Merging upstream version 126.0.

Signed-off-by: Daniel Baumann <daniel.baumann@progress-linux.org>
author: Daniel Baumann <daniel.baumann@progress-linux.org> 2024-05-15 03:35:49 +0000
committer: Daniel Baumann <daniel.baumann@progress-linux.org> 2024-05-15 03:35:49 +0000
commit: d8bbc7858622b6d9c278469aab701ca0b609cddf (patch)
tree: eff41dc61d9f714852212739e6b3738b82a2af87 /media/libaom
parent: Releasing progress-linux version 125.0.3-1~progress7.99u1. (diff)
download: firefox-d8bbc7858622b6d9c278469aab701ca0b609cddf.tar.xz
firefox-d8bbc7858622b6d9c278469aab701ca0b609cddf.zip
37 files changed, 526 insertions, 775 deletions
diff --git a/media/libaom/0001-errno.patch b/media/libaom/0001-errno.patch
new file mode 100644
index 0000000000..6040c42e38
--- /dev/null
+++ b/media/libaom/0001-errno.patch
@@ -0,0 +1,22 @@
+diff --git a/aom_util/aom_pthread.h b/aom/aom_util/aom_pthread.h
+--- a/aom_util/aom_pthread.h
++++ b/aom_util/aom_pthread.h
+@@ -30,16 +30,18 @@ extern "C" {
+ #define WIN32_LEAN_AND_MEAN
+ #include <process.h>  // NOLINT
+ #include <stddef.h>   // NOLINT
+ #include <windows.h>  // NOLINT
+ typedef HANDLE pthread_t;
+ typedef int pthread_attr_t;
+ typedef CRITICAL_SECTION pthread_mutex_t;
+ 
++#include <errno.h>
++
+ #if _WIN32_WINNT < 0x0600
+ #error _WIN32_WINNT must target Windows Vista / Server 2008 or newer.
+ #endif
+ typedef CONDITION_VARIABLE pthread_cond_t;
+ 
+ #ifndef WINAPI_FAMILY_PARTITION
+ #define WINAPI_PARTITION_DESKTOP 1
+ #define WINAPI_FAMILY_PARTITION(x) x
diff --git a/media/libaom/0002-mmloadusi64.patch b/media/libaom/0002-mmloadusi64.patch
new file mode 100644
index 0000000000..9d23c90f22
--- /dev/null
+++ b/media/libaom/0002-mmloadusi64.patch
@@ -0,0 +1,79 @@
+diff --git a/aom_dsp/x86/synonyms.h b/aom_dsp/x86/synonyms.h
+--- a/aom_dsp/x86/synonyms.h
++++ b/aom_dsp/x86/synonyms.h
+@@ -41,22 +41,35 @@ static INLINE __m128i xx_loadl_64(const 
+ static INLINE __m128i xx_load_128(const void *a) {
+   return _mm_load_si128((const __m128i *)a);
+ }
+ 
+ static INLINE __m128i xx_loadu_128(const void *a) {
+   return _mm_loadu_si128((const __m128i *)a);
+ }
+ 
++
++// _mm_loadu_si64 has been introduced in GCC 9, reimplement the function
++// manually on older compilers.
++#if !defined(__clang__) && __GNUC_MAJOR__ < 9
++static INLINE __m128i xx_loadu_2x64(const void *hi, const void *lo) {
++  __m64 hi_, lo_;
++  memcpy(&hi_, hi, sizeof(hi_));
++  memcpy(&lo_, lo, sizeof(lo_));
++  return _mm_set_epi64(hi_, lo_);
++}
++#endif
++#else
+ // Load 64 bits from each of hi and low, and pack into an SSE register
+ // Since directly loading as `int64_t`s and using _mm_set_epi64 may violate
+ // the strict aliasing rule, this takes a different approach
+ static INLINE __m128i xx_loadu_2x64(const void *hi, const void *lo) {
+   return _mm_unpacklo_epi64(_mm_loadu_si64(lo), _mm_loadu_si64(hi));
+ }
++#endif
+ 
+ static INLINE void xx_storel_32(void *const a, const __m128i v) {
+   const int val = _mm_cvtsi128_si32(v);
+   memcpy(a, &val, sizeof(val));
+ }
+ 
+ static INLINE void xx_storel_64(void *const a, const __m128i v) {
+   _mm_storel_epi64((__m128i *)a, v);
+diff --git a/aom_dsp/x86/synonyms_avx2.h b/aom_dsp/x86/synonyms_avx2.h
+--- a/aom_dsp/x86/synonyms_avx2.h
++++ b/aom_dsp/x86/synonyms_avx2.h
+@@ -66,21 +66,36 @@ static INLINE __m256i yy_set1_64_from_32
+ 
+ // Some compilers don't have _mm256_set_m128i defined in immintrin.h. We
+ // therefore define an equivalent function using a different intrinsic.
+ // ([ hi ], [ lo ]) -> [ hi ][ lo ]
+ static INLINE __m256i yy_set_m128i(__m128i hi, __m128i lo) {
+   return _mm256_insertf128_si256(_mm256_castsi128_si256(lo), hi, 1);
+ }
+ 
++#define GCC_VERSION (__GNUC__ * 10000 \
++                     + __GNUC_MINOR__ * 100 \
++                     + __GNUC_PATCHLEVEL__)
++
++// _mm256_loadu2_m128i has been introduced in GCC 10.1
++#if !defined(__clang__) && GCC_VERSION < 101000
++static INLINE __m256i yy_loadu2_128(const void *hi, const void *lo) {
++  __m128i mhi = _mm_loadu_si128((const __m128i *)(hi));
++  __m128i mlo = _mm_loadu_si128((const __m128i *)(lo));
++  return _mm256_set_m128i(mhi, mlo);
++}
++#else
+ static INLINE __m256i yy_loadu2_128(const void *hi, const void *lo) {
+   __m128i mhi = _mm_loadu_si128((const __m128i *)(hi));
+   __m128i mlo = _mm_loadu_si128((const __m128i *)(lo));
+   return yy_set_m128i(mhi, mlo);
+ }
++#endif
++
++#undef GCC_VERSION
+ 
+ static INLINE void yy_storeu2_128(void *hi, void *lo, const __m256i a) {
+   _mm_storeu_si128((__m128i *)hi, _mm256_extracti128_si256(a, 1));
+   _mm_storeu_si128((__m128i *)lo, _mm256_castsi256_si128(a));
+ }
+ 
+ static INLINE __m256i yy_roundn_epu16(__m256i v_val_w, int bits) {
+   const __m256i v_s_w = _mm256_srli_epi16(v_val_w, bits - 1);
diff --git a/media/libaom/config/generic/config/aom_config.asm b/media/libaom/config/generic/config/aom_config.asm
index be0715562c..0f329a7df5 100644
--- a/media/libaom/config/generic/config/aom_config.asm
+++ b/media/libaom/config/generic/config/aom_config.asm
@@ -53,6 +53,7 @@ CONFIG_OS_SUPPORT equ 1
 CONFIG_OUTPUT_FRAME_SIZE equ 0
 CONFIG_PARTITION_SEARCH_ORDER equ 0
 CONFIG_PIC equ 0
+CONFIG_QUANT_MATRIX equ 1
 CONFIG_RATECTRL_LOG equ 0
 CONFIG_RD_COMMAND equ 0
 CONFIG_RD_DEBUG equ 0
@@ -87,6 +88,7 @@ HAVE_SSE4_1 equ 0
 HAVE_SSE4_2 equ 0
 HAVE_SSSE3 equ 0
 HAVE_SVE equ 0
+HAVE_SVE2 equ 0
 HAVE_VSX equ 0
 HAVE_WXWIDGETS equ 0
 STATIC_LINK_JXL equ 0
diff --git a/media/libaom/config/generic/config/aom_config.h b/media/libaom/config/generic/config/aom_config.h
index a695b0b3e6..c89e1d755c 100644
--- a/media/libaom/config/generic/config/aom_config.h
+++ b/media/libaom/config/generic/config/aom_config.h
@@ -55,6 +55,7 @@
 #define CONFIG_OUTPUT_FRAME_SIZE 0
 #define CONFIG_PARTITION_SEARCH_ORDER 0
 #define CONFIG_PIC 0
+#define CONFIG_QUANT_MATRIX 1
 #define CONFIG_RATECTRL_LOG 0
 #define CONFIG_RD_COMMAND 0
 #define CONFIG_RD_DEBUG 0
@@ -89,6 +90,7 @@
 #define HAVE_SSE4_2 0
 #define HAVE_SSSE3 0
 #define HAVE_SVE 0
+#define HAVE_SVE2 0
 #define HAVE_VSX 0
 #define HAVE_WXWIDGETS 0
 #define INLINE inline
diff --git a/media/libaom/config/generic/config/aom_dsp_rtcd.h b/media/libaom/config/generic/config/aom_dsp_rtcd.h
index 0418b3568e..a61dc47a47 100644
--- a/media/libaom/config/generic/config/aom_dsp_rtcd.h
+++ b/media/libaom/config/generic/config/aom_dsp_rtcd.h
@@ -46,9 +46,15 @@ void aom_comp_avg_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int
 void aom_comp_mask_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
 #define aom_comp_mask_pred aom_comp_mask_pred_c
 
+double aom_compute_correlation_c(const unsigned char *frame1, int stride1, int x1, int y1, double mean1, double one_over_stddev1, const unsigned char *frame2, int stride2, int x2, int y2, double mean2, double one_over_stddev2);
+#define aom_compute_correlation aom_compute_correlation_c
+
 void aom_compute_flow_at_point_c(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
 #define aom_compute_flow_at_point aom_compute_flow_at_point_c
 
+bool aom_compute_mean_stddev_c(const unsigned char *frame, int stride, int x, int y, double *mean, double *one_over_stddev);
+#define aom_compute_mean_stddev aom_compute_mean_stddev_c
+
 void aom_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h);
 #define aom_convolve8 aom_convolve8_c
 
@@ -4693,9 +4699,6 @@ unsigned int aom_variance8x8_c(const uint8_t *src_ptr, int source_stride, const
 int aom_vector_var_c(const int16_t *ref, const int16_t *src, int bwl);
 #define aom_vector_var aom_vector_var_c
 
-double av1_compute_cross_correlation_c(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
-#define av1_compute_cross_correlation av1_compute_cross_correlation_c
-
 void aom_dsp_rtcd(void);
 
 #include "config/aom_config.h"
diff --git a/media/libaom/config/generic/config/aom_scale_rtcd.h b/media/libaom/config/generic/config/aom_scale_rtcd.h
index 733b2d9ea1..dd09c4e3a6 100644
--- a/media/libaom/config/generic/config/aom_scale_rtcd.h
+++ b/media/libaom/config/generic/config/aom_scale_rtcd.h
@@ -8,13 +8,15 @@
 #define RTCD_EXTERN extern
 #endif
 
+#include <stdbool.h>
+
 struct yv12_buffer_config;
 
 #ifdef __cplusplus
 extern "C" {
 #endif
 
-void aom_extend_frame_borders_c(struct yv12_buffer_config *ybf, const int num_planes);
+void aom_extend_frame_borders_c(struct yv12_buffer_config *ybf, int num_planes);
 #define aom_extend_frame_borders aom_extend_frame_borders_c
 
 void aom_extend_frame_borders_plane_row_c(const struct yv12_buffer_config *ybf, int plane, int v_start, int v_end);
@@ -50,13 +52,13 @@ void aom_vertical_band_5_4_scale_c(unsigned char *source, int src_pitch, unsigne
 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);
+void aom_yv12_copy_u_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc, int use_crop);
 #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);
+void aom_yv12_copy_v_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc, int use_crop);
 #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);
+void aom_yv12_copy_y_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc, int use_crop);
 #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);
@@ -80,7 +82,7 @@ void aom_yv12_partial_copy_v_c(const struct yv12_buffer_config *src_bc, int hsta
 void aom_yv12_partial_copy_y_c(const struct yv12_buffer_config *src_ybc, int hstart1, int hend1, int vstart1, int vend1, struct yv12_buffer_config *dst_ybc, int hstart2, int vstart2);
 #define aom_yv12_partial_copy_y aom_yv12_partial_copy_y_c
 
-int aom_yv12_realloc_with_new_border_c(struct yv12_buffer_config *ybf, int new_border, int byte_alignment, int num_pyramid_levels, int num_planes);
+int aom_yv12_realloc_with_new_border_c(struct yv12_buffer_config *ybf, int new_border, int byte_alignment, bool alloc_pyramid, int num_planes);
 #define aom_yv12_realloc_with_new_border aom_yv12_realloc_with_new_border_c
 
 void aom_scale_rtcd(void);
diff --git a/media/libaom/config/linux/arm/config/aom_config.asm b/media/libaom/config/linux/arm/config/aom_config.asm
index 63034fd7e2..1ec673f263 100644
--- a/media/libaom/config/linux/arm/config/aom_config.asm
+++ b/media/libaom/config/linux/arm/config/aom_config.asm
@@ -53,6 +53,7 @@
 .equ CONFIG_OUTPUT_FRAME_SIZE, 0
 .equ CONFIG_PARTITION_SEARCH_ORDER, 0
 .equ CONFIG_PIC, 1
+.equ CONFIG_QUANT_MATRIX, 1
 .equ CONFIG_RATECTRL_LOG, 0
 .equ CONFIG_RD_COMMAND, 0
 .equ CONFIG_RD_DEBUG, 0
@@ -87,6 +88,7 @@
 .equ HAVE_SSE4_2, 0
 .equ HAVE_SSSE3, 0
 .equ HAVE_SVE, 0
+.equ HAVE_SVE2, 0
 .equ HAVE_VSX, 0
 .equ HAVE_WXWIDGETS, 0
 .equ STATIC_LINK_JXL, 0
diff --git a/media/libaom/config/linux/arm/config/aom_config.h b/media/libaom/config/linux/arm/config/aom_config.h
index 3cbe7bf169..fb73e8431e 100644
--- a/media/libaom/config/linux/arm/config/aom_config.h
+++ b/media/libaom/config/linux/arm/config/aom_config.h
@@ -55,6 +55,7 @@
 #define CONFIG_OUTPUT_FRAME_SIZE 0
 #define CONFIG_PARTITION_SEARCH_ORDER 0
 #define CONFIG_PIC 1
+#define CONFIG_QUANT_MATRIX 1
 #define CONFIG_RATECTRL_LOG 0
 #define CONFIG_RD_COMMAND 0
 #define CONFIG_RD_DEBUG 0
@@ -89,6 +90,7 @@
 #define HAVE_SSE4_2 0
 #define HAVE_SSSE3 0
 #define HAVE_SVE 0
+#define HAVE_SVE2 0
 #define HAVE_VSX 0
 #define HAVE_WXWIDGETS 0
 #define INLINE inline
diff --git a/media/libaom/config/linux/arm/config/aom_dsp_rtcd.h b/media/libaom/config/linux/arm/config/aom_dsp_rtcd.h
index 50ee78932c..fffcc5a3e9 100644
--- a/media/libaom/config/linux/arm/config/aom_dsp_rtcd.h
+++ b/media/libaom/config/linux/arm/config/aom_dsp_rtcd.h
@@ -54,10 +54,16 @@ void aom_comp_mask_pred_c(uint8_t *comp_pred, const uint8_t *pred, int width, in
 void aom_comp_mask_pred_neon(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
 RTCD_EXTERN void (*aom_comp_mask_pred)(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
 
+double aom_compute_correlation_c(const unsigned char *frame1, int stride1, int x1, int y1, double mean1, double one_over_stddev1, const unsigned char *frame2, int stride2, int x2, int y2, double mean2, double one_over_stddev2);
+#define aom_compute_correlation aom_compute_correlation_c
+
 void aom_compute_flow_at_point_c(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
 void aom_compute_flow_at_point_neon(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
 RTCD_EXTERN void (*aom_compute_flow_at_point)(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
 
+bool aom_compute_mean_stddev_c(const unsigned char *frame, int stride, int x, int y, double *mean, double *one_over_stddev);
+#define aom_compute_mean_stddev aom_compute_mean_stddev_c
+
 void aom_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h);
 #define aom_convolve8 aom_convolve8_c
 
@@ -6212,9 +6218,6 @@ int aom_vector_var_c(const int16_t *ref, const int16_t *src, int bwl);
 int aom_vector_var_neon(const int16_t *ref, const int16_t *src, int bwl);
 RTCD_EXTERN int (*aom_vector_var)(const int16_t *ref, const int16_t *src, int bwl);
 
-double av1_compute_cross_correlation_c(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
-#define av1_compute_cross_correlation av1_compute_cross_correlation_c
-
 void aom_dsp_rtcd(void);
 
 #include "config/aom_config.h"
diff --git a/media/libaom/config/linux/arm/config/aom_scale_rtcd.h b/media/libaom/config/linux/arm/config/aom_scale_rtcd.h
index d296957f84..1024a666fe 100644
--- a/media/libaom/config/linux/arm/config/aom_scale_rtcd.h
+++ b/media/libaom/config/linux/arm/config/aom_scale_rtcd.h
@@ -8,13 +8,15 @@
 #define RTCD_EXTERN extern
 #endif
 
+#include <stdbool.h>
+
 struct yv12_buffer_config;
 
 #ifdef __cplusplus
 extern "C" {
 #endif
 
-void aom_extend_frame_borders_c(struct yv12_buffer_config *ybf, const int num_planes);
+void aom_extend_frame_borders_c(struct yv12_buffer_config *ybf, int num_planes);
 #define aom_extend_frame_borders aom_extend_frame_borders_c
 
 void aom_extend_frame_borders_plane_row_c(const struct yv12_buffer_config *ybf, int plane, int v_start, int v_end);
@@ -50,13 +52,13 @@ void aom_vertical_band_5_4_scale_c(unsigned char *source, int src_pitch, unsigne
 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);
+void aom_yv12_copy_u_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc, int use_crop);
 #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);
+void aom_yv12_copy_v_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc, int use_crop);
 #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);
+void aom_yv12_copy_y_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc, int use_crop);
 #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);
@@ -80,7 +82,7 @@ void aom_yv12_partial_copy_v_c(const struct yv12_buffer_config *src_bc, int hsta
 void aom_yv12_partial_copy_y_c(const struct yv12_buffer_config *src_ybc, int hstart1, int hend1, int vstart1, int vend1, struct yv12_buffer_config *dst_ybc, int hstart2, int vstart2);
 #define aom_yv12_partial_copy_y aom_yv12_partial_copy_y_c
 
-int aom_yv12_realloc_with_new_border_c(struct yv12_buffer_config *ybf, int new_border, int byte_alignment, int num_pyramid_levels, int num_planes);
+int aom_yv12_realloc_with_new_border_c(struct yv12_buffer_config *ybf, int new_border, int byte_alignment, bool alloc_pyramid, int num_planes);
 #define aom_yv12_realloc_with_new_border aom_yv12_realloc_with_new_border_c
 
 void aom_scale_rtcd(void);
diff --git a/media/libaom/config/linux/ia32/config/aom_config.asm b/media/libaom/config/linux/ia32/config/aom_config.asm
index e75260cb09..4fd596e34b 100644
--- a/media/libaom/config/linux/ia32/config/aom_config.asm
+++ b/media/libaom/config/linux/ia32/config/aom_config.asm
@@ -53,6 +53,7 @@ CONFIG_OS_SUPPORT equ 1
 CONFIG_OUTPUT_FRAME_SIZE equ 0
 CONFIG_PARTITION_SEARCH_ORDER equ 0
 CONFIG_PIC equ 1
+CONFIG_QUANT_MATRIX equ 1
 CONFIG_RATECTRL_LOG equ 0
 CONFIG_RD_COMMAND equ 0
 CONFIG_RD_DEBUG equ 0
@@ -87,6 +88,7 @@ HAVE_SSE4_1 equ 1
 HAVE_SSE4_2 equ 1
 HAVE_SSSE3 equ 1
 HAVE_SVE equ 0
+HAVE_SVE2 equ 0
 HAVE_VSX equ 0
 HAVE_WXWIDGETS equ 0
 STATIC_LINK_JXL equ 0
diff --git a/media/libaom/config/linux/ia32/config/aom_config.h b/media/libaom/config/linux/ia32/config/aom_config.h
index b0e5b5cabc..256f556662 100644
--- a/media/libaom/config/linux/ia32/config/aom_config.h
+++ b/media/libaom/config/linux/ia32/config/aom_config.h
@@ -55,6 +55,7 @@
 #define CONFIG_OUTPUT_FRAME_SIZE 0
 #define CONFIG_PARTITION_SEARCH_ORDER 0
 #define CONFIG_PIC 1
+#define CONFIG_QUANT_MATRIX 1
 #define CONFIG_RATECTRL_LOG 0
 #define CONFIG_RD_COMMAND 0
 #define CONFIG_RD_DEBUG 0
@@ -89,6 +90,7 @@
 #define HAVE_SSE4_2 1
 #define HAVE_SSSE3 1
 #define HAVE_SVE 0
+#define HAVE_SVE2 0
 #define HAVE_VSX 0
 #define HAVE_WXWIDGETS 0
 #define INLINE inline
diff --git a/media/libaom/config/linux/ia32/config/aom_dsp_rtcd.h b/media/libaom/config/linux/ia32/config/aom_dsp_rtcd.h
index a19adf5f61..93472f0e92 100644
--- a/media/libaom/config/linux/ia32/config/aom_dsp_rtcd.h
+++ b/media/libaom/config/linux/ia32/config/aom_dsp_rtcd.h
@@ -57,21 +57,30 @@ void aom_comp_mask_pred_ssse3(uint8_t *comp_pred, const uint8_t *pred, int width
 void aom_comp_mask_pred_avx2(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
 RTCD_EXTERN void (*aom_comp_mask_pred)(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
 
+double aom_compute_correlation_c(const unsigned char *frame1, int stride1, int x1, int y1, double mean1, double one_over_stddev1, const unsigned char *frame2, int stride2, int x2, int y2, double mean2, double one_over_stddev2);
+double aom_compute_correlation_sse4_1(const unsigned char *frame1, int stride1, int x1, int y1, double mean1, double one_over_stddev1, const unsigned char *frame2, int stride2, int x2, int y2, double mean2, double one_over_stddev2);
+double aom_compute_correlation_avx2(const unsigned char *frame1, int stride1, int x1, int y1, double mean1, double one_over_stddev1, const unsigned char *frame2, int stride2, int x2, int y2, double mean2, double one_over_stddev2);
+RTCD_EXTERN double (*aom_compute_correlation)(const unsigned char *frame1, int stride1, int x1, int y1, double mean1, double one_over_stddev1, const unsigned char *frame2, int stride2, int x2, int y2, double mean2, double one_over_stddev2);
+
 void aom_compute_flow_at_point_c(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
 void aom_compute_flow_at_point_sse4_1(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
+void aom_compute_flow_at_point_avx2(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
 RTCD_EXTERN void (*aom_compute_flow_at_point)(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
 
+bool aom_compute_mean_stddev_c(const unsigned char *frame, int stride, int x, int y, double *mean, double *one_over_stddev);
+bool aom_compute_mean_stddev_sse4_1(const unsigned char *frame, int stride, int x, int y, double *mean, double *one_over_stddev);
+bool aom_compute_mean_stddev_avx2(const unsigned char *frame, int stride, int x, int y, double *mean, double *one_over_stddev);
+RTCD_EXTERN bool (*aom_compute_mean_stddev)(const unsigned char *frame, int stride, int x, int y, double *mean, double *one_over_stddev);
+
 void aom_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h);
 #define aom_convolve8 aom_convolve8_c
 
 void aom_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
-void aom_convolve8_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void aom_convolve8_horiz_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void aom_convolve8_horiz_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 RTCD_EXTERN void (*aom_convolve8_horiz)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 
 void aom_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
-void aom_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void aom_convolve8_vert_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void aom_convolve8_vert_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 RTCD_EXTERN void (*aom_convolve8_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
@@ -903,6 +912,7 @@ RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance8x8)(const uin
 
 unsigned int aom_highbd_10_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
 unsigned int aom_highbd_10_mse16x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_10_mse16x16_avx2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
 RTCD_EXTERN unsigned int (*aom_highbd_10_mse16x16)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
 
 unsigned int aom_highbd_10_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
@@ -5130,7 +5140,8 @@ unsigned int aom_sad16x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const
 RTCD_EXTERN unsigned int (*aom_sad16x4_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 
 void aom_sad16x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
-#define aom_sad16x4x3d aom_sad16x4x3d_c
+void aom_sad16x4x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x4x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
 
 void aom_sad16x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
 void aom_sad16x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
@@ -5466,7 +5477,8 @@ unsigned int aom_sad_skip_16x4_c(const uint8_t *src_ptr, int src_stride, const u
 #define aom_sad_skip_16x4 aom_sad_skip_16x4_c
 
 void aom_sad_skip_16x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
-#define aom_sad_skip_16x4x4d aom_sad_skip_16x4x4d_c
+void aom_sad_skip_16x4x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_16x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
 
 unsigned int aom_sad_skip_16x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
 unsigned int aom_sad_skip_16x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
@@ -5867,243 +5879,199 @@ void aom_ssim_parms_8x8_c(const uint8_t *s, int sp, const uint8_t *r, int rp, ui
 #define aom_ssim_parms_8x8 aom_ssim_parms_8x8_c
 
 uint32_t aom_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance128x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance128x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance128x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance128x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance16x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance16x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance16x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance16x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance16x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance32x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance32x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance32x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance32x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance32x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance32x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance32x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance4x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance4x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance4x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance4x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance4x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance4x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance4x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance64x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance64x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance64x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance64x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance64x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance64x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance64x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance8x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance8x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance8x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance8x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance128x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance128x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance128x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance128x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x4_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x8_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance32x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance32x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance32x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance32x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance32x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance32x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance32x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance4x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance4x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance4x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance4x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance4x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance4x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance4x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance64x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance64x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance64x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance64x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance64x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance64x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance64x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance8x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance8x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance8x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance8x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
@@ -6326,11 +6294,6 @@ int aom_vector_var_sse4_1(const int16_t *ref, const int16_t *src, int bwl);
 int aom_vector_var_avx2(const int16_t *ref, const int16_t *src, int bwl);
 RTCD_EXTERN int (*aom_vector_var)(const int16_t *ref, const int16_t *src, int bwl);
 
-double av1_compute_cross_correlation_c(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
-double av1_compute_cross_correlation_sse4_1(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
-double av1_compute_cross_correlation_avx2(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
-RTCD_EXTERN double (*av1_compute_cross_correlation)(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
-
 void aom_dsp_rtcd(void);
 
