Update aom to commit id e87fb2378f01103d5d6e477a4ef6892dc714e614

4 files changed, 694 insertions, 362 deletions

diff --git a/media/libaom/config/linux/x64/aom_config.asm b/media/libaom/config/linux/x64/aom_config.asm
index 2821f150f..455683b8f 100644
--- a/media/libaom/config/linux/x64/aom_config.asm
+++ b/media/libaom/config/linux/x64/aom_config.asm
@@ -46,8 +46,6 @@ CONFIG_AV1 equ 1
 CONFIG_STATIC_MSVCRT equ 0
 CONFIG_SPATIAL_RESAMPLING equ 1
 CONFIG_REALTIME_ONLY equ 0
-CONFIG_ONTHEFLY_BITPACKING equ 0
-CONFIG_ERROR_CONCEALMENT equ 0
 CONFIG_SHARED equ 0
 CONFIG_STATIC equ 1
 CONFIG_SMALL equ 0
@@ -60,73 +58,71 @@ CONFIG_ACCOUNTING equ 0
 CONFIG_INSPECTION equ 0
 CONFIG_DECODE_PERF_TESTS equ 0
 CONFIG_ENCODE_PERF_TESTS equ 0
+CONFIG_BITSTREAM_DEBUG equ 0
+CONFIG_SYMBOLRATE equ 0
 CONFIG_COEFFICIENT_RANGE_CHECKING equ 0
 CONFIG_LOWBITDEPTH equ 1
 CONFIG_HIGHBITDEPTH equ 1
 CONFIG_EXPERIMENTAL equ 0
 CONFIG_SIZE_LIMIT equ 1
-CONFIG_COLORSPACE_HEADERS equ 0
 CONFIG_FP_MB_STATS equ 0
 CONFIG_CDEF equ 1
+CONFIG_CDEF_SINGLEPASS equ 1
 CONFIG_VAR_TX equ 1
 CONFIG_RECT_TX equ 1
 CONFIG_RECT_TX_EXT equ 0
 CONFIG_TPL_MV equ 0
 CONFIG_DUAL_FILTER equ 1
-CONFIG_CONVOLVE_ROUND equ 0
+CONFIG_CONVOLVE_ROUND equ 1
 CONFIG_COMPOUND_ROUND equ 0
 CONFIG_EXT_TX equ 1
-CONFIG_DPCM_INTRA equ 0
 CONFIG_TX64X64 equ 0
 CONFIG_EXT_INTRA equ 1
 CONFIG_INTRA_INTERP equ 0
 CONFIG_FILTER_INTRA equ 0
-CONFIG_INTRA_EDGE equ 0
+CONFIG_INTRA_EDGE equ 1
 CONFIG_INTRABC equ 0
-CONFIG_EXT_INTER equ 1
 CONFIG_INTERINTRA equ 1
 CONFIG_WEDGE equ 1
 CONFIG_COMPOUND_SEGMENT equ 1
 CONFIG_EXT_REFS equ 1
-CONFIG_ALTREF2 equ 0
-CONFIG_SPEED_REFS equ 0
-CONFIG_GF_GROUPS equ 0
-CONFIG_FLEX_REFS equ 0
 CONFIG_GLOBAL_MOTION equ 1
 CONFIG_NEW_QUANT equ 0
 CONFIG_SUPERTX equ 0
 CONFIG_ANS equ 0
-CONFIG_LOOP_RESTORATION equ 0
+CONFIG_LOOP_RESTORATION equ 1
+CONFIG_STRIPED_LOOP_RESTORATION equ 0
 CONFIG_EXT_PARTITION equ 0
 CONFIG_EXT_PARTITION_TYPES equ 0
+CONFIG_EXT_PARTITION_TYPES_AB equ 0
 CONFIG_UNPOISON_PARTITION_CTX equ 0
 CONFIG_EXT_TILE equ 0
 CONFIG_MOTION_VAR equ 1
 CONFIG_NCOBMC equ 0
 CONFIG_WARPED_MOTION equ 1
 CONFIG_Q_ADAPT_PROBS equ 0
-CONFIG_BITSTREAM_DEBUG equ 0
 CONFIG_INTER_STATS_ONLY equ 0
-CONFIG_ALT_INTRA equ 1
-CONFIG_PALETTE equ 1
 CONFIG_PALETTE_DELTA_ENCODING equ 0
 CONFIG_RAWBITS equ 0
-CONFIG_EC_SMALLMUL equ 1
+CONFIG_KF_CTX equ 0
 CONFIG_PVQ equ 0
 CONFIG_CFL equ 0
 CONFIG_XIPHRC equ 0
 CONFIG_DCT_ONLY equ 0
+CONFIG_DAALA_TX equ 0
 CONFIG_DAALA_DCT4 equ 0
 CONFIG_DAALA_DCT8 equ 0
+CONFIG_DAALA_DCT16 equ 0
+CONFIG_DAALA_DCT32 equ 0
+CONFIG_DAALA_DCT64 equ 0
 CONFIG_CB4X4 equ 1
 CONFIG_CHROMA_2X2 equ 0
 CONFIG_CHROMA_SUB8X8 equ 1
 CONFIG_FRAME_SIZE equ 0
-CONFIG_DELTA_Q equ 1
 CONFIG_EXT_DELTA_Q equ 1
 CONFIG_ADAPT_SCAN equ 0
-CONFIG_FILTER_7BIT equ 1
 CONFIG_PARALLEL_DEBLOCKING equ 1
+CONFIG_DEBLOCK_13TAP equ 0
 CONFIG_LOOPFILTERING_ACROSS_TILES equ 1
 CONFIG_TEMPMV_SIGNALING equ 1
 CONFIG_RD_DEBUG equ 0
@@ -135,29 +131,46 @@ CONFIG_COEF_INTERLEAVE equ 0
 CONFIG_ENTROPY_STATS equ 0
 CONFIG_MASKED_TX equ 0
 CONFIG_DEPENDENT_HORZTILES equ 0
-CONFIG_DIST_8X8 equ 0
-CONFIG_DAALA_DIST equ 0
-CONFIG_TRIPRED equ 0
+CONFIG_DIST_8X8 equ 1
 CONFIG_PALETTE_THROUGHPUT equ 1
 CONFIG_REF_ADAPT equ 0
 CONFIG_LV_MAP equ 0
+CONFIG_CTX1D equ 0
 CONFIG_TXK_SEL equ 0
 CONFIG_MV_COMPRESS equ 1
+CONFIG_SEGMENT_ZEROMV equ 0
 CONFIG_FRAME_SUPERRES equ 0
 CONFIG_NEW_MULTISYMBOL equ 0
 CONFIG_COMPOUND_SINGLEREF equ 0
-CONFIG_AOM_QM equ 0
+CONFIG_AOM_QM equ 1
 CONFIG_ONE_SIDED_COMPOUND equ 1
-CONFIG_EXT_COMP_REFS equ 0
+CONFIG_EXT_COMP_REFS equ 1
 CONFIG_SMOOTH_HV equ 1
 CONFIG_VAR_REFS equ 0
-CONFIG_RECT_INTRA_PRED equ 1
 CONFIG_LGT equ 0
+CONFIG_LGT_FROM_PRED equ 0
 CONFIG_SBL_SYMBOL equ 0
 CONFIG_NCOBMC_ADAPT_WEIGHT equ 0
 CONFIG_BGSPRITE equ 0
 CONFIG_VAR_TX_NO_TX_MODE equ 0
 CONFIG_MRC_TX equ 0
 CONFIG_LPF_DIRECT equ 0
-CONFIG_UV_LVL equ 0
+CONFIG_LOOPFILTER_LEVEL equ 0
+CONFIG_NO_FRAME_CONTEXT_SIGNALING equ 0
+CONFIG_TXMG equ 1
+CONFIG_MAX_TILE equ 0
+CONFIG_HASH_ME equ 0
+CONFIG_COLORSPACE_HEADERS equ 0
+CONFIG_MFMV equ 0
+CONFIG_FRAME_MARKER equ 0
+CONFIG_JNT_COMP equ 0
+CONFIG_FRAME_SIGN_BIAS equ 0
+CONFIG_EXT_SKIP equ 0
+CONFIG_OBU equ 0
+CONFIG_AMVR equ 0
+CONFIG_LPF_SB equ 0
+CONFIG_OPT_REF_MV equ 0
+CONFIG_TMV equ 0
+CONFIG_RESTRICT_COMPRESSED_HDR equ 0
+CONFIG_HORZONLY_FRAME_SUPERRES equ 0
 CONFIG_ANALYZER equ 0
diff --git a/media/libaom/config/linux/x64/aom_config.h b/media/libaom/config/linux/x64/aom_config.h
index f9f2caadd..f3416f3dd 100644
--- a/media/libaom/config/linux/x64/aom_config.h
+++ b/media/libaom/config/linux/x64/aom_config.h
@@ -59,8 +59,6 @@
 #define CONFIG_STATIC_MSVCRT 0
 #define CONFIG_SPATIAL_RESAMPLING 1
 #define CONFIG_REALTIME_ONLY 0
-#define CONFIG_ONTHEFLY_BITPACKING 0
-#define CONFIG_ERROR_CONCEALMENT 0
 #define CONFIG_SHARED 0
 #define CONFIG_STATIC 1
 #define CONFIG_SMALL 0
@@ -73,73 +71,71 @@
 #define CONFIG_INSPECTION 0
