13 files changed, 252 insertions, 102 deletions
diff --git a/media/libaom/config/linux/arm/config/aom_config.asm b/media/libaom/config/linux/arm/config/aom_config.asm
index 80b4a0e06..f5db6f6ca 100644
--- a/media/libaom/config/linux/arm/config/aom_config.asm
+++ b/media/libaom/config/linux/arm/config/aom_config.asm
@@ -14,6 +14,7 @@
 .equ ARCH_PPC, 0
 .equ ARCH_X86, 0
 .equ ARCH_X86_64, 0
+.equ CONFIG_2PASS_PARTITION_SEARCH_LVL, 1
 .equ CONFIG_ACCOUNTING, 0
 .equ CONFIG_ANALYZER, 0
 .equ CONFIG_AV1_DECODER, 1
@@ -21,30 +22,35 @@
 .equ CONFIG_BIG_ENDIAN, 0
 .equ CONFIG_BITSTREAM_DEBUG, 0
 .equ CONFIG_COEFFICIENT_RANGE_CHECKING, 0
-.equ CONFIG_COLLECT_INTER_MODE_RD_STATS, 1
+.equ CONFIG_COLLECT_INTER_MODE_RD_STATS, 0
 .equ CONFIG_COLLECT_RD_STATS, 0
 .equ CONFIG_DEBUG, 0
-.equ CONFIG_DENOISE, 0
-.equ CONFIG_DIST_8X8, 1
+.equ CONFIG_DENOISE, 1
+.equ CONFIG_DIST_8X8, 0
 .equ CONFIG_ENTROPY_STATS, 0
 .equ CONFIG_FILEOPTIONS, 1
+.equ CONFIG_FIX_GF_LENGTH, 1
 .equ CONFIG_FP_MB_STATS, 0
 .equ CONFIG_GCC, 1
 .equ CONFIG_GCOV, 0
+.equ CONFIG_GLOBAL_MOTION_SEARCH, 1
 .equ CONFIG_GPROF, 0
 .equ CONFIG_INSPECTION, 0
 .equ CONFIG_INTERNAL_STATS, 0
 .equ CONFIG_INTER_STATS_ONLY, 0
 .equ CONFIG_LIBYUV, 0
-.equ CONFIG_LOWBITDEPTH, 0
+.equ CONFIG_LOWBITDEPTH, 1
+.equ CONFIG_MAX_DECODE_PROFILE, 2
 .equ CONFIG_MISMATCH_DEBUG, 0
-.equ CONFIG_MSVS, 0
 .equ CONFIG_MULTITHREAD, 1
+.equ CONFIG_NORMAL_TILE_MODE, 0
 .equ CONFIG_OS_SUPPORT, 1
-.equ CONFIG_PIC, 0
+.equ CONFIG_PIC, 1
 .equ CONFIG_RD_DEBUG, 0
+.equ CONFIG_REDUCED_ENCODER_BORDER, 0
 .equ CONFIG_RUNTIME_CPU_DETECT, 1
 .equ CONFIG_SHARED, 0
+.equ CONFIG_SHARP_SETTINGS, 0
 .equ CONFIG_SIZE_LIMIT, 0
 .equ CONFIG_SPATIAL_RESAMPLING, 1
 .equ CONFIG_STATIC, 1
diff --git a/media/libaom/config/linux/arm/config/aom_config.h b/media/libaom/config/linux/arm/config/aom_config.h
index 20bb0c68e..850eebfbd 100644
--- a/media/libaom/config/linux/arm/config/aom_config.h
+++ b/media/libaom/config/linux/arm/config/aom_config.h
@@ -16,6 +16,7 @@
 #define ARCH_PPC 0
 #define ARCH_X86 0
 #define ARCH_X86_64 0
+#define CONFIG_2PASS_PARTITION_SEARCH_LVL 1
 #define CONFIG_ACCOUNTING 0
 #define CONFIG_ANALYZER 0
 #define CONFIG_AV1_DECODER 1
@@ -23,30 +24,35 @@
 #define CONFIG_BIG_ENDIAN 0
 #define CONFIG_BITSTREAM_DEBUG 0
 #define CONFIG_COEFFICIENT_RANGE_CHECKING 0
-#define CONFIG_COLLECT_INTER_MODE_RD_STATS 1
+#define CONFIG_COLLECT_INTER_MODE_RD_STATS 0
 #define CONFIG_COLLECT_RD_STATS 0
 #define CONFIG_DEBUG 0
-#define CONFIG_DENOISE 0
-#define CONFIG_DIST_8X8 1
+#define CONFIG_DENOISE 1
+#define CONFIG_DIST_8X8 0
 #define CONFIG_ENTROPY_STATS 0
 #define CONFIG_FILEOPTIONS 1
+#define CONFIG_FIX_GF_LENGTH 1
 #define CONFIG_FP_MB_STATS 0
 #define CONFIG_GCC 1
 #define CONFIG_GCOV 0
+#define CONFIG_GLOBAL_MOTION_SEARCH 1
 #define CONFIG_GPROF 0
 #define CONFIG_INSPECTION 0
 #define CONFIG_INTERNAL_STATS 0
 #define CONFIG_INTER_STATS_ONLY 0
 #define CONFIG_LIBYUV 0
-#define CONFIG_LOWBITDEPTH 0
+#define CONFIG_LOWBITDEPTH 1
+#define CONFIG_MAX_DECODE_PROFILE 2
 #define CONFIG_MISMATCH_DEBUG 0
-#define CONFIG_MSVS 0
 #define CONFIG_MULTITHREAD 1
+#define CONFIG_NORMAL_TILE_MODE 0
 #define CONFIG_OS_SUPPORT 1
-#define CONFIG_PIC 0
+#define CONFIG_PIC 1
 #define CONFIG_RD_DEBUG 0
+#define CONFIG_REDUCED_ENCODER_BORDER 0
 #define CONFIG_RUNTIME_CPU_DETECT 1
 #define CONFIG_SHARED 0
+#define CONFIG_SHARP_SETTINGS 0
 #define CONFIG_SIZE_LIMIT 0
 #define CONFIG_SPATIAL_RESAMPLING 1
 #define CONFIG_STATIC 1
diff --git a/media/libaom/config/linux/arm/config/aom_dsp_rtcd.h b/media/libaom/config/linux/arm/config/aom_dsp_rtcd.h
index 291d8b8ca..d80f8972b 100644
--- a/media/libaom/config/linux/arm/config/aom_dsp_rtcd.h
+++ b/media/libaom/config/linux/arm/config/aom_dsp_rtcd.h
@@ -1412,7 +1412,7 @@ void aom_dsp_rtcd(void);
 #include "aom_ports/arm.h"
 static void setup_rtcd_internal(void)
 {
-    int flags = arm_cpu_caps();
+    int flags = aom_arm_cpu_caps();
 
     (void)flags;
 
diff --git a/media/libaom/config/linux/arm/config/aom_scale_rtcd.h b/media/libaom/config/linux/arm/config/aom_scale_rtcd.h
index 52aaa83b0..165d73093 100644
--- a/media/libaom/config/linux/arm/config/aom_scale_rtcd.h
+++ b/media/libaom/config/linux/arm/config/aom_scale_rtcd.h
@@ -76,7 +76,7 @@ void aom_scale_rtcd(void);
 #include "aom_ports/arm.h"
 static void setup_rtcd_internal(void)
 {
-    int flags = arm_cpu_caps();
+    int flags = aom_arm_cpu_caps();
 