 #ifdef RTCD_C
@@ -6360,14 +6323,19 @@ static void setup_rtcd_internal(void)
     aom_comp_mask_pred = aom_comp_mask_pred_c;
     if (flags & HAS_SSSE3) aom_comp_mask_pred = aom_comp_mask_pred_ssse3;
     if (flags & HAS_AVX2) aom_comp_mask_pred = aom_comp_mask_pred_avx2;
+    aom_compute_correlation = aom_compute_correlation_c;
+    if (flags & HAS_SSE4_1) aom_compute_correlation = aom_compute_correlation_sse4_1;
+    if (flags & HAS_AVX2) aom_compute_correlation = aom_compute_correlation_avx2;
     aom_compute_flow_at_point = aom_compute_flow_at_point_c;
     if (flags & HAS_SSE4_1) aom_compute_flow_at_point = aom_compute_flow_at_point_sse4_1;
+    if (flags & HAS_AVX2) aom_compute_flow_at_point = aom_compute_flow_at_point_avx2;
+    aom_compute_mean_stddev = aom_compute_mean_stddev_c;
+    if (flags & HAS_SSE4_1) aom_compute_mean_stddev = aom_compute_mean_stddev_sse4_1;
+    if (flags & HAS_AVX2) aom_compute_mean_stddev = aom_compute_mean_stddev_avx2;
     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;
@@ -6768,6 +6736,7 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance8x8 = aom_highbd_10_masked_sub_pixel_variance8x8_ssse3;
     aom_highbd_10_mse16x16 = aom_highbd_10_mse16x16_c;
     if (flags & HAS_SSE2) aom_highbd_10_mse16x16 = aom_highbd_10_mse16x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_mse16x16 = aom_highbd_10_mse16x16_avx2;
     aom_highbd_10_mse8x8 = aom_highbd_10_mse8x8_c;
     if (flags & HAS_SSE2) aom_highbd_10_mse8x8 = aom_highbd_10_mse8x8_sse2;
     aom_highbd_10_obmc_variance128x128 = aom_highbd_10_obmc_variance128x128_c;
@@ -8526,6 +8495,8 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_SSE2) aom_sad16x4 = aom_sad16x4_sse2;
     aom_sad16x4_avg = aom_sad16x4_avg_c;
     if (flags & HAS_SSE2) aom_sad16x4_avg = aom_sad16x4_avg_sse2;
+    aom_sad16x4x3d = aom_sad16x4x3d_c;
+    if (flags & HAS_AVX2) aom_sad16x4x3d = aom_sad16x4x3d_avx2;
     aom_sad16x4x4d = aom_sad16x4x4d_c;
     if (flags & HAS_SSE2) aom_sad16x4x4d = aom_sad16x4x4d_sse2;
     if (flags & HAS_AVX2) aom_sad16x4x4d = aom_sad16x4x4d_avx2;
@@ -8695,6 +8666,8 @@ static void setup_rtcd_internal(void)
     aom_sad_skip_16x32x4d = aom_sad_skip_16x32x4d_c;
     if (flags & HAS_SSE2) aom_sad_skip_16x32x4d = aom_sad_skip_16x32x4d_sse2;
     if (flags & HAS_AVX2) aom_sad_skip_16x32x4d = aom_sad_skip_16x32x4d_avx2;
+    aom_sad_skip_16x4x4d = aom_sad_skip_16x4x4d_c;
+    if (flags & HAS_AVX2) aom_sad_skip_16x4x4d = aom_sad_skip_16x4x4d_avx2;
     aom_sad_skip_16x64 = aom_sad_skip_16x64_c;
     if (flags & HAS_SSE2) aom_sad_skip_16x64 = aom_sad_skip_16x64_sse2;
     aom_sad_skip_16x64x4d = aom_sad_skip_16x64x4d_c;
@@ -8897,157 +8870,113 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_SSE4_1) aom_sse = aom_sse_sse4_1;
     if (flags & HAS_AVX2) aom_sse = aom_sse_avx2;
     aom_sub_pixel_avg_variance128x128 = aom_sub_pixel_avg_variance128x128_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance128x128 = aom_sub_pixel_avg_variance128x128_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance128x128 = aom_sub_pixel_avg_variance128x128_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_avg_variance128x128 = aom_sub_pixel_avg_variance128x128_avx2;
     aom_sub_pixel_avg_variance128x64 = aom_sub_pixel_avg_variance128x64_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance128x64 = aom_sub_pixel_avg_variance128x64_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance128x64 = aom_sub_pixel_avg_variance128x64_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_avg_variance128x64 = aom_sub_pixel_avg_variance128x64_avx2;
     aom_sub_pixel_avg_variance16x16 = aom_sub_pixel_avg_variance16x16_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance16x16 = aom_sub_pixel_avg_variance16x16_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x16 = aom_sub_pixel_avg_variance16x16_ssse3;
     aom_sub_pixel_avg_variance16x32 = aom_sub_pixel_avg_variance16x32_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance16x32 = aom_sub_pixel_avg_variance16x32_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x32 = aom_sub_pixel_avg_variance16x32_ssse3;
     aom_sub_pixel_avg_variance16x4 = aom_sub_pixel_avg_variance16x4_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance16x4 = aom_sub_pixel_avg_variance16x4_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x4 = aom_sub_pixel_avg_variance16x4_ssse3;
     aom_sub_pixel_avg_variance16x64 = aom_sub_pixel_avg_variance16x64_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance16x64 = aom_sub_pixel_avg_variance16x64_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x64 = aom_sub_pixel_avg_variance16x64_ssse3;
     aom_sub_pixel_avg_variance16x8 = aom_sub_pixel_avg_variance16x8_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance16x8 = aom_sub_pixel_avg_variance16x8_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x8 = aom_sub_pixel_avg_variance16x8_ssse3;
     aom_sub_pixel_avg_variance32x16 = aom_sub_pixel_avg_variance32x16_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance32x16 = aom_sub_pixel_avg_variance32x16_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x16 = aom_sub_pixel_avg_variance32x16_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_avg_variance32x16 = aom_sub_pixel_avg_variance32x16_avx2;
     aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_avx2;
     aom_sub_pixel_avg_variance32x64 = aom_sub_pixel_avg_variance32x64_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance32x64 = aom_sub_pixel_avg_variance32x64_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x64 = aom_sub_pixel_avg_variance32x64_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_avg_variance32x64 = aom_sub_pixel_avg_variance32x64_avx2;
     aom_sub_pixel_avg_variance32x8 = aom_sub_pixel_avg_variance32x8_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance32x8 = aom_sub_pixel_avg_variance32x8_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x8 = aom_sub_pixel_avg_variance32x8_ssse3;
     aom_sub_pixel_avg_variance4x16 = aom_sub_pixel_avg_variance4x16_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance4x16 = aom_sub_pixel_avg_variance4x16_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance4x16 = aom_sub_pixel_avg_variance4x16_ssse3;
     aom_sub_pixel_avg_variance4x4 = aom_sub_pixel_avg_variance4x4_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance4x4 = aom_sub_pixel_avg_variance4x4_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance4x4 = aom_sub_pixel_avg_variance4x4_ssse3;
     aom_sub_pixel_avg_variance4x8 = aom_sub_pixel_avg_variance4x8_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance4x8 = aom_sub_pixel_avg_variance4x8_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance4x8 = aom_sub_pixel_avg_variance4x8_ssse3;
     aom_sub_pixel_avg_variance64x128 = aom_sub_pixel_avg_variance64x128_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance64x128 = aom_sub_pixel_avg_variance64x128_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance64x128 = aom_sub_pixel_avg_variance64x128_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_avg_variance64x128 = aom_sub_pixel_avg_variance64x128_avx2;
     aom_sub_pixel_avg_variance64x16 = aom_sub_pixel_avg_variance64x16_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance64x16 = aom_sub_pixel_avg_variance64x16_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance64x16 = aom_sub_pixel_avg_variance64x16_ssse3;
     aom_sub_pixel_avg_variance64x32 = aom_sub_pixel_avg_variance64x32_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance64x32 = aom_sub_pixel_avg_variance64x32_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance64x32 = aom_sub_pixel_avg_variance64x32_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_avg_variance64x32 = aom_sub_pixel_avg_variance64x32_avx2;
     aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_avx2;
     aom_sub_pixel_avg_variance8x16 = aom_sub_pixel_avg_variance8x16_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance8x16 = aom_sub_pixel_avg_variance8x16_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x16 = aom_sub_pixel_avg_variance8x16_ssse3;
     aom_sub_pixel_avg_variance8x32 = aom_sub_pixel_avg_variance8x32_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance8x32 = aom_sub_pixel_avg_variance8x32_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x32 = aom_sub_pixel_avg_variance8x32_ssse3;
     aom_sub_pixel_avg_variance8x4 = aom_sub_pixel_avg_variance8x4_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance8x4 = aom_sub_pixel_avg_variance8x4_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x4 = aom_sub_pixel_avg_variance8x4_ssse3;
     aom_sub_pixel_avg_variance8x8 = aom_sub_pixel_avg_variance8x8_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance8x8 = aom_sub_pixel_avg_variance8x8_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x8 = aom_sub_pixel_avg_variance8x8_ssse3;
     aom_sub_pixel_variance128x128 = aom_sub_pixel_variance128x128_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance128x128 = aom_sub_pixel_variance128x128_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance128x128 = aom_sub_pixel_variance128x128_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance128x128 = aom_sub_pixel_variance128x128_avx2;
     aom_sub_pixel_variance128x64 = aom_sub_pixel_variance128x64_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance128x64 = aom_sub_pixel_variance128x64_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance128x64 = aom_sub_pixel_variance128x64_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance128x64 = aom_sub_pixel_variance128x64_avx2;
     aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_avx2;
     aom_sub_pixel_variance16x32 = aom_sub_pixel_variance16x32_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance16x32 = aom_sub_pixel_variance16x32_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance16x32 = aom_sub_pixel_variance16x32_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance16x32 = aom_sub_pixel_variance16x32_avx2;
     aom_sub_pixel_variance16x4 = aom_sub_pixel_variance16x4_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance16x4 = aom_sub_pixel_variance16x4_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance16x4 = aom_sub_pixel_variance16x4_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance16x4 = aom_sub_pixel_variance16x4_avx2;
     aom_sub_pixel_variance16x64 = aom_sub_pixel_variance16x64_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance16x64 = aom_sub_pixel_variance16x64_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance16x64 = aom_sub_pixel_variance16x64_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance16x64 = aom_sub_pixel_variance16x64_avx2;
     aom_sub_pixel_variance16x8 = aom_sub_pixel_variance16x8_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance16x8 = aom_sub_pixel_variance16x8_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance16x8 = aom_sub_pixel_variance16x8_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance16x8 = aom_sub_pixel_variance16x8_avx2;
     aom_sub_pixel_variance32x16 = aom_sub_pixel_variance32x16_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance32x16 = aom_sub_pixel_variance32x16_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance32x16 = aom_sub_pixel_variance32x16_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance32x16 = aom_sub_pixel_variance32x16_avx2;
     aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_avx2;
     aom_sub_pixel_variance32x64 = aom_sub_pixel_variance32x64_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance32x64 = aom_sub_pixel_variance32x64_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance32x64 = aom_sub_pixel_variance32x64_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance32x64 = aom_sub_pixel_variance32x64_avx2;
     aom_sub_pixel_variance32x8 = aom_sub_pixel_variance32x8_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance32x8 = aom_sub_pixel_variance32x8_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance32x8 = aom_sub_pixel_variance32x8_ssse3;
     aom_sub_pixel_variance4x16 = aom_sub_pixel_variance4x16_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance4x16 = aom_sub_pixel_variance4x16_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance4x16 = aom_sub_pixel_variance4x16_ssse3;
     aom_sub_pixel_variance4x4 = aom_sub_pixel_variance4x4_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance4x4 = aom_sub_pixel_variance4x4_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance4x4 = aom_sub_pixel_variance4x4_ssse3;
     aom_sub_pixel_variance4x8 = aom_sub_pixel_variance4x8_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance4x8 = aom_sub_pixel_variance4x8_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance4x8 = aom_sub_pixel_variance4x8_ssse3;
     aom_sub_pixel_variance64x128 = aom_sub_pixel_variance64x128_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance64x128 = aom_sub_pixel_variance64x128_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance64x128 = aom_sub_pixel_variance64x128_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance64x128 = aom_sub_pixel_variance64x128_avx2;
     aom_sub_pixel_variance64x16 = aom_sub_pixel_variance64x16_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance64x16 = aom_sub_pixel_variance64x16_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance64x16 = aom_sub_pixel_variance64x16_ssse3;
     aom_sub_pixel_variance64x32 = aom_sub_pixel_variance64x32_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance64x32 = aom_sub_pixel_variance64x32_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance64x32 = aom_sub_pixel_variance64x32_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance64x32 = aom_sub_pixel_variance64x32_avx2;
     aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_avx2;
     aom_sub_pixel_variance8x16 = aom_sub_pixel_variance8x16_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance8x16 = aom_sub_pixel_variance8x16_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance8x16 = aom_sub_pixel_variance8x16_ssse3;
     aom_sub_pixel_variance8x32 = aom_sub_pixel_variance8x32_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance8x32 = aom_sub_pixel_variance8x32_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance8x32 = aom_sub_pixel_variance8x32_ssse3;
     aom_sub_pixel_variance8x4 = aom_sub_pixel_variance8x4_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance8x4 = aom_sub_pixel_variance8x4_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance8x4 = aom_sub_pixel_variance8x4_ssse3;
     aom_sub_pixel_variance8x8 = aom_sub_pixel_variance8x8_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance8x8 = aom_sub_pixel_variance8x8_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance8x8 = aom_sub_pixel_variance8x8_ssse3;
     aom_subtract_block = aom_subtract_block_c;
     if (flags & HAS_SSE2) aom_subtract_block = aom_subtract_block_sse2;
@@ -9172,9 +9101,6 @@ static void setup_rtcd_internal(void)
     aom_vector_var = aom_vector_var_c;
     if (flags & HAS_SSE4_1) aom_vector_var = aom_vector_var_sse4_1;
     if (flags & HAS_AVX2) aom_vector_var = aom_vector_var_avx2;
-    av1_compute_cross_correlation = av1_compute_cross_correlation_c;
-    if (flags & HAS_SSE4_1) av1_compute_cross_correlation = av1_compute_cross_correlation_sse4_1;
-    if (flags & HAS_AVX2) av1_compute_cross_correlation = av1_compute_cross_correlation_avx2;
 }
 #endif
 
diff --git a/media/libaom/config/linux/ia32/config/aom_scale_rtcd.h b/media/libaom/config/linux/ia32/config/aom_scale_rtcd.h
index 3b70fb47c3..cdabb21106 100644
--- a/media/libaom/config/linux/ia32/config/aom_scale_rtcd.h
+++ b/media/libaom/config/linux/ia32/config/aom_scale_rtcd.h
@@ -8,13 +8,15 @@
 #define RTCD_EXTERN extern
 #endif
 
+#include <stdbool.h>
+
 struct yv12_buffer_config;
 
 #ifdef __cplusplus
 extern "C" {
 #endif
 
-void aom_extend_frame_borders_c(struct yv12_buffer_config *ybf, const int num_planes);
+void aom_extend_frame_borders_c(struct yv12_buffer_config *ybf, int num_planes);
 #define aom_extend_frame_borders aom_extend_frame_borders_c
 
 void aom_extend_frame_borders_plane_row_c(const struct yv12_buffer_config *ybf, int plane, int v_start, int v_end);
@@ -50,13 +52,13 @@ void aom_vertical_band_5_4_scale_c(unsigned char *source, int src_pitch, unsigne
 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);
+void aom_yv12_copy_u_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc, int use_crop);
 #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);
+void aom_yv12_copy_v_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc, int use_crop);
 #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);
+void aom_yv12_copy_y_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc, int use_crop);
 #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);
@@ -80,7 +82,7 @@ void aom_yv12_partial_copy_v_c(const struct yv12_buffer_config *src_bc, int hsta
 void aom_yv12_partial_copy_y_c(const struct yv12_buffer_config *src_ybc, int hstart1, int hend1, int vstart1, int vend1, struct yv12_buffer_config *dst_ybc, int hstart2, int vstart2);
 #define aom_yv12_partial_copy_y aom_yv12_partial_copy_y_c
 
-int aom_yv12_realloc_with_new_border_c(struct yv12_buffer_config *ybf, int new_border, int byte_alignment, int num_pyramid_levels, int num_planes);
+int aom_yv12_realloc_with_new_border_c(struct yv12_buffer_config *ybf, int new_border, int byte_alignment, bool alloc_pyramid, int num_planes);
 #define aom_yv12_realloc_with_new_border aom_yv12_realloc_with_new_border_c
 
 void aom_scale_rtcd(void);
diff --git a/media/libaom/config/linux/ia32/config/av1_rtcd.h b/media/libaom/config/linux/ia32/config/av1_rtcd.h
index 3f404f61c8..37716517bf 100644
--- a/media/libaom/config/linux/ia32/config/av1_rtcd.h
+++ b/media/libaom/config/linux/ia32/config/av1_rtcd.h
@@ -265,7 +265,6 @@ void av1_convolve_y_sr_intrabc_c(const uint8_t *src, int src_stride, uint8_t *ds
 #define av1_convolve_y_sr_intrabc av1_convolve_y_sr_intrabc_c
 
 void av1_dist_wtd_convolve_2d_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
-void av1_dist_wtd_convolve_2d_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
 void av1_dist_wtd_convolve_2d_ssse3(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
 void av1_dist_wtd_convolve_2d_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
 RTCD_EXTERN void (*av1_dist_wtd_convolve_2d)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
@@ -764,84 +763,72 @@ void av1_wiener_convolve_add_src_avx2(const uint8_t *src, ptrdiff_t src_stride,
 RTCD_EXTERN void (*av1_wiener_convolve_add_src)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params);
 
 void cdef_copy_rect8_16bit_to_16bit_c(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
-void cdef_copy_rect8_16bit_to_16bit_sse2(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
 void cdef_copy_rect8_16bit_to_16bit_ssse3(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
 void cdef_copy_rect8_16bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
 void cdef_copy_rect8_16bit_to_16bit_avx2(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
 RTCD_EXTERN void (*cdef_copy_rect8_16bit_to_16bit)(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
 
 void cdef_copy_rect8_8bit_to_16bit_c(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
-void cdef_copy_rect8_8bit_to_16bit_sse2(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
 void cdef_copy_rect8_8bit_to_16bit_ssse3(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
 void cdef_copy_rect8_8bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
 void cdef_copy_rect8_8bit_to_16bit_avx2(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
 RTCD_EXTERN void (*cdef_copy_rect8_8bit_to_16bit)(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
 
 void cdef_filter_16_0_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_16_0_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_0_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_0_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_0_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 RTCD_EXTERN void (*cdef_filter_16_0)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 
 void cdef_filter_16_1_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_16_1_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_1_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_1_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_1_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 RTCD_EXTERN void (*cdef_filter_16_1)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 
 void cdef_filter_16_2_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_16_2_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_2_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_2_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_2_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 RTCD_EXTERN void (*cdef_filter_16_2)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 
 void cdef_filter_16_3_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_16_3_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_3_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_3_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_3_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 RTCD_EXTERN void (*cdef_filter_16_3)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 
 void cdef_filter_8_0_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_8_0_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_0_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_0_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_0_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 RTCD_EXTERN void (*cdef_filter_8_0)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 
 void cdef_filter_8_1_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_8_1_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_1_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_1_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_1_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 RTCD_EXTERN void (*cdef_filter_8_1)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 
 void cdef_filter_8_2_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_8_2_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_2_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_2_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_2_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 RTCD_EXTERN void (*cdef_filter_8_2)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 
 void cdef_filter_8_3_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_8_3_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_3_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_3_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_3_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 RTCD_EXTERN void (*cdef_filter_8_3)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 
 int cdef_find_dir_c(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
-int cdef_find_dir_sse2(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
 int cdef_find_dir_ssse3(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
 int cdef_find_dir_sse4_1(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
 int cdef_find_dir_avx2(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
 RTCD_EXTERN int (*cdef_find_dir)(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
 
 void cdef_find_dir_dual_c(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
-void cdef_find_dir_dual_sse2(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
 void cdef_find_dir_dual_ssse3(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
 void cdef_find_dir_dual_sse4_1(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
 void cdef_find_dir_dual_avx2(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
@@ -969,7 +956,6 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_SSE2) av1_convolve_y_sr = av1_convolve_y_sr_sse2;
     if (flags & HAS_AVX2) av1_convolve_y_sr = av1_convolve_y_sr_avx2;
     av1_dist_wtd_convolve_2d = av1_dist_wtd_convolve_2d_c;
-    if (flags & HAS_SSE2) av1_dist_wtd_convolve_2d = av1_dist_wtd_convolve_2d_sse2;
     if (flags & HAS_SSSE3) av1_dist_wtd_convolve_2d = av1_dist_wtd_convolve_2d_ssse3;
     if (flags & HAS_AVX2) av1_dist_wtd_convolve_2d = av1_dist_wtd_convolve_2d_avx2;
     av1_dist_wtd_convolve_2d_copy = av1_dist_wtd_convolve_2d_copy_c;
@@ -1176,62 +1162,50 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_SSE2) av1_wiener_convolve_add_src = av1_wiener_convolve_add_src_sse2;
     if (flags & HAS_AVX2) av1_wiener_convolve_add_src = av1_wiener_convolve_add_src_avx2;
     cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_c;
-    if (flags & HAS_SSE2) cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_sse2;
     if (flags & HAS_SSSE3) cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_ssse3;
     if (flags & HAS_SSE4_1) cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_sse4_1;
     if (flags & HAS_AVX2) cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_avx2;
     cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_c;
-    if (flags & HAS_SSE2) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_sse2;
     if (flags & HAS_SSSE3) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_ssse3;
     if (flags & HAS_SSE4_1) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_sse4_1;
     if (flags & HAS_AVX2) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_avx2;
     cdef_filter_16_0 = cdef_filter_16_0_c;
-    if (flags & HAS_SSE2) cdef_filter_16_0 = cdef_filter_16_0_sse2;
     if (flags & HAS_SSSE3) cdef_filter_16_0 = cdef_filter_16_0_ssse3;
     if (flags & HAS_SSE4_1) cdef_filter_16_0 = cdef_filter_16_0_sse4_1;
     if (flags & HAS_AVX2) cdef_filter_16_0 = cdef_filter_16_0_avx2;
     cdef_filter_16_1 = cdef_filter_16_1_c;
-    if (flags & HAS_SSE2) cdef_filter_16_1 = cdef_filter_16_1_sse2;
     if (flags & HAS_SSSE3) cdef_filter_16_1 = cdef_filter_16_1_ssse3;
     if (flags & HAS_SSE4_1) cdef_filter_16_1 = cdef_filter_16_1_sse4_1;
     if (flags & HAS_AVX2) cdef_filter_16_1 = cdef_filter_16_1_avx2;
     cdef_filter_16_2 = cdef_filter_16_2_c;
-    if (flags & HAS_SSE2) cdef_filter_16_2 = cdef_filter_16_2_sse2;
     if (flags & HAS_SSSE3) cdef_filter_16_2 = cdef_filter_16_2_ssse3;
     if (flags & HAS_SSE4_1) cdef_filter_16_2 = cdef_filter_16_2_sse4_1;
     if (flags & HAS_AVX2) cdef_filter_16_2 = cdef_filter_16_2_avx2;
     cdef_filter_16_3 = cdef_filter_16_3_c;
-    if (flags & HAS_SSE2) cdef_filter_16_3 = cdef_filter_16_3_sse2;
     if (flags & HAS_SSSE3) cdef_filter_16_3 = cdef_filter_16_3_ssse3;
     if (flags & HAS_SSE4_1) cdef_filter_16_3 = cdef_filter_16_3_sse4_1;
     if (flags & HAS_AVX2) cdef_filter_16_3 = cdef_filter_16_3_avx2;
     cdef_filter_8_0 = cdef_filter_8_0_c;
-    if (flags & HAS_SSE2) cdef_filter_8_0 = cdef_filter_8_0_sse2;
     if (flags & HAS_SSSE3) cdef_filter_8_0 = cdef_filter_8_0_ssse3;
     if (flags & HAS_SSE4_1) cdef_filter_8_0 = cdef_filter_8_0_sse4_1;
     if (flags & HAS_AVX2) cdef_filter_8_0 = cdef_filter_8_0_avx2;
     cdef_filter_8_1 = cdef_filter_8_1_c;
-    if (flags & HAS_SSE2) cdef_filter_8_1 = cdef_filter_8_1_sse2;
     if (flags & HAS_SSSE3) cdef_filter_8_1 = cdef_filter_8_1_ssse3;
     if (flags & HAS_SSE4_1) cdef_filter_8_1 = cdef_filter_8_1_sse4_1;
     if (flags & HAS_AVX2) cdef_filter_8_1 = cdef_filter_8_1_avx2;
     cdef_filter_8_2 = cdef_filter_8_2_c;
-    if (flags & HAS_SSE2) cdef_filter_8_2 = cdef_filter_8_2_sse2;
     if (flags & HAS_SSSE3) cdef_filter_8_2 = cdef_filter_8_2_ssse3;
     if (flags & HAS_SSE4_1) cdef_filter_8_2 = cdef_filter_8_2_sse4_1;
     if (flags & HAS_AVX2) cdef_filter_8_2 = cdef_filter_8_2_avx2;
     cdef_filter_8_3 = cdef_filter_8_3_c;
-    if (flags & HAS_SSE2) cdef_filter_8_3 = cdef_filter_8_3_sse2;
     if (flags & HAS_SSSE3) cdef_filter_8_3 = cdef_filter_8_3_ssse3;
     if (flags & HAS_SSE4_1) cdef_filter_8_3 = cdef_filter_8_3_sse4_1;
     if (flags & HAS_AVX2) cdef_filter_8_3 = cdef_filter_8_3_avx2;
     cdef_find_dir = cdef_find_dir_c;
-    if (flags & HAS_SSE2) cdef_find_dir = cdef_find_dir_sse2;
     if (flags & HAS_SSSE3) cdef_find_dir = cdef_find_dir_ssse3;
     if (flags & HAS_SSE4_1) cdef_find_dir = cdef_find_dir_sse4_1;
     if (flags & HAS_AVX2) cdef_find_dir = cdef_find_dir_avx2;
     cdef_find_dir_dual = cdef_find_dir_dual_c;
-    if (flags & HAS_SSE2) cdef_find_dir_dual = cdef_find_dir_dual_sse2;
     if (flags & HAS_SSSE3) cdef_find_dir_dual = cdef_find_dir_dual_ssse3;
     if (flags & HAS_SSE4_1) cdef_find_dir_dual = cdef_find_dir_dual_sse4_1;
     if (flags & HAS_AVX2) cdef_find_dir_dual = cdef_find_dir_dual_avx2;
diff --git a/media/libaom/config/linux/x64/config/aom_config.asm b/media/libaom/config/linux/x64/config/aom_config.asm
index f793ff3c6d..3f470f3a5f 100644
--- a/media/libaom/config/linux/x64/config/aom_config.asm
+++ b/media/libaom/config/linux/x64/config/aom_config.asm
@@ -53,6 +53,7 @@ CONFIG_OS_SUPPORT equ 1
 CONFIG_OUTPUT_FRAME_SIZE equ 0
 CONFIG_PARTITION_SEARCH_ORDER equ 0
 CONFIG_PIC equ 0
+CONFIG_QUANT_MATRIX equ 1
 CONFIG_RATECTRL_LOG equ 0
 CONFIG_RD_COMMAND equ 0
 CONFIG_RD_DEBUG equ 0
@@ -87,6 +88,7 @@ HAVE_SSE4_1 equ 1
 HAVE_SSE4_2 equ 1
 HAVE_SSSE3 equ 1
 HAVE_SVE equ 0
+HAVE_SVE2 equ 0
 HAVE_VSX equ 0
 HAVE_WXWIDGETS equ 0
 STATIC_LINK_JXL equ 0
diff --git a/media/libaom/config/linux/x64/config/aom_config.h b/media/libaom/config/linux/x64/config/aom_config.h
index 670d2ffe56..6d96b65b07 100644
--- a/media/libaom/config/linux/x64/config/aom_config.h
+++ b/media/libaom/config/linux/x64/config/aom_config.h
@@ -55,6 +55,7 @@
 #define CONFIG_OUTPUT_FRAME_SIZE 0
 #define CONFIG_PARTITION_SEARCH_ORDER 0
 #define CONFIG_PIC 0
+#define CONFIG_QUANT_MATRIX 1
 #define CONFIG_RATECTRL_LOG 0
 #define CONFIG_RD_COMMAND 0
 #define CONFIG_RD_DEBUG 0
@@ -89,6 +90,7 @@
 #define HAVE_SSE4_2 1
 #define HAVE_SSSE3 1
 #define HAVE_SVE 0
+#define HAVE_SVE2 0
 #define HAVE_VSX 0
 #define HAVE_WXWIDGETS 0
 #define INLINE inline
diff --git a/media/libaom/config/linux/x64/config/aom_dsp_rtcd.h b/media/libaom/config/linux/x64/config/aom_dsp_rtcd.h
index 8e979cc189..9135c6f423 100644
--- a/media/libaom/config/linux/x64/config/aom_dsp_rtcd.h
+++ b/media/libaom/config/linux/x64/config/aom_dsp_rtcd.h
@@ -57,21 +57,30 @@ void aom_comp_mask_pred_ssse3(uint8_t *comp_pred, const uint8_t *pred, int width
 void aom_comp_mask_pred_avx2(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
 RTCD_EXTERN void (*aom_comp_mask_pred)(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
 
