Update libaom to commit ID 1e227d41f0616de9548a673a83a21ef990b62591

11 files changed, 1431 insertions, 240 deletions

diff --git a/third_party/aom/av1/encoder/x86/av1_fwd_txfm1d_sse4.c b/third_party/aom/av1/encoder/x86/av1_fwd_txfm1d_sse4.c
index c71f2e74c..07615543c 100644
--- a/third_party/aom/av1/encoder/x86/av1_fwd_txfm1d_sse4.c
+++ b/third_party/aom/av1/encoder/x86/av1_fwd_txfm1d_sse4.c
@@ -395,7 +395,8 @@ void av1_fadst4_new_sse4_1(const __m128i *input, __m128i *output,
 }
 
 void av1_fdct64_new_sse4_1(const __m128i *input, __m128i *output,
-                           int8_t cos_bit) {
+                           int8_t cos_bit, const int instride,
+                           const int outstride) {
   const int32_t *cospi = cospi_arr(cos_bit);
   const __m128i __rounding = _mm_set1_epi32(1 << (cos_bit - 1));
 
@@ -480,70 +481,70 @@ void av1_fdct64_new_sse4_1(const __m128i *input, __m128i *output,
 
   // stage 1
   __m128i x1[64];
-  x1[0] = _mm_add_epi32(input[0], input[63]);
-  x1[63] = _mm_sub_epi32(input[0], input[63]);
-  x1[1] = _mm_add_epi32(input[1], input[62]);
-  x1[62] = _mm_sub_epi32(input[1], input[62]);
-  x1[2] = _mm_add_epi32(input[2], input[61]);
-  x1[61] = _mm_sub_epi32(input[2], input[61]);
-  x1[3] = _mm_add_epi32(input[3], input[60]);
-  x1[60] = _mm_sub_epi32(input[3], input[60]);
-  x1[4] = _mm_add_epi32(input[4], input[59]);
-  x1[59] = _mm_sub_epi32(input[4], input[59]);
-  x1[5] = _mm_add_epi32(input[5], input[58]);
-  x1[58] = _mm_sub_epi32(input[5], input[58]);
-  x1[6] = _mm_add_epi32(input[6], input[57]);
-  x1[57] = _mm_sub_epi32(input[6], input[57]);
-  x1[7] = _mm_add_epi32(input[7], input[56]);
-  x1[56] = _mm_sub_epi32(input[7], input[56]);
-  x1[8] = _mm_add_epi32(input[8], input[55]);
-  x1[55] = _mm_sub_epi32(input[8], input[55]);
-  x1[9] = _mm_add_epi32(input[9], input[54]);
-  x1[54] = _mm_sub_epi32(input[9], input[54]);
-  x1[10] = _mm_add_epi32(input[10], input[53]);
-  x1[53] = _mm_sub_epi32(input[10], input[53]);
-  x1[11] = _mm_add_epi32(input[11], input[52]);
-  x1[52] = _mm_sub_epi32(input[11], input[52]);
-  x1[12] = _mm_add_epi32(input[12], input[51]);
-  x1[51] = _mm_sub_epi32(input[12], input[51]);
-  x1[13] = _mm_add_epi32(input[13], input[50]);
-  x1[50] = _mm_sub_epi32(input[13], input[50]);
-  x1[14] = _mm_add_epi32(input[14], input[49]);
-  x1[49] = _mm_sub_epi32(input[14], input[49]);
-  x1[15] = _mm_add_epi32(input[15], input[48]);
-  x1[48] = _mm_sub_epi32(input[15], input[48]);
-  x1[16] = _mm_add_epi32(input[16], input[47]);
-  x1[47] = _mm_sub_epi32(input[16], input[47]);
-  x1[17] = _mm_add_epi32(input[17], input[46]);
-  x1[46] = _mm_sub_epi32(input[17], input[46]);
-  x1[18] = _mm_add_epi32(input[18], input[45]);
-  x1[45] = _mm_sub_epi32(input[18], input[45]);
-  x1[19] = _mm_add_epi32(input[19], input[44]);
-  x1[44] = _mm_sub_epi32(input[19], input[44]);
-  x1[20] = _mm_add_epi32(input[20], input[43]);
-  x1[43] = _mm_sub_epi32(input[20], input[43]);
-  x1[21] = _mm_add_epi32(input[21], input[42]);
-  x1[42] = _mm_sub_epi32(input[21], input[42]);
-  x1[22] = _mm_add_epi32(input[22], input[41]);
-  x1[41] = _mm_sub_epi32(input[22], input[41]);
-  x1[23] = _mm_add_epi32(input[23], input[40]);
-  x1[40] = _mm_sub_epi32(input[23], input[40]);
-  x1[24] = _mm_add_epi32(input[24], input[39]);
-  x1[39] = _mm_sub_epi32(input[24], input[39]);
-  x1[25] = _mm_add_epi32(input[25], input[38]);
-  x1[38] = _mm_sub_epi32(input[25], input[38]);
-  x1[26] = _mm_add_epi32(input[26], input[37]);
-  x1[37] = _mm_sub_epi32(input[26], input[37]);
-  x1[27] = _mm_add_epi32(input[27], input[36]);
-  x1[36] = _mm_sub_epi32(input[27], input[36]);
-  x1[28] = _mm_add_epi32(input[28], input[35]);
-  x1[35] = _mm_sub_epi32(input[28], input[35]);
-  x1[29] = _mm_add_epi32(input[29], input[34]);
-  x1[34] = _mm_sub_epi32(input[29], input[34]);
-  x1[30] = _mm_add_epi32(input[30], input[33]);
-  x1[33] = _mm_sub_epi32(input[30], input[33]);
-  x1[31] = _mm_add_epi32(input[31], input[32]);
-  x1[32] = _mm_sub_epi32(input[31], input[32]);
+  x1[0] = _mm_add_epi32(input[0 * instride], input[63 * instride]);
+  x1[63] = _mm_sub_epi32(input[0 * instride], input[63 * instride]);
+  x1[1] = _mm_add_epi32(input[1 * instride], input[62 * instride]);
+  x1[62] = _mm_sub_epi32(input[1 * instride], input[62 * instride]);
+  x1[2] = _mm_add_epi32(input[2 * instride], input[61 * instride]);
+  x1[61] = _mm_sub_epi32(input[2 * instride], input[61 * instride]);
+  x1[3] = _mm_add_epi32(input[3 * instride], input[60 * instride]);
+  x1[60] = _mm_sub_epi32(input[3 * instride], input[60 * instride]);
+  x1[4] = _mm_add_epi32(input[4 * instride], input[59 * instride]);
+  x1[59] = _mm_sub_epi32(input[4 * instride], input[59 * instride]);
+  x1[5] = _mm_add_epi32(input[5 * instride], input[58 * instride]);
+  x1[58] = _mm_sub_epi32(input[5 * instride], input[58 * instride]);
+  x1[6] = _mm_add_epi32(input[6 * instride], input[57 * instride]);
+  x1[57] = _mm_sub_epi32(input[6 * instride], input[57 * instride]);
+  x1[7] = _mm_add_epi32(input[7 * instride], input[56 * instride]);
+  x1[56] = _mm_sub_epi32(input[7 * instride], input[56 * instride]);
+  x1[8] = _mm_add_epi32(input[8 * instride], input[55 * instride]);
+  x1[55] = _mm_sub_epi32(input[8 * instride], input[55 * instride]);
+  x1[9] = _mm_add_epi32(input[9 * instride], input[54 * instride]);
+  x1[54] = _mm_sub_epi32(input[9 * instride], input[54 * instride]);
+  x1[10] = _mm_add_epi32(input[10 * instride], input[53 * instride]);
+  x1[53] = _mm_sub_epi32(input[10 * instride], input[53 * instride]);
+  x1[11] = _mm_add_epi32(input[11 * instride], input[52 * instride]);
+  x1[52] = _mm_sub_epi32(input[11 * instride], input[52 * instride]);
+  x1[12] = _mm_add_epi32(input[12 * instride], input[51 * instride]);
+  x1[51] = _mm_sub_epi32(input[12 * instride], input[51 * instride]);
+  x1[13] = _mm_add_epi32(input[13 * instride], input[50 * instride]);
+  x1[50] = _mm_sub_epi32(input[13 * instride], input[50 * instride]);
+  x1[14] = _mm_add_epi32(input[14 * instride], input[49 * instride]);
+  x1[49] = _mm_sub_epi32(input[14 * instride], input[49 * instride]);
+  x1[15] = _mm_add_epi32(input[15 * instride], input[48 * instride]);
+  x1[48] = _mm_sub_epi32(input[15 * instride], input[48 * instride]);
+  x1[16] = _mm_add_epi32(input[16 * instride], input[47 * instride]);
+  x1[47] = _mm_sub_epi32(input[16 * instride], input[47 * instride]);
+  x1[17] = _mm_add_epi32(input[17 * instride], input[46 * instride]);
+  x1[46] = _mm_sub_epi32(input[17 * instride], input[46 * instride]);
+  x1[18] = _mm_add_epi32(input[18 * instride], input[45 * instride]);
+  x1[45] = _mm_sub_epi32(input[18 * instride], input[45 * instride]);
+  x1[19] = _mm_add_epi32(input[19 * instride], input[44 * instride]);
+  x1[44] = _mm_sub_epi32(input[19 * instride], input[44 * instride]);
+  x1[20] = _mm_add_epi32(input[20 * instride], input[43 * instride]);
+  x1[43] = _mm_sub_epi32(input[20 * instride], input[43 * instride]);
+  x1[21] = _mm_add_epi32(input[21 * instride], input[42 * instride]);
+  x1[42] = _mm_sub_epi32(input[21 * instride], input[42 * instride]);
+  x1[22] = _mm_add_epi32(input[22 * instride], input[41 * instride]);
+  x1[41] = _mm_sub_epi32(input[22 * instride], input[41 * instride]);
+  x1[23] = _mm_add_epi32(input[23 * instride], input[40 * instride]);
+  x1[40] = _mm_sub_epi32(input[23 * instride], input[40 * instride]);
+  x1[24] = _mm_add_epi32(input[24 * instride], input[39 * instride]);
+  x1[39] = _mm_sub_epi32(input[24 * instride], input[39 * instride]);
+  x1[25] = _mm_add_epi32(input[25 * instride], input[38 * instride]);
+  x1[38] = _mm_sub_epi32(input[25 * instride], input[38 * instride]);
+  x1[26] = _mm_add_epi32(input[26 * instride], input[37 * instride]);
+  x1[37] = _mm_sub_epi32(input[26 * instride], input[37 * instride]);
+  x1[27] = _mm_add_epi32(input[27 * instride], input[36 * instride]);
+  x1[36] = _mm_sub_epi32(input[27 * instride], input[36 * instride]);
+  x1[28] = _mm_add_epi32(input[28 * instride], input[35 * instride]);
+  x1[35] = _mm_sub_epi32(input[28 * instride], input[35 * instride]);
+  x1[29] = _mm_add_epi32(input[29 * instride], input[34 * instride]);
+  x1[34] = _mm_sub_epi32(input[29 * instride], input[34 * instride]);
+  x1[30] = _mm_add_epi32(input[30 * instride], input[33 * instride]);
+  x1[33] = _mm_sub_epi32(input[30 * instride], input[33 * instride]);
+  x1[31] = _mm_add_epi32(input[31 * instride], input[32 * instride]);
+  x1[32] = _mm_sub_epi32(input[31 * instride], input[32 * instride]);
 
   // stage 2
   __m128i x2[64];
@@ -1149,68 +1150,68 @@ void av1_fdct64_new_sse4_1(const __m128i *input, __m128i *output,
                           x10[48], __rounding, cos_bit);
 