     (void)flags;
 
diff --git a/media/libaom/config/linux/arm/config/av1_rtcd.h b/media/libaom/config/linux/arm/config/av1_rtcd.h
index 036c0f545..b54b4514d 100644
--- a/media/libaom/config/linux/arm/config/av1_rtcd.h
+++ b/media/libaom/config/linux/arm/config/av1_rtcd.h
@@ -1,6 +1,6 @@
 // This file is generated. Do not edit.
-#ifndef AOM_RTCD_H_
-#define AOM_RTCD_H_
+#ifndef AV1_RTCD_H_
+#define AV1_RTCD_H_
 
 #ifdef RTCD_C
 #define RTCD_EXTERN
@@ -147,6 +147,27 @@ void av1_highbd_dr_prediction_z2_c(uint16_t *dst, ptrdiff_t stride, int bw, int
 void av1_highbd_dr_prediction_z3_c(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_left, int dx, int dy, int bd);
 #define av1_highbd_dr_prediction_z3 av1_highbd_dr_prediction_z3_c
 
+void av1_highbd_inv_txfm_add_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add av1_highbd_inv_txfm_add_c
+
+void av1_highbd_inv_txfm_add_16x16_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_16x16 av1_highbd_inv_txfm_add_16x16_c
+
+void av1_highbd_inv_txfm_add_16x8_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_16x8 av1_highbd_inv_txfm_add_16x8_c
+
+void av1_highbd_inv_txfm_add_32x32_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_32x32 av1_highbd_inv_txfm_add_32x32_c
+
+void av1_highbd_inv_txfm_add_4x4_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_4x4 av1_highbd_inv_txfm_add_4x4_c
+
+void av1_highbd_inv_txfm_add_8x16_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_8x16 av1_highbd_inv_txfm_add_8x16_c
+
+void av1_highbd_inv_txfm_add_8x8_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+#define av1_highbd_inv_txfm_add_8x8 av1_highbd_inv_txfm_add_8x8_c
+
 void av1_highbd_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
 #define av1_highbd_iwht4x4_16_add av1_highbd_iwht4x4_16_add_c
 
@@ -248,15 +269,15 @@ void av1_jnt_convolve_y_c(const uint8_t *src, int src_stride, uint8_t *dst, int
 void av1_jnt_convolve_y_neon(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
 RTCD_EXTERN void (*av1_jnt_convolve_y)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
 
-void av1_selfguided_restoration_c(const uint8_t *dgd8, int width, int height,
-                                  int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
-                                  int sgr_params_idx, int bit_depth, int highbd);
-void av1_selfguided_restoration_neon(const uint8_t *dgd8, int width, int height,
-                                  int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
-                                  int sgr_params_idx, int bit_depth, int highbd);
-RTCD_EXTERN void (*av1_selfguided_restoration)(const uint8_t *dgd8, int width, int height,
-                                  int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
-                                  int sgr_params_idx, int bit_depth, int highbd);
+int av1_selfguided_restoration_c(const uint8_t *dgd8, int width, int height,
+                                 int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
+                                 int sgr_params_idx, int bit_depth, int highbd);
+int av1_selfguided_restoration_neon(const uint8_t *dgd8, int width, int height,
+                                 int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
+                                 int sgr_params_idx, int bit_depth, int highbd);
+RTCD_EXTERN int (*av1_selfguided_restoration)(const uint8_t *dgd8, int width, int height,
+                                 int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
+                                 int sgr_params_idx, int bit_depth, int highbd);
 
 void av1_upsample_intra_edge_c(uint8_t *p, int sz);
 #define av1_upsample_intra_edge av1_upsample_intra_edge_c
@@ -265,7 +286,8 @@ void av1_upsample_intra_edge_high_c(uint16_t *p, int sz, int bd);
 #define av1_upsample_intra_edge_high av1_upsample_intra_edge_high_c
 
 void av1_warp_affine_c(const int32_t *mat, const uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
-#define av1_warp_affine av1_warp_affine_c
+void av1_warp_affine_neon(const int32_t *mat, const uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
+RTCD_EXTERN void (*av1_warp_affine)(const int32_t *mat, const uint8_t *ref, int width, int height, int stride, uint8_t *pred, int p_col, int p_row, int p_width, int p_height, int p_stride, int subsampling_x, int subsampling_y, ConvolveParams *conv_params, int16_t alpha, int16_t beta, int16_t gamma, int16_t delta);
 
 void av1_wiener_convolve_add_src_c(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const ConvolveParams *conv_params);
 void av1_wiener_convolve_add_src_neon(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const ConvolveParams *conv_params);
@@ -323,7 +345,7 @@ cfl_subtract_average_fn get_subtract_average_fn_c(TX_SIZE tx_size);
 cfl_subtract_average_fn get_subtract_average_fn_neon(TX_SIZE tx_size);
 RTCD_EXTERN cfl_subtract_average_fn (*get_subtract_average_fn)(TX_SIZE tx_size);
 
-void aom_rtcd(void);
+void av1_rtcd(void);
 
 #include "config/aom_config.h"
 
@@ -331,7 +353,7 @@ void aom_rtcd(void);
 #include "aom_ports/arm.h"
 static void setup_rtcd_internal(void)
 {
-    int flags = arm_cpu_caps();
+    int flags = aom_arm_cpu_caps();
 
     (void)flags;
 