+double aom_compute_correlation_c(const unsigned char *frame1, int stride1, int x1, int y1, double mean1, double one_over_stddev1, const unsigned char *frame2, int stride2, int x2, int y2, double mean2, double one_over_stddev2);
+double aom_compute_correlation_sse4_1(const unsigned char *frame1, int stride1, int x1, int y1, double mean1, double one_over_stddev1, const unsigned char *frame2, int stride2, int x2, int y2, double mean2, double one_over_stddev2);
+double aom_compute_correlation_avx2(const unsigned char *frame1, int stride1, int x1, int y1, double mean1, double one_over_stddev1, const unsigned char *frame2, int stride2, int x2, int y2, double mean2, double one_over_stddev2);
+RTCD_EXTERN double (*aom_compute_correlation)(const unsigned char *frame1, int stride1, int x1, int y1, double mean1, double one_over_stddev1, const unsigned char *frame2, int stride2, int x2, int y2, double mean2, double one_over_stddev2);
+
 void aom_compute_flow_at_point_c(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
 void aom_compute_flow_at_point_sse4_1(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
+void aom_compute_flow_at_point_avx2(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
 RTCD_EXTERN void (*aom_compute_flow_at_point)(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
 
+bool aom_compute_mean_stddev_c(const unsigned char *frame, int stride, int x, int y, double *mean, double *one_over_stddev);
+bool aom_compute_mean_stddev_sse4_1(const unsigned char *frame, int stride, int x, int y, double *mean, double *one_over_stddev);
+bool aom_compute_mean_stddev_avx2(const unsigned char *frame, int stride, int x, int y, double *mean, double *one_over_stddev);
+RTCD_EXTERN bool (*aom_compute_mean_stddev)(const unsigned char *frame, int stride, int x, int y, double *mean, double *one_over_stddev);
+
 void aom_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h);
 #define aom_convolve8 aom_convolve8_c
 
 void aom_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
-void aom_convolve8_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void aom_convolve8_horiz_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void aom_convolve8_horiz_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 RTCD_EXTERN void (*aom_convolve8_horiz)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 
 void aom_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
-void aom_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void aom_convolve8_vert_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void aom_convolve8_vert_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 RTCD_EXTERN void (*aom_convolve8_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
@@ -903,7 +912,8 @@ RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance8x8)(const uin
 
 unsigned int aom_highbd_10_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
 unsigned int aom_highbd_10_mse16x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
-#define aom_highbd_10_mse16x16 aom_highbd_10_mse16x16_sse2
+unsigned int aom_highbd_10_mse16x16_avx2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_mse16x16)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
 
 unsigned int aom_highbd_10_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
 #define aom_highbd_10_mse16x8 aom_highbd_10_mse16x8_c
@@ -5132,7 +5142,8 @@ unsigned int aom_sad16x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const
 #define aom_sad16x4_avg aom_sad16x4_avg_sse2
 
 void aom_sad16x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
-#define aom_sad16x4x3d aom_sad16x4x3d_c
+void aom_sad16x4x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x4x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
 
 void aom_sad16x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
 void aom_sad16x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
@@ -5468,7 +5479,8 @@ unsigned int aom_sad_skip_16x4_c(const uint8_t *src_ptr, int src_stride, const u
 #define aom_sad_skip_16x4 aom_sad_skip_16x4_c
 
 void aom_sad_skip_16x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
-#define aom_sad_skip_16x4x4d aom_sad_skip_16x4x4d_c
+void aom_sad_skip_16x4x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_16x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
 
 unsigned int aom_sad_skip_16x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
 unsigned int aom_sad_skip_16x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
@@ -5870,243 +5882,199 @@ void aom_ssim_parms_8x8_sse2(const uint8_t *s, int sp, const uint8_t *r, int rp,
 #define aom_ssim_parms_8x8 aom_ssim_parms_8x8_sse2
 
 uint32_t aom_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance128x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance128x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance128x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance128x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance16x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance16x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance16x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance16x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance16x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance32x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance32x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance32x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance32x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance32x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance32x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance32x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance4x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance4x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance4x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance4x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance4x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance4x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance4x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance64x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance64x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance64x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance64x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance64x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance64x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance64x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance8x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance8x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance8x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance8x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance128x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance128x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance128x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance128x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x4_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x8_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance32x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance32x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance32x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance32x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance32x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance32x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance32x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance4x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance4x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance4x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance4x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance4x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance4x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance4x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance64x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance64x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance64x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance64x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance64x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance64x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance64x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance8x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance8x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance8x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance8x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
@@ -6329,11 +6297,6 @@ int aom_vector_var_sse4_1(const int16_t *ref, const int16_t *src, int bwl);
 int aom_vector_var_avx2(const int16_t *ref, const int16_t *src, int bwl);
 RTCD_EXTERN int (*aom_vector_var)(const int16_t *ref, const int16_t *src, int bwl);
 
-double av1_compute_cross_correlation_c(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
-double av1_compute_cross_correlation_sse4_1(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
-double av1_compute_cross_correlation_avx2(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
-RTCD_EXTERN double (*av1_compute_cross_correlation)(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
-
 void aom_dsp_rtcd(void);
 
 #ifdef RTCD_C
@@ -6358,12 +6321,19 @@ static void setup_rtcd_internal(void)
     aom_comp_mask_pred = aom_comp_mask_pred_c;
     if (flags & HAS_SSSE3) aom_comp_mask_pred = aom_comp_mask_pred_ssse3;
     if (flags & HAS_AVX2) aom_comp_mask_pred = aom_comp_mask_pred_avx2;
+    aom_compute_correlation = aom_compute_correlation_c;
+    if (flags & HAS_SSE4_1) aom_compute_correlation = aom_compute_correlation_sse4_1;
+    if (flags & HAS_AVX2) aom_compute_correlation = aom_compute_correlation_avx2;
     aom_compute_flow_at_point = aom_compute_flow_at_point_c;
     if (flags & HAS_SSE4_1) aom_compute_flow_at_point = aom_compute_flow_at_point_sse4_1;
-    aom_convolve8_horiz = aom_convolve8_horiz_sse2;
+    if (flags & HAS_AVX2) aom_compute_flow_at_point = aom_compute_flow_at_point_avx2;
+    aom_compute_mean_stddev = aom_compute_mean_stddev_c;
+    if (flags & HAS_SSE4_1) aom_compute_mean_stddev = aom_compute_mean_stddev_sse4_1;
+    if (flags & HAS_AVX2) aom_compute_mean_stddev = aom_compute_mean_stddev_avx2;
+    aom_convolve8_horiz = aom_convolve8_horiz_c;
     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;
+    aom_convolve8_vert = aom_convolve8_vert_c;
     if (flags & HAS_SSSE3) aom_convolve8_vert = aom_convolve8_vert_ssse3;
     if (flags & HAS_AVX2) aom_convolve8_vert = aom_convolve8_vert_avx2;
     aom_convolve_copy = aom_convolve_copy_sse2;
@@ -6528,6 +6498,8 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance8x4 = aom_highbd_10_masked_sub_pixel_variance8x4_ssse3;
     aom_highbd_10_masked_sub_pixel_variance8x8 = aom_highbd_10_masked_sub_pixel_variance8x8_c;
     if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance8x8 = aom_highbd_10_masked_sub_pixel_variance8x8_ssse3;
+    aom_highbd_10_mse16x16 = aom_highbd_10_mse16x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_mse16x16 = aom_highbd_10_mse16x16_avx2;
     aom_highbd_10_obmc_variance128x128 = aom_highbd_10_obmc_variance128x128_c;
     if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance128x128 = aom_highbd_10_obmc_variance128x128_sse4_1;
     aom_highbd_10_obmc_variance128x64 = aom_highbd_10_obmc_variance128x64_c;
@@ -7626,6 +7598,8 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_AVX2) aom_sad16x32x3d = aom_sad16x32x3d_avx2;
     aom_sad16x32x4d = aom_sad16x32x4d_sse2;
     if (flags & HAS_AVX2) aom_sad16x32x4d = aom_sad16x32x4d_avx2;
+    aom_sad16x4x3d = aom_sad16x4x3d_c;
+    if (flags & HAS_AVX2) aom_sad16x4x3d = aom_sad16x4x3d_avx2;
     aom_sad16x4x4d = aom_sad16x4x4d_sse2;
     if (flags & HAS_AVX2) aom_sad16x4x4d = aom_sad16x4x4d_avx2;
     aom_sad16x64x3d = aom_sad16x64x3d_c;
@@ -7704,6 +7678,8 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_AVX2) aom_sad_skip_16x16x4d = aom_sad_skip_16x16x4d_avx2;
     aom_sad_skip_16x32x4d = aom_sad_skip_16x32x4d_sse2;
     if (flags & HAS_AVX2) aom_sad_skip_16x32x4d = aom_sad_skip_16x32x4d_avx2;
+    aom_sad_skip_16x4x4d = aom_sad_skip_16x4x4d_c;
+    if (flags & HAS_AVX2) aom_sad_skip_16x4x4d = aom_sad_skip_16x4x4d_avx2;
     aom_sad_skip_16x64x4d = aom_sad_skip_16x64x4d_sse2;
     if (flags & HAS_AVX2) aom_sad_skip_16x64x4d = aom_sad_skip_16x64x4d_avx2;
     aom_sad_skip_16x8x4d = aom_sad_skip_16x8x4d_sse2;
@@ -7859,114 +7835,114 @@ static void setup_rtcd_internal(void)
     aom_sse = aom_sse_c;
     if (flags & HAS_SSE4_1) aom_sse = aom_sse_sse4_1;
     if (flags & HAS_AVX2) aom_sse = aom_sse_avx2;
-    aom_sub_pixel_avg_variance128x128 = aom_sub_pixel_avg_variance128x128_sse2;
+    aom_sub_pixel_avg_variance128x128 = aom_sub_pixel_avg_variance128x128_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance128x128 = aom_sub_pixel_avg_variance128x128_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_avg_variance128x128 = aom_sub_pixel_avg_variance128x128_avx2;
-    aom_sub_pixel_avg_variance128x64 = aom_sub_pixel_avg_variance128x64_sse2;
+    aom_sub_pixel_avg_variance128x64 = aom_sub_pixel_avg_variance128x64_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance128x64 = aom_sub_pixel_avg_variance128x64_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_avg_variance128x64 = aom_sub_pixel_avg_variance128x64_avx2;
-    aom_sub_pixel_avg_variance16x16 = aom_sub_pixel_avg_variance16x16_sse2;
+    aom_sub_pixel_avg_variance16x16 = aom_sub_pixel_avg_variance16x16_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x16 = aom_sub_pixel_avg_variance16x16_ssse3;
-    aom_sub_pixel_avg_variance16x32 = aom_sub_pixel_avg_variance16x32_sse2;
+    aom_sub_pixel_avg_variance16x32 = aom_sub_pixel_avg_variance16x32_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x32 = aom_sub_pixel_avg_variance16x32_ssse3;
-    aom_sub_pixel_avg_variance16x4 = aom_sub_pixel_avg_variance16x4_sse2;
+    aom_sub_pixel_avg_variance16x4 = aom_sub_pixel_avg_variance16x4_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x4 = aom_sub_pixel_avg_variance16x4_ssse3;
-    aom_sub_pixel_avg_variance16x64 = aom_sub_pixel_avg_variance16x64_sse2;
+    aom_sub_pixel_avg_variance16x64 = aom_sub_pixel_avg_variance16x64_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x64 = aom_sub_pixel_avg_variance16x64_ssse3;
-    aom_sub_pixel_avg_variance16x8 = aom_sub_pixel_avg_variance16x8_sse2;
+    aom_sub_pixel_avg_variance16x8 = aom_sub_pixel_avg_variance16x8_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x8 = aom_sub_pixel_avg_variance16x8_ssse3;
-    aom_sub_pixel_avg_variance32x16 = aom_sub_pixel_avg_variance32x16_sse2;
+    aom_sub_pixel_avg_variance32x16 = aom_sub_pixel_avg_variance32x16_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x16 = aom_sub_pixel_avg_variance32x16_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_avg_variance32x16 = aom_sub_pixel_avg_variance32x16_avx2;
-    aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_sse2;
+    aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_avx2;
-    aom_sub_pixel_avg_variance32x64 = aom_sub_pixel_avg_variance32x64_sse2;
+    aom_sub_pixel_avg_variance32x64 = aom_sub_pixel_avg_variance32x64_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x64 = aom_sub_pixel_avg_variance32x64_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_avg_variance32x64 = aom_sub_pixel_avg_variance32x64_avx2;
-    aom_sub_pixel_avg_variance32x8 = aom_sub_pixel_avg_variance32x8_sse2;
+    aom_sub_pixel_avg_variance32x8 = aom_sub_pixel_avg_variance32x8_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x8 = aom_sub_pixel_avg_variance32x8_ssse3;
-    aom_sub_pixel_avg_variance4x16 = aom_sub_pixel_avg_variance4x16_sse2;
+    aom_sub_pixel_avg_variance4x16 = aom_sub_pixel_avg_variance4x16_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance4x16 = aom_sub_pixel_avg_variance4x16_ssse3;
-    aom_sub_pixel_avg_variance4x4 = aom_sub_pixel_avg_variance4x4_sse2;
+    aom_sub_pixel_avg_variance4x4 = aom_sub_pixel_avg_variance4x4_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance4x4 = aom_sub_pixel_avg_variance4x4_ssse3;
-    aom_sub_pixel_avg_variance4x8 = aom_sub_pixel_avg_variance4x8_sse2;
+    aom_sub_pixel_avg_variance4x8 = aom_sub_pixel_avg_variance4x8_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance4x8 = aom_sub_pixel_avg_variance4x8_ssse3;
-    aom_sub_pixel_avg_variance64x128 = aom_sub_pixel_avg_variance64x128_sse2;
+    aom_sub_pixel_avg_variance64x128 = aom_sub_pixel_avg_variance64x128_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance64x128 = aom_sub_pixel_avg_variance64x128_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_avg_variance64x128 = aom_sub_pixel_avg_variance64x128_avx2;
-    aom_sub_pixel_avg_variance64x16 = aom_sub_pixel_avg_variance64x16_sse2;
+    aom_sub_pixel_avg_variance64x16 = aom_sub_pixel_avg_variance64x16_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance64x16 = aom_sub_pixel_avg_variance64x16_ssse3;
-    aom_sub_pixel_avg_variance64x32 = aom_sub_pixel_avg_variance64x32_sse2;
+    aom_sub_pixel_avg_variance64x32 = aom_sub_pixel_avg_variance64x32_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance64x32 = aom_sub_pixel_avg_variance64x32_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_avg_variance64x32 = aom_sub_pixel_avg_variance64x32_avx2;
-    aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_sse2;
+    aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_avx2;
-    aom_sub_pixel_avg_variance8x16 = aom_sub_pixel_avg_variance8x16_sse2;
+    aom_sub_pixel_avg_variance8x16 = aom_sub_pixel_avg_variance8x16_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x16 = aom_sub_pixel_avg_variance8x16_ssse3;
-    aom_sub_pixel_avg_variance8x32 = aom_sub_pixel_avg_variance8x32_sse2;
+    aom_sub_pixel_avg_variance8x32 = aom_sub_pixel_avg_variance8x32_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x32 = aom_sub_pixel_avg_variance8x32_ssse3;
-    aom_sub_pixel_avg_variance8x4 = aom_sub_pixel_avg_variance8x4_sse2;
+    aom_sub_pixel_avg_variance8x4 = aom_sub_pixel_avg_variance8x4_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x4 = aom_sub_pixel_avg_variance8x4_ssse3;
-    aom_sub_pixel_avg_variance8x8 = aom_sub_pixel_avg_variance8x8_sse2;
+    aom_sub_pixel_avg_variance8x8 = aom_sub_pixel_avg_variance8x8_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x8 = aom_sub_pixel_avg_variance8x8_ssse3;
-    aom_sub_pixel_variance128x128 = aom_sub_pixel_variance128x128_sse2;
+    aom_sub_pixel_variance128x128 = aom_sub_pixel_variance128x128_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance128x128 = aom_sub_pixel_variance128x128_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance128x128 = aom_sub_pixel_variance128x128_avx2;
-    aom_sub_pixel_variance128x64 = aom_sub_pixel_variance128x64_sse2;
+    aom_sub_pixel_variance128x64 = aom_sub_pixel_variance128x64_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance128x64 = aom_sub_pixel_variance128x64_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance128x64 = aom_sub_pixel_variance128x64_avx2;
-    aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_sse2;
+    aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_avx2;
-    aom_sub_pixel_variance16x32 = aom_sub_pixel_variance16x32_sse2;
+    aom_sub_pixel_variance16x32 = aom_sub_pixel_variance16x32_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance16x32 = aom_sub_pixel_variance16x32_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance16x32 = aom_sub_pixel_variance16x32_avx2;
-    aom_sub_pixel_variance16x4 = aom_sub_pixel_variance16x4_sse2;
+    aom_sub_pixel_variance16x4 = aom_sub_pixel_variance16x4_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance16x4 = aom_sub_pixel_variance16x4_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance16x4 = aom_sub_pixel_variance16x4_avx2;
-    aom_sub_pixel_variance16x64 = aom_sub_pixel_variance16x64_sse2;
+    aom_sub_pixel_variance16x64 = aom_sub_pixel_variance16x64_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance16x64 = aom_sub_pixel_variance16x64_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance16x64 = aom_sub_pixel_variance16x64_avx2;
-    aom_sub_pixel_variance16x8 = aom_sub_pixel_variance16x8_sse2;
+    aom_sub_pixel_variance16x8 = aom_sub_pixel_variance16x8_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance16x8 = aom_sub_pixel_variance16x8_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance16x8 = aom_sub_pixel_variance16x8_avx2;
-    aom_sub_pixel_variance32x16 = aom_sub_pixel_variance32x16_sse2;
+    aom_sub_pixel_variance32x16 = aom_sub_pixel_variance32x16_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance32x16 = aom_sub_pixel_variance32x16_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance32x16 = aom_sub_pixel_variance32x16_avx2;
-    aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_sse2;
+    aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_avx2;
-    aom_sub_pixel_variance32x64 = aom_sub_pixel_variance32x64_sse2;
+    aom_sub_pixel_variance32x64 = aom_sub_pixel_variance32x64_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance32x64 = aom_sub_pixel_variance32x64_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance32x64 = aom_sub_pixel_variance32x64_avx2;
-    aom_sub_pixel_variance32x8 = aom_sub_pixel_variance32x8_sse2;
+    aom_sub_pixel_variance32x8 = aom_sub_pixel_variance32x8_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance32x8 = aom_sub_pixel_variance32x8_ssse3;
-    aom_sub_pixel_variance4x16 = aom_sub_pixel_variance4x16_sse2;
+    aom_sub_pixel_variance4x16 = aom_sub_pixel_variance4x16_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance4x16 = aom_sub_pixel_variance4x16_ssse3;
-    aom_sub_pixel_variance4x4 = aom_sub_pixel_variance4x4_sse2;
+    aom_sub_pixel_variance4x4 = aom_sub_pixel_variance4x4_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance4x4 = aom_sub_pixel_variance4x4_ssse3;
-    aom_sub_pixel_variance4x8 = aom_sub_pixel_variance4x8_sse2;
+    aom_sub_pixel_variance4x8 = aom_sub_pixel_variance4x8_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance4x8 = aom_sub_pixel_variance4x8_ssse3;
-    aom_sub_pixel_variance64x128 = aom_sub_pixel_variance64x128_sse2;
+    aom_sub_pixel_variance64x128 = aom_sub_pixel_variance64x128_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance64x128 = aom_sub_pixel_variance64x128_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance64x128 = aom_sub_pixel_variance64x128_avx2;
-    aom_sub_pixel_variance64x16 = aom_sub_pixel_variance64x16_sse2;
+    aom_sub_pixel_variance64x16 = aom_sub_pixel_variance64x16_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance64x16 = aom_sub_pixel_variance64x16_ssse3;
-    aom_sub_pixel_variance64x32 = aom_sub_pixel_variance64x32_sse2;
+    aom_sub_pixel_variance64x32 = aom_sub_pixel_variance64x32_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance64x32 = aom_sub_pixel_variance64x32_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance64x32 = aom_sub_pixel_variance64x32_avx2;
-    aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_sse2;
+    aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_avx2;
-    aom_sub_pixel_variance8x16 = aom_sub_pixel_variance8x16_sse2;
+    aom_sub_pixel_variance8x16 = aom_sub_pixel_variance8x16_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance8x16 = aom_sub_pixel_variance8x16_ssse3;
-    aom_sub_pixel_variance8x32 = aom_sub_pixel_variance8x32_sse2;
+    aom_sub_pixel_variance8x32 = aom_sub_pixel_variance8x32_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance8x32 = aom_sub_pixel_variance8x32_ssse3;
-    aom_sub_pixel_variance8x4 = aom_sub_pixel_variance8x4_sse2;
+    aom_sub_pixel_variance8x4 = aom_sub_pixel_variance8x4_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance8x4 = aom_sub_pixel_variance8x4_ssse3;
-    aom_sub_pixel_variance8x8 = aom_sub_pixel_variance8x8_sse2;
+    aom_sub_pixel_variance8x8 = aom_sub_pixel_variance8x8_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance8x8 = aom_sub_pixel_variance8x8_ssse3;
     aom_subtract_block = aom_subtract_block_sse2;
     if (flags & HAS_AVX2) aom_subtract_block = aom_subtract_block_avx2;
@@ -8023,9 +7999,6 @@ static void setup_rtcd_internal(void)
     aom_vector_var = aom_vector_var_c;
     if (flags & HAS_SSE4_1) aom_vector_var = aom_vector_var_sse4_1;
     if (flags & HAS_AVX2) aom_vector_var = aom_vector_var_avx2;
-    av1_compute_cross_correlation = av1_compute_cross_correlation_c;
-    if (flags & HAS_SSE4_1) av1_compute_cross_correlation = av1_compute_cross_correlation_sse4_1;
-    if (flags & HAS_AVX2) av1_compute_cross_correlation = av1_compute_cross_correlation_avx2;
 }
 #endif
 
diff --git a/media/libaom/config/linux/x64/config/aom_scale_rtcd.h b/media/libaom/config/linux/x64/config/aom_scale_rtcd.h
index 3b70fb47c3..cdabb21106 100644
--- a/media/libaom/config/linux/x64/config/aom_scale_rtcd.h
+++ b/media/libaom/config/linux/x64/config/aom_scale_rtcd.h
@@ -8,13 +8,15 @@
 #define RTCD_EXTERN extern
 #endif
 
+#include <stdbool.h>
+
 struct yv12_buffer_config;
 
 #ifdef __cplusplus
 extern "C" {
 #endif
 
-void aom_extend_frame_borders_c(struct yv12_buffer_config *ybf, const int num_planes);
+void aom_extend_frame_borders_c(struct yv12_buffer_config *ybf, int num_planes);
 #define aom_extend_frame_borders aom_extend_frame_borders_c
 
 void aom_extend_frame_borders_plane_row_c(const struct yv12_buffer_config *ybf, int plane, int v_start, int v_end);
@@ -50,13 +52,13 @@ void aom_vertical_band_5_4_scale_c(unsigned char *source, int src_pitch, unsigne
 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);
+void aom_yv12_copy_u_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc, int use_crop);
 #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);
+void aom_yv12_copy_v_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc, int use_crop);
 #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);
+void aom_yv12_copy_y_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc, int use_crop);
 #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);
@@ -80,7 +82,7 @@ void aom_yv12_partial_copy_v_c(const struct yv12_buffer_config *src_bc, int hsta
 void aom_yv12_partial_copy_y_c(const struct yv12_buffer_config *src_ybc, int hstart1, int hend1, int vstart1, int vend1, struct yv12_buffer_config *dst_ybc, int hstart2, int vstart2);
 #define aom_yv12_partial_copy_y aom_yv12_partial_copy_y_c
 