 #define CONFIG_DECODE_PERF_TESTS 0
 #define CONFIG_ENCODE_PERF_TESTS 0
+#define CONFIG_BITSTREAM_DEBUG 0
+#define CONFIG_SYMBOLRATE 0
 #define CONFIG_COEFFICIENT_RANGE_CHECKING 0
 #define CONFIG_LOWBITDEPTH 1
 #define CONFIG_HIGHBITDEPTH 1
 #define CONFIG_EXPERIMENTAL 0
 #define CONFIG_SIZE_LIMIT 1
-#define CONFIG_COLORSPACE_HEADERS 0
 #define CONFIG_FP_MB_STATS 0
 #define CONFIG_CDEF 1
+#define CONFIG_CDEF_SINGLEPASS 1
 #define CONFIG_VAR_TX 1
 #define CONFIG_RECT_TX 1
 #define CONFIG_RECT_TX_EXT 0
 #define CONFIG_TPL_MV 0
 #define CONFIG_DUAL_FILTER 1
-#define CONFIG_CONVOLVE_ROUND 0
+#define CONFIG_CONVOLVE_ROUND 1
 #define CONFIG_COMPOUND_ROUND 0
 #define CONFIG_EXT_TX 1
-#define CONFIG_DPCM_INTRA 0
 #define CONFIG_TX64X64 0
 #define CONFIG_EXT_INTRA 1
 #define CONFIG_INTRA_INTERP 0
 #define CONFIG_FILTER_INTRA 0
-#define CONFIG_INTRA_EDGE 0
+#define CONFIG_INTRA_EDGE 1
 #define CONFIG_INTRABC 0
-#define CONFIG_EXT_INTER 1
 #define CONFIG_INTERINTRA 1
 #define CONFIG_WEDGE 1
 #define CONFIG_COMPOUND_SEGMENT 1
 #define CONFIG_EXT_REFS 1
-#define CONFIG_ALTREF2 0
-#define CONFIG_SPEED_REFS 0
-#define CONFIG_GF_GROUPS 0
-#define CONFIG_FLEX_REFS 0
 #define CONFIG_GLOBAL_MOTION 1
 #define CONFIG_NEW_QUANT 0
 #define CONFIG_SUPERTX 0
 #define CONFIG_ANS 0
-#define CONFIG_LOOP_RESTORATION 0
+#define CONFIG_LOOP_RESTORATION 1
+#define CONFIG_STRIPED_LOOP_RESTORATION 0
 #define CONFIG_EXT_PARTITION 0
 #define CONFIG_EXT_PARTITION_TYPES 0
+#define CONFIG_EXT_PARTITION_TYPES_AB 0
 #define CONFIG_UNPOISON_PARTITION_CTX 0
 #define CONFIG_EXT_TILE 0
 #define CONFIG_MOTION_VAR 1
 #define CONFIG_NCOBMC 0
 #define CONFIG_WARPED_MOTION 1
 #define CONFIG_Q_ADAPT_PROBS 0
-#define CONFIG_BITSTREAM_DEBUG 0
 #define CONFIG_INTER_STATS_ONLY 0
-#define CONFIG_ALT_INTRA 1
-#define CONFIG_PALETTE 1
 #define CONFIG_PALETTE_DELTA_ENCODING 0
 #define CONFIG_RAWBITS 0
-#define CONFIG_EC_SMALLMUL 1
+#define CONFIG_KF_CTX 0
 #define CONFIG_PVQ 0
 #define CONFIG_CFL 0
 #define CONFIG_XIPHRC 0
 #define CONFIG_DCT_ONLY 0
+#define CONFIG_DAALA_TX 0
 #define CONFIG_DAALA_DCT4 0
 #define CONFIG_DAALA_DCT8 0
+#define CONFIG_DAALA_DCT16 0
+#define CONFIG_DAALA_DCT32 0
+#define CONFIG_DAALA_DCT64 0
 #define CONFIG_CB4X4 1
 #define CONFIG_CHROMA_2X2 0
 #define CONFIG_CHROMA_SUB8X8 1
 #define CONFIG_FRAME_SIZE 0
-#define CONFIG_DELTA_Q 1
 #define CONFIG_EXT_DELTA_Q 1
 #define CONFIG_ADAPT_SCAN 0
-#define CONFIG_FILTER_7BIT 1
 #define CONFIG_PARALLEL_DEBLOCKING 1
+#define CONFIG_DEBLOCK_13TAP 0
 #define CONFIG_LOOPFILTERING_ACROSS_TILES 1
 #define CONFIG_TEMPMV_SIGNALING 1
 #define CONFIG_RD_DEBUG 0
@@ -148,31 +144,48 @@
 #define CONFIG_ENTROPY_STATS 0
 #define CONFIG_MASKED_TX 0
 #define CONFIG_DEPENDENT_HORZTILES 0
-#define CONFIG_DIST_8X8 0
-#define CONFIG_DAALA_DIST 0
-#define CONFIG_TRIPRED 0
+#define CONFIG_DIST_8X8 1
 #define CONFIG_PALETTE_THROUGHPUT 1
 #define CONFIG_REF_ADAPT 0
 #define CONFIG_LV_MAP 0
+#define CONFIG_CTX1D 0
 #define CONFIG_TXK_SEL 0
 #define CONFIG_MV_COMPRESS 1
+#define CONFIG_SEGMENT_ZEROMV 0
 #define CONFIG_FRAME_SUPERRES 0
 #define CONFIG_NEW_MULTISYMBOL 0
 #define CONFIG_COMPOUND_SINGLEREF 0
-#define CONFIG_AOM_QM 0
+#define CONFIG_AOM_QM 1
 #define CONFIG_ONE_SIDED_COMPOUND 1
-#define CONFIG_EXT_COMP_REFS 0
+#define CONFIG_EXT_COMP_REFS 1
 #define CONFIG_SMOOTH_HV 1
 #define CONFIG_VAR_REFS 0
-#define CONFIG_RECT_INTRA_PRED 1
 #define CONFIG_LGT 0
+#define CONFIG_LGT_FROM_PRED 0
 #define CONFIG_SBL_SYMBOL 0
 #define CONFIG_NCOBMC_ADAPT_WEIGHT 0
 #define CONFIG_BGSPRITE 0
 #define CONFIG_VAR_TX_NO_TX_MODE 0
 #define CONFIG_MRC_TX 0
 #define CONFIG_LPF_DIRECT 0
-#define CONFIG_UV_LVL 0
+#define CONFIG_LOOPFILTER_LEVEL 0
+#define CONFIG_NO_FRAME_CONTEXT_SIGNALING 0
+#define CONFIG_TXMG 1
+#define CONFIG_MAX_TILE 0
+#define CONFIG_HASH_ME 0
+#define CONFIG_COLORSPACE_HEADERS 0
+#define CONFIG_MFMV 0
+#define CONFIG_FRAME_MARKER 0
+#define CONFIG_JNT_COMP 0
+#define CONFIG_FRAME_SIGN_BIAS 0
+#define CONFIG_EXT_SKIP 0
+#define CONFIG_OBU 0
+#define CONFIG_AMVR 0
+#define CONFIG_LPF_SB 0
+#define CONFIG_OPT_REF_MV 0
+#define CONFIG_TMV 0
+#define CONFIG_RESTRICT_COMPRESSED_HDR 0
+#define CONFIG_HORZONLY_FRAME_SUPERRES 0
 #define CONFIG_ANALYZER 0
 #define DECODE_WIDTH_LIMIT 8192
 #define DECODE_HEIGHT_LIMIT 4608
diff --git a/media/libaom/config/linux/x64/aom_dsp_rtcd.h b/media/libaom/config/linux/x64/aom_dsp_rtcd.h
index 6ee856f1d..d2bee385f 100644
--- a/media/libaom/config/linux/x64/aom_dsp_rtcd.h
+++ b/media/libaom/config/linux/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;
diff --git a/media/libaom/config/linux/x64/av1_rtcd.h b/media/libaom/config/linux/x64/av1_rtcd.h
index d4fc99c10..66c0349aa 100644
--- a/media/libaom/config/linux/x64/av1_rtcd.h
+++ b/media/libaom/config/linux/x64/av1_rtcd.h
@@ -31,44 +31,39 @@ struct search_site_config;
 struct mv;
 union int_mv;
 struct yv12_buffer_config;
-typedef uint16_t od_dering_in;
 
