse3;
+    if (flags & HAS_AVX2) get_predict_hbd_fn = get_predict_hbd_fn_avx2;
+    get_predict_lbd_fn = get_predict_lbd_fn_c;
+    if (flags & HAS_SSSE3) get_predict_lbd_fn = get_predict_lbd_fn_ssse3;
+    if (flags & HAS_AVX2) get_predict_lbd_fn = get_predict_lbd_fn_avx2;
+    get_subtract_average_fn = get_subtract_average_fn_sse2;
+    if (flags & HAS_AVX2) get_subtract_average_fn = get_subtract_average_fn_avx2;
+}
+#endif
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif