Adding upstream version 124.0.1.upstream/124.0.1

Signed-off-by: Daniel Baumann <daniel.baumann@progress-linux.org>

115 files changed, 58370 insertions, 0 deletions

diff --git a/third_party/aom/aom_dsp/x86/adaptive_quantize_avx2.c b/third_party/aom/aom_dsp/x86/adaptive_quantize_avx2.c
new file mode 100644
index 0000000000..b3dede75d5
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/adaptive_quantize_avx2.c
@@ -0,0 +1,244 @@
+/*
+ * Copyright (c) 2019, 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"
+#include "aom/aom_integer.h"
+#include "aom_dsp/quantize.h"
+#include "aom_dsp/x86/quantize_x86.h"
+
+static INLINE void load_b_values_avx2(const int16_t *zbin_ptr, __m256i *zbin,
+                                      const int16_t *round_ptr, __m256i *round,
+                                      const int16_t *quant_ptr, __m256i *quant,
+                                      const int16_t *dequant_ptr,
+                                      __m256i *dequant,
+                                      const int16_t *shift_ptr,
+                                      __m256i *shift) {
+  *zbin = _mm256_castsi128_si256(_mm_load_si128((const __m128i *)zbin_ptr));
+  *zbin = _mm256_permute4x64_epi64(*zbin, 0x54);
+  *zbin = _mm256_sub_epi16(*zbin, _mm256_set1_epi16(1));
+  *round = _mm256_castsi128_si256(_mm_load_si128((const __m128i *)round_ptr));
+  *round = _mm256_permute4x64_epi64(*round, 0x54);
+  *quant = _mm256_castsi128_si256(_mm_load_si128((const __m128i *)quant_ptr));
+  *quant = _mm256_permute4x64_epi64(*quant, 0x54);
+  *dequant =
+      _mm256_castsi128_si256(_mm_load_si128((const __m128i *)dequant_ptr));
+  *dequant = _mm256_permute4x64_epi64(*dequant, 0x54);
+  *shift = _mm256_castsi128_si256(_mm_load_si128((const __m128i *)shift_ptr));
+  *shift = _mm256_permute4x64_epi64(*shift, 0x54);
+}
+
+static INLINE __m256i load_coefficients_avx2(const tran_low_t *coeff_ptr) {
+  const __m256i coeff1 = _mm256_load_si256((__m256i *)(coeff_ptr));
+  const __m256i coeff2 = _mm256_load_si256((__m256i *)(coeff_ptr + 8));
+  return _mm256_packs_epi32(coeff1, coeff2);
+}
+
+static INLINE void update_mask1_avx2(__m256i *cmp_mask,
+                                     const int16_t *iscan_ptr, int *is_found,
+                                     __m256i *mask) {
+  __m256i temp_mask = _mm256_setzero_si256();
+  if (_mm256_movemask_epi8(*cmp_mask)) {
+    __m256i iscan = _mm256_loadu_si256((const __m256i *)(iscan_ptr));
+    temp_mask = _mm256_and_si256(*cmp_mask, iscan);
+    *is_found = 1;
+  }
+  *mask = _mm256_max_epi16(temp_mask, *mask);
+}
+
+static INLINE void update_mask0_avx2(__m256i *qcoeff, __m256i *threshold,
+                                     const int16_t *iscan_ptr, int *is_found,
+                                     __m256i *mask) {
+  __m256i zero = _mm256_setzero_si256();
+  __m256i coeff[2], cmp_mask0, cmp_mask1;
+  coeff[0] = _mm256_unpacklo_epi16(*qcoeff, zero);
+  coeff[1] = _mm256_unpackhi_epi16(*qcoeff, zero);