-int aom_yv12_realloc_with_new_border_c(struct yv12_buffer_config *ybf, int new_border, int byte_alignment, int num_pyramid_levels, int num_planes);
+int aom_yv12_realloc_with_new_border_c(struct yv12_buffer_config *ybf, int new_border, int byte_alignment, bool alloc_pyramid, int num_planes);
 #define aom_yv12_realloc_with_new_border aom_yv12_realloc_with_new_border_c
 
 void aom_scale_rtcd(void);
diff --git a/media/libaom/config/linux/x64/config/av1_rtcd.h b/media/libaom/config/linux/x64/config/av1_rtcd.h
index b1cdc99700..ad72985afe 100644
--- a/media/libaom/config/linux/x64/config/av1_rtcd.h
+++ b/media/libaom/config/linux/x64/config/av1_rtcd.h
@@ -253,7 +253,6 @@ void av1_convolve_y_sr_intrabc_c(const uint8_t *src, int src_stride, uint8_t *ds
 #define av1_convolve_y_sr_intrabc av1_convolve_y_sr_intrabc_c
 
 void av1_dist_wtd_convolve_2d_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
-void av1_dist_wtd_convolve_2d_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
 void av1_dist_wtd_convolve_2d_ssse3(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
 void av1_dist_wtd_convolve_2d_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
 RTCD_EXTERN void (*av1_dist_wtd_convolve_2d)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
@@ -659,7 +658,6 @@ void av1_inv_txfm_add_avx2(const tran_low_t *dqcoeff, uint8_t *dst, int stride,
 RTCD_EXTERN void (*av1_inv_txfm_add)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
 
 void av1_lowbd_fwd_txfm_c(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
-void av1_lowbd_fwd_txfm_sse2(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
 void av1_lowbd_fwd_txfm_sse4_1(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
 void av1_lowbd_fwd_txfm_avx2(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
 RTCD_EXTERN void (*av1_lowbd_fwd_txfm)(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
@@ -755,85 +753,61 @@ void av1_wiener_convolve_add_src_avx2(const uint8_t *src, ptrdiff_t src_stride,
 RTCD_EXTERN void (*av1_wiener_convolve_add_src)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params);
 
 void cdef_copy_rect8_16bit_to_16bit_c(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
-void cdef_copy_rect8_16bit_to_16bit_sse2(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
-void cdef_copy_rect8_16bit_to_16bit_ssse3(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
 void cdef_copy_rect8_16bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
 void cdef_copy_rect8_16bit_to_16bit_avx2(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
 RTCD_EXTERN void (*cdef_copy_rect8_16bit_to_16bit)(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
 
 void cdef_copy_rect8_8bit_to_16bit_c(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
-void cdef_copy_rect8_8bit_to_16bit_sse2(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
-void cdef_copy_rect8_8bit_to_16bit_ssse3(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
 void cdef_copy_rect8_8bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
 void cdef_copy_rect8_8bit_to_16bit_avx2(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
 RTCD_EXTERN void (*cdef_copy_rect8_8bit_to_16bit)(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
 
 void cdef_filter_16_0_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_16_0_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_16_0_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_0_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_0_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 RTCD_EXTERN void (*cdef_filter_16_0)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 
 void cdef_filter_16_1_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_16_1_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_16_1_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_1_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_1_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 RTCD_EXTERN void (*cdef_filter_16_1)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 
 void cdef_filter_16_2_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_16_2_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_16_2_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_2_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_2_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 RTCD_EXTERN void (*cdef_filter_16_2)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 
 void cdef_filter_16_3_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_16_3_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_16_3_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_3_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_3_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 RTCD_EXTERN void (*cdef_filter_16_3)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 
 void cdef_filter_8_0_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_8_0_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_8_0_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_0_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_0_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 RTCD_EXTERN void (*cdef_filter_8_0)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 
 void cdef_filter_8_1_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_8_1_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_8_1_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_1_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_1_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 RTCD_EXTERN void (*cdef_filter_8_1)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 
 void cdef_filter_8_2_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_8_2_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_8_2_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_2_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_2_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 RTCD_EXTERN void (*cdef_filter_8_2)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 
 void cdef_filter_8_3_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_8_3_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_8_3_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_3_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_3_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 RTCD_EXTERN void (*cdef_filter_8_3)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 
 int cdef_find_dir_c(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
-int cdef_find_dir_sse2(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
-int cdef_find_dir_ssse3(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
 int cdef_find_dir_sse4_1(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
 int cdef_find_dir_avx2(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
 RTCD_EXTERN int (*cdef_find_dir)(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
 
 void cdef_find_dir_dual_c(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
-void cdef_find_dir_dual_sse2(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
-void cdef_find_dir_dual_ssse3(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
 void cdef_find_dir_dual_sse4_1(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
 void cdef_find_dir_dual_avx2(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
 RTCD_EXTERN void (*cdef_find_dir_dual)(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
@@ -941,7 +915,7 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_AVX2) av1_convolve_x_sr = av1_convolve_x_sr_avx2;
     av1_convolve_y_sr = av1_convolve_y_sr_sse2;
     if (flags & HAS_AVX2) av1_convolve_y_sr = av1_convolve_y_sr_avx2;
-    av1_dist_wtd_convolve_2d = av1_dist_wtd_convolve_2d_sse2;
+    av1_dist_wtd_convolve_2d = av1_dist_wtd_convolve_2d_c;
     if (flags & HAS_SSSE3) av1_dist_wtd_convolve_2d = av1_dist_wtd_convolve_2d_ssse3;
     if (flags & HAS_AVX2) av1_dist_wtd_convolve_2d = av1_dist_wtd_convolve_2d_avx2;
     av1_dist_wtd_convolve_2d_copy = av1_dist_wtd_convolve_2d_copy_sse2;
@@ -1091,7 +1065,7 @@ static void setup_rtcd_internal(void)
     av1_inv_txfm_add = av1_inv_txfm_add_c;
     if (flags & HAS_SSSE3) av1_inv_txfm_add = av1_inv_txfm_add_ssse3;
     if (flags & HAS_AVX2) av1_inv_txfm_add = av1_inv_txfm_add_avx2;
-    av1_lowbd_fwd_txfm = av1_lowbd_fwd_txfm_sse2;
+    av1_lowbd_fwd_txfm = av1_lowbd_fwd_txfm_c;
     if (flags & HAS_SSE4_1) av1_lowbd_fwd_txfm = av1_lowbd_fwd_txfm_sse4_1;
     if (flags & HAS_AVX2) av1_lowbd_fwd_txfm = av1_lowbd_fwd_txfm_avx2;
     av1_lowbd_pixel_proj_error = av1_lowbd_pixel_proj_error_c;
@@ -1133,52 +1107,40 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_AVX2) av1_wedge_sse_from_residuals = av1_wedge_sse_from_residuals_avx2;
     av1_wiener_convolve_add_src = av1_wiener_convolve_add_src_sse2;
     if (flags & HAS_AVX2) av1_wiener_convolve_add_src = av1_wiener_convolve_add_src_avx2;
-    cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_sse2;
-    if (flags & HAS_SSSE3) cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_ssse3;
+    cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_c;
     if (flags & HAS_SSE4_1) cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_sse4_1;
     if (flags & HAS_AVX2) cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_avx2;
-    cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_sse2;
-    if (flags & HAS_SSSE3) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_ssse3;
+    cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_c;
     if (flags & HAS_SSE4_1) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_sse4_1;
     if (flags & HAS_AVX2) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_avx2;
-    cdef_filter_16_0 = cdef_filter_16_0_sse2;
-    if (flags & HAS_SSSE3) cdef_filter_16_0 = cdef_filter_16_0_ssse3;
+    cdef_filter_16_0 = cdef_filter_16_0_c;
     if (flags & HAS_SSE4_1) cdef_filter_16_0 = cdef_filter_16_0_sse4_1;
     if (flags & HAS_AVX2) cdef_filter_16_0 = cdef_filter_16_0_avx2;
-    cdef_filter_16_1 = cdef_filter_16_1_sse2;
-    if (flags & HAS_SSSE3) cdef_filter_16_1 = cdef_filter_16_1_ssse3;
+    cdef_filter_16_1 = cdef_filter_16_1_c;
     if (flags & HAS_SSE4_1) cdef_filter_16_1 = cdef_filter_16_1_sse4_1;
     if (flags & HAS_AVX2) cdef_filter_16_1 = cdef_filter_16_1_avx2;
-    cdef_filter_16_2 = cdef_filter_16_2_sse2;
-    if (flags & HAS_SSSE3) cdef_filter_16_2 = cdef_filter_16_2_ssse3;
+    cdef_filter_16_2 = cdef_filter_16_2_c;
     if (flags & HAS_SSE4_1) cdef_filter_16_2 = cdef_filter_16_2_sse4_1;
     if (flags & HAS_AVX2) cdef_filter_16_2 = cdef_filter_16_2_avx2;
-    cdef_filter_16_3 = cdef_filter_16_3_sse2;
-    if (flags & HAS_SSSE3) cdef_filter_16_3 = cdef_filter_16_3_ssse3;
+    cdef_filter_16_3 = cdef_filter_16_3_c;
     if (flags & HAS_SSE4_1) cdef_filter_16_3 = cdef_filter_16_3_sse4_1;
     if (flags & HAS_AVX2) cdef_filter_16_3 = cdef_filter_16_3_avx2;
-    cdef_filter_8_0 = cdef_filter_8_0_sse2;
-    if (flags & HAS_SSSE3) cdef_filter_8_0 = cdef_filter_8_0_ssse3;
+    cdef_filter_8_0 = cdef_filter_8_0_c;
     if (flags & HAS_SSE4_1) cdef_filter_8_0 = cdef_filter_8_0_sse4_1;
     if (flags & HAS_AVX2) cdef_filter_8_0 = cdef_filter_8_0_avx2;
-    cdef_filter_8_1 = cdef_filter_8_1_sse2;
-    if (flags & HAS_SSSE3) cdef_filter_8_1 = cdef_filter_8_1_ssse3;
+    cdef_filter_8_1 = cdef_filter_8_1_c;
     if (flags & HAS_SSE4_1) cdef_filter_8_1 = cdef_filter_8_1_sse4_1;
     if (flags & HAS_AVX2) cdef_filter_8_1 = cdef_filter_8_1_avx2;
-    cdef_filter_8_2 = cdef_filter_8_2_sse2;
-    if (flags & HAS_SSSE3) cdef_filter_8_2 = cdef_filter_8_2_ssse3;
+    cdef_filter_8_2 = cdef_filter_8_2_c;
     if (flags & HAS_SSE4_1) cdef_filter_8_2 = cdef_filter_8_2_sse4_1;
     if (flags & HAS_AVX2) cdef_filter_8_2 = cdef_filter_8_2_avx2;
-    cdef_filter_8_3 = cdef_filter_8_3_sse2;
-    if (flags & HAS_SSSE3) cdef_filter_8_3 = cdef_filter_8_3_ssse3;
+    cdef_filter_8_3 = cdef_filter_8_3_c;
     if (flags & HAS_SSE4_1) cdef_filter_8_3 = cdef_filter_8_3_sse4_1;
     if (flags & HAS_AVX2) cdef_filter_8_3 = cdef_filter_8_3_avx2;
-    cdef_find_dir = cdef_find_dir_sse2;
-    if (flags & HAS_SSSE3) cdef_find_dir = cdef_find_dir_ssse3;
+    cdef_find_dir = cdef_find_dir_c;
     if (flags & HAS_SSE4_1) cdef_find_dir = cdef_find_dir_sse4_1;
     if (flags & HAS_AVX2) cdef_find_dir = cdef_find_dir_avx2;
-    cdef_find_dir_dual = cdef_find_dir_dual_sse2;
-    if (flags & HAS_SSSE3) cdef_find_dir_dual = cdef_find_dir_dual_ssse3;
+    cdef_find_dir_dual = cdef_find_dir_dual_c;
     if (flags & HAS_SSE4_1) cdef_find_dir_dual = cdef_find_dir_dual_sse4_1;
     if (flags & HAS_AVX2) cdef_find_dir_dual = cdef_find_dir_dual_avx2;
     cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_c;
diff --git a/media/libaom/config/mac/x64/config/aom_config.asm b/media/libaom/config/mac/x64/config/aom_config.asm
index f793ff3c6d..3f470f3a5f 100644
--- a/media/libaom/config/mac/x64/config/aom_config.asm
+++ b/media/libaom/config/mac/x64/config/aom_config.asm
@@ -53,6 +53,7 @@ CONFIG_OS_SUPPORT equ 1
 CONFIG_OUTPUT_FRAME_SIZE equ 0
 CONFIG_PARTITION_SEARCH_ORDER equ 0
 CONFIG_PIC equ 0
+CONFIG_QUANT_MATRIX equ 1
 CONFIG_RATECTRL_LOG equ 0
 CONFIG_RD_COMMAND equ 0
 CONFIG_RD_DEBUG equ 0
@@ -87,6 +88,7 @@ HAVE_SSE4_1 equ 1
 HAVE_SSE4_2 equ 1
 HAVE_SSSE3 equ 1
 HAVE_SVE equ 0
+HAVE_SVE2 equ 0
 HAVE_VSX equ 0
 HAVE_WXWIDGETS equ 0
 STATIC_LINK_JXL equ 0
diff --git a/media/libaom/config/mac/x64/config/aom_config.h b/media/libaom/config/mac/x64/config/aom_config.h
index 670d2ffe56..6d96b65b07 100644
--- a/media/libaom/config/mac/x64/config/aom_config.h
+++ b/media/libaom/config/mac/x64/config/aom_config.h
@@ -55,6 +55,7 @@
 #define CONFIG_OUTPUT_FRAME_SIZE 0
 #define CONFIG_PARTITION_SEARCH_ORDER 0
 #define CONFIG_PIC 0
+#define CONFIG_QUANT_MATRIX 1
 #define CONFIG_RATECTRL_LOG 0
 #define CONFIG_RD_COMMAND 0
 #define CONFIG_RD_DEBUG 0
@@ -89,6 +90,7 @@
 #define HAVE_SSE4_2 1
 #define HAVE_SSSE3 1
 #define HAVE_SVE 0
+#define HAVE_SVE2 0
 #define HAVE_VSX 0
 #define HAVE_WXWIDGETS 0
 #define INLINE inline
diff --git a/media/libaom/config/mac/x64/config/aom_dsp_rtcd.h b/media/libaom/config/mac/x64/config/aom_dsp_rtcd.h
index 8e979cc189..9135c6f423 100644
--- a/media/libaom/config/mac/x64/config/aom_dsp_rtcd.h
+++ b/media/libaom/config/mac/x64/config/aom_dsp_rtcd.h
@@ -57,21 +57,30 @@ void aom_comp_mask_pred_ssse3(uint8_t *comp_pred, const uint8_t *pred, int width
 void aom_comp_mask_pred_avx2(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
 RTCD_EXTERN void (*aom_comp_mask_pred)(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
 
+double aom_compute_correlation_c(const unsigned char *frame1, int stride1, int x1, int y1, double mean1, double one_over_stddev1, const unsigned char *frame2, int stride2, int x2, int y2, double mean2, double one_over_stddev2);
+double aom_compute_correlation_sse4_1(const unsigned char *frame1, int stride1, int x1, int y1, double mean1, double one_over_stddev1, const unsigned char *frame2, int stride2, int x2, int y2, double mean2, double one_over_stddev2);
+double aom_compute_correlation_avx2(const unsigned char *frame1, int stride1, int x1, int y1, double mean1, double one_over_stddev1, const unsigned char *frame2, int stride2, int x2, int y2, double mean2, double one_over_stddev2);
+RTCD_EXTERN double (*aom_compute_correlation)(const unsigned char *frame1, int stride1, int x1, int y1, double mean1, double one_over_stddev1, const unsigned char *frame2, int stride2, int x2, int y2, double mean2, double one_over_stddev2);
+
 void aom_compute_flow_at_point_c(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
 void aom_compute_flow_at_point_sse4_1(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
+void aom_compute_flow_at_point_avx2(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
 RTCD_EXTERN void (*aom_compute_flow_at_point)(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
 
+bool aom_compute_mean_stddev_c(const unsigned char *frame, int stride, int x, int y, double *mean, double *one_over_stddev);
+bool aom_compute_mean_stddev_sse4_1(const unsigned char *frame, int stride, int x, int y, double *mean, double *one_over_stddev);
+bool aom_compute_mean_stddev_avx2(const unsigned char *frame, int stride, int x, int y, double *mean, double *one_over_stddev);
+RTCD_EXTERN bool (*aom_compute_mean_stddev)(const unsigned char *frame, int stride, int x, int y, double *mean, double *one_over_stddev);
+
 void aom_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h);
 #define aom_convolve8 aom_convolve8_c
 
 void aom_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
-void aom_convolve8_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void aom_convolve8_horiz_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void aom_convolve8_horiz_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 RTCD_EXTERN void (*aom_convolve8_horiz)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 
 void aom_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
-void aom_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void aom_convolve8_vert_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void aom_convolve8_vert_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 RTCD_EXTERN void (*aom_convolve8_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
@@ -903,7 +912,8 @@ RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance8x8)(const uin
 
 unsigned int aom_highbd_10_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
 unsigned int aom_highbd_10_mse16x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
-#define aom_highbd_10_mse16x16 aom_highbd_10_mse16x16_sse2
+unsigned int aom_highbd_10_mse16x16_avx2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_mse16x16)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
 
 unsigned int aom_highbd_10_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
 #define aom_highbd_10_mse16x8 aom_highbd_10_mse16x8_c
@@ -5132,7 +5142,8 @@ unsigned int aom_sad16x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const
 #define aom_sad16x4_avg aom_sad16x4_avg_sse2
 
 void aom_sad16x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
-#define aom_sad16x4x3d aom_sad16x4x3d_c
+void aom_sad16x4x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x4x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
 
 void aom_sad16x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
 void aom_sad16x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
@@ -5468,7 +5479,8 @@ unsigned int aom_sad_skip_16x4_c(const uint8_t *src_ptr, int src_stride, const u
 #define aom_sad_skip_16x4 aom_sad_skip_16x4_c
 
 void aom_sad_skip_16x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
-#define aom_sad_skip_16x4x4d aom_sad_skip_16x4x4d_c
+void aom_sad_skip_16x4x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_16x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
 
 unsigned int aom_sad_skip_16x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
 unsigned int aom_sad_skip_16x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
@@ -5870,243 +5882,199 @@ void aom_ssim_parms_8x8_sse2(const uint8_t *s, int sp, const uint8_t *r, int rp,
 #define aom_ssim_parms_8x8 aom_ssim_parms_8x8_sse2
 
 uint32_t aom_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance128x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance128x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance128x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance128x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance16x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance16x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance16x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance16x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance16x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance32x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance32x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance32x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance32x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance32x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance32x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance32x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance4x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance4x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance4x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance4x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance4x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance4x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance4x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance64x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance64x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance64x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance64x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance64x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance64x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance64x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance8x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance8x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance8x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance8x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance128x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance128x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance128x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance128x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x4_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x8_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance32x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance32x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance32x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance32x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance32x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance32x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance32x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance4x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance4x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance4x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance4x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance4x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance4x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance4x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance64x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance64x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance64x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance64x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance64x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance64x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance64x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance8x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance8x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance8x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance8x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
@@ -6329,11 +6297,6 @@ int aom_vector_var_sse4_1(const int16_t *ref, const int16_t *src, int bwl);
 int aom_vector_var_avx2(const int16_t *ref, const int16_t *src, int bwl);
 RTCD_EXTERN int (*aom_vector_var)(const int16_t *ref, const int16_t *src, int bwl);
 
-double av1_compute_cross_correlation_c(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
-double av1_compute_cross_correlation_sse4_1(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
-double av1_compute_cross_correlation_avx2(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
-RTCD_EXTERN double (*av1_compute_cross_correlation)(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
-
 void aom_dsp_rtcd(void);
 