 #ifdef __cplusplus
 extern "C" {
 #endif
 
-void aom_clpf_block_c(uint8_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
-void aom_clpf_block_sse2(uint8_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
-void aom_clpf_block_ssse3(uint8_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
-void aom_clpf_block_sse4_1(uint8_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
-RTCD_EXTERN void (*aom_clpf_block)(uint8_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
-
-void aom_clpf_block_hbd_c(uint16_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
-void aom_clpf_block_hbd_sse2(uint16_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
-void aom_clpf_block_hbd_ssse3(uint16_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
-void aom_clpf_block_hbd_sse4_1(uint16_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
-RTCD_EXTERN void (*aom_clpf_block_hbd)(uint16_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
-
-void aom_clpf_hblock_c(uint8_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
-void aom_clpf_hblock_sse2(uint8_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
-void aom_clpf_hblock_ssse3(uint8_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
-void aom_clpf_hblock_sse4_1(uint8_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
-RTCD_EXTERN void (*aom_clpf_hblock)(uint8_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
-
-void aom_clpf_hblock_hbd_c(uint16_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
-void aom_clpf_hblock_hbd_sse2(uint16_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
-void aom_clpf_hblock_hbd_ssse3(uint16_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
-void aom_clpf_hblock_hbd_sse4_1(uint16_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
-RTCD_EXTERN void (*aom_clpf_hblock_hbd)(uint16_t *dst, const uint16_t *src, int dstride, int sstride, int sizex, int sizey, unsigned int strength, unsigned int bd);
+void apply_selfguided_restoration_c(uint8_t *dat, int width, int height, int stride, int eps, int *xqd, uint8_t *dst, int dst_stride, int32_t *tmpbuf);
+void apply_selfguided_restoration_sse4_1(uint8_t *dat, int width, int height, int stride, int eps, int *xqd, uint8_t *dst, int dst_stride, int32_t *tmpbuf);
+RTCD_EXTERN void (*apply_selfguided_restoration)(uint8_t *dat, int width, int height, int stride, int eps, int *xqd, uint8_t *dst, int dst_stride, int32_t *tmpbuf);
+
+void apply_selfguided_restoration_highbd_c(uint16_t *dat, int width, int height, int stride, int bit_depth, int eps, int *xqd, uint16_t *dst, int dst_stride, int32_t *tmpbuf);
+void apply_selfguided_restoration_highbd_sse4_1(uint16_t *dat, int width, int height, int stride, int bit_depth, int eps, int *xqd, uint16_t *dst, int dst_stride, int32_t *tmpbuf);
+RTCD_EXTERN void (*apply_selfguided_restoration_highbd)(uint16_t *dat, int width, int height, int stride, int bit_depth, int eps, int *xqd, uint16_t *dst, int dst_stride, int32_t *tmpbuf);
 
 int64_t av1_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
 int64_t av1_block_error_avx2(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
 RTCD_EXTERN int64_t (*av1_block_error)(const tran_low_t *coeff, const tran_low_t *dqcoeff, intptr_t block_size, int64_t *ssz);
 
+void av1_convolve_2d_c(const uint8_t *src, int src_stride, CONV_BUF_TYPE *dst, int dst_stride, int w, int h, InterpFilterParams *filter_params_x, InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+void av1_convolve_2d_sse2(const uint8_t *src, int src_stride, CONV_BUF_TYPE *dst, int dst_stride, int w, int h, InterpFilterParams *filter_params_x, InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
+#define av1_convolve_2d av1_convolve_2d_sse2
+
+void av1_convolve_2d_scale_c(const uint8_t *src, int src_stride, CONV_BUF_TYPE *dst, int dst_stride, int w, int h, InterpFilterParams *filter_params_x, InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_q4, const int y_step_qn, ConvolveParams *conv_params);
+void av1_convolve_2d_scale_sse4_1(const uint8_t *src, int src_stride, CONV_BUF_TYPE *dst, int dst_stride, int w, int h, InterpFilterParams *filter_params_x, InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_q4, const int y_step_qn, ConvolveParams *conv_params);
+RTCD_EXTERN void (*av1_convolve_2d_scale)(const uint8_t *src, int src_stride, CONV_BUF_TYPE *dst, int dst_stride, int w, int h, InterpFilterParams *filter_params_x, InterpFilterParams *filter_params_y, const int subpel_x_qn, const int x_step_qn, const int subpel_y_q4, const int y_step_qn, ConvolveParams *conv_params);
+
 void av1_convolve_horiz_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams fp, const int subpel_x_q4, int x_step_q4, ConvolveParams *conv_params);
 void av1_convolve_horiz_ssse3(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams fp, const int subpel_x_q4, int x_step_q4, ConvolveParams *conv_params);
 RTCD_EXTERN void (*av1_convolve_horiz)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams fp, const int subpel_x_q4, int x_step_q4, ConvolveParams *conv_params);
 
