Update libaom to rev b25610052a1398032320008d69b51d2da94f5928

32 files changed, 2435 insertions, 387 deletions

diff --git a/third_party/aom/aom_dsp/aom_dsp.cmake b/third_party/aom/aom_dsp/aom_dsp.cmake
index 768875f7d..7c0111a69 100644
--- a/third_party/aom/aom_dsp/aom_dsp.cmake
+++ b/third_party/aom/aom_dsp/aom_dsp.cmake
@@ -83,6 +83,7 @@ list(APPEND AOM_DSP_COMMON_INTRIN_SSE4_1
 list(APPEND AOM_DSP_COMMON_INTRIN_AVX2
             "${AOM_ROOT}/aom_dsp/x86/aom_subpixel_8t_intrin_avx2.c"
             "${AOM_ROOT}/aom_dsp/x86/common_avx2.h"
+            "${AOM_ROOT}/aom_dsp/x86/txfm_common_avx2.h"
             "${AOM_ROOT}/aom_dsp/x86/convolve_avx2.h"
             "${AOM_ROOT}/aom_dsp/x86/fft_avx2.c"
             "${AOM_ROOT}/aom_dsp/x86/highbd_convolve_avx2.c"
@@ -190,13 +191,16 @@ if(CONFIG_AV1_ENCODER)
               "${AOM_ROOT}/aom_dsp/x86/ssim_opt_x86_64.asm")
 
   list(APPEND AOM_DSP_ENCODER_INTRIN_AVX2
+              "${AOM_ROOT}/aom_dsp/x86/masked_sad_intrin_avx2.c"
+              "${AOM_ROOT}/aom_dsp/x86/subtract_avx2.c"
               "${AOM_ROOT}/aom_dsp/x86/highbd_quantize_intrin_avx2.c"
               "${AOM_ROOT}/aom_dsp/x86/sad4d_avx2.c"
               "${AOM_ROOT}/aom_dsp/x86/sad_avx2.c"
               "${AOM_ROOT}/aom_dsp/x86/sad_highbd_avx2.c"
               "${AOM_ROOT}/aom_dsp/x86/sad_impl_avx2.c"
               "${AOM_ROOT}/aom_dsp/x86/variance_avx2.c"
-              "${AOM_ROOT}/aom_dsp/x86/variance_impl_avx2.c")
+              "${AOM_ROOT}/aom_dsp/x86/variance_impl_avx2.c"
+              "${AOM_ROOT}/aom_dsp/x86/obmc_sad_avx2.c")
 
   list(APPEND AOM_DSP_ENCODER_ASM_SSSE3_X86_64
               "${AOM_ROOT}/aom_dsp/x86/quantize_ssse3_x86_64.asm")
@@ -205,9 +209,11 @@ if(CONFIG_AV1_ENCODER)
               "${AOM_ROOT}/aom_dsp/x86/quantize_avx_x86_64.asm")
 
   list(APPEND AOM_DSP_ENCODER_INTRIN_SSSE3
+              "${AOM_ROOT}/aom_dsp/x86/masked_sad_intrin_ssse3.h"
               "${AOM_ROOT}/aom_dsp/x86/masked_sad_intrin_ssse3.c"
               "${AOM_ROOT}/aom_dsp/x86/masked_variance_intrin_ssse3.h"
               "${AOM_ROOT}/aom_dsp/x86/masked_variance_intrin_ssse3.c"
+              "${AOM_ROOT}/aom_dsp/x86/variance_impl_ssse3.c"
               "${AOM_ROOT}/aom_dsp/x86/jnt_variance_ssse3.c"
               "${AOM_ROOT}/aom_dsp/x86/jnt_sad_ssse3.c")
 
diff --git a/third_party/aom/aom_dsp/aom_dsp_rtcd.c b/third_party/aom/aom_dsp/aom_dsp_rtcd.c
index 5d7d4515b..1514bd64e 100644
--- a/third_party/aom/aom_dsp/aom_dsp_rtcd.c
+++ b/third_party/aom/aom_dsp/aom_dsp_rtcd.c
@@ -15,4 +15,4 @@
 
 #include "aom_ports/aom_once.h"
 
-void aom_dsp_rtcd() { once(setup_rtcd_internal); }
+void aom_dsp_rtcd() { aom_once(setup_rtcd_internal); }
diff --git a/third_party/aom/aom_dsp/aom_dsp_rtcd_defs.pl b/third_party/aom/aom_dsp/aom_dsp_rtcd_defs.pl
index a8ac5eb5c..1a9ac3184 100755
--- a/third_party/aom/aom_dsp/aom_dsp_rtcd_defs.pl
+++ b/third_party/aom/aom_dsp/aom_dsp_rtcd_defs.pl
@@ -377,7 +377,7 @@ add_proto qw/void aom_lpf_vertical_14_dual/, "uint8_t *s, int pitch, const uint8
 specialize qw/aom_lpf_vertical_14_dual sse2/;
 
 add_proto qw/void aom_lpf_vertical_6/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
-specialize qw/aom_lpf_vertical_6 sse2/;
+specialize qw/aom_lpf_vertical_6 sse2 neon/;
 
 add_proto qw/void aom_lpf_vertical_8/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
 specialize qw/aom_lpf_vertical_8 sse2 neon/;
@@ -386,13 +386,13 @@ add_proto qw/void aom_lpf_vertical_8_dual/, "uint8_t *s, int pitch, const uint8_
 specialize qw/aom_lpf_vertical_8_dual sse2/;
 
 add_proto qw/void aom_lpf_vertical_4/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
-specialize qw/aom_lpf_vertical_4 sse2/;
+specialize qw/aom_lpf_vertical_4 sse2 neon/;
 
 add_proto qw/void aom_lpf_vertical_4_dual/, "uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1";
 specialize qw/aom_lpf_vertical_4_dual sse2/;
 
 add_proto qw/void aom_lpf_horizontal_14/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
-specialize qw/aom_lpf_horizontal_14 sse2/;
+specialize qw/aom_lpf_horizontal_14 sse2 neon/;
 
 add_proto qw/void aom_lpf_horizontal_14_dual/, "uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1";
 specialize qw/aom_lpf_horizontal_14_dual sse2/;
@@ -410,7 +410,7 @@ add_proto qw/void aom_lpf_horizontal_8_dual/, "uint8_t *s, int pitch, const uint
 specialize qw/aom_lpf_horizontal_8_dual sse2/;
 
 add_proto qw/void aom_lpf_horizontal_4/, "uint8_t *s, int pitch, const uint8_t *blimit, const uint8_t *limit, const uint8_t *thresh";
-specialize qw/aom_lpf_horizontal_4 sse2/;
+specialize qw/aom_lpf_horizontal_4 sse2 neon/;
 
 add_proto qw/void aom_lpf_horizontal_4_dual/, "uint8_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1";
 specialize qw/aom_lpf_horizontal_4_dual sse2/;
@@ -564,7 +564,7 @@ if (aom_config("CONFIG_AV1_ENCODER") eq "yes") {
   # Block subtraction
   #
   add_proto qw/void aom_subtract_block/, "int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride";
-  specialize qw/aom_subtract_block neon msa sse2/;
+  specialize qw/aom_subtract_block neon msa sse2 avx2/;
 
   add_proto qw/void aom_highbd_subtract_block/, "int rows, int cols, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr, ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride, int bd";
   specialize qw/aom_highbd_subtract_block sse2/;
@@ -732,14 +732,14 @@ if (aom_config("CONFIG_AV1_ENCODER") eq "yes") {
   foreach (@block_sizes) {
     ($w, $h) = @$_;
     add_proto qw/unsigned int/, "aom_masked_sad${w}x${h}", "const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, const uint8_t *second_pred, const uint8_t *msk, int msk_stride, int invert_mask";
-    specialize "aom_masked_sad${w}x${h}", qw/ssse3/;
+    specialize "aom_masked_sad${w}x${h}", qw/ssse3 avx2/;
   }
 
 
     foreach (@block_sizes) {
       ($w, $h) = @$_;
       add_proto qw/unsigned int/, "aom_highbd_masked_sad${w}x${h}", "const uint8_t *src8, int src_stride, const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, int msk_stride, int invert_mask";
-      specialize "aom_highbd_masked_sad${w}x${h}", qw/ssse3/;
+      specialize "aom_highbd_masked_sad${w}x${h}", qw/ssse3 avx2/;
     }
 
 
@@ -750,7 +750,7 @@ if (aom_config("CONFIG_AV1_ENCODER") eq "yes") {
     ($w, $h) = @$_;
     add_proto qw/unsigned int/, "aom_obmc_sad${w}x${h}", "const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask";
     if (! (($w == 128 && $h == 32) || ($w == 32 && $h == 128))) {
-       specialize "aom_obmc_sad${w}x${h}", qw/sse4_1/;
+       specialize "aom_obmc_sad${w}x${h}", qw/sse4_1 avx2/;
     }
   }
 
@@ -759,7 +759,7 @@ if (aom_config("CONFIG_AV1_ENCODER") eq "yes") {
       ($w, $h) = @$_;
       add_proto qw/unsigned int/, "aom_highbd_obmc_sad${w}x${h}", "const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask";
       if (! (($w == 128 && $h == 32) || ($w == 32 && $h == 128))) {
-        specialize "aom_highbd_obmc_sad${w}x${h}", qw/sse4_1/;
+        specialize "aom_highbd_obmc_sad${w}x${h}", qw/sse4_1 avx2/;
       }
     }
 
@@ -1102,6 +1102,7 @@ if (aom_config("CONFIG_AV1_ENCODER") eq "yes") {
     add_proto qw/unsigned int/, "aom_obmc_variance${w}x${h}", "const uint8_t *pre, int pre_stride, const int32_t *wsrc, const int32_t *mask, unsigned int *sse";
     add_proto qw/unsigned int/, "aom_obmc_sub_pixel_variance${w}x${h}", "const uint8_t *pre, int pre_stride, int xoffset, int yoffset, const int32_t *wsrc, const int32_t *mask, unsigned int *sse";
     specialize "aom_obmc_variance${w}x${h}", q/sse4_1/;
+    specialize "aom_obmc_sub_pixel_variance${w}x${h}", q/sse4_1/;
   }
 
 
@@ -1539,9 +1540,7 @@ if (aom_config("CONFIG_AV1_ENCODER") eq "yes") {
   specialize qw/aom_comp_mask_pred ssse3 avx2/;
 
   add_proto qw/void aom_highbd_comp_mask_pred/, "uint16_t *comp_pred, const uint8_t *pred8, int width, int height, const uint8_t *ref8, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask";
-  add_proto qw/void aom_highbd_comp_mask_upsampled_pred/, "MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col, const MV *const mv, uint16_t *comp_pred, const uint8_t *pred8, int width,
-                                                           int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8, int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask, int bd";
-
+  specialize qw/aom_highbd_comp_mask_pred avx2/;
 
 }  # CONFIG_AV1_ENCODER
 
diff --git a/third_party/aom/aom_dsp/arm/intrapred_neon.c b/third_party/aom/aom_dsp/arm/intrapred_neon.c
index 69470eeb0..c85b1e910 100644
--- a/third_party/aom/aom_dsp/arm/intrapred_neon.c
+++ b/third_party/aom/aom_dsp/arm/intrapred_neon.c
@@ -528,3 +528,63 @@ void aom_h_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
     }
   }
 }
+
+static INLINE void highbd_dc_predictor(uint16_t *dst, ptrdiff_t stride, int bw,
+                                       const uint16_t *above,
+                                       const uint16_t *left) {
+  assert(bw >= 4);
+  assert(IS_POWER_OF_TWO(bw));
+  int expected_dc, sum = 0;
+  const int count = bw * 2;
+  uint32x4_t sum_q = vdupq_n_u32(0);
+  uint32x2_t sum_d;
+  uint16_t *dst_1;
+  if (bw >= 8) {
+    for (int i = 0; i < bw; i += 8) {
+      sum_q = vpadalq_u16(sum_q, vld1q_u16(above));
+      sum_q = vpadalq_u16(sum_q, vld1q_u16(left));
+      above += 8;
+      left += 8;
+    }
+    sum_d = vadd_u32(vget_low_u32(sum_q), vget_high_u32(sum_q));
+    sum = vget_lane_s32(vreinterpret_s32_u64(vpaddl_u32(sum_d)), 0);
+    expected_dc = (sum + (count >> 1)) / count;
+    const uint16x8_t dc = vdupq_n_u16((uint16_t)expected_dc);
+    for (int r = 0; r < bw; r++) {
+      dst_1 = dst;
+      for (int i = 0; i < bw; i += 8) {
+        vst1q_u16(dst_1, dc);
+        dst_1 += 8;
+      }
+      dst += stride;
+    }
+  } else {  // 4x4
+    sum_q = vaddl_u16(vld1_u16(above), vld1_u16(left));
+    sum_d = vadd_u32(vget_low_u32(sum_q), vget_high_u32(sum_q));
+    sum = vget_lane_s32(vreinterpret_s32_u64(vpaddl_u32(sum_d)), 0);
+    expected_dc = (sum + (count >> 1)) / count;
+    const uint16x4_t dc = vdup_n_u16((uint16_t)expected_dc);
+    for (int r = 0; r < bw; r++) {
+      vst1_u16(dst, dc);
+      dst += stride;
+    }
+  }
+}
+
+#define intra_pred_highbd_sized_neon(type, width)               \
+  void aom_highbd_##type##_predictor_##width##x##width##_neon(  \
+      uint16_t *dst, ptrdiff_t stride, const uint16_t *above,   \
+      const uint16_t *left, int bd) {                           \
+    (void)bd;                                                   \
+    highbd_##type##_predictor(dst, stride, width, above, left); \
+  }
+
+#define intra_pred_square(type)           \
+  intra_pred_highbd_sized_neon(type, 4);  \
+  intra_pred_highbd_sized_neon(type, 8);  \
+  intra_pred_highbd_sized_neon(type, 16); \
+  intra_pred_highbd_sized_neon(type, 32); \
+  intra_pred_highbd_sized_neon(type, 64);
+
+intra_pred_square(dc);
+#undef intra_pred_square
diff --git a/third_party/aom/aom_dsp/arm/loopfilter_neon.c b/third_party/aom/aom_dsp/arm/loopfilter_neon.c
index ee1a3c78f..bdc67626d 100644
--- a/third_party/aom/aom_dsp/arm/loopfilter_neon.c
+++ b/third_party/aom/aom_dsp/arm/loopfilter_neon.c
@@ -52,6 +52,36 @@ static INLINE uint8x8_t lpf_mask(uint8x8_t p3q3, uint8x8_t p2q2, uint8x8_t p1q1,
   return mask_8x8;
 }
 
+static INLINE uint8x8_t lpf_mask2(uint8x8_t p1q1, uint8x8_t p0q0,
+                                  const uint8_t blimit, const uint8_t limit) {
+  uint32x2x2_t p0q0_p1q1;
+  uint16x8_t temp_16x8;
+  uint16x4_t temp0_16x4, temp1_16x4;
+  const uint16x4_t blimit_16x4 = vdup_n_u16(blimit);
+  const uint8x8_t limit_8x8 = vdup_n_u8(limit);
+  uint8x8_t mask_8x8, temp_8x8;
+
+  mask_8x8 = vabd_u8(p1q1, p0q0);
+  mask_8x8 = vcle_u8(mask_8x8, limit_8x8);
+
+  temp_8x8 = vreinterpret_u8_u32(vrev64_u32(vreinterpret_u32_u8(mask_8x8)));
+  mask_8x8 = vand_u8(mask_8x8, temp_8x8);
+
+  p0q0_p1q1 = vtrn_u32(vreinterpret_u32_u8(p0q0), vreinterpret_u32_u8(p1q1));
+  temp_8x8 = vabd_u8(vreinterpret_u8_u32(p0q0_p1q1.val[0]),
+                     vreinterpret_u8_u32(p0q0_p1q1.val[1]));
+  temp_16x8 = vmovl_u8(temp_8x8);
+  temp0_16x4 = vshl_n_u16(vget_low_u16(temp_16x8), 1);
+  temp1_16x4 = vshr_n_u16(vget_high_u16(temp_16x8), 1);
+  temp0_16x4 = vadd_u16(temp0_16x4, temp1_16x4);
+  temp0_16x4 = vcle_u16(temp0_16x4, blimit_16x4);
+  temp_8x8 = vmovn_u16(vcombine_u16(temp0_16x4, temp0_16x4));
+
+  mask_8x8 = vand_u8(mask_8x8, temp_8x8);
+
+  return mask_8x8;
+}
+
 static INLINE uint8x8_t lpf_flat_mask4(uint8x8_t p3q3, uint8x8_t p2q2,
                                        uint8x8_t p1q1, uint8x8_t p0q0) {
   const uint8x8_t thresh_8x8 = vdup_n_u8(1);  // for bd==8 threshold is always 1
@@ -523,6 +553,68 @@ static void lpf_6_neon(uint8x8_t *p2q2, uint8x8_t *p1q1, uint8x8_t *p0q0,
   }
 }
 
+static void lpf_4_neon(uint8x8_t *p1q1, uint8x8_t *p0q0, const uint8_t blimit,
+                       const uint8_t limit, const uint8_t thresh) {
+  int32x2x2_t ps0_qs0, ps1_qs1;
+  int16x8_t filter_s16;
+  const uint8x8_t thresh_f4 = vdup_n_u8(thresh);
+  uint8x8_t mask_8x8, temp0_8x8, temp1_8x8;
+  int8x8_t ps0_s8, ps1_s8, qs0_s8, qs1_s8, temp_s8;
+  int8x8_t op0, oq0, op1, oq1;
+  int8x8_t pq_s0, pq_s1;
+  int8x8_t filter_s8, filter1_s8, filter2_s8;
+  int8x8_t hev_8x8;
+  const int8x8_t sign_mask = vdup_n_s8(0x80);
+  const int8x8_t val_4 = vdup_n_s8(4);
+  const int8x8_t val_3 = vdup_n_s8(3);
+
+  // Calculate filter mask
+  mask_8x8 = lpf_mask2(*p1q1, *p0q0, blimit, limit);
+
+  pq_s0 = veor_s8(vreinterpret_s8_u8(*p0q0), sign_mask);
+  pq_s1 = veor_s8(vreinterpret_s8_u8(*p1q1), sign_mask);
+
+  ps0_qs0 = vtrn_s32(vreinterpret_s32_s8(pq_s0), vreinterpret_s32_s8(pq_s0));
+  ps1_qs1 = vtrn_s32(vreinterpret_s32_s8(pq_s1), vreinterpret_s32_s8(pq_s1));
+  ps0_s8 = vreinterpret_s8_s32(ps0_qs0.val[0]);
+  qs0_s8 = vreinterpret_s8_s32(ps0_qs0.val[1]);
+  ps1_s8 = vreinterpret_s8_s32(ps1_qs1.val[0]);
+  qs1_s8 = vreinterpret_s8_s32(ps1_qs1.val[1]);
+
+  // hev_mask
+  temp0_8x8 = vcgt_u8(vabd_u8(*p0q0, *p1q1), thresh_f4);
+  temp1_8x8 = vreinterpret_u8_u32(vrev64_u32(vreinterpret_u32_u8(temp0_8x8)));
+  hev_8x8 = vreinterpret_s8_u8(vorr_u8(temp0_8x8, temp1_8x8));
+
+  // add outer taps if we have high edge variance
+  filter_s8 = vqsub_s8(ps1_s8, qs1_s8);
+  filter_s8 = vand_s8(filter_s8, hev_8x8);
+
+  // inner taps
+  temp_s8 = vqsub_s8(qs0_s8, ps0_s8);
+  filter_s16 = vmovl_s8(filter_s8);
+  filter_s16 = vmlal_s8(filter_s16, temp_s8, val_3);
+  filter_s8 = vqmovn_s16(filter_s16);
+  filter_s8 = vand_s8(filter_s8, vreinterpret_s8_u8(mask_8x8));
+
+  filter1_s8 = vqadd_s8(filter_s8, val_4);
+  filter2_s8 = vqadd_s8(filter_s8, val_3);
+  filter1_s8 = vshr_n_s8(filter1_s8, 3);
+  filter2_s8 = vshr_n_s8(filter2_s8, 3);
+
+  oq0 = veor_s8(vqsub_s8(qs0_s8, filter1_s8), sign_mask);
+  op0 = veor_s8(vqadd_s8(ps0_s8, filter2_s8), sign_mask);
+
+  filter_s8 = vrshr_n_s8(filter1_s8, 1);
+  filter_s8 = vbic_s8(filter_s8, hev_8x8);
+
+  oq1 = veor_s8(vqsub_s8(qs1_s8, filter_s8), sign_mask);
+  op1 = veor_s8(vqadd_s8(ps1_s8, filter_s8), sign_mask);
+
+  *p0q0 = vreinterpret_u8_s8(vext_s8(op0, oq0, 4));
+  *p1q1 = vreinterpret_u8_s8(vext_s8(op1, oq1, 4));
+}
+
 void aom_lpf_vertical_14_neon(uint8_t *src, int stride, const uint8_t *blimit,
                               const uint8_t *limit, const uint8_t *thresh) {
   uint8x16_t row0, row1, row2, row3;
@@ -646,6 +738,125 @@ void aom_lpf_vertical_8_neon(uint8_t *src, int stride, const uint8_t *blimit,
   store_u8_8x4(src - 4, stride, p3q0, p2q1, p1q2, p0q3);
 }
 