   // stage 11
-  output[0] = x10[0];
-  output[1] = x10[32];
-  output[2] = x10[16];
-  output[3] = x10[48];
-  output[4] = x10[8];
-  output[5] = x10[40];
-  output[6] = x10[24];
-  output[7] = x10[56];
-  output[8] = x10[4];
-  output[9] = x10[36];
-  output[10] = x10[20];
-  output[11] = x10[52];
-  output[12] = x10[12];
-  output[13] = x10[44];
-  output[14] = x10[28];
-  output[15] = x10[60];
-  output[16] = x10[2];
-  output[17] = x10[34];
-  output[18] = x10[18];
-  output[19] = x10[50];
-  output[20] = x10[10];
-  output[21] = x10[42];
-  output[22] = x10[26];
-  output[23] = x10[58];
-  output[24] = x10[6];
-  output[25] = x10[38];
-  output[26] = x10[22];
-  output[27] = x10[54];
-  output[28] = x10[14];
-  output[29] = x10[46];
-  output[30] = x10[30];
-  output[31] = x10[62];
-  output[32] = x10[1];
-  output[33] = x10[33];
-  output[34] = x10[17];
-  output[35] = x10[49];
-  output[36] = x10[9];
-  output[37] = x10[41];
-  output[38] = x10[25];
-  output[39] = x10[57];
-  output[40] = x10[5];
-  output[41] = x10[37];
-  output[42] = x10[21];
-  output[43] = x10[53];
-  output[44] = x10[13];
-  output[45] = x10[45];
-  output[46] = x10[29];
-  output[47] = x10[61];
-  output[48] = x10[3];
-  output[49] = x10[35];
-  output[50] = x10[19];
-  output[51] = x10[51];
-  output[52] = x10[11];
-  output[53] = x10[43];
-  output[54] = x10[27];
-  output[55] = x10[59];
-  output[56] = x10[7];
-  output[57] = x10[39];
-  output[58] = x10[23];
-  output[59] = x10[55];
-  output[60] = x10[15];
-  output[61] = x10[47];
-  output[62] = x10[31];
-  output[63] = x10[63];
+  output[0 * outstride] = x10[0];
+  output[1 * outstride] = x10[32];
+  output[2 * outstride] = x10[16];
+  output[3 * outstride] = x10[48];
+  output[4 * outstride] = x10[8];
+  output[5 * outstride] = x10[40];
+  output[6 * outstride] = x10[24];
+  output[7 * outstride] = x10[56];
+  output[8 * outstride] = x10[4];
+  output[9 * outstride] = x10[36];
+  output[10 * outstride] = x10[20];
+  output[11 * outstride] = x10[52];
+  output[12 * outstride] = x10[12];
+  output[13 * outstride] = x10[44];
+  output[14 * outstride] = x10[28];
+  output[15 * outstride] = x10[60];
+  output[16 * outstride] = x10[2];
+  output[17 * outstride] = x10[34];
+  output[18 * outstride] = x10[18];
+  output[19 * outstride] = x10[50];
+  output[20 * outstride] = x10[10];
+  output[21 * outstride] = x10[42];
+  output[22 * outstride] = x10[26];
+  output[23 * outstride] = x10[58];
+  output[24 * outstride] = x10[6];
+  output[25 * outstride] = x10[38];
+  output[26 * outstride] = x10[22];
+  output[27 * outstride] = x10[54];
+  output[28 * outstride] = x10[14];
+  output[29 * outstride] = x10[46];
+  output[30 * outstride] = x10[30];
+  output[31 * outstride] = x10[62];
+  output[32 * outstride] = x10[1];
+  output[33 * outstride] = x10[33];
+  output[34 * outstride] = x10[17];
+  output[35 * outstride] = x10[49];
+  output[36 * outstride] = x10[9];
+  output[37 * outstride] = x10[41];
+  output[38 * outstride] = x10[25];
+  output[39 * outstride] = x10[57];
+  output[40 * outstride] = x10[5];
+  output[41 * outstride] = x10[37];
+  output[42 * outstride] = x10[21];
+  output[43 * outstride] = x10[53];
+  output[44 * outstride] = x10[13];
+  output[45 * outstride] = x10[45];
+  output[46 * outstride] = x10[29];
+  output[47 * outstride] = x10[61];
+  output[48 * outstride] = x10[3];
+  output[49 * outstride] = x10[35];
+  output[50 * outstride] = x10[19];
+  output[51 * outstride] = x10[51];
+  output[52 * outstride] = x10[11];
+  output[53 * outstride] = x10[43];
+  output[54 * outstride] = x10[27];
+  output[55 * outstride] = x10[59];
+  output[56 * outstride] = x10[7];
+  output[57 * outstride] = x10[39];
+  output[58 * outstride] = x10[23];
+  output[59 * outstride] = x10[55];
+  output[60 * outstride] = x10[15];
+  output[61 * outstride] = x10[47];
+  output[62 * outstride] = x10[31];
+  output[63 * outstride] = x10[63];
 }
diff --git a/third_party/aom/av1/encoder/x86/av1_fwd_txfm2d_sse4.c b/third_party/aom/av1/encoder/x86/av1_fwd_txfm2d_sse4.c
index abb95f31e..8ec0256eb 100644
--- a/third_party/aom/av1/encoder/x86/av1_fwd_txfm2d_sse4.c
+++ b/third_party/aom/av1/encoder/x86/av1_fwd_txfm2d_sse4.c
@@ -14,6 +14,7 @@
 #include "av1/common/enums.h"
 #include "av1/common/av1_txfm.h"
 #include "av1/common/x86/av1_txfm_sse2.h"
+#include "av1/common/x86/highbd_txfm_utility_sse4.h"
 #include "av1/encoder/av1_fwd_txfm1d_cfg.h"
 #include "av1/encoder/x86/av1_txfm1d_sse4.h"
 #include "av1/encoder/x86/av1_fwd_txfm_sse2.h"
@@ -52,9 +53,22 @@ static void fdct32_new_sse4_1(const __m128i *input, __m128i *output,
   }
 }
 
+static void fdct64_new_sse4_1(const __m128i *input, __m128i *output,
+                              const int8_t cos_bit, const int8_t *stage_range) {
+  const int txfm_size = 64;
+  const int num_per_128 = 4;
+  int col_num = txfm_size / num_per_128;
+  (void)stage_range;
+  for (int col = 0; col < col_num; col++) {
+    av1_fdct64_new_sse4_1((input + col), (output + col), cos_bit, col_num,
+                          col_num);
+  }
+}
+
 static INLINE TxfmFuncSSE2 fwd_txfm_type_to_func(TXFM_TYPE txfm_type) {
   switch (txfm_type) {
     case TXFM_TYPE_DCT32: return fdct32_new_sse4_1; break;
+    case TXFM_TYPE_DCT64: return fdct64_new_sse4_1; break;
     default: assert(0);
   }
   return NULL;
@@ -95,6 +109,42 @@ static INLINE void fwd_txfm2d_sse4_1(const int16_t *input, int32_t *output,
   transpose_32(txfm_size, buf_128, out_128);
 }
 
+static INLINE void fwd_txfm2d_64x64_sse4_1(const int16_t *input,
+                                           int32_t *output, const int stride,
+                                           const TXFM_2D_FLIP_CFG *cfg,
+                                           int32_t *txfm_buf) {
+  assert(cfg->tx_size < TX_SIZES);
+  const int txfm_size = tx_size_wide[cfg->tx_size];
+  const int8_t *shift = cfg->shift;
+  const int8_t *stage_range_col = cfg->stage_range_col;
+  const int8_t cos_bit_col = cfg->cos_bit_col;
+  const int8_t cos_bit_row = cfg->cos_bit_row;
+  const TxfmFuncSSE2 txfm_func_col = fwd_txfm_type_to_func(cfg->txfm_type_col);
+  __m128i *buf_128 = (__m128i *)txfm_buf;
+  __m128i *out_128 = (__m128i *)output;
+
+  const int num_per_128 = 4;
+  int txfm2d_size_128 = txfm_size * txfm_size / num_per_128;
+  int col_num = txfm_size / num_per_128;
+
+  int16_array_with_stride_to_int32_array_without_stride(input, stride, output,
+                                                        txfm_size);
+  /*col wise transform*/
+  txfm_func_col(out_128, buf_128, cos_bit_col, stage_range_col);
+  av1_round_shift_array_32_sse4_1(buf_128, out_128, txfm2d_size_128, -shift[1]);
+  transpose_32(txfm_size, out_128, buf_128);
+
+  /*row wise transform*/
+  for (int col = 0; col < (col_num >> 1); col++) {
+    av1_fdct64_new_sse4_1((buf_128 + col), (out_128 + col), cos_bit_row,
+                          col_num, (col_num >> 1));
+  }
+
+  txfm2d_size_128 = (col_num >> 1) * (txfm_size >> 1);
+  av1_round_shift_array_32_sse4_1(out_128, buf_128, txfm2d_size_128, -shift[2]);
+  transpose_32x32(buf_128, out_128);
+}
+
 void av1_fwd_txfm2d_32x32_sse4_1(const int16_t *input, int32_t *output,
                                  int stride, TX_TYPE tx_type, int bd) {
   DECLARE_ALIGNED(16, int32_t, txfm_buf[1024]);
@@ -104,6 +154,15 @@ void av1_fwd_txfm2d_32x32_sse4_1(const int16_t *input, int32_t *output,
   fwd_txfm2d_sse4_1(input, output, stride, &cfg, txfm_buf);
 }
 
+void av1_fwd_txfm2d_64x64_sse4_1(const int16_t *input, int32_t *output,
+                                 int stride, TX_TYPE tx_type, int bd) {
+  DECLARE_ALIGNED(16, int32_t, txfm_buf[4096]);
+  TXFM_2D_FLIP_CFG cfg;
+  av1_get_fwd_txfm_cfg(tx_type, TX_64X64, &cfg);
+  (void)bd;
+  fwd_txfm2d_64x64_sse4_1(input, output, stride, &cfg, txfm_buf);
+}
+
 static INLINE void transpose_32_4x4x2(int stride, const __m128i *inputA,
                                       const __m128i *inputB, __m128i *output) {
   __m128i temp0 = _mm_unpacklo_epi32(inputA[0], inputA[2]);
@@ -162,8 +221,8 @@ static void lowbd_fwd_txfm2d_64x64_sse4_1(const int16_t *input, int32_t *output,
       bufA[j] = _mm_cvtepi16_epi32(buf[j]);
       bufB[j] = _mm_cvtepi16_epi32(_mm_unpackhi_epi64(buf[j], buf[j]));
     }
-    av1_fdct64_new_sse4_1(bufA, bufA, cos_bit_row);
-    av1_fdct64_new_sse4_1(bufB, bufB, cos_bit_row);
+    av1_fdct64_new_sse4_1(bufA, bufA, cos_bit_row, 1, 1);
+    av1_fdct64_new_sse4_1(bufB, bufB, cos_bit_row, 1, 1);
     av1_round_shift_array_32_sse4_1(bufA, bufA, 32, -shift[2]);
     av1_round_shift_array_32_sse4_1(bufB, bufB, 32, -shift[2]);
 
@@ -209,10 +268,10 @@ static void lowbd_fwd_txfm2d_64x32_sse4_1(const int16_t *input, int32_t *output,
       bufA[j] = _mm_cvtepi16_epi32(buf[j]);
       bufB[j] = _mm_cvtepi16_epi32(_mm_unpackhi_epi64(buf[j], buf[j]));
     }
-    av1_fdct64_new_sse4_1(bufA, bufA, cos_bit_row);
-    av1_fdct64_new_sse4_1(bufB, bufB, cos_bit_row);
-    av1_round_shift_rect_array_32_sse4_1(bufA, bufA, 32, -shift[2]);
-    av1_round_shift_rect_array_32_sse4_1(bufB, bufB, 32, -shift[2]);
+    av1_fdct64_new_sse4_1(bufA, bufA, cos_bit_row, 1, 1);
+    av1_fdct64_new_sse4_1(bufB, bufB, cos_bit_row, 1, 1);
+    av1_round_shift_rect_array_32_sse4_1(bufA, bufA, 32, -shift[2], NewSqrt2);
+    av1_round_shift_rect_array_32_sse4_1(bufB, bufB, 32, -shift[2], NewSqrt2);
 
     int32_t *output8 = output + 8 * 32 * i;
     for (int j = 0; j < width_div8; ++j) {
@@ -260,8 +319,8 @@ static void lowbd_fwd_txfm2d_32x64_sse4_1(const int16_t *input, int32_t *output,
     }
     av1_fdct32_new_sse4_1(bufA, bufA, cos_bit_row);
     av1_fdct32_new_sse4_1(bufB, bufB, cos_bit_row);
-    av1_round_shift_rect_array_32_sse4_1(bufA, bufA, 32, -shift[2]);
-    av1_round_shift_rect_array_32_sse4_1(bufB, bufB, 32, -shift[2]);
+    av1_round_shift_rect_array_32_sse4_1(bufA, bufA, 32, -shift[2], NewSqrt2);
+    av1_round_shift_rect_array_32_sse4_1(bufB, bufB, 32, -shift[2], NewSqrt2);
 
     int32_t *output8 = output + 8 * 32 * i;
     for (int j = 0; j < (32 / 4); ++j) {
diff --git a/third_party/aom/av1/encoder/x86/av1_fwd_txfm_avx2.h b/third_party/aom/av1/encoder/x86/av1_fwd_txfm_avx2.h
index c582ca0e3..38707137c 100644
--- a/third_party/aom/av1/encoder/x86/av1_fwd_txfm_avx2.h
+++ b/third_party/aom/av1/encoder/x86/av1_fwd_txfm_avx2.h
@@ -9,8 +9,8 @@
  * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
  */
 
-#ifndef AV1_FWD_TXFM_AVX2_H_
-#define AV1_FWD_TXFM_AVX2_H_
+#ifndef AOM_AV1_ENCODER_X86_AV1_FWD_TXFM_AVX2_H_
+#define AOM_AV1_ENCODER_X86_AV1_FWD_TXFM_AVX2_H_
 #include <immintrin.h>
 
 static INLINE __m256i av1_round_shift_32_avx2(__m256i vec, int bit) {
@@ -100,4 +100,4 @@ static INLINE void btf_32_avx2_type1_new(const __m256i ww0, const __m256i ww1,
   *in1 = _mm256_srai_epi32(temp1, cos_bit);
 }
 
-#endif  // AV1_FWD_TXFM_AVX2_H_
+#endif  // AOM_AV1_ENCODER_X86_AV1_FWD_TXFM_AVX2_H_
diff --git a/third_party/aom/av1/encoder/x86/av1_fwd_txfm_sse2.h b/third_party/aom/av1/encoder/x86/av1_fwd_txfm_sse2.h
index aa14d3ade..99a6b9082 100644
--- a/third_party/aom/av1/encoder/x86/av1_fwd_txfm_sse2.h
+++ b/third_party/aom/av1/encoder/x86/av1_fwd_txfm_sse2.h
@@ -8,8 +8,8 @@
  * Media Patent License 1.0 was not distributed with this source code in the
  * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
  */
-#ifndef AV1_COMMON_X86_AV1_FWD_TXFM_SSE2_H_
-#define AV1_COMMON_X86_AV1_FWD_TXFM_SSE2_H_
+#ifndef AOM_AV1_ENCODER_X86_AV1_FWD_TXFM_SSE2_H_
+#define AOM_AV1_ENCODER_X86_AV1_FWD_TXFM_SSE2_H_
 
 #include <immintrin.h>
 
@@ -114,4 +114,4 @@ static const transform_1d_sse2 col_txfm8x32_arr[TX_TYPES] = {
 }
 #endif
 
-#endif  // AV1_COMMON_X86_AV1_FWD_TXFM_SSE2_H_
+#endif  // AOM_AV1_ENCODER_X86_AV1_FWD_TXFM_SSE2_H_
diff --git a/third_party/aom/av1/encoder/x86/av1_txfm1d_sse4.h b/third_party/aom/av1/encoder/x86/av1_txfm1d_sse4.h
index 0adefecdb..6df2a8bdb 100644
--- a/third_party/aom/av1/encoder/x86/av1_txfm1d_sse4.h
+++ b/third_party/aom/av1/encoder/x86/av1_txfm1d_sse4.h
@@ -9,8 +9,8 @@
  * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
  */
 
-#ifndef AV1_TXMF1D_SSE2_H_
-#define AV1_TXMF1D_SSE2_H_
+#ifndef AOM_AV1_ENCODER_X86_AV1_TXFM1D_SSE4_H_
+#define AOM_AV1_ENCODER_X86_AV1_TXFM1D_SSE4_H_
 