@@ -359,6 +381,8 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_NEON) av1_jnt_convolve_y = av1_jnt_convolve_y_neon;
     av1_selfguided_restoration = av1_selfguided_restoration_c;
     if (flags & HAS_NEON) av1_selfguided_restoration = av1_selfguided_restoration_neon;
+    av1_warp_affine = av1_warp_affine_c;
+    if (flags & HAS_NEON) av1_warp_affine = av1_warp_affine_neon;
     av1_wiener_convolve_add_src = av1_wiener_convolve_add_src_c;
     if (flags & HAS_NEON) av1_wiener_convolve_add_src = av1_wiener_convolve_add_src_neon;
     cdef_filter_block = cdef_filter_block_c;
diff --git a/media/libaom/config/linux/ia32/config/aom_config.asm b/media/libaom/config/linux/ia32/config/aom_config.asm
index ee4ac9765..128f85ec2 100644
--- a/media/libaom/config/linux/ia32/config/aom_config.asm
+++ b/media/libaom/config/linux/ia32/config/aom_config.asm
@@ -14,6 +14,7 @@ ARCH_MIPS equ 0
 ARCH_PPC equ 0
 ARCH_X86 equ 1
 ARCH_X86_64 equ 0
+CONFIG_2PASS_PARTITION_SEARCH_LVL equ 1
 CONFIG_ACCOUNTING equ 0
 CONFIG_ANALYZER equ 0
 CONFIG_AV1_DECODER equ 1
@@ -21,30 +22,35 @@ CONFIG_AV1_ENCODER equ 0
 CONFIG_BIG_ENDIAN equ 0
 CONFIG_BITSTREAM_DEBUG equ 0
 CONFIG_COEFFICIENT_RANGE_CHECKING equ 0
-CONFIG_COLLECT_INTER_MODE_RD_STATS equ 1
+CONFIG_COLLECT_INTER_MODE_RD_STATS equ 0
 CONFIG_COLLECT_RD_STATS equ 0
 CONFIG_DEBUG equ 0
-CONFIG_DENOISE equ 0
-CONFIG_DIST_8X8 equ 1
+CONFIG_DENOISE equ 1
+CONFIG_DIST_8X8 equ 0
 CONFIG_ENTROPY_STATS equ 0
 CONFIG_FILEOPTIONS equ 1
+CONFIG_FIX_GF_LENGTH equ 1
 CONFIG_FP_MB_STATS equ 0
 CONFIG_GCC equ 1
 CONFIG_GCOV equ 0
+CONFIG_GLOBAL_MOTION_SEARCH equ 1
 CONFIG_GPROF equ 0
 CONFIG_INSPECTION equ 0
 CONFIG_INTERNAL_STATS equ 0
 CONFIG_INTER_STATS_ONLY equ 0
 CONFIG_LIBYUV equ 0
-CONFIG_LOWBITDEPTH equ 0
+CONFIG_LOWBITDEPTH equ 1
+CONFIG_MAX_DECODE_PROFILE equ 2
 CONFIG_MISMATCH_DEBUG equ 0
-CONFIG_MSVS equ 0
 CONFIG_MULTITHREAD equ 1
+CONFIG_NORMAL_TILE_MODE equ 0
 CONFIG_OS_SUPPORT equ 1
 CONFIG_PIC equ 1
 CONFIG_RD_DEBUG equ 0
+CONFIG_REDUCED_ENCODER_BORDER equ 0
 CONFIG_RUNTIME_CPU_DETECT equ 1
 CONFIG_SHARED equ 0
+CONFIG_SHARP_SETTINGS equ 0
 CONFIG_SIZE_LIMIT equ 0
 CONFIG_SPATIAL_RESAMPLING equ 1
 CONFIG_STATIC equ 1
diff --git a/media/libaom/config/linux/ia32/config/aom_config.h b/media/libaom/config/linux/ia32/config/aom_config.h
index a0066f30c..02f6f7243 100644
--- a/media/libaom/config/linux/ia32/config/aom_config.h
+++ b/media/libaom/config/linux/ia32/config/aom_config.h
@@ -16,6 +16,7 @@
 #define ARCH_PPC 0
 #define ARCH_X86 1
 #define ARCH_X86_64 0
+#define CONFIG_2PASS_PARTITION_SEARCH_LVL 1
 #define CONFIG_ACCOUNTING 0
 #define CONFIG_ANALYZER 0
 #define CONFIG_AV1_DECODER 1
@@ -23,30 +24,35 @@
 #define CONFIG_BIG_ENDIAN 0
 #define CONFIG_BITSTREAM_DEBUG 0
 #define CONFIG_COEFFICIENT_RANGE_CHECKING 0
-#define CONFIG_COLLECT_INTER_MODE_RD_STATS 1
+#define CONFIG_COLLECT_INTER_MODE_RD_STATS 0
 #define CONFIG_COLLECT_RD_STATS 0
 #define CONFIG_DEBUG 0
-#define CONFIG_DENOISE 0
-#define CONFIG_DIST_8X8 1
+#define CONFIG_DENOISE 1
+#define CONFIG_DIST_8X8 0
 #define CONFIG_ENTROPY_STATS 0
 #define CONFIG_FILEOPTIONS 1
+#define CONFIG_FIX_GF_LENGTH 1
 #define CONFIG_FP_MB_STATS 0
 #define CONFIG_GCC 1
 #define CONFIG_GCOV 0
+#define CONFIG_GLOBAL_MOTION_SEARCH 1
 #define CONFIG_GPROF 0
 #define CONFIG_INSPECTION 0
 #define CONFIG_INTERNAL_STATS 0
 #define CONFIG_INTER_STATS_ONLY 0
 #define CONFIG_LIBYUV 0
-#define CONFIG_LOWBITDEPTH 0
+#define CONFIG_LOWBITDEPTH 1
+#define CONFIG_MAX_DECODE_PROFILE 2
 #define CONFIG_MISMATCH_DEBUG 0
-#define CONFIG_MSVS 0
 #define CONFIG_MULTITHREAD 1
+#define CONFIG_NORMAL_TILE_MODE 0
 #define CONFIG_OS_SUPPORT 1
 #define CONFIG_PIC 1
 #define CONFIG_RD_DEBUG 0
+#define CONFIG_REDUCED_ENCODER_BORDER 0
 #define CONFIG_RUNTIME_CPU_DETECT 1
 #define CONFIG_SHARED 0
+#define CONFIG_SHARP_SETTINGS 0
 #define CONFIG_SIZE_LIMIT 0
 #define CONFIG_SPATIAL_RESAMPLING 1
 #define CONFIG_STATIC 1
diff --git a/media/libaom/config/linux/ia32/config/aom_dsp_rtcd.h b/media/libaom/config/linux/ia32/config/aom_dsp_rtcd.h
index 02abbdc39..5b3762f10 100644
--- a/media/libaom/config/linux/ia32/config/aom_dsp_rtcd.h
+++ b/media/libaom/config/linux/ia32/config/aom_dsp_rtcd.h
@@ -28,6 +28,7 @@ RTCD_EXTERN void (*aom_blend_a64_hmask)(uint8_t *dst, uint32_t dst_stride, const
 
 void aom_blend_a64_mask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subx, int suby);
 void aom_blend_a64_mask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subx, int suby);
+void aom_blend_a64_mask_avx2(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subx, int suby);
 RTCD_EXTERN void (*aom_blend_a64_mask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subx, int suby);
 
 void aom_blend_a64_vmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h);
@@ -1230,6 +1231,7 @@ RTCD_EXTERN void (*aom_highbd_v_predictor_8x8)(uint16_t *dst, ptrdiff_t y_stride
 
 void aom_lowbd_blend_a64_d16_mask_c(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subx, int suby, ConvolveParams *conv_params);
 void aom_lowbd_blend_a64_d16_mask_sse4_1(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subx, int suby, ConvolveParams *conv_params);
+void aom_lowbd_blend_a64_d16_mask_avx2(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subx, int suby, ConvolveParams *conv_params);
 RTCD_EXTERN void (*aom_lowbd_blend_a64_d16_mask)(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subx, int suby, ConvolveParams *conv_params);
 