 #ifdef RTCD_C
@@ -6358,12 +6321,19 @@ static void setup_rtcd_internal(void)
     aom_comp_mask_pred = aom_comp_mask_pred_c;
     if (flags & HAS_SSSE3) aom_comp_mask_pred = aom_comp_mask_pred_ssse3;
     if (flags & HAS_AVX2) aom_comp_mask_pred = aom_comp_mask_pred_avx2;
+    aom_compute_correlation = aom_compute_correlation_c;
+    if (flags & HAS_SSE4_1) aom_compute_correlation = aom_compute_correlation_sse4_1;
+    if (flags & HAS_AVX2) aom_compute_correlation = aom_compute_correlation_avx2;
     aom_compute_flow_at_point = aom_compute_flow_at_point_c;
     if (flags & HAS_SSE4_1) aom_compute_flow_at_point = aom_compute_flow_at_point_sse4_1;
-    aom_convolve8_horiz = aom_convolve8_horiz_sse2;
+    if (flags & HAS_AVX2) aom_compute_flow_at_point = aom_compute_flow_at_point_avx2;
+    aom_compute_mean_stddev = aom_compute_mean_stddev_c;
+    if (flags & HAS_SSE4_1) aom_compute_mean_stddev = aom_compute_mean_stddev_sse4_1;
+    if (flags & HAS_AVX2) aom_compute_mean_stddev = aom_compute_mean_stddev_avx2;
+    aom_convolve8_horiz = aom_convolve8_horiz_c;
     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;
+    aom_convolve8_vert = aom_convolve8_vert_c;
     if (flags & HAS_SSSE3) aom_convolve8_vert = aom_convolve8_vert_ssse3;
     if (flags & HAS_AVX2) aom_convolve8_vert = aom_convolve8_vert_avx2;
     aom_convolve_copy = aom_convolve_copy_sse2;
@@ -6528,6 +6498,8 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance8x4 = aom_highbd_10_masked_sub_pixel_variance8x4_ssse3;
     aom_highbd_10_masked_sub_pixel_variance8x8 = aom_highbd_10_masked_sub_pixel_variance8x8_c;
     if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance8x8 = aom_highbd_10_masked_sub_pixel_variance8x8_ssse3;
+    aom_highbd_10_mse16x16 = aom_highbd_10_mse16x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_mse16x16 = aom_highbd_10_mse16x16_avx2;
     aom_highbd_10_obmc_variance128x128 = aom_highbd_10_obmc_variance128x128_c;
     if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance128x128 = aom_highbd_10_obmc_variance128x128_sse4_1;
     aom_highbd_10_obmc_variance128x64 = aom_highbd_10_obmc_variance128x64_c;
@@ -7626,6 +7598,8 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_AVX2) aom_sad16x32x3d = aom_sad16x32x3d_avx2;
     aom_sad16x32x4d = aom_sad16x32x4d_sse2;
     if (flags & HAS_AVX2) aom_sad16x32x4d = aom_sad16x32x4d_avx2;
+    aom_sad16x4x3d = aom_sad16x4x3d_c;
+    if (flags & HAS_AVX2) aom_sad16x4x3d = aom_sad16x4x3d_avx2;
     aom_sad16x4x4d = aom_sad16x4x4d_sse2;
     if (flags & HAS_AVX2) aom_sad16x4x4d = aom_sad16x4x4d_avx2;
     aom_sad16x64x3d = aom_sad16x64x3d_c;
@@ -7704,6 +7678,8 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_AVX2) aom_sad_skip_16x16x4d = aom_sad_skip_16x16x4d_avx2;
     aom_sad_skip_16x32x4d = aom_sad_skip_16x32x4d_sse2;
     if (flags & HAS_AVX2) aom_sad_skip_16x32x4d = aom_sad_skip_16x32x4d_avx2;
+    aom_sad_skip_16x4x4d = aom_sad_skip_16x4x4d_c;
+    if (flags & HAS_AVX2) aom_sad_skip_16x4x4d = aom_sad_skip_16x4x4d_avx2;
     aom_sad_skip_16x64x4d = aom_sad_skip_16x64x4d_sse2;
     if (flags & HAS_AVX2) aom_sad_skip_16x64x4d = aom_sad_skip_16x64x4d_avx2;
     aom_sad_skip_16x8x4d = aom_sad_skip_16x8x4d_sse2;
@@ -7859,114 +7835,114 @@ static void setup_rtcd_internal(void)
     aom_sse = aom_sse_c;
     if (flags & HAS_SSE4_1) aom_sse = aom_sse_sse4_1;
     if (flags & HAS_AVX2) aom_sse = aom_sse_avx2;
-    aom_sub_pixel_avg_variance128x128 = aom_sub_pixel_avg_variance128x128_sse2;
+    aom_sub_pixel_avg_variance128x128 = aom_sub_pixel_avg_variance128x128_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance128x128 = aom_sub_pixel_avg_variance128x128_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_avg_variance128x128 = aom_sub_pixel_avg_variance128x128_avx2;
-    aom_sub_pixel_avg_variance128x64 = aom_sub_pixel_avg_variance128x64_sse2;
+    aom_sub_pixel_avg_variance128x64 = aom_sub_pixel_avg_variance128x64_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance128x64 = aom_sub_pixel_avg_variance128x64_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_avg_variance128x64 = aom_sub_pixel_avg_variance128x64_avx2;
-    aom_sub_pixel_avg_variance16x16 = aom_sub_pixel_avg_variance16x16_sse2;
+    aom_sub_pixel_avg_variance16x16 = aom_sub_pixel_avg_variance16x16_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x16 = aom_sub_pixel_avg_variance16x16_ssse3;
-    aom_sub_pixel_avg_variance16x32 = aom_sub_pixel_avg_variance16x32_sse2;
+    aom_sub_pixel_avg_variance16x32 = aom_sub_pixel_avg_variance16x32_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x32 = aom_sub_pixel_avg_variance16x32_ssse3;
-    aom_sub_pixel_avg_variance16x4 = aom_sub_pixel_avg_variance16x4_sse2;
+    aom_sub_pixel_avg_variance16x4 = aom_sub_pixel_avg_variance16x4_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x4 = aom_sub_pixel_avg_variance16x4_ssse3;
-    aom_sub_pixel_avg_variance16x64 = aom_sub_pixel_avg_variance16x64_sse2;
+    aom_sub_pixel_avg_variance16x64 = aom_sub_pixel_avg_variance16x64_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x64 = aom_sub_pixel_avg_variance16x64_ssse3;
-    aom_sub_pixel_avg_variance16x8 = aom_sub_pixel_avg_variance16x8_sse2;
+    aom_sub_pixel_avg_variance16x8 = aom_sub_pixel_avg_variance16x8_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x8 = aom_sub_pixel_avg_variance16x8_ssse3;
-    aom_sub_pixel_avg_variance32x16 = aom_sub_pixel_avg_variance32x16_sse2;
+    aom_sub_pixel_avg_variance32x16 = aom_sub_pixel_avg_variance32x16_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x16 = aom_sub_pixel_avg_variance32x16_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_avg_variance32x16 = aom_sub_pixel_avg_variance32x16_avx2;
-    aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_sse2;
+    aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_avx2;
-    aom_sub_pixel_avg_variance32x64 = aom_sub_pixel_avg_variance32x64_sse2;
+    aom_sub_pixel_avg_variance32x64 = aom_sub_pixel_avg_variance32x64_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x64 = aom_sub_pixel_avg_variance32x64_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_avg_variance32x64 = aom_sub_pixel_avg_variance32x64_avx2;
-    aom_sub_pixel_avg_variance32x8 = aom_sub_pixel_avg_variance32x8_sse2;
+    aom_sub_pixel_avg_variance32x8 = aom_sub_pixel_avg_variance32x8_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x8 = aom_sub_pixel_avg_variance32x8_ssse3;
-    aom_sub_pixel_avg_variance4x16 = aom_sub_pixel_avg_variance4x16_sse2;
+    aom_sub_pixel_avg_variance4x16 = aom_sub_pixel_avg_variance4x16_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance4x16 = aom_sub_pixel_avg_variance4x16_ssse3;
-    aom_sub_pixel_avg_variance4x4 = aom_sub_pixel_avg_variance4x4_sse2;
+    aom_sub_pixel_avg_variance4x4 = aom_sub_pixel_avg_variance4x4_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance4x4 = aom_sub_pixel_avg_variance4x4_ssse3;
-    aom_sub_pixel_avg_variance4x8 = aom_sub_pixel_avg_variance4x8_sse2;
+    aom_sub_pixel_avg_variance4x8 = aom_sub_pixel_avg_variance4x8_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance4x8 = aom_sub_pixel_avg_variance4x8_ssse3;
-    aom_sub_pixel_avg_variance64x128 = aom_sub_pixel_avg_variance64x128_sse2;
+    aom_sub_pixel_avg_variance64x128 = aom_sub_pixel_avg_variance64x128_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance64x128 = aom_sub_pixel_avg_variance64x128_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_avg_variance64x128 = aom_sub_pixel_avg_variance64x128_avx2;
-    aom_sub_pixel_avg_variance64x16 = aom_sub_pixel_avg_variance64x16_sse2;
+    aom_sub_pixel_avg_variance64x16 = aom_sub_pixel_avg_variance64x16_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance64x16 = aom_sub_pixel_avg_variance64x16_ssse3;
-    aom_sub_pixel_avg_variance64x32 = aom_sub_pixel_avg_variance64x32_sse2;
+    aom_sub_pixel_avg_variance64x32 = aom_sub_pixel_avg_variance64x32_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance64x32 = aom_sub_pixel_avg_variance64x32_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_avg_variance64x32 = aom_sub_pixel_avg_variance64x32_avx2;
-    aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_sse2;
+    aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_avx2;
-    aom_sub_pixel_avg_variance8x16 = aom_sub_pixel_avg_variance8x16_sse2;
+    aom_sub_pixel_avg_variance8x16 = aom_sub_pixel_avg_variance8x16_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x16 = aom_sub_pixel_avg_variance8x16_ssse3;
-    aom_sub_pixel_avg_variance8x32 = aom_sub_pixel_avg_variance8x32_sse2;
+    aom_sub_pixel_avg_variance8x32 = aom_sub_pixel_avg_variance8x32_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x32 = aom_sub_pixel_avg_variance8x32_ssse3;
-    aom_sub_pixel_avg_variance8x4 = aom_sub_pixel_avg_variance8x4_sse2;
+    aom_sub_pixel_avg_variance8x4 = aom_sub_pixel_avg_variance8x4_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x4 = aom_sub_pixel_avg_variance8x4_ssse3;
-    aom_sub_pixel_avg_variance8x8 = aom_sub_pixel_avg_variance8x8_sse2;
+    aom_sub_pixel_avg_variance8x8 = aom_sub_pixel_avg_variance8x8_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x8 = aom_sub_pixel_avg_variance8x8_ssse3;
-    aom_sub_pixel_variance128x128 = aom_sub_pixel_variance128x128_sse2;
+    aom_sub_pixel_variance128x128 = aom_sub_pixel_variance128x128_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance128x128 = aom_sub_pixel_variance128x128_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance128x128 = aom_sub_pixel_variance128x128_avx2;
-    aom_sub_pixel_variance128x64 = aom_sub_pixel_variance128x64_sse2;
+    aom_sub_pixel_variance128x64 = aom_sub_pixel_variance128x64_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance128x64 = aom_sub_pixel_variance128x64_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance128x64 = aom_sub_pixel_variance128x64_avx2;
-    aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_sse2;
+    aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_avx2;
-    aom_sub_pixel_variance16x32 = aom_sub_pixel_variance16x32_sse2;
+    aom_sub_pixel_variance16x32 = aom_sub_pixel_variance16x32_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance16x32 = aom_sub_pixel_variance16x32_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance16x32 = aom_sub_pixel_variance16x32_avx2;
-    aom_sub_pixel_variance16x4 = aom_sub_pixel_variance16x4_sse2;
+    aom_sub_pixel_variance16x4 = aom_sub_pixel_variance16x4_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance16x4 = aom_sub_pixel_variance16x4_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance16x4 = aom_sub_pixel_variance16x4_avx2;
-    aom_sub_pixel_variance16x64 = aom_sub_pixel_variance16x64_sse2;
+    aom_sub_pixel_variance16x64 = aom_sub_pixel_variance16x64_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance16x64 = aom_sub_pixel_variance16x64_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance16x64 = aom_sub_pixel_variance16x64_avx2;
-    aom_sub_pixel_variance16x8 = aom_sub_pixel_variance16x8_sse2;
+    aom_sub_pixel_variance16x8 = aom_sub_pixel_variance16x8_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance16x8 = aom_sub_pixel_variance16x8_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance16x8 = aom_sub_pixel_variance16x8_avx2;
-    aom_sub_pixel_variance32x16 = aom_sub_pixel_variance32x16_sse2;
+    aom_sub_pixel_variance32x16 = aom_sub_pixel_variance32x16_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance32x16 = aom_sub_pixel_variance32x16_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance32x16 = aom_sub_pixel_variance32x16_avx2;
-    aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_sse2;
+    aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_avx2;
-    aom_sub_pixel_variance32x64 = aom_sub_pixel_variance32x64_sse2;
+    aom_sub_pixel_variance32x64 = aom_sub_pixel_variance32x64_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance32x64 = aom_sub_pixel_variance32x64_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance32x64 = aom_sub_pixel_variance32x64_avx2;
-    aom_sub_pixel_variance32x8 = aom_sub_pixel_variance32x8_sse2;
+    aom_sub_pixel_variance32x8 = aom_sub_pixel_variance32x8_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance32x8 = aom_sub_pixel_variance32x8_ssse3;
-    aom_sub_pixel_variance4x16 = aom_sub_pixel_variance4x16_sse2;
+    aom_sub_pixel_variance4x16 = aom_sub_pixel_variance4x16_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance4x16 = aom_sub_pixel_variance4x16_ssse3;
-    aom_sub_pixel_variance4x4 = aom_sub_pixel_variance4x4_sse2;
+    aom_sub_pixel_variance4x4 = aom_sub_pixel_variance4x4_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance4x4 = aom_sub_pixel_variance4x4_ssse3;
-    aom_sub_pixel_variance4x8 = aom_sub_pixel_variance4x8_sse2;
+    aom_sub_pixel_variance4x8 = aom_sub_pixel_variance4x8_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance4x8 = aom_sub_pixel_variance4x8_ssse3;
-    aom_sub_pixel_variance64x128 = aom_sub_pixel_variance64x128_sse2;
+    aom_sub_pixel_variance64x128 = aom_sub_pixel_variance64x128_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance64x128 = aom_sub_pixel_variance64x128_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance64x128 = aom_sub_pixel_variance64x128_avx2;
-    aom_sub_pixel_variance64x16 = aom_sub_pixel_variance64x16_sse2;
+    aom_sub_pixel_variance64x16 = aom_sub_pixel_variance64x16_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance64x16 = aom_sub_pixel_variance64x16_ssse3;
-    aom_sub_pixel_variance64x32 = aom_sub_pixel_variance64x32_sse2;
+    aom_sub_pixel_variance64x32 = aom_sub_pixel_variance64x32_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance64x32 = aom_sub_pixel_variance64x32_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance64x32 = aom_sub_pixel_variance64x32_avx2;
-    aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_sse2;
+    aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_avx2;
-    aom_sub_pixel_variance8x16 = aom_sub_pixel_variance8x16_sse2;
+    aom_sub_pixel_variance8x16 = aom_sub_pixel_variance8x16_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance8x16 = aom_sub_pixel_variance8x16_ssse3;
-    aom_sub_pixel_variance8x32 = aom_sub_pixel_variance8x32_sse2;
+    aom_sub_pixel_variance8x32 = aom_sub_pixel_variance8x32_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance8x32 = aom_sub_pixel_variance8x32_ssse3;
-    aom_sub_pixel_variance8x4 = aom_sub_pixel_variance8x4_sse2;
+    aom_sub_pixel_variance8x4 = aom_sub_pixel_variance8x4_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance8x4 = aom_sub_pixel_variance8x4_ssse3;
-    aom_sub_pixel_variance8x8 = aom_sub_pixel_variance8x8_sse2;
+    aom_sub_pixel_variance8x8 = aom_sub_pixel_variance8x8_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance8x8 = aom_sub_pixel_variance8x8_ssse3;
     aom_subtract_block = aom_subtract_block_sse2;
     if (flags & HAS_AVX2) aom_subtract_block = aom_subtract_block_avx2;
@@ -8023,9 +7999,6 @@ static void setup_rtcd_internal(void)
     aom_vector_var = aom_vector_var_c;
     if (flags & HAS_SSE4_1) aom_vector_var = aom_vector_var_sse4_1;
     if (flags & HAS_AVX2) aom_vector_var = aom_vector_var_avx2;
-    av1_compute_cross_correlation = av1_compute_cross_correlation_c;
-    if (flags & HAS_SSE4_1) av1_compute_cross_correlation = av1_compute_cross_correlation_sse4_1;
-    if (flags & HAS_AVX2) av1_compute_cross_correlation = av1_compute_cross_correlation_avx2;
 }
 #endif
 
diff --git a/media/libaom/config/mac/x64/config/aom_scale_rtcd.h b/media/libaom/config/mac/x64/config/aom_scale_rtcd.h
index 3b70fb47c3..cdabb21106 100644
--- a/media/libaom/config/mac/x64/config/aom_scale_rtcd.h
+++ b/media/libaom/config/mac/x64/config/aom_scale_rtcd.h
@@ -8,13 +8,15 @@
 #define RTCD_EXTERN extern
 #endif
 
+#include <stdbool.h>
+
 struct yv12_buffer_config;
 
 #ifdef __cplusplus
 extern "C" {
 #endif
 
-void aom_extend_frame_borders_c(struct yv12_buffer_config *ybf, const int num_planes);
+void aom_extend_frame_borders_c(struct yv12_buffer_config *ybf, int num_planes);
 #define aom_extend_frame_borders aom_extend_frame_borders_c
 
 void aom_extend_frame_borders_plane_row_c(const struct yv12_buffer_config *ybf, int plane, int v_start, int v_end);
@@ -50,13 +52,13 @@ void aom_vertical_band_5_4_scale_c(unsigned char *source, int src_pitch, unsigne
 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);
+void aom_yv12_copy_u_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc, int use_crop);
 #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);
+void aom_yv12_copy_v_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc, int use_crop);
 #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);
+void aom_yv12_copy_y_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc, int use_crop);
 #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);
@@ -80,7 +82,7 @@ void aom_yv12_partial_copy_v_c(const struct yv12_buffer_config *src_bc, int hsta
 void aom_yv12_partial_copy_y_c(const struct yv12_buffer_config *src_ybc, int hstart1, int hend1, int vstart1, int vend1, struct yv12_buffer_config *dst_ybc, int hstart2, int vstart2);
 #define aom_yv12_partial_copy_y aom_yv12_partial_copy_y_c
 
-int aom_yv12_realloc_with_new_border_c(struct yv12_buffer_config *ybf, int new_border, int byte_alignment, int num_pyramid_levels, int num_planes);
+int aom_yv12_realloc_with_new_border_c(struct yv12_buffer_config *ybf, int new_border, int byte_alignment, bool alloc_pyramid, int num_planes);
 #define aom_yv12_realloc_with_new_border aom_yv12_realloc_with_new_border_c
 
 void aom_scale_rtcd(void);
diff --git a/media/libaom/config/mac/x64/config/av1_rtcd.h b/media/libaom/config/mac/x64/config/av1_rtcd.h
index b1cdc99700..ad72985afe 100644
--- a/media/libaom/config/mac/x64/config/av1_rtcd.h
+++ b/media/libaom/config/mac/x64/config/av1_rtcd.h
@@ -253,7 +253,6 @@ void av1_convolve_y_sr_intrabc_c(const uint8_t *src, int src_stride, uint8_t *ds
 #define av1_convolve_y_sr_intrabc av1_convolve_y_sr_intrabc_c
 
 void av1_dist_wtd_convolve_2d_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
-void av1_dist_wtd_convolve_2d_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
 void av1_dist_wtd_convolve_2d_ssse3(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
 void av1_dist_wtd_convolve_2d_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
 RTCD_EXTERN void (*av1_dist_wtd_convolve_2d)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
@@ -659,7 +658,6 @@ void av1_inv_txfm_add_avx2(const tran_low_t *dqcoeff, uint8_t *dst, int stride,
 RTCD_EXTERN void (*av1_inv_txfm_add)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
 
 void av1_lowbd_fwd_txfm_c(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
-void av1_lowbd_fwd_txfm_sse2(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
 void av1_lowbd_fwd_txfm_sse4_1(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
 void av1_lowbd_fwd_txfm_avx2(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
 RTCD_EXTERN void (*av1_lowbd_fwd_txfm)(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
@@ -755,85 +753,61 @@ void av1_wiener_convolve_add_src_avx2(const uint8_t *src, ptrdiff_t src_stride,
 RTCD_EXTERN void (*av1_wiener_convolve_add_src)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params);
 
 void cdef_copy_rect8_16bit_to_16bit_c(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
-void cdef_copy_rect8_16bit_to_16bit_sse2(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
-void cdef_copy_rect8_16bit_to_16bit_ssse3(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
 void cdef_copy_rect8_16bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
 void cdef_copy_rect8_16bit_to_16bit_avx2(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
 RTCD_EXTERN void (*cdef_copy_rect8_16bit_to_16bit)(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
 
 void cdef_copy_rect8_8bit_to_16bit_c(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
-void cdef_copy_rect8_8bit_to_16bit_sse2(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
-void cdef_copy_rect8_8bit_to_16bit_ssse3(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
 void cdef_copy_rect8_8bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
 void cdef_copy_rect8_8bit_to_16bit_avx2(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
 RTCD_EXTERN void (*cdef_copy_rect8_8bit_to_16bit)(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
 
 void cdef_filter_16_0_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_16_0_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_16_0_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_0_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_0_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 RTCD_EXTERN void (*cdef_filter_16_0)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 
 void cdef_filter_16_1_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_16_1_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_16_1_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_1_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_1_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 RTCD_EXTERN void (*cdef_filter_16_1)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 
 void cdef_filter_16_2_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_16_2_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_16_2_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_2_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_2_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 RTCD_EXTERN void (*cdef_filter_16_2)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 
 void cdef_filter_16_3_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_16_3_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_16_3_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_3_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_3_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 RTCD_EXTERN void (*cdef_filter_16_3)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 
 void cdef_filter_8_0_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_8_0_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_8_0_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_0_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_0_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 RTCD_EXTERN void (*cdef_filter_8_0)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 
 void cdef_filter_8_1_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_8_1_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_8_1_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_1_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_1_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 RTCD_EXTERN void (*cdef_filter_8_1)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 
 void cdef_filter_8_2_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_8_2_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_8_2_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_2_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_2_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 RTCD_EXTERN void (*cdef_filter_8_2)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 
 void cdef_filter_8_3_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_8_3_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_8_3_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_3_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_3_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 RTCD_EXTERN void (*cdef_filter_8_3)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 
 int cdef_find_dir_c(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
-int cdef_find_dir_sse2(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
-int cdef_find_dir_ssse3(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
 int cdef_find_dir_sse4_1(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
 int cdef_find_dir_avx2(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
 RTCD_EXTERN int (*cdef_find_dir)(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
 
 void cdef_find_dir_dual_c(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
-void cdef_find_dir_dual_sse2(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
-void cdef_find_dir_dual_ssse3(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
 void cdef_find_dir_dual_sse4_1(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
 void cdef_find_dir_dual_avx2(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
 RTCD_EXTERN void (*cdef_find_dir_dual)(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
@@ -941,7 +915,7 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_AVX2) av1_convolve_x_sr = av1_convolve_x_sr_avx2;
     av1_convolve_y_sr = av1_convolve_y_sr_sse2;
     if (flags & HAS_AVX2) av1_convolve_y_sr = av1_convolve_y_sr_avx2;
-    av1_dist_wtd_convolve_2d = av1_dist_wtd_convolve_2d_sse2;
+    av1_dist_wtd_convolve_2d = av1_dist_wtd_convolve_2d_c;
     if (flags & HAS_SSSE3) av1_dist_wtd_convolve_2d = av1_dist_wtd_convolve_2d_ssse3;
     if (flags & HAS_AVX2) av1_dist_wtd_convolve_2d = av1_dist_wtd_convolve_2d_avx2;
     av1_dist_wtd_convolve_2d_copy = av1_dist_wtd_convolve_2d_copy_sse2;
@@ -1091,7 +1065,7 @@ static void setup_rtcd_internal(void)
     av1_inv_txfm_add = av1_inv_txfm_add_c;
     if (flags & HAS_SSSE3) av1_inv_txfm_add = av1_inv_txfm_add_ssse3;
     if (flags & HAS_AVX2) av1_inv_txfm_add = av1_inv_txfm_add_avx2;
-    av1_lowbd_fwd_txfm = av1_lowbd_fwd_txfm_sse2;
+    av1_lowbd_fwd_txfm = av1_lowbd_fwd_txfm_c;
     if (flags & HAS_SSE4_1) av1_lowbd_fwd_txfm = av1_lowbd_fwd_txfm_sse4_1;
     if (flags & HAS_AVX2) av1_lowbd_fwd_txfm = av1_lowbd_fwd_txfm_avx2;
     av1_lowbd_pixel_proj_error = av1_lowbd_pixel_proj_error_c;
@@ -1133,52 +1107,40 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_AVX2) av1_wedge_sse_from_residuals = av1_wedge_sse_from_residuals_avx2;
     av1_wiener_convolve_add_src = av1_wiener_convolve_add_src_sse2;
     if (flags & HAS_AVX2) av1_wiener_convolve_add_src = av1_wiener_convolve_add_src_avx2;
-    cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_sse2;
-    if (flags & HAS_SSSE3) cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_ssse3;
+    cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_c;
     if (flags & HAS_SSE4_1) cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_sse4_1;
     if (flags & HAS_AVX2) cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_avx2;
-    cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_sse2;
-    if (flags & HAS_SSSE3) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_ssse3;
+    cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_c;
     if (flags & HAS_SSE4_1) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_sse4_1;
     if (flags & HAS_AVX2) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_avx2;
-    cdef_filter_16_0 = cdef_filter_16_0_sse2;
-    if (flags & HAS_SSSE3) cdef_filter_16_0 = cdef_filter_16_0_ssse3;
+    cdef_filter_16_0 = cdef_filter_16_0_c;
     if (flags & HAS_SSE4_1) cdef_filter_16_0 = cdef_filter_16_0_sse4_1;
     if (flags & HAS_AVX2) cdef_filter_16_0 = cdef_filter_16_0_avx2;
-    cdef_filter_16_1 = cdef_filter_16_1_sse2;
-    if (flags & HAS_SSSE3) cdef_filter_16_1 = cdef_filter_16_1_ssse3;
+    cdef_filter_16_1 = cdef_filter_16_1_c;
     if (flags & HAS_SSE4_1) cdef_filter_16_1 = cdef_filter_16_1_sse4_1;
     if (flags & HAS_AVX2) cdef_filter_16_1 = cdef_filter_16_1_avx2;
-    cdef_filter_16_2 = cdef_filter_16_2_sse2;
-    if (flags & HAS_SSSE3) cdef_filter_16_2 = cdef_filter_16_2_ssse3;
+    cdef_filter_16_2 = cdef_filter_16_2_c;
     if (flags & HAS_SSE4_1) cdef_filter_16_2 = cdef_filter_16_2_sse4_1;
     if (flags & HAS_AVX2) cdef_filter_16_2 = cdef_filter_16_2_avx2;
-    cdef_filter_16_3 = cdef_filter_16_3_sse2;
-    if (flags & HAS_SSSE3) cdef_filter_16_3 = cdef_filter_16_3_ssse3;
+    cdef_filter_16_3 = cdef_filter_16_3_c;
     if (flags & HAS_SSE4_1) cdef_filter_16_3 = cdef_filter_16_3_sse4_1;
     if (flags & HAS_AVX2) cdef_filter_16_3 = cdef_filter_16_3_avx2;
-    cdef_filter_8_0 = cdef_filter_8_0_sse2;
-    if (flags & HAS_SSSE3) cdef_filter_8_0 = cdef_filter_8_0_ssse3;
+    cdef_filter_8_0 = cdef_filter_8_0_c;
     if (flags & HAS_SSE4_1) cdef_filter_8_0 = cdef_filter_8_0_sse4_1;
     if (flags & HAS_AVX2) cdef_filter_8_0 = cdef_filter_8_0_avx2;
-    cdef_filter_8_1 = cdef_filter_8_1_sse2;
-    if (flags & HAS_SSSE3) cdef_filter_8_1 = cdef_filter_8_1_ssse3;
+    cdef_filter_8_1 = cdef_filter_8_1_c;
     if (flags & HAS_SSE4_1) cdef_filter_8_1 = cdef_filter_8_1_sse4_1;
     if (flags & HAS_AVX2) cdef_filter_8_1 = cdef_filter_8_1_avx2;
-    cdef_filter_8_2 = cdef_filter_8_2_sse2;
-    if (flags & HAS_SSSE3) cdef_filter_8_2 = cdef_filter_8_2_ssse3;
+    cdef_filter_8_2 = cdef_filter_8_2_c;
     if (flags & HAS_SSE4_1) cdef_filter_8_2 = cdef_filter_8_2_sse4_1;
     if (flags & HAS_AVX2) cdef_filter_8_2 = cdef_filter_8_2_avx2;
-    cdef_filter_8_3 = cdef_filter_8_3_sse2;
-    if (flags & HAS_SSSE3) cdef_filter_8_3 = cdef_filter_8_3_ssse3;
+    cdef_filter_8_3 = cdef_filter_8_3_c;
     if (flags & HAS_SSE4_1) cdef_filter_8_3 = cdef_filter_8_3_sse4_1;
     if (flags & HAS_AVX2) cdef_filter_8_3 = cdef_filter_8_3_avx2;
-    cdef_find_dir = cdef_find_dir_sse2;
-    if (flags & HAS_SSSE3) cdef_find_dir = cdef_find_dir_ssse3;
+    cdef_find_dir = cdef_find_dir_c;
     if (flags & HAS_SSE4_1) cdef_find_dir = cdef_find_dir_sse4_1;
     if (flags & HAS_AVX2) cdef_find_dir = cdef_find_dir_avx2;
-    cdef_find_dir_dual = cdef_find_dir_dual_sse2;
-    if (flags & HAS_SSSE3) cdef_find_dir_dual = cdef_find_dir_dual_ssse3;
+    cdef_find_dir_dual = cdef_find_dir_dual_c;
     if (flags & HAS_SSE4_1) cdef_find_dir_dual = cdef_find_dir_dual_sse4_1;
     if (flags & HAS_AVX2) cdef_find_dir_dual = cdef_find_dir_dual_avx2;
     cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_c;
diff --git a/media/libaom/config/win/ia32/config/aom_config.asm b/media/libaom/config/win/ia32/config/aom_config.asm
index af78328283..8f6c3592fa 100644
--- a/media/libaom/config/win/ia32/config/aom_config.asm
+++ b/media/libaom/config/win/ia32/config/aom_config.asm
@@ -53,6 +53,7 @@ CONFIG_OS_SUPPORT equ 1
 CONFIG_OUTPUT_FRAME_SIZE equ 0
 CONFIG_PARTITION_SEARCH_ORDER equ 0
 CONFIG_PIC equ 0
+CONFIG_QUANT_MATRIX equ 1
 CONFIG_RATECTRL_LOG equ 0
 CONFIG_RD_COMMAND equ 0
 CONFIG_RD_DEBUG equ 0
@@ -87,6 +88,7 @@ HAVE_SSE4_1 equ 1
 HAVE_SSE4_2 equ 1
 HAVE_SSSE3 equ 1
 HAVE_SVE equ 0
+HAVE_SVE2 equ 0
 HAVE_VSX equ 0
 HAVE_WXWIDGETS equ 0
 STATIC_LINK_JXL equ 0
diff --git a/media/libaom/config/win/ia32/config/aom_config.h b/media/libaom/config/win/ia32/config/aom_config.h
index dba805b1b6..7d1ce61373 100644
--- a/media/libaom/config/win/ia32/config/aom_config.h
+++ b/media/libaom/config/win/ia32/config/aom_config.h
@@ -55,6 +55,7 @@
 #define CONFIG_OUTPUT_FRAME_SIZE 0
 #define CONFIG_PARTITION_SEARCH_ORDER 0
 #define CONFIG_PIC 0
+#define CONFIG_QUANT_MATRIX 1
 #define CONFIG_RATECTRL_LOG 0
 #define CONFIG_RD_COMMAND 0
 #define CONFIG_RD_DEBUG 0
@@ -89,6 +90,7 @@
 #define HAVE_SSE4_2 1
 #define HAVE_SSSE3 1
 #define HAVE_SVE 0
+#define HAVE_SVE2 0
 #define HAVE_VSX 0
 #define HAVE_WXWIDGETS 0
 #define INLINE inline
diff --git a/media/libaom/config/win/ia32/config/aom_dsp_rtcd.h b/media/libaom/config/win/ia32/config/aom_dsp_rtcd.h
index a19adf5f61..93472f0e92 100644
--- a/media/libaom/config/win/ia32/config/aom_dsp_rtcd.h
+++ b/media/libaom/config/win/ia32/config/aom_dsp_rtcd.h
@@ -57,21 +57,30 @@ void aom_comp_mask_pred_ssse3(uint8_t *comp_pred, const uint8_t *pred, int width
 void aom_comp_mask_pred_avx2(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
 RTCD_EXTERN void (*aom_comp_mask_pred)(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
 
+double aom_compute_correlation_c(const unsigned char *frame1, int stride1, int x1, int y1, double mean1, double one_over_stddev1, const unsigned char *frame2, int stride2, int x2, int y2, double mean2, double one_over_stddev2);
+double aom_compute_correlation_sse4_1(const unsigned char *frame1, int stride1, int x1, int y1, double mean1, double one_over_stddev1, const unsigned char *frame2, int stride2, int x2, int y2, double mean2, double one_over_stddev2);
+double aom_compute_correlation_avx2(const unsigned char *frame1, int stride1, int x1, int y1, double mean1, double one_over_stddev1, const unsigned char *frame2, int stride2, int x2, int y2, double mean2, double one_over_stddev2);
+RTCD_EXTERN double (*aom_compute_correlation)(const unsigned char *frame1, int stride1, int x1, int y1, double mean1, double one_over_stddev1, const unsigned char *frame2, int stride2, int x2, int y2, double mean2, double one_over_stddev2);
+
 void aom_compute_flow_at_point_c(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
 void aom_compute_flow_at_point_sse4_1(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
+void aom_compute_flow_at_point_avx2(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
 RTCD_EXTERN void (*aom_compute_flow_at_point)(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
 