 #include <smmintrin.h>
 #include "av1/common/av1_txfm.h"
@@ -29,7 +29,8 @@ void av1_fdct16_new_sse4_1(const __m128i *input, __m128i *output,
 void av1_fdct32_new_sse4_1(const __m128i *input, __m128i *output,
                            int8_t cos_bit);
 void av1_fdct64_new_sse4_1(const __m128i *input, __m128i *output,
-                           int8_t cos_bit);
+                           int8_t cos_bit, const int instride,
+                           const int outstride);
 
 void av1_fadst4_new_sse4_1(const __m128i *input, __m128i *output,
                            const int8_t cos_bit, const int8_t *stage_range);
@@ -138,4 +139,4 @@ static INLINE void transpose_32(int txfm_size, const __m128i *input,
 }
 #endif
 
-#endif  // AV1_TXMF1D_SSE2_H_
+#endif  // AOM_AV1_ENCODER_X86_AV1_TXFM1D_SSE4_H_
diff --git a/third_party/aom/av1/encoder/x86/encodetxb_avx2.c b/third_party/aom/av1/encoder/x86/encodetxb_avx2.c
new file mode 100644
index 000000000..7642f57d1
--- /dev/null
+++ b/third_party/aom/av1/encoder/x86/encodetxb_avx2.c
@@ -0,0 +1,130 @@
+/*
+ * Copyright (c) 2018, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <emmintrin.h>  // SSE2
+#include <smmintrin.h>  /* SSE4.1 */
+#include <immintrin.h>  /* AVX2 */
+
+#include "aom/aom_integer.h"
+#include "aom_dsp/x86/mem_sse2.h"
+#include "av1/common/onyxc_int.h"
+#include "av1/common/txb_common.h"
+#include "aom_dsp/x86/synonyms.h"
+#include "aom_dsp/x86/synonyms_avx2.h"
+
+void av1_txb_init_levels_avx2(const tran_low_t *const coeff, const int width,
+                              const int height, uint8_t *const levels) {
+  const int stride = width + TX_PAD_HOR;
+  const __m256i y_zeros = _mm256_setzero_si256();
+
+  const int32_t pre_len = sizeof(*levels) * TX_PAD_TOP * stride;
+  uint8_t *pre_buf = levels - TX_PAD_TOP * stride;
+  uint8_t *pre_buf_end = pre_buf + pre_len;
+  do {
+    yy_storeu_256(pre_buf, y_zeros);
+    pre_buf += 32;
+  } while (pre_buf < pre_buf_end);
+
+  const int32_t bottom_len = sizeof(*levels) * (TX_PAD_BOTTOM * stride);
+  uint8_t *bottom_buf_end = levels + (height + TX_PAD_BOTTOM) * stride;
+  uint8_t *bottom_buf = bottom_buf_end - ((bottom_len + 31) & (~31));
+
+  do {
+    yy_storeu_256(bottom_buf, y_zeros);
+    bottom_buf += 32;
+  } while (bottom_buf < bottom_buf_end);
+
+  int i = 0;
+  uint8_t *ls = levels;
+  const tran_low_t *cf = coeff;
+  if (width == 4) {
+    do {
+      const __m256i c0 = yy_loadu_256(cf);
+      const __m256i c1 = yy_loadu_256(cf + 8);
+      const __m256i abs01 = _mm256_abs_epi16(_mm256_packs_epi32(c0, c1));
+      const __m256i abs01_8 = _mm256_packs_epi16(abs01, y_zeros);
+      const __m256i res_ = _mm256_shuffle_epi32(abs01_8, 0xd8);
+      const __m256i res = _mm256_permute4x64_epi64(res_, 0xd8);
+      yy_storeu_256(ls, res);
+      ls += 32;
+      cf += 16;
+      i += 4;
+    } while (i < height);
+  } else if (width == 8) {
+    do {
+      const __m256i coeffA = yy_loadu_256(cf);
+      const __m256i coeffB = yy_loadu_256(cf + 8);
+      const __m256i coeffC = yy_loadu_256(cf + 16);
+      const __m256i coeffD = yy_loadu_256(cf + 24);
+      const __m256i coeffAB = _mm256_packs_epi32(coeffA, coeffB);
+      const __m256i coeffCD = _mm256_packs_epi32(coeffC, coeffD);
+      const __m256i absAB = _mm256_abs_epi16(coeffAB);
+      const __m256i absCD = _mm256_abs_epi16(coeffCD);
+      const __m256i absABCD = _mm256_packs_epi16(absAB, absCD);
+      const __m256i res_ = _mm256_permute4x64_epi64(absABCD, 0xd8);
+      const __m256i res = _mm256_shuffle_epi32(res_, 0xd8);
+      const __m128i res0 = _mm256_castsi256_si128(res);
+      const __m128i res1 = _mm256_extracti128_si256(res, 1);
+      xx_storel_64(ls, res0);
+      *(int32_t *)(ls + width) = 0;
+      xx_storel_64(ls + stride, _mm_srli_si128(res0, 8));
+      *(int32_t *)(ls + width + stride) = 0;
+      xx_storel_64(ls + stride * 2, res1);
+      *(int32_t *)(ls + width + stride * 2) = 0;
+      xx_storel_64(ls + stride * 3, _mm_srli_si128(res1, 8));
+      *(int32_t *)(ls + width + stride * 3) = 0;
+      cf += 32;
+      ls += stride << 2;
+      i += 4;
+    } while (i < height);
+  } else if (width == 16) {
+    do {
+      const __m256i coeffA = yy_loadu_256(cf);
+      const __m256i coeffB = yy_loadu_256(cf + 8);
+      const __m256i coeffC = yy_loadu_256(cf + 16);
+      const __m256i coeffD = yy_loadu_256(cf + 24);
+      const __m256i coeffAB = _mm256_packs_epi32(coeffA, coeffB);
+      const __m256i coeffCD = _mm256_packs_epi32(coeffC, coeffD);
+      const __m256i absAB = _mm256_abs_epi16(coeffAB);
+      const __m256i absCD = _mm256_abs_epi16(coeffCD);
+      const __m256i absABCD = _mm256_packs_epi16(absAB, absCD);
+      const __m256i res_ = _mm256_permute4x64_epi64(absABCD, 0xd8);
+      const __m256i res = _mm256_shuffle_epi32(res_, 0xd8);
+      xx_storeu_128(ls, _mm256_castsi256_si128(res));
+      xx_storeu_128(ls + stride, _mm256_extracti128_si256(res, 1));
+      cf += 32;
+      *(int32_t *)(ls + width) = 0;
+      *(int32_t *)(ls + stride + width) = 0;
+      ls += stride << 1;
+      i += 2;
+    } while (i < height);
+  } else {
+    do {
+      const __m256i coeffA = yy_loadu_256(cf);
+      const __m256i coeffB = yy_loadu_256(cf + 8);
+      const __m256i coeffC = yy_loadu_256(cf + 16);
+      const __m256i coeffD = yy_loadu_256(cf + 24);
+      const __m256i coeffAB = _mm256_packs_epi32(coeffA, coeffB);
+      const __m256i coeffCD = _mm256_packs_epi32(coeffC, coeffD);
+      const __m256i absAB = _mm256_abs_epi16(coeffAB);
+      const __m256i absCD = _mm256_abs_epi16(coeffCD);
+      const __m256i absABCD = _mm256_packs_epi16(absAB, absCD);
+      const __m256i res_ = _mm256_permute4x64_epi64(absABCD, 0xd8);
+      const __m256i res = _mm256_shuffle_epi32(res_, 0xd8);
+      yy_storeu_256(ls, res);
+      cf += 32;
+      *(int32_t *)(ls + width) = 0;
+      ls += stride;
+      i += 1;
+    } while (i < height);
+  }
+}
diff --git a/third_party/aom/av1/encoder/x86/encodetxb_sse4.c b/third_party/aom/av1/encoder/x86/encodetxb_sse4.c
index b3a879b0f..5e0687cd3 100644
--- a/third_party/aom/av1/encoder/x86/encodetxb_sse4.c
+++ b/third_party/aom/av1/encoder/x86/encodetxb_sse4.c
@@ -14,43 +14,55 @@
 #include <smmintrin.h>  /* SSE4.1 */
 
 #include "aom/aom_integer.h"
-#include "aom_dsp/x86/mem_sse2.h"
 #include "av1/common/onyxc_int.h"
 #include "av1/common/txb_common.h"
+#include "aom_dsp/x86/synonyms.h"
 
 void av1_txb_init_levels_sse4_1(const tran_low_t *const coeff, const int width,
                                 const int height, uint8_t *const levels) {
   const int stride = width + TX_PAD_HOR;
-  memset(levels - TX_PAD_TOP * stride, 0,
-         sizeof(*levels) * TX_PAD_TOP * stride);
-  memset(levels + stride * height, 0,
-         sizeof(*levels) * (TX_PAD_BOTTOM * stride + TX_PAD_END));
-
   const __m128i zeros = _mm_setzero_si128();
+
+  const int32_t pre_len = sizeof(*levels) * TX_PAD_TOP * stride;
+  uint8_t *pre_buf = levels - TX_PAD_TOP * stride;
+  uint8_t *pre_buf_end = pre_buf + pre_len;
+  do {
+    _mm_storeu_si128((__m128i *)(pre_buf), zeros);
+    pre_buf += 16;
+  } while (pre_buf < pre_buf_end);
+
+  const int32_t bottom_len = sizeof(*levels) * (TX_PAD_BOTTOM * stride);
+  uint8_t *bottom_buf = levels + stride * height;
+  uint8_t *bottom_buf_end = bottom_buf + bottom_len;
+  do {
+    _mm_storeu_si128((__m128i *)(bottom_buf), zeros);
+    bottom_buf += 16;
+  } while (bottom_buf < bottom_buf_end);
+
   int i = 0;
   uint8_t *ls = levels;
   const tran_low_t *cf = coeff;
   if (width == 4) {
     do {
-      const __m128i coeffA = _mm_load_si128((__m128i *)(cf));
-      const __m128i coeffB = _mm_load_si128((__m128i *)(cf + width));
+      const __m128i coeffA = xx_loadu_128(cf);
+      const __m128i coeffB = xx_loadu_128(cf + 4);
       const __m128i coeffAB = _mm_packs_epi32(coeffA, coeffB);
       const __m128i absAB = _mm_abs_epi16(coeffAB);
       const __m128i absAB8 = _mm_packs_epi16(absAB, zeros);
       const __m128i lsAB = _mm_unpacklo_epi32(absAB8, zeros);
-      _mm_storeu_si128((__m128i *)ls, lsAB);
+      xx_storeu_128(ls, lsAB);
       ls += (stride << 1);
       cf += (width << 1);
       i += 2;
     } while (i < height);
   } else if (width == 8) {
     do {
-      const __m128i coeffA = _mm_load_si128((__m128i *)(cf));
-      const __m128i coeffB = _mm_load_si128((__m128i *)(cf + 4));
+      const __m128i coeffA = xx_loadu_128(cf);
+      const __m128i coeffB = xx_loadu_128(cf + 4);
       const __m128i coeffAB = _mm_packs_epi32(coeffA, coeffB);
       const __m128i absAB = _mm_abs_epi16(coeffAB);
       const __m128i absAB8 = _mm_packs_epi16(absAB, zeros);
-      _mm_storeu_si128((__m128i *)ls, absAB8);
+      xx_storeu_128(ls, absAB8);
       ls += stride;
       cf += width;
       i += 1;
@@ -59,16 +71,16 @@ void av1_txb_init_levels_sse4_1(const tran_low_t *const coeff, const int width,
     do {
       int j = 0;
       do {
-        const __m128i coeffA = _mm_load_si128((__m128i *)(cf));
-        const __m128i coeffB = _mm_load_si128((__m128i *)(cf + 4));
-        const __m128i coeffC = _mm_load_si128((__m128i *)(cf + 8));
-        const __m128i coeffD = _mm_load_si128((__m128i *)(cf + 12));
+        const __m128i coeffA = xx_loadu_128(cf);
+        const __m128i coeffB = xx_loadu_128(cf + 4);
+        const __m128i coeffC = xx_loadu_128(cf + 8);
+        const __m128i coeffD = xx_loadu_128(cf + 12);
         const __m128i coeffAB = _mm_packs_epi32(coeffA, coeffB);
         const __m128i coeffCD = _mm_packs_epi32(coeffC, coeffD);
         const __m128i absAB = _mm_abs_epi16(coeffAB);
         const __m128i absCD = _mm_abs_epi16(coeffCD);
         const __m128i absABCD = _mm_packs_epi16(absAB, absCD);
-        _mm_storeu_si128((__m128i *)(ls + j), absABCD);
+        xx_storeu_128(ls + j, absABCD);
         j += 16;
         cf += 16;
       } while (j < width);
diff --git a/third_party/aom/av1/encoder/x86/highbd_fwd_txfm_sse4.c b/third_party/aom/av1/encoder/x86/highbd_fwd_txfm_sse4.c
index 4cd6371a6..535485ae8 100644
--- a/third_party/aom/av1/encoder/x86/highbd_fwd_txfm_sse4.c
+++ b/third_party/aom/av1/encoder/x86/highbd_fwd_txfm_sse4.c
@@ -17,6 +17,7 @@
 #include "av1/common/av1_txfm.h"
 #include "av1/common/x86/highbd_txfm_utility_sse4.h"
 #include "av1/encoder/av1_fwd_txfm1d_cfg.h"
+#include "av1/encoder/x86/av1_txfm1d_sse4.h"
 #include "aom_dsp/txfm_common.h"
 #include "aom_dsp/x86/txfm_common_sse2.h"
 #include "aom_ports/mem.h"
@@ -393,7 +394,32 @@ static INLINE void write_buffer_8x8(const __m128i *res, int32_t *output) {
   _mm_store_si128((__m128i *)(output + 15 * 4), res[15]);
 }
 