 void aom_lpf_horizontal_14_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
@@ -1725,6 +1727,7 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_SSE4_1) aom_blend_a64_hmask = aom_blend_a64_hmask_sse4_1;
     aom_blend_a64_mask = aom_blend_a64_mask_c;
     if (flags & HAS_SSE4_1) aom_blend_a64_mask = aom_blend_a64_mask_sse4_1;
+    if (flags & HAS_AVX2) aom_blend_a64_mask = aom_blend_a64_mask_avx2;
     aom_blend_a64_vmask = aom_blend_a64_vmask_c;
     if (flags & HAS_SSE4_1) aom_blend_a64_vmask = aom_blend_a64_vmask_sse4_1;
     aom_convolve8_horiz = aom_convolve8_horiz_c;
@@ -2125,6 +2128,7 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_SSE2) aom_highbd_v_predictor_8x8 = aom_highbd_v_predictor_8x8_sse2;
     aom_lowbd_blend_a64_d16_mask = aom_lowbd_blend_a64_d16_mask_c;
     if (flags & HAS_SSE4_1) aom_lowbd_blend_a64_d16_mask = aom_lowbd_blend_a64_d16_mask_sse4_1;
+    if (flags & HAS_AVX2) aom_lowbd_blend_a64_d16_mask = aom_lowbd_blend_a64_d16_mask_avx2;
     aom_lpf_horizontal_14 = aom_lpf_horizontal_14_c;
     if (flags & HAS_SSE2) aom_lpf_horizontal_14 = aom_lpf_horizontal_14_sse2;
     aom_lpf_horizontal_14_dual = aom_lpf_horizontal_14_dual_c;
diff --git a/media/libaom/config/linux/ia32/config/av1_rtcd.h b/media/libaom/config/linux/ia32/config/av1_rtcd.h
index 9bb457db2..a2a6f14ff 100644
--- a/media/libaom/config/linux/ia32/config/av1_rtcd.h
+++ b/media/libaom/config/linux/ia32/config/av1_rtcd.h
@@ -1,6 +1,6 @@
 // This file is generated. Do not edit.
-#ifndef AOM_RTCD_H_
-#define AOM_RTCD_H_
+#ifndef AV1_RTCD_H_
+#define AV1_RTCD_H_
 
 #ifdef RTCD_C
 #define RTCD_EXTERN
@@ -58,10 +58,12 @@ RTCD_EXTERN void (*apply_selfguided_restoration)(const uint8_t *dat, int width,
 
 void av1_build_compound_diffwtd_mask_c(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w);
 void av1_build_compound_diffwtd_mask_sse4_1(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w);
+void av1_build_compound_diffwtd_mask_avx2(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w);
 RTCD_EXTERN void (*av1_build_compound_diffwtd_mask)(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w);
 
 void av1_build_compound_diffwtd_mask_d16_c(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const CONV_BUF_TYPE *src0, int src0_stride, const CONV_BUF_TYPE *src1, int src1_stride, int h, int w, ConvolveParams *conv_params, int bd);
 void av1_build_compound_diffwtd_mask_d16_sse4_1(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const CONV_BUF_TYPE *src0, int src0_stride, const CONV_BUF_TYPE *src1, int src1_stride, int h, int w, ConvolveParams *conv_params, int bd);
+void av1_build_compound_diffwtd_mask_d16_avx2(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const CONV_BUF_TYPE *src0, int src0_stride, const CONV_BUF_TYPE *src1, int src1_stride, int h, int w, ConvolveParams *conv_params, int bd);
 RTCD_EXTERN void (*av1_build_compound_diffwtd_mask_d16)(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const CONV_BUF_TYPE *src0, int src0_stride, const CONV_BUF_TYPE *src1, int src1_stride, int h, int w, ConvolveParams *conv_params, int bd);
 
 void av1_build_compound_diffwtd_mask_highbd_c(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w, int bd);
@@ -170,6 +172,36 @@ void av1_highbd_dr_prediction_z2_c(uint16_t *dst, ptrdiff_t stride, int bw, int
 void av1_highbd_dr_prediction_z3_c(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_left, int dx, int dy, int bd);
 #define av1_highbd_dr_prediction_z3 av1_highbd_dr_prediction_z3_c
 
+void av1_highbd_inv_txfm_add_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_sse4_1(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_avx2(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_16x16_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_16x16_sse4_1(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_16x16)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_16x8_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_16x8_sse4_1(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_16x8)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_32x32_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_32x32_sse4_1(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_32x32_avx2(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_32x32)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_4x4_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_4x4_sse4_1(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_4x4)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_8x16_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_8x16_sse4_1(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_8x16)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_8x8_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_8x8_sse4_1(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_8x8)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+
 void av1_highbd_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
 #define av1_highbd_iwht4x4_16_add av1_highbd_iwht4x4_16_add_c
 
@@ -206,8 +238,7 @@ void av1_highbd_wiener_convolve_add_src_avx2(const uint8_t *src, ptrdiff_t src_s
 RTCD_EXTERN void (*av1_highbd_wiener_convolve_add_src)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const ConvolveParams *conv_params, int bps);
 
 void av1_inv_txfm2d_add_16x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
-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);
+#define av1_inv_txfm2d_add_16x16 av1_inv_txfm2d_add_16x16_c
 
 void av1_inv_txfm2d_add_16x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
 #define av1_inv_txfm2d_add_16x32 av1_inv_txfm2d_add_16x32_c
@@ -225,8 +256,7 @@ void av1_inv_txfm2d_add_32x16_c(const int32_t *input, uint16_t *output, int stri
 #define av1_inv_txfm2d_add_32x16 av1_inv_txfm2d_add_32x16_c
 
 void av1_inv_txfm2d_add_32x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
-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);
+#define av1_inv_txfm2d_add_32x32 av1_inv_txfm2d_add_32x32_c
 
 void av1_inv_txfm2d_add_32x64_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
 #define av1_inv_txfm2d_add_32x64 av1_inv_txfm2d_add_32x64_c
@@ -251,8 +281,7 @@ void av1_inv_txfm2d_add_64x32_c(const int32_t *input, uint16_t *output, int stri
 #define av1_inv_txfm2d_add_64x32 av1_inv_txfm2d_add_64x32_c
 
 void av1_inv_txfm2d_add_64x64_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
-void av1_inv_txfm2d_add_64x64_sse4_1(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
-RTCD_EXTERN void (*av1_inv_txfm2d_add_64x64)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_64x64 av1_inv_txfm2d_add_64x64_c
 
 void av1_inv_txfm2d_add_8x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
 #define av1_inv_txfm2d_add_8x16 av1_inv_txfm2d_add_8x16_c
@@ -292,18 +321,18 @@ void av1_jnt_convolve_y_sse2(const uint8_t *src, int src_stride, uint8_t *dst, i
 void av1_jnt_convolve_y_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
 RTCD_EXTERN void (*av1_jnt_convolve_y)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
 
-void av1_selfguided_restoration_c(const uint8_t *dgd8, int width, int height,
-                                  int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
-                                  int sgr_params_idx, int bit_depth, int highbd);
-void av1_selfguided_restoration_sse4_1(const uint8_t *dgd8, int width, int height,
-                                  int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
-                                  int sgr_params_idx, int bit_depth, int highbd);
-void av1_selfguided_restoration_avx2(const uint8_t *dgd8, int width, int height,
-                                  int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
-                                  int sgr_params_idx, int bit_depth, int highbd);
-RTCD_EXTERN void (*av1_selfguided_restoration)(const uint8_t *dgd8, int width, int height,
-                                  int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
-                                  int sgr_params_idx, int bit_depth, int highbd);
+int av1_selfguided_restoration_c(const uint8_t *dgd8, int width, int height,
+                                 int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
+                                 int sgr_params_idx, int bit_depth, int highbd);
+int av1_selfguided_restoration_sse4_1(const uint8_t *dgd8, int width, int height,
+                                 int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
+                                 int sgr_params_idx, int bit_depth, int highbd);
+int av1_selfguided_restoration_avx2(const uint8_t *dgd8, int width, int height,
+                                 int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
+                                 int sgr_params_idx, int bit_depth, int highbd);
+RTCD_EXTERN int (*av1_selfguided_restoration)(const uint8_t *dgd8, int width, int height,
+                                 int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
+                                 int sgr_params_idx, int bit_depth, int highbd);
 
 void av1_upsample_intra_edge_c(uint8_t *p, int sz);
 void av1_upsample_intra_edge_sse4_1(uint8_t *p, int sz);
@@ -395,7 +424,7 @@ cfl_subtract_average_fn get_subtract_average_fn_sse2(TX_SIZE tx_size);
 cfl_subtract_average_fn get_subtract_average_fn_avx2(TX_SIZE tx_size);
 RTCD_EXTERN cfl_subtract_average_fn (*get_subtract_average_fn)(TX_SIZE tx_size);
 