+void av1_convolve_rounding_c(const int32_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, int bits);
+void av1_convolve_rounding_avx2(const int32_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, int bits);
+RTCD_EXTERN void (*av1_convolve_rounding)(const int32_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, int bits);
+
 void av1_convolve_vert_c(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams fp, const int subpel_x_q4, int x_step_q4, ConvolveParams *conv_params);
 void av1_convolve_vert_ssse3(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams fp, const int subpel_x_q4, int x_step_q4, ConvolveParams *conv_params);
 RTCD_EXTERN void (*av1_convolve_vert)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams fp, const int subpel_x_q4, int x_step_q4, ConvolveParams *conv_params);
@@ -76,9 +71,6 @@ RTCD_EXTERN void (*av1_convolve_vert)(const uint8_t *src, int src_stride, uint8_
 int av1_diamond_search_sad_c(struct macroblock *x, const struct search_site_config *cfg,  struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct aom_variance_vtable *fn_ptr, const struct mv *center_mv);
 #define av1_diamond_search_sad av1_diamond_search_sad_c
 
-void av1_fdct8x8_quant_c(const int16_t *input, int stride, 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);
-#define av1_fdct8x8_quant av1_fdct8x8_quant_c
-
 void av1_fht16x16_c(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
 void av1_fht16x16_sse2(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
 void av1_fht16x16_avx2(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
@@ -133,6 +125,14 @@ void av1_fht8x8_c(const int16_t *input, tran_low_t *output, int stride, struct t
 void av1_fht8x8_sse2(const int16_t *input, tran_low_t *output, int stride, struct txfm_param *param);
 #define av1_fht8x8 av1_fht8x8_sse2
 
+void av1_filter_intra_edge_c(uint8_t *p, int sz, int strength);
+void av1_filter_intra_edge_sse4_1(uint8_t *p, int sz, int strength);
+RTCD_EXTERN void (*av1_filter_intra_edge)(uint8_t *p, int sz, int strength);
+
+void av1_filter_intra_edge_high_c(uint16_t *p, int sz, int strength);
+void av1_filter_intra_edge_high_sse4_1(uint16_t *p, int sz, int strength);
+RTCD_EXTERN void (*av1_filter_intra_edge_high)(uint16_t *p, int sz, int strength);
+
 int av1_full_range_search_c(const struct macroblock *x, const struct search_site_config *cfg, struct mv *ref_mv, struct mv *best_mv, int search_param, int sad_per_bit, int *num00, const struct aom_variance_vtable *fn_ptr, const struct mv *center_mv);
 #define av1_full_range_search av1_full_range_search_c
 
@@ -141,46 +141,42 @@ int av1_full_search_sadx3(const struct macroblock *x, const struct mv *ref_mv, i
 int av1_full_search_sadx8(const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct aom_variance_vtable *fn_ptr, const struct mv *center_mv, struct mv *best_mv);
 RTCD_EXTERN int (*av1_full_search_sad)(const struct macroblock *x, const struct mv *ref_mv, int sad_per_bit, int distance, const struct aom_variance_vtable *fn_ptr, const struct mv *center_mv, struct mv *best_mv);
 
-void av1_fwd_idtx_c(const int16_t *src_diff, tran_low_t *coeff, int stride, int bs, int tx_type);
+void av1_fwd_idtx_c(const int16_t *src_diff, tran_low_t *coeff, int stride, int bsx, int bsy, TX_TYPE tx_type);
 #define av1_fwd_idtx av1_fwd_idtx_c
 
-void av1_fwd_txfm2d_16x16_c(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
-void av1_fwd_txfm2d_16x16_sse4_1(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
-RTCD_EXTERN void (*av1_fwd_txfm2d_16x16)(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
+void av1_fwd_txfm2d_16x16_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_16x16_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_16x16)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
 
-void av1_fwd_txfm2d_16x32_c(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
+void av1_fwd_txfm2d_16x32_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
 #define av1_fwd_txfm2d_16x32 av1_fwd_txfm2d_16x32_c
 
-void av1_fwd_txfm2d_16x8_c(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
+void av1_fwd_txfm2d_16x8_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
 #define av1_fwd_txfm2d_16x8 av1_fwd_txfm2d_16x8_c
 
-void av1_fwd_txfm2d_32x16_c(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
+void av1_fwd_txfm2d_32x16_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
 #define av1_fwd_txfm2d_32x16 av1_fwd_txfm2d_32x16_c
 
-void av1_fwd_txfm2d_32x32_c(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
-void av1_fwd_txfm2d_32x32_sse4_1(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
-RTCD_EXTERN void (*av1_fwd_txfm2d_32x32)(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
+void av1_fwd_txfm2d_32x32_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_32x32_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_32x32)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
 
-void av1_fwd_txfm2d_4x4_c(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
-void av1_fwd_txfm2d_4x4_sse4_1(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
-RTCD_EXTERN void (*av1_fwd_txfm2d_4x4)(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
+void av1_fwd_txfm2d_4x4_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_4x4_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_4x4)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
 
-void av1_fwd_txfm2d_4x8_c(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
+void av1_fwd_txfm2d_4x8_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
 #define av1_fwd_txfm2d_4x8 av1_fwd_txfm2d_4x8_c
 
-void av1_fwd_txfm2d_64x64_c(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
-void av1_fwd_txfm2d_64x64_sse4_1(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
-RTCD_EXTERN void (*av1_fwd_txfm2d_64x64)(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
-
-void av1_fwd_txfm2d_8x16_c(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
+void av1_fwd_txfm2d_8x16_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
 #define av1_fwd_txfm2d_8x16 av1_fwd_txfm2d_8x16_c
 
-void av1_fwd_txfm2d_8x4_c(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
+void av1_fwd_txfm2d_8x4_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
 #define av1_fwd_txfm2d_8x4 av1_fwd_txfm2d_8x4_c
 
-void av1_fwd_txfm2d_8x8_c(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
-void av1_fwd_txfm2d_8x8_sse4_1(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
-RTCD_EXTERN void (*av1_fwd_txfm2d_8x8)(const int16_t *input, int32_t *output, int stride, int tx_type, int bd);
+void av1_fwd_txfm2d_8x8_c(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_fwd_txfm2d_8x8_sse4_1(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_fwd_txfm2d_8x8)(const int16_t *input, int32_t *output, int stride, TX_TYPE tx_type, int bd);
 
 void av1_fwht4x4_c(const int16_t *input, tran_low_t *output, int stride);
 #define av1_fwht4x4 av1_fwht4x4_c
@@ -213,6 +209,14 @@ void av1_highbd_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride, uint8
 void av1_highbd_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
 #define av1_highbd_convolve8_vert av1_highbd_convolve8_vert_sse2
 
+void av1_highbd_convolve_2d_c(const uint16_t *src, int src_stride, CONV_BUF_TYPE *dst, int dst_stride, int w, int h, InterpFilterParams *filter_params_x, InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
+void av1_highbd_convolve_2d_ssse3(const uint16_t *src, int src_stride, CONV_BUF_TYPE *dst, int dst_stride, int w, int h, InterpFilterParams *filter_params_x, InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_convolve_2d)(const uint16_t *src, int src_stride, CONV_BUF_TYPE *dst, int dst_stride, int w, int h, InterpFilterParams *filter_params_x, InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params, int bd);
+
+void av1_highbd_convolve_2d_scale_c(const uint16_t *src, int src_stride, CONV_BUF_TYPE *dst, int dst_stride, int w, int h, InterpFilterParams *filter_params_x, InterpFilterParams *filter_params_y, const int subpel_x_q4, const int x_step_qn, const int subpel_y_q4, const int y_step_qn, ConvolveParams *conv_params, int bd);
+void av1_highbd_convolve_2d_scale_sse4_1(const uint16_t *src, int src_stride, CONV_BUF_TYPE *dst, int dst_stride, int w, int h, InterpFilterParams *filter_params_x, InterpFilterParams *filter_params_y, const int subpel_x_q4, const int x_step_qn, const int subpel_y_q4, const int y_step_qn, ConvolveParams *conv_params, int bd);
+RTCD_EXTERN void (*av1_highbd_convolve_2d_scale)(const uint16_t *src, int src_stride, CONV_BUF_TYPE *dst, int dst_stride, int w, int h, InterpFilterParams *filter_params_x, InterpFilterParams *filter_params_y, const int subpel_x_q4, const int x_step_qn, const int subpel_y_q4, const int y_step_qn, ConvolveParams *conv_params, int bd);
+
 void av1_highbd_convolve_avg_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps);
 #define av1_highbd_convolve_avg av1_highbd_convolve_avg_c
 