-static void fdct8x8_sse4_1(__m128i *in, __m128i *out, int bit) {
+static INLINE void write_buffer_16x8(const __m128i *res, int32_t *output,
+                                     const int stride) {
+  _mm_storeu_si128((__m128i *)(output), res[0]);
+  _mm_storeu_si128((__m128i *)(output + 4), res[1]);
+  _mm_storeu_si128((__m128i *)(output + stride), res[2]);
+  _mm_storeu_si128((__m128i *)(output + stride + 4), res[3]);
+
+  _mm_storeu_si128((__m128i *)(output + (stride * 2)), res[4]);
+  _mm_storeu_si128((__m128i *)(output + (stride * 2) + 4), res[5]);
+  _mm_storeu_si128((__m128i *)(output + (stride * 3)), res[6]);
+  _mm_storeu_si128((__m128i *)(output + (stride * 3) + 4), res[7]);
+
+  _mm_storeu_si128((__m128i *)(output + (stride * 4)), res[8]);
+  _mm_storeu_si128((__m128i *)(output + (stride * 4) + 4), res[9]);
+  _mm_storeu_si128((__m128i *)(output + (stride * 5)), res[10]);
+  _mm_storeu_si128((__m128i *)(output + (stride * 5) + 4), res[11]);
+
+  _mm_storeu_si128((__m128i *)(output + (stride * 6)), res[12]);
+  _mm_storeu_si128((__m128i *)(output + (stride * 6) + 4), res[13]);
+  _mm_storeu_si128((__m128i *)(output + (stride * 7)), res[14]);
+  _mm_storeu_si128((__m128i *)(output + (stride * 7) + 4), res[15]);
+}
+
+static void fdct8x8_sse4_1(__m128i *in, __m128i *out, int bit,
+                           const int col_num) {
+  (void)(col_num);
   const int32_t *cospi = cospi_arr(bit);
   const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
   const __m128i cospim32 = _mm_set1_epi32(-cospi[32]);
@@ -589,7 +615,9 @@ static void fdct8x8_sse4_1(__m128i *in, __m128i *out, int bit) {
   out[13] = u[3];  // buf0[3]
 }
 
-static void fadst8x8_sse4_1(__m128i *in, __m128i *out, int bit) {
+static void fadst8x8_sse4_1(__m128i *in, __m128i *out, int bit,
+                            const int col_num) {
+  (void)(col_num);
   const int32_t *cospi = cospi_arr(bit);
   const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
   const __m128i cospi16 = _mm_set1_epi32(cospi[16]);
@@ -780,82 +808,82 @@ void av1_fwd_txfm2d_8x8_sse4_1(const int16_t *input, int32_t *coeff, int stride,
   switch (tx_type) {
     case DCT_DCT:
       load_buffer_8x8(input, in, stride, 0, 0, shift[0]);
-      fdct8x8_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx]);
+      fdct8x8_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx], 0);
       col_txfm_8x8_rounding(out, -shift[1]);
       transpose_8x8(out, in);
-      fdct8x8_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx]);
+      fdct8x8_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx], 0);
       transpose_8x8(out, in);
       write_buffer_8x8(in, coeff);
       break;
     case ADST_DCT:
       load_buffer_8x8(input, in, stride, 0, 0, shift[0]);
-      fadst8x8_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx]);
+      fadst8x8_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx], 0);
       col_txfm_8x8_rounding(out, -shift[1]);
       transpose_8x8(out, in);
-      fdct8x8_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx]);
+      fdct8x8_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx], 0);
       transpose_8x8(out, in);
       write_buffer_8x8(in, coeff);
       break;
     case DCT_ADST:
       load_buffer_8x8(input, in, stride, 0, 0, shift[0]);
-      fdct8x8_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx]);
+      fdct8x8_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx], 0);
       col_txfm_8x8_rounding(out, -shift[1]);
       transpose_8x8(out, in);
-      fadst8x8_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx]);
+      fadst8x8_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx], 0);
       transpose_8x8(out, in);
       write_buffer_8x8(in, coeff);
       break;
     case ADST_ADST:
       load_buffer_8x8(input, in, stride, 0, 0, shift[0]);
-      fadst8x8_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx]);
+      fadst8x8_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx], 0);
       col_txfm_8x8_rounding(out, -shift[1]);
       transpose_8x8(out, in);
-      fadst8x8_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx]);
+      fadst8x8_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx], 0);
       transpose_8x8(out, in);
       write_buffer_8x8(in, coeff);
       break;
     case FLIPADST_DCT:
       load_buffer_8x8(input, in, stride, 1, 0, shift[0]);
-      fadst8x8_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx]);
+      fadst8x8_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx], 0);
       col_txfm_8x8_rounding(out, -shift[1]);
       transpose_8x8(out, in);
-      fdct8x8_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx]);
+      fdct8x8_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx], 0);
       transpose_8x8(out, in);
       write_buffer_8x8(in, coeff);
       break;
     case DCT_FLIPADST:
       load_buffer_8x8(input, in, stride, 0, 1, shift[0]);
-      fdct8x8_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx]);
+      fdct8x8_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx], 0);
       col_txfm_8x8_rounding(out, -shift[1]);
       transpose_8x8(out, in);
-      fadst8x8_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx]);
+      fadst8x8_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx], 0);
       transpose_8x8(out, in);
       write_buffer_8x8(in, coeff);
       break;
     case FLIPADST_FLIPADST:
       load_buffer_8x8(input, in, stride, 1, 1, shift[0]);
-      fadst8x8_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx]);
+      fadst8x8_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx], 0);
       col_txfm_8x8_rounding(out, -shift[1]);
       transpose_8x8(out, in);
-      fadst8x8_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx]);
+      fadst8x8_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx], 0);
       transpose_8x8(out, in);
       write_buffer_8x8(in, coeff);
       break;
     case ADST_FLIPADST:
       load_buffer_8x8(input, in, stride, 0, 1, shift[0]);
-      fadst8x8_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx]);
+      fadst8x8_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx], 0);
       col_txfm_8x8_rounding(out, -shift[1]);
       transpose_8x8(out, in);
-      fadst8x8_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx]);
+      fadst8x8_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx], 0);
       transpose_8x8(out, in);
       write_buffer_8x8(in, coeff);
       break;
     case FLIPADST_ADST:
       load_buffer_8x8(input, in, stride, 1, 0, shift[0]);
-      fadst8x8_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx]);
+      fadst8x8_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx], 0);
       col_txfm_8x8_rounding(out, -shift[1]);
       transpose_8x8(out, in);
-      fadst8x8_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx]);
+      fadst8x8_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx], 0);
       transpose_8x8(out, in);
       write_buffer_8x8(in, coeff);
       break;
@@ -940,7 +968,26 @@ static INLINE void load_buffer_16x16(const int16_t *input, __m128i *out,
   convert_8x8_to_16x16(in, out);
 }
 
-static void fdct16x16_sse4_1(__m128i *in, __m128i *out, int bit) {
+static INLINE void load_buffer_8x16(const int16_t *input, __m128i *out,
+                                    int stride, int flipud, int fliplr,
+                                    int shift) {
+  const int16_t *topL = input;
+  const int16_t *botL = input + 8 * stride;
+
+  const int16_t *tmp;
+
+  if (flipud) {
+    tmp = topL;
+    topL = botL;
+    botL = tmp;
+  }
+
+  load_buffer_8x8(topL, out, stride, flipud, fliplr, shift);
+  load_buffer_8x8(botL, out + 16, stride, flipud, fliplr, shift);
+}
+
+static void fdct16x16_sse4_1(__m128i *in, __m128i *out, int bit,
+                             const int col_num) {
   const int32_t *cospi = cospi_arr(bit);
   const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
   const __m128i cospim32 = _mm_set1_epi32(-cospi[32]);
@@ -962,7 +1009,6 @@ static void fdct16x16_sse4_1(__m128i *in, __m128i *out, int bit) {
   const __m128i cospi52 = _mm_set1_epi32(cospi[52]);
   const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
   __m128i u[16], v[16], x;
-  const int col_num = 4;
   int col;
 
   // Calculate the column 0, 1, 2, 3
@@ -1226,7 +1272,8 @@ static void fdct16x16_sse4_1(__m128i *in, __m128i *out, int bit) {
   }
 }
 
-static void fadst16x16_sse4_1(__m128i *in, __m128i *out, int bit) {
+static void fadst16x16_sse4_1(__m128i *in, __m128i *out, int bit,
+                              const int num_cols) {
   const int32_t *cospi = cospi_arr(bit);
   const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
   const __m128i cospi48 = _mm_set1_epi32(cospi[48]);
@@ -1271,25 +1318,25 @@ static void fadst16x16_sse4_1(__m128i *in, __m128i *out, int bit) {
   __m128i u[16], v[16], x, y;
   int col;
 
-  for (col = 0; col < 4; ++col) {
+  for (col = 0; col < num_cols; ++col) {
     // stage 0
     // stage 1
-    u[0] = in[0 * 4 + col];
-    u[1] = _mm_sub_epi32(zero, in[15 * 4 + col]);
-    u[2] = _mm_sub_epi32(zero, in[7 * 4 + col]);
-    u[3] = in[8 * 4 + col];
-    u[4] = _mm_sub_epi32(zero, in[3 * 4 + col]);
-    u[5] = in[12 * 4 + col];
-    u[6] = in[4 * 4 + col];
-    u[7] = _mm_sub_epi32(zero, in[11 * 4 + col]);
-    u[8] = _mm_sub_epi32(zero, in[1 * 4 + col]);
-    u[9] = in[14 * 4 + col];
-    u[10] = in[6 * 4 + col];
-    u[11] = _mm_sub_epi32(zero, in[9 * 4 + col]);
-    u[12] = in[2 * 4 + col];
-    u[13] = _mm_sub_epi32(zero, in[13 * 4 + col]);
-    u[14] = _mm_sub_epi32(zero, in[5 * 4 + col]);
-    u[15] = in[10 * 4 + col];
+    u[0] = in[0 * num_cols + col];
+    u[1] = _mm_sub_epi32(zero, in[15 * num_cols + col]);
+    u[2] = _mm_sub_epi32(zero, in[7 * num_cols + col]);
+    u[3] = in[8 * num_cols + col];
+    u[4] = _mm_sub_epi32(zero, in[3 * num_cols + col]);
+    u[5] = in[12 * num_cols + col];
+    u[6] = in[4 * num_cols + col];
+    u[7] = _mm_sub_epi32(zero, in[11 * num_cols + col]);
+    u[8] = _mm_sub_epi32(zero, in[1 * num_cols + col]);
+    u[9] = in[14 * num_cols + col];
+    u[10] = in[6 * num_cols + col];
+    u[11] = _mm_sub_epi32(zero, in[9 * num_cols + col]);
+    u[12] = in[2 * num_cols + col];
+    u[13] = _mm_sub_epi32(zero, in[13 * num_cols + col]);
+    u[14] = _mm_sub_epi32(zero, in[5 * num_cols + col]);
+    u[15] = in[10 * num_cols + col];
 
     // stage 2
     v[0] = u[0];
@@ -1453,22 +1500,22 @@ static void fadst16x16_sse4_1(__m128i *in, __m128i *out, int bit) {
     v[15] = half_btf_sse4_1(&cospi6, &u[14], &cospim58, &u[15], &rnding, bit);
 
     // stage 9
-    out[0 * 4 + col] = v[1];
-    out[1 * 4 + col] = v[14];
-    out[2 * 4 + col] = v[3];
-    out[3 * 4 + col] = v[12];
-    out[4 * 4 + col] = v[5];
-    out[5 * 4 + col] = v[10];
-    out[6 * 4 + col] = v[7];
-    out[7 * 4 + col] = v[8];
-    out[8 * 4 + col] = v[9];
-    out[9 * 4 + col] = v[6];
-    out[10 * 4 + col] = v[11];
-    out[11 * 4 + col] = v[4];
-    out[12 * 4 + col] = v[13];
-    out[13 * 4 + col] = v[2];
-    out[14 * 4 + col] = v[15];
-    out[15 * 4 + col] = v[0];
+    out[0 * num_cols + col] = v[1];
+    out[1 * num_cols + col] = v[14];
+    out[2 * num_cols + col] = v[3];
+    out[3 * num_cols + col] = v[12];
+    out[4 * num_cols + col] = v[5];
+    out[5 * num_cols + col] = v[10];
+    out[6 * num_cols + col] = v[7];
+    out[7 * num_cols + col] = v[8];
+    out[8 * num_cols + col] = v[9];
+    out[9 * num_cols + col] = v[6];
+    out[10 * num_cols + col] = v[11];
+    out[11 * num_cols + col] = v[4];
+    out[12 * num_cols + col] = v[13];
+    out[13 * num_cols + col] = v[2];
+    out[14 * num_cols + col] = v[15];
+    out[15 * num_cols + col] = v[0];
   }
 }
 
@@ -1482,6 +1529,11 @@ static void col_txfm_16x16_rounding(__m128i *in, int shift) {
   col_txfm_8x8_rounding(&in[48], shift);
 }
 