-void aom_rtcd(void);
+void av1_rtcd(void);
 
 #ifdef RTCD_C
 #include "aom_ports/x86.h"
@@ -410,8 +439,10 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_AVX2) apply_selfguided_restoration = apply_selfguided_restoration_avx2;
     av1_build_compound_diffwtd_mask = av1_build_compound_diffwtd_mask_c;
     if (flags & HAS_SSE4_1) av1_build_compound_diffwtd_mask = av1_build_compound_diffwtd_mask_sse4_1;
+    if (flags & HAS_AVX2) av1_build_compound_diffwtd_mask = av1_build_compound_diffwtd_mask_avx2;
     av1_build_compound_diffwtd_mask_d16 = av1_build_compound_diffwtd_mask_d16_c;
     if (flags & HAS_SSE4_1) av1_build_compound_diffwtd_mask_d16 = av1_build_compound_diffwtd_mask_d16_sse4_1;
+    if (flags & HAS_AVX2) av1_build_compound_diffwtd_mask_d16 = av1_build_compound_diffwtd_mask_d16_avx2;
     av1_build_compound_diffwtd_mask_highbd = av1_build_compound_diffwtd_mask_highbd_c;
     if (flags & HAS_SSSE3) av1_build_compound_diffwtd_mask_highbd = av1_build_compound_diffwtd_mask_highbd_ssse3;
     if (flags & HAS_AVX2) av1_build_compound_diffwtd_mask_highbd = av1_build_compound_diffwtd_mask_highbd_avx2;
@@ -453,6 +484,22 @@ static void setup_rtcd_internal(void)
     av1_highbd_convolve_y_sr = av1_highbd_convolve_y_sr_c;
     if (flags & HAS_SSSE3) av1_highbd_convolve_y_sr = av1_highbd_convolve_y_sr_ssse3;
     if (flags & HAS_AVX2) av1_highbd_convolve_y_sr = av1_highbd_convolve_y_sr_avx2;
+    av1_highbd_inv_txfm_add = av1_highbd_inv_txfm_add_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add = av1_highbd_inv_txfm_add_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_inv_txfm_add = av1_highbd_inv_txfm_add_avx2;
+    av1_highbd_inv_txfm_add_16x16 = av1_highbd_inv_txfm_add_16x16_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_16x16 = av1_highbd_inv_txfm_add_16x16_sse4_1;
+    av1_highbd_inv_txfm_add_16x8 = av1_highbd_inv_txfm_add_16x8_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_16x8 = av1_highbd_inv_txfm_add_16x8_sse4_1;
+    av1_highbd_inv_txfm_add_32x32 = av1_highbd_inv_txfm_add_32x32_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_32x32 = av1_highbd_inv_txfm_add_32x32_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_inv_txfm_add_32x32 = av1_highbd_inv_txfm_add_32x32_avx2;
+    av1_highbd_inv_txfm_add_4x4 = av1_highbd_inv_txfm_add_4x4_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_4x4 = av1_highbd_inv_txfm_add_4x4_sse4_1;
+    av1_highbd_inv_txfm_add_8x16 = av1_highbd_inv_txfm_add_8x16_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_8x16 = av1_highbd_inv_txfm_add_8x16_sse4_1;
+    av1_highbd_inv_txfm_add_8x8 = av1_highbd_inv_txfm_add_8x8_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_8x8 = av1_highbd_inv_txfm_add_8x8_sse4_1;
     av1_highbd_jnt_convolve_2d = av1_highbd_jnt_convolve_2d_c;
     if (flags & HAS_SSE4_1) av1_highbd_jnt_convolve_2d = av1_highbd_jnt_convolve_2d_sse4_1;
     if (flags & HAS_AVX2) av1_highbd_jnt_convolve_2d = av1_highbd_jnt_convolve_2d_avx2;
@@ -470,14 +517,8 @@ static void setup_rtcd_internal(void)
     av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_c;
     if (flags & HAS_SSSE3) av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_ssse3;
     if (flags & HAS_AVX2) av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_avx2;
-    av1_inv_txfm2d_add_16x16 = av1_inv_txfm2d_add_16x16_c;
-    if (flags & HAS_SSE4_1) av1_inv_txfm2d_add_16x16 = av1_inv_txfm2d_add_16x16_sse4_1;
-    av1_inv_txfm2d_add_32x32 = av1_inv_txfm2d_add_32x32_c;
-    if (flags & HAS_AVX2) av1_inv_txfm2d_add_32x32 = av1_inv_txfm2d_add_32x32_avx2;
     av1_inv_txfm2d_add_4x4 = av1_inv_txfm2d_add_4x4_c;
     if (flags & HAS_SSE4_1) av1_inv_txfm2d_add_4x4 = av1_inv_txfm2d_add_4x4_sse4_1;
-    av1_inv_txfm2d_add_64x64 = av1_inv_txfm2d_add_64x64_c;
-    if (flags & HAS_SSE4_1) av1_inv_txfm2d_add_64x64 = av1_inv_txfm2d_add_64x64_sse4_1;