+bool aom_compute_mean_stddev_c(const unsigned char *frame, int stride, int x, int y, double *mean, double *one_over_stddev);
+bool aom_compute_mean_stddev_sse4_1(const unsigned char *frame, int stride, int x, int y, double *mean, double *one_over_stddev);
+bool aom_compute_mean_stddev_avx2(const unsigned char *frame, int stride, int x, int y, double *mean, double *one_over_stddev);
+RTCD_EXTERN bool (*aom_compute_mean_stddev)(const unsigned char *frame, int stride, int x, int y, double *mean, double *one_over_stddev);
+
 void aom_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h);
 #define aom_convolve8 aom_convolve8_c
 
 void aom_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
-void aom_convolve8_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void aom_convolve8_horiz_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void aom_convolve8_horiz_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 RTCD_EXTERN void (*aom_convolve8_horiz)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 
 void aom_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
-void aom_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void aom_convolve8_vert_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void aom_convolve8_vert_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 RTCD_EXTERN void (*aom_convolve8_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
@@ -903,6 +912,7 @@ RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance8x8)(const uin
 
 unsigned int aom_highbd_10_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
 unsigned int aom_highbd_10_mse16x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+unsigned int aom_highbd_10_mse16x16_avx2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
 RTCD_EXTERN unsigned int (*aom_highbd_10_mse16x16)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
 
 unsigned int aom_highbd_10_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
@@ -5130,7 +5140,8 @@ unsigned int aom_sad16x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const
 RTCD_EXTERN unsigned int (*aom_sad16x4_avg)(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride, const uint8_t *second_pred);
 
 void aom_sad16x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
-#define aom_sad16x4x3d aom_sad16x4x3d_c
+void aom_sad16x4x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x4x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
 
 void aom_sad16x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
 void aom_sad16x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
@@ -5466,7 +5477,8 @@ unsigned int aom_sad_skip_16x4_c(const uint8_t *src_ptr, int src_stride, const u
 #define aom_sad_skip_16x4 aom_sad_skip_16x4_c
 
 void aom_sad_skip_16x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
-#define aom_sad_skip_16x4x4d aom_sad_skip_16x4x4d_c
+void aom_sad_skip_16x4x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_16x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
 
 unsigned int aom_sad_skip_16x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
 unsigned int aom_sad_skip_16x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
@@ -5867,243 +5879,199 @@ void aom_ssim_parms_8x8_c(const uint8_t *s, int sp, const uint8_t *r, int rp, ui
 #define aom_ssim_parms_8x8 aom_ssim_parms_8x8_c
 
 uint32_t aom_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance128x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance128x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance128x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance128x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance16x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance16x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance16x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance16x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance16x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance32x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance32x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance32x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance32x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance32x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance32x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance32x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance4x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance4x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance4x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance4x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance4x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance4x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance4x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance64x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance64x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance64x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance64x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance64x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance64x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance64x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance8x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance8x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance8x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance8x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance128x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance128x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance128x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance128x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x4_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x8_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance32x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance32x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance32x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance32x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance32x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance32x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance32x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance4x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance4x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance4x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance4x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance4x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance4x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance4x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance64x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance64x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance64x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance64x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance64x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance64x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance64x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance8x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance8x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance8x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance8x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
@@ -6326,11 +6294,6 @@ int aom_vector_var_sse4_1(const int16_t *ref, const int16_t *src, int bwl);
 int aom_vector_var_avx2(const int16_t *ref, const int16_t *src, int bwl);
 RTCD_EXTERN int (*aom_vector_var)(const int16_t *ref, const int16_t *src, int bwl);
 
-double av1_compute_cross_correlation_c(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
-double av1_compute_cross_correlation_sse4_1(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
-double av1_compute_cross_correlation_avx2(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
-RTCD_EXTERN double (*av1_compute_cross_correlation)(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
-
 void aom_dsp_rtcd(void);
 
 #ifdef RTCD_C
@@ -6360,14 +6323,19 @@ static void setup_rtcd_internal(void)
     aom_comp_mask_pred = aom_comp_mask_pred_c;
     if (flags & HAS_SSSE3) aom_comp_mask_pred = aom_comp_mask_pred_ssse3;
     if (flags & HAS_AVX2) aom_comp_mask_pred = aom_comp_mask_pred_avx2;
+    aom_compute_correlation = aom_compute_correlation_c;
+    if (flags & HAS_SSE4_1) aom_compute_correlation = aom_compute_correlation_sse4_1;
+    if (flags & HAS_AVX2) aom_compute_correlation = aom_compute_correlation_avx2;
     aom_compute_flow_at_point = aom_compute_flow_at_point_c;
     if (flags & HAS_SSE4_1) aom_compute_flow_at_point = aom_compute_flow_at_point_sse4_1;
+    if (flags & HAS_AVX2) aom_compute_flow_at_point = aom_compute_flow_at_point_avx2;
+    aom_compute_mean_stddev = aom_compute_mean_stddev_c;
+    if (flags & HAS_SSE4_1) aom_compute_mean_stddev = aom_compute_mean_stddev_sse4_1;
+    if (flags & HAS_AVX2) aom_compute_mean_stddev = aom_compute_mean_stddev_avx2;
     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;
@@ -6768,6 +6736,7 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance8x8 = aom_highbd_10_masked_sub_pixel_variance8x8_ssse3;
     aom_highbd_10_mse16x16 = aom_highbd_10_mse16x16_c;
     if (flags & HAS_SSE2) aom_highbd_10_mse16x16 = aom_highbd_10_mse16x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_mse16x16 = aom_highbd_10_mse16x16_avx2;
     aom_highbd_10_mse8x8 = aom_highbd_10_mse8x8_c;
     if (flags & HAS_SSE2) aom_highbd_10_mse8x8 = aom_highbd_10_mse8x8_sse2;
     aom_highbd_10_obmc_variance128x128 = aom_highbd_10_obmc_variance128x128_c;
@@ -8526,6 +8495,8 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_SSE2) aom_sad16x4 = aom_sad16x4_sse2;
     aom_sad16x4_avg = aom_sad16x4_avg_c;
     if (flags & HAS_SSE2) aom_sad16x4_avg = aom_sad16x4_avg_sse2;
+    aom_sad16x4x3d = aom_sad16x4x3d_c;
+    if (flags & HAS_AVX2) aom_sad16x4x3d = aom_sad16x4x3d_avx2;
     aom_sad16x4x4d = aom_sad16x4x4d_c;
     if (flags & HAS_SSE2) aom_sad16x4x4d = aom_sad16x4x4d_sse2;
     if (flags & HAS_AVX2) aom_sad16x4x4d = aom_sad16x4x4d_avx2;
@@ -8695,6 +8666,8 @@ static void setup_rtcd_internal(void)
     aom_sad_skip_16x32x4d = aom_sad_skip_16x32x4d_c;
     if (flags & HAS_SSE2) aom_sad_skip_16x32x4d = aom_sad_skip_16x32x4d_sse2;
     if (flags & HAS_AVX2) aom_sad_skip_16x32x4d = aom_sad_skip_16x32x4d_avx2;
+    aom_sad_skip_16x4x4d = aom_sad_skip_16x4x4d_c;
+    if (flags & HAS_AVX2) aom_sad_skip_16x4x4d = aom_sad_skip_16x4x4d_avx2;
     aom_sad_skip_16x64 = aom_sad_skip_16x64_c;
     if (flags & HAS_SSE2) aom_sad_skip_16x64 = aom_sad_skip_16x64_sse2;
     aom_sad_skip_16x64x4d = aom_sad_skip_16x64x4d_c;
@@ -8897,157 +8870,113 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_SSE4_1) aom_sse = aom_sse_sse4_1;
     if (flags & HAS_AVX2) aom_sse = aom_sse_avx2;
     aom_sub_pixel_avg_variance128x128 = aom_sub_pixel_avg_variance128x128_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance128x128 = aom_sub_pixel_avg_variance128x128_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance128x128 = aom_sub_pixel_avg_variance128x128_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_avg_variance128x128 = aom_sub_pixel_avg_variance128x128_avx2;
     aom_sub_pixel_avg_variance128x64 = aom_sub_pixel_avg_variance128x64_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance128x64 = aom_sub_pixel_avg_variance128x64_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance128x64 = aom_sub_pixel_avg_variance128x64_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_avg_variance128x64 = aom_sub_pixel_avg_variance128x64_avx2;
     aom_sub_pixel_avg_variance16x16 = aom_sub_pixel_avg_variance16x16_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance16x16 = aom_sub_pixel_avg_variance16x16_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x16 = aom_sub_pixel_avg_variance16x16_ssse3;
     aom_sub_pixel_avg_variance16x32 = aom_sub_pixel_avg_variance16x32_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance16x32 = aom_sub_pixel_avg_variance16x32_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x32 = aom_sub_pixel_avg_variance16x32_ssse3;
     aom_sub_pixel_avg_variance16x4 = aom_sub_pixel_avg_variance16x4_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance16x4 = aom_sub_pixel_avg_variance16x4_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x4 = aom_sub_pixel_avg_variance16x4_ssse3;
     aom_sub_pixel_avg_variance16x64 = aom_sub_pixel_avg_variance16x64_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance16x64 = aom_sub_pixel_avg_variance16x64_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x64 = aom_sub_pixel_avg_variance16x64_ssse3;
     aom_sub_pixel_avg_variance16x8 = aom_sub_pixel_avg_variance16x8_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance16x8 = aom_sub_pixel_avg_variance16x8_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x8 = aom_sub_pixel_avg_variance16x8_ssse3;
     aom_sub_pixel_avg_variance32x16 = aom_sub_pixel_avg_variance32x16_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance32x16 = aom_sub_pixel_avg_variance32x16_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x16 = aom_sub_pixel_avg_variance32x16_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_avg_variance32x16 = aom_sub_pixel_avg_variance32x16_avx2;
     aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_avx2;
     aom_sub_pixel_avg_variance32x64 = aom_sub_pixel_avg_variance32x64_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance32x64 = aom_sub_pixel_avg_variance32x64_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x64 = aom_sub_pixel_avg_variance32x64_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_avg_variance32x64 = aom_sub_pixel_avg_variance32x64_avx2;
     aom_sub_pixel_avg_variance32x8 = aom_sub_pixel_avg_variance32x8_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance32x8 = aom_sub_pixel_avg_variance32x8_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x8 = aom_sub_pixel_avg_variance32x8_ssse3;
     aom_sub_pixel_avg_variance4x16 = aom_sub_pixel_avg_variance4x16_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance4x16 = aom_sub_pixel_avg_variance4x16_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance4x16 = aom_sub_pixel_avg_variance4x16_ssse3;
     aom_sub_pixel_avg_variance4x4 = aom_sub_pixel_avg_variance4x4_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance4x4 = aom_sub_pixel_avg_variance4x4_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance4x4 = aom_sub_pixel_avg_variance4x4_ssse3;
     aom_sub_pixel_avg_variance4x8 = aom_sub_pixel_avg_variance4x8_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance4x8 = aom_sub_pixel_avg_variance4x8_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance4x8 = aom_sub_pixel_avg_variance4x8_ssse3;
     aom_sub_pixel_avg_variance64x128 = aom_sub_pixel_avg_variance64x128_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance64x128 = aom_sub_pixel_avg_variance64x128_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance64x128 = aom_sub_pixel_avg_variance64x128_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_avg_variance64x128 = aom_sub_pixel_avg_variance64x128_avx2;
     aom_sub_pixel_avg_variance64x16 = aom_sub_pixel_avg_variance64x16_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance64x16 = aom_sub_pixel_avg_variance64x16_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance64x16 = aom_sub_pixel_avg_variance64x16_ssse3;
     aom_sub_pixel_avg_variance64x32 = aom_sub_pixel_avg_variance64x32_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance64x32 = aom_sub_pixel_avg_variance64x32_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance64x32 = aom_sub_pixel_avg_variance64x32_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_avg_variance64x32 = aom_sub_pixel_avg_variance64x32_avx2;
     aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_avx2;
     aom_sub_pixel_avg_variance8x16 = aom_sub_pixel_avg_variance8x16_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance8x16 = aom_sub_pixel_avg_variance8x16_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x16 = aom_sub_pixel_avg_variance8x16_ssse3;
     aom_sub_pixel_avg_variance8x32 = aom_sub_pixel_avg_variance8x32_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance8x32 = aom_sub_pixel_avg_variance8x32_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x32 = aom_sub_pixel_avg_variance8x32_ssse3;
     aom_sub_pixel_avg_variance8x4 = aom_sub_pixel_avg_variance8x4_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance8x4 = aom_sub_pixel_avg_variance8x4_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x4 = aom_sub_pixel_avg_variance8x4_ssse3;
     aom_sub_pixel_avg_variance8x8 = aom_sub_pixel_avg_variance8x8_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_avg_variance8x8 = aom_sub_pixel_avg_variance8x8_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x8 = aom_sub_pixel_avg_variance8x8_ssse3;
     aom_sub_pixel_variance128x128 = aom_sub_pixel_variance128x128_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance128x128 = aom_sub_pixel_variance128x128_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance128x128 = aom_sub_pixel_variance128x128_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance128x128 = aom_sub_pixel_variance128x128_avx2;
     aom_sub_pixel_variance128x64 = aom_sub_pixel_variance128x64_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance128x64 = aom_sub_pixel_variance128x64_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance128x64 = aom_sub_pixel_variance128x64_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance128x64 = aom_sub_pixel_variance128x64_avx2;
     aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_avx2;
     aom_sub_pixel_variance16x32 = aom_sub_pixel_variance16x32_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance16x32 = aom_sub_pixel_variance16x32_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance16x32 = aom_sub_pixel_variance16x32_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance16x32 = aom_sub_pixel_variance16x32_avx2;
     aom_sub_pixel_variance16x4 = aom_sub_pixel_variance16x4_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance16x4 = aom_sub_pixel_variance16x4_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance16x4 = aom_sub_pixel_variance16x4_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance16x4 = aom_sub_pixel_variance16x4_avx2;
     aom_sub_pixel_variance16x64 = aom_sub_pixel_variance16x64_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance16x64 = aom_sub_pixel_variance16x64_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance16x64 = aom_sub_pixel_variance16x64_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance16x64 = aom_sub_pixel_variance16x64_avx2;
     aom_sub_pixel_variance16x8 = aom_sub_pixel_variance16x8_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance16x8 = aom_sub_pixel_variance16x8_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance16x8 = aom_sub_pixel_variance16x8_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance16x8 = aom_sub_pixel_variance16x8_avx2;
     aom_sub_pixel_variance32x16 = aom_sub_pixel_variance32x16_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance32x16 = aom_sub_pixel_variance32x16_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance32x16 = aom_sub_pixel_variance32x16_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance32x16 = aom_sub_pixel_variance32x16_avx2;
     aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_avx2;
     aom_sub_pixel_variance32x64 = aom_sub_pixel_variance32x64_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance32x64 = aom_sub_pixel_variance32x64_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance32x64 = aom_sub_pixel_variance32x64_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance32x64 = aom_sub_pixel_variance32x64_avx2;
     aom_sub_pixel_variance32x8 = aom_sub_pixel_variance32x8_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance32x8 = aom_sub_pixel_variance32x8_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance32x8 = aom_sub_pixel_variance32x8_ssse3;
     aom_sub_pixel_variance4x16 = aom_sub_pixel_variance4x16_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance4x16 = aom_sub_pixel_variance4x16_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance4x16 = aom_sub_pixel_variance4x16_ssse3;
     aom_sub_pixel_variance4x4 = aom_sub_pixel_variance4x4_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance4x4 = aom_sub_pixel_variance4x4_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance4x4 = aom_sub_pixel_variance4x4_ssse3;
     aom_sub_pixel_variance4x8 = aom_sub_pixel_variance4x8_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance4x8 = aom_sub_pixel_variance4x8_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance4x8 = aom_sub_pixel_variance4x8_ssse3;
     aom_sub_pixel_variance64x128 = aom_sub_pixel_variance64x128_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance64x128 = aom_sub_pixel_variance64x128_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance64x128 = aom_sub_pixel_variance64x128_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance64x128 = aom_sub_pixel_variance64x128_avx2;
     aom_sub_pixel_variance64x16 = aom_sub_pixel_variance64x16_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance64x16 = aom_sub_pixel_variance64x16_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance64x16 = aom_sub_pixel_variance64x16_ssse3;
     aom_sub_pixel_variance64x32 = aom_sub_pixel_variance64x32_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance64x32 = aom_sub_pixel_variance64x32_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance64x32 = aom_sub_pixel_variance64x32_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance64x32 = aom_sub_pixel_variance64x32_avx2;
     aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_avx2;
     aom_sub_pixel_variance8x16 = aom_sub_pixel_variance8x16_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance8x16 = aom_sub_pixel_variance8x16_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance8x16 = aom_sub_pixel_variance8x16_ssse3;
     aom_sub_pixel_variance8x32 = aom_sub_pixel_variance8x32_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance8x32 = aom_sub_pixel_variance8x32_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance8x32 = aom_sub_pixel_variance8x32_ssse3;
     aom_sub_pixel_variance8x4 = aom_sub_pixel_variance8x4_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance8x4 = aom_sub_pixel_variance8x4_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance8x4 = aom_sub_pixel_variance8x4_ssse3;
     aom_sub_pixel_variance8x8 = aom_sub_pixel_variance8x8_c;
-    if (flags & HAS_SSE2) aom_sub_pixel_variance8x8 = aom_sub_pixel_variance8x8_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance8x8 = aom_sub_pixel_variance8x8_ssse3;
     aom_subtract_block = aom_subtract_block_c;
     if (flags & HAS_SSE2) aom_subtract_block = aom_subtract_block_sse2;
@@ -9172,9 +9101,6 @@ static void setup_rtcd_internal(void)
     aom_vector_var = aom_vector_var_c;
     if (flags & HAS_SSE4_1) aom_vector_var = aom_vector_var_sse4_1;
     if (flags & HAS_AVX2) aom_vector_var = aom_vector_var_avx2;
-    av1_compute_cross_correlation = av1_compute_cross_correlation_c;
-    if (flags & HAS_SSE4_1) av1_compute_cross_correlation = av1_compute_cross_correlation_sse4_1;
-    if (flags & HAS_AVX2) av1_compute_cross_correlation = av1_compute_cross_correlation_avx2;
 }
 #endif
 
diff --git a/media/libaom/config/win/ia32/config/aom_scale_rtcd.h b/media/libaom/config/win/ia32/config/aom_scale_rtcd.h
index 3b70fb47c3..cdabb21106 100644
--- a/media/libaom/config/win/ia32/config/aom_scale_rtcd.h
+++ b/media/libaom/config/win/ia32/config/aom_scale_rtcd.h
@@ -8,13 +8,15 @@
 #define RTCD_EXTERN extern
 #endif
 
+#include <stdbool.h>
+
 struct yv12_buffer_config;
 
 #ifdef __cplusplus
 extern "C" {
 #endif
 
-void aom_extend_frame_borders_c(struct yv12_buffer_config *ybf, const int num_planes);
+void aom_extend_frame_borders_c(struct yv12_buffer_config *ybf, int num_planes);
 #define aom_extend_frame_borders aom_extend_frame_borders_c
 
 void aom_extend_frame_borders_plane_row_c(const struct yv12_buffer_config *ybf, int plane, int v_start, int v_end);
@@ -50,13 +52,13 @@ void aom_vertical_band_5_4_scale_c(unsigned char *source, int src_pitch, unsigne
 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);
+void aom_yv12_copy_u_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc, int use_crop);
 #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);
+void aom_yv12_copy_v_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc, int use_crop);
 #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);
+void aom_yv12_copy_y_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc, int use_crop);
 #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);
@@ -80,7 +82,7 @@ void aom_yv12_partial_copy_v_c(const struct yv12_buffer_config *src_bc, int hsta
 void aom_yv12_partial_copy_y_c(const struct yv12_buffer_config *src_ybc, int hstart1, int hend1, int vstart1, int vend1, struct yv12_buffer_config *dst_ybc, int hstart2, int vstart2);
 #define aom_yv12_partial_copy_y aom_yv12_partial_copy_y_c
 
-int aom_yv12_realloc_with_new_border_c(struct yv12_buffer_config *ybf, int new_border, int byte_alignment, int num_pyramid_levels, int num_planes);
+int aom_yv12_realloc_with_new_border_c(struct yv12_buffer_config *ybf, int new_border, int byte_alignment, bool alloc_pyramid, int num_planes);
 #define aom_yv12_realloc_with_new_border aom_yv12_realloc_with_new_border_c
 
 void aom_scale_rtcd(void);
diff --git a/media/libaom/config/win/ia32/config/av1_rtcd.h b/media/libaom/config/win/ia32/config/av1_rtcd.h
index 3f404f61c8..37716517bf 100644
--- a/media/libaom/config/win/ia32/config/av1_rtcd.h
+++ b/media/libaom/config/win/ia32/config/av1_rtcd.h
@@ -265,7 +265,6 @@ void av1_convolve_y_sr_intrabc_c(const uint8_t *src, int src_stride, uint8_t *ds
 #define av1_convolve_y_sr_intrabc av1_convolve_y_sr_intrabc_c
 
 void av1_dist_wtd_convolve_2d_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
-void av1_dist_wtd_convolve_2d_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
 void av1_dist_wtd_convolve_2d_ssse3(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
 void av1_dist_wtd_convolve_2d_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
 RTCD_EXTERN void (*av1_dist_wtd_convolve_2d)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
@@ -764,84 +763,72 @@ void av1_wiener_convolve_add_src_avx2(const uint8_t *src, ptrdiff_t src_stride,
 RTCD_EXTERN void (*av1_wiener_convolve_add_src)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params);
 
 void cdef_copy_rect8_16bit_to_16bit_c(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
-void cdef_copy_rect8_16bit_to_16bit_sse2(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
 void cdef_copy_rect8_16bit_to_16bit_ssse3(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
 void cdef_copy_rect8_16bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
 void cdef_copy_rect8_16bit_to_16bit_avx2(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
 RTCD_EXTERN void (*cdef_copy_rect8_16bit_to_16bit)(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
 
 void cdef_copy_rect8_8bit_to_16bit_c(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
-void cdef_copy_rect8_8bit_to_16bit_sse2(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
 void cdef_copy_rect8_8bit_to_16bit_ssse3(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
 void cdef_copy_rect8_8bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
 void cdef_copy_rect8_8bit_to_16bit_avx2(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
 RTCD_EXTERN void (*cdef_copy_rect8_8bit_to_16bit)(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
 
 void cdef_filter_16_0_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_16_0_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_0_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_0_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_0_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 RTCD_EXTERN void (*cdef_filter_16_0)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 
 void cdef_filter_16_1_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_16_1_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_1_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_1_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_1_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 RTCD_EXTERN void (*cdef_filter_16_1)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 
 void cdef_filter_16_2_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_16_2_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_2_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_2_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_2_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 RTCD_EXTERN void (*cdef_filter_16_2)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 
 void cdef_filter_16_3_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_16_3_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_3_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_3_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_3_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 RTCD_EXTERN void (*cdef_filter_16_3)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 
 void cdef_filter_8_0_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_8_0_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_0_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_0_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_0_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 RTCD_EXTERN void (*cdef_filter_8_0)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 
 void cdef_filter_8_1_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_8_1_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_1_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_1_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_1_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 RTCD_EXTERN void (*cdef_filter_8_1)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 
 void cdef_filter_8_2_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_8_2_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_2_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_2_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_2_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 RTCD_EXTERN void (*cdef_filter_8_2)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 
 void cdef_filter_8_3_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_8_3_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_3_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_3_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_3_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 RTCD_EXTERN void (*cdef_filter_8_3)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 
 int cdef_find_dir_c(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
-int cdef_find_dir_sse2(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
 int cdef_find_dir_ssse3(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
 int cdef_find_dir_sse4_1(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
 int cdef_find_dir_avx2(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
 RTCD_EXTERN int (*cdef_find_dir)(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
 