@@ -227,6 +231,10 @@ void av1_highbd_convolve_init_c(void);
 void av1_highbd_convolve_init_sse4_1(void);
 RTCD_EXTERN void (*av1_highbd_convolve_init)(void);
 
+void av1_highbd_convolve_rounding_c(const int32_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, int bits, int bd);
+void av1_highbd_convolve_rounding_avx2(const int32_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, int bits, int bd);
+RTCD_EXTERN void (*av1_highbd_convolve_rounding)(const int32_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, int bits, int bd);
+
 void av1_highbd_convolve_vert_c(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams fp, const int subpel_x_q4, int x_step_q4, int avg, int bd);
 void av1_highbd_convolve_vert_sse4_1(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams fp, const int subpel_x_q4, int x_step_q4, int avg, int bd);
 RTCD_EXTERN void (*av1_highbd_convolve_vert)(const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w, int h, const InterpFilterParams fp, const int subpel_x_q4, int x_step_q4, int avg, int bd);
@@ -273,9 +281,6 @@ void av1_highbd_iht8x4_32_add_c(const tran_low_t *input, uint8_t *dest, int dest
 void av1_highbd_iht8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_highbd_iht8x8_64_add av1_highbd_iht8x8_64_add_c
 
-void av1_highbd_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, int log_scale);
-#define av1_highbd_quantize_b av1_highbd_quantize_b_c
-
 void av1_highbd_quantize_fp_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, int log_scale);
 void av1_highbd_quantize_fp_sse4_1(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, int log_scale);
 void av1_highbd_quantize_fp_avx2(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, int log_scale);
@@ -288,6 +293,14 @@ void av1_highbd_warp_affine_c(const int32_t *mat, const uint16_t *ref, int width
 void av1_highbd_warp_affine_ssse3(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
 RTCD_EXTERN void (*av1_highbd_warp_affine)(const int32_t *mat, const uint16_t *ref, int width, int height, int stride, uint16_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, int bd, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
 
+void av1_highpass_filter_c(uint8_t *dgd, int width, int height, int stride, int32_t *dst, int dst_stride, int r, int eps);
+void av1_highpass_filter_sse4_1(uint8_t *dgd, int width, int height, int stride, int32_t *dst, int dst_stride, int r, int eps);
+RTCD_EXTERN void (*av1_highpass_filter)(uint8_t *dgd, int width, int height, int stride, int32_t *dst, int dst_stride, int r, int eps);
+
+void av1_highpass_filter_highbd_c(uint16_t *dgd, int width, int height, int stride, int32_t *dst, int dst_stride, int r, int eps);
+void av1_highpass_filter_highbd_sse4_1(uint16_t *dgd, int width, int height, int stride, int32_t *dst, int dst_stride, int r, int eps);
+RTCD_EXTERN void (*av1_highpass_filter_highbd)(uint16_t *dgd, int width, int height, int stride, int32_t *dst, int dst_stride, int r, int eps);
+
 void av1_iht16x16_256_add_c(const tran_low_t *input, uint8_t *output, int pitch, const struct txfm_param *param);
 void av1_iht16x16_256_add_sse2(const tran_low_t *input, uint8_t *output, int pitch, const struct txfm_param *param);
 void av1_iht16x16_256_add_avx2(const tran_low_t *input, uint8_t *output, int pitch, const struct txfm_param *param);
@@ -340,48 +353,45 @@ void av1_iht8x8_64_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride
 void av1_iht8x8_64_add_sse2(const tran_low_t *input, uint8_t *dest, int dest_stride, const struct txfm_param *param);
 #define av1_iht8x8_64_add av1_iht8x8_64_add_sse2
 
-void av1_inv_txfm2d_add_16x16_c(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
-void av1_inv_txfm2d_add_16x16_sse4_1(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
-RTCD_EXTERN void (*av1_inv_txfm2d_add_16x16)(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
+void av1_inv_txfm2d_add_16x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_inv_txfm2d_add_16x16_sse4_1(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_inv_txfm2d_add_16x16)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
 
-void av1_inv_txfm2d_add_16x32_c(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
+void av1_inv_txfm2d_add_16x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
 #define av1_inv_txfm2d_add_16x32 av1_inv_txfm2d_add_16x32_c
 
-void av1_inv_txfm2d_add_16x8_c(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
+void av1_inv_txfm2d_add_16x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
 #define av1_inv_txfm2d_add_16x8 av1_inv_txfm2d_add_16x8_c
 
-void av1_inv_txfm2d_add_32x16_c(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
+void av1_inv_txfm2d_add_32x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
 #define av1_inv_txfm2d_add_32x16 av1_inv_txfm2d_add_32x16_c
 
-void av1_inv_txfm2d_add_32x32_c(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
-void av1_inv_txfm2d_add_32x32_avx2(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
-RTCD_EXTERN void (*av1_inv_txfm2d_add_32x32)(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
+void av1_inv_txfm2d_add_32x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_inv_txfm2d_add_32x32_avx2(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_inv_txfm2d_add_32x32)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
 
-void av1_inv_txfm2d_add_4x4_c(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
-void av1_inv_txfm2d_add_4x4_sse4_1(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
-RTCD_EXTERN void (*av1_inv_txfm2d_add_4x4)(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
+void av1_inv_txfm2d_add_4x4_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_inv_txfm2d_add_4x4_sse4_1(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_inv_txfm2d_add_4x4)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
 
-void av1_inv_txfm2d_add_4x8_c(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
+void av1_inv_txfm2d_add_4x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
 #define av1_inv_txfm2d_add_4x8 av1_inv_txfm2d_add_4x8_c
 
-void av1_inv_txfm2d_add_64x64_c(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
-#define av1_inv_txfm2d_add_64x64 av1_inv_txfm2d_add_64x64_c
-
-void av1_inv_txfm2d_add_8x16_c(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
+void av1_inv_txfm2d_add_8x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
 #define av1_inv_txfm2d_add_8x16 av1_inv_txfm2d_add_8x16_c
 
-void av1_inv_txfm2d_add_8x4_c(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
+void av1_inv_txfm2d_add_8x4_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
 #define av1_inv_txfm2d_add_8x4 av1_inv_txfm2d_add_8x4_c
 
-void av1_inv_txfm2d_add_8x8_c(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
-void av1_inv_txfm2d_add_8x8_sse4_1(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
-RTCD_EXTERN void (*av1_inv_txfm2d_add_8x8)(const int32_t *input, uint16_t *output, int stride, int tx_type, int bd);
+void av1_inv_txfm2d_add_8x8_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+void av1_inv_txfm2d_add_8x8_sse4_1(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+RTCD_EXTERN void (*av1_inv_txfm2d_add_8x8)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
 
 void av1_lowbd_convolve_init_c(void);
 void av1_lowbd_convolve_init_ssse3(void);
 RTCD_EXTERN void (*av1_lowbd_convolve_init)(void);
 