     av1_inv_txfm2d_add_8x8 = av1_inv_txfm2d_add_8x8_c;
     if (flags & HAS_SSE4_1) av1_inv_txfm2d_add_8x8 = av1_inv_txfm2d_add_8x8_sse4_1;
     av1_inv_txfm_add = av1_inv_txfm_add_c;
diff --git a/media/libaom/config/linux/x64/config/aom_config.asm b/media/libaom/config/linux/x64/config/aom_config.asm
index f00704849..4ab2dedb4 100644
--- a/media/libaom/config/linux/x64/config/aom_config.asm
+++ b/media/libaom/config/linux/x64/config/aom_config.asm
@@ -14,6 +14,7 @@ ARCH_MIPS equ 0
 ARCH_PPC equ 0
 ARCH_X86 equ 0
 ARCH_X86_64 equ 1
+CONFIG_2PASS_PARTITION_SEARCH_LVL equ 1
 CONFIG_ACCOUNTING equ 0
 CONFIG_ANALYZER equ 0
 CONFIG_AV1_DECODER equ 1
@@ -21,30 +22,35 @@ CONFIG_AV1_ENCODER equ 0
 CONFIG_BIG_ENDIAN equ 0
 CONFIG_BITSTREAM_DEBUG equ 0
 CONFIG_COEFFICIENT_RANGE_CHECKING equ 0
-CONFIG_COLLECT_INTER_MODE_RD_STATS equ 1
+CONFIG_COLLECT_INTER_MODE_RD_STATS equ 0
 CONFIG_COLLECT_RD_STATS equ 0
 CONFIG_DEBUG equ 0
-CONFIG_DENOISE equ 0
-CONFIG_DIST_8X8 equ 1
+CONFIG_DENOISE equ 1
+CONFIG_DIST_8X8 equ 0
 CONFIG_ENTROPY_STATS equ 0
 CONFIG_FILEOPTIONS equ 1
+CONFIG_FIX_GF_LENGTH equ 1
 CONFIG_FP_MB_STATS equ 0
 CONFIG_GCC equ 1
 CONFIG_GCOV equ 0
+CONFIG_GLOBAL_MOTION_SEARCH equ 1
 CONFIG_GPROF equ 0
 CONFIG_INSPECTION equ 0
 CONFIG_INTERNAL_STATS equ 0
 CONFIG_INTER_STATS_ONLY equ 0
 CONFIG_LIBYUV equ 0
-CONFIG_LOWBITDEPTH equ 0
+CONFIG_LOWBITDEPTH equ 1
+CONFIG_MAX_DECODE_PROFILE equ 2
 CONFIG_MISMATCH_DEBUG equ 0
-CONFIG_MSVS equ 0
 CONFIG_MULTITHREAD equ 1
+CONFIG_NORMAL_TILE_MODE equ 0
 CONFIG_OS_SUPPORT equ 1
 CONFIG_PIC equ 0
 CONFIG_RD_DEBUG equ 0
+CONFIG_REDUCED_ENCODER_BORDER equ 0
 CONFIG_RUNTIME_CPU_DETECT equ 1
 CONFIG_SHARED equ 0
+CONFIG_SHARP_SETTINGS equ 0
 CONFIG_SIZE_LIMIT equ 0
 CONFIG_SPATIAL_RESAMPLING equ 1
 CONFIG_STATIC equ 1
diff --git a/media/libaom/config/linux/x64/config/aom_config.h b/media/libaom/config/linux/x64/config/aom_config.h
index a036dc904..9a0be7cd1 100644
--- a/media/libaom/config/linux/x64/config/aom_config.h
+++ b/media/libaom/config/linux/x64/config/aom_config.h
@@ -16,6 +16,7 @@
 #define ARCH_PPC 0
 #define ARCH_X86 0
 #define ARCH_X86_64 1
+#define CONFIG_2PASS_PARTITION_SEARCH_LVL 1
 #define CONFIG_ACCOUNTING 0
 #define CONFIG_ANALYZER 0
 #define CONFIG_AV1_DECODER 1
@@ -23,30 +24,35 @@
 #define CONFIG_BIG_ENDIAN 0
 #define CONFIG_BITSTREAM_DEBUG 0
 #define CONFIG_COEFFICIENT_RANGE_CHECKING 0
-#define CONFIG_COLLECT_INTER_MODE_RD_STATS 1
+#define CONFIG_COLLECT_INTER_MODE_RD_STATS 0
 #define CONFIG_COLLECT_RD_STATS 0
 #define CONFIG_DEBUG 0
-#define CONFIG_DENOISE 0
-#define CONFIG_DIST_8X8 1
+#define CONFIG_DENOISE 1
+#define CONFIG_DIST_8X8 0
 #define CONFIG_ENTROPY_STATS 0
 #define CONFIG_FILEOPTIONS 1
+#define CONFIG_FIX_GF_LENGTH 1
 #define CONFIG_FP_MB_STATS 0
 #define CONFIG_GCC 1
 #define CONFIG_GCOV 0
+#define CONFIG_GLOBAL_MOTION_SEARCH 1
 #define CONFIG_GPROF 0
 #define CONFIG_INSPECTION 0
 #define CONFIG_INTERNAL_STATS 0
 #define CONFIG_INTER_STATS_ONLY 0
 #define CONFIG_LIBYUV 0
-#define CONFIG_LOWBITDEPTH 0
+#define CONFIG_LOWBITDEPTH 1
+#define CONFIG_MAX_DECODE_PROFILE 2
 #define CONFIG_MISMATCH_DEBUG 0
-#define CONFIG_MSVS 0
 #define CONFIG_MULTITHREAD 1
+#define CONFIG_NORMAL_TILE_MODE 0
 #define CONFIG_OS_SUPPORT 1
 #define CONFIG_PIC 0
 #define CONFIG_RD_DEBUG 0
+#define CONFIG_REDUCED_ENCODER_BORDER 0
 #define CONFIG_RUNTIME_CPU_DETECT 1
 #define CONFIG_SHARED 0
+#define CONFIG_SHARP_SETTINGS 0
 #define CONFIG_SIZE_LIMIT 0
 #define CONFIG_SPATIAL_RESAMPLING 1
 #define CONFIG_STATIC 1
diff --git a/media/libaom/config/linux/x64/config/aom_dsp_rtcd.h b/media/libaom/config/linux/x64/config/aom_dsp_rtcd.h
index 494e10f98..2856d4ede 100644
--- a/media/libaom/config/linux/x64/config/aom_dsp_rtcd.h
+++ b/media/libaom/config/linux/x64/config/aom_dsp_rtcd.h
@@ -28,6 +28,7 @@ RTCD_EXTERN void (*aom_blend_a64_hmask)(uint8_t *dst, uint32_t dst_stride, const
 