+void aom_lpf_vertical_6_neon(uint8_t *src, int stride, const uint8_t *blimit,
+                             const uint8_t *limit, const uint8_t *thresh) {
+  uint32x2x2_t p2q2_p1q1, pxqy_p0q0;
+  uint32x2_t pq_rev;
+  uint8x8_t pxq0, p2q1, p1q2, p0qy;
+  uint8x8_t p0q0, p1q1, p2q2, pxqy;
+
+  // row0: px p2 p1 p0 | q0 q1 q2 qy
+  // row1: px p2 p1 p0 | q0 q1 q2 qy
+  // row2: px p2 p1 p0 | q0 q1 q2 qy
+  // row3: px p2 p1 p0 | q0 q1 q2 qy
+  load_u8_8x4(src - 4, stride, &pxq0, &p2q1, &p1q2, &p0qy);
+
+  transpose_u8_8x4(&pxq0, &p2q1, &p1q2, &p0qy);
+
+  pq_rev = vrev64_u32(vreinterpret_u32_u8(p0qy));
+  pxqy_p0q0 = vtrn_u32(vreinterpret_u32_u8(pxq0), pq_rev);
+
+  pq_rev = vrev64_u32(vreinterpret_u32_u8(p1q2));
+  p2q2_p1q1 = vtrn_u32(vreinterpret_u32_u8(p2q1), pq_rev);
+
+  p0q0 = vreinterpret_u8_u32(vrev64_u32(pxqy_p0q0.val[1]));
+  p1q1 = vreinterpret_u8_u32(vrev64_u32(p2q2_p1q1.val[1]));
+  p2q2 = vreinterpret_u8_u32(p2q2_p1q1.val[0]);
+  pxqy = vreinterpret_u8_u32(pxqy_p0q0.val[0]);
+
+  lpf_6_neon(&p2q2, &p1q1, &p0q0, *blimit, *limit, *thresh);
+
+  pq_rev = vrev64_u32(vreinterpret_u32_u8(p0q0));
+  pxqy_p0q0 = vtrn_u32(vreinterpret_u32_u8(pxqy), pq_rev);
+
+  pq_rev = vrev64_u32(vreinterpret_u32_u8(p1q1));
+  p2q2_p1q1 = vtrn_u32(vreinterpret_u32_u8(p2q2), pq_rev);
+
+  p0qy = vreinterpret_u8_u32(vrev64_u32(pxqy_p0q0.val[1]));
+  p1q2 = vreinterpret_u8_u32(vrev64_u32(p2q2_p1q1.val[1]));
+  p2q1 = vreinterpret_u8_u32(p2q2_p1q1.val[0]);
+  pxq0 = vreinterpret_u8_u32(pxqy_p0q0.val[0]);
+  transpose_u8_8x4(&pxq0, &p2q1, &p1q2, &p0qy);
+
+  store_u8_8x4(src - 4, stride, pxq0, p2q1, p1q2, p0qy);
+}
+
+void aom_lpf_vertical_4_neon(uint8_t *src, int stride, const uint8_t *blimit,
+                             const uint8_t *limit, const uint8_t *thresh) {
+  uint32x2x2_t p1q0_p0q1, p1q1_p0q0, p1p0_q1q0;
+  uint32x2_t pq_rev;
+  uint8x8_t UNINITIALIZED_IS_SAFE(p1p0), q0q1, p0q0, p1q1;
+
+  // row0: p1 p0 | q0 q1
+  // row1: p1 p0 | q0 q1
+  // row2: p1 p0 | q0 q1
+  // row3: p1 p0 | q0 q1
+  load_u8_4x1(src - 2, &p1p0, 0);
+  load_u8_4x1((src - 2) + 1 * stride, &p1p0, 1);
+  load_u8_4x1((src - 2) + 2 * stride, &q0q1, 0);
+  load_u8_4x1((src - 2) + 3 * stride, &q0q1, 1);
+
+  transpose_u8_4x4(&p1p0, &q0q1);
+
+  p1q0_p0q1 = vtrn_u32(vreinterpret_u32_u8(p1p0), vreinterpret_u32_u8(q0q1));
+
+  pq_rev = vrev64_u32(p1q0_p0q1.val[1]);
+  p1q1_p0q0 = vtrn_u32(p1q0_p0q1.val[0], pq_rev);
+
+  p1q1 = vreinterpret_u8_u32(p1q1_p0q0.val[0]);
+  p0q0 = vreinterpret_u8_u32(p1q1_p0q0.val[1]);
+
+  lpf_4_neon(&p1q1, &p0q0, *blimit, *limit, *thresh);
+
+  p1p0_q1q0 = vtrn_u32(vreinterpret_u32_u8(p1q1), vreinterpret_u32_u8(p0q0));
+
+  p1p0 = vreinterpret_u8_u32(p1p0_q1q0.val[0]);
+  q0q1 = vreinterpret_u8_u32(vrev64_u32(p1p0_q1q0.val[1]));
+
+  transpose_u8_4x4(&p1p0, &q0q1);
+
+  store_u8_4x1(src - 2, p1p0, 0);
+  store_u8_4x1((src - 2) + 1 * stride, q0q1, 0);
+  store_u8_4x1((src - 2) + 2 * stride, p1p0, 1);
+  store_u8_4x1((src - 2) + 3 * stride, q0q1, 1);
+}
+
+void aom_lpf_horizontal_14_neon(uint8_t *src, int stride, const uint8_t *blimit,
+                                const uint8_t *limit, const uint8_t *thresh) {
+  uint8x8_t p0q0, p1q1, p2q2, p3q3, p4q4, p5q5, UNINITIALIZED_IS_SAFE(p6q6);
+
+  load_u8_4x1(src - 7 * stride, &p6q6, 0);
+  load_u8_4x1(src - 6 * stride, &p5q5, 0);
+  load_u8_4x1(src - 5 * stride, &p4q4, 0);
+  load_u8_4x1(src - 4 * stride, &p3q3, 0);
+  load_u8_4x1(src - 3 * stride, &p2q2, 0);
+  load_u8_4x1(src - 2 * stride, &p1q1, 0);
+  load_u8_4x1(src - 1 * stride, &p0q0, 0);
+  load_u8_4x1(src + 0 * stride, &p0q0, 1);
+  load_u8_4x1(src + 1 * stride, &p1q1, 1);
+  load_u8_4x1(src + 2 * stride, &p2q2, 1);
+  load_u8_4x1(src + 3 * stride, &p3q3, 1);
+  load_u8_4x1(src + 4 * stride, &p4q4, 1);
+  load_u8_4x1(src + 5 * stride, &p5q5, 1);
+  load_u8_4x1(src + 6 * stride, &p6q6, 1);
+
+  lpf_14_neon(&p6q6, &p5q5, &p4q4, &p3q3, &p2q2, &p1q1, &p0q0, *blimit, *limit,
+              *thresh);
+
+  store_u8_4x1(src - 6 * stride, p5q5, 0);
+  store_u8_4x1(src - 5 * stride, p4q4, 0);
+  store_u8_4x1(src - 4 * stride, p3q3, 0);
+  store_u8_4x1(src - 3 * stride, p2q2, 0);
+  store_u8_4x1(src - 2 * stride, p1q1, 0);
+  store_u8_4x1(src - 1 * stride, p0q0, 0);
+  store_u8_4x1(src + 0 * stride, p0q0, 1);
+  store_u8_4x1(src + 1 * stride, p1q1, 1);
+  store_u8_4x1(src + 2 * stride, p2q2, 1);
+  store_u8_4x1(src + 3 * stride, p3q3, 1);
+  store_u8_4x1(src + 4 * stride, p4q4, 1);
+  store_u8_4x1(src + 5 * stride, p5q5, 1);
+}
+
 void aom_lpf_horizontal_8_neon(uint8_t *src, int stride, const uint8_t *blimit,
                                const uint8_t *limit, const uint8_t *thresh) {
   uint8x8_t p0q0, p1q1, p2q2, p3q3;
@@ -698,3 +909,20 @@ void aom_lpf_horizontal_6_neon(uint8_t *src, int stride, const uint8_t *blimit,
   vst1_lane_u32((uint32_t *)(src + 1 * stride), vreinterpret_u32_u8(p1q1), 1);
   vst1_lane_u32((uint32_t *)(src + 2 * stride), vreinterpret_u32_u8(p2q2), 1);
 }
+
+void aom_lpf_horizontal_4_neon(uint8_t *src, int stride, const uint8_t *blimit,
+                               const uint8_t *limit, const uint8_t *thresh) {
+  uint8x8_t p0q0, UNINITIALIZED_IS_SAFE(p1q1);
+
+  load_u8_4x1(src - 2 * stride, &p1q1, 0);
+  load_u8_4x1(src - 1 * stride, &p0q0, 0);
+  load_u8_4x1(src + 0 * stride, &p0q0, 1);
+  load_u8_4x1(src + 1 * stride, &p1q1, 1);
+
+  lpf_4_neon(&p1q1, &p0q0, *blimit, *limit, *thresh);
+
+  store_u8_4x1(src - 2 * stride, p1q1, 0);
+  store_u8_4x1(src - 1 * stride, p0q0, 0);
+  store_u8_4x1(src + 0 * stride, p0q0, 1);
+  store_u8_4x1(src + 1 * stride, p1q1, 1);
+}
diff --git a/third_party/aom/aom_dsp/bitreader_buffer.c b/third_party/aom/aom_dsp/bitreader_buffer.c
index 68fc381f2..02b5ef924 100644
--- a/third_party/aom/aom_dsp/bitreader_buffer.c
+++ b/third_party/aom/aom_dsp/bitreader_buffer.c
@@ -8,11 +8,14 @@
  * 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 "config/aom_config.h"
 
 #include "aom_dsp/bitreader_buffer.h"
 
-size_t aom_rb_bytes_read(struct aom_read_bit_buffer *rb) {
+size_t aom_rb_bytes_read(const struct aom_read_bit_buffer *rb) {
   return (rb->bit_offset + 7) >> 3;
 }
 
@@ -31,6 +34,7 @@ int aom_rb_read_bit(struct aom_read_bit_buffer *rb) {
 }
 
 int aom_rb_read_literal(struct aom_read_bit_buffer *rb, int bits) {
+  assert(bits <= 31);
   int value = 0, bit;
   for (bit = bits - 1; bit >= 0; bit--) value |= aom_rb_read_bit(rb) << bit;
   return value;
@@ -38,6 +42,7 @@ int aom_rb_read_literal(struct aom_read_bit_buffer *rb, int bits) {
 
 uint32_t aom_rb_read_unsigned_literal(struct aom_read_bit_buffer *rb,
                                       int bits) {
+  assert(bits <= 32);
   uint32_t value = 0;
   int bit;
   for (bit = bits - 1; bit >= 0; bit--)
diff --git a/third_party/aom/aom_dsp/bitreader_buffer.h b/third_party/aom/aom_dsp/bitreader_buffer.h
index 2dafe11ad..5c94ab883 100644
--- a/third_party/aom/aom_dsp/bitreader_buffer.h
+++ b/third_party/aom/aom_dsp/bitreader_buffer.h
@@ -31,7 +31,7 @@ struct aom_read_bit_buffer {
   aom_rb_error_handler error_handler;
 };
 
-size_t aom_rb_bytes_read(struct aom_read_bit_buffer *rb);
+size_t aom_rb_bytes_read(const struct aom_read_bit_buffer *rb);
 
 int aom_rb_read_bit(struct aom_read_bit_buffer *rb);
 
diff --git a/third_party/aom/aom_dsp/bitwriter_buffer.c b/third_party/aom/aom_dsp/bitwriter_buffer.c
index 21314eb2a..a563bf684 100644
--- a/third_party/aom/aom_dsp/bitwriter_buffer.c
+++ b/third_party/aom/aom_dsp/bitwriter_buffer.c
@@ -9,6 +9,7 @@
  * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
  */
 
+#include <assert.h>
 #include <limits.h>
 #include <stdlib.h>
 
@@ -49,12 +50,14 @@ void aom_wb_overwrite_bit(struct aom_write_bit_buffer *wb, int bit) {
 }
 
 void aom_wb_write_literal(struct aom_write_bit_buffer *wb, int data, int bits) {
+  assert(bits <= 31);
   int bit;
   for (bit = bits - 1; bit >= 0; bit--) aom_wb_write_bit(wb, (data >> bit) & 1);
 }
 
 void aom_wb_write_unsigned_literal(struct aom_write_bit_buffer *wb,
                                    uint32_t data, int bits) {
+  assert(bits <= 32);
   int bit;
   for (bit = bits - 1; bit >= 0; bit--) aom_wb_write_bit(wb, (data >> bit) & 1);
 }
diff --git a/third_party/aom/aom_dsp/grain_synthesis.c b/third_party/aom/aom_dsp/grain_synthesis.c
index fcb6c290e..ff1ec41a2 100644
--- a/third_party/aom/aom_dsp/grain_synthesis.c
+++ b/third_party/aom/aom_dsp/grain_synthesis.c
@@ -17,6 +17,7 @@
 #include <stdio.h>
 #include <string.h>
 #include <stdlib.h>
+#include <assert.h>
 #include "aom_dsp/grain_synthesis.h"
 #include "aom_mem/aom_mem.h"
 
@@ -237,7 +238,7 @@ static int grain_max;
 
 static uint16_t random_register = 0;  // random number generator register
 
-static void init_arrays(aom_film_grain_t *params, int luma_stride,
+static void init_arrays(const aom_film_grain_t *params, int luma_stride,
                         int chroma_stride, int ***pred_pos_luma_p,
                         int ***pred_pos_chroma_p, int **luma_grain_block,
                         int **cb_grain_block, int **cr_grain_block,
@@ -331,7 +332,7 @@ static void init_arrays(aom_film_grain_t *params, int luma_stride,
       (int *)aom_malloc(sizeof(**cr_grain_block) * chroma_grain_samples);
 }
 
-static void dealloc_arrays(aom_film_grain_t *params, int ***pred_pos_luma,
+static void dealloc_arrays(const aom_film_grain_t *params, int ***pred_pos_luma,
                            int ***pred_pos_chroma, int **luma_grain_block,
                            int **cb_grain_block, int **cr_grain_block,
                            int **y_line_buf, int **cb_line_buf,
@@ -396,10 +397,14 @@ static void init_random_generator(int luma_line, uint16_t seed) {
 }
 
 static void generate_luma_grain_block(
-    aom_film_grain_t *params, int **pred_pos_luma, int *luma_grain_block,
+    const aom_film_grain_t *params, int **pred_pos_luma, int *luma_grain_block,
     int luma_block_size_y, int luma_block_size_x, int luma_grain_stride,
     int left_pad, int top_pad, int right_pad, int bottom_pad) {
-  if (params->num_y_points == 0) return;
+  if (params->num_y_points == 0) {
+    memset(luma_grain_block, 0,
+           sizeof(*luma_grain_block) * luma_block_size_y * luma_grain_stride);
+    return;
+  }
 
   int bit_depth = params->bit_depth;
   int gauss_sec_shift = 12 - bit_depth + params->grain_scale_shift;
@@ -431,7 +436,7 @@ static void generate_luma_grain_block(
 }
 
 static void generate_chroma_grain_blocks(
-    aom_film_grain_t *params,
+    const aom_film_grain_t *params,
     //                                  int** pred_pos_luma,
     int **pred_pos_chroma, int *luma_grain_block, int *cb_grain_block,
     int *cr_grain_block, int luma_grain_stride, int chroma_block_size_y,
@@ -443,7 +448,7 @@ static void generate_chroma_grain_blocks(
   int num_pos_chroma = 2 * params->ar_coeff_lag * (params->ar_coeff_lag + 1);
   if (params->num_y_points > 0) ++num_pos_chroma;
   int rounding_offset = (1 << (params->ar_coeff_shift - 1));
-  int chroma_grain_samples = chroma_block_size_y * chroma_block_size_x;
+  int chroma_grain_block_size = chroma_block_size_y * chroma_grain_stride;
 
   if (params->num_cb_points || params->chroma_scaling_from_luma) {
     init_random_generator(7 << 5, params->random_seed);
@@ -455,7 +460,8 @@ static void generate_chroma_grain_blocks(
              ((1 << gauss_sec_shift) >> 1)) >>
             gauss_sec_shift;
   } else {
-    memset(cr_grain_block, 0, sizeof(*cr_grain_block) * chroma_grain_samples);
+    memset(cb_grain_block, 0,
+           sizeof(*cb_grain_block) * chroma_grain_block_size);
   }
 
   if (params->num_cr_points || params->chroma_scaling_from_luma) {
@@ -468,7 +474,8 @@ static void generate_chroma_grain_blocks(
              ((1 << gauss_sec_shift) >> 1)) >>
             gauss_sec_shift;
   } else {
-    memset(cb_grain_block, 0, sizeof(*cb_grain_block) * chroma_grain_samples);
+    memset(cr_grain_block, 0,
+           sizeof(*cr_grain_block) * chroma_grain_block_size);
   }
 
   for (int i = top_pad; i < chroma_block_size_y - bottom_pad; i++)
@@ -522,7 +529,7 @@ static void generate_chroma_grain_blocks(
     }
 }
 
-static void init_scaling_function(int scaling_points[][2], int num_points,
+static void init_scaling_function(const int scaling_points[][2], int num_points,
                                   int scaling_lut[]) {
   if (num_points == 0) return;
 
@@ -559,7 +566,7 @@ static int scale_LUT(int *scaling_lut, int index, int bit_depth) {
                              (bit_depth - 8));
 }
 
-static void add_noise_to_block(aom_film_grain_t *params, uint8_t *luma,
+static void add_noise_to_block(const aom_film_grain_t *params, uint8_t *luma,
                                uint8_t *cb, uint8_t *cr, int luma_stride,
                                int chroma_stride, int *luma_grain,
                                int *cb_grain, int *cr_grain,
@@ -675,7 +682,7 @@ static void add_noise_to_block(aom_film_grain_t *params, uint8_t *luma,
 }
 
 static void add_noise_to_block_hbd(
-    aom_film_grain_t *params, uint16_t *luma, uint16_t *cb, uint16_t *cr,
+    const aom_film_grain_t *params, uint16_t *luma, uint16_t *cb, uint16_t *cr,
     int luma_stride, int chroma_stride, int *luma_grain, int *cb_grain,
     int *cr_grain, int luma_grain_stride, int chroma_grain_stride,
     int half_luma_height, int half_luma_width, int bit_depth,
@@ -903,7 +910,7 @@ static void hor_boundary_overlap(int *top_block, int top_stride,
   }
 }
 
-void av1_add_film_grain(aom_film_grain_t *params, aom_image_t *src,
+void av1_add_film_grain(const aom_film_grain_t *params, const aom_image_t *src,
                         aom_image_t *dst) {
   uint8_t *luma, *cb, *cr;
   int height, width, luma_stride, chroma_stride;
@@ -950,6 +957,11 @@ void av1_add_film_grain(aom_film_grain_t *params, aom_image_t *src,
       exit(1);
   }
 
+  assert(params->bit_depth == src->bit_depth);
+
+  dst->fmt = src->fmt;
+  dst->bit_depth = src->bit_depth;
+
   dst->r_w = src->r_w;
   dst->r_h = src->r_h;
   dst->d_w = src->d_w;
@@ -999,15 +1011,13 @@ void av1_add_film_grain(aom_film_grain_t *params, aom_image_t *src,
   luma_stride = dst->stride[AOM_PLANE_Y] >> use_high_bit_depth;
   chroma_stride = dst->stride[AOM_PLANE_U] >> use_high_bit_depth;
 
-  params->bit_depth = dst->bit_depth;
-
   av1_add_film_grain_run(params, luma, cb, cr, height, width, luma_stride,
                          chroma_stride, use_high_bit_depth, chroma_subsamp_y,
                          chroma_subsamp_x, mc_identity);
   return;
 }
 
-void av1_add_film_grain_run(aom_film_grain_t *params, uint8_t *luma,
+void av1_add_film_grain_run(const aom_film_grain_t *params, uint8_t *luma,
                             uint8_t *cb, uint8_t *cr, int height, int width,
                             int luma_stride, int chroma_stride,
                             int use_high_bit_depth, int chroma_subsamp_y,
diff --git a/third_party/aom/aom_dsp/grain_synthesis.h b/third_party/aom/aom_dsp/grain_synthesis.h
index 016cb12d7..65feb6068 100644
--- a/third_party/aom/aom_dsp/grain_synthesis.h
+++ b/third_party/aom/aom_dsp/grain_synthesis.h
@@ -72,7 +72,7 @@ typedef struct {
 
   int clip_to_restricted_range;
 
-  int bit_depth;  // video bit depth
+  unsigned int bit_depth;  // video bit depth
 
   int chroma_scaling_from_luma;
 
@@ -94,7 +94,7 @@ typedef struct {
  * \param[in]    luma_stride      luma plane stride
  * \param[in]    chroma_stride    chroma plane stride
  */
-void av1_add_film_grain_run(aom_film_grain_t *grain_params, uint8_t *luma,
+void av1_add_film_grain_run(const aom_film_grain_t *grain_params, uint8_t *luma,
                             uint8_t *cb, uint8_t *cr, int height, int width,
                             int luma_stride, int chroma_stride,
                             int use_high_bit_depth, int chroma_subsamp_y,
@@ -106,10 +106,10 @@ void av1_add_film_grain_run(aom_film_grain_t *grain_params, uint8_t *luma,
  *
  * \param[in]    grain_params     Grain parameters
  * \param[in]    src              Source image
- * \param[in]    dst              Resulting image with grain
+ * \param[out]   dst              Resulting image with grain
  */
-void av1_add_film_grain(aom_film_grain_t *grain_params, aom_image_t *src,
-                        aom_image_t *dst);
+void av1_add_film_grain(const aom_film_grain_t *grain_params,
+                        const aom_image_t *src, aom_image_t *dst);
 
 #ifdef __cplusplus
 }  // extern "C"
diff --git a/third_party/aom/aom_dsp/noise_model.c b/third_party/aom/aom_dsp/noise_model.c
index a1287f74f..5975c62e8 100644
--- a/third_party/aom/aom_dsp/noise_model.c
+++ b/third_party/aom/aom_dsp/noise_model.c
@@ -1458,3 +1458,189 @@ int aom_wiener_denoise_2d(const uint8_t *const data[3], uint8_t *denoised[3],
   }
   return init_success;
 }
+
+struct aom_denoise_and_model_t {
+  int block_size;
+  int bit_depth;
+  float noise_level;
+
+  // Size of current denoised buffer and flat_block buffer
+  int width;
+  int height;
+  int y_stride;
+  int uv_stride;
+  int num_blocks_w;
+  int num_blocks_h;
+
+  // Buffers for image and noise_psd allocated on the fly
+  float *noise_psd[3];
+  uint8_t *denoised[3];
+  uint8_t *flat_blocks;
+
+  aom_flat_block_finder_t flat_block_finder;
+  aom_noise_model_t noise_model;
+};
+
+struct aom_denoise_and_model_t *aom_denoise_and_model_alloc(int bit_depth,
+                                                            int block_size,
+                                                            float noise_level) {
+  struct aom_denoise_and_model_t *ctx =
+      (struct aom_denoise_and_model_t *)aom_malloc(
+          sizeof(struct aom_denoise_and_model_t));
+  if (!ctx) {
+    fprintf(stderr, "Unable to allocate denoise_and_model struct\n");
+    return NULL;
+  }
+  memset(ctx, 0, sizeof(*ctx));
+
+  ctx->block_size = block_size;
+  ctx->noise_level = noise_level;
+  ctx->bit_depth = bit_depth;
+
+  ctx->noise_psd[0] =
+      aom_malloc(sizeof(*ctx->noise_psd[0]) * block_size * block_size);
+  ctx->noise_psd[1] =
+      aom_malloc(sizeof(*ctx->noise_psd[1]) * block_size * block_size);
+  ctx->noise_psd[2] =
+      aom_malloc(sizeof(*ctx->noise_psd[2]) * block_size * block_size);
+  if (!ctx->noise_psd[0] || !ctx->noise_psd[1] || !ctx->noise_psd[2]) {
+    fprintf(stderr, "Unable to allocate noise PSD buffers\n");
+    aom_denoise_and_model_free(ctx);
+    return NULL;
+  }
+  return ctx;
+}
+
+void aom_denoise_and_model_free(struct aom_denoise_and_model_t *ctx) {
+  aom_free(ctx->flat_blocks);
+  for (int i = 0; i < 3; ++i) {
+    aom_free(ctx->denoised[i]);
+    aom_free(ctx->noise_psd[i]);
+  }
+  aom_noise_model_free(&ctx->noise_model);
+  aom_flat_block_finder_free(&ctx->flat_block_finder);
+  aom_free(ctx);
+}
+
+static int denoise_and_model_realloc_if_necessary(
+    struct aom_denoise_and_model_t *ctx, YV12_BUFFER_CONFIG *sd) {
+  if (ctx->width == sd->y_width && ctx->height == sd->y_height &&
+      ctx->y_stride == sd->y_stride && ctx->uv_stride == sd->uv_stride)
+    return 1;
+  const int use_highbd = (sd->flags & YV12_FLAG_HIGHBITDEPTH) != 0;
+  const int block_size = ctx->block_size;
+
+  ctx->width = sd->y_width;
+  ctx->height = sd->y_height;
+  ctx->y_stride = sd->y_stride;
+  ctx->uv_stride = sd->uv_stride;
+
+  for (int i = 0; i < 3; ++i) {
+    aom_free(ctx->denoised[i]);
+    ctx->denoised[i] = NULL;
+  }
+  aom_free(ctx->flat_blocks);
+  ctx->flat_blocks = NULL;
+
+  ctx->denoised[0] = aom_malloc((sd->y_stride * sd->y_height) << use_highbd);
+  ctx->denoised[1] = aom_malloc((sd->uv_stride * sd->uv_height) << use_highbd);
+  ctx->denoised[2] = aom_malloc((sd->uv_stride * sd->uv_height) << use_highbd);
+  if (!ctx->denoised[0] || !ctx->denoised[1] || !ctx->denoised[2]) {
+    fprintf(stderr, "Unable to allocate denoise buffers\n");
+    return 0;
+  }
+  ctx->num_blocks_w = (sd->y_width + ctx->block_size - 1) / ctx->block_size;
+  ctx->num_blocks_h = (sd->y_height + ctx->block_size - 1) / ctx->block_size;
+  ctx->flat_blocks = aom_malloc(ctx->num_blocks_w * ctx->num_blocks_h);
+
+  aom_flat_block_finder_free(&ctx->flat_block_finder);
+  if (!aom_flat_block_finder_init(&ctx->flat_block_finder, ctx->block_size,
+                                  ctx->bit_depth, use_highbd)) {
+    fprintf(stderr, "Unable to init flat block finder\n");
+    return 0;
+  }
+
+  const aom_noise_model_params_t params = { AOM_NOISE_SHAPE_SQUARE, 3,
+                                            ctx->bit_depth, use_highbd };
+  aom_noise_model_free(&ctx->noise_model);
+  if (!aom_noise_model_init(&ctx->noise_model, params)) {
+    fprintf(stderr, "Unable to init noise model\n");
+    return 0;
+  }
+
+  // Simply use a flat PSD (although we could use the flat blocks to estimate
+  // PSD) those to estimate an actual noise PSD)
+  const float y_noise_level =
+      aom_noise_psd_get_default_value(ctx->block_size, ctx->noise_level);
+  const float uv_noise_level = aom_noise_psd_get_default_value(
+      ctx->block_size >> sd->subsampling_x, ctx->noise_level);
+  for (int i = 0; i < block_size * block_size; ++i) {
+    ctx->noise_psd[0][i] = y_noise_level;
+    ctx->noise_psd[1][i] = ctx->noise_psd[2][i] = uv_noise_level;
+  }
+  return 1;
+}
+
+int aom_denoise_and_model_run(struct aom_denoise_and_model_t *ctx,
+                              YV12_BUFFER_CONFIG *sd,
+                              aom_film_grain_t *film_grain) {
+  const int block_size = ctx->block_size;
+  const int use_highbd = (sd->flags & YV12_FLAG_HIGHBITDEPTH) != 0;
+  uint8_t *raw_data[3] = {
+    use_highbd ? (uint8_t *)CONVERT_TO_SHORTPTR(sd->y_buffer) : sd->y_buffer,
+    use_highbd ? (uint8_t *)CONVERT_TO_SHORTPTR(sd->u_buffer) : sd->u_buffer,
+    use_highbd ? (uint8_t *)CONVERT_TO_SHORTPTR(sd->v_buffer) : sd->v_buffer,
+  };
+  const uint8_t *const data[3] = { raw_data[0], raw_data[1], raw_data[2] };
+  int strides[3] = { sd->y_stride, sd->uv_stride, sd->uv_stride };
+  int chroma_sub_log2[2] = { sd->subsampling_x, sd->subsampling_y };
+
+  if (!denoise_and_model_realloc_if_necessary(ctx, sd)) {
+    fprintf(stderr, "Unable to realloc buffers\n");
+    return 0;
+  }
+
+  aom_flat_block_finder_run(&ctx->flat_block_finder, data[0], sd->y_width,
+                            sd->y_height, strides[0], ctx->flat_blocks);
+
+  if (!aom_wiener_denoise_2d(data, ctx->denoised, sd->y_width, sd->y_height,
+                             strides, chroma_sub_log2, ctx->noise_psd,
+                             block_size, ctx->bit_depth, use_highbd)) {
+    fprintf(stderr, "Unable to denoise image\n");
+    return 0;
+  }
+
+  const aom_noise_status_t status = aom_noise_model_update(
+      &ctx->noise_model, data, (const uint8_t *const *)ctx->denoised,
+      sd->y_width, sd->y_height, strides, chroma_sub_log2, ctx->flat_blocks,
+      block_size);
+  int have_noise_estimate = 0;
+  if (status == AOM_NOISE_STATUS_OK) {
+    have_noise_estimate = 1;
+  } else if (status == AOM_NOISE_STATUS_DIFFERENT_NOISE_TYPE) {
+    aom_noise_model_save_latest(&ctx->noise_model);
+    have_noise_estimate = 1;
+  } else {
+    // Unable to update noise model; proceed if we have a previous estimate.
+    have_noise_estimate =
+        (ctx->noise_model.combined_state[0].strength_solver.num_equations > 0);
+  }
+
+  film_grain->apply_grain = 0;
+  if (have_noise_estimate) {
+    if (!aom_noise_model_get_grain_parameters(&ctx->noise_model, film_grain)) {
+      fprintf(stderr, "Unable to get grain parameters.\n");
+      return 0;
+    }
+    if (!film_grain->random_seed) {
+      film_grain->random_seed = 1071;
+    }
+    memcpy(raw_data[0], ctx->denoised[0],
+           (strides[0] * sd->y_height) << use_highbd);
+    memcpy(raw_data[1], ctx->denoised[1],
+           (strides[1] * sd->uv_height) << use_highbd);
+    memcpy(raw_data[2], ctx->denoised[2],
+           (strides[2] * sd->uv_height) << use_highbd);
+  }
+  return 1;
+}
diff --git a/third_party/aom/aom_dsp/noise_model.h b/third_party/aom/aom_dsp/noise_model.h
index dabeacc14..b07bf8617 100644
--- a/third_party/aom/aom_dsp/noise_model.h
+++ b/third_party/aom/aom_dsp/noise_model.h
@@ -18,6 +18,7 @@ extern "C" {
 