-void av1_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, int log_scale);
+void av1_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, const qm_val_t * qm_ptr, const qm_val_t * iqm_ptr, int log_scale);
 #define av1_quantize_b av1_quantize_b_c
 
 void av1_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, 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);
@@ -393,10 +403,26 @@ void av1_quantize_fp_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs, int
 void av1_quantize_fp_32x32_avx2(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);
 RTCD_EXTERN void (*av1_quantize_fp_32x32)(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 av1_selfguided_restoration_c(uint8_t *dgd, int width, int height, int stride, int32_t *dst, int dst_stride, int r, int eps);
+void av1_selfguided_restoration_sse4_1(uint8_t *dgd, int width, int height, int stride, int32_t *dst, int dst_stride, int r, int eps);
+RTCD_EXTERN void (*av1_selfguided_restoration)(uint8_t *dgd, int width, int height, int stride, int32_t *dst, int dst_stride, int r, int eps);
+
+void av1_selfguided_restoration_highbd_c(uint16_t *dgd, int width, int height, int stride, int32_t *dst, int dst_stride, int bit_depth, int r, int eps);
+void av1_selfguided_restoration_highbd_sse4_1(uint16_t *dgd, int width, int height, int stride, int32_t *dst, int dst_stride, int bit_depth, int r, int eps);
+RTCD_EXTERN void (*av1_selfguided_restoration_highbd)(uint16_t *dgd, int width, int height, int stride, int32_t *dst, int dst_stride, int bit_depth, int r, int eps);
+
 void av1_temporal_filter_apply_c(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
 void av1_temporal_filter_apply_sse2(uint8_t *frame1, unsigned int stride, uint8_t *frame2, unsigned int block_width, unsigned int block_height, int strength, int filter_weight, unsigned int *accumulator, uint16_t *count);
 #define av1_temporal_filter_apply av1_temporal_filter_apply_sse2
 
+void av1_upsample_intra_edge_c(uint8_t *p, int sz);
+void av1_upsample_intra_edge_sse4_1(uint8_t *p, int sz);
+RTCD_EXTERN void (*av1_upsample_intra_edge)(uint8_t *p, int sz);
+
+void av1_upsample_intra_edge_high_c(uint16_t *p, int sz, int bd);
+void av1_upsample_intra_edge_high_sse4_1(uint16_t *p, int sz, int bd);
+RTCD_EXTERN void (*av1_upsample_intra_edge_high)(uint16_t *p, int sz, int bd);
+
 void av1_warp_affine_c(const int32_t *mat, const uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
 void av1_warp_affine_sse2(const int32_t *mat, const uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
 void av1_warp_affine_ssse3(const int32_t *mat, const uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
@@ -414,64 +440,38 @@ uint64_t av1_wedge_sse_from_residuals_c(const int16_t *r1, const int16_t *d, con
 uint64_t av1_wedge_sse_from_residuals_sse2(const int16_t *r1, const int16_t *d, const uint8_t *m, int N);
 #define av1_wedge_sse_from_residuals av1_wedge_sse_from_residuals_sse2
 
+void cdef_filter_block_c(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int max);
+void cdef_filter_block_sse2(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int max);
+void cdef_filter_block_ssse3(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int max);
+void cdef_filter_block_sse4_1(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int max);
+void cdef_filter_block_avx2(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int max);
+RTCD_EXTERN void (*cdef_filter_block)(uint8_t *dst8, uint16_t *dst16, int dstride, const uint16_t *in, int pri_strength, int sec_strength, int dir, int pri_damping, int sec_damping, int bsize, int max);
+
+int cdef_find_dir_c(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
+int cdef_find_dir_sse2(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
+int cdef_find_dir_ssse3(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
+int cdef_find_dir_sse4_1(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
+int cdef_find_dir_avx2(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
+RTCD_EXTERN int (*cdef_find_dir)(const uint16_t *img, int stride, int32_t *var, int coeff_shift);
+
 double compute_cross_correlation_c(unsigned char *im1, int stride1, int x1, int y1, unsigned char *im2, int stride2, int x2, int y2);
 double compute_cross_correlation_sse4_1(unsigned char *im1, int stride1, int x1, int y1, unsigned char *im2, int stride2, int x2, int y2);
 RTCD_EXTERN double (*compute_cross_correlation)(unsigned char *im1, int stride1, int x1, int y1, unsigned char *im2, int stride2, int x2, int y2);
 
-void copy_4x4_16bit_to_16bit_c(uint16_t *dst, int dstride, const uint16_t *src, int sstride);
-void copy_4x4_16bit_to_16bit_sse2(uint16_t *dst, int dstride, const uint16_t *src, int sstride);
-void copy_4x4_16bit_to_16bit_ssse3(uint16_t *dst, int dstride, const uint16_t *src, int sstride);
-void copy_4x4_16bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint16_t *src, int sstride);
-RTCD_EXTERN void (*copy_4x4_16bit_to_16bit)(uint16_t *dst, int dstride, const uint16_t *src, int sstride);
-
-void copy_4x4_16bit_to_8bit_c(uint8_t *dst, int dstride, const uint16_t *src, int sstride);
-void copy_4x4_16bit_to_8bit_sse2(uint8_t *dst, int dstride, const uint16_t *src, int sstride);
-void copy_4x4_16bit_to_8bit_ssse3(uint8_t *dst, int dstride, const uint16_t *src, int sstride);
-void copy_4x4_16bit_to_8bit_sse4_1(uint8_t *dst, int dstride, const uint16_t *src, int sstride);
-RTCD_EXTERN void (*copy_4x4_16bit_to_8bit)(uint8_t *dst, int dstride, const uint16_t *src, int sstride);
-
-void copy_8x8_16bit_to_16bit_c(uint16_t *dst, int dstride, const uint16_t *src, int sstride);
-void copy_8x8_16bit_to_16bit_sse2(uint16_t *dst, int dstride, const uint16_t *src, int sstride);
-void copy_8x8_16bit_to_16bit_ssse3(uint16_t *dst, int dstride, const uint16_t *src, int sstride);
-void copy_8x8_16bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint16_t *src, int sstride);
-RTCD_EXTERN void (*copy_8x8_16bit_to_16bit)(uint16_t *dst, int dstride, const uint16_t *src, int sstride);
-
-void copy_8x8_16bit_to_8bit_c(uint8_t *dst, int dstride, const uint16_t *src, int sstride);
-void copy_8x8_16bit_to_8bit_sse2(uint8_t *dst, int dstride, const uint16_t *src, int sstride);
-void copy_8x8_16bit_to_8bit_ssse3(uint8_t *dst, int dstride, const uint16_t *src, int sstride);
-void copy_8x8_16bit_to_8bit_sse4_1(uint8_t *dst, int dstride, const uint16_t *src, int sstride);
-RTCD_EXTERN void (*copy_8x8_16bit_to_8bit)(uint8_t *dst, int dstride, const uint16_t *src, int sstride);
-
 void copy_rect8_16bit_to_16bit_c(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int v, int h);
 void copy_rect8_16bit_to_16bit_sse2(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int v, int h);
 void copy_rect8_16bit_to_16bit_ssse3(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int v, int h);
 void copy_rect8_16bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int v, int h);
+void copy_rect8_16bit_to_16bit_avx2(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int v, int h);
 RTCD_EXTERN void (*copy_rect8_16bit_to_16bit)(uint16_t *dst, int dstride, const uint16_t *src, int sstride, int v, int h);
 
 void copy_rect8_8bit_to_16bit_c(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h);
 void copy_rect8_8bit_to_16bit_sse2(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h);
 void copy_rect8_8bit_to_16bit_ssse3(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h);
 void copy_rect8_8bit_to_16bit_sse4_1(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h);
+void copy_rect8_8bit_to_16bit_avx2(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h);
 RTCD_EXTERN void (*copy_rect8_8bit_to_16bit)(uint16_t *dst, int dstride, const uint8_t *src, int sstride, int v, int h);
 