 void aom_blend_a64_mask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subx, int suby);
 void aom_blend_a64_mask_sse4_1(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subx, int suby);
+void aom_blend_a64_mask_avx2(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subx, int suby);
 RTCD_EXTERN void (*aom_blend_a64_mask)(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subx, int suby);
 
 void aom_blend_a64_vmask_c(uint8_t *dst, uint32_t dst_stride, const uint8_t *src0, uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride, const uint8_t *mask, int w, int h);
@@ -1232,6 +1233,7 @@ void aom_highbd_v_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t y_stride, const ui
 
 void aom_lowbd_blend_a64_d16_mask_c(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subx, int suby, ConvolveParams *conv_params);
 void aom_lowbd_blend_a64_d16_mask_sse4_1(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subx, int suby, ConvolveParams *conv_params);
+void aom_lowbd_blend_a64_d16_mask_avx2(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subx, int suby, ConvolveParams *conv_params);
 RTCD_EXTERN void (*aom_lowbd_blend_a64_d16_mask)(uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0, uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride, const uint8_t *mask, uint32_t mask_stride, int w, int h, int subx, int suby, ConvolveParams *conv_params);
 
 void aom_lpf_horizontal_14_c(uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh);
@@ -1727,6 +1729,7 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_SSE4_1) aom_blend_a64_hmask = aom_blend_a64_hmask_sse4_1;
     aom_blend_a64_mask = aom_blend_a64_mask_c;
     if (flags & HAS_SSE4_1) aom_blend_a64_mask = aom_blend_a64_mask_sse4_1;
+    if (flags & HAS_AVX2) aom_blend_a64_mask = aom_blend_a64_mask_avx2;
     aom_blend_a64_vmask = aom_blend_a64_vmask_c;
     if (flags & HAS_SSE4_1) aom_blend_a64_vmask = aom_blend_a64_vmask_sse4_1;
     aom_convolve8_horiz = aom_convolve8_horiz_sse2;
@@ -1811,6 +1814,7 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_AVX2) aom_highbd_lpf_vertical_8_dual = aom_highbd_lpf_vertical_8_dual_avx2;
     aom_lowbd_blend_a64_d16_mask = aom_lowbd_blend_a64_d16_mask_c;
     if (flags & HAS_SSE4_1) aom_lowbd_blend_a64_d16_mask = aom_lowbd_blend_a64_d16_mask_sse4_1;
+    if (flags & HAS_AVX2) aom_lowbd_blend_a64_d16_mask = aom_lowbd_blend_a64_d16_mask_avx2;
     aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_c;
     if (flags & HAS_SSSE3) aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_ssse3;
     if (flags & HAS_AVX2) aom_paeth_predictor_16x16 = aom_paeth_predictor_16x16_avx2;
diff --git a/media/libaom/config/linux/x64/config/av1_rtcd.h b/media/libaom/config/linux/x64/config/av1_rtcd.h
index d3a2ac579..d27318208 100644
--- a/media/libaom/config/linux/x64/config/av1_rtcd.h
+++ b/media/libaom/config/linux/x64/config/av1_rtcd.h
@@ -1,6 +1,6 @@
 // This file is generated. Do not edit.
-#ifndef AOM_RTCD_H_
-#define AOM_RTCD_H_
+#ifndef AV1_RTCD_H_
+#define AV1_RTCD_H_
 
 #ifdef RTCD_C
 #define RTCD_EXTERN
@@ -58,10 +58,12 @@ RTCD_EXTERN void (*apply_selfguided_restoration)(const uint8_t *dat, int width,
 
 void av1_build_compound_diffwtd_mask_c(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w);
 void av1_build_compound_diffwtd_mask_sse4_1(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w);
+void av1_build_compound_diffwtd_mask_avx2(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w);
 RTCD_EXTERN void (*av1_build_compound_diffwtd_mask)(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w);
 
 void av1_build_compound_diffwtd_mask_d16_c(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const CONV_BUF_TYPE *src0, int src0_stride, const CONV_BUF_TYPE *src1, int src1_stride, int h, int w, ConvolveParams *conv_params, int bd);
 void av1_build_compound_diffwtd_mask_d16_sse4_1(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const CONV_BUF_TYPE *src0, int src0_stride, const CONV_BUF_TYPE *src1, int src1_stride, int h, int w, ConvolveParams *conv_params, int bd);
+void av1_build_compound_diffwtd_mask_d16_avx2(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const CONV_BUF_TYPE *src0, int src0_stride, const CONV_BUF_TYPE *src1, int src1_stride, int h, int w, ConvolveParams *conv_params, int bd);
 RTCD_EXTERN void (*av1_build_compound_diffwtd_mask_d16)(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const CONV_BUF_TYPE *src0, int src0_stride, const CONV_BUF_TYPE *src1, int src1_stride, int h, int w, ConvolveParams *conv_params, int bd);
 
 void av1_build_compound_diffwtd_mask_highbd_c(uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0, int src0_stride, const uint8_t *src1, int src1_stride, int h, int w, int bd);
@@ -173,6 +175,36 @@ void av1_highbd_dr_prediction_z2_c(uint16_t *dst, ptrdiff_t stride, int bw, int
 void av1_highbd_dr_prediction_z3_c(uint16_t *dst, ptrdiff_t stride, int bw, int bh, const uint16_t *above, const uint16_t *left, int upsample_left, int dx, int dy, int bd);
 #define av1_highbd_dr_prediction_z3 av1_highbd_dr_prediction_z3_c
 
+void av1_highbd_inv_txfm_add_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_sse4_1(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_avx2(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_16x16_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_16x16_sse4_1(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_16x16)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_16x8_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_16x8_sse4_1(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_16x8)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_32x32_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_32x32_sse4_1(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_32x32_avx2(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_32x32)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_4x4_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_4x4_sse4_1(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_4x4)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_8x16_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_8x16_sse4_1(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_8x16)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+
+void av1_highbd_inv_txfm_add_8x8_c(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+void av1_highbd_inv_txfm_add_8x8_sse4_1(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+RTCD_EXTERN void (*av1_highbd_inv_txfm_add_8x8)(const tran_low_t *dqcoeff, uint8_t *dst, int stride, const TxfmParam *txfm_param);
+
 void av1_highbd_iwht4x4_16_add_c(const tran_low_t *input, uint8_t *dest, int dest_stride, int bd);
 #define av1_highbd_iwht4x4_16_add av1_highbd_iwht4x4_16_add_c
 
@@ -209,8 +241,7 @@ void av1_highbd_wiener_convolve_add_src_avx2(const uint8_t *src, ptrdiff_t src_s
 RTCD_EXTERN void (*av1_highbd_wiener_convolve_add_src)(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, const ConvolveParams *conv_params, int bps);
 
 void av1_inv_txfm2d_add_16x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
-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);
+#define av1_inv_txfm2d_add_16x16 av1_inv_txfm2d_add_16x16_c
 
 void av1_inv_txfm2d_add_16x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
 #define av1_inv_txfm2d_add_16x32 av1_inv_txfm2d_add_16x32_c
@@ -228,8 +259,7 @@ void av1_inv_txfm2d_add_32x16_c(const int32_t *input, uint16_t *output, int stri
 #define av1_inv_txfm2d_add_32x16 av1_inv_txfm2d_add_32x16_c
 
 void av1_inv_txfm2d_add_32x32_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
-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);
+#define av1_inv_txfm2d_add_32x32 av1_inv_txfm2d_add_32x32_c
 
 void av1_inv_txfm2d_add_32x64_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
 #define av1_inv_txfm2d_add_32x64 av1_inv_txfm2d_add_32x64_c
@@ -254,8 +284,7 @@ void av1_inv_txfm2d_add_64x32_c(const int32_t *input, uint16_t *output, int stri
 #define av1_inv_txfm2d_add_64x32 av1_inv_txfm2d_add_64x32_c
 
 void av1_inv_txfm2d_add_64x64_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
-void av1_inv_txfm2d_add_64x64_sse4_1(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
-RTCD_EXTERN void (*av1_inv_txfm2d_add_64x64)(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
+#define av1_inv_txfm2d_add_64x64 av1_inv_txfm2d_add_64x64_c
 
 void av1_inv_txfm2d_add_8x16_c(const int32_t *input, uint16_t *output, int stride, TX_TYPE tx_type, int bd);
 #define av1_inv_txfm2d_add_8x16 av1_inv_txfm2d_add_8x16_c
@@ -295,18 +324,18 @@ void av1_jnt_convolve_y_sse2(const uint8_t *src, int src_stride, uint8_t *dst, i
 void av1_jnt_convolve_y_avx2(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
 RTCD_EXTERN void (*av1_jnt_convolve_y)(const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w, int h, const InterpFilterParams *filter_params_x, const InterpFilterParams *filter_params_y, const int subpel_x_q4, const int subpel_y_q4, ConvolveParams *conv_params);
 