 #include <stdint.h>
 #include "aom_dsp/grain_synthesis.h"
+#include "aom_scale/yv12config.h"
 
 /*!\brief Wrapper of data required to represent linear system of eqns and soln.
  */
@@ -280,6 +281,42 @@ int aom_wiener_denoise_2d(const uint8_t *const data[3], uint8_t *denoised[3],
                           int w, int h, int stride[3], int chroma_sub_log2[2],
                           float *noise_psd[3], int block_size, int bit_depth,
                           int use_highbd);
+
+struct aom_denoise_and_model_t;
+
+/*!\brief Denoise the buffer and model the residual noise.
+ *
+ * This is meant to be called sequentially on input frames. The input buffer
+ * is denoised and the residual noise is modelled. The current noise estimate
+ * is populated in film_grain. Returns true on success. The grain.apply_grain
+ * parameter will be true when the input buffer was successfully denoised and
+ * grain was modelled. Returns false on error.
+ *
+ * \param[in]      ctx   Struct allocated with aom_denoise_and_model_alloc
+ *                       that holds some buffers for denoising and the current
+ *                       noise estimate.
+ * \param[in/out]   buf  The raw input buffer to be denoised.
+ * \param[out]    grain  Output film grain parameters
+ */
+int aom_denoise_and_model_run(struct aom_denoise_and_model_t *ctx,
+                              YV12_BUFFER_CONFIG *buf, aom_film_grain_t *grain);
+
+/*!\brief Allocates a context that can be used for denoising and noise modeling.
+ *
+ * \param[in]  bit_depth   Bit depth of buffers this will be run on.
+ * \param[in]  block_size  Block size for noise modeling and flat block
+ *                         estimation
+ * \param[in]  noise_level The noise_level (2.5 for moderate noise, and 5 for
+ *                         higher levels of noise)
+ */
+struct aom_denoise_and_model_t *aom_denoise_and_model_alloc(int bit_depth,
+                                                            int block_size,
+                                                            float noise_level);
+
+/*!\brief Frees the denoise context allocated with aom_denoise_and_model_alloc
+ */
+void aom_denoise_and_model_free(struct aom_denoise_and_model_t *denoise_model);
+
 #ifdef __cplusplus
 }  // extern "C"
 #endif  // __cplusplus
diff --git a/third_party/aom/aom_dsp/simd/v256_intrinsics.h b/third_party/aom/aom_dsp/simd/v256_intrinsics.h
index 0e5ae5b68..4b70cc57b 100644
--- a/third_party/aom/aom_dsp/simd/v256_intrinsics.h
+++ b/third_party/aom/aom_dsp/simd/v256_intrinsics.h
@@ -289,6 +289,15 @@ SIMD_INLINE v256 v256_shr_u32(v256 a, unsigned int c) {
 SIMD_INLINE v256 v256_shr_s32(v256 a, unsigned int c) {
   return c_v256_shr_s32(a, c);
 }
+SIMD_INLINE v256 v256_shl_64(v256 a, unsigned int c) {
+  return c_v256_shl_64(a, c);
+}
+SIMD_INLINE v256 v256_shr_u64(v256 a, unsigned int c) {
+  return c_v256_shr_u64(a, c);
+}
+SIMD_INLINE v256 v256_shr_s64(v256 a, unsigned int c) {
+  return c_v256_shr_s64(a, c);
+}
 
 SIMD_INLINE v256 v256_shr_n_byte(v256 a, unsigned int n) {
   return c_v256_shr_n_byte(a, n);
diff --git a/third_party/aom/aom_dsp/variance.c b/third_party/aom/aom_dsp/variance.c
index d367905bc..817ebe15d 100644
--- a/third_party/aom/aom_dsp/variance.c
+++ b/third_party/aom/aom_dsp/variance.c
@@ -386,7 +386,7 @@ void aom_upsampled_pred_c(MACROBLOCKD *xd, const AV1_COMMON *const cm,
     }
   }
 
-  const InterpFilterParams filter =
+  const InterpFilterParams *filter =
       av1_get_interp_filter_params_with_block_size(EIGHTTAP_REGULAR, 8);
 
   if (!subpel_x_q3 && !subpel_y_q3) {
@@ -413,12 +413,12 @@ void aom_upsampled_pred_c(MACROBLOCKD *xd, const AV1_COMMON *const cm,
     const int16_t *const kernel_y =
         av1_get_interp_filter_subpel_kernel(filter, subpel_y_q3 << 1);
     const int intermediate_height =
-        (((height - 1) * 8 + subpel_y_q3) >> 3) + filter.taps;
+        (((height - 1) * 8 + subpel_y_q3) >> 3) + filter->taps;
     assert(intermediate_height <= (MAX_SB_SIZE * 2 + 16) + 16);
-    aom_convolve8_horiz(ref - ref_stride * ((filter.taps >> 1) - 1), ref_stride,
-                        temp, MAX_SB_SIZE, kernel_x, 16, NULL, -1, width,
-                        intermediate_height);
-    aom_convolve8_vert(temp + MAX_SB_SIZE * ((filter.taps >> 1) - 1),
+    aom_convolve8_horiz(ref - ref_stride * ((filter->taps >> 1) - 1),
+                        ref_stride, temp, MAX_SB_SIZE, kernel_x, 16, NULL, -1,
+                        width, intermediate_height);
+    aom_convolve8_vert(temp + MAX_SB_SIZE * ((filter->taps >> 1) - 1),
                        MAX_SB_SIZE, comp_pred, width, NULL, -1, kernel_y, 16,
                        width, height);
   }
@@ -974,7 +974,7 @@ void aom_highbd_upsampled_pred_c(MACROBLOCKD *xd,
     }
   }
 
-  const InterpFilterParams filter =
+  const InterpFilterParams *filter =
       av1_get_interp_filter_params_with_block_size(EIGHTTAP_REGULAR, 8);
 
   if (!subpel_x_q3 && !subpel_y_q3) {
@@ -1004,14 +1004,14 @@ void aom_highbd_upsampled_pred_c(MACROBLOCKD *xd,
     const int16_t *const kernel_y =
         av1_get_interp_filter_subpel_kernel(filter, subpel_y_q3 << 1);
     const int intermediate_height =
-        (((height - 1) * 8 + subpel_y_q3) >> 3) + filter.taps;
+        (((height - 1) * 8 + subpel_y_q3) >> 3) + filter->taps;
     assert(intermediate_height <= (MAX_SB_SIZE * 2 + 16) + 16);
-    aom_highbd_convolve8_horiz(ref8 - ref_stride * ((filter.taps >> 1) - 1),
+    aom_highbd_convolve8_horiz(ref8 - ref_stride * ((filter->taps >> 1) - 1),
                                ref_stride, CONVERT_TO_BYTEPTR(temp),
                                MAX_SB_SIZE, kernel_x, 16, NULL, -1, width,
                                intermediate_height, bd);
     aom_highbd_convolve8_vert(
-        CONVERT_TO_BYTEPTR(temp + MAX_SB_SIZE * ((filter.taps >> 1) - 1)),
+        CONVERT_TO_BYTEPTR(temp + MAX_SB_SIZE * ((filter->taps >> 1) - 1)),
         MAX_SB_SIZE, CONVERT_TO_BYTEPTR(comp_pred), width, NULL, -1, kernel_y,
         16, width, height, bd);
   }
@@ -1185,29 +1185,18 @@ void aom_highbd_comp_mask_pred_c(uint16_t *comp_pred, const uint8_t *pred8,
   }
 }
 
-void aom_highbd_comp_mask_upsampled_pred_c(
+void aom_highbd_comp_mask_upsampled_pred(
     MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
     const MV *const mv, uint16_t *comp_pred, const uint8_t *pred8, int width,
     int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8,
     int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask,
     int bd) {
-  int i, j;
-
-  uint16_t *pred = CONVERT_TO_SHORTPTR(pred8);
   aom_highbd_upsampled_pred(xd, cm, mi_row, mi_col, mv, comp_pred, width,
                             height, subpel_x_q3, subpel_y_q3, ref8, ref_stride,
                             bd);
-  for (i = 0; i < height; ++i) {
-    for (j = 0; j < width; ++j) {
-      if (!invert_mask)
-        comp_pred[j] = AOM_BLEND_A64(mask[j], comp_pred[j], pred[j]);
-      else
-        comp_pred[j] = AOM_BLEND_A64(mask[j], pred[j], comp_pred[j]);
-    }
-    comp_pred += width;
-    pred += width;
-    mask += mask_stride;
-  }
+  aom_highbd_comp_mask_pred(comp_pred, pred8, width, height,
+                            CONVERT_TO_BYTEPTR(comp_pred), width, mask,
+                            mask_stride, invert_mask);
 }
 
 #define HIGHBD_MASK_SUBPIX_VAR(W, H)                                           \
diff --git a/third_party/aom/aom_dsp/variance.h b/third_party/aom/aom_dsp/variance.h
index 544dda944..b954470de 100644
--- a/third_party/aom/aom_dsp/variance.h
+++ b/third_party/aom/aom_dsp/variance.h
@@ -76,6 +76,13 @@ void aom_comp_mask_upsampled_pred(
     int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref,
     int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask);
 
+void aom_highbd_comp_mask_upsampled_pred(
+    MACROBLOCKD *xd, const struct AV1Common *const cm, int mi_row, int mi_col,
+    const MV *const mv, uint16_t *comp_pred, const uint8_t *pred8, int width,
+    int height, int subpel_x_q3, int subpel_y_q3, const uint8_t *ref8,
+    int ref_stride, const uint8_t *mask, int mask_stride, int invert_mask,
+    int bd);
+
 typedef unsigned int (*aom_obmc_sad_fn_t)(const uint8_t *pred, int pred_stride,
                                           const int32_t *wsrc,
                                           const int32_t *msk);
diff --git a/third_party/aom/aom_dsp/x86/aom_subpixel_8t_intrin_avx2.c b/third_party/aom/aom_dsp/x86/aom_subpixel_8t_intrin_avx2.c
index af45a03ac..f3fe50372 100644
--- a/third_party/aom/aom_dsp/x86/aom_subpixel_8t_intrin_avx2.c
+++ b/third_party/aom/aom_dsp/x86/aom_subpixel_8t_intrin_avx2.c
@@ -41,20 +41,290 @@
 #define MM256_BROADCASTSI128_SI256(x) _mm256_broadcastsi128_si256(x)
 #endif  // __clang__
 
+static INLINE void xx_storeu2_epi32(const uint8_t *output_ptr,
+                                    const ptrdiff_t stride, const __m256i *a) {
+  *((uint32_t *)(output_ptr)) = _mm_cvtsi128_si32(_mm256_castsi256_si128(*a));
+  *((uint32_t *)(output_ptr + stride)) =
+      _mm_cvtsi128_si32(_mm256_extracti128_si256(*a, 1));
+}
+
+static INLINE __m256i xx_loadu2_epi64(const void *hi, const void *lo) {
+  __m256i a = _mm256_castsi128_si256(_mm_loadl_epi64((const __m128i *)(lo)));
+  a = _mm256_inserti128_si256(a, _mm_loadl_epi64((const __m128i *)(hi)), 1);
+  return a;
+}
+
+static INLINE void xx_storeu2_epi64(const uint8_t *output_ptr,
+                                    const ptrdiff_t stride, const __m256i *a) {
+  _mm_storel_epi64((__m128i *)output_ptr, _mm256_castsi256_si128(*a));
+  _mm_storel_epi64((__m128i *)(output_ptr + stride),
+                   _mm256_extractf128_si256(*a, 1));
+}
+
+static INLINE __m256i xx_loadu2_mi128(const void *hi, const void *lo) {
+  __m256i a = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(lo)));
+  a = _mm256_inserti128_si256(a, _mm_loadu_si128((const __m128i *)(hi)), 1);
+  return a;
+}
+
+static INLINE void xx_store2_mi128(const uint8_t *output_ptr,
+                                   const ptrdiff_t stride, const __m256i *a) {
+  _mm_store_si128((__m128i *)output_ptr, _mm256_castsi256_si128(*a));
+  _mm_store_si128((__m128i *)(output_ptr + stride),
+                  _mm256_extractf128_si256(*a, 1));
+}
+
+static void aom_filter_block1d4_h8_avx2(
+    const uint8_t *src_ptr, ptrdiff_t src_pixels_per_line, uint8_t *output_ptr,
+    ptrdiff_t output_pitch, uint32_t output_height, const int16_t *filter) {
+  __m128i filtersReg;
+  __m256i addFilterReg32, filt1Reg, filt2Reg;
+  __m256i firstFilters, secondFilters;
+  __m256i srcRegFilt32b1_1, srcRegFilt32b2;
+  __m256i srcReg32b1;
+  unsigned int i;
+  ptrdiff_t src_stride, dst_stride;
+  src_ptr -= 3;
+  addFilterReg32 = _mm256_set1_epi16(32);
+  filtersReg = _mm_loadu_si128((const __m128i *)filter);
+  filtersReg = _mm_srai_epi16(filtersReg, 1);
+  // converting the 16 bit (short) to 8 bit (byte) and have the same data
+  // in both lanes of 128 bit register.
+  filtersReg = _mm_packs_epi16(filtersReg, filtersReg);
+  // have the same data in both lanes of a 256 bit register
+  const __m256i filtersReg32 = MM256_BROADCASTSI128_SI256(filtersReg);
+
+  // duplicate only the first 32 bits
+  firstFilters = _mm256_shuffle_epi32(filtersReg32, 0);
+  // duplicate only the second 32 bits
+  secondFilters = _mm256_shuffle_epi32(filtersReg32, 0x55);
+
+  filt1Reg = _mm256_load_si256((__m256i const *)filt_d4_global_avx2);
+  filt2Reg = _mm256_load_si256((__m256i const *)(filt_d4_global_avx2 + 32));
+
+  // multiple the size of the source and destination stride by two
+  src_stride = src_pixels_per_line << 1;
+  dst_stride = output_pitch << 1;
+  for (i = output_height; i > 1; i -= 2) {
+    // load the 2 strides of source
+    srcReg32b1 = xx_loadu2_mi128(src_ptr + src_pixels_per_line, src_ptr);
+
+    // filter the source buffer
+    srcRegFilt32b1_1 = _mm256_shuffle_epi8(srcReg32b1, filt1Reg);
+
+    // multiply 4 adjacent elements with the filter and add the result
+    srcRegFilt32b1_1 = _mm256_maddubs_epi16(srcRegFilt32b1_1, firstFilters);
+
+    // filter the source buffer
+    srcRegFilt32b2 = _mm256_shuffle_epi8(srcReg32b1, filt2Reg);
+
+    // multiply 4 adjacent elements with the filter and add the result
+    srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, secondFilters);
+
+    srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1, srcRegFilt32b2);
+
+    srcRegFilt32b1_1 =
+        _mm256_hadds_epi16(srcRegFilt32b1_1, _mm256_setzero_si256());
+
+    // shift by 6 bit each 16 bit
+    srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1, addFilterReg32);
+    srcRegFilt32b1_1 = _mm256_srai_epi16(srcRegFilt32b1_1, 6);
+
+    // shrink to 8 bit each 16 bits, the first lane contain the first
+    // convolve result and the second lane contain the second convolve result
+    srcRegFilt32b1_1 =
+        _mm256_packus_epi16(srcRegFilt32b1_1, _mm256_setzero_si256());
+
+    src_ptr += src_stride;
+
+    xx_storeu2_epi32(output_ptr, output_pitch, &srcRegFilt32b1_1);
+    output_ptr += dst_stride;
+  }
+
+  // if the number of strides is odd.
+  // process only 4 bytes
+  if (i > 0) {
+    __m128i srcReg1, srcRegFilt1_1;
+    __m128i srcRegFilt2;
+
+    srcReg1 = _mm_loadu_si128((const __m128i *)(src_ptr));
+
+    // filter the source buffer
+    srcRegFilt1_1 = _mm_shuffle_epi8(srcReg1, _mm256_castsi256_si128(filt1Reg));
+
+    // multiply 4 adjacent elements with the filter and add the result
+    srcRegFilt1_1 =
+        _mm_maddubs_epi16(srcRegFilt1_1, _mm256_castsi256_si128(firstFilters));
+
+    // filter the source buffer
+    srcRegFilt2 = _mm_shuffle_epi8(srcReg1, _mm256_castsi256_si128(filt2Reg));
+
+    // multiply 4 adjacent elements with the filter and add the result
+    srcRegFilt2 =
+        _mm_maddubs_epi16(srcRegFilt2, _mm256_castsi256_si128(secondFilters));
+
+    srcRegFilt1_1 = _mm_adds_epi16(srcRegFilt1_1, srcRegFilt2);
+    srcRegFilt1_1 = _mm_hadds_epi16(srcRegFilt1_1, _mm_setzero_si128());
+    // shift by 6 bit each 16 bit
+    srcRegFilt1_1 =
+        _mm_adds_epi16(srcRegFilt1_1, _mm256_castsi256_si128(addFilterReg32));
+    srcRegFilt1_1 = _mm_srai_epi16(srcRegFilt1_1, 6);
+
+    // shrink to 8 bit each 16 bits, the first lane contain the first
+    // convolve result and the second lane contain the second convolve result
+    srcRegFilt1_1 = _mm_packus_epi16(srcRegFilt1_1, _mm_setzero_si128());
+
+    // save 4 bytes
+    *((uint32_t *)(output_ptr)) = _mm_cvtsi128_si32(srcRegFilt1_1);
+  }
+}
+
+static void aom_filter_block1d8_h8_avx2(
+    const uint8_t *src_ptr, ptrdiff_t src_pixels_per_line, uint8_t *output_ptr,
+    ptrdiff_t output_pitch, uint32_t output_height, const int16_t *filter) {
+  __m128i filtersReg;
+  __m256i addFilterReg32, filt1Reg, filt2Reg, filt3Reg, filt4Reg;
+  __m256i firstFilters, secondFilters, thirdFilters, forthFilters;
+  __m256i srcRegFilt32b1_1, srcRegFilt32b2, srcRegFilt32b3;
+  __m256i srcReg32b1;
+  unsigned int i;
+  ptrdiff_t src_stride, dst_stride;
+  src_ptr -= 3;
+  addFilterReg32 = _mm256_set1_epi16(32);
+  filtersReg = _mm_loadu_si128((const __m128i *)filter);
+  filtersReg = _mm_srai_epi16(filtersReg, 1);
+  // converting the 16 bit (short) to 8 bit (byte) and have the same data
+  // in both lanes of 128 bit register.
+  filtersReg = _mm_packs_epi16(filtersReg, filtersReg);
+  // have the same data in both lanes of a 256 bit register
+  const __m256i filtersReg32 = MM256_BROADCASTSI128_SI256(filtersReg);
+
+  // duplicate only the first 16 bits (first and second byte)
+  // across 256 bit register
+  firstFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x100u));
+  // duplicate only the second 16 bits (third and forth byte)
+  // across 256 bit register
+  secondFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x302u));
+  // duplicate only the third 16 bits (fifth and sixth byte)
+  // across 256 bit register
+  thirdFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x504u));
+  // duplicate only the forth 16 bits (seventh and eighth byte)
+  // across 256 bit register
+  forthFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x706u));
+
+  filt1Reg = _mm256_load_si256((__m256i const *)filt_global_avx2);
+  filt2Reg = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32));
+  filt3Reg = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32 * 2));
+  filt4Reg = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32 * 3));
+
+  // multiple the size of the source and destination stride by two
+  src_stride = src_pixels_per_line << 1;
+  dst_stride = output_pitch << 1;
+  for (i = output_height; i > 1; i -= 2) {
+    // load the 2 strides of source
+    srcReg32b1 = xx_loadu2_mi128(src_ptr + src_pixels_per_line, src_ptr);
+
+    // filter the source buffer
+    srcRegFilt32b1_1 = _mm256_shuffle_epi8(srcReg32b1, filt1Reg);
+    srcRegFilt32b2 = _mm256_shuffle_epi8(srcReg32b1, filt4Reg);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt32b1_1 = _mm256_maddubs_epi16(srcRegFilt32b1_1, firstFilters);
+    srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, forthFilters);
+
+    // add and saturate the results together
+    srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1, srcRegFilt32b2);
+
+    // filter the source buffer
+    srcRegFilt32b3 = _mm256_shuffle_epi8(srcReg32b1, filt2Reg);
+    srcRegFilt32b2 = _mm256_shuffle_epi8(srcReg32b1, filt3Reg);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt32b3 = _mm256_maddubs_epi16(srcRegFilt32b3, secondFilters);
+    srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, thirdFilters);
+
+    __m256i sum23 = _mm256_adds_epi16(srcRegFilt32b3, srcRegFilt32b2);
+    srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1, sum23);
+
+    // shift by 6 bit each 16 bit
+    srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1, addFilterReg32);
+    srcRegFilt32b1_1 = _mm256_srai_epi16(srcRegFilt32b1_1, 6);
+
+    // shrink to 8 bit each 16 bits, the first lane contain the first
+    // convolve result and the second lane contain the second convolve result
+    srcRegFilt32b1_1 =
+        _mm256_packus_epi16(srcRegFilt32b1_1, _mm256_setzero_si256());
+
+    src_ptr += src_stride;
+
+    xx_storeu2_epi64(output_ptr, output_pitch, &srcRegFilt32b1_1);
+    output_ptr += dst_stride;
+  }
+
+  // if the number of strides is odd.
+  // process only 8 bytes
+  if (i > 0) {
+    __m128i srcReg1, srcRegFilt1_1;
+    __m128i srcRegFilt2, srcRegFilt3;
+
+    srcReg1 = _mm_loadu_si128((const __m128i *)(src_ptr));
+
+    // filter the source buffer
+    srcRegFilt1_1 = _mm_shuffle_epi8(srcReg1, _mm256_castsi256_si128(filt1Reg));
+    srcRegFilt2 = _mm_shuffle_epi8(srcReg1, _mm256_castsi256_si128(filt4Reg));
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt1_1 =
+        _mm_maddubs_epi16(srcRegFilt1_1, _mm256_castsi256_si128(firstFilters));
+    srcRegFilt2 =
+        _mm_maddubs_epi16(srcRegFilt2, _mm256_castsi256_si128(forthFilters));
+
+    // add and saturate the results together
+    srcRegFilt1_1 = _mm_adds_epi16(srcRegFilt1_1, srcRegFilt2);
+
+    // filter the source buffer
+    srcRegFilt3 = _mm_shuffle_epi8(srcReg1, _mm256_castsi256_si128(filt2Reg));
+    srcRegFilt2 = _mm_shuffle_epi8(srcReg1, _mm256_castsi256_si128(filt3Reg));
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt3 =
+        _mm_maddubs_epi16(srcRegFilt3, _mm256_castsi256_si128(secondFilters));
+    srcRegFilt2 =
+        _mm_maddubs_epi16(srcRegFilt2, _mm256_castsi256_si128(thirdFilters));
+
+    // add and saturate the results together
+    srcRegFilt1_1 =
+        _mm_adds_epi16(srcRegFilt1_1, _mm_adds_epi16(srcRegFilt3, srcRegFilt2));
+
+    // shift by 6 bit each 16 bit
+    srcRegFilt1_1 =
+        _mm_adds_epi16(srcRegFilt1_1, _mm256_castsi256_si128(addFilterReg32));
+    srcRegFilt1_1 = _mm_srai_epi16(srcRegFilt1_1, 6);
+
+    // shrink to 8 bit each 16 bits, the first lane contain the first
+    // convolve result and the second lane contain the second convolve
+    // result
+    srcRegFilt1_1 = _mm_packus_epi16(srcRegFilt1_1, _mm_setzero_si128());
+
+    // save 8 bytes
+    _mm_storel_epi64((__m128i *)output_ptr, srcRegFilt1_1);
+  }
+}
+
 static void aom_filter_block1d16_h8_avx2(
     const uint8_t *src_ptr, ptrdiff_t src_pixels_per_line, uint8_t *output_ptr,
     ptrdiff_t output_pitch, uint32_t output_height, const int16_t *filter) {
   __m128i filtersReg;
-  __m256i addFilterReg64, filt1Reg, filt2Reg, filt3Reg, filt4Reg;
+  __m256i addFilterReg32, filt1Reg, filt2Reg, filt3Reg, filt4Reg;
   __m256i firstFilters, secondFilters, thirdFilters, forthFilters;
   __m256i srcRegFilt32b1_1, srcRegFilt32b2_1, srcRegFilt32b2, srcRegFilt32b3;
   __m256i srcReg32b1, srcReg32b2, filtersReg32;
   unsigned int i;
   ptrdiff_t src_stride, dst_stride;
-
-  // create a register with 0,64,0,64,0,64,0,64,0,64,0,64,0,64,0,64
-  addFilterReg64 = _mm256_set1_epi32((int)0x0400040u);
+  src_ptr -= 3;
+  addFilterReg32 = _mm256_set1_epi16(32);
   filtersReg = _mm_loadu_si128((const __m128i *)filter);
+  filtersReg = _mm_srai_epi16(filtersReg, 1);
   // converting the 16 bit (short) to 8 bit (byte) and have the same data
   // in both lanes of 128 bit register.
   filtersReg = _mm_packs_epi16(filtersReg, filtersReg);
@@ -74,22 +344,17 @@ static void aom_filter_block1d16_h8_avx2(
   // across 256 bit register
   forthFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x706u));
 
