Update aom to commit id e87fb2378f01103d5d6e477a4ef6892dc714e614

1 files changed, 461 insertions, 160 deletions

diff --git a/media/libaom/config/win/x64/aom_dsp_rtcd.h b/media/libaom/config/win/x64/aom_dsp_rtcd.h
index 6ee856f1d..d2bee385f 100644
--- a/media/libaom/config/win/x64/aom_dsp_rtcd.h
+++ b/media/libaom/config/win/x64/aom_dsp_rtcd.h
@@ -20,6 +20,9 @@
 extern "C" {
 #endif
 
+void aom_blend_a64_d32_mask_c(int32_t *dst, uint32_t dst_stride, const int32_t *src0, uint32_t src0_stride, const int32_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int h, int w, int suby, int subx);
+#define aom_blend_a64_d32_mask aom_blend_a64_d32_mask_c
+
 void aom_blend_a64_hmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int h, int w);
 void aom_blend_a64_hmask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int h, int w);
 RTCD_EXTERN void (*aom_blend_a64_hmask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int h, int w);
@@ -51,6 +54,28 @@ void aom_convolve8_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
 void aom_convolve8_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)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t 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_add_src_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+void aom_convolve8_add_src_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+RTCD_EXTERN void (*aom_convolve8_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);
+
+void aom_convolve8_add_src_hip_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_add_src_hip_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define aom_convolve8_add_src_hip aom_convolve8_add_src_hip_sse2
+
+void aom_convolve8_add_src_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_add_src_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);
+RTCD_EXTERN void (*aom_convolve8_add_src_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_add_src_horiz_hip_c(const uint8_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define aom_convolve8_add_src_horiz_hip aom_convolve8_add_src_horiz_hip_c
+
+void aom_convolve8_add_src_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_add_src_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);
+RTCD_EXTERN void (*aom_convolve8_add_src_vert)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+
+void aom_convolve8_add_src_vert_hip_c(const uint16_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
+#define aom_convolve8_add_src_vert_hip aom_convolve8_add_src_vert_hip_c
+
 void aom_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h);
 void aom_convolve8_avg_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_avg_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);
@@ -312,33 +337,41 @@ void aom_dc_128_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uin
 #define aom_dc_128_predictor_16x16 aom_dc_128_predictor_16x16_sse2
 
 void aom_dc_128_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_dc_128_predictor_16x32 aom_dc_128_predictor_16x32_c
+void aom_dc_128_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_16x32 aom_dc_128_predictor_16x32_sse2
 
 void aom_dc_128_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_dc_128_predictor_16x8 aom_dc_128_predictor_16x8_c
+void aom_dc_128_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_16x8 aom_dc_128_predictor_16x8_sse2
 
 void aom_dc_128_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_dc_128_predictor_2x2 aom_dc_128_predictor_2x2_c
 
 void aom_dc_128_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_dc_128_predictor_32x16 aom_dc_128_predictor_32x16_c
+void aom_dc_128_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_128_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void aom_dc_128_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void aom_dc_128_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_dc_128_predictor_32x32 aom_dc_128_predictor_32x32_sse2
+void aom_dc_128_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_128_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void aom_dc_128_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void aom_dc_128_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_dc_128_predictor_4x4 aom_dc_128_predictor_4x4_sse2
 
 void aom_dc_128_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_dc_128_predictor_4x8 aom_dc_128_predictor_4x8_c
+void aom_dc_128_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_4x8 aom_dc_128_predictor_4x8_sse2
 
 void aom_dc_128_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_dc_128_predictor_8x16 aom_dc_128_predictor_8x16_c
+void aom_dc_128_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_8x16 aom_dc_128_predictor_8x16_sse2
 
 void aom_dc_128_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_dc_128_predictor_8x4 aom_dc_128_predictor_8x4_c
+void aom_dc_128_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_128_predictor_8x4 aom_dc_128_predictor_8x4_sse2
 
 void aom_dc_128_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void aom_dc_128_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
@@ -349,33 +382,41 @@ void aom_dc_left_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const ui
 #define aom_dc_left_predictor_16x16 aom_dc_left_predictor_16x16_sse2
 
 void aom_dc_left_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_dc_left_predictor_16x32 aom_dc_left_predictor_16x32_c
+void aom_dc_left_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_16x32 aom_dc_left_predictor_16x32_sse2
 
 void aom_dc_left_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_dc_left_predictor_16x8 aom_dc_left_predictor_16x8_c
+void aom_dc_left_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_16x8 aom_dc_left_predictor_16x8_sse2
 
 void aom_dc_left_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_dc_left_predictor_2x2 aom_dc_left_predictor_2x2_c
 
 void aom_dc_left_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_dc_left_predictor_32x16 aom_dc_left_predictor_32x16_c
+void aom_dc_left_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_left_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void aom_dc_left_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void aom_dc_left_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_dc_left_predictor_32x32 aom_dc_left_predictor_32x32_sse2
+void aom_dc_left_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_left_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void aom_dc_left_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void aom_dc_left_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_dc_left_predictor_4x4 aom_dc_left_predictor_4x4_sse2
 
 void aom_dc_left_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_dc_left_predictor_4x8 aom_dc_left_predictor_4x8_c
+void aom_dc_left_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_4x8 aom_dc_left_predictor_4x8_sse2
 
 void aom_dc_left_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_dc_left_predictor_8x16 aom_dc_left_predictor_8x16_c
+void aom_dc_left_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_8x16 aom_dc_left_predictor_8x16_sse2
 
 void aom_dc_left_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_dc_left_predictor_8x4 aom_dc_left_predictor_8x4_c
+void aom_dc_left_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_left_predictor_8x4 aom_dc_left_predictor_8x4_sse2
 
 void aom_dc_left_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void aom_dc_left_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
@@ -386,33 +427,41 @@ void aom_dc_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t
 #define aom_dc_predictor_16x16 aom_dc_predictor_16x16_sse2
 
 void aom_dc_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_dc_predictor_16x32 aom_dc_predictor_16x32_c
+void aom_dc_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_16x32 aom_dc_predictor_16x32_sse2
 
 void aom_dc_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_dc_predictor_16x8 aom_dc_predictor_16x8_c
+void aom_dc_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_16x8 aom_dc_predictor_16x8_sse2
 
 void aom_dc_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_dc_predictor_2x2 aom_dc_predictor_2x2_c
 
 void aom_dc_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_dc_predictor_32x16 aom_dc_predictor_32x16_c
+void aom_dc_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void aom_dc_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void aom_dc_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_dc_predictor_32x32 aom_dc_predictor_32x32_sse2
+void aom_dc_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void aom_dc_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void aom_dc_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_dc_predictor_4x4 aom_dc_predictor_4x4_sse2
 
 void aom_dc_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_dc_predictor_4x8 aom_dc_predictor_4x8_c
+void aom_dc_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_4x8 aom_dc_predictor_4x8_sse2
 
 void aom_dc_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_dc_predictor_8x16 aom_dc_predictor_8x16_c
+void aom_dc_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_8x16 aom_dc_predictor_8x16_sse2
 
 void aom_dc_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_dc_predictor_8x4 aom_dc_predictor_8x4_c
+void aom_dc_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_predictor_8x4 aom_dc_predictor_8x4_sse2
 
 void aom_dc_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void aom_dc_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
@@ -423,33 +472,41 @@ void aom_dc_top_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uin
 #define aom_dc_top_predictor_16x16 aom_dc_top_predictor_16x16_sse2
 
 void aom_dc_top_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_dc_top_predictor_16x32 aom_dc_top_predictor_16x32_c
+void aom_dc_top_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_16x32 aom_dc_top_predictor_16x32_sse2
 
 void aom_dc_top_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_dc_top_predictor_16x8 aom_dc_top_predictor_16x8_c
+void aom_dc_top_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_16x8 aom_dc_top_predictor_16x8_sse2
 
 void aom_dc_top_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_dc_top_predictor_2x2 aom_dc_top_predictor_2x2_c
 