-int od_dir_find8_c(const od_dering_in *img, int stride, int32_t *var, int coeff_shift);
-int od_dir_find8_sse2(const od_dering_in *img, int stride, int32_t *var, int coeff_shift);
-int od_dir_find8_ssse3(const od_dering_in *img, int stride, int32_t *var, int coeff_shift);
-int od_dir_find8_sse4_1(const od_dering_in *img, int stride, int32_t *var, int coeff_shift);
-RTCD_EXTERN int (*od_dir_find8)(const od_dering_in *img, int stride, int32_t *var, int coeff_shift);
-
-void od_filter_dering_direction_4x4_c(uint16_t *y, int ystride, const uint16_t *in, int threshold, int dir, int damping);
-void od_filter_dering_direction_4x4_sse2(uint16_t *y, int ystride, const uint16_t *in, int threshold, int dir, int damping);
-void od_filter_dering_direction_4x4_ssse3(uint16_t *y, int ystride, const uint16_t *in, int threshold, int dir, int damping);
-void od_filter_dering_direction_4x4_sse4_1(uint16_t *y, int ystride, const uint16_t *in, int threshold, int dir, int damping);
-RTCD_EXTERN void (*od_filter_dering_direction_4x4)(uint16_t *y, int ystride, const uint16_t *in, int threshold, int dir, int damping);
-
-void od_filter_dering_direction_8x8_c(uint16_t *y, int ystride, const uint16_t *in, int threshold, int dir, int damping);
-void od_filter_dering_direction_8x8_sse2(uint16_t *y, int ystride, const uint16_t *in, int threshold, int dir, int damping);
-void od_filter_dering_direction_8x8_ssse3(uint16_t *y, int ystride, const uint16_t *in, int threshold, int dir, int damping);
-void od_filter_dering_direction_8x8_sse4_1(uint16_t *y, int ystride, const uint16_t *in, int threshold, int dir, int damping);
-RTCD_EXTERN void (*od_filter_dering_direction_8x8)(uint16_t *y, int ystride, const uint16_t *in, int threshold, int dir, int damping);
-
 void aom_rtcd(void);
 
 #ifdef RTCD_C
@@ -482,28 +482,28 @@ static void setup_rtcd_internal(void)
 
     (void)flags;
 