-  filt1Reg = _mm256_load_si256((__m256i const *)filt1_global_avx2);
-  filt2Reg = _mm256_load_si256((__m256i const *)filt2_global_avx2);
-  filt3Reg = _mm256_load_si256((__m256i const *)filt3_global_avx2);
-  filt4Reg = _mm256_load_si256((__m256i const *)filt4_global_avx2);
+  filt1Reg = _mm256_load_si256((__m256i const *)filt_global_avx2);
+  filt2Reg = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32));
+  filt3Reg = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32 * 2));
+  filt4Reg = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32 * 3));
 
   // multiple the size of the source and destination stride by two
   src_stride = src_pixels_per_line << 1;
   dst_stride = output_pitch << 1;
   for (i = output_height; i > 1; i -= 2) {
     // load the 2 strides of source
-    srcReg32b1 =
-        _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(src_ptr - 3)));
-    srcReg32b1 = _mm256_inserti128_si256(
-        srcReg32b1,
-        _mm_loadu_si128((const __m128i *)(src_ptr + src_pixels_per_line - 3)),
-        1);
+    srcReg32b1 = xx_loadu2_mi128(src_ptr + src_pixels_per_line, src_ptr);
 
     // filter the source buffer
     srcRegFilt32b1_1 = _mm256_shuffle_epi8(srcReg32b1, filt1Reg);
@@ -110,22 +375,13 @@ static void aom_filter_block1d16_h8_avx2(
     srcRegFilt32b3 = _mm256_maddubs_epi16(srcRegFilt32b3, secondFilters);
     srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, thirdFilters);
 
-    // add and saturate the results together
-    srcRegFilt32b1_1 = _mm256_adds_epi16(
-        srcRegFilt32b1_1, _mm256_min_epi16(srcRegFilt32b3, srcRegFilt32b2));
+    __m256i sum23 = _mm256_adds_epi16(srcRegFilt32b3, srcRegFilt32b2);
+    srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1, sum23);
 
     // reading 2 strides of the next 16 bytes
     // (part of it was being read by earlier read)
     srcReg32b2 =
-        _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(src_ptr + 5)));
-    srcReg32b2 = _mm256_inserti128_si256(
-        srcReg32b2,
-        _mm_loadu_si128((const __m128i *)(src_ptr + src_pixels_per_line + 5)),
-        1);
-
-    // add and saturate the results together
-    srcRegFilt32b1_1 = _mm256_adds_epi16(
-        srcRegFilt32b1_1, _mm256_max_epi16(srcRegFilt32b3, srcRegFilt32b2));
+        xx_loadu2_mi128(src_ptr + src_pixels_per_line + 8, src_ptr + 8);
 
     // filter the source buffer
     srcRegFilt32b2_1 = _mm256_shuffle_epi8(srcReg32b2, filt1Reg);
@@ -148,32 +404,21 @@ static void aom_filter_block1d16_h8_avx2(
 
     // add and saturate the results together
     srcRegFilt32b2_1 = _mm256_adds_epi16(
-        srcRegFilt32b2_1, _mm256_min_epi16(srcRegFilt32b3, srcRegFilt32b2));
-    srcRegFilt32b2_1 = _mm256_adds_epi16(
-        srcRegFilt32b2_1, _mm256_max_epi16(srcRegFilt32b3, srcRegFilt32b2));
-
-    srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1, addFilterReg64);
+        srcRegFilt32b2_1, _mm256_adds_epi16(srcRegFilt32b3, srcRegFilt32b2));
 
-    srcRegFilt32b2_1 = _mm256_adds_epi16(srcRegFilt32b2_1, addFilterReg64);
-
-    // shift by 7 bit each 16 bit
-    srcRegFilt32b1_1 = _mm256_srai_epi16(srcRegFilt32b1_1, 7);
-    srcRegFilt32b2_1 = _mm256_srai_epi16(srcRegFilt32b2_1, 7);
+    // shift by 6 bit each 16 bit
+    srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b1_1, addFilterReg32);
+    srcRegFilt32b2_1 = _mm256_adds_epi16(srcRegFilt32b2_1, addFilterReg32);
+    srcRegFilt32b1_1 = _mm256_srai_epi16(srcRegFilt32b1_1, 6);
+    srcRegFilt32b2_1 = _mm256_srai_epi16(srcRegFilt32b2_1, 6);
 
     // shrink to 8 bit each 16 bits, the first lane contain the first
-    // convolve result and the second lane contain the second convolve
-    // result
+    // convolve result and the second lane contain the second convolve result
     srcRegFilt32b1_1 = _mm256_packus_epi16(srcRegFilt32b1_1, srcRegFilt32b2_1);
 
     src_ptr += src_stride;
 
-    // save 16 bytes
-    _mm_store_si128((__m128i *)output_ptr,
-                    _mm256_castsi256_si128(srcRegFilt32b1_1));
-
-    // save the next 16 bits
-    _mm_store_si128((__m128i *)(output_ptr + output_pitch),
-                    _mm256_extractf128_si256(srcRegFilt32b1_1, 1));
+    xx_store2_mi128(output_ptr, output_pitch, &srcRegFilt32b1_1);
     output_ptr += dst_stride;
   }
 
@@ -183,7 +428,7 @@ static void aom_filter_block1d16_h8_avx2(
     __m128i srcReg1, srcReg2, srcRegFilt1_1, srcRegFilt2_1;
     __m128i srcRegFilt2, srcRegFilt3;
 
-    srcReg1 = _mm_loadu_si128((const __m128i *)(src_ptr - 3));
+    srcReg1 = _mm_loadu_si128((const __m128i *)(src_ptr));
 
     // filter the source buffer
     srcRegFilt1_1 = _mm_shuffle_epi8(srcReg1, _mm256_castsi256_si128(filt1Reg));
@@ -210,15 +455,11 @@ static void aom_filter_block1d16_h8_avx2(
 
     // add and saturate the results together
     srcRegFilt1_1 =
-        _mm_adds_epi16(srcRegFilt1_1, _mm_min_epi16(srcRegFilt3, srcRegFilt2));
+        _mm_adds_epi16(srcRegFilt1_1, _mm_adds_epi16(srcRegFilt3, srcRegFilt2));
 
     // reading the next 16 bytes
     // (part of it was being read by earlier read)
-    srcReg2 = _mm_loadu_si128((const __m128i *)(src_ptr + 5));
-
-    // add and saturate the results together
-    srcRegFilt1_1 =
-        _mm_adds_epi16(srcRegFilt1_1, _mm_max_epi16(srcRegFilt3, srcRegFilt2));
+    srcReg2 = _mm_loadu_si128((const __m128i *)(src_ptr + 8));
 
     // filter the source buffer
     srcRegFilt2_1 = _mm_shuffle_epi8(srcReg2, _mm256_castsi256_si128(filt1Reg));
@@ -245,19 +486,16 @@ static void aom_filter_block1d16_h8_avx2(
 
     // add and saturate the results together
     srcRegFilt2_1 =
-        _mm_adds_epi16(srcRegFilt2_1, _mm_min_epi16(srcRegFilt3, srcRegFilt2));
-    srcRegFilt2_1 =
-        _mm_adds_epi16(srcRegFilt2_1, _mm_max_epi16(srcRegFilt3, srcRegFilt2));
+        _mm_adds_epi16(srcRegFilt2_1, _mm_adds_epi16(srcRegFilt3, srcRegFilt2));
 
+    // shift by 6 bit each 16 bit
     srcRegFilt1_1 =
-        _mm_adds_epi16(srcRegFilt1_1, _mm256_castsi256_si128(addFilterReg64));
+        _mm_adds_epi16(srcRegFilt1_1, _mm256_castsi256_si128(addFilterReg32));
+    srcRegFilt1_1 = _mm_srai_epi16(srcRegFilt1_1, 6);
 
     srcRegFilt2_1 =
-        _mm_adds_epi16(srcRegFilt2_1, _mm256_castsi256_si128(addFilterReg64));
-
-    // shift by 7 bit each 16 bit
-    srcRegFilt1_1 = _mm_srai_epi16(srcRegFilt1_1, 7);
-    srcRegFilt2_1 = _mm_srai_epi16(srcRegFilt2_1, 7);
+        _mm_adds_epi16(srcRegFilt2_1, _mm256_castsi256_si128(addFilterReg32));
+    srcRegFilt2_1 = _mm_srai_epi16(srcRegFilt2_1, 6);
 
     // shrink to 8 bit each 16 bits, the first lane contain the first
     // convolve result and the second lane contain the second convolve
@@ -269,11 +507,163 @@ static void aom_filter_block1d16_h8_avx2(
   }
 }
 
+static void aom_filter_block1d8_v8_avx2(
+    const uint8_t *src_ptr, ptrdiff_t src_pitch, uint8_t *output_ptr,
+    ptrdiff_t out_pitch, uint32_t output_height, const int16_t *filter) {
+  __m128i filtersReg;
+  __m256i addFilterReg32;
+  __m256i srcReg32b1, srcReg32b2, srcReg32b3, srcReg32b4, srcReg32b5;
+  __m256i srcReg32b6, srcReg32b7, srcReg32b8, srcReg32b9, srcReg32b10;
+  __m256i srcReg32b11, srcReg32b12, filtersReg32;
+  __m256i firstFilters, secondFilters, thirdFilters, forthFilters;
+  unsigned int i;
+  ptrdiff_t src_stride, dst_stride;
+
+  addFilterReg32 = _mm256_set1_epi16(32);
+  filtersReg = _mm_loadu_si128((const __m128i *)filter);
+  // converting the 16 bit (short) to  8 bit (byte) and have the
+  // same data in both lanes of 128 bit register.
+  filtersReg = _mm_srai_epi16(filtersReg, 1);
+  filtersReg = _mm_packs_epi16(filtersReg, filtersReg);
+  // have the same data in both lanes of a 256 bit register
+  filtersReg32 = MM256_BROADCASTSI128_SI256(filtersReg);
+
+  // duplicate only the first 16 bits (first and second byte)
+  // across 256 bit register
+  firstFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x100u));
+  // duplicate only the second 16 bits (third and forth byte)
+  // across 256 bit register
+  secondFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x302u));
+  // duplicate only the third 16 bits (fifth and sixth byte)
+  // across 256 bit register
+  thirdFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x504u));
+  // duplicate only the forth 16 bits (seventh and eighth byte)
+  // across 256 bit register
+  forthFilters = _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi16(0x706u));
+
+  // multiple the size of the source and destination stride by two
+  src_stride = src_pitch << 1;
+  dst_stride = out_pitch << 1;
+
+  // load 16 bytes 7 times in stride of src_pitch
+  srcReg32b1 = xx_loadu2_epi64(src_ptr + src_pitch, src_ptr);
+  srcReg32b3 =
+      xx_loadu2_epi64(src_ptr + src_pitch * 3, src_ptr + src_pitch * 2);
+  srcReg32b5 =
+      xx_loadu2_epi64(src_ptr + src_pitch * 5, src_ptr + src_pitch * 4);
+  srcReg32b7 = _mm256_castsi128_si256(
+      _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 6)));
+
+  // have each consecutive loads on the same 256 register
+  srcReg32b2 = _mm256_permute2x128_si256(srcReg32b1, srcReg32b3, 0x21);
+  srcReg32b4 = _mm256_permute2x128_si256(srcReg32b3, srcReg32b5, 0x21);
+  srcReg32b6 = _mm256_permute2x128_si256(srcReg32b5, srcReg32b7, 0x21);
+  // merge every two consecutive registers except the last one
+  srcReg32b10 = _mm256_unpacklo_epi8(srcReg32b1, srcReg32b2);
+  srcReg32b11 = _mm256_unpacklo_epi8(srcReg32b3, srcReg32b4);
+  srcReg32b2 = _mm256_unpacklo_epi8(srcReg32b5, srcReg32b6);
+
+  for (i = output_height; i > 1; i -= 2) {
+    // load the last 2 loads of 16 bytes and have every two
+    // consecutive loads in the same 256 bit register
+    srcReg32b8 = _mm256_castsi128_si256(
+        _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 7)));
+    srcReg32b7 = _mm256_inserti128_si256(srcReg32b7,
+                                         _mm256_castsi256_si128(srcReg32b8), 1);
+    srcReg32b9 = _mm256_castsi128_si256(
+        _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 8)));
+    srcReg32b8 = _mm256_inserti128_si256(srcReg32b8,
+                                         _mm256_castsi256_si128(srcReg32b9), 1);
+
+    // merge every two consecutive registers
+    // save
+    srcReg32b4 = _mm256_unpacklo_epi8(srcReg32b7, srcReg32b8);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcReg32b10 = _mm256_maddubs_epi16(srcReg32b10, firstFilters);
+    srcReg32b6 = _mm256_maddubs_epi16(srcReg32b4, forthFilters);
+
+    // add and saturate the results together
+    srcReg32b10 = _mm256_adds_epi16(srcReg32b10, srcReg32b6);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcReg32b8 = _mm256_maddubs_epi16(srcReg32b11, secondFilters);
+    srcReg32b12 = _mm256_maddubs_epi16(srcReg32b2, thirdFilters);
+
+    // add and saturate the results together
+    srcReg32b10 = _mm256_adds_epi16(srcReg32b10,
+                                    _mm256_adds_epi16(srcReg32b8, srcReg32b12));
+
+    // shift by 6 bit each 16 bit
+    srcReg32b10 = _mm256_adds_epi16(srcReg32b10, addFilterReg32);
+    srcReg32b10 = _mm256_srai_epi16(srcReg32b10, 6);
+
+    // shrink to 8 bit each 16 bits, the first lane contain the first
+    // convolve result and the second lane contain the second convolve
+    // result
+    srcReg32b1 = _mm256_packus_epi16(srcReg32b10, _mm256_setzero_si256());
+
+    src_ptr += src_stride;
+
+    xx_storeu2_epi64(output_ptr, out_pitch, &srcReg32b1);
+
+    output_ptr += dst_stride;
+
+    // save part of the registers for next strides
+    srcReg32b10 = srcReg32b11;
+    srcReg32b11 = srcReg32b2;
+    srcReg32b2 = srcReg32b4;
+    srcReg32b7 = srcReg32b9;
+  }
+  if (i > 0) {
+    __m128i srcRegFilt1, srcRegFilt4, srcRegFilt6, srcRegFilt8;
+    // load the last 16 bytes
+    srcRegFilt8 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 7));
+
+    // merge the last 2 results together
+    srcRegFilt4 =
+        _mm_unpacklo_epi8(_mm256_castsi256_si128(srcReg32b7), srcRegFilt8);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt1 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b10),
+                                    _mm256_castsi256_si128(firstFilters));
+    srcRegFilt4 =
+        _mm_maddubs_epi16(srcRegFilt4, _mm256_castsi256_si128(forthFilters));
+
+    // add and saturate the results together
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt4);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt4 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b11),
+                                    _mm256_castsi256_si128(secondFilters));
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt6 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b2),
+                                    _mm256_castsi256_si128(thirdFilters));
+
+    // add and saturate the results together
+    srcRegFilt1 =
+        _mm_adds_epi16(srcRegFilt1, _mm_adds_epi16(srcRegFilt4, srcRegFilt6));
+
+    // shift by 6 bit each 16 bit
+    srcRegFilt1 =
+        _mm_adds_epi16(srcRegFilt1, _mm256_castsi256_si128(addFilterReg32));
+    srcRegFilt1 = _mm_srai_epi16(srcRegFilt1, 6);
+
+    // shrink to 8 bit each 16 bits, the first lane contain the first
+    // convolve result and the second lane contain the second convolve result
+    srcRegFilt1 = _mm_packus_epi16(srcRegFilt1, _mm_setzero_si128());
+
+    // save 8 bytes
+    _mm_storel_epi64((__m128i *)output_ptr, srcRegFilt1);
+  }
+}
+
 static void aom_filter_block1d16_v8_avx2(
     const uint8_t *src_ptr, ptrdiff_t src_pitch, uint8_t *output_ptr,
     ptrdiff_t out_pitch, uint32_t output_height, const int16_t *filter) {
   __m128i filtersReg;
-  __m256i addFilterReg64;
+  __m256i addFilterReg32;
   __m256i srcReg32b1, srcReg32b2, srcReg32b3, srcReg32b4, srcReg32b5;
   __m256i srcReg32b6, srcReg32b7, srcReg32b8, srcReg32b9, srcReg32b10;
   __m256i srcReg32b11, srcReg32b12, filtersReg32;
@@ -281,11 +671,11 @@ static void aom_filter_block1d16_v8_avx2(
   unsigned int i;
   ptrdiff_t src_stride, dst_stride;
 
-  // create a register with 0,64,0,64,0,64,0,64,0,64,0,64,0,64,0,64
-  addFilterReg64 = _mm256_set1_epi32((int)0x0400040u);
+  addFilterReg32 = _mm256_set1_epi16(32);
   filtersReg = _mm_loadu_si128((const __m128i *)filter);
   // converting the 16 bit (short) to  8 bit (byte) and have the
   // same data in both lanes of 128 bit register.
+  filtersReg = _mm_srai_epi16(filtersReg, 1);
   filtersReg = _mm_packs_epi16(filtersReg, filtersReg);
   // have the same data in both lanes of a 256 bit register
   filtersReg32 = MM256_BROADCASTSI128_SI256(filtersReg);
@@ -308,49 +698,26 @@ static void aom_filter_block1d16_v8_avx2(
   dst_stride = out_pitch << 1;
 
   // load 16 bytes 7 times in stride of src_pitch
-  srcReg32b1 =
-      _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(src_ptr)));
-  srcReg32b2 = _mm256_castsi128_si256(
-      _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch)));
-  srcReg32b3 = _mm256_castsi128_si256(
-      _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 2)));
-  srcReg32b4 = _mm256_castsi128_si256(
-      _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 3)));
-  srcReg32b5 = _mm256_castsi128_si256(
-      _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 4)));
-  srcReg32b6 = _mm256_castsi128_si256(
-      _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 5)));
+  srcReg32b1 = xx_loadu2_mi128(src_ptr + src_pitch, src_ptr);
+  srcReg32b3 =
+      xx_loadu2_mi128(src_ptr + src_pitch * 3, src_ptr + src_pitch * 2);
+  srcReg32b5 =
+      xx_loadu2_mi128(src_ptr + src_pitch * 5, src_ptr + src_pitch * 4);
   srcReg32b7 = _mm256_castsi128_si256(
       _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 6)));
 
   // have each consecutive loads on the same 256 register
-  srcReg32b1 = _mm256_inserti128_si256(srcReg32b1,
-                                       _mm256_castsi256_si128(srcReg32b2), 1);
-  srcReg32b2 = _mm256_inserti128_si256(srcReg32b2,
-                                       _mm256_castsi256_si128(srcReg32b3), 1);
-  srcReg32b3 = _mm256_inserti128_si256(srcReg32b3,
-                                       _mm256_castsi256_si128(srcReg32b4), 1);
-  srcReg32b4 = _mm256_inserti128_si256(srcReg32b4,
-                                       _mm256_castsi256_si128(srcReg32b5), 1);
-  srcReg32b5 = _mm256_inserti128_si256(srcReg32b5,
-                                       _mm256_castsi256_si128(srcReg32b6), 1);
-  srcReg32b6 = _mm256_inserti128_si256(srcReg32b6,
-                                       _mm256_castsi256_si128(srcReg32b7), 1);
-
+  srcReg32b2 = _mm256_permute2x128_si256(srcReg32b1, srcReg32b3, 0x21);
+  srcReg32b4 = _mm256_permute2x128_si256(srcReg32b3, srcReg32b5, 0x21);
+  srcReg32b6 = _mm256_permute2x128_si256(srcReg32b5, srcReg32b7, 0x21);
   // merge every two consecutive registers except the last one
   srcReg32b10 = _mm256_unpacklo_epi8(srcReg32b1, srcReg32b2);
   srcReg32b1 = _mm256_unpackhi_epi8(srcReg32b1, srcReg32b2);
 
   // save
   srcReg32b11 = _mm256_unpacklo_epi8(srcReg32b3, srcReg32b4);
-
-  // save
   srcReg32b3 = _mm256_unpackhi_epi8(srcReg32b3, srcReg32b4);
-
-  // save
   srcReg32b2 = _mm256_unpacklo_epi8(srcReg32b5, srcReg32b6);
-
-  // save
   srcReg32b5 = _mm256_unpackhi_epi8(srcReg32b5, srcReg32b6);
 
   for (i = output_height; i > 1; i -= 2) {
@@ -383,9 +750,7 @@ static void aom_filter_block1d16_v8_avx2(
 