+static void col_txfm_8x16_rounding(__m128i *in, int shift) {
+  col_txfm_8x8_rounding(&in[0], shift);
+  col_txfm_8x8_rounding(&in[16], shift);
+}
+
 static void write_buffer_16x16(const __m128i *in, int32_t *output) {
   const int size_8x8 = 16 * 4;
   write_buffer_8x8(&in[0], output);
@@ -1499,85 +1551,86 @@ void av1_fwd_txfm2d_16x16_sse4_1(const int16_t *input, int32_t *coeff,
   const int8_t *shift = fwd_txfm_shift_ls[TX_16X16];
   const int txw_idx = get_txw_idx(TX_16X16);
   const int txh_idx = get_txh_idx(TX_16X16);
+  const int col_num = 4;
   switch (tx_type) {
     case DCT_DCT:
       load_buffer_16x16(input, in, stride, 0, 0, shift[0]);
-      fdct16x16_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx]);
+      fdct16x16_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx], col_num);
       col_txfm_16x16_rounding(out, -shift[1]);
       transpose_16x16(out, in);
-      fdct16x16_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx]);
+      fdct16x16_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx], col_num);
       transpose_16x16(out, in);
       write_buffer_16x16(in, coeff);
       break;
     case ADST_DCT:
       load_buffer_16x16(input, in, stride, 0, 0, shift[0]);
-      fadst16x16_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx]);
+      fadst16x16_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx], col_num);
       col_txfm_16x16_rounding(out, -shift[1]);
       transpose_16x16(out, in);
-      fdct16x16_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx]);
+      fdct16x16_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx], col_num);
       transpose_16x16(out, in);
       write_buffer_16x16(in, coeff);
       break;
     case DCT_ADST:
       load_buffer_16x16(input, in, stride, 0, 0, shift[0]);
-      fdct16x16_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx]);
+      fdct16x16_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx], col_num);
       col_txfm_16x16_rounding(out, -shift[1]);
       transpose_16x16(out, in);
-      fadst16x16_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx]);
+      fadst16x16_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx], col_num);
       transpose_16x16(out, in);
       write_buffer_16x16(in, coeff);
       break;
     case ADST_ADST:
       load_buffer_16x16(input, in, stride, 0, 0, shift[0]);
-      fadst16x16_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx]);
+      fadst16x16_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx], col_num);
       col_txfm_16x16_rounding(out, -shift[1]);
       transpose_16x16(out, in);
-      fadst16x16_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx]);
+      fadst16x16_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx], col_num);
       transpose_16x16(out, in);
       write_buffer_16x16(in, coeff);
       break;
     case FLIPADST_DCT:
       load_buffer_16x16(input, in, stride, 1, 0, shift[0]);
-      fadst16x16_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx]);
+      fadst16x16_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx], col_num);
       col_txfm_16x16_rounding(out, -shift[1]);
       transpose_16x16(out, in);
-      fdct16x16_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx]);
+      fdct16x16_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx], col_num);
       transpose_16x16(out, in);
       write_buffer_16x16(in, coeff);
       break;
     case DCT_FLIPADST:
       load_buffer_16x16(input, in, stride, 0, 1, shift[0]);
-      fdct16x16_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx]);
+      fdct16x16_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx], col_num);
       col_txfm_16x16_rounding(out, -shift[1]);
       transpose_16x16(out, in);
-      fadst16x16_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx]);
+      fadst16x16_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx], col_num);
       transpose_16x16(out, in);
       write_buffer_16x16(in, coeff);
       break;
     case FLIPADST_FLIPADST:
       load_buffer_16x16(input, in, stride, 1, 1, shift[0]);
-      fadst16x16_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx]);
+      fadst16x16_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx], col_num);
       col_txfm_16x16_rounding(out, -shift[1]);
       transpose_16x16(out, in);
-      fadst16x16_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx]);
+      fadst16x16_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx], col_num);
       transpose_16x16(out, in);
       write_buffer_16x16(in, coeff);
       break;
     case ADST_FLIPADST:
       load_buffer_16x16(input, in, stride, 0, 1, shift[0]);
-      fadst16x16_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx]);
+      fadst16x16_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx], col_num);
       col_txfm_16x16_rounding(out, -shift[1]);
       transpose_16x16(out, in);
-      fadst16x16_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx]);
+      fadst16x16_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx], col_num);
       transpose_16x16(out, in);
       write_buffer_16x16(in, coeff);
       break;
     case FLIPADST_ADST:
       load_buffer_16x16(input, in, stride, 1, 0, shift[0]);
-      fadst16x16_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx]);
+      fadst16x16_sse4_1(in, out, fwd_cos_bit_col[txw_idx][txh_idx], col_num);
       col_txfm_16x16_rounding(out, -shift[1]);
       transpose_16x16(out, in);
-      fadst16x16_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx]);
+      fadst16x16_sse4_1(in, out, fwd_cos_bit_row[txw_idx][txh_idx], col_num);
       transpose_16x16(out, in);
       write_buffer_16x16(in, coeff);
       break;
@@ -1585,3 +1638,146 @@ void av1_fwd_txfm2d_16x16_sse4_1(const int16_t *input, int32_t *coeff,
   }
   (void)bd;
 }
+
+static INLINE void flip_buf_sse4_1(__m128i *in, __m128i *out, int size) {
+  for (int i = 0; i < size; i += 2) in[30 - i] = out[i];
+  for (int i = 1; i < size; i += 2) in[size - i] = out[i];
+}
+
+static const fwd_transform_1d_sse4_1 col_highbd_txfm8x8_arr[TX_TYPES] = {
+  fdct8x8_sse4_1,   // DCT_DCT
+  fadst8x8_sse4_1,  // ADST_DCT
+  fdct8x8_sse4_1,   // DCT_ADST
+  fadst8x8_sse4_1,  // ADST_ADST
+  fadst8x8_sse4_1,  // FLIPADST_DCT
+  fdct8x8_sse4_1,   // DCT_FLIPADST
+  fadst8x8_sse4_1,  // FLIPADST_FLIPADST
+  fadst8x8_sse4_1,  // ADST_FLIPADST
+  fadst8x8_sse4_1,  // FLIPADST_ADST
+  NULL,             // IDTX
+  NULL,             // V_DCT
+  NULL,             // H_DCT
+  NULL,             // V_ADST
+  NULL,             // H_ADST
+  NULL,             // V_FLIPADST
+  NULL              // H_FLIPADST
+};
+
+static const fwd_transform_1d_sse4_1 row_highbd_txfm8x16_arr[TX_TYPES] = {
+  fdct16x16_sse4_1,   // DCT_DCT
+  fdct16x16_sse4_1,   // ADST_DCT
+  fadst16x16_sse4_1,  // DCT_ADST
+  fadst16x16_sse4_1,  // ADST_ADST
+  fdct16x16_sse4_1,   // FLIPADST_DCT
+  fadst16x16_sse4_1,  // DCT_FLIPADST
+  fadst16x16_sse4_1,  // FLIPADST_FLIPADST
+  fadst16x16_sse4_1,  // ADST_FLIPADST
+  fadst16x16_sse4_1,  // FLIPADST_ADST
+  NULL,               // IDTX
+  NULL,               // V_DCT
+  NULL,               // H_DCT
+  NULL,               // V_ADST
+  NULL,               // H_ADST
+  NULL,               // V_FLIPADST
+  NULL                // H_FLIPADST
+};
+
+static const fwd_transform_1d_sse4_1 col_highbd_txfm8x16_arr[TX_TYPES] = {
+  fdct16x16_sse4_1,   // DCT_DCT
+  fadst16x16_sse4_1,  // ADST_DCT
+  fdct16x16_sse4_1,   // DCT_ADST
+  fadst16x16_sse4_1,  // ADST_ADST
+  fadst16x16_sse4_1,  // FLIPADST_DCT
+  fdct16x16_sse4_1,   // DCT_FLIPADST
+  fadst16x16_sse4_1,  // FLIPADST_FLIPADST
+  fadst16x16_sse4_1,  // ADST_FLIPADST
+  fadst16x16_sse4_1,  // FLIPADST_ADST
+  NULL,               // IDTX
+  NULL,               // V_DCT
+  NULL,               // H_DCT
+  NULL,               // V_ADST
+  NULL,               // H_ADST
+  NULL,               // V_FLIPADST
+  NULL                // H_FLIPADST
+};
+static const fwd_transform_1d_sse4_1 row_highbd_txfm8x8_arr[TX_TYPES] = {
+  fdct8x8_sse4_1,   // DCT_DCT
+  fdct8x8_sse4_1,   // ADST_DCT
+  fadst8x8_sse4_1,  // DCT_ADST
+  fadst8x8_sse4_1,  // ADST_ADST
+  fdct8x8_sse4_1,   // FLIPADST_DCT
+  fadst8x8_sse4_1,  // DCT_FLIPADST
+  fadst8x8_sse4_1,  // FLIPADST_FLIPADST
+  fadst8x8_sse4_1,  // ADST_FLIPADST
+  fadst8x8_sse4_1,  // FLIPADST_ADST
+  NULL,             // IDTX
+  NULL,             // V_DCT
+  NULL,             // H_DCT
+  NULL,             // V_ADST
+  NULL,             // H_ADST
+  NULL,             // V_FLIPADST
+  NULL              // H_FLIPADST
+};
+
+void av1_fwd_txfm2d_16x8_sse4_1(const int16_t *input, int32_t *coeff,
+                                int stride, TX_TYPE tx_type, int bd) {
+  __m128i in[32], out[32];
+  const int8_t *shift = fwd_txfm_shift_ls[TX_16X8];
+  const int txw_idx = get_txw_idx(TX_16X8);
+  const int txh_idx = get_txh_idx(TX_16X8);
+  const fwd_transform_1d_sse4_1 col_txfm = col_highbd_txfm8x8_arr[tx_type];
+  const fwd_transform_1d_sse4_1 row_txfm = row_highbd_txfm8x16_arr[tx_type];
+  int bit = fwd_cos_bit_col[txw_idx][txh_idx];
+  int ud_flip, lr_flip;
+  get_flip_cfg(tx_type, &ud_flip, &lr_flip);
+
+  for (int i = 0; i < 2; i++) {
+    load_buffer_8x8(input + i * 8, in, stride, ud_flip, 0, shift[0]);
+    col_txfm(in, in, bit, 0);
+    col_txfm_8x8_rounding(in, -shift[1]);
+    transpose_8x8(in, out + i * 16);
+  }
+
+  if (lr_flip) {
+    flip_buf_sse4_1(in, out, 32);
+    row_txfm(in, out, bit, 2);
+  } else {
+    row_txfm(out, out, bit, 2);
+  }
+
+  for (int i = 0; i < 2; i++) {
+    transpose_8x8(out + i * 16, in);
+    av1_round_shift_rect_array_32_sse4_1(in, in, 16, -shift[2], NewSqrt2);
+    write_buffer_16x8(in, coeff + i * 8, 16);
+  }
+
+  (void)bd;
+}
+
+void av1_fwd_txfm2d_8x16_sse4_1(const int16_t *input, int32_t *coeff,
+                                int stride, TX_TYPE tx_type, int bd) {
+  __m128i in[32], out[32];
+  const int8_t *shift = fwd_txfm_shift_ls[TX_8X16];
+  const int txw_idx = get_txw_idx(TX_8X16);
+  const int txh_idx = get_txh_idx(TX_8X16);
+  const fwd_transform_1d_sse4_1 col_txfm = col_highbd_txfm8x16_arr[tx_type];
+  const fwd_transform_1d_sse4_1 row_txfm = row_highbd_txfm8x8_arr[tx_type];
+  int bit = fwd_cos_bit_col[txw_idx][txh_idx];
+  int ud_flip, lr_flip;
+  get_flip_cfg(tx_type, &ud_flip, &lr_flip);
+
+  load_buffer_8x16(input, in, stride, ud_flip, lr_flip, shift[0]);
+  col_txfm(in, in, bit, 2);
+  col_txfm_8x16_rounding(in, -shift[1]);
+  transpose_8x8(in, out);
+  transpose_8x8(in + 16, out + 16);
+
+  for (int i = 0; i < 2; i++) {
+    row_txfm(out + i * 16, out, bit, 0);
+    transpose_8x8(out, in);
+    av1_round_shift_rect_array_32_sse4_1(in, in, 16, -shift[2], NewSqrt2);
+    write_buffer_8x8(in, coeff + i * 64);
+  }
+
+  (void)bd;
+}
diff --git a/third_party/aom/av1/encoder/x86/pickrst_avx2.c b/third_party/aom/av1/encoder/x86/pickrst_avx2.c
new file mode 100644
index 000000000..06aaaa7ee
--- /dev/null
+++ b/third_party/aom/av1/encoder/x86/pickrst_avx2.c
@@ -0,0 +1,403 @@
+/*
+ * Copyright (c) 2018, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <immintrin.h>  // AVX2
+#include "aom_dsp/x86/synonyms.h"
+#include "aom_dsp/x86/synonyms_avx2.h"
+#include "aom_dsp/x86/transpose_sse2.h"
+
+#include "config/av1_rtcd.h"
+#include "av1/common/restoration.h"
+#include "av1/encoder/pickrst.h"
+