 void cdef_find_dir_dual_c(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
-void cdef_find_dir_dual_sse2(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
 void cdef_find_dir_dual_ssse3(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
 void cdef_find_dir_dual_sse4_1(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
 void cdef_find_dir_dual_avx2(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
@@ -969,7 +956,6 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_SSE2) av1_convolve_y_sr = av1_convolve_y_sr_sse2;
     if (flags & HAS_AVX2) av1_convolve_y_sr = av1_convolve_y_sr_avx2;
     av1_dist_wtd_convolve_2d = av1_dist_wtd_convolve_2d_c;
-    if (flags & HAS_SSE2) av1_dist_wtd_convolve_2d = av1_dist_wtd_convolve_2d_sse2;
     if (flags & HAS_SSSE3) av1_dist_wtd_convolve_2d = av1_dist_wtd_convolve_2d_ssse3;
     if (flags & HAS_AVX2) av1_dist_wtd_convolve_2d = av1_dist_wtd_convolve_2d_avx2;
     av1_dist_wtd_convolve_2d_copy = av1_dist_wtd_convolve_2d_copy_c;
@@ -1176,62 +1162,50 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_SSE2) av1_wiener_convolve_add_src = av1_wiener_convolve_add_src_sse2;
     if (flags & HAS_AVX2) av1_wiener_convolve_add_src = av1_wiener_convolve_add_src_avx2;
     cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_c;
-    if (flags & HAS_SSE2) cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_sse2;
     if (flags & HAS_SSSE3) cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_ssse3;
     if (flags & HAS_SSE4_1) cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_sse4_1;
     if (flags & HAS_AVX2) cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_avx2;
     cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_c;
-    if (flags & HAS_SSE2) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_sse2;
     if (flags & HAS_SSSE3) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_ssse3;
     if (flags & HAS_SSE4_1) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_sse4_1;
     if (flags & HAS_AVX2) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_avx2;
     cdef_filter_16_0 = cdef_filter_16_0_c;
-    if (flags & HAS_SSE2) cdef_filter_16_0 = cdef_filter_16_0_sse2;
     if (flags & HAS_SSSE3) cdef_filter_16_0 = cdef_filter_16_0_ssse3;
     if (flags & HAS_SSE4_1) cdef_filter_16_0 = cdef_filter_16_0_sse4_1;
     if (flags & HAS_AVX2) cdef_filter_16_0 = cdef_filter_16_0_avx2;
     cdef_filter_16_1 = cdef_filter_16_1_c;
-    if (flags & HAS_SSE2) cdef_filter_16_1 = cdef_filter_16_1_sse2;
     if (flags & HAS_SSSE3) cdef_filter_16_1 = cdef_filter_16_1_ssse3;
     if (flags & HAS_SSE4_1) cdef_filter_16_1 = cdef_filter_16_1_sse4_1;
     if (flags & HAS_AVX2) cdef_filter_16_1 = cdef_filter_16_1_avx2;
     cdef_filter_16_2 = cdef_filter_16_2_c;
-    if (flags & HAS_SSE2) cdef_filter_16_2 = cdef_filter_16_2_sse2;
     if (flags & HAS_SSSE3) cdef_filter_16_2 = cdef_filter_16_2_ssse3;
     if (flags & HAS_SSE4_1) cdef_filter_16_2 = cdef_filter_16_2_sse4_1;
     if (flags & HAS_AVX2) cdef_filter_16_2 = cdef_filter_16_2_avx2;
     cdef_filter_16_3 = cdef_filter_16_3_c;
-    if (flags & HAS_SSE2) cdef_filter_16_3 = cdef_filter_16_3_sse2;
     if (flags & HAS_SSSE3) cdef_filter_16_3 = cdef_filter_16_3_ssse3;
     if (flags & HAS_SSE4_1) cdef_filter_16_3 = cdef_filter_16_3_sse4_1;
     if (flags & HAS_AVX2) cdef_filter_16_3 = cdef_filter_16_3_avx2;
     cdef_filter_8_0 = cdef_filter_8_0_c;
-    if (flags & HAS_SSE2) cdef_filter_8_0 = cdef_filter_8_0_sse2;
     if (flags & HAS_SSSE3) cdef_filter_8_0 = cdef_filter_8_0_ssse3;
     if (flags & HAS_SSE4_1) cdef_filter_8_0 = cdef_filter_8_0_sse4_1;
     if (flags & HAS_AVX2) cdef_filter_8_0 = cdef_filter_8_0_avx2;
     cdef_filter_8_1 = cdef_filter_8_1_c;
-    if (flags & HAS_SSE2) cdef_filter_8_1 = cdef_filter_8_1_sse2;
     if (flags & HAS_SSSE3) cdef_filter_8_1 = cdef_filter_8_1_ssse3;
     if (flags & HAS_SSE4_1) cdef_filter_8_1 = cdef_filter_8_1_sse4_1;
     if (flags & HAS_AVX2) cdef_filter_8_1 = cdef_filter_8_1_avx2;
     cdef_filter_8_2 = cdef_filter_8_2_c;
-    if (flags & HAS_SSE2) cdef_filter_8_2 = cdef_filter_8_2_sse2;
     if (flags & HAS_SSSE3) cdef_filter_8_2 = cdef_filter_8_2_ssse3;
     if (flags & HAS_SSE4_1) cdef_filter_8_2 = cdef_filter_8_2_sse4_1;
     if (flags & HAS_AVX2) cdef_filter_8_2 = cdef_filter_8_2_avx2;
     cdef_filter_8_3 = cdef_filter_8_3_c;
-    if (flags & HAS_SSE2) cdef_filter_8_3 = cdef_filter_8_3_sse2;
     if (flags & HAS_SSSE3) cdef_filter_8_3 = cdef_filter_8_3_ssse3;
     if (flags & HAS_SSE4_1) cdef_filter_8_3 = cdef_filter_8_3_sse4_1;
     if (flags & HAS_AVX2) cdef_filter_8_3 = cdef_filter_8_3_avx2;
     cdef_find_dir = cdef_find_dir_c;
-    if (flags & HAS_SSE2) cdef_find_dir = cdef_find_dir_sse2;
     if (flags & HAS_SSSE3) cdef_find_dir = cdef_find_dir_ssse3;
     if (flags & HAS_SSE4_1) cdef_find_dir = cdef_find_dir_sse4_1;
     if (flags & HAS_AVX2) cdef_find_dir = cdef_find_dir_avx2;
     cdef_find_dir_dual = cdef_find_dir_dual_c;
-    if (flags & HAS_SSE2) cdef_find_dir_dual = cdef_find_dir_dual_sse2;
     if (flags & HAS_SSSE3) cdef_find_dir_dual = cdef_find_dir_dual_ssse3;
     if (flags & HAS_SSE4_1) cdef_find_dir_dual = cdef_find_dir_dual_sse4_1;
     if (flags & HAS_AVX2) cdef_find_dir_dual = cdef_find_dir_dual_avx2;
diff --git a/media/libaom/config/win/x64/config/aom_config.asm b/media/libaom/config/win/x64/config/aom_config.asm
index f793ff3c6d..3f470f3a5f 100644
--- a/media/libaom/config/win/x64/config/aom_config.asm
+++ b/media/libaom/config/win/x64/config/aom_config.asm
@@ -53,6 +53,7 @@ CONFIG_OS_SUPPORT equ 1
 CONFIG_OUTPUT_FRAME_SIZE equ 0
 CONFIG_PARTITION_SEARCH_ORDER equ 0
 CONFIG_PIC equ 0
+CONFIG_QUANT_MATRIX equ 1
 CONFIG_RATECTRL_LOG equ 0
 CONFIG_RD_COMMAND equ 0
 CONFIG_RD_DEBUG equ 0
@@ -87,6 +88,7 @@ HAVE_SSE4_1 equ 1
 HAVE_SSE4_2 equ 1
 HAVE_SSSE3 equ 1
 HAVE_SVE equ 0
+HAVE_SVE2 equ 0
 HAVE_VSX equ 0
 HAVE_WXWIDGETS equ 0
 STATIC_LINK_JXL equ 0
diff --git a/media/libaom/config/win/x64/config/aom_config.h b/media/libaom/config/win/x64/config/aom_config.h
index 670d2ffe56..6d96b65b07 100644
--- a/media/libaom/config/win/x64/config/aom_config.h
+++ b/media/libaom/config/win/x64/config/aom_config.h
@@ -55,6 +55,7 @@
 #define CONFIG_OUTPUT_FRAME_SIZE 0
 #define CONFIG_PARTITION_SEARCH_ORDER 0
 #define CONFIG_PIC 0
+#define CONFIG_QUANT_MATRIX 1
 #define CONFIG_RATECTRL_LOG 0
 #define CONFIG_RD_COMMAND 0
 #define CONFIG_RD_DEBUG 0
@@ -89,6 +90,7 @@
 #define HAVE_SSE4_2 1
 #define HAVE_SSSE3 1
 #define HAVE_SVE 0
+#define HAVE_SVE2 0
 #define HAVE_VSX 0
 #define HAVE_WXWIDGETS 0
 #define INLINE inline
diff --git a/media/libaom/config/win/x64/config/aom_dsp_rtcd.h b/media/libaom/config/win/x64/config/aom_dsp_rtcd.h
index 8e979cc189..9135c6f423 100644
--- a/media/libaom/config/win/x64/config/aom_dsp_rtcd.h
+++ b/media/libaom/config/win/x64/config/aom_dsp_rtcd.h
@@ -57,21 +57,30 @@ void aom_comp_mask_pred_ssse3(uint8_t *comp_pred, const uint8_t *pred, int width
 void aom_comp_mask_pred_avx2(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
 RTCD_EXTERN void (*aom_comp_mask_pred)(uint8_t *comp_pred, const uint8_t *pred, int width, int height, const uint8_t *ref, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
 
+double aom_compute_correlation_c(const unsigned char *frame1, int stride1, int x1, int y1, double mean1, double one_over_stddev1, const unsigned char *frame2, int stride2, int x2, int y2, double mean2, double one_over_stddev2);
+double aom_compute_correlation_sse4_1(const unsigned char *frame1, int stride1, int x1, int y1, double mean1, double one_over_stddev1, const unsigned char *frame2, int stride2, int x2, int y2, double mean2, double one_over_stddev2);
+double aom_compute_correlation_avx2(const unsigned char *frame1, int stride1, int x1, int y1, double mean1, double one_over_stddev1, const unsigned char *frame2, int stride2, int x2, int y2, double mean2, double one_over_stddev2);
+RTCD_EXTERN double (*aom_compute_correlation)(const unsigned char *frame1, int stride1, int x1, int y1, double mean1, double one_over_stddev1, const unsigned char *frame2, int stride2, int x2, int y2, double mean2, double one_over_stddev2);
+
 void aom_compute_flow_at_point_c(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
 void aom_compute_flow_at_point_sse4_1(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
+void aom_compute_flow_at_point_avx2(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
 RTCD_EXTERN void (*aom_compute_flow_at_point)(const uint8_t *src, const uint8_t *ref, int x, int y, int width, int height, int stride, double *u, double *v);
 
+bool aom_compute_mean_stddev_c(const unsigned char *frame, int stride, int x, int y, double *mean, double *one_over_stddev);
+bool aom_compute_mean_stddev_sse4_1(const unsigned char *frame, int stride, int x, int y, double *mean, double *one_over_stddev);
+bool aom_compute_mean_stddev_avx2(const unsigned char *frame, int stride, int x, int y, double *mean, double *one_over_stddev);
+RTCD_EXTERN bool (*aom_compute_mean_stddev)(const unsigned char *frame, int stride, int x, int y, double *mean, double *one_over_stddev);
+
 void aom_convolve8_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const InterpKernel *filter, int x0_q4, int x_step_q4, int y0_q4, int y_step_q4, int w, int h);
 #define aom_convolve8 aom_convolve8_c
 
 void aom_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
-void aom_convolve8_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void aom_convolve8_horiz_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void aom_convolve8_horiz_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 RTCD_EXTERN void (*aom_convolve8_horiz)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 
 void aom_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
-void aom_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void aom_convolve8_vert_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void aom_convolve8_vert_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 RTCD_EXTERN void (*aom_convolve8_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
@@ -903,7 +912,8 @@ RTCD_EXTERN unsigned int (*aom_highbd_10_masked_sub_pixel_variance8x8)(const uin
 
 unsigned int aom_highbd_10_mse16x16_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
 unsigned int aom_highbd_10_mse16x16_sse2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
-#define aom_highbd_10_mse16x16 aom_highbd_10_mse16x16_sse2
+unsigned int aom_highbd_10_mse16x16_avx2(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
+RTCD_EXTERN unsigned int (*aom_highbd_10_mse16x16)(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
 
 unsigned int aom_highbd_10_mse16x8_c(const uint8_t *src_ptr, int  source_stride, const uint8_t *ref_ptr, int  recon_stride, unsigned int *sse);
 #define aom_highbd_10_mse16x8 aom_highbd_10_mse16x8_c
@@ -5132,7 +5142,8 @@ unsigned int aom_sad16x4_avg_sse2(const uint8_t *src_ptr, int src_stride, const
 #define aom_sad16x4_avg aom_sad16x4_avg_sse2
 
 void aom_sad16x4x3d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
-#define aom_sad16x4x3d aom_sad16x4x3d_c
+void aom_sad16x4x3d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad16x4x3d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
 
 void aom_sad16x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
 void aom_sad16x4x4d_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
@@ -5468,7 +5479,8 @@ unsigned int aom_sad_skip_16x4_c(const uint8_t *src_ptr, int src_stride, const u
 #define aom_sad_skip_16x4 aom_sad_skip_16x4_c
 
 void aom_sad_skip_16x4x4d_c(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
-#define aom_sad_skip_16x4x4d aom_sad_skip_16x4x4d_c
+void aom_sad_skip_16x4x4d_avx2(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
+RTCD_EXTERN void (*aom_sad_skip_16x4x4d)(const uint8_t *src_ptr, int src_stride, const uint8_t * const ref_ptr[4], int ref_stride, uint32_t sad_array[4]);
 
 unsigned int aom_sad_skip_16x64_c(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
 unsigned int aom_sad_skip_16x64_sse2(const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr, int ref_stride);
@@ -5870,243 +5882,199 @@ void aom_ssim_parms_8x8_sse2(const uint8_t *s, int sp, const uint8_t *r, int rp,
 #define aom_ssim_parms_8x8 aom_ssim_parms_8x8_sse2
 
 uint32_t aom_sub_pixel_avg_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance128x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance128x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance128x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance128x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance16x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance16x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance16x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance16x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance16x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance32x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance32x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance32x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance32x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance32x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance32x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance32x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance4x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance4x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance4x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance4x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance4x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance4x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance4x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance64x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance64x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance64x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance64x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance64x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance64x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance64x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance8x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance8x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance8x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_avg_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
-uint32_t aom_sub_pixel_avg_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 uint32_t aom_sub_pixel_avg_variance8x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_avg_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse, const uint8_t *second_pred);
 
 uint32_t aom_sub_pixel_variance128x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance128x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance128x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance128x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance128x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance128x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance128x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance128x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance128x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance128x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance16x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance16x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance16x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance16x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance16x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance16x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x4_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance16x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance16x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance16x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance16x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance16x8_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance16x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance32x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance32x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance32x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance32x16_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance32x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance32x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance32x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance32x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance32x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance32x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance32x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance32x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance32x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance32x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance32x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance32x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance4x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance4x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance4x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance4x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance4x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance4x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance4x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance4x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance4x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance4x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance4x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance4x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance64x128_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance64x128_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance64x128_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance64x128_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x128)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance64x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance64x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance64x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance64x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance64x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance64x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance64x32_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance64x64_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance64x64_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance64x64_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance64x64_avx2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance64x64)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance8x16_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance8x16_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance8x16_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x16)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance8x32_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance8x32_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance8x32_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x32)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance8x4_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance8x4_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance8x4_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x4)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
 uint32_t aom_sub_pixel_variance8x8_c(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
-uint32_t aom_sub_pixel_variance8x8_sse2(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 uint32_t aom_sub_pixel_variance8x8_ssse3(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 RTCD_EXTERN uint32_t (*aom_sub_pixel_variance8x8)(const uint8_t *src_ptr, int source_stride, int xoffset, int  yoffset, const uint8_t *ref_ptr, int ref_stride, uint32_t *sse);
 
@@ -6329,11 +6297,6 @@ int aom_vector_var_sse4_1(const int16_t *ref, const int16_t *src, int bwl);
 int aom_vector_var_avx2(const int16_t *ref, const int16_t *src, int bwl);
 RTCD_EXTERN int (*aom_vector_var)(const int16_t *ref, const int16_t *src, int bwl);
 
-double av1_compute_cross_correlation_c(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
-double av1_compute_cross_correlation_sse4_1(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
-double av1_compute_cross_correlation_avx2(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
-RTCD_EXTERN double (*av1_compute_cross_correlation)(const unsigned char *frame1, int stride1, int x1, int y1, const unsigned char *frame2, int stride2, int x2, int y2);
-
 void aom_dsp_rtcd(void);
 
 #ifdef RTCD_C
@@ -6358,12 +6321,19 @@ static void setup_rtcd_internal(void)
     aom_comp_mask_pred = aom_comp_mask_pred_c;
     if (flags & HAS_SSSE3) aom_comp_mask_pred = aom_comp_mask_pred_ssse3;
     if (flags & HAS_AVX2) aom_comp_mask_pred = aom_comp_mask_pred_avx2;
+    aom_compute_correlation = aom_compute_correlation_c;
+    if (flags & HAS_SSE4_1) aom_compute_correlation = aom_compute_correlation_sse4_1;
+    if (flags & HAS_AVX2) aom_compute_correlation = aom_compute_correlation_avx2;
     aom_compute_flow_at_point = aom_compute_flow_at_point_c;
     if (flags & HAS_SSE4_1) aom_compute_flow_at_point = aom_compute_flow_at_point_sse4_1;
-    aom_convolve8_horiz = aom_convolve8_horiz_sse2;
+    if (flags & HAS_AVX2) aom_compute_flow_at_point = aom_compute_flow_at_point_avx2;
+    aom_compute_mean_stddev = aom_compute_mean_stddev_c;
+    if (flags & HAS_SSE4_1) aom_compute_mean_stddev = aom_compute_mean_stddev_sse4_1;
+    if (flags & HAS_AVX2) aom_compute_mean_stddev = aom_compute_mean_stddev_avx2;
+    aom_convolve8_horiz = aom_convolve8_horiz_c;
     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;
+    aom_convolve8_vert = aom_convolve8_vert_c;
     if (flags & HAS_SSSE3) aom_convolve8_vert = aom_convolve8_vert_ssse3;
     if (flags & HAS_AVX2) aom_convolve8_vert = aom_convolve8_vert_avx2;
     aom_convolve_copy = aom_convolve_copy_sse2;
@@ -6528,6 +6498,8 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance8x4 = aom_highbd_10_masked_sub_pixel_variance8x4_ssse3;
     aom_highbd_10_masked_sub_pixel_variance8x8 = aom_highbd_10_masked_sub_pixel_variance8x8_c;
     if (flags & HAS_SSSE3) aom_highbd_10_masked_sub_pixel_variance8x8 = aom_highbd_10_masked_sub_pixel_variance8x8_ssse3;
+    aom_highbd_10_mse16x16 = aom_highbd_10_mse16x16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_10_mse16x16 = aom_highbd_10_mse16x16_avx2;
     aom_highbd_10_obmc_variance128x128 = aom_highbd_10_obmc_variance128x128_c;
     if (flags & HAS_SSE4_1) aom_highbd_10_obmc_variance128x128 = aom_highbd_10_obmc_variance128x128_sse4_1;
     aom_highbd_10_obmc_variance128x64 = aom_highbd_10_obmc_variance128x64_c;
@@ -7626,6 +7598,8 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_AVX2) aom_sad16x32x3d = aom_sad16x32x3d_avx2;
     aom_sad16x32x4d = aom_sad16x32x4d_sse2;
     if (flags & HAS_AVX2) aom_sad16x32x4d = aom_sad16x32x4d_avx2;
+    aom_sad16x4x3d = aom_sad16x4x3d_c;
+    if (flags & HAS_AVX2) aom_sad16x4x3d = aom_sad16x4x3d_avx2;
     aom_sad16x4x4d = aom_sad16x4x4d_sse2;
     if (flags & HAS_AVX2) aom_sad16x4x4d = aom_sad16x4x4d_avx2;
     aom_sad16x64x3d = aom_sad16x64x3d_c;
@@ -7704,6 +7678,8 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_AVX2) aom_sad_skip_16x16x4d = aom_sad_skip_16x16x4d_avx2;
     aom_sad_skip_16x32x4d = aom_sad_skip_16x32x4d_sse2;
     if (flags & HAS_AVX2) aom_sad_skip_16x32x4d = aom_sad_skip_16x32x4d_avx2;
+    aom_sad_skip_16x4x4d = aom_sad_skip_16x4x4d_c;
+    if (flags & HAS_AVX2) aom_sad_skip_16x4x4d = aom_sad_skip_16x4x4d_avx2;
     aom_sad_skip_16x64x4d = aom_sad_skip_16x64x4d_sse2;
     if (flags & HAS_AVX2) aom_sad_skip_16x64x4d = aom_sad_skip_16x64x4d_avx2;
     aom_sad_skip_16x8x4d = aom_sad_skip_16x8x4d_sse2;
@@ -7859,114 +7835,114 @@ static void setup_rtcd_internal(void)
     aom_sse = aom_sse_c;
     if (flags & HAS_SSE4_1) aom_sse = aom_sse_sse4_1;
     if (flags & HAS_AVX2) aom_sse = aom_sse_avx2;
-    aom_sub_pixel_avg_variance128x128 = aom_sub_pixel_avg_variance128x128_sse2;
+    aom_sub_pixel_avg_variance128x128 = aom_sub_pixel_avg_variance128x128_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance128x128 = aom_sub_pixel_avg_variance128x128_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_avg_variance128x128 = aom_sub_pixel_avg_variance128x128_avx2;
-    aom_sub_pixel_avg_variance128x64 = aom_sub_pixel_avg_variance128x64_sse2;
+    aom_sub_pixel_avg_variance128x64 = aom_sub_pixel_avg_variance128x64_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance128x64 = aom_sub_pixel_avg_variance128x64_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_avg_variance128x64 = aom_sub_pixel_avg_variance128x64_avx2;
-    aom_sub_pixel_avg_variance16x16 = aom_sub_pixel_avg_variance16x16_sse2;
+    aom_sub_pixel_avg_variance16x16 = aom_sub_pixel_avg_variance16x16_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x16 = aom_sub_pixel_avg_variance16x16_ssse3;
-    aom_sub_pixel_avg_variance16x32 = aom_sub_pixel_avg_variance16x32_sse2;
+    aom_sub_pixel_avg_variance16x32 = aom_sub_pixel_avg_variance16x32_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x32 = aom_sub_pixel_avg_variance16x32_ssse3;
-    aom_sub_pixel_avg_variance16x4 = aom_sub_pixel_avg_variance16x4_sse2;
+    aom_sub_pixel_avg_variance16x4 = aom_sub_pixel_avg_variance16x4_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x4 = aom_sub_pixel_avg_variance16x4_ssse3;
-    aom_sub_pixel_avg_variance16x64 = aom_sub_pixel_avg_variance16x64_sse2;
+    aom_sub_pixel_avg_variance16x64 = aom_sub_pixel_avg_variance16x64_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x64 = aom_sub_pixel_avg_variance16x64_ssse3;
-    aom_sub_pixel_avg_variance16x8 = aom_sub_pixel_avg_variance16x8_sse2;
+    aom_sub_pixel_avg_variance16x8 = aom_sub_pixel_avg_variance16x8_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x8 = aom_sub_pixel_avg_variance16x8_ssse3;
-    aom_sub_pixel_avg_variance32x16 = aom_sub_pixel_avg_variance32x16_sse2;
+    aom_sub_pixel_avg_variance32x16 = aom_sub_pixel_avg_variance32x16_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x16 = aom_sub_pixel_avg_variance32x16_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_avg_variance32x16 = aom_sub_pixel_avg_variance32x16_avx2;
-    aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_sse2;
+    aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_avg_variance32x32 = aom_sub_pixel_avg_variance32x32_avx2;
-    aom_sub_pixel_avg_variance32x64 = aom_sub_pixel_avg_variance32x64_sse2;
+    aom_sub_pixel_avg_variance32x64 = aom_sub_pixel_avg_variance32x64_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x64 = aom_sub_pixel_avg_variance32x64_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_avg_variance32x64 = aom_sub_pixel_avg_variance32x64_avx2;
-    aom_sub_pixel_avg_variance32x8 = aom_sub_pixel_avg_variance32x8_sse2;
+    aom_sub_pixel_avg_variance32x8 = aom_sub_pixel_avg_variance32x8_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance32x8 = aom_sub_pixel_avg_variance32x8_ssse3;
-    aom_sub_pixel_avg_variance4x16 = aom_sub_pixel_avg_variance4x16_sse2;
+    aom_sub_pixel_avg_variance4x16 = aom_sub_pixel_avg_variance4x16_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance4x16 = aom_sub_pixel_avg_variance4x16_ssse3;
-    aom_sub_pixel_avg_variance4x4 = aom_sub_pixel_avg_variance4x4_sse2;
+    aom_sub_pixel_avg_variance4x4 = aom_sub_pixel_avg_variance4x4_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance4x4 = aom_sub_pixel_avg_variance4x4_ssse3;
-    aom_sub_pixel_avg_variance4x8 = aom_sub_pixel_avg_variance4x8_sse2;
+    aom_sub_pixel_avg_variance4x8 = aom_sub_pixel_avg_variance4x8_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance4x8 = aom_sub_pixel_avg_variance4x8_ssse3;
-    aom_sub_pixel_avg_variance64x128 = aom_sub_pixel_avg_variance64x128_sse2;
+    aom_sub_pixel_avg_variance64x128 = aom_sub_pixel_avg_variance64x128_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance64x128 = aom_sub_pixel_avg_variance64x128_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_avg_variance64x128 = aom_sub_pixel_avg_variance64x128_avx2;
-    aom_sub_pixel_avg_variance64x16 = aom_sub_pixel_avg_variance64x16_sse2;
+    aom_sub_pixel_avg_variance64x16 = aom_sub_pixel_avg_variance64x16_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance64x16 = aom_sub_pixel_avg_variance64x16_ssse3;
-    aom_sub_pixel_avg_variance64x32 = aom_sub_pixel_avg_variance64x32_sse2;
+    aom_sub_pixel_avg_variance64x32 = aom_sub_pixel_avg_variance64x32_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance64x32 = aom_sub_pixel_avg_variance64x32_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_avg_variance64x32 = aom_sub_pixel_avg_variance64x32_avx2;
-    aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_sse2;
+    aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_avg_variance64x64 = aom_sub_pixel_avg_variance64x64_avx2;
-    aom_sub_pixel_avg_variance8x16 = aom_sub_pixel_avg_variance8x16_sse2;
+    aom_sub_pixel_avg_variance8x16 = aom_sub_pixel_avg_variance8x16_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x16 = aom_sub_pixel_avg_variance8x16_ssse3;
-    aom_sub_pixel_avg_variance8x32 = aom_sub_pixel_avg_variance8x32_sse2;
+    aom_sub_pixel_avg_variance8x32 = aom_sub_pixel_avg_variance8x32_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x32 = aom_sub_pixel_avg_variance8x32_ssse3;
-    aom_sub_pixel_avg_variance8x4 = aom_sub_pixel_avg_variance8x4_sse2;
+    aom_sub_pixel_avg_variance8x4 = aom_sub_pixel_avg_variance8x4_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x4 = aom_sub_pixel_avg_variance8x4_ssse3;
-    aom_sub_pixel_avg_variance8x8 = aom_sub_pixel_avg_variance8x8_sse2;
+    aom_sub_pixel_avg_variance8x8 = aom_sub_pixel_avg_variance8x8_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance8x8 = aom_sub_pixel_avg_variance8x8_ssse3;
-    aom_sub_pixel_variance128x128 = aom_sub_pixel_variance128x128_sse2;
+    aom_sub_pixel_variance128x128 = aom_sub_pixel_variance128x128_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance128x128 = aom_sub_pixel_variance128x128_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance128x128 = aom_sub_pixel_variance128x128_avx2;
-    aom_sub_pixel_variance128x64 = aom_sub_pixel_variance128x64_sse2;
+    aom_sub_pixel_variance128x64 = aom_sub_pixel_variance128x64_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance128x64 = aom_sub_pixel_variance128x64_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance128x64 = aom_sub_pixel_variance128x64_avx2;
-    aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_sse2;
+    aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance16x16 = aom_sub_pixel_variance16x16_avx2;
-    aom_sub_pixel_variance16x32 = aom_sub_pixel_variance16x32_sse2;
+    aom_sub_pixel_variance16x32 = aom_sub_pixel_variance16x32_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance16x32 = aom_sub_pixel_variance16x32_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance16x32 = aom_sub_pixel_variance16x32_avx2;
-    aom_sub_pixel_variance16x4 = aom_sub_pixel_variance16x4_sse2;
+    aom_sub_pixel_variance16x4 = aom_sub_pixel_variance16x4_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance16x4 = aom_sub_pixel_variance16x4_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance16x4 = aom_sub_pixel_variance16x4_avx2;
-    aom_sub_pixel_variance16x64 = aom_sub_pixel_variance16x64_sse2;
+    aom_sub_pixel_variance16x64 = aom_sub_pixel_variance16x64_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance16x64 = aom_sub_pixel_variance16x64_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance16x64 = aom_sub_pixel_variance16x64_avx2;
-    aom_sub_pixel_variance16x8 = aom_sub_pixel_variance16x8_sse2;
+    aom_sub_pixel_variance16x8 = aom_sub_pixel_variance16x8_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance16x8 = aom_sub_pixel_variance16x8_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance16x8 = aom_sub_pixel_variance16x8_avx2;
-    aom_sub_pixel_variance32x16 = aom_sub_pixel_variance32x16_sse2;
+    aom_sub_pixel_variance32x16 = aom_sub_pixel_variance32x16_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance32x16 = aom_sub_pixel_variance32x16_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance32x16 = aom_sub_pixel_variance32x16_avx2;
-    aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_sse2;
+    aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance32x32 = aom_sub_pixel_variance32x32_avx2;
-    aom_sub_pixel_variance32x64 = aom_sub_pixel_variance32x64_sse2;
+    aom_sub_pixel_variance32x64 = aom_sub_pixel_variance32x64_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance32x64 = aom_sub_pixel_variance32x64_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance32x64 = aom_sub_pixel_variance32x64_avx2;
-    aom_sub_pixel_variance32x8 = aom_sub_pixel_variance32x8_sse2;
+    aom_sub_pixel_variance32x8 = aom_sub_pixel_variance32x8_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance32x8 = aom_sub_pixel_variance32x8_ssse3;
-    aom_sub_pixel_variance4x16 = aom_sub_pixel_variance4x16_sse2;
+    aom_sub_pixel_variance4x16 = aom_sub_pixel_variance4x16_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance4x16 = aom_sub_pixel_variance4x16_ssse3;
-    aom_sub_pixel_variance4x4 = aom_sub_pixel_variance4x4_sse2;
+    aom_sub_pixel_variance4x4 = aom_sub_pixel_variance4x4_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance4x4 = aom_sub_pixel_variance4x4_ssse3;
-    aom_sub_pixel_variance4x8 = aom_sub_pixel_variance4x8_sse2;
+    aom_sub_pixel_variance4x8 = aom_sub_pixel_variance4x8_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance4x8 = aom_sub_pixel_variance4x8_ssse3;
-    aom_sub_pixel_variance64x128 = aom_sub_pixel_variance64x128_sse2;
+    aom_sub_pixel_variance64x128 = aom_sub_pixel_variance64x128_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance64x128 = aom_sub_pixel_variance64x128_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance64x128 = aom_sub_pixel_variance64x128_avx2;
-    aom_sub_pixel_variance64x16 = aom_sub_pixel_variance64x16_sse2;
+    aom_sub_pixel_variance64x16 = aom_sub_pixel_variance64x16_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance64x16 = aom_sub_pixel_variance64x16_ssse3;
-    aom_sub_pixel_variance64x32 = aom_sub_pixel_variance64x32_sse2;
+    aom_sub_pixel_variance64x32 = aom_sub_pixel_variance64x32_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance64x32 = aom_sub_pixel_variance64x32_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance64x32 = aom_sub_pixel_variance64x32_avx2;
-    aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_sse2;
+    aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_ssse3;
     if (flags & HAS_AVX2) aom_sub_pixel_variance64x64 = aom_sub_pixel_variance64x64_avx2;
-    aom_sub_pixel_variance8x16 = aom_sub_pixel_variance8x16_sse2;
+    aom_sub_pixel_variance8x16 = aom_sub_pixel_variance8x16_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance8x16 = aom_sub_pixel_variance8x16_ssse3;
-    aom_sub_pixel_variance8x32 = aom_sub_pixel_variance8x32_sse2;
+    aom_sub_pixel_variance8x32 = aom_sub_pixel_variance8x32_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance8x32 = aom_sub_pixel_variance8x32_ssse3;
-    aom_sub_pixel_variance8x4 = aom_sub_pixel_variance8x4_sse2;
+    aom_sub_pixel_variance8x4 = aom_sub_pixel_variance8x4_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance8x4 = aom_sub_pixel_variance8x4_ssse3;
-    aom_sub_pixel_variance8x8 = aom_sub_pixel_variance8x8_sse2;
+    aom_sub_pixel_variance8x8 = aom_sub_pixel_variance8x8_c;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance8x8 = aom_sub_pixel_variance8x8_ssse3;
     aom_subtract_block = aom_subtract_block_sse2;
     if (flags & HAS_AVX2) aom_subtract_block = aom_subtract_block_avx2;
@@ -8023,9 +7999,6 @@ static void setup_rtcd_internal(void)
     aom_vector_var = aom_vector_var_c;
     if (flags & HAS_SSE4_1) aom_vector_var = aom_vector_var_sse4_1;
     if (flags & HAS_AVX2) aom_vector_var = aom_vector_var_avx2;
-    av1_compute_cross_correlation = av1_compute_cross_correlation_c;
-    if (flags & HAS_SSE4_1) av1_compute_cross_correlation = av1_compute_cross_correlation_sse4_1;
-    if (flags & HAS_AVX2) av1_compute_cross_correlation = av1_compute_cross_correlation_avx2;
 }
 #endif
 