     // add and saturate the results together
     srcReg32b10 = _mm256_adds_epi16(srcReg32b10,
-                                    _mm256_min_epi16(srcReg32b8, srcReg32b12));
-    srcReg32b10 = _mm256_adds_epi16(srcReg32b10,
-                                    _mm256_max_epi16(srcReg32b8, srcReg32b12));
+                                    _mm256_adds_epi16(srcReg32b8, srcReg32b12));
 
     // multiply 2 adjacent elements with the filter and add the result
     srcReg32b1 = _mm256_maddubs_epi16(srcReg32b1, firstFilters);
@@ -399,16 +764,13 @@ static void aom_filter_block1d16_v8_avx2(
 
     // add and saturate the results together
     srcReg32b1 = _mm256_adds_epi16(srcReg32b1,
-                                   _mm256_min_epi16(srcReg32b8, srcReg32b12));
-    srcReg32b1 = _mm256_adds_epi16(srcReg32b1,
-                                   _mm256_max_epi16(srcReg32b8, srcReg32b12));
-
-    srcReg32b10 = _mm256_adds_epi16(srcReg32b10, addFilterReg64);
-    srcReg32b1 = _mm256_adds_epi16(srcReg32b1, addFilterReg64);
+                                   _mm256_adds_epi16(srcReg32b8, srcReg32b12));
 
-    // shift by 7 bit each 16 bit
-    srcReg32b10 = _mm256_srai_epi16(srcReg32b10, 7);
-    srcReg32b1 = _mm256_srai_epi16(srcReg32b1, 7);
+    // shift by 6 bit each 16 bit
+    srcReg32b10 = _mm256_adds_epi16(srcReg32b10, addFilterReg32);
+    srcReg32b1 = _mm256_adds_epi16(srcReg32b1, addFilterReg32);
+    srcReg32b10 = _mm256_srai_epi16(srcReg32b10, 6);
+    srcReg32b1 = _mm256_srai_epi16(srcReg32b1, 6);
 
     // shrink to 8 bit each 16 bits, the first lane contain the first
     // convolve result and the second lane contain the second convolve
@@ -417,12 +779,7 @@ static void aom_filter_block1d16_v8_avx2(
 
     src_ptr += src_stride;
 
-    // save 16 bytes
-    _mm_store_si128((__m128i *)output_ptr, _mm256_castsi256_si128(srcReg32b1));
-
-    // save the next 16 bits
-    _mm_store_si128((__m128i *)(output_ptr + out_pitch),
-                    _mm256_extractf128_si256(srcReg32b1, 1));
+    xx_store2_mi128(output_ptr, out_pitch, &srcReg32b1);
 
     output_ptr += dst_stride;
 
@@ -475,24 +832,17 @@ static void aom_filter_block1d16_v8_avx2(
 
     // add and saturate the results together
     srcRegFilt1 =
-        _mm_adds_epi16(srcRegFilt1, _mm_min_epi16(srcRegFilt4, srcRegFilt6));
+        _mm_adds_epi16(srcRegFilt1, _mm_adds_epi16(srcRegFilt4, srcRegFilt6));
     srcRegFilt3 =
-        _mm_adds_epi16(srcRegFilt3, _mm_min_epi16(srcRegFilt5, srcRegFilt7));
+        _mm_adds_epi16(srcRegFilt3, _mm_adds_epi16(srcRegFilt5, srcRegFilt7));
 
-    // add and saturate the results together
+    // shift by 6 bit each 16 bit
     srcRegFilt1 =
-        _mm_adds_epi16(srcRegFilt1, _mm_max_epi16(srcRegFilt4, srcRegFilt6));
+        _mm_adds_epi16(srcRegFilt1, _mm256_castsi256_si128(addFilterReg32));
     srcRegFilt3 =
-        _mm_adds_epi16(srcRegFilt3, _mm_max_epi16(srcRegFilt5, srcRegFilt7));
-
-    srcRegFilt1 =
-        _mm_adds_epi16(srcRegFilt1, _mm256_castsi256_si128(addFilterReg64));
-    srcRegFilt3 =
-        _mm_adds_epi16(srcRegFilt3, _mm256_castsi256_si128(addFilterReg64));
-
-    // shift by 7 bit each 16 bit
-    srcRegFilt1 = _mm_srai_epi16(srcRegFilt1, 7);
-    srcRegFilt3 = _mm_srai_epi16(srcRegFilt3, 7);
+        _mm_adds_epi16(srcRegFilt3, _mm256_castsi256_si128(addFilterReg32));
+    srcRegFilt1 = _mm_srai_epi16(srcRegFilt1, 6);
+    srcRegFilt3 = _mm_srai_epi16(srcRegFilt3, 6);
 
     // shrink to 8 bit each 16 bits, the first lane contain the first
     // convolve result and the second lane contain the second convolve
@@ -506,21 +856,6 @@ static void aom_filter_block1d16_v8_avx2(
 
 #if HAVE_AVX2 && HAVE_SSSE3
 filter8_1dfunction aom_filter_block1d4_v8_ssse3;
-#if ARCH_X86_64
-filter8_1dfunction aom_filter_block1d8_v8_intrin_ssse3;
-filter8_1dfunction aom_filter_block1d8_h8_intrin_ssse3;
-filter8_1dfunction aom_filter_block1d4_h8_intrin_ssse3;
-#define aom_filter_block1d8_v8_avx2 aom_filter_block1d8_v8_intrin_ssse3
-#define aom_filter_block1d8_h8_avx2 aom_filter_block1d8_h8_intrin_ssse3
-#define aom_filter_block1d4_h8_avx2 aom_filter_block1d4_h8_intrin_ssse3
-#else  // ARCH_X86
-filter8_1dfunction aom_filter_block1d8_v8_ssse3;
-filter8_1dfunction aom_filter_block1d8_h8_ssse3;
-filter8_1dfunction aom_filter_block1d4_h8_ssse3;
-#define aom_filter_block1d8_v8_avx2 aom_filter_block1d8_v8_ssse3
-#define aom_filter_block1d8_h8_avx2 aom_filter_block1d8_h8_ssse3
-#define aom_filter_block1d4_h8_avx2 aom_filter_block1d4_h8_ssse3
-#endif  // ARCH_X86_64
 filter8_1dfunction aom_filter_block1d16_v2_ssse3;
 filter8_1dfunction aom_filter_block1d16_h2_ssse3;
 filter8_1dfunction aom_filter_block1d8_v2_ssse3;
diff --git a/third_party/aom/aom_dsp/x86/convolve_avx2.h b/third_party/aom/aom_dsp/x86/convolve_avx2.h
index 7790baf2e..72fabd236 100644
--- a/third_party/aom/aom_dsp/x86/convolve_avx2.h
+++ b/third_party/aom/aom_dsp/x86/convolve_avx2.h
@@ -13,31 +13,27 @@
 #define AOM_DSP_X86_CONVOLVE_AVX2_H_
 
 // filters for 16
-DECLARE_ALIGNED(32, static const uint8_t, filt1_global_avx2[32]) = {
-  0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8,
-  0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8
+DECLARE_ALIGNED(32, static const uint8_t, filt_global_avx2[]) = {
+  0,  1,  1,  2,  2, 3,  3,  4,  4,  5,  5,  6,  6,  7,  7,  8,  0,  1,  1,
+  2,  2,  3,  3,  4, 4,  5,  5,  6,  6,  7,  7,  8,  2,  3,  3,  4,  4,  5,
+  5,  6,  6,  7,  7, 8,  8,  9,  9,  10, 2,  3,  3,  4,  4,  5,  5,  6,  6,
+  7,  7,  8,  8,  9, 9,  10, 4,  5,  5,  6,  6,  7,  7,  8,  8,  9,  9,  10,
+  10, 11, 11, 12, 4, 5,  5,  6,  6,  7,  7,  8,  8,  9,  9,  10, 10, 11, 11,
+  12, 6,  7,  7,  8, 8,  9,  9,  10, 10, 11, 11, 12, 12, 13, 13, 14, 6,  7,
+  7,  8,  8,  9,  9, 10, 10, 11, 11, 12, 12, 13, 13, 14
 };
 
-DECLARE_ALIGNED(32, static const uint8_t, filt2_global_avx2[32]) = {
-  2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10,
-  2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10
-};
-
-DECLARE_ALIGNED(32, static const uint8_t, filt3_global_avx2[32]) = {
-  4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12,
-  4, 5, 5, 6, 6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12
-};
-
-DECLARE_ALIGNED(32, static const uint8_t, filt4_global_avx2[32]) = {
-  6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14,
-  6, 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14
+DECLARE_ALIGNED(32, static const uint8_t, filt_d4_global_avx2[]) = {
+  0, 1, 2, 3,  1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6, 0, 1, 2, 3,  1, 2,
+  3, 4, 2, 3,  4, 5, 3, 4, 5, 6, 4, 5, 6, 7, 5, 6, 7, 8, 6, 7,  8, 9,
+  7, 8, 9, 10, 4, 5, 6, 7, 5, 6, 7, 8, 6, 7, 8, 9, 7, 8, 9, 10,
 };
 
 static INLINE void prepare_coeffs_lowbd(
     const InterpFilterParams *const filter_params, const int subpel_q4,
     __m256i *const coeffs /* [4] */) {
   const int16_t *const filter = av1_get_interp_filter_subpel_kernel(
-      *filter_params, subpel_q4 & SUBPEL_MASK);
+      filter_params, subpel_q4 & SUBPEL_MASK);
   const __m128i coeffs_8 = _mm_loadu_si128((__m128i *)filter);
   const __m256i filter_coeffs = _mm256_broadcastsi128_si256(coeffs_8);
 
@@ -65,7 +61,7 @@ static INLINE void prepare_coeffs(const InterpFilterParams *const filter_params,
                                   const int subpel_q4,
                                   __m256i *const coeffs /* [4] */) {
   const int16_t *filter = av1_get_interp_filter_subpel_kernel(
-      *filter_params, subpel_q4 & SUBPEL_MASK);
+      filter_params, subpel_q4 & SUBPEL_MASK);
 
   const __m128i coeff_8 = _mm_loadu_si128((__m128i *)filter);
   const __m256i coeff = _mm256_broadcastsi128_si256(coeff_8);
diff --git a/third_party/aom/aom_dsp/x86/convolve_sse2.h b/third_party/aom/aom_dsp/x86/convolve_sse2.h
index 846fe7bb4..399df5d6d 100644
--- a/third_party/aom/aom_dsp/x86/convolve_sse2.h
+++ b/third_party/aom/aom_dsp/x86/convolve_sse2.h
@@ -19,7 +19,7 @@ static INLINE void prepare_coeffs(const InterpFilterParams *const filter_params,
                                   const int subpel_q4,
                                   __m128i *const coeffs /* [4] */) {
   const int16_t *filter = av1_get_interp_filter_subpel_kernel(
-      *filter_params, subpel_q4 & SUBPEL_MASK);
+      filter_params, subpel_q4 & SUBPEL_MASK);
   const __m128i coeff = _mm_loadu_si128((__m128i *)filter);
 
   // coeffs 0 1 0 1 0 1 0 1
diff --git a/third_party/aom/aom_dsp/x86/highbd_convolve_avx2.c b/third_party/aom/aom_dsp/x86/highbd_convolve_avx2.c
index e5e3238d5..099fcf7fc 100644
--- a/third_party/aom/aom_dsp/x86/highbd_convolve_avx2.c
+++ b/third_party/aom/aom_dsp/x86/highbd_convolve_avx2.c
@@ -105,8 +105,8 @@ void aom_highbd_convolve_copy_avx2(const uint8_t *src8, ptrdiff_t src_stride,
 
 void av1_highbd_convolve_y_sr_avx2(const uint16_t *src, int src_stride,
                                    uint16_t *dst, int dst_stride, int w, int h,
-                                   InterpFilterParams *filter_params_x,
-                                   InterpFilterParams *filter_params_y,
+                                   const InterpFilterParams *filter_params_x,
+                                   const InterpFilterParams *filter_params_y,
                                    const int subpel_x_q4, const int subpel_y_q4,
                                    ConvolveParams *conv_params, int bd) {
   int i, j;
@@ -254,8 +254,8 @@ void av1_highbd_convolve_y_sr_avx2(const uint16_t *src, int src_stride,
 
 void av1_highbd_convolve_x_sr_avx2(const uint16_t *src, int src_stride,
                                    uint16_t *dst, int dst_stride, int w, int h,
-                                   InterpFilterParams *filter_params_x,
-                                   InterpFilterParams *filter_params_y,
+                                   const InterpFilterParams *filter_params_x,
+                                   const InterpFilterParams *filter_params_y,
                                    const int subpel_x_q4, const int subpel_y_q4,
                                    ConvolveParams *conv_params, int bd) {
   int i, j;
diff --git a/third_party/aom/aom_dsp/x86/highbd_convolve_ssse3.c b/third_party/aom/aom_dsp/x86/highbd_convolve_ssse3.c
index f7ac9b496..e7b33d1c4 100644
--- a/third_party/aom/aom_dsp/x86/highbd_convolve_ssse3.c
+++ b/third_party/aom/aom_dsp/x86/highbd_convolve_ssse3.c
@@ -18,8 +18,8 @@
 
 void av1_highbd_convolve_y_sr_ssse3(const uint16_t *src, int src_stride,
                                     uint16_t *dst, int dst_stride, int w, int h,
-                                    InterpFilterParams *filter_params_x,
-                                    InterpFilterParams *filter_params_y,
+                                    const InterpFilterParams *filter_params_x,
+                                    const InterpFilterParams *filter_params_y,
                                     const int subpel_x_q4,
                                     const int subpel_y_q4,
                                     ConvolveParams *conv_params, int bd) {
@@ -166,8 +166,8 @@ void av1_highbd_convolve_y_sr_ssse3(const uint16_t *src, int src_stride,
 
 void av1_highbd_convolve_x_sr_ssse3(const uint16_t *src, int src_stride,
                                     uint16_t *dst, int dst_stride, int w, int h,
-                                    InterpFilterParams *filter_params_x,
-                                    InterpFilterParams *filter_params_y,
+                                    const InterpFilterParams *filter_params_x,
+                                    const InterpFilterParams *filter_params_y,
                                     const int subpel_x_q4,
                                     const int subpel_y_q4,
                                     ConvolveParams *conv_params, int bd) {
diff --git a/third_party/aom/aom_dsp/x86/highbd_variance_sse2.c b/third_party/aom/aom_dsp/x86/highbd_variance_sse2.c
index fdfadc886..131c16aa9 100644
--- a/third_party/aom/aom_dsp/x86/highbd_variance_sse2.c
+++ b/third_party/aom/aom_dsp/x86/highbd_variance_sse2.c
@@ -676,7 +676,7 @@ void aom_highbd_upsampled_pred_sse2(MACROBLOCKD *xd,
     }
   }
 
-  const InterpFilterParams filter =
+  const InterpFilterParams *filter =
       av1_get_interp_filter_params_with_block_size(EIGHTTAP_REGULAR, 8);
 
   if (!subpel_x_q3 && !subpel_y_q3) {
@@ -726,14 +726,14 @@ void aom_highbd_upsampled_pred_sse2(MACROBLOCKD *xd,
     const int16_t *const kernel_y =
         av1_get_interp_filter_subpel_kernel(filter, subpel_y_q3 << 1);
     const int intermediate_height =
-        (((height - 1) * 8 + subpel_y_q3) >> 3) + filter.taps;
+        (((height - 1) * 8 + subpel_y_q3) >> 3) + filter->taps;
     assert(intermediate_height <= (MAX_SB_SIZE * 2 + 16) + 16);
-    aom_highbd_convolve8_horiz(ref8 - ref_stride * ((filter.taps >> 1) - 1),
+    aom_highbd_convolve8_horiz(ref8 - ref_stride * ((filter->taps >> 1) - 1),
                                ref_stride, CONVERT_TO_BYTEPTR(temp),
                                MAX_SB_SIZE, kernel_x, 16, NULL, -1, width,
                                intermediate_height, bd);
     aom_highbd_convolve8_vert(
-        CONVERT_TO_BYTEPTR(temp + MAX_SB_SIZE * ((filter.taps >> 1) - 1)),
+        CONVERT_TO_BYTEPTR(temp + MAX_SB_SIZE * ((filter->taps >> 1) - 1)),
         MAX_SB_SIZE, CONVERT_TO_BYTEPTR(comp_pred), width, NULL, -1, kernel_y,
         16, width, height, bd);
   }
diff --git a/third_party/aom/aom_dsp/x86/jnt_variance_ssse3.c b/third_party/aom/aom_dsp/x86/jnt_variance_ssse3.c
index 9801e285c..eaf1f347b 100644
--- a/third_party/aom/aom_dsp/x86/jnt_variance_ssse3.c
+++ b/third_party/aom/aom_dsp/x86/jnt_variance_ssse3.c
@@ -22,118 +22,12 @@
 void aom_var_filter_block2d_bil_first_pass_ssse3(
     const uint8_t *a, uint16_t *b, unsigned int src_pixels_per_line,
     unsigned int pixel_step, unsigned int output_height,
-    unsigned int output_width, const uint8_t *filter) {
-  // Note: filter[0], filter[1] could be {128, 0}, where 128 will overflow
-  // in computation using _mm_maddubs_epi16.
-  // Change {128, 0} to {64, 0} and reduce FILTER_BITS by 1 to avoid overflow.
-  const int16_t round = (1 << (FILTER_BITS - 1)) >> 1;
-  const __m128i r = _mm_set1_epi16(round);
-  const uint8_t f0 = filter[0] >> 1;
-  const uint8_t f1 = filter[1] >> 1;
-  const __m128i filters = _mm_setr_epi8(f0, f1, f0, f1, f0, f1, f0, f1, f0, f1,
-                                        f0, f1, f0, f1, f0, f1);
-  const __m128i shuffle_mask =
-      _mm_setr_epi8(0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8);
-  unsigned int i, j;
-  (void)pixel_step;
-
-  if (output_width >= 8) {
-    for (i = 0; i < output_height; ++i) {
-      for (j = 0; j < output_width; j += 8) {
-        // load source
-        __m128i source_low = xx_loadl_64(a);
-        __m128i source_hi = _mm_setzero_si128();
-
-        // avoid load undefined memory
-        if (a + 8 != NULL) source_hi = xx_loadl_64(a + 8);
-        __m128i source = _mm_unpacklo_epi64(source_low, source_hi);
-
-        // shuffle to:
-        // { a[0], a[1], a[1], a[2], a[2], a[3], a[3], a[4],
-        //   a[4], a[5], a[5], a[6], a[6], a[7], a[7], a[8] }
-        __m128i source_shuffle = _mm_shuffle_epi8(source, shuffle_mask);
-
-        // b[i] = a[i] * filter[0] + a[i + 1] * filter[1]
-        __m128i res = _mm_maddubs_epi16(source_shuffle, filters);
-
-        // round
-        res = _mm_srai_epi16(_mm_add_epi16(res, r), FILTER_BITS - 1);
-
-        xx_storeu_128(b, res);
-
-        a += 8;
-        b += 8;
-      }
-
-      a += src_pixels_per_line - output_width;
-    }
-  } else {
-    for (i = 0; i < output_height; ++i) {
-      // load source, only first 5 values are meaningful:
-      // { a[0], a[1], a[2], a[3], a[4], xxxx }
-      __m128i source = xx_loadl_64(a);
-
-      // shuffle, up to the first 8 are useful
-      // { a[0], a[1], a[1], a[2], a[2], a[3], a[3], a[4],
-      //   a[4], a[5], a[5], a[6], a[6], a[7], a[7], a[8] }
-      __m128i source_shuffle = _mm_shuffle_epi8(source, shuffle_mask);
-
-      __m128i res = _mm_maddubs_epi16(source_shuffle, filters);
-      res = _mm_srai_epi16(_mm_add_epi16(res, r), FILTER_BITS - 1);
-
-      xx_storel_64(b, res);
-
-      a += src_pixels_per_line;
-      b += output_width;
-    }
-  }
-}
+    unsigned int output_width, const uint8_t *filter);
 
 void aom_var_filter_block2d_bil_second_pass_ssse3(
     const uint16_t *a, uint8_t *b, unsigned int src_pixels_per_line,
     unsigned int pixel_step, unsigned int output_height,
-    unsigned int output_width, const uint8_t *filter) {
-  const int16_t round = (1 << FILTER_BITS) >> 1;
-  const __m128i r = _mm_set1_epi32(round);
-  const __m128i filters =
-      _mm_setr_epi16(filter[0], filter[1], filter[0], filter[1], filter[0],
-                     filter[1], filter[0], filter[1]);
-  const __m128i shuffle_mask =
-      _mm_setr_epi8(0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15);
-  const __m128i mask =
-      _mm_setr_epi8(0, 4, 8, 12, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1);
-  unsigned int i, j;
-
-  for (i = 0; i < output_height; ++i) {
-    for (j = 0; j < output_width; j += 4) {
-      // load source as:
-      // { a[0], a[1], a[2], a[3], a[w], a[w+1], a[w+2], a[w+3] }
-      __m128i source1 = xx_loadl_64(a);
-      __m128i source2 = xx_loadl_64(a + pixel_step);
-      __m128i source = _mm_unpacklo_epi64(source1, source2);
-
-      // shuffle source to:
-      // { a[0], a[w], a[1], a[w+1], a[2], a[w+2], a[3], a[w+3] }
-      __m128i source_shuffle = _mm_shuffle_epi8(source, shuffle_mask);
-
-      // b[i] = a[i] * filter[0] + a[w + i] * filter[1]
-      __m128i res = _mm_madd_epi16(source_shuffle, filters);
-
-      // round
-      res = _mm_srai_epi32(_mm_add_epi32(res, r), FILTER_BITS);
-
-      // shuffle to get each lower 8 bit of every 32 bit
-      res = _mm_shuffle_epi8(res, mask);
-
-      xx_storel_32(b, res);
-
-      a += 4;
-      b += 4;
-    }
-
-    a += src_pixels_per_line - output_width;
-  }
-}
+    unsigned int output_width, const uint8_t *filter);
 