 void aom_dc_top_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_dc_top_predictor_32x16 aom_dc_top_predictor_32x16_c
+void aom_dc_top_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_dc_top_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void aom_dc_top_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void aom_dc_top_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_dc_top_predictor_32x32 aom_dc_top_predictor_32x32_sse2
+void aom_dc_top_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_dc_top_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void aom_dc_top_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void aom_dc_top_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_dc_top_predictor_4x4 aom_dc_top_predictor_4x4_sse2
 
 void aom_dc_top_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_dc_top_predictor_4x8 aom_dc_top_predictor_4x8_c
+void aom_dc_top_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_4x8 aom_dc_top_predictor_4x8_sse2
 
 void aom_dc_top_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_dc_top_predictor_8x16 aom_dc_top_predictor_8x16_c
+void aom_dc_top_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_8x16 aom_dc_top_predictor_8x16_sse2
 
 void aom_dc_top_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_dc_top_predictor_8x4 aom_dc_top_predictor_8x4_c
+void aom_dc_top_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_dc_top_predictor_8x4 aom_dc_top_predictor_8x4_sse2
 
 void aom_dc_top_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void aom_dc_top_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
@@ -503,33 +560,40 @@ void aom_h_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t
 #define aom_h_predictor_16x16 aom_h_predictor_16x16_sse2
 
 void aom_h_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_h_predictor_16x32 aom_h_predictor_16x32_c
+void aom_h_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_16x32 aom_h_predictor_16x32_sse2
 
 void aom_h_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_h_predictor_16x8 aom_h_predictor_16x8_c
+void aom_h_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_16x8 aom_h_predictor_16x8_sse2
 
 void aom_h_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_h_predictor_2x2 aom_h_predictor_2x2_c
 
 void aom_h_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_h_predictor_32x16 aom_h_predictor_32x16_c
+void aom_h_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_32x16 aom_h_predictor_32x16_sse2
 
 void aom_h_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void aom_h_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_h_predictor_32x32 aom_h_predictor_32x32_sse2
+void aom_h_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_h_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void aom_h_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void aom_h_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_h_predictor_4x4 aom_h_predictor_4x4_sse2
 
 void aom_h_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_h_predictor_4x8 aom_h_predictor_4x8_c
+void aom_h_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_4x8 aom_h_predictor_4x8_sse2
 
 void aom_h_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_h_predictor_8x16 aom_h_predictor_8x16_c
+void aom_h_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_8x16 aom_h_predictor_8x16_sse2
 
 void aom_h_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_h_predictor_8x4 aom_h_predictor_8x4_c
+void aom_h_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_h_predictor_8x4 aom_h_predictor_8x4_sse2
 
 void aom_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void aom_h_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
@@ -1455,6 +1519,26 @@ void aom_highbd_convolve8_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t
 void aom_highbd_convolve8_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
 RTCD_EXTERN void (*aom_highbd_convolve8)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
 
+void aom_highbd_convolve8_add_src_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+void aom_highbd_convolve8_add_src_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define aom_highbd_convolve8_add_src aom_highbd_convolve8_add_src_sse2
+
+void aom_highbd_convolve8_add_src_hip_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+void aom_highbd_convolve8_add_src_hip_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, int bps);
+RTCD_EXTERN void (*aom_highbd_convolve8_add_src_hip)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+
+void aom_highbd_convolve8_add_src_horiz_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define aom_highbd_convolve8_add_src_horiz aom_highbd_convolve8_add_src_horiz_c
+
+void aom_highbd_convolve8_add_src_horiz_hip_c(const uint8_t *src, ptrdiff_t src_stride, uint16_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define aom_highbd_convolve8_add_src_horiz_hip aom_highbd_convolve8_add_src_horiz_hip_c
+
+void aom_highbd_convolve8_add_src_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define aom_highbd_convolve8_add_src_vert aom_highbd_convolve8_add_src_vert_c
+
+void aom_highbd_convolve8_add_src_vert_hip_c(const uint16_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
+#define aom_highbd_convolve8_add_src_vert_hip aom_highbd_convolve8_add_src_vert_hip_c
+
 void aom_highbd_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
 void aom_highbd_convolve8_avg_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
 void aom_highbd_convolve8_avg_avx2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
@@ -1491,7 +1575,8 @@ void aom_highbd_convolve_copy_avx2(const uint8_t *src, ptrdiff_t src_stride, uin
 RTCD_EXTERN void (*aom_highbd_convolve_copy)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
 
 void aom_highbd_d117_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_d117_predictor_16x16 aom_highbd_d117_predictor_16x16_c
+void aom_highbd_d117_predictor_16x16_ssse3(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_d117_predictor_16x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 
 void aom_highbd_d117_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d117_predictor_16x32 aom_highbd_d117_predictor_16x32_c
@@ -1506,10 +1591,12 @@ void aom_highbd_d117_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const
 #define aom_highbd_d117_predictor_32x16 aom_highbd_d117_predictor_32x16_c
 
 void aom_highbd_d117_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_d117_predictor_32x32 aom_highbd_d117_predictor_32x32_c
+void aom_highbd_d117_predictor_32x32_ssse3(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_d117_predictor_32x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 
 void aom_highbd_d117_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_d117_predictor_4x4 aom_highbd_d117_predictor_4x4_c
+void aom_highbd_d117_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d117_predictor_4x4 aom_highbd_d117_predictor_4x4_sse2
 
 void aom_highbd_d117_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d117_predictor_4x8 aom_highbd_d117_predictor_4x8_c
@@ -1521,10 +1608,12 @@ void aom_highbd_d117_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const ui
 #define aom_highbd_d117_predictor_8x4 aom_highbd_d117_predictor_8x4_c
 
 void aom_highbd_d117_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_d117_predictor_8x8 aom_highbd_d117_predictor_8x8_c
+void aom_highbd_d117_predictor_8x8_ssse3(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_d117_predictor_8x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 
 void aom_highbd_d135_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_d135_predictor_16x16 aom_highbd_d135_predictor_16x16_c
+void aom_highbd_d135_predictor_16x16_ssse3(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_d135_predictor_16x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 
 void aom_highbd_d135_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d135_predictor_16x32 aom_highbd_d135_predictor_16x32_c
@@ -1539,10 +1628,12 @@ void aom_highbd_d135_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const
 #define aom_highbd_d135_predictor_32x16 aom_highbd_d135_predictor_32x16_c
 
 void aom_highbd_d135_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_d135_predictor_32x32 aom_highbd_d135_predictor_32x32_c
+void aom_highbd_d135_predictor_32x32_ssse3(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_d135_predictor_32x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 
 void aom_highbd_d135_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_d135_predictor_4x4 aom_highbd_d135_predictor_4x4_c
+void aom_highbd_d135_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d135_predictor_4x4 aom_highbd_d135_predictor_4x4_sse2
 
 void aom_highbd_d135_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d135_predictor_4x8 aom_highbd_d135_predictor_4x8_c
@@ -1554,10 +1645,12 @@ void aom_highbd_d135_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const ui
 #define aom_highbd_d135_predictor_8x4 aom_highbd_d135_predictor_8x4_c
 
 void aom_highbd_d135_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_d135_predictor_8x8 aom_highbd_d135_predictor_8x8_c
+void aom_highbd_d135_predictor_8x8_ssse3(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_d135_predictor_8x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 
 void aom_highbd_d153_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_d153_predictor_16x16 aom_highbd_d153_predictor_16x16_c
+void aom_highbd_d153_predictor_16x16_ssse3(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_d153_predictor_16x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 
 void aom_highbd_d153_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d153_predictor_16x32 aom_highbd_d153_predictor_16x32_c
@@ -1572,10 +1665,12 @@ void aom_highbd_d153_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const
 #define aom_highbd_d153_predictor_32x16 aom_highbd_d153_predictor_32x16_c
 