+static INLINE void acc_stat_avx2(int32_t *dst, const uint8_t *src,
+                                 const __m128i *shuffle, const __m256i *kl) {
+  const __m128i s = _mm_shuffle_epi8(xx_loadu_128(src), *shuffle);
+  const __m256i d0 = _mm256_madd_epi16(*kl, _mm256_cvtepu8_epi16(s));
+  const __m256i dst0 = yy_loadu_256(dst);
+  const __m256i r0 = _mm256_add_epi32(dst0, d0);
+  yy_storeu_256(dst, r0);
+}
+
+static INLINE void acc_stat_win7_one_line_avx2(
+    const uint8_t *dgd, const uint8_t *src, int h_start, int h_end,
+    int dgd_stride, const __m128i *shuffle, int32_t *sumX,
+    int32_t sumY[WIENER_WIN][WIENER_WIN], int32_t M_int[WIENER_WIN][WIENER_WIN],
+    int32_t H_int[WIENER_WIN2][WIENER_WIN * 8]) {
+  int j, k, l;
+  const int wiener_win = WIENER_WIN;
+  for (j = h_start; j < h_end; j += 2) {
+    const uint8_t X1 = src[j];
+    const uint8_t X2 = src[j + 1];
+    *sumX += X1 + X2;
+    const uint8_t *dgd_ij = dgd + j;
+    for (k = 0; k < wiener_win; k++) {
+      const uint8_t *dgd_ijk = dgd_ij + k * dgd_stride;
+      for (l = 0; l < wiener_win; l++) {
+        int32_t *H_ = &H_int[(l * wiener_win + k)][0];
+        const uint8_t D1 = dgd_ijk[l];
+        const uint8_t D2 = dgd_ijk[l + 1];
+        sumY[k][l] += D1 + D2;
+        M_int[k][l] += D1 * X1 + D2 * X2;
+
+        const __m256i kl =
+            _mm256_cvtepu8_epi16(_mm_set1_epi16(*((uint16_t *)(dgd_ijk + l))));
+        acc_stat_avx2(H_ + 0 * 8, dgd_ij + 0 * dgd_stride, shuffle, &kl);
+        acc_stat_avx2(H_ + 1 * 8, dgd_ij + 1 * dgd_stride, shuffle, &kl);
+        acc_stat_avx2(H_ + 2 * 8, dgd_ij + 2 * dgd_stride, shuffle, &kl);
+        acc_stat_avx2(H_ + 3 * 8, dgd_ij + 3 * dgd_stride, shuffle, &kl);
+        acc_stat_avx2(H_ + 4 * 8, dgd_ij + 4 * dgd_stride, shuffle, &kl);
+        acc_stat_avx2(H_ + 5 * 8, dgd_ij + 5 * dgd_stride, shuffle, &kl);
+        acc_stat_avx2(H_ + 6 * 8, dgd_ij + 6 * dgd_stride, shuffle, &kl);
+      }
+    }
+  }
+}
+
+static INLINE void compute_stats_win7_opt_avx2(
+    const uint8_t *dgd, const uint8_t *src, int h_start, int h_end, int v_start,
+    int v_end, int dgd_stride, int src_stride, double *M, double *H) {
+  int i, j, k, l, m, n;
+  const int wiener_win = WIENER_WIN;
+  const int pixel_count = (h_end - h_start) * (v_end - v_start);
+  const int wiener_win2 = wiener_win * wiener_win;
+  const int wiener_halfwin = (wiener_win >> 1);
+  const double avg =
+      find_average(dgd, h_start, h_end, v_start, v_end, dgd_stride);
+
+  int32_t M_int32[WIENER_WIN][WIENER_WIN] = { { 0 } };
+  int64_t M_int64[WIENER_WIN][WIENER_WIN] = { { 0 } };
+  int32_t H_int32[WIENER_WIN2][WIENER_WIN * 8] = { { 0 } };
+  int64_t H_int64[WIENER_WIN2][WIENER_WIN * 8] = { { 0 } };
+  int32_t sumY[WIENER_WIN][WIENER_WIN] = { { 0 } };
+  int32_t sumX = 0;
+  const uint8_t *dgd_win = dgd - wiener_halfwin * dgd_stride - wiener_halfwin;
+
+  const __m128i shuffle = xx_loadu_128(g_shuffle_stats_data);
+  for (j = v_start; j < v_end; j += 64) {
+    const int vert_end = AOMMIN(64, v_end - j) + j;
+    for (i = j; i < vert_end; i++) {
+      acc_stat_win7_one_line_avx2(
+          dgd_win + i * dgd_stride, src + i * src_stride, h_start, h_end,
+          dgd_stride, &shuffle, &sumX, sumY, M_int32, H_int32);
+    }
+    for (k = 0; k < wiener_win; ++k) {
+      for (l = 0; l < wiener_win; ++l) {
+        M_int64[k][l] += M_int32[k][l];
+        M_int32[k][l] = 0;
+      }
+    }
+    for (k = 0; k < WIENER_WIN2; ++k) {
+      for (l = 0; l < WIENER_WIN * 8; ++l) {
+        H_int64[k][l] += H_int32[k][l];
+        H_int32[k][l] = 0;
+      }
+    }
+  }
+
+  const double avg_square_sum = avg * avg * pixel_count;
+  for (k = 0; k < wiener_win; k++) {
+    for (l = 0; l < wiener_win; l++) {
+      const int32_t idx0 = l * wiener_win + k;
+      M[idx0] = M_int64[k][l] + avg_square_sum - avg * (sumX + sumY[k][l]);
+      double *H_ = H + idx0 * wiener_win2;
+      int64_t *H_int_ = &H_int64[idx0][0];
+      for (m = 0; m < wiener_win; m++) {
+        for (n = 0; n < wiener_win; n++) {
+          H_[m * wiener_win + n] = H_int_[n * 8 + m] + avg_square_sum -
+                                   avg * (sumY[k][l] + sumY[n][m]);
+        }
+      }
+    }
+  }
+}
+
+static INLINE void acc_stat_win5_one_line_avx2(
+    const uint8_t *dgd, const uint8_t *src, int h_start, int h_end,
+    int dgd_stride, const __m128i *shuffle, int32_t *sumX,
+    int32_t sumY[WIENER_WIN_CHROMA][WIENER_WIN_CHROMA],
+    int32_t M_int[WIENER_WIN_CHROMA][WIENER_WIN_CHROMA],
+    int32_t H_int[WIENER_WIN2_CHROMA][WIENER_WIN_CHROMA * 8]) {
+  int j, k, l;
+  const int wiener_win = WIENER_WIN_CHROMA;
+  for (j = h_start; j < h_end; j += 2) {
+    const uint8_t X1 = src[j];
+    const uint8_t X2 = src[j + 1];
+    *sumX += X1 + X2;
+    const uint8_t *dgd_ij = dgd + j;
+    for (k = 0; k < wiener_win; k++) {
+      const uint8_t *dgd_ijk = dgd_ij + k * dgd_stride;
+      for (l = 0; l < wiener_win; l++) {
+        int32_t *H_ = &H_int[(l * wiener_win + k)][0];
+        const uint8_t D1 = dgd_ijk[l];
+        const uint8_t D2 = dgd_ijk[l + 1];
+        sumY[k][l] += D1 + D2;
+        M_int[k][l] += D1 * X1 + D2 * X2;
+
+        const __m256i kl =
+            _mm256_cvtepu8_epi16(_mm_set1_epi16(*((uint16_t *)(dgd_ijk + l))));
+        acc_stat_avx2(H_ + 0 * 8, dgd_ij + 0 * dgd_stride, shuffle, &kl);
+        acc_stat_avx2(H_ + 1 * 8, dgd_ij + 1 * dgd_stride, shuffle, &kl);
+        acc_stat_avx2(H_ + 2 * 8, dgd_ij + 2 * dgd_stride, shuffle, &kl);
+        acc_stat_avx2(H_ + 3 * 8, dgd_ij + 3 * dgd_stride, shuffle, &kl);
+        acc_stat_avx2(H_ + 4 * 8, dgd_ij + 4 * dgd_stride, shuffle, &kl);
+      }
+    }
+  }
+}
+
+static INLINE void compute_stats_win5_opt_avx2(
+    const uint8_t *dgd, const uint8_t *src, int h_start, int h_end, int v_start,
+    int v_end, int dgd_stride, int src_stride, double *M, double *H) {
+  int i, j, k, l, m, n;
+  const int wiener_win = WIENER_WIN_CHROMA;
+  const int pixel_count = (h_end - h_start) * (v_end - v_start);
+  const int wiener_win2 = wiener_win * wiener_win;
+  const int wiener_halfwin = (wiener_win >> 1);
+  const double avg =
+      find_average(dgd, h_start, h_end, v_start, v_end, dgd_stride);
+
+  int32_t M_int32[WIENER_WIN_CHROMA][WIENER_WIN_CHROMA] = { { 0 } };
+  int64_t M_int64[WIENER_WIN_CHROMA][WIENER_WIN_CHROMA] = { { 0 } };
+  int32_t H_int32[WIENER_WIN2_CHROMA][WIENER_WIN_CHROMA * 8] = { { 0 } };
+  int64_t H_int64[WIENER_WIN2_CHROMA][WIENER_WIN_CHROMA * 8] = { { 0 } };
+  int32_t sumY[WIENER_WIN_CHROMA][WIENER_WIN_CHROMA] = { { 0 } };
+  int32_t sumX = 0;
+  const uint8_t *dgd_win = dgd - wiener_halfwin * dgd_stride - wiener_halfwin;
+
+  const __m128i shuffle = xx_loadu_128(g_shuffle_stats_data);
+  for (j = v_start; j < v_end; j += 64) {
+    const int vert_end = AOMMIN(64, v_end - j) + j;
+    for (i = j; i < vert_end; i++) {
+      acc_stat_win5_one_line_avx2(
+          dgd_win + i * dgd_stride, src + i * src_stride, h_start, h_end,
+          dgd_stride, &shuffle, &sumX, sumY, M_int32, H_int32);
+    }
+    for (k = 0; k < wiener_win; ++k) {
+      for (l = 0; l < wiener_win; ++l) {
+        M_int64[k][l] += M_int32[k][l];
+        M_int32[k][l] = 0;
+      }
+    }
+    for (k = 0; k < WIENER_WIN2_CHROMA; ++k) {
+      for (l = 0; l < WIENER_WIN_CHROMA * 8; ++l) {
+        H_int64[k][l] += H_int32[k][l];
+        H_int32[k][l] = 0;
+      }
+    }
+  }
+
+  const double avg_square_sum = avg * avg * pixel_count;
+  for (k = 0; k < wiener_win; k++) {
+    for (l = 0; l < wiener_win; l++) {
+      const int32_t idx0 = l * wiener_win + k;
+      M[idx0] = M_int64[k][l] + avg_square_sum - avg * (sumX + sumY[k][l]);
+      double *H_ = H + idx0 * wiener_win2;
+      int64_t *H_int_ = &H_int64[idx0][0];
+      for (m = 0; m < wiener_win; m++) {
+        for (n = 0; n < wiener_win; n++) {
+          H_[m * wiener_win + n] = H_int_[n * 8 + m] + avg_square_sum -
+                                   avg * (sumY[k][l] + sumY[n][m]);
+        }
+      }
+    }
+  }
+}
+
+void av1_compute_stats_avx2(int wiener_win, const uint8_t *dgd,
+                            const uint8_t *src, int h_start, int h_end,
+                            int v_start, int v_end, int dgd_stride,
+                            int src_stride, double *M, double *H) {
+  if (wiener_win == WIENER_WIN) {
+    compute_stats_win7_opt_avx2(dgd, src, h_start, h_end, v_start, v_end,
+                                dgd_stride, src_stride, M, H);
+  } else if (wiener_win == WIENER_WIN_CHROMA) {
+    compute_stats_win5_opt_avx2(dgd, src, h_start, h_end, v_start, v_end,
+                                dgd_stride, src_stride, M, H);
+  } else {
+    av1_compute_stats_c(wiener_win, dgd, src, h_start, h_end, v_start, v_end,
+                        dgd_stride, src_stride, M, H);
+  }
+}
+
+static INLINE __m256i pair_set_epi16(uint16_t a, uint16_t b) {
+  return _mm256_set1_epi32(
+      (int32_t)(((uint16_t)(a)) | (((uint32_t)(b)) << 16)));
+}
+
+int64_t av1_lowbd_pixel_proj_error_avx2(
+    const uint8_t *src8, int width, int height, int src_stride,
+    const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride,
+    int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params) {
+  int i, j, k;
+  const int32_t shift = SGRPROJ_RST_BITS + SGRPROJ_PRJ_BITS;
+  const __m256i rounding = _mm256_set1_epi32(1 << (shift - 1));
+  __m256i sum64 = _mm256_setzero_si256();
+  const uint8_t *src = src8;
+  const uint8_t *dat = dat8;
+  int64_t err = 0;
+  if (params->r[0] > 0 && params->r[1] > 0) {
+    __m256i xq_coeff = pair_set_epi16(xq[0], xq[1]);
+    for (i = 0; i < height; ++i) {
+      __m256i sum32 = _mm256_setzero_si256();
+      for (j = 0; j <= width - 16; j += 16) {
+        const __m256i d0 = _mm256_cvtepu8_epi16(xx_loadu_128(dat + j));
+        const __m256i s0 = _mm256_cvtepu8_epi16(xx_loadu_128(src + j));
+        const __m256i flt0_16b = _mm256_permute4x64_epi64(
+            _mm256_packs_epi32(yy_loadu_256(flt0 + j),
+                               yy_loadu_256(flt0 + j + 8)),
+            0xd8);
+        const __m256i flt1_16b = _mm256_permute4x64_epi64(
+            _mm256_packs_epi32(yy_loadu_256(flt1 + j),
+                               yy_loadu_256(flt1 + j + 8)),
+            0xd8);
+        const __m256i u0 = _mm256_slli_epi16(d0, SGRPROJ_RST_BITS);
+        const __m256i flt0_0_sub_u = _mm256_sub_epi16(flt0_16b, u0);
+        const __m256i flt1_0_sub_u = _mm256_sub_epi16(flt1_16b, u0);
+        const __m256i v0 = _mm256_madd_epi16(
+            xq_coeff, _mm256_unpacklo_epi16(flt0_0_sub_u, flt1_0_sub_u));
+        const __m256i v1 = _mm256_madd_epi16(
+            xq_coeff, _mm256_unpackhi_epi16(flt0_0_sub_u, flt1_0_sub_u));
+        const __m256i vr0 =
+            _mm256_srai_epi32(_mm256_add_epi32(v0, rounding), shift);
+        const __m256i vr1 =
+            _mm256_srai_epi32(_mm256_add_epi32(v1, rounding), shift);
+        const __m256i e0 = _mm256_sub_epi16(
+            _mm256_add_epi16(_mm256_packs_epi32(vr0, vr1), d0), s0);
+        const __m256i err0 = _mm256_madd_epi16(e0, e0);
+        sum32 = _mm256_add_epi32(sum32, err0);
+      }
+      for (k = j; k < width; ++k) {
+        const int32_t u = (int32_t)(dat[k] << SGRPROJ_RST_BITS);
+        int32_t v = xq[0] * (flt0[k] - u) + xq[1] * (flt1[k] - u);
+        const int32_t e = ROUND_POWER_OF_TWO(v, shift) + dat[k] - src[k];