-void av1_selfguided_restoration_c(const uint8_t *dgd8, int width, int height,
-                                  int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
-                                  int sgr_params_idx, int bit_depth, int highbd);
-void av1_selfguided_restoration_sse4_1(const uint8_t *dgd8, int width, int height,
-                                  int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
-                                  int sgr_params_idx, int bit_depth, int highbd);
-void av1_selfguided_restoration_avx2(const uint8_t *dgd8, int width, int height,
-                                  int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
-                                  int sgr_params_idx, int bit_depth, int highbd);
-RTCD_EXTERN void (*av1_selfguided_restoration)(const uint8_t *dgd8, int width, int height,
-                                  int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
-                                  int sgr_params_idx, int bit_depth, int highbd);
+int av1_selfguided_restoration_c(const uint8_t *dgd8, int width, int height,
+                                 int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
+                                 int sgr_params_idx, int bit_depth, int highbd);
+int av1_selfguided_restoration_sse4_1(const uint8_t *dgd8, int width, int height,
+                                 int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
+                                 int sgr_params_idx, int bit_depth, int highbd);
+int av1_selfguided_restoration_avx2(const uint8_t *dgd8, int width, int height,
+                                 int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
+                                 int sgr_params_idx, int bit_depth, int highbd);
+RTCD_EXTERN int (*av1_selfguided_restoration)(const uint8_t *dgd8, int width, int height,
+                                 int dgd_stride, int32_t *flt0, int32_t *flt1, int flt_stride,
+                                 int sgr_params_idx, int bit_depth, int highbd);
 
 void av1_upsample_intra_edge_c(uint8_t *p, int sz);
 void av1_upsample_intra_edge_sse4_1(uint8_t *p, int sz);
@@ -398,7 +427,7 @@ cfl_subtract_average_fn get_subtract_average_fn_sse2(TX_SIZE tx_size);
 cfl_subtract_average_fn get_subtract_average_fn_avx2(TX_SIZE tx_size);
 RTCD_EXTERN cfl_subtract_average_fn (*get_subtract_average_fn)(TX_SIZE tx_size);
 
-void aom_rtcd(void);
+void av1_rtcd(void);
 
 #ifdef RTCD_C
 #include "aom_ports/x86.h"
@@ -413,8 +442,10 @@ static void setup_rtcd_internal(void)
     if (flags & HAS_AVX2) apply_selfguided_restoration = apply_selfguided_restoration_avx2;
     av1_build_compound_diffwtd_mask = av1_build_compound_diffwtd_mask_c;
     if (flags & HAS_SSE4_1) av1_build_compound_diffwtd_mask = av1_build_compound_diffwtd_mask_sse4_1;
+    if (flags & HAS_AVX2) av1_build_compound_diffwtd_mask = av1_build_compound_diffwtd_mask_avx2;
     av1_build_compound_diffwtd_mask_d16 = av1_build_compound_diffwtd_mask_d16_c;
     if (flags & HAS_SSE4_1) av1_build_compound_diffwtd_mask_d16 = av1_build_compound_diffwtd_mask_d16_sse4_1;
+    if (flags & HAS_AVX2) av1_build_compound_diffwtd_mask_d16 = av1_build_compound_diffwtd_mask_d16_avx2;
     av1_build_compound_diffwtd_mask_highbd = av1_build_compound_diffwtd_mask_highbd_c;
     if (flags & HAS_SSSE3) av1_build_compound_diffwtd_mask_highbd = av1_build_compound_diffwtd_mask_highbd_ssse3;
     if (flags & HAS_AVX2) av1_build_compound_diffwtd_mask_highbd = av1_build_compound_diffwtd_mask_highbd_avx2;
@@ -451,6 +482,22 @@ static void setup_rtcd_internal(void)
     av1_highbd_convolve_y_sr = av1_highbd_convolve_y_sr_c;
     if (flags & HAS_SSSE3) av1_highbd_convolve_y_sr = av1_highbd_convolve_y_sr_ssse3;
     if (flags & HAS_AVX2) av1_highbd_convolve_y_sr = av1_highbd_convolve_y_sr_avx2;
+    av1_highbd_inv_txfm_add = av1_highbd_inv_txfm_add_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add = av1_highbd_inv_txfm_add_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_inv_txfm_add = av1_highbd_inv_txfm_add_avx2;
+    av1_highbd_inv_txfm_add_16x16 = av1_highbd_inv_txfm_add_16x16_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_16x16 = av1_highbd_inv_txfm_add_16x16_sse4_1;
+    av1_highbd_inv_txfm_add_16x8 = av1_highbd_inv_txfm_add_16x8_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_16x8 = av1_highbd_inv_txfm_add_16x8_sse4_1;
+    av1_highbd_inv_txfm_add_32x32 = av1_highbd_inv_txfm_add_32x32_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_32x32 = av1_highbd_inv_txfm_add_32x32_sse4_1;
+    if (flags & HAS_AVX2) av1_highbd_inv_txfm_add_32x32 = av1_highbd_inv_txfm_add_32x32_avx2;
+    av1_highbd_inv_txfm_add_4x4 = av1_highbd_inv_txfm_add_4x4_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_4x4 = av1_highbd_inv_txfm_add_4x4_sse4_1;
+    av1_highbd_inv_txfm_add_8x16 = av1_highbd_inv_txfm_add_8x16_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_8x16 = av1_highbd_inv_txfm_add_8x16_sse4_1;
+    av1_highbd_inv_txfm_add_8x8 = av1_highbd_inv_txfm_add_8x8_c;
+    if (flags & HAS_SSE4_1) av1_highbd_inv_txfm_add_8x8 = av1_highbd_inv_txfm_add_8x8_sse4_1;
     av1_highbd_jnt_convolve_2d = av1_highbd_jnt_convolve_2d_c;
     if (flags & HAS_SSE4_1) av1_highbd_jnt_convolve_2d = av1_highbd_jnt_convolve_2d_sse4_1;
     if (flags & HAS_AVX2) av1_highbd_jnt_convolve_2d = av1_highbd_jnt_convolve_2d_avx2;
@@ -468,14 +515,8 @@ static void setup_rtcd_internal(void)
     av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_c;
     if (flags & HAS_SSSE3) av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_ssse3;
     if (flags & HAS_AVX2) av1_highbd_wiener_convolve_add_src = av1_highbd_wiener_convolve_add_src_avx2;
-    av1_inv_txfm2d_add_16x16 = av1_inv_txfm2d_add_16x16_c;
-    if (flags & HAS_SSE4_1) av1_inv_txfm2d_add_16x16 = av1_inv_txfm2d_add_16x16_sse4_1;
-    av1_inv_txfm2d_add_32x32 = av1_inv_txfm2d_add_32x32_c;
-    if (flags & HAS_AVX2) av1_inv_txfm2d_add_32x32 = av1_inv_txfm2d_add_32x32_avx2;
     av1_inv_txfm2d_add_4x4 = av1_inv_txfm2d_add_4x4_c;
     if (flags & HAS_SSE4_1) av1_inv_txfm2d_add_4x4 = av1_inv_txfm2d_add_4x4_sse4_1;
-    av1_inv_txfm2d_add_64x64 = av1_inv_txfm2d_add_64x64_c;
-    if (flags & HAS_SSE4_1) av1_inv_txfm2d_add_64x64 = av1_inv_txfm2d_add_64x64_sse4_1;
     av1_inv_txfm2d_add_8x8 = av1_inv_txfm2d_add_8x8_c;
     if (flags & HAS_SSE4_1) av1_inv_txfm2d_add_8x8 = av1_inv_txfm2d_add_8x8_sse4_1;
     av1_inv_txfm_add = av1_inv_txfm_add_c;