 void aom_highbd_d153_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_d153_predictor_32x32 aom_highbd_d153_predictor_32x32_c
+void aom_highbd_d153_predictor_32x32_ssse3(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_d153_predictor_32x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 
 void aom_highbd_d153_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_d153_predictor_4x4 aom_highbd_d153_predictor_4x4_c
+void aom_highbd_d153_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d153_predictor_4x4 aom_highbd_d153_predictor_4x4_sse2
 
 void aom_highbd_d153_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d153_predictor_4x8 aom_highbd_d153_predictor_4x8_c
@@ -1587,7 +1682,8 @@ void aom_highbd_d153_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const ui
 #define aom_highbd_d153_predictor_8x4 aom_highbd_d153_predictor_8x4_c
 
 void aom_highbd_d153_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_d153_predictor_8x8 aom_highbd_d153_predictor_8x8_c
+void aom_highbd_d153_predictor_8x8_ssse3(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_d153_predictor_8x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 
 void aom_highbd_d207e_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d207e_predictor_16x16 aom_highbd_d207e_predictor_16x16_c
@@ -1623,37 +1719,47 @@ void aom_highbd_d207e_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const u
 #define aom_highbd_d207e_predictor_8x8 aom_highbd_d207e_predictor_8x8_c
 
 void aom_highbd_d45e_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_d45e_predictor_16x16 aom_highbd_d45e_predictor_16x16_c
+void aom_highbd_d45e_predictor_16x16_avx2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_d45e_predictor_16x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 
 void aom_highbd_d45e_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_d45e_predictor_16x32 aom_highbd_d45e_predictor_16x32_c
+void aom_highbd_d45e_predictor_16x32_avx2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_d45e_predictor_16x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 
 void aom_highbd_d45e_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_d45e_predictor_16x8 aom_highbd_d45e_predictor_16x8_c
+void aom_highbd_d45e_predictor_16x8_avx2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_d45e_predictor_16x8)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 
 void aom_highbd_d45e_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d45e_predictor_2x2 aom_highbd_d45e_predictor_2x2_c
 
 void aom_highbd_d45e_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_d45e_predictor_32x16 aom_highbd_d45e_predictor_32x16_c
+void aom_highbd_d45e_predictor_32x16_avx2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_d45e_predictor_32x16)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 
 void aom_highbd_d45e_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_d45e_predictor_32x32 aom_highbd_d45e_predictor_32x32_c
+void aom_highbd_d45e_predictor_32x32_avx2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+RTCD_EXTERN void (*aom_highbd_d45e_predictor_32x32)(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 
 void aom_highbd_d45e_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_d45e_predictor_4x4 aom_highbd_d45e_predictor_4x4_c
+void aom_highbd_d45e_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d45e_predictor_4x4 aom_highbd_d45e_predictor_4x4_sse2
 
 void aom_highbd_d45e_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_d45e_predictor_4x8 aom_highbd_d45e_predictor_4x8_c
+void aom_highbd_d45e_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d45e_predictor_4x8 aom_highbd_d45e_predictor_4x8_sse2
 
 void aom_highbd_d45e_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_d45e_predictor_8x16 aom_highbd_d45e_predictor_8x16_c
+void aom_highbd_d45e_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d45e_predictor_8x16 aom_highbd_d45e_predictor_8x16_sse2
 
 void aom_highbd_d45e_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_d45e_predictor_8x4 aom_highbd_d45e_predictor_8x4_c
+void aom_highbd_d45e_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d45e_predictor_8x4 aom_highbd_d45e_predictor_8x4_sse2
 
 void aom_highbd_d45e_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_d45e_predictor_8x8 aom_highbd_d45e_predictor_8x8_c
+void aom_highbd_d45e_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_d45e_predictor_8x8 aom_highbd_d45e_predictor_8x8_sse2
 
 void aom_highbd_d63e_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_d63e_predictor_16x16 aom_highbd_d63e_predictor_16x16_c
@@ -1689,86 +1795,109 @@ void aom_highbd_d63e_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const ui
 #define aom_highbd_d63e_predictor_8x8 aom_highbd_d63e_predictor_8x8_c
 
 void aom_highbd_dc_128_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_dc_128_predictor_16x16 aom_highbd_dc_128_predictor_16x16_c
+void aom_highbd_dc_128_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_16x16 aom_highbd_dc_128_predictor_16x16_sse2
 
 void aom_highbd_dc_128_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_dc_128_predictor_16x32 aom_highbd_dc_128_predictor_16x32_c
+void aom_highbd_dc_128_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_16x32 aom_highbd_dc_128_predictor_16x32_sse2
 
 void aom_highbd_dc_128_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_dc_128_predictor_16x8 aom_highbd_dc_128_predictor_16x8_c
+void aom_highbd_dc_128_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_16x8 aom_highbd_dc_128_predictor_16x8_sse2
 
 void aom_highbd_dc_128_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_128_predictor_2x2 aom_highbd_dc_128_predictor_2x2_c
 
 void aom_highbd_dc_128_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_dc_128_predictor_32x16 aom_highbd_dc_128_predictor_32x16_c
+void aom_highbd_dc_128_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_32x16 aom_highbd_dc_128_predictor_32x16_sse2
 
 void aom_highbd_dc_128_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_dc_128_predictor_32x32 aom_highbd_dc_128_predictor_32x32_c
+void aom_highbd_dc_128_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_32x32 aom_highbd_dc_128_predictor_32x32_sse2
 
 void aom_highbd_dc_128_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_dc_128_predictor_4x4 aom_highbd_dc_128_predictor_4x4_c
+void aom_highbd_dc_128_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_4x4 aom_highbd_dc_128_predictor_4x4_sse2
 
 void aom_highbd_dc_128_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_dc_128_predictor_4x8 aom_highbd_dc_128_predictor_4x8_c
+void aom_highbd_dc_128_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_4x8 aom_highbd_dc_128_predictor_4x8_sse2
 
 void aom_highbd_dc_128_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_dc_128_predictor_8x16 aom_highbd_dc_128_predictor_8x16_c
+void aom_highbd_dc_128_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_8x16 aom_highbd_dc_128_predictor_8x16_sse2
 
 void aom_highbd_dc_128_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_dc_128_predictor_8x4 aom_highbd_dc_128_predictor_8x4_c
+void aom_highbd_dc_128_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_8x4 aom_highbd_dc_128_predictor_8x4_sse2
 
 void aom_highbd_dc_128_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_dc_128_predictor_8x8 aom_highbd_dc_128_predictor_8x8_c
+void aom_highbd_dc_128_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_128_predictor_8x8 aom_highbd_dc_128_predictor_8x8_sse2
 
 void aom_highbd_dc_left_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_dc_left_predictor_16x16 aom_highbd_dc_left_predictor_16x16_c
+void aom_highbd_dc_left_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_16x16 aom_highbd_dc_left_predictor_16x16_sse2
 
 void aom_highbd_dc_left_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_dc_left_predictor_16x32 aom_highbd_dc_left_predictor_16x32_c
+void aom_highbd_dc_left_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_16x32 aom_highbd_dc_left_predictor_16x32_sse2
 
 void aom_highbd_dc_left_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_dc_left_predictor_16x8 aom_highbd_dc_left_predictor_16x8_c
+void aom_highbd_dc_left_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_16x8 aom_highbd_dc_left_predictor_16x8_sse2
 
 void aom_highbd_dc_left_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_left_predictor_2x2 aom_highbd_dc_left_predictor_2x2_c
 
 void aom_highbd_dc_left_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_dc_left_predictor_32x16 aom_highbd_dc_left_predictor_32x16_c
+void aom_highbd_dc_left_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_32x16 aom_highbd_dc_left_predictor_32x16_sse2
 
 void aom_highbd_dc_left_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_dc_left_predictor_32x32 aom_highbd_dc_left_predictor_32x32_c
+void aom_highbd_dc_left_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_32x32 aom_highbd_dc_left_predictor_32x32_sse2
 