diff --git a/media/libaom/config/win/x64/config/aom_scale_rtcd.h b/media/libaom/config/win/x64/config/aom_scale_rtcd.h
index 3b70fb47c3..cdabb21106 100644
--- a/media/libaom/config/win/x64/config/aom_scale_rtcd.h
+++ b/media/libaom/config/win/x64/config/aom_scale_rtcd.h
@@ -8,13 +8,15 @@
 #define RTCD_EXTERN extern
 #endif
 
+#include <stdbool.h>
+
 struct yv12_buffer_config;
 
 #ifdef __cplusplus
 extern "C" {
 #endif
 
-void aom_extend_frame_borders_c(struct yv12_buffer_config *ybf, const int num_planes);
+void aom_extend_frame_borders_c(struct yv12_buffer_config *ybf, int num_planes);
 #define aom_extend_frame_borders aom_extend_frame_borders_c
 
 void aom_extend_frame_borders_plane_row_c(const struct yv12_buffer_config *ybf, int plane, int v_start, int v_end);
@@ -50,13 +52,13 @@ void aom_vertical_band_5_4_scale_c(unsigned char *source, int src_pitch, unsigne
 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);
+void aom_yv12_copy_u_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc, int use_crop);
 #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);
+void aom_yv12_copy_v_c(const struct yv12_buffer_config *src_bc, struct yv12_buffer_config *dst_bc, int use_crop);
 #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);
+void aom_yv12_copy_y_c(const struct yv12_buffer_config *src_ybc, struct yv12_buffer_config *dst_ybc, int use_crop);
 #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);
@@ -80,7 +82,7 @@ void aom_yv12_partial_copy_v_c(const struct yv12_buffer_config *src_bc, int hsta
 void aom_yv12_partial_copy_y_c(const struct yv12_buffer_config *src_ybc, int hstart1, int hend1, int vstart1, int vend1, struct yv12_buffer_config *dst_ybc, int hstart2, int vstart2);
 #define aom_yv12_partial_copy_y aom_yv12_partial_copy_y_c
 
-int aom_yv12_realloc_with_new_border_c(struct yv12_buffer_config *ybf, int new_border, int byte_alignment, int num_pyramid_levels, int num_planes);
+int aom_yv12_realloc_with_new_border_c(struct yv12_buffer_config *ybf, int new_border, int byte_alignment, bool alloc_pyramid, int num_planes);
 #define aom_yv12_realloc_with_new_border aom_yv12_realloc_with_new_border_c
 
 void aom_scale_rtcd(void);
diff --git a/media/libaom/config/win/x64/config/av1_rtcd.h b/media/libaom/config/win/x64/config/av1_rtcd.h
index b1cdc99700..ad72985afe 100644
--- a/media/libaom/config/win/x64/config/av1_rtcd.h
+++ b/media/libaom/config/win/x64/config/av1_rtcd.h
@@ -253,7 +253,6 @@ void av1_convolve_y_sr_intrabc_c(const uint8_t *src, int src_stride, uint8_t *ds
 #define av1_convolve_y_sr_intrabc av1_convolve_y_sr_intrabc_c
 
 void av1_dist_wtd_convolve_2d_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
-void av1_dist_wtd_convolve_2d_sse2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
 void av1_dist_wtd_convolve_2d_ssse3(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
 void av1_dist_wtd_convolve_2d_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
 RTCD_EXTERN void (*av1_dist_wtd_convolve_2d)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_qn, const int subpel_y_qn, ConvolveParams *conv_params);
@@ -659,7 +658,6 @@ void av1_inv_txfm_add_avx2(const tran_low_t *dqcoeff, uint8_t *dst, int stride,
 RTCD_EXTERN void (*av1_inv_txfm_add)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
 
 void av1_lowbd_fwd_txfm_c(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
-void av1_lowbd_fwd_txfm_sse2(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
 void av1_lowbd_fwd_txfm_sse4_1(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
 void av1_lowbd_fwd_txfm_avx2(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
 RTCD_EXTERN void (*av1_lowbd_fwd_txfm)(const int16_t *src_diff, tran_low_t *coeff, int diff_stride, TxfmParam *txfm_param);
@@ -755,85 +753,61 @@ void av1_wiener_convolve_add_src_avx2(const uint8_t *src, ptrdiff_t src_stride,
 RTCD_EXTERN void (*av1_wiener_convolve_add_src)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const WienerConvolveParams *conv_params);
 
 void cdef_copy_rect8_16bit_to_16bit_c(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
-void cdef_copy_rect8_16bit_to_16bit_sse2(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
-void cdef_copy_rect8_16bit_to_16bit_ssse3(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
 void cdef_copy_rect8_16bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
 void cdef_copy_rect8_16bit_to_16bit_avx2(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
 RTCD_EXTERN void (*cdef_copy_rect8_16bit_to_16bit)(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int width, int height);
 
 void cdef_copy_rect8_8bit_to_16bit_c(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
-void cdef_copy_rect8_8bit_to_16bit_sse2(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
-void cdef_copy_rect8_8bit_to_16bit_ssse3(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
 void cdef_copy_rect8_8bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
 void cdef_copy_rect8_8bit_to_16bit_avx2(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
 RTCD_EXTERN void (*cdef_copy_rect8_8bit_to_16bit)(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int width, int height);
 
 void cdef_filter_16_0_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_16_0_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_16_0_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_0_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_0_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 RTCD_EXTERN void (*cdef_filter_16_0)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 
 void cdef_filter_16_1_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_16_1_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_16_1_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_1_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_1_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 RTCD_EXTERN void (*cdef_filter_16_1)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 
 void cdef_filter_16_2_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_16_2_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_16_2_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_2_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_2_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 RTCD_EXTERN void (*cdef_filter_16_2)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 
 void cdef_filter_16_3_c(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_16_3_sse2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_16_3_ssse3(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_3_sse4_1(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_16_3_avx2(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 RTCD_EXTERN void (*cdef_filter_16_3)(void *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 
 void cdef_filter_8_0_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_8_0_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_8_0_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_0_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_0_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 RTCD_EXTERN void (*cdef_filter_8_0)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 
 void cdef_filter_8_1_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_8_1_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_8_1_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_1_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_1_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 RTCD_EXTERN void (*cdef_filter_8_1)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 
 void cdef_filter_8_2_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_8_2_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_8_2_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_2_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_2_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 RTCD_EXTERN void (*cdef_filter_8_2)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 
 void cdef_filter_8_3_c(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_8_3_sse2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
-void cdef_filter_8_3_ssse3(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_3_sse4_1(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 void cdef_filter_8_3_avx2(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 RTCD_EXTERN void (*cdef_filter_8_3)(void *dst8, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int coeff_shift, int block_width, int block_height);
 
 int cdef_find_dir_c(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
-int cdef_find_dir_sse2(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
-int cdef_find_dir_ssse3(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
 int cdef_find_dir_sse4_1(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
 int cdef_find_dir_avx2(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
 RTCD_EXTERN int (*cdef_find_dir)(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
 
 void cdef_find_dir_dual_c(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
-void cdef_find_dir_dual_sse2(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
-void cdef_find_dir_dual_ssse3(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
 void cdef_find_dir_dual_sse4_1(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
 void cdef_find_dir_dual_avx2(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
 RTCD_EXTERN void (*cdef_find_dir_dual)(const uint16_t *img1, const uint16_t *img2, int stride, int32_t *var1, int32_t *var2, int coeff_shift, int *out1, int *out2);
@@ -941,7 +915,7 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_AVX2) av1_convolve_x_sr = av1_convolve_x_sr_avx2;
     av1_convolve_y_sr = av1_convolve_y_sr_sse2;
     if (flags & HAS_AVX2) av1_convolve_y_sr = av1_convolve_y_sr_avx2;
-    av1_dist_wtd_convolve_2d = av1_dist_wtd_convolve_2d_sse2;
+    av1_dist_wtd_convolve_2d = av1_dist_wtd_convolve_2d_c;
     if (flags & HAS_SSSE3) av1_dist_wtd_convolve_2d = av1_dist_wtd_convolve_2d_ssse3;
     if (flags & HAS_AVX2) av1_dist_wtd_convolve_2d = av1_dist_wtd_convolve_2d_avx2;
     av1_dist_wtd_convolve_2d_copy = av1_dist_wtd_convolve_2d_copy_sse2;
@@ -1091,7 +1065,7 @@ static void setup_rtcd_internal(void)
     av1_inv_txfm_add = av1_inv_txfm_add_c;
     if (flags & HAS_SSSE3) av1_inv_txfm_add = av1_inv_txfm_add_ssse3;
     if (flags & HAS_AVX2) av1_inv_txfm_add = av1_inv_txfm_add_avx2;
-    av1_lowbd_fwd_txfm = av1_lowbd_fwd_txfm_sse2;
+    av1_lowbd_fwd_txfm = av1_lowbd_fwd_txfm_c;
     if (flags & HAS_SSE4_1) av1_lowbd_fwd_txfm = av1_lowbd_fwd_txfm_sse4_1;
     if (flags & HAS_AVX2) av1_lowbd_fwd_txfm = av1_lowbd_fwd_txfm_avx2;
     av1_lowbd_pixel_proj_error = av1_lowbd_pixel_proj_error_c;
@@ -1133,52 +1107,40 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_AVX2) av1_wedge_sse_from_residuals = av1_wedge_sse_from_residuals_avx2;
     av1_wiener_convolve_add_src = av1_wiener_convolve_add_src_sse2;
     if (flags & HAS_AVX2) av1_wiener_convolve_add_src = av1_wiener_convolve_add_src_avx2;
-    cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_sse2;
-    if (flags & HAS_SSSE3) cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_ssse3;
+    cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_c;
     if (flags & HAS_SSE4_1) cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_sse4_1;
     if (flags & HAS_AVX2) cdef_copy_rect8_16bit_to_16bit = cdef_copy_rect8_16bit_to_16bit_avx2;
-    cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_sse2;
-    if (flags & HAS_SSSE3) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_ssse3;
+    cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_c;
     if (flags & HAS_SSE4_1) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_sse4_1;
     if (flags & HAS_AVX2) cdef_copy_rect8_8bit_to_16bit = cdef_copy_rect8_8bit_to_16bit_avx2;
-    cdef_filter_16_0 = cdef_filter_16_0_sse2;
-    if (flags & HAS_SSSE3) cdef_filter_16_0 = cdef_filter_16_0_ssse3;
+    cdef_filter_16_0 = cdef_filter_16_0_c;
     if (flags & HAS_SSE4_1) cdef_filter_16_0 = cdef_filter_16_0_sse4_1;
     if (flags & HAS_AVX2) cdef_filter_16_0 = cdef_filter_16_0_avx2;
-    cdef_filter_16_1 = cdef_filter_16_1_sse2;
-    if (flags & HAS_SSSE3) cdef_filter_16_1 = cdef_filter_16_1_ssse3;
+    cdef_filter_16_1 = cdef_filter_16_1_c;
     if (flags & HAS_SSE4_1) cdef_filter_16_1 = cdef_filter_16_1_sse4_1;
     if (flags & HAS_AVX2) cdef_filter_16_1 = cdef_filter_16_1_avx2;
-    cdef_filter_16_2 = cdef_filter_16_2_sse2;
-    if (flags & HAS_SSSE3) cdef_filter_16_2 = cdef_filter_16_2_ssse3;
+    cdef_filter_16_2 = cdef_filter_16_2_c;
     if (flags & HAS_SSE4_1) cdef_filter_16_2 = cdef_filter_16_2_sse4_1;
     if (flags & HAS_AVX2) cdef_filter_16_2 = cdef_filter_16_2_avx2;
-    cdef_filter_16_3 = cdef_filter_16_3_sse2;
-    if (flags & HAS_SSSE3) cdef_filter_16_3 = cdef_filter_16_3_ssse3;
+    cdef_filter_16_3 = cdef_filter_16_3_c;
     if (flags & HAS_SSE4_1) cdef_filter_16_3 = cdef_filter_16_3_sse4_1;
     if (flags & HAS_AVX2) cdef_filter_16_3 = cdef_filter_16_3_avx2;
-    cdef_filter_8_0 = cdef_filter_8_0_sse2;
-    if (flags & HAS_SSSE3) cdef_filter_8_0 = cdef_filter_8_0_ssse3;
+    cdef_filter_8_0 = cdef_filter_8_0_c;
     if (flags & HAS_SSE4_1) cdef_filter_8_0 = cdef_filter_8_0_sse4_1;
     if (flags & HAS_AVX2) cdef_filter_8_0 = cdef_filter_8_0_avx2;
-    cdef_filter_8_1 = cdef_filter_8_1_sse2;
-    if (flags & HAS_SSSE3) cdef_filter_8_1 = cdef_filter_8_1_ssse3;
+    cdef_filter_8_1 = cdef_filter_8_1_c;
     if (flags & HAS_SSE4_1) cdef_filter_8_1 = cdef_filter_8_1_sse4_1;
     if (flags & HAS_AVX2) cdef_filter_8_1 = cdef_filter_8_1_avx2;
-    cdef_filter_8_2 = cdef_filter_8_2_sse2;
-    if (flags & HAS_SSSE3) cdef_filter_8_2 = cdef_filter_8_2_ssse3;
+    cdef_filter_8_2 = cdef_filter_8_2_c;
     if (flags & HAS_SSE4_1) cdef_filter_8_2 = cdef_filter_8_2_sse4_1;
     if (flags & HAS_AVX2) cdef_filter_8_2 = cdef_filter_8_2_avx2;
-    cdef_filter_8_3 = cdef_filter_8_3_sse2;
-    if (flags & HAS_SSSE3) cdef_filter_8_3 = cdef_filter_8_3_ssse3;
+    cdef_filter_8_3 = cdef_filter_8_3_c;
     if (flags & HAS_SSE4_1) cdef_filter_8_3 = cdef_filter_8_3_sse4_1;
     if (flags & HAS_AVX2) cdef_filter_8_3 = cdef_filter_8_3_avx2;
-    cdef_find_dir = cdef_find_dir_sse2;
-    if (flags & HAS_SSSE3) cdef_find_dir = cdef_find_dir_ssse3;
+    cdef_find_dir = cdef_find_dir_c;
     if (flags & HAS_SSE4_1) cdef_find_dir = cdef_find_dir_sse4_1;
     if (flags & HAS_AVX2) cdef_find_dir = cdef_find_dir_avx2;
-    cdef_find_dir_dual = cdef_find_dir_dual_sse2;
-    if (flags & HAS_SSSE3) cdef_find_dir_dual = cdef_find_dir_dual_ssse3;
+    cdef_find_dir_dual = cdef_find_dir_dual_c;
     if (flags & HAS_SSE4_1) cdef_find_dir_dual = cdef_find_dir_dual_sse4_1;
     if (flags & HAS_AVX2) cdef_find_dir_dual = cdef_find_dir_dual_avx2;
     cfl_get_luma_subsampling_420_hbd = cfl_get_luma_subsampling_420_hbd_c;
diff --git a/media/libaom/moz.yaml b/media/libaom/moz.yaml
index b06ee5115a..a37ab1e904 100644
--- a/media/libaom/moz.yaml
+++ b/media/libaom/moz.yaml
@@ -20,11 +20,11 @@ origin:
 
   # Human-readable identifier for this version/release
   # Generally "version NNN", "tag SSS", "bookmark SSS"
-  release: 11631186b36e96afce18808ebebb17cc23a010ef (Fri Jan 19 23:29:34 2024 +0000).
+  release: 879d14159441796c92f3bbba7f8965e1bcf320ca (Tue Apr 02 21:57:54 2024 +0000).
 
   # Revision to pull in
   # Must be a long or short commit SHA (long preferred)
-  revision: 11631186b36e96afce18808ebebb17cc23a010ef
+  revision: 879d14159441796c92f3bbba7f8965e1bcf320ca
 
   # The package's license, where possible using the mnemonic from
   # https://spdx.org/licenses/
@@ -54,3 +54,7 @@ vendoring:
     - action: run-script
       script: '{yaml_dir}/generate_sources_mozbuild.sh'
       cwd: '{yaml_dir}'
+
+  patches:
+    - 0001-errno.patch
+    - 0002-mmloadusi64.patch
diff --git a/media/libaom/sources.mozbuild b/media/libaom/sources.mozbuild
index b29ddd5c97..187bf97f8a 100644
--- a/media/libaom/sources.mozbuild
+++ b/media/libaom/sources.mozbuild
@@ -506,6 +506,7 @@ files = {
     '../../third_party/aom/aom_dsp/flow_estimation/ransac.c',
     '../../third_party/aom/aom_dsp/flow_estimation/x86/corner_match_avx2.c',
     '../../third_party/aom/aom_dsp/flow_estimation/x86/corner_match_sse4.c',
+    '../../third_party/aom/aom_dsp/flow_estimation/x86/disflow_avx2.c',
     '../../third_party/aom/aom_dsp/flow_estimation/x86/disflow_sse4.c',
     '../../third_party/aom/aom_dsp/fwd_txfm.c',
     '../../third_party/aom/aom_dsp/grain_table.c',
@@ -533,11 +534,8 @@ files = {
     '../../third_party/aom/aom_dsp/x86/aom_high_subpixel_bilinear_sse2.asm',
     '../../third_party/aom/aom_dsp/x86/aom_quantize_avx.c',
     '../../third_party/aom/aom_dsp/x86/aom_subpixel_8t_intrin_avx2.c',
-    '../../third_party/aom/aom_dsp/x86/aom_subpixel_8t_intrin_sse2.c',
     '../../third_party/aom/aom_dsp/x86/aom_subpixel_8t_intrin_ssse3.c',
-    '../../third_party/aom/aom_dsp/x86/aom_subpixel_8t_sse2.asm',
     '../../third_party/aom/aom_dsp/x86/aom_subpixel_8t_ssse3.asm',
-    '../../third_party/aom/aom_dsp/x86/aom_subpixel_bilinear_sse2.asm',
     '../../third_party/aom/aom_dsp/x86/aom_subpixel_bilinear_ssse3.asm',
     '../../third_party/aom/aom_dsp/x86/avg_intrin_avx2.c',
     '../../third_party/aom/aom_dsp/x86/avg_intrin_sse2.c',
@@ -599,7 +597,7 @@ files = {
     '../../third_party/aom/aom_dsp/x86/sad_sse2.asm',
     '../../third_party/aom/aom_dsp/x86/sse_avx2.c',
     '../../third_party/aom/aom_dsp/x86/sse_sse4.c',
-    '../../third_party/aom/aom_dsp/x86/subpel_variance_sse2.asm',
+    '../../third_party/aom/aom_dsp/x86/subpel_variance_ssse3.asm',
     '../../third_party/aom/aom_dsp/x86/subtract_avx2.c',
     '../../third_party/aom/aom_dsp/x86/subtract_sse2.asm',
     '../../third_party/aom/aom_dsp/x86/sum_squares_avx2.c',
@@ -658,7 +656,6 @@ files = {
     '../../third_party/aom/av1/common/x86/av1_inv_txfm_ssse3.c',
     '../../third_party/aom/av1/common/x86/av1_txfm_sse4.c',
     '../../third_party/aom/av1/common/x86/cdef_block_avx2.c',
-    '../../third_party/aom/av1/common/x86/cdef_block_sse2.c',
     '../../third_party/aom/av1/common/x86/cdef_block_sse4.c',
     '../../third_party/aom/av1/common/x86/cdef_block_ssse3.c',
     '../../third_party/aom/av1/common/x86/cfl_avx2.c',
@@ -859,6 +856,7 @@ files = {
     '../../third_party/aom/aom_dsp/flow_estimation/ransac.c',
     '../../third_party/aom/aom_dsp/flow_estimation/x86/corner_match_avx2.c',
     '../../third_party/aom/aom_dsp/flow_estimation/x86/corner_match_sse4.c',
+    '../../third_party/aom/aom_dsp/flow_estimation/x86/disflow_avx2.c',
     '../../third_party/aom/aom_dsp/flow_estimation/x86/disflow_sse4.c',
     '../../third_party/aom/aom_dsp/fwd_txfm.c',
     '../../third_party/aom/aom_dsp/grain_table.c',
@@ -886,11 +884,8 @@ files = {
     '../../third_party/aom/aom_dsp/x86/aom_high_subpixel_bilinear_sse2.asm',
     '../../third_party/aom/aom_dsp/x86/aom_quantize_avx.c',
     '../../third_party/aom/aom_dsp/x86/aom_subpixel_8t_intrin_avx2.c',
-    '../../third_party/aom/aom_dsp/x86/aom_subpixel_8t_intrin_sse2.c',
     '../../third_party/aom/aom_dsp/x86/aom_subpixel_8t_intrin_ssse3.c',
-    '../../third_party/aom/aom_dsp/x86/aom_subpixel_8t_sse2.asm',
     '../../third_party/aom/aom_dsp/x86/aom_subpixel_8t_ssse3.asm',
-    '../../third_party/aom/aom_dsp/x86/aom_subpixel_bilinear_sse2.asm',
     '../../third_party/aom/aom_dsp/x86/aom_subpixel_bilinear_ssse3.asm',
     '../../third_party/aom/aom_dsp/x86/avg_intrin_avx2.c',
     '../../third_party/aom/aom_dsp/x86/avg_intrin_sse2.c',
@@ -955,7 +950,7 @@ files = {
     '../../third_party/aom/aom_dsp/x86/sse_avx2.c',
     '../../third_party/aom/aom_dsp/x86/sse_sse4.c',
     '../../third_party/aom/aom_dsp/x86/ssim_sse2_x86_64.asm',
-    '../../third_party/aom/aom_dsp/x86/subpel_variance_sse2.asm',
+    '../../third_party/aom/aom_dsp/x86/subpel_variance_ssse3.asm',
     '../../third_party/aom/aom_dsp/x86/subtract_avx2.c',
     '../../third_party/aom/aom_dsp/x86/subtract_sse2.asm',
     '../../third_party/aom/aom_dsp/x86/sum_squares_avx2.c',
@@ -1014,9 +1009,7 @@ files = {
     '../../third_party/aom/av1/common/x86/av1_inv_txfm_ssse3.c',
     '../../third_party/aom/av1/common/x86/av1_txfm_sse4.c',
     '../../third_party/aom/av1/common/x86/cdef_block_avx2.c',
-    '../../third_party/aom/av1/common/x86/cdef_block_sse2.c',
     '../../third_party/aom/av1/common/x86/cdef_block_sse4.c',
-    '../../third_party/aom/av1/common/x86/cdef_block_ssse3.c',
     '../../third_party/aom/av1/common/x86/cfl_avx2.c',
     '../../third_party/aom/av1/common/x86/cfl_sse2.c',
     '../../third_party/aom/av1/common/x86/cfl_ssse3.c',
author	Daniel Baumann <daniel.baumann@progress-linux.org>	2024-05-15 03:35:49 +0000
committer	Daniel Baumann <daniel.baumann@progress-linux.org>	2024-05-15 03:35:49 +0000
commit	d8bbc7858622b6d9c278469aab701ca0b609cddf (patch)
tree	eff41dc61d9f714852212739e6b3738b82a2af87 /media/libaom
parent	Releasing progress-linux version 125.0.3-1~progress7.99u1. (diff)
download	firefox-d8bbc7858622b6d9c278469aab701ca0b609cddf.tar.xz firefox-d8bbc7858622b6d9c278469aab701ca0b609cddf.zip