 static INLINE void compute_jnt_comp_avg(__m128i *p0, __m128i *p1,
                                         const __m128i *w, const __m128i *r,
diff --git a/third_party/aom/aom_dsp/x86/masked_sad_intrin_avx2.c b/third_party/aom/aom_dsp/x86/masked_sad_intrin_avx2.c
new file mode 100644
index 000000000..6538e4d5e
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/masked_sad_intrin_avx2.c
@@ -0,0 +1,390 @@
+/*
+ * 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 <stdio.h>
+#include <tmmintrin.h>
+
+#include "config/aom_config.h"
+#include "config/aom_dsp_rtcd.h"
+
+#include "aom_dsp/blend.h"
+#include "aom/aom_integer.h"
+#include "aom_dsp/x86/synonyms.h"
+#include "aom_dsp/x86//masked_sad_intrin_ssse3.h"
+
+static INLINE unsigned int masked_sad32xh_avx2(
+    const uint8_t *src_ptr, int src_stride, const uint8_t *a_ptr, int a_stride,
+    const uint8_t *b_ptr, int b_stride, const uint8_t *m_ptr, int m_stride,
+    int width, int height) {
+  int x, y;
+  __m256i res = _mm256_setzero_si256();
+  const __m256i mask_max = _mm256_set1_epi8((1 << AOM_BLEND_A64_ROUND_BITS));
+  const __m256i round_scale =
+      _mm256_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS));
+  for (y = 0; y < height; y++) {
+    for (x = 0; x < width; x += 32) {
+      const __m256i src = _mm256_lddqu_si256((const __m256i *)&src_ptr[x]);
+      const __m256i a = _mm256_lddqu_si256((const __m256i *)&a_ptr[x]);
+      const __m256i b = _mm256_lddqu_si256((const __m256i *)&b_ptr[x]);
+      const __m256i m = _mm256_lddqu_si256((const __m256i *)&m_ptr[x]);
+      const __m256i m_inv = _mm256_sub_epi8(mask_max, m);
+
+      // Calculate 16 predicted pixels.
+      // Note that the maximum value of any entry of 'pred_l' or 'pred_r'
+      // is 64 * 255, so we have plenty of space to add rounding constants.
+      const __m256i data_l = _mm256_unpacklo_epi8(a, b);
+      const __m256i mask_l = _mm256_unpacklo_epi8(m, m_inv);
+      __m256i pred_l = _mm256_maddubs_epi16(data_l, mask_l);
+      pred_l = _mm256_mulhrs_epi16(pred_l, round_scale);
+
+      const __m256i data_r = _mm256_unpackhi_epi8(a, b);
+      const __m256i mask_r = _mm256_unpackhi_epi8(m, m_inv);
+      __m256i pred_r = _mm256_maddubs_epi16(data_r, mask_r);
+      pred_r = _mm256_mulhrs_epi16(pred_r, round_scale);
+
+      const __m256i pred = _mm256_packus_epi16(pred_l, pred_r);
+      res = _mm256_add_epi32(res, _mm256_sad_epu8(pred, src));
+    }
+
+    src_ptr += src_stride;
+    a_ptr += a_stride;
+    b_ptr += b_stride;
+    m_ptr += m_stride;
+  }
+  // At this point, we have two 32-bit partial SADs in lanes 0 and 2 of 'res'.
+  res = _mm256_shuffle_epi32(res, 0xd8);
+  res = _mm256_permute4x64_epi64(res, 0xd8);
+  res = _mm256_hadd_epi32(res, res);
+  res = _mm256_hadd_epi32(res, res);
+  int32_t sad = _mm256_extract_epi32(res, 0);
+  return (sad + 31) >> 6;
+}
+
+static INLINE __m256i xx_loadu2_m128i(const void *hi, const void *lo) {
+  __m128i a0 = _mm_lddqu_si128((const __m128i *)(lo));
+  __m128i a1 = _mm_lddqu_si128((const __m128i *)(hi));
+  __m256i a = _mm256_castsi128_si256(a0);
+  return _mm256_inserti128_si256(a, a1, 1);
+}
+
+static INLINE unsigned int masked_sad16xh_avx2(
+    const uint8_t *src_ptr, int src_stride, const uint8_t *a_ptr, int a_stride,
+    const uint8_t *b_ptr, int b_stride, const uint8_t *m_ptr, int m_stride,
+    int height) {
+  int y;
+  __m256i res = _mm256_setzero_si256();
+  const __m256i mask_max = _mm256_set1_epi8((1 << AOM_BLEND_A64_ROUND_BITS));
+  const __m256i round_scale =
+      _mm256_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS));
+  for (y = 0; y < height; y += 2) {
+    const __m256i src = xx_loadu2_m128i(src_ptr + src_stride, src_ptr);
+    const __m256i a = xx_loadu2_m128i(a_ptr + a_stride, a_ptr);
+    const __m256i b = xx_loadu2_m128i(b_ptr + b_stride, b_ptr);
+    const __m256i m = xx_loadu2_m128i(m_ptr + m_stride, m_ptr);
+    const __m256i m_inv = _mm256_sub_epi8(mask_max, m);
+
+    // Calculate 16 predicted pixels.
+    // Note that the maximum value of any entry of 'pred_l' or 'pred_r'
+    // is 64 * 255, so we have plenty of space to add rounding constants.
+    const __m256i data_l = _mm256_unpacklo_epi8(a, b);
+    const __m256i mask_l = _mm256_unpacklo_epi8(m, m_inv);
+    __m256i pred_l = _mm256_maddubs_epi16(data_l, mask_l);
+    pred_l = _mm256_mulhrs_epi16(pred_l, round_scale);
+
+    const __m256i data_r = _mm256_unpackhi_epi8(a, b);
+    const __m256i mask_r = _mm256_unpackhi_epi8(m, m_inv);
+    __m256i pred_r = _mm256_maddubs_epi16(data_r, mask_r);
+    pred_r = _mm256_mulhrs_epi16(pred_r, round_scale);
+
+    const __m256i pred = _mm256_packus_epi16(pred_l, pred_r);
+    res = _mm256_add_epi32(res, _mm256_sad_epu8(pred, src));
+
+    src_ptr += src_stride << 1;
+    a_ptr += a_stride << 1;
+    b_ptr += b_stride << 1;
+    m_ptr += m_stride << 1;
+  }
+  // At this point, we have two 32-bit partial SADs in lanes 0 and 2 of 'res'.
+  res = _mm256_shuffle_epi32(res, 0xd8);
+  res = _mm256_permute4x64_epi64(res, 0xd8);
+  res = _mm256_hadd_epi32(res, res);
+  res = _mm256_hadd_epi32(res, res);
+  int32_t sad = _mm256_extract_epi32(res, 0);
+  return (sad + 31) >> 6;
+}
+
+static INLINE unsigned int aom_masked_sad_avx2(
+    const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride,
+    const uint8_t *second_pred, const uint8_t *msk, int msk_stride,
+    int invert_mask, int m, int n) {
+  unsigned int sad;
+  if (!invert_mask) {
+    switch (m) {
+      case 4:
+        sad = aom_masked_sad4xh_ssse3(src, src_stride, ref, ref_stride,
+                                      second_pred, m, msk, msk_stride, n);
+        break;
+      case 8:
+        sad = aom_masked_sad8xh_ssse3(src, src_stride, ref, ref_stride,
+                                      second_pred, m, msk, msk_stride, n);
+        break;
+      case 16:
+        sad = masked_sad16xh_avx2(src, src_stride, ref, ref_stride, second_pred,
+                                  m, msk, msk_stride, n);
+        break;
+      default:
+        sad = masked_sad32xh_avx2(src, src_stride, ref, ref_stride, second_pred,
+                                  m, msk, msk_stride, m, n);
+        break;
+    }
+  } else {
+    switch (m) {
+      case 4:
+        sad = aom_masked_sad4xh_ssse3(src, src_stride, second_pred, m, ref,
+                                      ref_stride, msk, msk_stride, n);
+        break;
+      case 8:
+        sad = aom_masked_sad8xh_ssse3(src, src_stride, second_pred, m, ref,
+                                      ref_stride, msk, msk_stride, n);
+        break;
+      case 16:
+        sad = masked_sad16xh_avx2(src, src_stride, second_pred, m, ref,
+                                  ref_stride, msk, msk_stride, n);
+        break;
+      default:
+        sad = masked_sad32xh_avx2(src, src_stride, second_pred, m, ref,
+                                  ref_stride, msk, msk_stride, m, n);
+        break;
+    }
+  }
+  return sad;
+}
+
+#define MASKSADMXN_AVX2(m, n)                                                 \
+  unsigned int aom_masked_sad##m##x##n##_avx2(                                \
+      const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, \
+      const uint8_t *second_pred, const uint8_t *msk, int msk_stride,         \
+      int invert_mask) {                                                      \
+    return aom_masked_sad_avx2(src, src_stride, ref, ref_stride, second_pred, \
+                               msk, msk_stride, invert_mask, m, n);           \
+  }
+
+MASKSADMXN_AVX2(4, 4)
+MASKSADMXN_AVX2(4, 8)
+MASKSADMXN_AVX2(8, 4)
+MASKSADMXN_AVX2(8, 8)
+MASKSADMXN_AVX2(8, 16)
+MASKSADMXN_AVX2(16, 8)
+MASKSADMXN_AVX2(16, 16)
+MASKSADMXN_AVX2(16, 32)
+MASKSADMXN_AVX2(32, 16)
+MASKSADMXN_AVX2(32, 32)
+MASKSADMXN_AVX2(32, 64)
+MASKSADMXN_AVX2(64, 32)
+MASKSADMXN_AVX2(64, 64)
+MASKSADMXN_AVX2(64, 128)
+MASKSADMXN_AVX2(128, 64)
+MASKSADMXN_AVX2(128, 128)
+MASKSADMXN_AVX2(4, 16)
+MASKSADMXN_AVX2(16, 4)
+MASKSADMXN_AVX2(8, 32)
+MASKSADMXN_AVX2(32, 8)
+MASKSADMXN_AVX2(16, 64)
+MASKSADMXN_AVX2(64, 16)
+
+static INLINE unsigned int highbd_masked_sad8xh_avx2(
+    const uint8_t *src8, int src_stride, const uint8_t *a8, int a_stride,
+    const uint8_t *b8, int b_stride, const uint8_t *m_ptr, int m_stride,
+    int height) {
+  const uint16_t *src_ptr = CONVERT_TO_SHORTPTR(src8);
+  const uint16_t *a_ptr = CONVERT_TO_SHORTPTR(a8);
+  const uint16_t *b_ptr = CONVERT_TO_SHORTPTR(b8);
+  int y;
+  __m256i res = _mm256_setzero_si256();
+  const __m256i mask_max = _mm256_set1_epi16((1 << AOM_BLEND_A64_ROUND_BITS));
+  const __m256i round_const =
+      _mm256_set1_epi32((1 << AOM_BLEND_A64_ROUND_BITS) >> 1);
+  const __m256i one = _mm256_set1_epi16(1);
+
+  for (y = 0; y < height; y += 2) {
+    const __m256i src = xx_loadu2_m128i(src_ptr + src_stride, src_ptr);
+    const __m256i a = xx_loadu2_m128i(a_ptr + a_stride, a_ptr);
+    const __m256i b = xx_loadu2_m128i(b_ptr + b_stride, b_ptr);
+    // Zero-extend mask to 16 bits
+    const __m256i m = _mm256_cvtepu8_epi16(_mm_unpacklo_epi64(
+        _mm_loadl_epi64((const __m128i *)(m_ptr)),
+        _mm_loadl_epi64((const __m128i *)(m_ptr + m_stride))));
+    const __m256i m_inv = _mm256_sub_epi16(mask_max, m);
+
+    const __m256i data_l = _mm256_unpacklo_epi16(a, b);
+    const __m256i mask_l = _mm256_unpacklo_epi16(m, m_inv);
+    __m256i pred_l = _mm256_madd_epi16(data_l, mask_l);
+    pred_l = _mm256_srai_epi32(_mm256_add_epi32(pred_l, round_const),
+                               AOM_BLEND_A64_ROUND_BITS);
+
+    const __m256i data_r = _mm256_unpackhi_epi16(a, b);
+    const __m256i mask_r = _mm256_unpackhi_epi16(m, m_inv);
+    __m256i pred_r = _mm256_madd_epi16(data_r, mask_r);
+    pred_r = _mm256_srai_epi32(_mm256_add_epi32(pred_r, round_const),
+                               AOM_BLEND_A64_ROUND_BITS);
+
+    // Note: the maximum value in pred_l/r is (2^bd)-1 < 2^15,
+    // so it is safe to do signed saturation here.
+    const __m256i pred = _mm256_packs_epi32(pred_l, pred_r);
+    // There is no 16-bit SAD instruction, so we have to synthesize
+    // an 8-element SAD. We do this by storing 4 32-bit partial SADs,
+    // and accumulating them at the end
+    const __m256i diff = _mm256_abs_epi16(_mm256_sub_epi16(pred, src));
+    res = _mm256_add_epi32(res, _mm256_madd_epi16(diff, one));
+
+    src_ptr += src_stride << 1;
+    a_ptr += a_stride << 1;
+    b_ptr += b_stride << 1;
+    m_ptr += m_stride << 1;
+  }
+  // At this point, we have four 32-bit partial SADs stored in 'res'.
+  res = _mm256_hadd_epi32(res, res);
+  res = _mm256_hadd_epi32(res, res);
+  int sad = _mm256_extract_epi32(res, 0) + _mm256_extract_epi32(res, 4);
+  return (sad + 31) >> 6;
+}
+
+static INLINE unsigned int highbd_masked_sad16xh_avx2(
+    const uint8_t *src8, int src_stride, const uint8_t *a8, int a_stride,
+    const uint8_t *b8, int b_stride, const uint8_t *m_ptr, int m_stride,
+    int width, int height) {
+  const uint16_t *src_ptr = CONVERT_TO_SHORTPTR(src8);
+  const uint16_t *a_ptr = CONVERT_TO_SHORTPTR(a8);
+  const uint16_t *b_ptr = CONVERT_TO_SHORTPTR(b8);
+  int x, y;
+  __m256i res = _mm256_setzero_si256();
+  const __m256i mask_max = _mm256_set1_epi16((1 << AOM_BLEND_A64_ROUND_BITS));
+  const __m256i round_const =
+      _mm256_set1_epi32((1 << AOM_BLEND_A64_ROUND_BITS) >> 1);
+  const __m256i one = _mm256_set1_epi16(1);
+
+  for (y = 0; y < height; y++) {
+    for (x = 0; x < width; x += 16) {
+      const __m256i src = _mm256_lddqu_si256((const __m256i *)&src_ptr[x]);
+      const __m256i a = _mm256_lddqu_si256((const __m256i *)&a_ptr[x]);
+      const __m256i b = _mm256_lddqu_si256((const __m256i *)&b_ptr[x]);
+      // Zero-extend mask to 16 bits
+      const __m256i m =
+          _mm256_cvtepu8_epi16(_mm_lddqu_si128((const __m128i *)&m_ptr[x]));
+      const __m256i m_inv = _mm256_sub_epi16(mask_max, m);
+
+      const __m256i data_l = _mm256_unpacklo_epi16(a, b);
+      const __m256i mask_l = _mm256_unpacklo_epi16(m, m_inv);
+      __m256i pred_l = _mm256_madd_epi16(data_l, mask_l);
+      pred_l = _mm256_srai_epi32(_mm256_add_epi32(pred_l, round_const),
+                                 AOM_BLEND_A64_ROUND_BITS);
+
+      const __m256i data_r = _mm256_unpackhi_epi16(a, b);
+      const __m256i mask_r = _mm256_unpackhi_epi16(m, m_inv);
+      __m256i pred_r = _mm256_madd_epi16(data_r, mask_r);
+      pred_r = _mm256_srai_epi32(_mm256_add_epi32(pred_r, round_const),
+                                 AOM_BLEND_A64_ROUND_BITS);
+
+      // Note: the maximum value in pred_l/r is (2^bd)-1 < 2^15,
+      // so it is safe to do signed saturation here.
+      const __m256i pred = _mm256_packs_epi32(pred_l, pred_r);
+      // There is no 16-bit SAD instruction, so we have to synthesize
+      // an 8-element SAD. We do this by storing 4 32-bit partial SADs,
+      // and accumulating them at the end
+      const __m256i diff = _mm256_abs_epi16(_mm256_sub_epi16(pred, src));
+      res = _mm256_add_epi32(res, _mm256_madd_epi16(diff, one));
+    }
+
+    src_ptr += src_stride;
+    a_ptr += a_stride;
+    b_ptr += b_stride;
+    m_ptr += m_stride;
+  }
+  // At this point, we have four 32-bit partial SADs stored in 'res'.
+  res = _mm256_hadd_epi32(res, res);
+  res = _mm256_hadd_epi32(res, res);
+  int sad = _mm256_extract_epi32(res, 0) + _mm256_extract_epi32(res, 4);
+  return (sad + 31) >> 6;
+}
+
+static INLINE unsigned int aom_highbd_masked_sad_avx2(
+    const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride,
+    const uint8_t *second_pred, const uint8_t *msk, int msk_stride,
+    int invert_mask, int m, int n) {
+  unsigned int sad;
+  if (!invert_mask) {
+    switch (m) {
+      case 4:
+        sad =
+            aom_highbd_masked_sad4xh_ssse3(src, src_stride, ref, ref_stride,
+                                           second_pred, m, msk, msk_stride, n);
+        break;
+      case 8:
+        sad = highbd_masked_sad8xh_avx2(src, src_stride, ref, ref_stride,
+                                        second_pred, m, msk, msk_stride, n);
+        break;
+      default:
+        sad = highbd_masked_sad16xh_avx2(src, src_stride, ref, ref_stride,
+                                         second_pred, m, msk, msk_stride, m, n);
+        break;
+    }
+  } else {
+    switch (m) {
+      case 4:
+        sad =
+            aom_highbd_masked_sad4xh_ssse3(src, src_stride, second_pred, m, ref,
+                                           ref_stride, msk, msk_stride, n);
+        break;
+      case 8:
+        sad = highbd_masked_sad8xh_avx2(src, src_stride, second_pred, m, ref,
+                                        ref_stride, msk, msk_stride, n);
+        break;
+      default:
+        sad = highbd_masked_sad16xh_avx2(src, src_stride, second_pred, m, ref,
+                                         ref_stride, msk, msk_stride, m, n);
+        break;
+    }
+  }
+  return sad;
+}
+
+#define HIGHBD_MASKSADMXN_AVX2(m, n)                                      \
+  unsigned int aom_highbd_masked_sad##m##x##n##_avx2(                     \
+      const uint8_t *src8, int src_stride, const uint8_t *ref8,           \
+      int ref_stride, const uint8_t *second_pred8, const uint8_t *msk,    \
+      int msk_stride, int invert_mask) {                                  \
+    return aom_highbd_masked_sad_avx2(src8, src_stride, ref8, ref_stride, \
+                                      second_pred8, msk, msk_stride,      \
+                                      invert_mask, m, n);                 \
+  }
+
+HIGHBD_MASKSADMXN_AVX2(4, 4);
+HIGHBD_MASKSADMXN_AVX2(4, 8);
+HIGHBD_MASKSADMXN_AVX2(8, 4);
+HIGHBD_MASKSADMXN_AVX2(8, 8);
+HIGHBD_MASKSADMXN_AVX2(8, 16);
+HIGHBD_MASKSADMXN_AVX2(16, 8);
+HIGHBD_MASKSADMXN_AVX2(16, 16);
+HIGHBD_MASKSADMXN_AVX2(16, 32);
+HIGHBD_MASKSADMXN_AVX2(32, 16);
+HIGHBD_MASKSADMXN_AVX2(32, 32);
+HIGHBD_MASKSADMXN_AVX2(32, 64);
+HIGHBD_MASKSADMXN_AVX2(64, 32);
+HIGHBD_MASKSADMXN_AVX2(64, 64);
+HIGHBD_MASKSADMXN_AVX2(64, 128);
+HIGHBD_MASKSADMXN_AVX2(128, 64);
+HIGHBD_MASKSADMXN_AVX2(128, 128);
+HIGHBD_MASKSADMXN_AVX2(4, 16);
+HIGHBD_MASKSADMXN_AVX2(16, 4);
+HIGHBD_MASKSADMXN_AVX2(8, 32);
+HIGHBD_MASKSADMXN_AVX2(32, 8);
+HIGHBD_MASKSADMXN_AVX2(16, 64);
+HIGHBD_MASKSADMXN_AVX2(64, 16);
diff --git a/third_party/aom/aom_dsp/x86/masked_sad_intrin_ssse3.c b/third_party/aom/aom_dsp/x86/masked_sad_intrin_ssse3.c
index 1f42eec2f..493f9bd8f 100644
--- a/third_party/aom/aom_dsp/x86/masked_sad_intrin_ssse3.c
+++ b/third_party/aom/aom_dsp/x86/masked_sad_intrin_ssse3.c
@@ -19,6 +19,8 @@
 #include "aom/aom_integer.h"
 #include "aom_dsp/x86/synonyms.h"
 
+#include "aom_dsp/x86//masked_sad_intrin_ssse3.h"
+
 // For width a multiple of 16
 static INLINE unsigned int masked_sad_ssse3(const uint8_t *src_ptr,
                                             int src_stride,
@@ -27,16 +29,6 @@ static INLINE unsigned int masked_sad_ssse3(const uint8_t *src_ptr,
                                             const uint8_t *m_ptr, int m_stride,
                                             int width, int height);
 
-static INLINE unsigned int masked_sad8xh_ssse3(
-    const uint8_t *src_ptr, int src_stride, const uint8_t *a_ptr, int a_stride,
-    const uint8_t *b_ptr, int b_stride, const uint8_t *m_ptr, int m_stride,
-    int height);
-
-static INLINE unsigned int masked_sad4xh_ssse3(
-    const uint8_t *src_ptr, int src_stride, const uint8_t *a_ptr, int a_stride,
-    const uint8_t *b_ptr, int b_stride, const uint8_t *m_ptr, int m_stride,
-    int height);
-
 #define MASKSADMXN_SSSE3(m, n)                                                \
   unsigned int aom_masked_sad##m##x##n##_ssse3(                               \
       const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, \
@@ -56,11 +48,11 @@ static INLINE unsigned int masked_sad4xh_ssse3(
       const uint8_t *second_pred, const uint8_t *msk, int msk_stride,         \
       int invert_mask) {                                                      \
     if (!invert_mask)                                                         \
-      return masked_sad8xh_ssse3(src, src_stride, ref, ref_stride,            \
-                                 second_pred, 8, msk, msk_stride, n);         \
+      return aom_masked_sad8xh_ssse3(src, src_stride, ref, ref_stride,        \
+                                     second_pred, 8, msk, msk_stride, n);     \
     else                                                                      \
-      return masked_sad8xh_ssse3(src, src_stride, second_pred, 8, ref,        \
-                                 ref_stride, msk, msk_stride, n);             \
+      return aom_masked_sad8xh_ssse3(src, src_stride, second_pred, 8, ref,    \
+                                     ref_stride, msk, msk_stride, n);         \
   }
 
 #define MASKSAD4XN_SSSE3(n)                                                   \
@@ -69,11 +61,11 @@ static INLINE unsigned int masked_sad4xh_ssse3(
       const uint8_t *second_pred, const uint8_t *msk, int msk_stride,         \
       int invert_mask) {                                                      \
     if (!invert_mask)                                                         \
-      return masked_sad4xh_ssse3(src, src_stride, ref, ref_stride,            \
-                                 second_pred, 4, msk, msk_stride, n);         \
+      return aom_masked_sad4xh_ssse3(src, src_stride, ref, ref_stride,        \
+                                     second_pred, 4, msk, msk_stride, n);     \
     else                                                                      \
-      return masked_sad4xh_ssse3(src, src_stride, second_pred, 4, ref,        \
-                                 ref_stride, msk, msk_stride, n);             \
+      return aom_masked_sad4xh_ssse3(src, src_stride, second_pred, 4, ref,    \
+                                     ref_stride, msk, msk_stride, n);         \
   }
 
 MASKSADMXN_SSSE3(128, 128)
@@ -145,10 +137,11 @@ static INLINE unsigned int masked_sad_ssse3(const uint8_t *src_ptr,
   return (sad + 31) >> 6;
 }
 
-static INLINE unsigned int masked_sad8xh_ssse3(
-    const uint8_t *src_ptr, int src_stride, const uint8_t *a_ptr, int a_stride,
-    const uint8_t *b_ptr, int b_stride, const uint8_t *m_ptr, int m_stride,
-    int height) {
+unsigned int aom_masked_sad8xh_ssse3(const uint8_t *src_ptr, int src_stride,
+                                     const uint8_t *a_ptr, int a_stride,
+                                     const uint8_t *b_ptr, int b_stride,
+                                     const uint8_t *m_ptr, int m_stride,
+                                     int height) {
   int y;
   __m128i res = _mm_setzero_si128();
   const __m128i mask_max = _mm_set1_epi8((1 << AOM_BLEND_A64_ROUND_BITS));
@@ -189,10 +182,11 @@ static INLINE unsigned int masked_sad8xh_ssse3(
   return (sad + 31) >> 6;
 }
 
-static INLINE unsigned int masked_sad4xh_ssse3(
-    const uint8_t *src_ptr, int src_stride, const uint8_t *a_ptr, int a_stride,
-    const uint8_t *b_ptr, int b_stride, const uint8_t *m_ptr, int m_stride,
-    int height) {
+unsigned int aom_masked_sad4xh_ssse3(const uint8_t *src_ptr, int src_stride,
+                                     const uint8_t *a_ptr, int a_stride,
+                                     const uint8_t *b_ptr, int b_stride,
+                                     const uint8_t *m_ptr, int m_stride,
+                                     int height) {
   int y;
   __m128i res = _mm_setzero_si128();
   const __m128i mask_max = _mm_set1_epi8((1 << AOM_BLEND_A64_ROUND_BITS));
@@ -238,11 +232,6 @@ static INLINE unsigned int highbd_masked_sad_ssse3(
     const uint8_t *b8, int b_stride, const uint8_t *m_ptr, int m_stride,
     int width, int height);
 
-static INLINE unsigned int highbd_masked_sad4xh_ssse3(
-    const uint8_t *src8, int src_stride, const uint8_t *a8, int a_stride,
-    const uint8_t *b8, int b_stride, const uint8_t *m_ptr, int m_stride,
-    int height);
-
 #define HIGHBD_MASKSADMXN_SSSE3(m, n)                                         \
   unsigned int aom_highbd_masked_sad##m##x##n##_ssse3(                        \
       const uint8_t *src8, int src_stride, const uint8_t *ref8,               \
@@ -262,11 +251,13 @@ static INLINE unsigned int highbd_masked_sad4xh_ssse3(
       int ref_stride, const uint8_t *second_pred8, const uint8_t *msk,         \
       int msk_stride, int invert_mask) {                                       \
     if (!invert_mask)                                                          \
-      return highbd_masked_sad4xh_ssse3(src8, src_stride, ref8, ref_stride,    \
-                                        second_pred8, 4, msk, msk_stride, n);  \
+      return aom_highbd_masked_sad4xh_ssse3(src8, src_stride, ref8,            \
+                                            ref_stride, second_pred8, 4, msk,  \
+                                            msk_stride, n);                    \
     else                                                                       \
-      return highbd_masked_sad4xh_ssse3(src8, src_stride, second_pred8, 4,     \
-                                        ref8, ref_stride, msk, msk_stride, n); \
+      return aom_highbd_masked_sad4xh_ssse3(src8, src_stride, second_pred8, 4, \
+                                            ref8, ref_stride, msk, msk_stride, \
+                                            n);                                \
   }
 