 void aom_highbd_dc_left_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_dc_left_predictor_4x4 aom_highbd_dc_left_predictor_4x4_c
+void aom_highbd_dc_left_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_4x4 aom_highbd_dc_left_predictor_4x4_sse2
 
 void aom_highbd_dc_left_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_dc_left_predictor_4x8 aom_highbd_dc_left_predictor_4x8_c
+void aom_highbd_dc_left_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_4x8 aom_highbd_dc_left_predictor_4x8_sse2
 
 void aom_highbd_dc_left_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_dc_left_predictor_8x16 aom_highbd_dc_left_predictor_8x16_c
+void aom_highbd_dc_left_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_8x16 aom_highbd_dc_left_predictor_8x16_sse2
 
 void aom_highbd_dc_left_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_dc_left_predictor_8x4 aom_highbd_dc_left_predictor_8x4_c
+void aom_highbd_dc_left_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_8x4 aom_highbd_dc_left_predictor_8x4_sse2
 
 void aom_highbd_dc_left_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_dc_left_predictor_8x8 aom_highbd_dc_left_predictor_8x8_c
+void aom_highbd_dc_left_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_left_predictor_8x8 aom_highbd_dc_left_predictor_8x8_sse2
 
 void aom_highbd_dc_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 void aom_highbd_dc_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_predictor_16x16 aom_highbd_dc_predictor_16x16_sse2
 
 void aom_highbd_dc_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_dc_predictor_16x32 aom_highbd_dc_predictor_16x32_c
+void aom_highbd_dc_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_16x32 aom_highbd_dc_predictor_16x32_sse2
 
 void aom_highbd_dc_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_dc_predictor_16x8 aom_highbd_dc_predictor_16x8_c
+void aom_highbd_dc_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_16x8 aom_highbd_dc_predictor_16x8_sse2
 
 void aom_highbd_dc_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_predictor_2x2 aom_highbd_dc_predictor_2x2_c
 
 void aom_highbd_dc_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_dc_predictor_32x16 aom_highbd_dc_predictor_32x16_c
+void aom_highbd_dc_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_32x16 aom_highbd_dc_predictor_32x16_sse2
 
 void aom_highbd_dc_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 void aom_highbd_dc_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
@@ -1779,50 +1908,63 @@ void aom_highbd_dc_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const u
 #define aom_highbd_dc_predictor_4x4 aom_highbd_dc_predictor_4x4_sse2
 
 void aom_highbd_dc_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_dc_predictor_4x8 aom_highbd_dc_predictor_4x8_c
+void aom_highbd_dc_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_4x8 aom_highbd_dc_predictor_4x8_sse2
 
 void aom_highbd_dc_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_dc_predictor_8x16 aom_highbd_dc_predictor_8x16_c
+void aom_highbd_dc_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_8x16 aom_highbd_dc_predictor_8x16_sse2
 
 void aom_highbd_dc_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_dc_predictor_8x4 aom_highbd_dc_predictor_8x4_c
+void aom_highbd_dc_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_predictor_8x4 aom_highbd_dc_predictor_8x4_sse2
 
 void aom_highbd_dc_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 void aom_highbd_dc_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_predictor_8x8 aom_highbd_dc_predictor_8x8_sse2
 
 void aom_highbd_dc_top_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_dc_top_predictor_16x16 aom_highbd_dc_top_predictor_16x16_c
+void aom_highbd_dc_top_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_16x16 aom_highbd_dc_top_predictor_16x16_sse2
 
 void aom_highbd_dc_top_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_dc_top_predictor_16x32 aom_highbd_dc_top_predictor_16x32_c
+void aom_highbd_dc_top_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_16x32 aom_highbd_dc_top_predictor_16x32_sse2
 
 void aom_highbd_dc_top_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_dc_top_predictor_16x8 aom_highbd_dc_top_predictor_16x8_c
+void aom_highbd_dc_top_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_16x8 aom_highbd_dc_top_predictor_16x8_sse2
 
 void aom_highbd_dc_top_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_dc_top_predictor_2x2 aom_highbd_dc_top_predictor_2x2_c
 
 void aom_highbd_dc_top_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_dc_top_predictor_32x16 aom_highbd_dc_top_predictor_32x16_c
+void aom_highbd_dc_top_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_32x16 aom_highbd_dc_top_predictor_32x16_sse2
 
 void aom_highbd_dc_top_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_dc_top_predictor_32x32 aom_highbd_dc_top_predictor_32x32_c
+void aom_highbd_dc_top_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_32x32 aom_highbd_dc_top_predictor_32x32_sse2
 
 void aom_highbd_dc_top_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_dc_top_predictor_4x4 aom_highbd_dc_top_predictor_4x4_c
+void aom_highbd_dc_top_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_4x4 aom_highbd_dc_top_predictor_4x4_sse2
 
 void aom_highbd_dc_top_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_dc_top_predictor_4x8 aom_highbd_dc_top_predictor_4x8_c
+void aom_highbd_dc_top_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_4x8 aom_highbd_dc_top_predictor_4x8_sse2
 
 void aom_highbd_dc_top_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_dc_top_predictor_8x16 aom_highbd_dc_top_predictor_8x16_c
+void aom_highbd_dc_top_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_8x16 aom_highbd_dc_top_predictor_8x16_sse2
 
 void aom_highbd_dc_top_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_dc_top_predictor_8x4 aom_highbd_dc_top_predictor_8x4_c
+void aom_highbd_dc_top_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_8x4 aom_highbd_dc_top_predictor_8x4_sse2
 
 void aom_highbd_dc_top_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_dc_top_predictor_8x8 aom_highbd_dc_top_predictor_8x8_c
+void aom_highbd_dc_top_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_dc_top_predictor_8x8 aom_highbd_dc_top_predictor_8x8_sse2
 
 void aom_highbd_fdct16x16_c(const int16_t *input, tran_low_t *output, int stride);
 void aom_highbd_fdct16x16_sse2(const int16_t *input, tran_low_t *output, int stride);
@@ -1845,37 +1987,47 @@ void aom_highbd_fdct8x8_sse2(const int16_t *input, tran_low_t *output, int strid
 #define aom_highbd_fdct8x8 aom_highbd_fdct8x8_sse2
 
 void aom_highbd_h_predictor_16x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_h_predictor_16x16 aom_highbd_h_predictor_16x16_c
+void aom_highbd_h_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_16x16 aom_highbd_h_predictor_16x16_sse2
 
 void aom_highbd_h_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_h_predictor_16x32 aom_highbd_h_predictor_16x32_c
+void aom_highbd_h_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_16x32 aom_highbd_h_predictor_16x32_sse2
 
 void aom_highbd_h_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_h_predictor_16x8 aom_highbd_h_predictor_16x8_c
+void aom_highbd_h_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_16x8 aom_highbd_h_predictor_16x8_sse2
 
 void aom_highbd_h_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_h_predictor_2x2 aom_highbd_h_predictor_2x2_c
 
 void aom_highbd_h_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_h_predictor_32x16 aom_highbd_h_predictor_32x16_c
+void aom_highbd_h_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_32x16 aom_highbd_h_predictor_32x16_sse2
 
 void aom_highbd_h_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_h_predictor_32x32 aom_highbd_h_predictor_32x32_c
+void aom_highbd_h_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_32x32 aom_highbd_h_predictor_32x32_sse2
 
 void aom_highbd_h_predictor_4x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_h_predictor_4x4 aom_highbd_h_predictor_4x4_c
+void aom_highbd_h_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_4x4 aom_highbd_h_predictor_4x4_sse2
 
 void aom_highbd_h_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_h_predictor_4x8 aom_highbd_h_predictor_4x8_c
+void aom_highbd_h_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_4x8 aom_highbd_h_predictor_4x8_sse2
 
 void aom_highbd_h_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_h_predictor_8x16 aom_highbd_h_predictor_8x16_c
+void aom_highbd_h_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_8x16 aom_highbd_h_predictor_8x16_sse2
 