+        err += e * e;
+      }
+      dat += dat_stride;
+      src += src_stride;
+      flt0 += flt0_stride;
+      flt1 += flt1_stride;
+      const __m256i sum64_0 =
+          _mm256_cvtepi32_epi64(_mm256_castsi256_si128(sum32));
+      const __m256i sum64_1 =
+          _mm256_cvtepi32_epi64(_mm256_extracti128_si256(sum32, 1));
+      sum64 = _mm256_add_epi64(sum64, sum64_0);
+      sum64 = _mm256_add_epi64(sum64, sum64_1);
+    }
+  } else if (params->r[0] > 0) {
+    __m256i xq_coeff =
+        pair_set_epi16(xq[0], (-xq[0] * (1 << SGRPROJ_RST_BITS)));
+    for (i = 0; i < height; ++i) {
+      __m256i sum32 = _mm256_setzero_si256();
+      for (j = 0; j <= width - 16; j += 16) {
+        const __m256i d0 = _mm256_cvtepu8_epi16(xx_loadu_128(dat + j));
+        const __m256i s0 = _mm256_cvtepu8_epi16(xx_loadu_128(src + j));
+        const __m256i flt0_16b = _mm256_permute4x64_epi64(
+            _mm256_packs_epi32(yy_loadu_256(flt0 + j),
+                               yy_loadu_256(flt0 + j + 8)),
+            0xd8);
+        const __m256i v0 =
+            _mm256_madd_epi16(xq_coeff, _mm256_unpacklo_epi16(flt0_16b, d0));
+        const __m256i v1 =
+            _mm256_madd_epi16(xq_coeff, _mm256_unpackhi_epi16(flt0_16b, d0));
+        const __m256i vr0 =
+            _mm256_srai_epi32(_mm256_add_epi32(v0, rounding), shift);
+        const __m256i vr1 =
+            _mm256_srai_epi32(_mm256_add_epi32(v1, rounding), shift);
+        const __m256i e0 = _mm256_sub_epi16(
+            _mm256_add_epi16(_mm256_packs_epi32(vr0, vr1), d0), s0);
+        const __m256i err0 = _mm256_madd_epi16(e0, e0);
+        sum32 = _mm256_add_epi32(sum32, err0);
+      }
+      for (k = j; k < width; ++k) {
+        const int32_t u = (int32_t)(dat[k] << SGRPROJ_RST_BITS);
+        int32_t v = xq[0] * (flt0[k] - u);
+        const int32_t e = ROUND_POWER_OF_TWO(v, shift) + dat[k] - src[k];
+        err += e * e;
+      }
+      dat += dat_stride;
+      src += src_stride;
+      flt0 += flt0_stride;
+      const __m256i sum64_0 =
+          _mm256_cvtepi32_epi64(_mm256_castsi256_si128(sum32));
+      const __m256i sum64_1 =
+          _mm256_cvtepi32_epi64(_mm256_extracti128_si256(sum32, 1));
+      sum64 = _mm256_add_epi64(sum64, sum64_0);
+      sum64 = _mm256_add_epi64(sum64, sum64_1);
+    }
+  } else if (params->r[1] > 0) {
+    __m256i xq_coeff = pair_set_epi16(xq[1], -(xq[1] << SGRPROJ_RST_BITS));
+    for (i = 0; i < height; ++i) {
+      __m256i sum32 = _mm256_setzero_si256();
+      for (j = 0; j <= width - 16; j += 16) {
+        const __m256i d0 = _mm256_cvtepu8_epi16(xx_loadu_128(dat + j));
+        const __m256i s0 = _mm256_cvtepu8_epi16(xx_loadu_128(src + j));
+        const __m256i flt1_16b = _mm256_permute4x64_epi64(
+            _mm256_packs_epi32(yy_loadu_256(flt1 + j),
+                               yy_loadu_256(flt1 + j + 8)),
+            0xd8);
+        const __m256i v0 =
+            _mm256_madd_epi16(xq_coeff, _mm256_unpacklo_epi16(flt1_16b, d0));
+        const __m256i v1 =
+            _mm256_madd_epi16(xq_coeff, _mm256_unpackhi_epi16(flt1_16b, d0));
+        const __m256i vr0 =
+            _mm256_srai_epi32(_mm256_add_epi32(v0, rounding), shift);
+        const __m256i vr1 =
+            _mm256_srai_epi32(_mm256_add_epi32(v1, rounding), shift);
+        const __m256i e0 = _mm256_sub_epi16(
+            _mm256_add_epi16(_mm256_packs_epi32(vr0, vr1), d0), s0);
+        const __m256i err0 = _mm256_madd_epi16(e0, e0);
+        sum32 = _mm256_add_epi32(sum32, err0);
+      }
+      for (k = j; k < width; ++k) {
+        const int32_t u = (int32_t)(dat[k] << SGRPROJ_RST_BITS);
+        int32_t v = xq[1] * (flt1[k] - u);
+        const int32_t e = ROUND_POWER_OF_TWO(v, shift) + dat[k] - src[k];
+        err += e * e;
+      }
+      dat += dat_stride;
+      src += src_stride;
+      flt1 += flt1_stride;
+      const __m256i sum64_0 =
+          _mm256_cvtepi32_epi64(_mm256_castsi256_si128(sum32));
+      const __m256i sum64_1 =
+          _mm256_cvtepi32_epi64(_mm256_extracti128_si256(sum32, 1));
+      sum64 = _mm256_add_epi64(sum64, sum64_0);
+      sum64 = _mm256_add_epi64(sum64, sum64_1);
+    }
+  } else {
+    __m256i sum32 = _mm256_setzero_si256();
+    for (i = 0; i < height; ++i) {
+      for (j = 0; j <= width - 16; j += 16) {
+        const __m256i d0 = _mm256_cvtepu8_epi16(xx_loadu_128(dat + j));
+        const __m256i s0 = _mm256_cvtepu8_epi16(xx_loadu_128(src + j));
+        const __m256i diff0 = _mm256_sub_epi16(d0, s0);
+        const __m256i err0 = _mm256_madd_epi16(diff0, diff0);
+        sum32 = _mm256_add_epi32(sum32, err0);
+      }
+      for (k = j; k < width; ++k) {
+        const int32_t e = (int32_t)(dat[k]) - src[k];
+        err += e * e;
+      }
+      dat += dat_stride;
+      src += src_stride;
+    }
+    const __m256i sum64_0 =
+        _mm256_cvtepi32_epi64(_mm256_castsi256_si128(sum32));
+    const __m256i sum64_1 =
+        _mm256_cvtepi32_epi64(_mm256_extracti128_si256(sum32, 1));
+    sum64 = _mm256_add_epi64(sum64_0, sum64_1);
+  }
+  int64_t sum[4];
+  yy_storeu_256(sum, sum64);
+  err += sum[0] + sum[1] + sum[2] + sum[3];
+  return err;
+}
diff --git a/third_party/aom/av1/encoder/x86/pickrst_sse4.c b/third_party/aom/av1/encoder/x86/pickrst_sse4.c
new file mode 100644
index 000000000..04e4d1afc
--- /dev/null
+++ b/third_party/aom/av1/encoder/x86/pickrst_sse4.c
@@ -0,0 +1,389 @@
+/*
+ * Copyright (c) 2018, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <emmintrin.h>
+#include "aom_dsp/x86/synonyms.h"
+
+#include "config/av1_rtcd.h"
+#include "av1/common/restoration.h"
+#include "av1/encoder/pickrst.h"
+
+static INLINE void acc_stat_sse41(int32_t *dst, const uint8_t *src,
+                                  const __m128i *shuffle, const __m128i *kl) {
+  const __m128i s = _mm_shuffle_epi8(xx_loadu_128(src), *shuffle);
+  const __m128i d0 = _mm_madd_epi16(*kl, _mm_cvtepu8_epi16(s));
+  const __m128i d1 =
+      _mm_madd_epi16(*kl, _mm_cvtepu8_epi16(_mm_srli_si128(s, 8)));
+  const __m128i dst0 = xx_loadu_128(dst);
+  const __m128i dst1 = xx_loadu_128(dst + 4);
+  const __m128i r0 = _mm_add_epi32(dst0, d0);
+  const __m128i r1 = _mm_add_epi32(dst1, d1);
+  xx_storeu_128(dst, r0);
+  xx_storeu_128(dst + 4, r1);
+}
+
+static INLINE void acc_stat_win7_one_line_sse4_1(
+    const uint8_t *dgd, const uint8_t *src, int h_start, int h_end,
+    int dgd_stride, const __m128i *shuffle, int32_t *sumX,
+    int32_t sumY[WIENER_WIN][WIENER_WIN], int32_t M_int[WIENER_WIN][WIENER_WIN],
+    int32_t H_int[WIENER_WIN2][WIENER_WIN * 8]) {
+  const int wiener_win = 7;
+  int j, k, l;
+  for (j = h_start; j < h_end; j += 2) {
+    const uint8_t *dgd_ij = dgd + j;
+    const uint8_t X1 = src[j];
+    const uint8_t X2 = src[j + 1];
+    *sumX += X1 + X2;
+    for (k = 0; k < wiener_win; k++) {
+      const uint8_t *dgd_ijk = dgd_ij + k * dgd_stride;
+      for (l = 0; l < wiener_win; l++) {
+        int32_t *H_ = &H_int[(l * wiener_win + k)][0];
+        const uint8_t D1 = dgd_ijk[l];
+        const uint8_t D2 = dgd_ijk[l + 1];
+        sumY[k][l] += D1 + D2;
+        M_int[k][l] += D1 * X1 + D2 * X2;
+
+        const __m128i kl =
+            _mm_cvtepu8_epi16(_mm_set1_epi16(*((uint16_t *)(dgd_ijk + l))));
+        acc_stat_sse41(H_ + 0 * 8, dgd_ij + 0 * dgd_stride, shuffle, &kl);
+        acc_stat_sse41(H_ + 1 * 8, dgd_ij + 1 * dgd_stride, shuffle, &kl);
+        acc_stat_sse41(H_ + 2 * 8, dgd_ij + 2 * dgd_stride, shuffle, &kl);
+        acc_stat_sse41(H_ + 3 * 8, dgd_ij + 3 * dgd_stride, shuffle, &kl);
+        acc_stat_sse41(H_ + 4 * 8, dgd_ij + 4 * dgd_stride, shuffle, &kl);
+        acc_stat_sse41(H_ + 5 * 8, dgd_ij + 5 * dgd_stride, shuffle, &kl);
+        acc_stat_sse41(H_ + 6 * 8, dgd_ij + 6 * dgd_stride, shuffle, &kl);
+      }
+    }
+  }
+}
+
+static INLINE void compute_stats_win7_opt_sse4_1(
+    const uint8_t *dgd, const uint8_t *src, int h_start, int h_end, int v_start,
+    int v_end, int dgd_stride, int src_stride, double *M, double *H) {
+  int i, j, k, l, m, n;
+  const int wiener_win = WIENER_WIN;
+  const int pixel_count = (h_end - h_start) * (v_end - v_start);
+  const int wiener_win2 = wiener_win * wiener_win;
+  const int wiener_halfwin = (wiener_win >> 1);
+  const double avg =
+      find_average(dgd, h_start, h_end, v_start, v_end, dgd_stride);
+
+  int32_t M_int32[WIENER_WIN][WIENER_WIN] = { { 0 } };
+  int64_t M_int64[WIENER_WIN][WIENER_WIN] = { { 0 } };
+  int32_t H_int32[WIENER_WIN2][WIENER_WIN * 8] = { { 0 } };
+  int64_t H_int64[WIENER_WIN2][WIENER_WIN * 8] = { { 0 } };
+  int32_t sumY[WIENER_WIN][WIENER_WIN] = { { 0 } };
+  int32_t sumX = 0;
+  const uint8_t *dgd_win = dgd - wiener_halfwin * dgd_stride - wiener_halfwin;
+
+  const __m128i shuffle = xx_loadu_128(g_shuffle_stats_data);
+  for (j = v_start; j < v_end; j += 64) {
+    const int vert_end = AOMMIN(64, v_end - j) + j;
+    for (i = j; i < vert_end; i++) {
+      acc_stat_win7_one_line_sse4_1(
+          dgd_win + i * dgd_stride, src + i * src_stride, h_start, h_end,
+          dgd_stride, &shuffle, &sumX, sumY, M_int32, H_int32);
+    }
+    for (k = 0; k < wiener_win; ++k) {
+      for (l = 0; l < wiener_win; ++l) {
+        M_int64[k][l] += M_int32[k][l];
+        M_int32[k][l] = 0;
+      }
+    }
+    for (k = 0; k < WIENER_WIN2; ++k) {
+      for (l = 0; l < WIENER_WIN * 8; ++l) {
+        H_int64[k][l] += H_int32[k][l];
+        H_int32[k][l] = 0;
+      }
+    }
+  }
+
+  const double avg_square_sum = avg * avg * pixel_count;
+  for (k = 0; k < wiener_win; k++) {
+    for (l = 0; l < wiener_win; l++) {
+      const int32_t idx0 = l * wiener_win + k;
+      M[idx0] = M_int64[k][l] + avg_square_sum - avg * (sumX + sumY[k][l]);
+      double *H_ = H + idx0 * wiener_win2;
+      int64_t *H_int_ = &H_int64[idx0][0];
+      for (m = 0; m < wiener_win; m++) {
+        for (n = 0; n < wiener_win; n++) {
+          H_[m * wiener_win + n] = H_int_[n * 8 + m] + avg_square_sum -
+                                   avg * (sumY[k][l] + sumY[n][m]);
+        }
+      }
+    }
+  }
+}
+
+static INLINE void acc_stat_win5_one_line_sse4_1(
+    const uint8_t *dgd, const uint8_t *src, int h_start, int h_end,
+    int dgd_stride, const __m128i *shuffle, int32_t *sumX,
+    int32_t sumY[WIENER_WIN_CHROMA][WIENER_WIN_CHROMA],
+    int32_t M_int[WIENER_WIN_CHROMA][WIENER_WIN_CHROMA],
+    int32_t H_int[WIENER_WIN2_CHROMA][WIENER_WIN_CHROMA * 8]) {
+  const int wiener_win = WIENER_WIN_CHROMA;
+  int j, k, l;
+  for (j = h_start; j < h_end; j += 2) {
+    const uint8_t *dgd_ij = dgd + j;
+    const uint8_t X1 = src[j];
+    const uint8_t X2 = src[j + 1];
+    *sumX += X1 + X2;
+    for (k = 0; k < wiener_win; k++) {
+      const uint8_t *dgd_ijk = dgd_ij + k * dgd_stride;
+      for (l = 0; l < wiener_win; l++) {
+        int32_t *H_ = &H_int[(l * wiener_win + k)][0];
+        const uint8_t D1 = dgd_ijk[l];
+        const uint8_t D2 = dgd_ijk[l + 1];
+        sumY[k][l] += D1 + D2;
+        M_int[k][l] += D1 * X1 + D2 * X2;
+
+        const __m128i kl =
+            _mm_cvtepu8_epi16(_mm_set1_epi16(*((uint16_t *)(dgd_ijk + l))));
+        acc_stat_sse41(H_ + 0 * 8, dgd_ij + 0 * dgd_stride, shuffle, &kl);
+        acc_stat_sse41(H_ + 1 * 8, dgd_ij + 1 * dgd_stride, shuffle, &kl);
+        acc_stat_sse41(H_ + 2 * 8, dgd_ij + 2 * dgd_stride, shuffle, &kl);