 HIGHBD_MASKSADMXN_SSSE3(128, 128)
@@ -350,10 +341,11 @@ static INLINE unsigned int highbd_masked_sad_ssse3(
   return (sad + 31) >> 6;
 }
 
-static INLINE unsigned int highbd_masked_sad4xh_ssse3(
-    const uint8_t *src8, int src_stride, const uint8_t *a8, int a_stride,
-    const uint8_t *b8, int b_stride, const uint8_t *m_ptr, int m_stride,
-    int height) {
+unsigned int aom_highbd_masked_sad4xh_ssse3(const uint8_t *src8, int src_stride,
+                                            const uint8_t *a8, int a_stride,
+                                            const uint8_t *b8, int b_stride,
+                                            const uint8_t *m_ptr, int m_stride,
+                                            int height) {
   const uint16_t *src_ptr = CONVERT_TO_SHORTPTR(src8);
   const uint16_t *a_ptr = CONVERT_TO_SHORTPTR(a8);
   const uint16_t *b_ptr = CONVERT_TO_SHORTPTR(b8);
diff --git a/third_party/aom/aom_dsp/x86/masked_sad_intrin_ssse3.h b/third_party/aom/aom_dsp/x86/masked_sad_intrin_ssse3.h
new file mode 100644
index 000000000..19b429d91
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/masked_sad_intrin_ssse3.h
@@ -0,0 +1,33 @@
+/*
+ * 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.
+ */
+
+#ifndef _AOM_DSP_X86_MASKED_SAD_INTRIN_SSSE3_H
+#define _AOM_DSP_X86_MASKED_SAD_INTRIN_SSSE3_H
+
+unsigned int aom_masked_sad8xh_ssse3(const uint8_t *src_ptr, int src_stride,
+                                     const uint8_t *a_ptr, int a_stride,
+                                     const uint8_t *b_ptr, int b_stride,
+                                     const uint8_t *m_ptr, int m_stride,
+                                     int height);
+
+unsigned int aom_masked_sad4xh_ssse3(const uint8_t *src_ptr, int src_stride,
+                                     const uint8_t *a_ptr, int a_stride,
+                                     const uint8_t *b_ptr, int b_stride,
+                                     const uint8_t *m_ptr, int m_stride,
+                                     int height);
+
+unsigned int aom_highbd_masked_sad4xh_ssse3(const uint8_t *src8, int src_stride,
+                                            const uint8_t *a8, int a_stride,
+                                            const uint8_t *b8, int b_stride,
+                                            const uint8_t *m_ptr, int m_stride,
+                                            int height);
+
+#endif
diff --git a/third_party/aom/aom_dsp/x86/obmc_sad_avx2.c b/third_party/aom/aom_dsp/x86/obmc_sad_avx2.c
new file mode 100644
index 000000000..2aa2a0555
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/obmc_sad_avx2.c
@@ -0,0 +1,270 @@
+/*
+ * 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 <immintrin.h>
+
+#include "config/aom_config.h"
+
+#include "aom_ports/mem.h"
+#include "aom/aom_integer.h"
+
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/x86/obmc_intrinsic_ssse3.h"
+#include "aom_dsp/x86/synonyms.h"
+
+////////////////////////////////////////////////////////////////////////////////
+// 8 bit
+////////////////////////////////////////////////////////////////////////////////
+
+static INLINE unsigned int obmc_sad_w4_avx2(const uint8_t *pre,
+                                            const int pre_stride,
+                                            const int32_t *wsrc,
+                                            const int32_t *mask,
+                                            const int height) {
+  int n = 0;
+  __m256i v_sad_d = _mm256_setzero_si256();
+  const __m256i v_bias_d = _mm256_set1_epi32((1 << 12) >> 1);
+
+  do {
+    const __m128i v_p_b_0 = xx_loadl_32(pre);
+    const __m128i v_p_b_1 = xx_loadl_32(pre + pre_stride);
+    const __m128i v_p_b = _mm_unpacklo_epi32(v_p_b_0, v_p_b_1);
+    const __m256i v_m_d = _mm256_lddqu_si256((__m256i *)(mask + n));
+    const __m256i v_w_d = _mm256_lddqu_si256((__m256i *)(wsrc + n));
+
+    const __m256i v_p_d = _mm256_cvtepu8_epi32(v_p_b);
+
+    // Values in both pre and mask fit in 15 bits, and are packed at 32 bit
+    // boundaries. We use pmaddwd, as it has lower latency on Haswell
+    // than pmulld but produces the same result with these inputs.
+    const __m256i v_pm_d = _mm256_madd_epi16(v_p_d, v_m_d);
+
+    const __m256i v_diff_d = _mm256_sub_epi32(v_w_d, v_pm_d);
+    const __m256i v_absdiff_d = _mm256_abs_epi32(v_diff_d);
+
+    // Rounded absolute difference
+    const __m256i v_tmp_d = _mm256_add_epi32(v_absdiff_d, v_bias_d);
+    const __m256i v_rad_d = _mm256_srli_epi32(v_tmp_d, 12);
+
+    v_sad_d = _mm256_add_epi32(v_sad_d, v_rad_d);
+
+    n += 8;
+    pre += pre_stride << 1;
+  } while (n < 8 * (height >> 1));
+
+  __m128i v_sad_d_0 = _mm256_castsi256_si128(v_sad_d);
+  __m128i v_sad_d_1 = _mm256_extracti128_si256(v_sad_d, 1);
+  v_sad_d_0 = _mm_add_epi32(v_sad_d_0, v_sad_d_1);
+  return xx_hsum_epi32_si32(v_sad_d_0);
+}
+
+static INLINE unsigned int obmc_sad_w8n_avx2(
+    const uint8_t *pre, const int pre_stride, const int32_t *wsrc,
+    const int32_t *mask, const int width, const int height) {
+  const int pre_step = pre_stride - width;
+  int n = 0;
+  __m256i v_sad_d = _mm256_setzero_si256();
+  const __m256i v_bias_d = _mm256_set1_epi32((1 << 12) >> 1);
+  assert(width >= 8);
+  assert(IS_POWER_OF_TWO(width));
+
+  do {
+    const __m128i v_p0_b = xx_loadl_64(pre + n);
+    const __m256i v_m0_d = _mm256_lddqu_si256((__m256i *)(mask + n));
+    const __m256i v_w0_d = _mm256_lddqu_si256((__m256i *)(wsrc + n));
+
+    const __m256i v_p0_d = _mm256_cvtepu8_epi32(v_p0_b);
+
+    // Values in both pre and mask fit in 15 bits, and are packed at 32 bit
+    // boundaries. We use pmaddwd, as it has lower latency on Haswell
+    // than pmulld but produces the same result with these inputs.
+    const __m256i v_pm0_d = _mm256_madd_epi16(v_p0_d, v_m0_d);
+
+    const __m256i v_diff0_d = _mm256_sub_epi32(v_w0_d, v_pm0_d);
+    const __m256i v_absdiff0_d = _mm256_abs_epi32(v_diff0_d);
+
+    // Rounded absolute difference
+    const __m256i v_tmp_d = _mm256_add_epi32(v_absdiff0_d, v_bias_d);
+    const __m256i v_rad0_d = _mm256_srli_epi32(v_tmp_d, 12);
+
+    v_sad_d = _mm256_add_epi32(v_sad_d, v_rad0_d);
+
+    n += 8;
+
+    if ((n & (width - 1)) == 0) pre += pre_step;
+  } while (n < width * height);
+
+  __m128i v_sad_d_0 = _mm256_castsi256_si128(v_sad_d);
+  __m128i v_sad_d_1 = _mm256_extracti128_si256(v_sad_d, 1);
+  v_sad_d_0 = _mm_add_epi32(v_sad_d_0, v_sad_d_1);
+  return xx_hsum_epi32_si32(v_sad_d_0);
+}
+
+#define OBMCSADWXH(w, h)                                          \
+  unsigned int aom_obmc_sad##w##x##h##_avx2(                      \
+      const uint8_t *pre, int pre_stride, const int32_t *wsrc,    \
+      const int32_t *msk) {                                       \
+    if (w == 4) {                                                 \
+      return obmc_sad_w4_avx2(pre, pre_stride, wsrc, msk, h);     \
+    } else {                                                      \
+      return obmc_sad_w8n_avx2(pre, pre_stride, wsrc, msk, w, h); \
+    }                                                             \
+  }
+
+OBMCSADWXH(128, 128)
+OBMCSADWXH(128, 64)
+OBMCSADWXH(64, 128)
+OBMCSADWXH(64, 64)
+OBMCSADWXH(64, 32)
+OBMCSADWXH(32, 64)
+OBMCSADWXH(32, 32)
+OBMCSADWXH(32, 16)
+OBMCSADWXH(16, 32)
+OBMCSADWXH(16, 16)
+OBMCSADWXH(16, 8)
+OBMCSADWXH(8, 16)
+OBMCSADWXH(8, 8)
+OBMCSADWXH(8, 4)
+OBMCSADWXH(4, 8)
+OBMCSADWXH(4, 4)
+OBMCSADWXH(4, 16)
+OBMCSADWXH(16, 4)
+OBMCSADWXH(8, 32)
+OBMCSADWXH(32, 8)
+OBMCSADWXH(16, 64)
+OBMCSADWXH(64, 16)
+
+////////////////////////////////////////////////////////////////////////////////
+// High bit-depth
+////////////////////////////////////////////////////////////////////////////////
+
+static INLINE unsigned int hbd_obmc_sad_w4_avx2(const uint8_t *pre8,
+                                                const int pre_stride,
+                                                const int32_t *wsrc,
+                                                const int32_t *mask,
+                                                const int height) {
+  const uint16_t *pre = CONVERT_TO_SHORTPTR(pre8);
+  int n = 0;
+  __m256i v_sad_d = _mm256_setzero_si256();
+  const __m256i v_bias_d = _mm256_set1_epi32((1 << 12) >> 1);
+  do {
+    const __m128i v_p_w_0 = xx_loadl_64(pre);
+    const __m128i v_p_w_1 = xx_loadl_64(pre + pre_stride);
+    const __m128i v_p_w = _mm_unpacklo_epi64(v_p_w_0, v_p_w_1);
+    const __m256i v_m_d = _mm256_lddqu_si256((__m256i *)(mask + n));
+    const __m256i v_w_d = _mm256_lddqu_si256((__m256i *)(wsrc + n));
+
+    const __m256i v_p_d = _mm256_cvtepu16_epi32(v_p_w);
+
+    // Values in both pre and mask fit in 15 bits, and are packed at 32 bit
+    // boundaries. We use pmaddwd, as it has lower latency on Haswell
+    // than pmulld but produces the same result with these inputs.
+    const __m256i v_pm_d = _mm256_madd_epi16(v_p_d, v_m_d);
+
+    const __m256i v_diff_d = _mm256_sub_epi32(v_w_d, v_pm_d);
+    const __m256i v_absdiff_d = _mm256_abs_epi32(v_diff_d);
+
+    // Rounded absolute difference
+
+    const __m256i v_tmp_d = _mm256_add_epi32(v_absdiff_d, v_bias_d);
+    const __m256i v_rad_d = _mm256_srli_epi32(v_tmp_d, 12);
+
+    v_sad_d = _mm256_add_epi32(v_sad_d, v_rad_d);
+
+    n += 8;
+
+    pre += pre_stride << 1;
+  } while (n < 8 * (height >> 1));
+
+  __m128i v_sad_d_0 = _mm256_castsi256_si128(v_sad_d);
+  __m128i v_sad_d_1 = _mm256_extracti128_si256(v_sad_d, 1);
+  v_sad_d_0 = _mm_add_epi32(v_sad_d_0, v_sad_d_1);
+  return xx_hsum_epi32_si32(v_sad_d_0);
+}
+
+static INLINE unsigned int hbd_obmc_sad_w8n_avx2(
+    const uint8_t *pre8, const int pre_stride, const int32_t *wsrc,
+    const int32_t *mask, const int width, const int height) {
+  const uint16_t *pre = CONVERT_TO_SHORTPTR(pre8);
+  const int pre_step = pre_stride - width;
+  int n = 0;
+  __m256i v_sad_d = _mm256_setzero_si256();
+  const __m256i v_bias_d = _mm256_set1_epi32((1 << 12) >> 1);
+
+  assert(width >= 8);
+  assert(IS_POWER_OF_TWO(width));
+
+  do {
+    const __m128i v_p0_w = _mm_lddqu_si128((__m128i *)(pre + n));
+    const __m256i v_m0_d = _mm256_lddqu_si256((__m256i *)(mask + n));
+    const __m256i v_w0_d = _mm256_lddqu_si256((__m256i *)(wsrc + n));
+
+    const __m256i v_p0_d = _mm256_cvtepu16_epi32(v_p0_w);
+
+    // Values in both pre and mask fit in 15 bits, and are packed at 32 bit
+    // boundaries. We use pmaddwd, as it has lower latency on Haswell
+    // than pmulld but produces the same result with these inputs.
+    const __m256i v_pm0_d = _mm256_madd_epi16(v_p0_d, v_m0_d);
+
+    const __m256i v_diff0_d = _mm256_sub_epi32(v_w0_d, v_pm0_d);
+    const __m256i v_absdiff0_d = _mm256_abs_epi32(v_diff0_d);
+
+    // Rounded absolute difference
+    const __m256i v_tmp_d = _mm256_add_epi32(v_absdiff0_d, v_bias_d);
+    const __m256i v_rad0_d = _mm256_srli_epi32(v_tmp_d, 12);
+
+    v_sad_d = _mm256_add_epi32(v_sad_d, v_rad0_d);
+
+    n += 8;
+
+    if (n % width == 0) pre += pre_step;
+  } while (n < width * height);
+
+  __m128i v_sad_d_0 = _mm256_castsi256_si128(v_sad_d);
+  __m128i v_sad_d_1 = _mm256_extracti128_si256(v_sad_d, 1);
+  v_sad_d_0 = _mm_add_epi32(v_sad_d_0, v_sad_d_1);
+  return xx_hsum_epi32_si32(v_sad_d_0);
+}
+
+#define HBD_OBMCSADWXH(w, h)                                           \
+  unsigned int aom_highbd_obmc_sad##w##x##h##_avx2(                    \
+      const uint8_t *pre, int pre_stride, const int32_t *wsrc,         \
+      const int32_t *mask) {                                           \
+    if (w == 4) {                                                      \
+      return hbd_obmc_sad_w4_avx2(pre, pre_stride, wsrc, mask, h);     \
+    } else {                                                           \
+      return hbd_obmc_sad_w8n_avx2(pre, pre_stride, wsrc, mask, w, h); \
+    }                                                                  \
+  }
+
+HBD_OBMCSADWXH(128, 128)
+HBD_OBMCSADWXH(128, 64)
+HBD_OBMCSADWXH(64, 128)
+HBD_OBMCSADWXH(64, 64)
+HBD_OBMCSADWXH(64, 32)
+HBD_OBMCSADWXH(32, 64)
+HBD_OBMCSADWXH(32, 32)
+HBD_OBMCSADWXH(32, 16)
+HBD_OBMCSADWXH(16, 32)
+HBD_OBMCSADWXH(16, 16)
+HBD_OBMCSADWXH(16, 8)
+HBD_OBMCSADWXH(8, 16)
+HBD_OBMCSADWXH(8, 8)
+HBD_OBMCSADWXH(8, 4)
+HBD_OBMCSADWXH(4, 8)
+HBD_OBMCSADWXH(4, 4)
+HBD_OBMCSADWXH(4, 16)
+HBD_OBMCSADWXH(16, 4)
+HBD_OBMCSADWXH(8, 32)
+HBD_OBMCSADWXH(32, 8)
+HBD_OBMCSADWXH(16, 64)
+HBD_OBMCSADWXH(64, 16)
diff --git a/third_party/aom/aom_dsp/x86/obmc_variance_sse4.c b/third_party/aom/aom_dsp/x86/obmc_variance_sse4.c
index 571aa770b..2e2f6e09f 100644
--- a/third_party/aom/aom_dsp/x86/obmc_variance_sse4.c
+++ b/third_party/aom/aom_dsp/x86/obmc_variance_sse4.c
@@ -26,6 +26,16 @@
 // 8 bit
 ////////////////////////////////////////////////////////////////////////////////
 
+void aom_var_filter_block2d_bil_first_pass_ssse3(
+    const uint8_t *a, uint16_t *b, unsigned int src_pixels_per_line,
+    unsigned int pixel_step, unsigned int output_height,
+    unsigned int output_width, const uint8_t *filter);
+
+void aom_var_filter_block2d_bil_second_pass_ssse3(
+    const uint16_t *a, uint8_t *b, unsigned int src_pixels_per_line,
+    unsigned int pixel_step, unsigned int output_height,
+    unsigned int output_width, const uint8_t *filter);
+
 static INLINE void obmc_variance_w4(const uint8_t *pre, const int pre_stride,
                                     const int32_t *wsrc, const int32_t *mask,
                                     unsigned int *const sse, int *const sum,
@@ -152,6 +162,46 @@ OBMCVARWXH(32, 8)
 OBMCVARWXH(16, 64)
 OBMCVARWXH(64, 16)
 