 void aom_highbd_h_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_h_predictor_8x4 aom_highbd_h_predictor_8x4_c
+void aom_highbd_h_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_8x4 aom_highbd_h_predictor_8x4_sse2
 
 void aom_highbd_h_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_h_predictor_8x8 aom_highbd_h_predictor_8x8_c
+void aom_highbd_h_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_h_predictor_8x8 aom_highbd_h_predictor_8x8_sse2
 
 void aom_highbd_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
 #define aom_highbd_iwht4x4_16_add aom_highbd_iwht4x4_16_add_c
@@ -1889,7 +2041,8 @@ void aom_highbd_lpf_horizontal_4_sse2(uint16_t *s, int pitch, const uint8_t *bli
 
 void aom_highbd_lpf_horizontal_4_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
 void aom_highbd_lpf_horizontal_4_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
-#define aom_highbd_lpf_horizontal_4_dual aom_highbd_lpf_horizontal_4_dual_sse2
+void aom_highbd_lpf_horizontal_4_dual_avx2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_horizontal_4_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
 
 void aom_highbd_lpf_horizontal_8_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
 void aom_highbd_lpf_horizontal_8_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
@@ -1897,11 +2050,13 @@ void aom_highbd_lpf_horizontal_8_sse2(uint16_t *s, int pitch, const uint8_t *bli
 
 void aom_highbd_lpf_horizontal_8_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
 void aom_highbd_lpf_horizontal_8_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
-#define aom_highbd_lpf_horizontal_8_dual aom_highbd_lpf_horizontal_8_dual_sse2
+void aom_highbd_lpf_horizontal_8_dual_avx2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_horizontal_8_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
 
 void aom_highbd_lpf_horizontal_edge_16_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
 void aom_highbd_lpf_horizontal_edge_16_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
-#define aom_highbd_lpf_horizontal_edge_16 aom_highbd_lpf_horizontal_edge_16_sse2
+void aom_highbd_lpf_horizontal_edge_16_avx2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_horizontal_edge_16)(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
 
 void aom_highbd_lpf_horizontal_edge_8_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
 void aom_highbd_lpf_horizontal_edge_8_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
@@ -1913,7 +2068,8 @@ void aom_highbd_lpf_vertical_16_sse2(uint16_t *s, int pitch, const uint8_t *blim
 
 void aom_highbd_lpf_vertical_16_dual_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
 void aom_highbd_lpf_vertical_16_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
-#define aom_highbd_lpf_vertical_16_dual aom_highbd_lpf_vertical_16_dual_sse2
+void aom_highbd_lpf_vertical_16_dual_avx2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_vertical_16_dual)(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
 
 void aom_highbd_lpf_vertical_4_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
 void aom_highbd_lpf_vertical_4_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
@@ -1921,7 +2077,8 @@ void aom_highbd_lpf_vertical_4_sse2(uint16_t *s, int pitch, const uint8_t *blimi
 
 void aom_highbd_lpf_vertical_4_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
 void aom_highbd_lpf_vertical_4_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
-#define aom_highbd_lpf_vertical_4_dual aom_highbd_lpf_vertical_4_dual_sse2
+void aom_highbd_lpf_vertical_4_dual_avx2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_vertical_4_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
 
 void aom_highbd_lpf_vertical_8_c(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
 void aom_highbd_lpf_vertical_8_sse2(uint16_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh, int bd);
@@ -1929,7 +2086,8 @@ void aom_highbd_lpf_vertical_8_sse2(uint16_t *s, int pitch, const uint8_t *blimi
 
 void aom_highbd_lpf_vertical_8_dual_c(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
 void aom_highbd_lpf_vertical_8_dual_sse2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
-#define aom_highbd_lpf_vertical_8_dual aom_highbd_lpf_vertical_8_dual_sse2
+void aom_highbd_lpf_vertical_8_dual_avx2(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
+RTCD_EXTERN void (*aom_highbd_lpf_vertical_8_dual)(uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1, int bd);
 
 unsigned int aom_highbd_masked_sad16x16_c(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
 unsigned int aom_highbd_masked_sad16x16_ssse3(const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask);
@@ -2510,16 +2668,19 @@ void aom_highbd_v_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const
 #define aom_highbd_v_predictor_16x16 aom_highbd_v_predictor_16x16_sse2
 
 void aom_highbd_v_predictor_16x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_v_predictor_16x32 aom_highbd_v_predictor_16x32_c
+void aom_highbd_v_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_16x32 aom_highbd_v_predictor_16x32_sse2
 
 void aom_highbd_v_predictor_16x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_v_predictor_16x8 aom_highbd_v_predictor_16x8_c
+void aom_highbd_v_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_16x8 aom_highbd_v_predictor_16x8_sse2
 
 void aom_highbd_v_predictor_2x2_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 #define aom_highbd_v_predictor_2x2 aom_highbd_v_predictor_2x2_c
 
 void aom_highbd_v_predictor_32x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_v_predictor_32x16 aom_highbd_v_predictor_32x16_c
+void aom_highbd_v_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_32x16 aom_highbd_v_predictor_32x16_sse2
 
 void aom_highbd_v_predictor_32x32_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 void aom_highbd_v_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
@@ -2530,13 +2691,16 @@ void aom_highbd_v_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const ui
 #define aom_highbd_v_predictor_4x4 aom_highbd_v_predictor_4x4_sse2
 
 void aom_highbd_v_predictor_4x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_v_predictor_4x8 aom_highbd_v_predictor_4x8_c
+void aom_highbd_v_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_4x8 aom_highbd_v_predictor_4x8_sse2
 
 void aom_highbd_v_predictor_8x16_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_v_predictor_8x16 aom_highbd_v_predictor_8x16_c
+void aom_highbd_v_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_8x16 aom_highbd_v_predictor_8x16_sse2
 
 void aom_highbd_v_predictor_8x4_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
-#define aom_highbd_v_predictor_8x4 aom_highbd_v_predictor_8x4_c
+void aom_highbd_v_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
+#define aom_highbd_v_predictor_8x4 aom_highbd_v_predictor_8x4_sse2
 
 void aom_highbd_v_predictor_8x8_c(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
 void aom_highbd_v_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bd);
@@ -2626,50 +2790,43 @@ void aom_lpf_horizontal_4_sse2(uint8_t *s, int pitch, const uint8_t *blimit, con
 #define aom_lpf_horizontal_4 aom_lpf_horizontal_4_sse2
 
 void aom_lpf_horizontal_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
-void aom_lpf_horizontal_4_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
-#define aom_lpf_horizontal_4_dual aom_lpf_horizontal_4_dual_sse2
+#define aom_lpf_horizontal_4_dual aom_lpf_horizontal_4_dual_c
 
 void aom_lpf_horizontal_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
 void aom_lpf_horizontal_8_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
 #define aom_lpf_horizontal_8 aom_lpf_horizontal_8_sse2
 
 void aom_lpf_horizontal_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
-void aom_lpf_horizontal_8_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
-#define aom_lpf_horizontal_8_dual aom_lpf_horizontal_8_dual_sse2
+#define aom_lpf_horizontal_8_dual aom_lpf_horizontal_8_dual_c
 
 void aom_lpf_horizontal_edge_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
 void aom_lpf_horizontal_edge_16_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
-void aom_lpf_horizontal_edge_16_avx2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
-RTCD_EXTERN void (*aom_lpf_horizontal_edge_16)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define aom_lpf_horizontal_edge_16 aom_lpf_horizontal_edge_16_sse2
 
 void aom_lpf_horizontal_edge_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
 void aom_lpf_horizontal_edge_8_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
-void aom_lpf_horizontal_edge_8_avx2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
-RTCD_EXTERN void (*aom_lpf_horizontal_edge_8)(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
+#define aom_lpf_horizontal_edge_8 aom_lpf_horizontal_edge_8_sse2
 
 void aom_lpf_vertical_16_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
 void aom_lpf_vertical_16_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
 #define aom_lpf_vertical_16 aom_lpf_vertical_16_sse2
 