-    aom_clpf_block = aom_clpf_block_sse2;
-    if (flags & HAS_SSSE3) aom_clpf_block = aom_clpf_block_ssse3;
-    if (flags & HAS_SSE4_1) aom_clpf_block = aom_clpf_block_sse4_1;
-    aom_clpf_block_hbd = aom_clpf_block_hbd_sse2;
-    if (flags & HAS_SSSE3) aom_clpf_block_hbd = aom_clpf_block_hbd_ssse3;
-    if (flags & HAS_SSE4_1) aom_clpf_block_hbd = aom_clpf_block_hbd_sse4_1;
-    aom_clpf_hblock = aom_clpf_hblock_sse2;
-    if (flags & HAS_SSSE3) aom_clpf_hblock = aom_clpf_hblock_ssse3;
-    if (flags & HAS_SSE4_1) aom_clpf_hblock = aom_clpf_hblock_sse4_1;
-    aom_clpf_hblock_hbd = aom_clpf_hblock_hbd_sse2;
-    if (flags & HAS_SSSE3) aom_clpf_hblock_hbd = aom_clpf_hblock_hbd_ssse3;
-    if (flags & HAS_SSE4_1) aom_clpf_hblock_hbd = aom_clpf_hblock_hbd_sse4_1;
+    apply_selfguided_restoration = apply_selfguided_restoration_c;
+    if (flags & HAS_SSE4_1) apply_selfguided_restoration = apply_selfguided_restoration_sse4_1;
+    apply_selfguided_restoration_highbd = apply_selfguided_restoration_highbd_c;
+    if (flags & HAS_SSE4_1) apply_selfguided_restoration_highbd = apply_selfguided_restoration_highbd_sse4_1;
     av1_block_error = av1_block_error_c;
     if (flags & HAS_AVX2) av1_block_error = av1_block_error_avx2;
+    av1_convolve_2d_scale = av1_convolve_2d_scale_c;
+    if (flags & HAS_SSE4_1) av1_convolve_2d_scale = av1_convolve_2d_scale_sse4_1;
     av1_convolve_horiz = av1_convolve_horiz_c;
     if (flags & HAS_SSSE3) av1_convolve_horiz = av1_convolve_horiz_ssse3;
+    av1_convolve_rounding = av1_convolve_rounding_c;
+    if (flags & HAS_AVX2) av1_convolve_rounding = av1_convolve_rounding_avx2;
     av1_convolve_vert = av1_convolve_vert_c;
     if (flags & HAS_SSSE3) av1_convolve_vert = av1_convolve_vert_ssse3;
     av1_fht16x16 = av1_fht16x16_sse2;
     if (flags & HAS_AVX2) av1_fht16x16 = av1_fht16x16_avx2;
     av1_fht32x32 = av1_fht32x32_sse2;
     if (flags & HAS_AVX2) av1_fht32x32 = av1_fht32x32_avx2;
+    av1_filter_intra_edge = av1_filter_intra_edge_c;
+    if (flags & HAS_SSE4_1) av1_filter_intra_edge = av1_filter_intra_edge_sse4_1;
+    av1_filter_intra_edge_high = av1_filter_intra_edge_high_c;
+    if (flags & HAS_SSE4_1) av1_filter_intra_edge_high = av1_filter_intra_edge_high_sse4_1;
     av1_full_search_sad = av1_full_search_sad_c;
     if (flags & HAS_SSE3) av1_full_search_sad = av1_full_search_sadx3;
     if (flags & HAS_SSE4_1) av1_full_search_sad = av1_full_search_sadx8;
@@ -513,14 +513,18 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_SSE4_1) av1_fwd_txfm2d_32x32 = av1_fwd_txfm2d_32x32_sse4_1;
     av1_fwd_txfm2d_4x4 = av1_fwd_txfm2d_4x4_c;
     if (flags & HAS_SSE4_1) av1_fwd_txfm2d_4x4 = av1_fwd_txfm2d_4x4_sse4_1;
-    av1_fwd_txfm2d_64x64 = av1_fwd_txfm2d_64x64_c;
-    if (flags & HAS_SSE4_1) av1_fwd_txfm2d_64x64 = av1_fwd_txfm2d_64x64_sse4_1;
     av1_fwd_txfm2d_8x8 = av1_fwd_txfm2d_8x8_c;
     if (flags & HAS_SSE4_1) av1_fwd_txfm2d_8x8 = av1_fwd_txfm2d_8x8_sse4_1;
+    av1_highbd_convolve_2d = av1_highbd_convolve_2d_c;
+    if (flags & HAS_SSSE3) av1_highbd_convolve_2d = av1_highbd_convolve_2d_ssse3;
+    av1_highbd_convolve_2d_scale = av1_highbd_convolve_2d_scale_c;
+    if (flags & HAS_SSE4_1) av1_highbd_convolve_2d_scale = av1_highbd_convolve_2d_scale_sse4_1;
     av1_highbd_convolve_horiz = av1_highbd_convolve_horiz_c;
     if (flags & HAS_SSE4_1) av1_highbd_convolve_horiz = av1_highbd_convolve_horiz_sse4_1;
     av1_highbd_convolve_init = av1_highbd_convolve_init_c;
     if (flags & HAS_SSE4_1) av1_highbd_convolve_init = av1_highbd_convolve_init_sse4_1;
+    av1_highbd_convolve_rounding = av1_highbd_convolve_rounding_c;
+    if (flags & HAS_AVX2) av1_highbd_convolve_rounding = av1_highbd_convolve_rounding_avx2;
     av1_highbd_convolve_vert = av1_highbd_convolve_vert_c;
     if (flags & HAS_SSE4_1) av1_highbd_convolve_vert = av1_highbd_convolve_vert_sse4_1;
     av1_highbd_quantize_fp = av1_highbd_quantize_fp_c;
@@ -528,6 +532,10 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_AVX2) av1_highbd_quantize_fp = av1_highbd_quantize_fp_avx2;
     av1_highbd_warp_affine = av1_highbd_warp_affine_c;
     if (flags & HAS_SSSE3) av1_highbd_warp_affine = av1_highbd_warp_affine_ssse3;
+    av1_highpass_filter = av1_highpass_filter_c;
+    if (flags & HAS_SSE4_1) av1_highpass_filter = av1_highpass_filter_sse4_1;
+    av1_highpass_filter_highbd = av1_highpass_filter_highbd_c;
+    if (flags & HAS_SSE4_1) av1_highpass_filter_highbd = av1_highpass_filter_highbd_sse4_1;
     av1_iht16x16_256_add = av1_iht16x16_256_add_sse2;
     if (flags & HAS_AVX2) av1_iht16x16_256_add = av1_iht16x16_256_add_avx2;
     av1_inv_txfm2d_add_16x16 = av1_inv_txfm2d_add_16x16_c;
@@ -544,37 +552,34 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_AVX2) av1_quantize_fp = av1_quantize_fp_avx2;
     av1_quantize_fp_32x32 = av1_quantize_fp_32x32_c;
     if (flags & HAS_AVX2) av1_quantize_fp_32x32 = av1_quantize_fp_32x32_avx2;
+    av1_selfguided_restoration = av1_selfguided_restoration_c;
+    if (flags & HAS_SSE4_1) av1_selfguided_restoration = av1_selfguided_restoration_sse4_1;
+    av1_selfguided_restoration_highbd = av1_selfguided_restoration_highbd_c;
+    if (flags & HAS_SSE4_1) av1_selfguided_restoration_highbd = av1_selfguided_restoration_highbd_sse4_1;
+    av1_upsample_intra_edge = av1_upsample_intra_edge_c;
+    if (flags & HAS_SSE4_1) av1_upsample_intra_edge = av1_upsample_intra_edge_sse4_1;
+    av1_upsample_intra_edge_high = av1_upsample_intra_edge_high_c;
+    if (flags & HAS_SSE4_1) av1_upsample_intra_edge_high = av1_upsample_intra_edge_high_sse4_1;
     av1_warp_affine = av1_warp_affine_sse2;
     if (flags & HAS_SSSE3) av1_warp_affine = av1_warp_affine_ssse3;
+    cdef_filter_block = cdef_filter_block_sse2;
+    if (flags & HAS_SSSE3) cdef_filter_block = cdef_filter_block_ssse3;
+    if (flags & HAS_SSE4_1) cdef_filter_block = cdef_filter_block_sse4_1;
+    if (flags & HAS_AVX2) cdef_filter_block = cdef_filter_block_avx2;
+    cdef_find_dir = cdef_find_dir_sse2;
+    if (flags & HAS_SSSE3) cdef_find_dir = cdef_find_dir_ssse3;
+    if (flags & HAS_SSE4_1) cdef_find_dir = cdef_find_dir_sse4_1;
+    if (flags & HAS_AVX2) cdef_find_dir = cdef_find_dir_avx2;
     compute_cross_correlation = compute_cross_correlation_c;
     if (flags & HAS_SSE4_1) compute_cross_correlation = compute_cross_correlation_sse4_1;
-    copy_4x4_16bit_to_16bit = copy_4x4_16bit_to_16bit_sse2;
-    if (flags & HAS_SSSE3) copy_4x4_16bit_to_16bit = copy_4x4_16bit_to_16bit_ssse3;
-    if (flags & HAS_SSE4_1) copy_4x4_16bit_to_16bit = copy_4x4_16bit_to_16bit_sse4_1;
-    copy_4x4_16bit_to_8bit = copy_4x4_16bit_to_8bit_sse2;
-    if (flags & HAS_SSSE3) copy_4x4_16bit_to_8bit = copy_4x4_16bit_to_8bit_ssse3;
-    if (flags & HAS_SSE4_1) copy_4x4_16bit_to_8bit = copy_4x4_16bit_to_8bit_sse4_1;
-    copy_8x8_16bit_to_16bit = copy_8x8_16bit_to_16bit_sse2;
-    if (flags & HAS_SSSE3) copy_8x8_16bit_to_16bit = copy_8x8_16bit_to_16bit_ssse3;
-    if (flags & HAS_SSE4_1) copy_8x8_16bit_to_16bit = copy_8x8_16bit_to_16bit_sse4_1;
-    copy_8x8_16bit_to_8bit = copy_8x8_16bit_to_8bit_sse2;
-    if (flags & HAS_SSSE3) copy_8x8_16bit_to_8bit = copy_8x8_16bit_to_8bit_ssse3;
-    if (flags & HAS_SSE4_1) copy_8x8_16bit_to_8bit = copy_8x8_16bit_to_8bit_sse4_1;
     copy_rect8_16bit_to_16bit = copy_rect8_16bit_to_16bit_sse2;
     if (flags & HAS_SSSE3) copy_rect8_16bit_to_16bit = copy_rect8_16bit_to_16bit_ssse3;
     if (flags & HAS_SSE4_1) copy_rect8_16bit_to_16bit = copy_rect8_16bit_to_16bit_sse4_1;
+    if (flags & HAS_AVX2) copy_rect8_16bit_to_16bit = copy_rect8_16bit_to_16bit_avx2;
     copy_rect8_8bit_to_16bit = copy_rect8_8bit_to_16bit_sse2;
     if (flags & HAS_SSSE3) copy_rect8_8bit_to_16bit = copy_rect8_8bit_to_16bit_ssse3;
     if (flags & HAS_SSE4_1) copy_rect8_8bit_to_16bit = copy_rect8_8bit_to_16bit_sse4_1;
-    od_dir_find8 = od_dir_find8_sse2;
-    if (flags & HAS_SSSE3) od_dir_find8 = od_dir_find8_ssse3;
-    if (flags & HAS_SSE4_1) od_dir_find8 = od_dir_find8_sse4_1;
-    od_filter_dering_direction_4x4 = od_filter_dering_direction_4x4_sse2;
-    if (flags & HAS_SSSE3) od_filter_dering_direction_4x4 = od_filter_dering_direction_4x4_ssse3;
-    if (flags & HAS_SSE4_1) od_filter_dering_direction_4x4 = od_filter_dering_direction_4x4_sse4_1;
-    od_filter_dering_direction_8x8 = od_filter_dering_direction_8x8_sse2;
-    if (flags & HAS_SSSE3) od_filter_dering_direction_8x8 = od_filter_dering_direction_8x8_ssse3;
-    if (flags & HAS_SSE4_1) od_filter_dering_direction_8x8 = od_filter_dering_direction_8x8_sse4_1;
+    if (flags & HAS_AVX2) copy_rect8_8bit_to_16bit = copy_rect8_8bit_to_16bit_avx2;
 }
 #endif