+#include "config/aom_dsp_rtcd.h"
+
+#define OBMC_SUBPIX_VAR(W, H)                                                \
+  uint32_t aom_obmc_sub_pixel_variance##W##x##H##_sse4_1(                    \
+      const uint8_t *pre, int pre_stride, int xoffset, int yoffset,          \
+      const int32_t *wsrc, const int32_t *mask, unsigned int *sse) {         \
+    uint16_t fdata3[(H + 1) * W];                                            \
+    uint8_t temp2[H * W];                                                    \
+                                                                             \
+    aom_var_filter_block2d_bil_first_pass_ssse3(                             \
+        pre, fdata3, pre_stride, 1, H + 1, W, bilinear_filters_2t[xoffset]); \
+    aom_var_filter_block2d_bil_second_pass_ssse3(                            \
+        fdata3, temp2, W, W, H, W, bilinear_filters_2t[yoffset]);            \
+                                                                             \
+    return aom_obmc_variance##W##x##H##_sse4_1(temp2, W, wsrc, mask, sse);   \
+  }
+
+OBMC_SUBPIX_VAR(128, 128)
+OBMC_SUBPIX_VAR(128, 64)
+OBMC_SUBPIX_VAR(64, 128)
+OBMC_SUBPIX_VAR(64, 64)
+OBMC_SUBPIX_VAR(64, 32)
+OBMC_SUBPIX_VAR(32, 64)
+OBMC_SUBPIX_VAR(32, 32)
+OBMC_SUBPIX_VAR(32, 16)
+OBMC_SUBPIX_VAR(16, 32)
+OBMC_SUBPIX_VAR(16, 16)
+OBMC_SUBPIX_VAR(16, 8)
+OBMC_SUBPIX_VAR(8, 16)
+OBMC_SUBPIX_VAR(8, 8)
+OBMC_SUBPIX_VAR(8, 4)
+OBMC_SUBPIX_VAR(4, 8)
+OBMC_SUBPIX_VAR(4, 4)
+OBMC_SUBPIX_VAR(4, 16)
+OBMC_SUBPIX_VAR(16, 4)
+OBMC_SUBPIX_VAR(8, 32)
+OBMC_SUBPIX_VAR(32, 8)
+OBMC_SUBPIX_VAR(16, 64)
+OBMC_SUBPIX_VAR(64, 16)
+
 ////////////////////////////////////////////////////////////////////////////////
 // High bit-depth
 ////////////////////////////////////////////////////////////////////////////////
diff --git a/third_party/aom/aom_dsp/x86/subtract_avx2.c b/third_party/aom/aom_dsp/x86/subtract_avx2.c
new file mode 100644
index 000000000..4389d123d
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/subtract_avx2.c
@@ -0,0 +1,108 @@
+/*
+ * 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>
+
+#include "config/aom_dsp_rtcd.h"
+
+static INLINE void subtract32_avx2(int16_t *diff_ptr, const uint8_t *src_ptr,
+                                   const uint8_t *pred_ptr) {
+  __m256i s = _mm256_lddqu_si256((__m256i *)(src_ptr));
+  __m256i p = _mm256_lddqu_si256((__m256i *)(pred_ptr));
+  __m256i s_0 = _mm256_cvtepu8_epi16(_mm256_castsi256_si128(s));
+  __m256i s_1 = _mm256_cvtepu8_epi16(_mm256_extracti128_si256(s, 1));
+  __m256i p_0 = _mm256_cvtepu8_epi16(_mm256_castsi256_si128(p));
+  __m256i p_1 = _mm256_cvtepu8_epi16(_mm256_extracti128_si256(p, 1));
+  const __m256i d_0 = _mm256_sub_epi16(s_0, p_0);
+  const __m256i d_1 = _mm256_sub_epi16(s_1, p_1);
+  _mm256_store_si256((__m256i *)(diff_ptr), d_0);
+  _mm256_store_si256((__m256i *)(diff_ptr + 16), d_1);
+}
+
+static INLINE void aom_subtract_block_16xn_avx2(
+    int rows, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr,
+    ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride) {
+  for (int32_t j = 0; j < rows; ++j) {
+    __m128i s = _mm_lddqu_si128((__m128i *)(src_ptr));
+    __m128i p = _mm_lddqu_si128((__m128i *)(pred_ptr));
+    __m256i s_0 = _mm256_cvtepu8_epi16(s);
+    __m256i p_0 = _mm256_cvtepu8_epi16(p);
+    const __m256i d_0 = _mm256_sub_epi16(s_0, p_0);
+    _mm256_store_si256((__m256i *)(diff_ptr), d_0);
+    src_ptr += src_stride;
+    pred_ptr += pred_stride;
+    diff_ptr += diff_stride;
+  }
+}
+
+static INLINE void aom_subtract_block_32xn_avx2(
+    int rows, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr,
+    ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride) {
+  for (int32_t j = 0; j < rows; ++j) {
+    subtract32_avx2(diff_ptr, src_ptr, pred_ptr);
+    src_ptr += src_stride;
+    pred_ptr += pred_stride;
+    diff_ptr += diff_stride;
+  }
+}
+
+static INLINE void aom_subtract_block_64xn_avx2(
+    int rows, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr,
+    ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride) {
+  for (int32_t j = 0; j < rows; ++j) {
+    subtract32_avx2(diff_ptr, src_ptr, pred_ptr);
+    subtract32_avx2(diff_ptr + 32, src_ptr + 32, pred_ptr + 32);
+    src_ptr += src_stride;
+    pred_ptr += pred_stride;
+    diff_ptr += diff_stride;
+  }
+}
+
+static INLINE void aom_subtract_block_128xn_avx2(
+    int rows, int16_t *diff_ptr, ptrdiff_t diff_stride, const uint8_t *src_ptr,
+    ptrdiff_t src_stride, const uint8_t *pred_ptr, ptrdiff_t pred_stride) {
+  for (int32_t j = 0; j < rows; ++j) {
+    subtract32_avx2(diff_ptr, src_ptr, pred_ptr);
+    subtract32_avx2(diff_ptr + 32, src_ptr + 32, pred_ptr + 32);
+    subtract32_avx2(diff_ptr + 64, src_ptr + 64, pred_ptr + 64);
+    subtract32_avx2(diff_ptr + 96, src_ptr + 96, pred_ptr + 96);
+    src_ptr += src_stride;
+    pred_ptr += pred_stride;
+    diff_ptr += diff_stride;
+  }
+}
+
+void aom_subtract_block_avx2(int rows, int cols, int16_t *diff_ptr,
+                             ptrdiff_t diff_stride, const uint8_t *src_ptr,
+                             ptrdiff_t src_stride, const uint8_t *pred_ptr,
+                             ptrdiff_t pred_stride) {
+  switch (cols) {
+    case 16:
+      aom_subtract_block_16xn_avx2(rows, diff_ptr, diff_stride, src_ptr,
+                                   src_stride, pred_ptr, pred_stride);
+      break;
+    case 32:
+      aom_subtract_block_32xn_avx2(rows, diff_ptr, diff_stride, src_ptr,
+                                   src_stride, pred_ptr, pred_stride);
+      break;
+    case 64:
+      aom_subtract_block_64xn_avx2(rows, diff_ptr, diff_stride, src_ptr,
+                                   src_stride, pred_ptr, pred_stride);
+      break;
+    case 128:
+      aom_subtract_block_128xn_avx2(rows, diff_ptr, diff_stride, src_ptr,
+                                    src_stride, pred_ptr, pred_stride);
+      break;
+    default:
+      aom_subtract_block_sse2(rows, cols, diff_ptr, diff_stride, src_ptr,
+                              src_stride, pred_ptr, pred_stride);
+      break;
+  }
+}
diff --git a/third_party/aom/aom_dsp/x86/txfm_common_avx2.h b/third_party/aom/aom_dsp/x86/txfm_common_avx2.h
new file mode 100644
index 000000000..bdff64b8f
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/txfm_common_avx2.h
@@ -0,0 +1,199 @@
+/*
+ * 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.
+ */
+
+#ifndef AOM_DSP_X86_TXFM_COMMON_AVX2_H_
+#define AOM_DSP_X86_TXFM_COMMON_AVX2_H_
+
+#include <emmintrin.h>
+#include "aom/aom_integer.h"
+#include "aom_dsp/x86/synonyms.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef void (*transform_1d_avx2)(const __m256i *input, __m256i *output,
+                                  int8_t cos_bit);
+
+static INLINE __m256i pair_set_w16_epi16(int16_t a, int16_t b) {
+  return _mm256_set1_epi32(
+      (int32_t)(((uint16_t)(a)) | (((uint32_t)(b)) << 16)));
+}
+
+static INLINE void btf_16_w16_avx2(const __m256i w0, const __m256i w1,
+                                   __m256i *in0, __m256i *in1, const __m256i _r,
+                                   const int32_t cos_bit) {
+  __m256i t0 = _mm256_unpacklo_epi16(*in0, *in1);
+  __m256i t1 = _mm256_unpackhi_epi16(*in0, *in1);
+  __m256i u0 = _mm256_madd_epi16(t0, w0);
+  __m256i u1 = _mm256_madd_epi16(t1, w0);
+  __m256i v0 = _mm256_madd_epi16(t0, w1);
+  __m256i v1 = _mm256_madd_epi16(t1, w1);
+
+  __m256i a0 = _mm256_add_epi32(u0, _r);
+  __m256i a1 = _mm256_add_epi32(u1, _r);
+  __m256i b0 = _mm256_add_epi32(v0, _r);
+  __m256i b1 = _mm256_add_epi32(v1, _r);
+
+  __m256i c0 = _mm256_srai_epi32(a0, cos_bit);
+  __m256i c1 = _mm256_srai_epi32(a1, cos_bit);
+  __m256i d0 = _mm256_srai_epi32(b0, cos_bit);
+  __m256i d1 = _mm256_srai_epi32(b1, cos_bit);
+
+  *in0 = _mm256_packs_epi32(c0, c1);
+  *in1 = _mm256_packs_epi32(d0, d1);
+}
+
+static INLINE void btf_16_adds_subs_avx2(__m256i *in0, __m256i *in1) {
+  const __m256i _in0 = *in0;
+  const __m256i _in1 = *in1;
+  *in0 = _mm256_adds_epi16(_in0, _in1);
+  *in1 = _mm256_subs_epi16(_in0, _in1);
+}
+
+static INLINE void btf_32_add_sub_avx2(__m256i *in0, __m256i *in1) {
+  const __m256i _in0 = *in0;
+  const __m256i _in1 = *in1;
+  *in0 = _mm256_add_epi32(_in0, _in1);
+  *in1 = _mm256_sub_epi32(_in0, _in1);
+}
+
+static INLINE void btf_16_adds_subs_out_avx2(__m256i *out0, __m256i *out1,
+                                             __m256i in0, __m256i in1) {
+  const __m256i _in0 = in0;
+  const __m256i _in1 = in1;
+  *out0 = _mm256_adds_epi16(_in0, _in1);
+  *out1 = _mm256_subs_epi16(_in0, _in1);
+}
+
+static INLINE void btf_32_add_sub_out_avx2(__m256i *out0, __m256i *out1,
+                                           __m256i in0, __m256i in1) {
+  const __m256i _in0 = in0;
+  const __m256i _in1 = in1;
+  *out0 = _mm256_add_epi32(_in0, _in1);
+  *out1 = _mm256_sub_epi32(_in0, _in1);
+}
+
+static INLINE __m256i load_16bit_to_16bit_avx2(const int16_t *a) {
+  return _mm256_load_si256((const __m256i *)a);
+}
+
+static INLINE void load_buffer_16bit_to_16bit_avx2(const int16_t *in,
+                                                   int stride, __m256i *out,
+                                                   int out_size) {
+  for (int i = 0; i < out_size; ++i) {
+    out[i] = load_16bit_to_16bit_avx2(in + i * stride);
+  }
+}
+
+static INLINE void load_buffer_16bit_to_16bit_flip_avx2(const int16_t *in,
+                                                        int stride,
+                                                        __m256i *out,
+                                                        int out_size) {
+  for (int i = 0; i < out_size; ++i) {
+    out[out_size - i - 1] = load_16bit_to_16bit_avx2(in + i * stride);
+  }
+}
+
+static INLINE __m256i load_32bit_to_16bit_w16_avx2(const int32_t *a) {
+  const __m256i a_low = _mm256_lddqu_si256((const __m256i *)a);
+  const __m256i b = _mm256_packs_epi32(a_low, *(const __m256i *)(a + 8));
+  return _mm256_permute4x64_epi64(b, 0xD8);
+}
+
+static INLINE void load_buffer_32bit_to_16bit_w16_avx2(const int32_t *in,
+                                                       int stride, __m256i *out,
+                                                       int out_size) {
+  for (int i = 0; i < out_size; ++i) {
+    out[i] = load_32bit_to_16bit_w16_avx2(in + i * stride);
+  }
+}
+
+static INLINE void transpose_16bit_16x16_avx2(const __m256i *const in,
+                                              __m256i *const out) {
+  // Unpack 16 bit elements. Goes from:
+  // in[0]: 00 01 02 03  08 09 0a 0b  04 05 06 07  0c 0d 0e 0f
+  // in[1]: 10 11 12 13  18 19 1a 1b  14 15 16 17  1c 1d 1e 1f
+  // in[2]: 20 21 22 23  28 29 2a 2b  24 25 26 27  2c 2d 2e 2f
+  // in[3]: 30 31 32 33  38 39 3a 3b  34 35 36 37  3c 3d 3e 3f
+  // in[4]: 40 41 42 43  48 49 4a 4b  44 45 46 47  4c 4d 4e 4f
+  // in[5]: 50 51 52 53  58 59 5a 5b  54 55 56 57  5c 5d 5e 5f
+  // in[6]: 60 61 62 63  68 69 6a 6b  64 65 66 67  6c 6d 6e 6f
+  // in[7]: 70 71 72 73  78 79 7a 7b  74 75 76 77  7c 7d 7e 7f
+  // in[8]: 80 81 82 83  88 89 8a 8b  84 85 86 87  8c 8d 8e 8f
+  // to:
+  // a0:    00 10 01 11  02 12 03 13  04 14 05 15  06 16 07 17
+  // a1:    20 30 21 31  22 32 23 33  24 34 25 35  26 36 27 37
+  // a2:    40 50 41 51  42 52 43 53  44 54 45 55  46 56 47 57
+  // a3:    60 70 61 71  62 72 63 73  64 74 65 75  66 76 67 77
+  // ...
+  __m256i a[16];
+  for (int i = 0; i < 16; i += 2) {
+    a[i / 2 + 0] = _mm256_unpacklo_epi16(in[i], in[i + 1]);
+    a[i / 2 + 8] = _mm256_unpackhi_epi16(in[i], in[i + 1]);
+  }
+  __m256i b[16];
+  for (int i = 0; i < 16; i += 2) {
+    b[i / 2 + 0] = _mm256_unpacklo_epi32(a[i], a[i + 1]);
+    b[i / 2 + 8] = _mm256_unpackhi_epi32(a[i], a[i + 1]);
+  }
+  __m256i c[16];
+  for (int i = 0; i < 16; i += 2) {
+    c[i / 2 + 0] = _mm256_unpacklo_epi64(b[i], b[i + 1]);
+    c[i / 2 + 8] = _mm256_unpackhi_epi64(b[i], b[i + 1]);
+  }
+  out[0 + 0] = _mm256_permute2x128_si256(c[0], c[1], 0x20);
+  out[1 + 0] = _mm256_permute2x128_si256(c[8], c[9], 0x20);
+  out[2 + 0] = _mm256_permute2x128_si256(c[4], c[5], 0x20);
+  out[3 + 0] = _mm256_permute2x128_si256(c[12], c[13], 0x20);
+
+  out[0 + 8] = _mm256_permute2x128_si256(c[0], c[1], 0x31);
+  out[1 + 8] = _mm256_permute2x128_si256(c[8], c[9], 0x31);
+  out[2 + 8] = _mm256_permute2x128_si256(c[4], c[5], 0x31);
+  out[3 + 8] = _mm256_permute2x128_si256(c[12], c[13], 0x31);
+
+  out[4 + 0] = _mm256_permute2x128_si256(c[0 + 2], c[1 + 2], 0x20);
+  out[5 + 0] = _mm256_permute2x128_si256(c[8 + 2], c[9 + 2], 0x20);
+  out[6 + 0] = _mm256_permute2x128_si256(c[4 + 2], c[5 + 2], 0x20);
+  out[7 + 0] = _mm256_permute2x128_si256(c[12 + 2], c[13 + 2], 0x20);
+
+  out[4 + 8] = _mm256_permute2x128_si256(c[0 + 2], c[1 + 2], 0x31);
+  out[5 + 8] = _mm256_permute2x128_si256(c[8 + 2], c[9 + 2], 0x31);
+  out[6 + 8] = _mm256_permute2x128_si256(c[4 + 2], c[5 + 2], 0x31);
+  out[7 + 8] = _mm256_permute2x128_si256(c[12 + 2], c[13 + 2], 0x31);
+}
+
+static INLINE void flip_buf_avx2(__m256i *in, __m256i *out, int size) {
+  for (int i = 0; i < size; ++i) {
+    out[size - i - 1] = in[i];
+  }
+}
+
+static INLINE void round_shift_16bit_w16_avx2(__m256i *in, int size, int bit) {
+  if (bit < 0) {
+    bit = -bit;
+    __m256i round = _mm256_set1_epi16(1 << (bit - 1));
+    for (int i = 0; i < size; ++i) {
+      in[i] = _mm256_adds_epi16(in[i], round);
+      in[i] = _mm256_srai_epi16(in[i], bit);
+    }
+  } else if (bit > 0) {
+    for (int i = 0; i < size; ++i) {
+      in[i] = _mm256_slli_epi16(in[i], bit);
+    }
+  }
+}
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif  // AOM_DSP_X86_TXFM_COMMON_AVX2_H_
diff --git a/third_party/aom/aom_dsp/x86/variance_avx2.c b/third_party/aom/aom_dsp/x86/variance_avx2.c
index 7d6b7d287..a7ac2c93d 100644
--- a/third_party/aom/aom_dsp/x86/variance_avx2.c
+++ b/third_party/aom/aom_dsp/x86/variance_avx2.c
@@ -324,6 +324,12 @@ static INLINE __m256i mm256_loadu2(const uint8_t *p0, const uint8_t *p1) {
   return _mm256_insertf128_si256(d, _mm_loadu_si128((const __m128i *)p0), 1);
 }
 
+static INLINE __m256i mm256_loadu2_16(const uint16_t *p0, const uint16_t *p1) {
+  const __m256i d =
+      _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)p1));
+  return _mm256_insertf128_si256(d, _mm_loadu_si128((const __m128i *)p0), 1);
+}
+
 static INLINE void comp_mask_pred_line_avx2(const __m256i s0, const __m256i s1,
                                             const __m256i a,
                                             uint8_t *comp_pred) {
@@ -401,3 +407,110 @@ void aom_comp_mask_pred_avx2(uint8_t *comp_pred, const uint8_t *pred, int width,
     } while (i < height);
   }
 }
+
+static INLINE __m256i highbd_comp_mask_pred_line_avx2(const __m256i s0,
+                                                      const __m256i s1,
+                                                      const __m256i a) {
+  const __m256i alpha_max = _mm256_set1_epi16((1 << AOM_BLEND_A64_ROUND_BITS));
+  const __m256i round_const =
+      _mm256_set1_epi32((1 << AOM_BLEND_A64_ROUND_BITS) >> 1);
+  const __m256i a_inv = _mm256_sub_epi16(alpha_max, a);
+
+  const __m256i s_lo = _mm256_unpacklo_epi16(s0, s1);
+  const __m256i a_lo = _mm256_unpacklo_epi16(a, a_inv);
+  const __m256i pred_lo = _mm256_madd_epi16(s_lo, a_lo);
+  const __m256i pred_l = _mm256_srai_epi32(
+      _mm256_add_epi32(pred_lo, round_const), AOM_BLEND_A64_ROUND_BITS);
+
+  const __m256i s_hi = _mm256_unpackhi_epi16(s0, s1);
+  const __m256i a_hi = _mm256_unpackhi_epi16(a, a_inv);
+  const __m256i pred_hi = _mm256_madd_epi16(s_hi, a_hi);
+  const __m256i pred_h = _mm256_srai_epi32(
+      _mm256_add_epi32(pred_hi, round_const), AOM_BLEND_A64_ROUND_BITS);
+
+  const __m256i comp = _mm256_packs_epi32(pred_l, pred_h);
+
+  return comp;
+}
+
+void aom_highbd_comp_mask_pred_avx2(uint16_t *comp_pred, const uint8_t *pred8,
+                                    int width, int height, const uint8_t *ref8,
+                                    int ref_stride, const uint8_t *mask,
+                                    int mask_stride, int invert_mask) {
+  int i = 0;
+  uint16_t *pred = CONVERT_TO_SHORTPTR(pred8);
+  uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);
+  const uint16_t *src0 = invert_mask ? pred : ref;
+  const uint16_t *src1 = invert_mask ? ref : pred;
+  const int stride0 = invert_mask ? width : ref_stride;
+  const int stride1 = invert_mask ? ref_stride : width;
+  const __m256i zero = _mm256_setzero_si256();
+
+  if (width == 8) {
+    do {
+      const __m256i s0 = mm256_loadu2_16(src0 + stride0, src0);
+      const __m256i s1 = mm256_loadu2_16(src1 + stride1, src1);
+
+      const __m128i m_l = _mm_loadl_epi64((const __m128i *)mask);
+      const __m128i m_h = _mm_loadl_epi64((const __m128i *)(mask + 8));
+
+      __m256i m = _mm256_castsi128_si256(m_l);
+      m = _mm256_insertf128_si256(m, m_h, 1);
+      const __m256i m_16 = _mm256_unpacklo_epi8(m, zero);
+
+      const __m256i comp = highbd_comp_mask_pred_line_avx2(s0, s1, m_16);
+
+      _mm_storeu_si128((__m128i *)(comp_pred), _mm256_castsi256_si128(comp));
+
+      _mm_storeu_si128((__m128i *)(comp_pred + width),
+                       _mm256_extractf128_si256(comp, 1));
+
+      src0 += (stride0 << 1);
+      src1 += (stride1 << 1);
+      mask += (mask_stride << 1);
+      comp_pred += (width << 1);
+      i += 2;
+    } while (i < height);
+  } else if (width == 16) {
+    do {
+      const __m256i s0 = _mm256_loadu_si256((const __m256i *)(src0));
+      const __m256i s1 = _mm256_loadu_si256((const __m256i *)(src1));
+      const __m256i m_16 =
+          _mm256_cvtepu8_epi16(_mm_loadu_si128((const __m128i *)mask));
+
+      const __m256i comp = highbd_comp_mask_pred_line_avx2(s0, s1, m_16);
+
+      _mm256_storeu_si256((__m256i *)comp_pred, comp);
+
+      src0 += stride0;
+      src1 += stride1;
+      mask += mask_stride;
+      comp_pred += width;
+      i += 1;
+    } while (i < height);
+  } else if (width == 32) {
+    do {
+      const __m256i s0 = _mm256_loadu_si256((const __m256i *)src0);
+      const __m256i s2 = _mm256_loadu_si256((const __m256i *)(src0 + 16));
+      const __m256i s1 = _mm256_loadu_si256((const __m256i *)src1);
+      const __m256i s3 = _mm256_loadu_si256((const __m256i *)(src1 + 16));
+
+      const __m256i m01_16 =
+          _mm256_cvtepu8_epi16(_mm_loadu_si128((const __m128i *)mask));
+      const __m256i m23_16 =
+          _mm256_cvtepu8_epi16(_mm_loadu_si128((const __m128i *)(mask + 16)));
+
+      const __m256i comp = highbd_comp_mask_pred_line_avx2(s0, s1, m01_16);
+      const __m256i comp1 = highbd_comp_mask_pred_line_avx2(s2, s3, m23_16);
+
+      _mm256_storeu_si256((__m256i *)comp_pred, comp);
+      _mm256_storeu_si256((__m256i *)(comp_pred + 16), comp1);
+
+      src0 += stride0;
+      src1 += stride1;
+      mask += mask_stride;
+      comp_pred += width;
+      i += 1;
+    } while (i < height);
+  }
+}
diff --git a/third_party/aom/aom_dsp/x86/variance_impl_ssse3.c b/third_party/aom/aom_dsp/x86/variance_impl_ssse3.c
new file mode 100644
index 000000000..66b0d7d84
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/variance_impl_ssse3.c
@@ -0,0 +1,129 @@
+/*
+ * 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 <tmmintrin.h>
+
+#include "config/aom_config.h"
+#include "config/aom_dsp_rtcd.h"
+
+#include "aom_dsp/x86/synonyms.h"
+
+void aom_var_filter_block2d_bil_first_pass_ssse3(
+    const uint8_t *a, uint16_t *b, unsigned int src_pixels_per_line,
+    unsigned int pixel_step, unsigned int output_height,
+    unsigned int output_width, const uint8_t *filter) {
+  // Note: filter[0], filter[1] could be {128, 0}, where 128 will overflow
+  // in computation using _mm_maddubs_epi16.
+  // Change {128, 0} to {64, 0} and reduce FILTER_BITS by 1 to avoid overflow.
+  const int16_t round = (1 << (FILTER_BITS - 1)) >> 1;
+  const __m128i r = _mm_set1_epi16(round);
+  const uint8_t f0 = filter[0] >> 1;
+  const uint8_t f1 = filter[1] >> 1;
+  const __m128i filters = _mm_setr_epi8(f0, f1, f0, f1, f0, f1, f0, f1, f0, f1,
+                                        f0, f1, f0, f1, f0, f1);
+  unsigned int i, j;
+  (void)pixel_step;
+
+  if (output_width >= 8) {
+    for (i = 0; i < output_height; ++i) {
+      for (j = 0; j < output_width; j += 8) {
+        // load source
+        __m128i source_low = xx_loadl_64(a);
+        __m128i source_hi = xx_loadl_64(a + 1);
+
+        // unpack to:
+        // { a[0], a[1], a[1], a[2], a[2], a[3], a[3], a[4],
+        //   a[4], a[5], a[5], a[6], a[6], a[7], a[7], a[8] }
+        __m128i source = _mm_unpacklo_epi8(source_low, source_hi);
+
+        // b[i] = a[i] * filter[0] + a[i + 1] * filter[1]
+        __m128i res = _mm_maddubs_epi16(source, filters);
+
+        // round
+        res = _mm_srai_epi16(_mm_add_epi16(res, r), FILTER_BITS - 1);
+
+        xx_storeu_128(b, res);
+
+        a += 8;
+        b += 8;
+      }
+
+      a += src_pixels_per_line - output_width;
+    }
+  } else {
+    const __m128i shuffle_mask =
+        _mm_setr_epi8(0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7, 8);
+    for (i = 0; i < output_height; ++i) {
+      // load source, only first 5 values are meaningful:
+      // { a[0], a[1], a[2], a[3], a[4], xxxx }
+      __m128i source = xx_loadl_64(a);
+
+      // shuffle, up to the first 8 are useful
+      // { a[0], a[1], a[1], a[2], a[2], a[3], a[3], a[4],
+      //   a[4], a[5], a[5], a[6], a[6], a[7], a[7], a[8] }
+      __m128i source_shuffle = _mm_shuffle_epi8(source, shuffle_mask);
+
+      __m128i res = _mm_maddubs_epi16(source_shuffle, filters);
+      res = _mm_srai_epi16(_mm_add_epi16(res, r), FILTER_BITS - 1);
+
+      xx_storel_64(b, res);
+
+      a += src_pixels_per_line;
+      b += output_width;
+    }
+  }
+}
+
+void aom_var_filter_block2d_bil_second_pass_ssse3(
+    const uint16_t *a, uint8_t *b, unsigned int src_pixels_per_line,
+    unsigned int pixel_step, unsigned int output_height,
+    unsigned int output_width, const uint8_t *filter) {
+  const int16_t round = (1 << FILTER_BITS) >> 1;
+  const __m128i r = _mm_set1_epi32(round);
+  const __m128i filters =
+      _mm_setr_epi16(filter[0], filter[1], filter[0], filter[1], filter[0],
+                     filter[1], filter[0], filter[1]);
+  const __m128i shuffle_mask =
+      _mm_setr_epi8(0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15);
+  const __m128i mask =
+      _mm_setr_epi8(0, 4, 8, 12, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1);
+  unsigned int i, j;
+
+  for (i = 0; i < output_height; ++i) {
+    for (j = 0; j < output_width; j += 4) {
+      // load source as:
+      // { a[0], a[1], a[2], a[3], a[w], a[w+1], a[w+2], a[w+3] }
+      __m128i source1 = xx_loadl_64(a);
+      __m128i source2 = xx_loadl_64(a + pixel_step);
+      __m128i source = _mm_unpacklo_epi64(source1, source2);
+
+      // shuffle source to:
+      // { a[0], a[w], a[1], a[w+1], a[2], a[w+2], a[3], a[w+3] }
+      __m128i source_shuffle = _mm_shuffle_epi8(source, shuffle_mask);
+
+      // b[i] = a[i] * filter[0] + a[w + i] * filter[1]
+      __m128i res = _mm_madd_epi16(source_shuffle, filters);
+
+      // round
+      res = _mm_srai_epi32(_mm_add_epi32(res, r), FILTER_BITS);
+
+      // shuffle to get each lower 8 bit of every 32 bit
+      res = _mm_shuffle_epi8(res, mask);
+
+      xx_storel_32(b, res);
+
+      a += 4;
+      b += 4;
+    }
+
+    a += src_pixels_per_line - output_width;
+  }
+}
diff --git a/third_party/aom/aom_dsp/x86/variance_sse2.c b/third_party/aom/aom_dsp/x86/variance_sse2.c
index c8c90a7dc..7e3c5d5db 100644
--- a/third_party/aom/aom_dsp/x86/variance_sse2.c
+++ b/third_party/aom/aom_dsp/x86/variance_sse2.c
@@ -569,7 +569,7 @@ void aom_upsampled_pred_sse2(MACROBLOCKD *xd, const struct AV1Common *const cm,
     }
   }
 
-  const InterpFilterParams filter =
+  const InterpFilterParams *filter =
       av1_get_interp_filter_params_with_block_size(EIGHTTAP_REGULAR, 8);
 
   if (!subpel_x_q3 && !subpel_y_q3) {
@@ -633,12 +633,12 @@ void aom_upsampled_pred_sse2(MACROBLOCKD *xd, const struct AV1Common *const cm,
     const int16_t *const kernel_y =
         av1_get_interp_filter_subpel_kernel(filter, subpel_y_q3 << 1);
     const int intermediate_height =
-        (((height - 1) * 8 + subpel_y_q3) >> 3) + filter.taps;
+        (((height - 1) * 8 + subpel_y_q3) >> 3) + filter->taps;
     assert(intermediate_height <= (MAX_SB_SIZE * 2 + 16) + 16);
-    aom_convolve8_horiz(ref - ref_stride * ((filter.taps >> 1) - 1), ref_stride,
-                        temp, MAX_SB_SIZE, kernel_x, 16, NULL, -1, width,
-                        intermediate_height);
-    aom_convolve8_vert(temp + MAX_SB_SIZE * ((filter.taps >> 1) - 1),
+    aom_convolve8_horiz(ref - ref_stride * ((filter->taps >> 1) - 1),
+                        ref_stride, temp, MAX_SB_SIZE, kernel_x, 16, NULL, -1,
+                        width, intermediate_height);
+    aom_convolve8_vert(temp + MAX_SB_SIZE * ((filter->taps >> 1) - 1),
                        MAX_SB_SIZE, comp_pred, width, NULL, -1, kernel_y, 16,
                        width, height);
   }