 void aom_lpf_vertical_16_dual_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
-void aom_lpf_vertical_16_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
-#define aom_lpf_vertical_16_dual aom_lpf_vertical_16_dual_sse2
+#define aom_lpf_vertical_16_dual aom_lpf_vertical_16_dual_c
 
 void aom_lpf_vertical_4_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
 void aom_lpf_vertical_4_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
 #define aom_lpf_vertical_4 aom_lpf_vertical_4_sse2
 
 void aom_lpf_vertical_4_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
-void aom_lpf_vertical_4_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
-#define aom_lpf_vertical_4_dual aom_lpf_vertical_4_dual_sse2
+#define aom_lpf_vertical_4_dual aom_lpf_vertical_4_dual_c
 
 void aom_lpf_vertical_8_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
 void aom_lpf_vertical_8_sse2(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
 #define aom_lpf_vertical_8 aom_lpf_vertical_8_sse2
 
 void aom_lpf_vertical_8_dual_c(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
-void aom_lpf_vertical_8_dual_sse2(uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1);
-#define aom_lpf_vertical_8_dual aom_lpf_vertical_8_dual_sse2
+#define aom_lpf_vertical_8_dual aom_lpf_vertical_8_dual_c
 
 unsigned int aom_masked_sad16x16_c(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
 unsigned int aom_masked_sad16x16_ssse3(const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask);
@@ -2940,37 +3097,52 @@ unsigned int aom_obmc_variance8x8_sse4_1(const uint8_t *pre, int pre_stride, con
 RTCD_EXTERN unsigned int (*aom_obmc_variance8x8)(const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse);
 
 void aom_paeth_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_paeth_predictor_16x16 aom_paeth_predictor_16x16_c
+void aom_paeth_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void aom_paeth_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_paeth_predictor_16x32 aom_paeth_predictor_16x32_c
+void aom_paeth_predictor_16x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void aom_paeth_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_paeth_predictor_16x8 aom_paeth_predictor_16x8_c
+void aom_paeth_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_16x8_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void aom_paeth_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_paeth_predictor_2x2 aom_paeth_predictor_2x2_c
 
 void aom_paeth_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_paeth_predictor_32x16 aom_paeth_predictor_32x16_c
+void aom_paeth_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void aom_paeth_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_paeth_predictor_32x32 aom_paeth_predictor_32x32_c
+void aom_paeth_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_paeth_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void aom_paeth_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_paeth_predictor_4x4 aom_paeth_predictor_4x4_c
+void aom_paeth_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void aom_paeth_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_paeth_predictor_4x8 aom_paeth_predictor_4x8_c
+void aom_paeth_predictor_4x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void aom_paeth_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_paeth_predictor_8x16 aom_paeth_predictor_8x16_c
+void aom_paeth_predictor_8x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void aom_paeth_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_paeth_predictor_8x4 aom_paeth_predictor_8x4_c
+void aom_paeth_predictor_8x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void aom_paeth_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_paeth_predictor_8x8 aom_paeth_predictor_8x8_c
+void aom_paeth_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_paeth_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void aom_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
 void aom_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block, const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr, const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr, const int16_t *scan, const int16_t *iscan);
@@ -3273,37 +3445,47 @@ void aom_smooth_h_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_
 #define aom_smooth_h_predictor_8x8 aom_smooth_h_predictor_8x8_c
 
 void aom_smooth_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_smooth_predictor_16x16 aom_smooth_predictor_16x16_c
+void aom_smooth_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void aom_smooth_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_smooth_predictor_16x32 aom_smooth_predictor_16x32_c
+void aom_smooth_predictor_16x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void aom_smooth_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_smooth_predictor_16x8 aom_smooth_predictor_16x8_c
+void aom_smooth_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_16x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void aom_smooth_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_smooth_predictor_2x2 aom_smooth_predictor_2x2_c
 
 void aom_smooth_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_smooth_predictor_32x16 aom_smooth_predictor_32x16_c
+void aom_smooth_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void aom_smooth_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_smooth_predictor_32x32 aom_smooth_predictor_32x32_c
+void aom_smooth_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void aom_smooth_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_smooth_predictor_4x4 aom_smooth_predictor_4x4_c
+void aom_smooth_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_4x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void aom_smooth_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_smooth_predictor_4x8 aom_smooth_predictor_4x8_c
+void aom_smooth_predictor_4x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_4x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void aom_smooth_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_smooth_predictor_8x16 aom_smooth_predictor_8x16_c
+void aom_smooth_predictor_8x16_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_8x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void aom_smooth_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_smooth_predictor_8x4 aom_smooth_predictor_8x4_c
+void aom_smooth_predictor_8x4_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_8x4)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void aom_smooth_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_smooth_predictor_8x8 aom_smooth_predictor_8x8_c
+void aom_smooth_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_smooth_predictor_8x8)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void aom_smooth_v_predictor_16x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_smooth_v_predictor_16x16 aom_smooth_v_predictor_16x16_c
@@ -3493,33 +3675,41 @@ void aom_v_predictor_16x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t
 #define aom_v_predictor_16x16 aom_v_predictor_16x16_sse2
 
 void aom_v_predictor_16x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_v_predictor_16x32 aom_v_predictor_16x32_c
+void aom_v_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_16x32 aom_v_predictor_16x32_sse2
 
 void aom_v_predictor_16x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_v_predictor_16x8 aom_v_predictor_16x8_c
+void aom_v_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_16x8 aom_v_predictor_16x8_sse2
 
 void aom_v_predictor_2x2_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_v_predictor_2x2 aom_v_predictor_2x2_c
 
 void aom_v_predictor_32x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_v_predictor_32x16 aom_v_predictor_32x16_c
+void aom_v_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+void aom_v_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_32x16)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void aom_v_predictor_32x32_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void aom_v_predictor_32x32_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_v_predictor_32x32 aom_v_predictor_32x32_sse2
+void aom_v_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+RTCD_EXTERN void (*aom_v_predictor_32x32)(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 
 void aom_v_predictor_4x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void aom_v_predictor_4x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 #define aom_v_predictor_4x4 aom_v_predictor_4x4_sse2
 
 void aom_v_predictor_4x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_v_predictor_4x8 aom_v_predictor_4x8_c
+void aom_v_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_4x8 aom_v_predictor_4x8_sse2
 
 void aom_v_predictor_8x16_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_v_predictor_8x16 aom_v_predictor_8x16_c
+void aom_v_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_8x16 aom_v_predictor_8x16_sse2
 
 void aom_v_predictor_8x4_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
-#define aom_v_predictor_8x4 aom_v_predictor_8x4_c
+void aom_v_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
+#define aom_v_predictor_8x4 aom_v_predictor_8x4_sse2
 