+        acc_stat_sse41(H_ + 3 * 8, dgd_ij + 3 * dgd_stride, shuffle, &kl);
+        acc_stat_sse41(H_ + 4 * 8, dgd_ij + 4 * dgd_stride, shuffle, &kl);
+      }
+    }
+  }
+}
+
+static INLINE void compute_stats_win5_opt_sse4_1(
+    const uint8_t *dgd, const uint8_t *src, int h_start, int h_end, int v_start,
+    int v_end, int dgd_stride, int src_stride, double *M, double *H) {
+  int i, j, k, l, m, n;
+  const int wiener_win = WIENER_WIN_CHROMA;
+  const int pixel_count = (h_end - h_start) * (v_end - v_start);
+  const int wiener_win2 = wiener_win * wiener_win;
+  const int wiener_halfwin = (wiener_win >> 1);
+  const double avg =
+      find_average(dgd, h_start, h_end, v_start, v_end, dgd_stride);
+
+  int32_t M_int32[WIENER_WIN_CHROMA][WIENER_WIN_CHROMA] = { { 0 } };
+  int64_t M_int64[WIENER_WIN_CHROMA][WIENER_WIN_CHROMA] = { { 0 } };
+  int32_t H_int32[WIENER_WIN2_CHROMA][WIENER_WIN_CHROMA * 8] = { { 0 } };
+  int64_t H_int64[WIENER_WIN2_CHROMA][WIENER_WIN_CHROMA * 8] = { { 0 } };
+  int32_t sumY[WIENER_WIN_CHROMA][WIENER_WIN_CHROMA] = { { 0 } };
+  int32_t sumX = 0;
+  const uint8_t *dgd_win = dgd - wiener_halfwin * dgd_stride - wiener_halfwin;
+
+  const __m128i shuffle = xx_loadu_128(g_shuffle_stats_data);
+  for (j = v_start; j < v_end; j += 64) {
+    const int vert_end = AOMMIN(64, v_end - j) + j;
+    for (i = j; i < vert_end; i++) {
+      acc_stat_win5_one_line_sse4_1(
+          dgd_win + i * dgd_stride, src + i * src_stride, h_start, h_end,
+          dgd_stride, &shuffle, &sumX, sumY, M_int32, H_int32);
+    }
+    for (k = 0; k < wiener_win; ++k) {
+      for (l = 0; l < wiener_win; ++l) {
+        M_int64[k][l] += M_int32[k][l];
+        M_int32[k][l] = 0;
+      }
+    }
+    for (k = 0; k < WIENER_WIN_CHROMA * WIENER_WIN_CHROMA; ++k) {
+      for (l = 0; l < WIENER_WIN_CHROMA * 8; ++l) {
+        H_int64[k][l] += H_int32[k][l];
+        H_int32[k][l] = 0;
+      }
+    }
+  }
+
+  const double avg_square_sum = avg * avg * pixel_count;
+  for (k = 0; k < wiener_win; k++) {
+    for (l = 0; l < wiener_win; l++) {
+      const int32_t idx0 = l * wiener_win + k;
+      M[idx0] = M_int64[k][l] + avg_square_sum - avg * (sumX + sumY[k][l]);
+      double *H_ = H + idx0 * wiener_win2;
+      int64_t *H_int_ = &H_int64[idx0][0];
+      for (m = 0; m < wiener_win; m++) {
+        for (n = 0; n < wiener_win; n++) {
+          H_[m * wiener_win + n] = H_int_[n * 8 + m] + avg_square_sum -
+                                   avg * (sumY[k][l] + sumY[n][m]);
+        }
+      }
+    }
+  }
+}
+void av1_compute_stats_sse4_1(int wiener_win, const uint8_t *dgd,
+                              const uint8_t *src, int h_start, int h_end,
+                              int v_start, int v_end, int dgd_stride,
+                              int src_stride, double *M, double *H) {
+  if (wiener_win == WIENER_WIN) {
+    compute_stats_win7_opt_sse4_1(dgd, src, h_start, h_end, v_start, v_end,
+                                  dgd_stride, src_stride, M, H);
+  } else if (wiener_win == WIENER_WIN_CHROMA) {
+    compute_stats_win5_opt_sse4_1(dgd, src, h_start, h_end, v_start, v_end,
+                                  dgd_stride, src_stride, M, H);
+  } else {
+    av1_compute_stats_c(wiener_win, dgd, src, h_start, h_end, v_start, v_end,
+                        dgd_stride, src_stride, M, H);
+  }
+}
+
+static INLINE __m128i pair_set_epi16(uint16_t a, uint16_t b) {
+  return _mm_set1_epi32((int32_t)(((uint16_t)(a)) | (((uint32_t)(b)) << 16)));
+}
+
+int64_t av1_lowbd_pixel_proj_error_sse4_1(
+    const uint8_t *src8, int width, int height, int src_stride,
+    const uint8_t *dat8, int dat_stride, int32_t *flt0, int flt0_stride,
+    int32_t *flt1, int flt1_stride, int xq[2], const sgr_params_type *params) {
+  int i, j, k;
+  const int32_t shift = SGRPROJ_RST_BITS + SGRPROJ_PRJ_BITS;
+  const __m128i rounding = _mm_set1_epi32(1 << (shift - 1));
+  __m128i sum64 = _mm_setzero_si128();
+  const uint8_t *src = src8;
+  const uint8_t *dat = dat8;
+  int64_t err = 0;
+  if (params->r[0] > 0 && params->r[1] > 0) {
+    __m128i xq_coeff = pair_set_epi16(xq[0], xq[1]);
+    for (i = 0; i < height; ++i) {
+      __m128i sum32 = _mm_setzero_si128();
+      for (j = 0; j < width - 8; j += 8) {
+        const __m128i d0 = _mm_cvtepu8_epi16(xx_loadl_64(dat + j));
+        const __m128i s0 = _mm_cvtepu8_epi16(xx_loadl_64(src + j));
+        const __m128i flt0_16b =
+            _mm_packs_epi32(xx_loadu_128(flt0 + j), xx_loadu_128(flt0 + j + 4));
+        const __m128i flt1_16b =
+            _mm_packs_epi32(xx_loadu_128(flt1 + j), xx_loadu_128(flt1 + j + 4));
+        const __m128i u0 = _mm_slli_epi16(d0, SGRPROJ_RST_BITS);
+        const __m128i flt0_0_sub_u = _mm_sub_epi16(flt0_16b, u0);
+        const __m128i flt1_0_sub_u = _mm_sub_epi16(flt1_16b, u0);
+        const __m128i v0 = _mm_madd_epi16(
+            xq_coeff, _mm_unpacklo_epi16(flt0_0_sub_u, flt1_0_sub_u));
+        const __m128i v1 = _mm_madd_epi16(
+            xq_coeff, _mm_unpackhi_epi16(flt0_0_sub_u, flt1_0_sub_u));
+        const __m128i vr0 = _mm_srai_epi32(_mm_add_epi32(v0, rounding), shift);
+        const __m128i vr1 = _mm_srai_epi32(_mm_add_epi32(v1, rounding), shift);
+        const __m128i e0 =
+            _mm_sub_epi16(_mm_add_epi16(_mm_packs_epi32(vr0, vr1), d0), s0);
+        const __m128i err0 = _mm_madd_epi16(e0, e0);
+        sum32 = _mm_add_epi32(sum32, err0);
+      }
+      for (k = j; k < width; ++k) {
+        const int32_t u = (int32_t)(dat[k] << SGRPROJ_RST_BITS);
+        int32_t v = xq[0] * (flt0[k] - u) + xq[1] * (flt1[k] - u);
+        const int32_t e = ROUND_POWER_OF_TWO(v, shift) + dat[k] - src[k];
+        err += e * e;
+      }
+      dat += dat_stride;
+      src += src_stride;
+      flt0 += flt0_stride;
+      flt1 += flt1_stride;
+      const __m128i sum64_0 = _mm_cvtepi32_epi64(sum32);
+      const __m128i sum64_1 = _mm_cvtepi32_epi64(_mm_srli_si128(sum32, 8));
+      sum64 = _mm_add_epi64(sum64, sum64_0);
+      sum64 = _mm_add_epi64(sum64, sum64_1);
+    }
+  } else if (params->r[0] > 0) {
+    __m128i xq_coeff = pair_set_epi16(xq[0], -(xq[0] << SGRPROJ_RST_BITS));
+    for (i = 0; i < height; ++i) {
+      __m128i sum32 = _mm_setzero_si128();
+      for (j = 0; j < width - 8; j += 8) {
+        const __m128i d0 = _mm_cvtepu8_epi16(xx_loadl_64(dat + j));
+        const __m128i s0 = _mm_cvtepu8_epi16(xx_loadl_64(src + j));
+        const __m128i flt0_16b =
+            _mm_packs_epi32(xx_loadu_128(flt0 + j), xx_loadu_128(flt0 + j + 4));
+        const __m128i v0 =
+            _mm_madd_epi16(xq_coeff, _mm_unpacklo_epi16(flt0_16b, d0));
+        const __m128i v1 =
+            _mm_madd_epi16(xq_coeff, _mm_unpackhi_epi16(flt0_16b, d0));
+        const __m128i vr0 = _mm_srai_epi32(_mm_add_epi32(v0, rounding), shift);
+        const __m128i vr1 = _mm_srai_epi32(_mm_add_epi32(v1, rounding), shift);
+        const __m128i e0 =
+            _mm_sub_epi16(_mm_add_epi16(_mm_packs_epi32(vr0, vr1), d0), s0);
+        const __m128i err0 = _mm_madd_epi16(e0, e0);
+        sum32 = _mm_add_epi32(sum32, err0);
+      }
+      for (k = j; k < width; ++k) {
+        const int32_t u = (int32_t)(dat[k] << SGRPROJ_RST_BITS);
+        int32_t v = xq[0] * (flt0[k] - u);
+        const int32_t e = ROUND_POWER_OF_TWO(v, shift) + dat[k] - src[k];
+        err += e * e;
+      }
+      dat += dat_stride;
+      src += src_stride;
+      flt0 += flt0_stride;
+      const __m128i sum64_0 = _mm_cvtepi32_epi64(sum32);
+      const __m128i sum64_1 = _mm_cvtepi32_epi64(_mm_srli_si128(sum32, 8));
+      sum64 = _mm_add_epi64(sum64, sum64_0);
+      sum64 = _mm_add_epi64(sum64, sum64_1);
+    }
+  } else if (params->r[1] > 0) {
+    __m128i xq_coeff = pair_set_epi16(xq[1], -(xq[1] << SGRPROJ_RST_BITS));
+    for (i = 0; i < height; ++i) {
+      __m128i sum32 = _mm_setzero_si128();
+      for (j = 0; j < width - 8; j += 8) {
+        const __m128i d0 = _mm_cvtepu8_epi16(xx_loadl_64(dat + j));
+        const __m128i s0 = _mm_cvtepu8_epi16(xx_loadl_64(src + j));
+        const __m128i flt1_16b =
+            _mm_packs_epi32(xx_loadu_128(flt1 + j), xx_loadu_128(flt1 + j + 4));
+        const __m128i v0 =
+            _mm_madd_epi16(xq_coeff, _mm_unpacklo_epi16(flt1_16b, d0));
+        const __m128i v1 =
+            _mm_madd_epi16(xq_coeff, _mm_unpackhi_epi16(flt1_16b, d0));
+        const __m128i vr0 = _mm_srai_epi32(_mm_add_epi32(v0, rounding), shift);
+        const __m128i vr1 = _mm_srai_epi32(_mm_add_epi32(v1, rounding), shift);
+        const __m128i e0 =
+            _mm_sub_epi16(_mm_add_epi16(_mm_packs_epi32(vr0, vr1), d0), s0);
+        const __m128i err0 = _mm_madd_epi16(e0, e0);
+        sum32 = _mm_add_epi32(sum32, err0);
+      }
+      for (k = j; k < width; ++k) {
+        const int32_t u = (int32_t)(dat[k] << SGRPROJ_RST_BITS);
+        int32_t v = xq[1] * (flt1[k] - u);
+        const int32_t e = ROUND_POWER_OF_TWO(v, shift) + dat[k] - src[k];
+        err += e * e;
+      }
+      dat += dat_stride;
+      src += src_stride;
+      flt1 += flt1_stride;
+      const __m128i sum64_0 = _mm_cvtepi32_epi64(sum32);
+      const __m128i sum64_1 = _mm_cvtepi32_epi64(_mm_srli_si128(sum32, 8));
+      sum64 = _mm_add_epi64(sum64, sum64_0);
+      sum64 = _mm_add_epi64(sum64, sum64_1);
+    }
+  } else {
+    __m128i sum32 = _mm_setzero_si128();
+    for (i = 0; i < height; ++i) {
+      for (j = 0; j < width - 16; j += 16) {
+        const __m128i d = xx_loadu_128(dat + j);
+        const __m128i s = xx_loadu_128(src + j);
+        const __m128i d0 = _mm_cvtepu8_epi16(d);
+        const __m128i d1 = _mm_cvtepu8_epi16(_mm_srli_si128(d, 8));
+        const __m128i s0 = _mm_cvtepu8_epi16(s);
+        const __m128i s1 = _mm_cvtepu8_epi16(_mm_srli_si128(s, 8));
+        const __m128i diff0 = _mm_sub_epi16(d0, s0);
+        const __m128i diff1 = _mm_sub_epi16(d1, s1);
+        const __m128i err0 = _mm_madd_epi16(diff0, diff0);
+        const __m128i err1 = _mm_madd_epi16(diff1, diff1);
+        sum32 = _mm_add_epi32(sum32, err0);
+        sum32 = _mm_add_epi32(sum32, err1);
+      }
+      for (k = j; k < width; ++k) {
+        const int32_t e = (int32_t)(dat[k]) - src[k];
+        err += e * e;
+      }
+      dat += dat_stride;
+      src += src_stride;
+    }
+    const __m128i sum64_0 = _mm_cvtepi32_epi64(sum32);
+    const __m128i sum64_1 = _mm_cvtepi32_epi64(_mm_srli_si128(sum32, 8));
+    sum64 = _mm_add_epi64(sum64_0, sum64_1);
+  }
+  int64_t sum[2];
+  xx_storeu_128(sum, sum64);
+  err += sum[0] + sum[1];
+  return err;
+}
diff --git a/third_party/aom/av1/encoder/x86/wedge_utils_avx2.c b/third_party/aom/av1/encoder/x86/wedge_utils_avx2.c
index f776e84c7..2a792f14e 100644
--- a/third_party/aom/av1/encoder/x86/wedge_utils_avx2.c
+++ b/third_party/aom/av1/encoder/x86/wedge_utils_avx2.c
@@ -14,7 +14,7 @@
 #include <smmintrin.h>
 
 #include "aom_dsp/x86/synonyms.h"
-
+#include "aom_dsp/x86/synonyms_avx2.h"
 #include "aom/aom_integer.h"
 
 #include "av1/common/reconinter.h"
@@ -31,7 +31,7 @@ uint64_t av1_wedge_sse_from_residuals_avx2(const int16_t *r1, const int16_t *d,
   uint64_t csse;
 
   const __m256i v_mask_max_w = _mm256_set1_epi16(MAX_MASK_VALUE);
-  const __m256i v_zext_q = _mm256_set1_epi64x(0xffffffff);
+  const __m256i v_zext_q = yy_set1_64_from_32i(0xffffffff);
 
   __m256i v_acc0_q = _mm256_setzero_si256();