 void aom_v_predictor_8x8_c(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
 void aom_v_predictor_8x8_sse2(uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left);
@@ -3626,6 +3816,12 @@ static void setup_rtcd_internal(void)
     aom_convolve8 = aom_convolve8_sse2;
     if (flags & HAS_SSSE3) aom_convolve8 = aom_convolve8_ssse3;
     if (flags & HAS_AVX2) aom_convolve8 = aom_convolve8_avx2;
+    aom_convolve8_add_src = aom_convolve8_add_src_c;
+    if (flags & HAS_SSSE3) aom_convolve8_add_src = aom_convolve8_add_src_ssse3;
+    aom_convolve8_add_src_horiz = aom_convolve8_add_src_horiz_c;
+    if (flags & HAS_SSSE3) aom_convolve8_add_src_horiz = aom_convolve8_add_src_horiz_ssse3;
+    aom_convolve8_add_src_vert = aom_convolve8_add_src_vert_c;
+    if (flags & HAS_SSSE3) aom_convolve8_add_src_vert = aom_convolve8_add_src_vert_ssse3;
     aom_convolve8_avg = aom_convolve8_avg_sse2;
     if (flags & HAS_SSSE3) aom_convolve8_avg = aom_convolve8_avg_ssse3;
     aom_convolve8_avg_horiz = aom_convolve8_avg_horiz_sse2;
@@ -3648,6 +3844,22 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_SSSE3) aom_d153_predictor_8x8 = aom_d153_predictor_8x8_ssse3;
     aom_d63e_predictor_4x4 = aom_d63e_predictor_4x4_c;
     if (flags & HAS_SSSE3) aom_d63e_predictor_4x4 = aom_d63e_predictor_4x4_ssse3;
+    aom_dc_128_predictor_32x16 = aom_dc_128_predictor_32x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_128_predictor_32x16 = aom_dc_128_predictor_32x16_avx2;
+    aom_dc_128_predictor_32x32 = aom_dc_128_predictor_32x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_128_predictor_32x32 = aom_dc_128_predictor_32x32_avx2;
+    aom_dc_left_predictor_32x16 = aom_dc_left_predictor_32x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_left_predictor_32x16 = aom_dc_left_predictor_32x16_avx2;
+    aom_dc_left_predictor_32x32 = aom_dc_left_predictor_32x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_left_predictor_32x32 = aom_dc_left_predictor_32x32_avx2;
+    aom_dc_predictor_32x16 = aom_dc_predictor_32x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_predictor_32x16 = aom_dc_predictor_32x16_avx2;
+    aom_dc_predictor_32x32 = aom_dc_predictor_32x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_predictor_32x32 = aom_dc_predictor_32x32_avx2;
+    aom_dc_top_predictor_32x16 = aom_dc_top_predictor_32x16_sse2;
+    if (flags & HAS_AVX2) aom_dc_top_predictor_32x16 = aom_dc_top_predictor_32x16_avx2;
+    aom_dc_top_predictor_32x32 = aom_dc_top_predictor_32x32_sse2;
+    if (flags & HAS_AVX2) aom_dc_top_predictor_32x32 = aom_dc_top_predictor_32x32_avx2;
     aom_fdct32x32 = aom_fdct32x32_sse2;
     if (flags & HAS_AVX2) aom_fdct32x32 = aom_fdct32x32_avx2;
     aom_fdct32x32_rd = aom_fdct32x32_rd_sse2;
@@ -3656,6 +3868,8 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_SSSE3) aom_fdct8x8 = aom_fdct8x8_ssse3;
     aom_get16x16var = aom_get16x16var_sse2;
     if (flags & HAS_AVX2) aom_get16x16var = aom_get16x16var_avx2;
+    aom_h_predictor_32x32 = aom_h_predictor_32x32_sse2;
+    if (flags & HAS_AVX2) aom_h_predictor_32x32 = aom_h_predictor_32x32_avx2;
     aom_hadamard_8x8 = aom_hadamard_8x8_sse2;
     if (flags & HAS_SSSE3) aom_hadamard_8x8 = aom_hadamard_8x8_ssse3;
     aom_highbd_10_masked_sub_pixel_variance16x16 = aom_highbd_10_masked_sub_pixel_variance16x16_c;
@@ -3814,6 +4028,8 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_SSE4_1) aom_highbd_blend_a64_vmask = aom_highbd_blend_a64_vmask_sse4_1;
     aom_highbd_convolve8 = aom_highbd_convolve8_sse2;
     if (flags & HAS_AVX2) aom_highbd_convolve8 = aom_highbd_convolve8_avx2;
+    aom_highbd_convolve8_add_src_hip = aom_highbd_convolve8_add_src_hip_c;
+    if (flags & HAS_SSSE3) aom_highbd_convolve8_add_src_hip = aom_highbd_convolve8_add_src_hip_ssse3;
     aom_highbd_convolve8_avg = aom_highbd_convolve8_avg_sse2;
     if (flags & HAS_AVX2) aom_highbd_convolve8_avg = aom_highbd_convolve8_avg_avx2;
     aom_highbd_convolve8_avg_horiz = aom_highbd_convolve8_avg_horiz_sse2;
@@ -3828,6 +4044,46 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_AVX2) aom_highbd_convolve_avg = aom_highbd_convolve_avg_avx2;
     aom_highbd_convolve_copy = aom_highbd_convolve_copy_sse2;
     if (flags & HAS_AVX2) aom_highbd_convolve_copy = aom_highbd_convolve_copy_avx2;
+    aom_highbd_d117_predictor_16x16 = aom_highbd_d117_predictor_16x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_d117_predictor_16x16 = aom_highbd_d117_predictor_16x16_ssse3;
+    aom_highbd_d117_predictor_32x32 = aom_highbd_d117_predictor_32x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_d117_predictor_32x32 = aom_highbd_d117_predictor_32x32_ssse3;
+    aom_highbd_d117_predictor_8x8 = aom_highbd_d117_predictor_8x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_d117_predictor_8x8 = aom_highbd_d117_predictor_8x8_ssse3;
+    aom_highbd_d135_predictor_16x16 = aom_highbd_d135_predictor_16x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_d135_predictor_16x16 = aom_highbd_d135_predictor_16x16_ssse3;
+    aom_highbd_d135_predictor_32x32 = aom_highbd_d135_predictor_32x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_d135_predictor_32x32 = aom_highbd_d135_predictor_32x32_ssse3;
+    aom_highbd_d135_predictor_8x8 = aom_highbd_d135_predictor_8x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_d135_predictor_8x8 = aom_highbd_d135_predictor_8x8_ssse3;
+    aom_highbd_d153_predictor_16x16 = aom_highbd_d153_predictor_16x16_c;
+    if (flags & HAS_SSSE3) aom_highbd_d153_predictor_16x16 = aom_highbd_d153_predictor_16x16_ssse3;
+    aom_highbd_d153_predictor_32x32 = aom_highbd_d153_predictor_32x32_c;
+    if (flags & HAS_SSSE3) aom_highbd_d153_predictor_32x32 = aom_highbd_d153_predictor_32x32_ssse3;
+    aom_highbd_d153_predictor_8x8 = aom_highbd_d153_predictor_8x8_c;
+    if (flags & HAS_SSSE3) aom_highbd_d153_predictor_8x8 = aom_highbd_d153_predictor_8x8_ssse3;
+    aom_highbd_d45e_predictor_16x16 = aom_highbd_d45e_predictor_16x16_c;
+    if (flags & HAS_AVX2) aom_highbd_d45e_predictor_16x16 = aom_highbd_d45e_predictor_16x16_avx2;
+    aom_highbd_d45e_predictor_16x32 = aom_highbd_d45e_predictor_16x32_c;
+    if (flags & HAS_AVX2) aom_highbd_d45e_predictor_16x32 = aom_highbd_d45e_predictor_16x32_avx2;
+    aom_highbd_d45e_predictor_16x8 = aom_highbd_d45e_predictor_16x8_c;
+    if (flags & HAS_AVX2) aom_highbd_d45e_predictor_16x8 = aom_highbd_d45e_predictor_16x8_avx2;
+    aom_highbd_d45e_predictor_32x16 = aom_highbd_d45e_predictor_32x16_c;
+    if (flags & HAS_AVX2) aom_highbd_d45e_predictor_32x16 = aom_highbd_d45e_predictor_32x16_avx2;
+    aom_highbd_d45e_predictor_32x32 = aom_highbd_d45e_predictor_32x32_c;
+    if (flags & HAS_AVX2) aom_highbd_d45e_predictor_32x32 = aom_highbd_d45e_predictor_32x32_avx2;
+    aom_highbd_lpf_horizontal_4_dual = aom_highbd_lpf_horizontal_4_dual_sse2;
+    if (flags & HAS_AVX2) aom_highbd_lpf_horizontal_4_dual = aom_highbd_lpf_horizontal_4_dual_avx2;
+    aom_highbd_lpf_horizontal_8_dual = aom_highbd_lpf_horizontal_8_dual_sse2;
+    if (flags & HAS_AVX2) aom_highbd_lpf_horizontal_8_dual = aom_highbd_lpf_horizontal_8_dual_avx2;
+    aom_highbd_lpf_horizontal_edge_16 = aom_highbd_lpf_horizontal_edge_16_sse2;
+    if (flags & HAS_AVX2) aom_highbd_lpf_horizontal_edge_16 = aom_highbd_lpf_horizontal_edge_16_avx2;
+    aom_highbd_lpf_vertical_16_dual = aom_highbd_lpf_vertical_16_dual_sse2;
+    if (flags & HAS_AVX2) aom_highbd_lpf_vertical_16_dual = aom_highbd_lpf_vertical_16_dual_avx2;
+    aom_highbd_lpf_vertical_4_dual = aom_highbd_lpf_vertical_4_dual_sse2;
+    if (flags & HAS_AVX2) aom_highbd_lpf_vertical_4_dual = aom_highbd_lpf_vertical_4_dual_avx2;
+    aom_highbd_lpf_vertical_8_dual = aom_highbd_lpf_vertical_8_dual_sse2;
+    if (flags & HAS_AVX2) aom_highbd_lpf_vertical_8_dual = aom_highbd_lpf_vertical_8_dual_avx2;
     aom_highbd_masked_sad16x16 = aom_highbd_masked_sad16x16_c;
     if (flags & HAS_SSSE3) aom_highbd_masked_sad16x16 = aom_highbd_masked_sad16x16_ssse3;
     aom_highbd_masked_sad16x32 = aom_highbd_masked_sad16x32_c;
@@ -3979,10 +4235,6 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_SSSE3) aom_idct8x8_12_add = aom_idct8x8_12_add_ssse3;
     aom_idct8x8_64_add = aom_idct8x8_64_add_sse2;
     if (flags & HAS_SSSE3) aom_idct8x8_64_add = aom_idct8x8_64_add_ssse3;
-    aom_lpf_horizontal_edge_16 = aom_lpf_horizontal_edge_16_sse2;
-    if (flags & HAS_AVX2) aom_lpf_horizontal_edge_16 = aom_lpf_horizontal_edge_16_avx2;
-    aom_lpf_horizontal_edge_8 = aom_lpf_horizontal_edge_8_sse2;
-    if (flags & HAS_AVX2) aom_lpf_horizontal_edge_8 = aom_lpf_horizontal_edge_8_avx2;
     aom_masked_sad16x16 = aom_masked_sad16x16_c;
     if (flags & HAS_SSSE3) aom_masked_sad16x16 = aom_masked_sad16x16_ssse3;
     aom_masked_sad16x32 = aom_masked_sad16x32_c;
@@ -4089,6 +4341,31 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_SSE4_1) aom_obmc_variance8x4 = aom_obmc_variance8x4_sse4_1;
     aom_obmc_variance8x8 = aom_obmc_variance8x8_c;
     if (flags & HAS_SSE4_1) aom_obmc_variance8x8 = aom_obmc_variance8x8_sse4_1;
+    aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_avx2;
+    aom_paeth_predictor_16x32 = aom_paeth_predictor_16x32_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_16x32 = aom_paeth_predictor_16x32_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_16x32 = aom_paeth_predictor_16x32_avx2;
+    aom_paeth_predictor_16x8 = aom_paeth_predictor_16x8_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_16x8 = aom_paeth_predictor_16x8_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_16x8 = aom_paeth_predictor_16x8_avx2;
+    aom_paeth_predictor_32x16 = aom_paeth_predictor_32x16_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_32x16 = aom_paeth_predictor_32x16_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_32x16 = aom_paeth_predictor_32x16_avx2;
+    aom_paeth_predictor_32x32 = aom_paeth_predictor_32x32_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_32x32 = aom_paeth_predictor_32x32_ssse3;
+    if (flags & HAS_AVX2) aom_paeth_predictor_32x32 = aom_paeth_predictor_32x32_avx2;
+    aom_paeth_predictor_4x4 = aom_paeth_predictor_4x4_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_4x4 = aom_paeth_predictor_4x4_ssse3;
+    aom_paeth_predictor_4x8 = aom_paeth_predictor_4x8_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_4x8 = aom_paeth_predictor_4x8_ssse3;
+    aom_paeth_predictor_8x16 = aom_paeth_predictor_8x16_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_8x16 = aom_paeth_predictor_8x16_ssse3;
+    aom_paeth_predictor_8x4 = aom_paeth_predictor_8x4_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_8x4 = aom_paeth_predictor_8x4_ssse3;
+    aom_paeth_predictor_8x8 = aom_paeth_predictor_8x8_c;
+    if (flags & HAS_SSSE3) aom_paeth_predictor_8x8 = aom_paeth_predictor_8x8_ssse3;
     aom_quantize_b = aom_quantize_b_sse2;
     if (flags & HAS_SSSE3) aom_quantize_b = aom_quantize_b_ssse3;
     if (flags & HAS_AVX) aom_quantize_b = aom_quantize_b_avx;
@@ -4147,6 +4424,26 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_SSE4_1) aom_sad8x8x8 = aom_sad8x8x8_sse4_1;
     aom_scaled_2d = aom_scaled_2d_c;
     if (flags & HAS_SSSE3) aom_scaled_2d = aom_scaled_2d_ssse3;
+    aom_smooth_predictor_16x16 = aom_smooth_predictor_16x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_16x16 = aom_smooth_predictor_16x16_ssse3;
+    aom_smooth_predictor_16x32 = aom_smooth_predictor_16x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_16x32 = aom_smooth_predictor_16x32_ssse3;
+    aom_smooth_predictor_16x8 = aom_smooth_predictor_16x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_16x8 = aom_smooth_predictor_16x8_ssse3;
+    aom_smooth_predictor_32x16 = aom_smooth_predictor_32x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_32x16 = aom_smooth_predictor_32x16_ssse3;
+    aom_smooth_predictor_32x32 = aom_smooth_predictor_32x32_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_32x32 = aom_smooth_predictor_32x32_ssse3;
+    aom_smooth_predictor_4x4 = aom_smooth_predictor_4x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_4x4 = aom_smooth_predictor_4x4_ssse3;
+    aom_smooth_predictor_4x8 = aom_smooth_predictor_4x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_4x8 = aom_smooth_predictor_4x8_ssse3;
+    aom_smooth_predictor_8x16 = aom_smooth_predictor_8x16_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_8x16 = aom_smooth_predictor_8x16_ssse3;
+    aom_smooth_predictor_8x4 = aom_smooth_predictor_8x4_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_8x4 = aom_smooth_predictor_8x4_ssse3;
+    aom_smooth_predictor_8x8 = aom_smooth_predictor_8x8_c;
+    if (flags & HAS_SSSE3) aom_smooth_predictor_8x8 = aom_smooth_predictor_8x8_ssse3;
     aom_sub_pixel_avg_variance16x16 = aom_sub_pixel_avg_variance16x16_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_avg_variance16x16 = aom_sub_pixel_avg_variance16x16_ssse3;
     aom_sub_pixel_avg_variance16x32 = aom_sub_pixel_avg_variance16x32_sse2;
@@ -4203,6 +4500,10 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_SSSE3) aom_sub_pixel_variance8x4 = aom_sub_pixel_variance8x4_ssse3;
     aom_sub_pixel_variance8x8 = aom_sub_pixel_variance8x8_sse2;
     if (flags & HAS_SSSE3) aom_sub_pixel_variance8x8 = aom_sub_pixel_variance8x8_ssse3;
+    aom_v_predictor_32x16 = aom_v_predictor_32x16_sse2;
+    if (flags & HAS_AVX2) aom_v_predictor_32x16 = aom_v_predictor_32x16_avx2;
+    aom_v_predictor_32x32 = aom_v_predictor_32x32_sse2;
+    if (flags & HAS_AVX2) aom_v_predictor_32x32 = aom_v_predictor_32x32_avx2;
     aom_variance16x16 = aom_variance16x16_sse2;
     if (flags & HAS_AVX2) aom_variance16x16 = aom_variance16x16_avx2;
     aom_variance32x16 = aom_variance32x16_sse2;