+  coeff[0] = _mm256_slli_epi32(coeff[0], AOM_QM_BITS);
+  cmp_mask0 = _mm256_cmpgt_epi32(coeff[0], threshold[0]);
+  coeff[1] = _mm256_slli_epi32(coeff[1], AOM_QM_BITS);
+  cmp_mask1 = _mm256_cmpgt_epi32(coeff[1], threshold[1]);
+  cmp_mask0 =
+      _mm256_permute4x64_epi64(_mm256_packs_epi32(cmp_mask0, cmp_mask1), 0xd8);
+  update_mask1_avx2(&cmp_mask0, iscan_ptr, is_found, mask);
+}
+
+static INLINE void calculate_qcoeff_avx2(__m256i *coeff, const __m256i *round,
+                                         const __m256i *quant,
+                                         const __m256i *shift) {
+  __m256i tmp, qcoeff;
+  qcoeff = _mm256_adds_epi16(*coeff, *round);
+  tmp = _mm256_mulhi_epi16(qcoeff, *quant);
+  qcoeff = _mm256_add_epi16(tmp, qcoeff);
+  *coeff = _mm256_mulhi_epi16(qcoeff, *shift);
+}
+
+static INLINE __m256i calculate_dqcoeff_avx2(__m256i qcoeff, __m256i dequant) {
+  return _mm256_mullo_epi16(qcoeff, dequant);
+}
+
+static INLINE void store_coefficients_avx2(__m256i coeff_vals,
+                                           tran_low_t *coeff_ptr) {
+  __m256i coeff_sign = _mm256_srai_epi16(coeff_vals, 15);
+  __m256i coeff_vals_lo = _mm256_unpacklo_epi16(coeff_vals, coeff_sign);
+  __m256i coeff_vals_hi = _mm256_unpackhi_epi16(coeff_vals, coeff_sign);
+  _mm256_store_si256((__m256i *)(coeff_ptr), coeff_vals_lo);
+  _mm256_store_si256((__m256i *)(coeff_ptr + 8), coeff_vals_hi);
+}
+
+void aom_quantize_b_adaptive_avx2(
+    const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr,
+    const int16_t *round_ptr, const int16_t *quant_ptr,
+    const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
+    tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
+    const int16_t *scan, const int16_t *iscan) {
+  int index = 16;
+  int non_zero_count = 0;
+  int non_zero_count_prescan_add_zero = 0;
+  int is_found0 = 0, is_found1 = 0;
+  int eob = -1;
+  const __m256i zero = _mm256_setzero_si256();
+  __m256i zbin, round, quant, dequant, shift;
+  __m256i coeff, qcoeff;
+  __m256i cmp_mask, mask0 = zero, mask1 = zero;
+  __m128i temp_mask0, temp_mask1;
+  int prescan_add[2];
+  int thresh[2];
+  const qm_val_t wt = (1 << AOM_QM_BITS);
+  for (int i = 0; i < 2; ++i) {
+    prescan_add[i] = ROUND_POWER_OF_TWO(dequant_ptr[i] * EOB_FACTOR, 7);
+    thresh[i] = (zbin_ptr[i] * wt + prescan_add[i]) - 1;
+  }
+  __m256i threshold[2];
+  threshold[0] = _mm256_set1_epi32(thresh[0]);
+  threshold[1] = _mm256_set1_epi32(thresh[1]);
+  threshold[0] = _mm256_blend_epi32(threshold[0], threshold[1], 0xfe);
+
+#if SKIP_EOB_FACTOR_ADJUST
+  int first = -1;
+#endif
+
+  // Setup global values.
+  load_b_values_avx2(zbin_ptr, &zbin, round_ptr, &round, quant_ptr, &quant,
+                     dequant_ptr, &dequant, quant_shift_ptr, &shift);
+
+  // Do DC and first 15 AC.
+  coeff = load_coefficients_avx2(coeff_ptr);
+  qcoeff = _mm256_abs_epi16(coeff);
+  update_mask0_avx2(&qcoeff, threshold, iscan, &is_found0, &mask0);
+  __m256i temp0 = _mm256_cmpgt_epi16(qcoeff, zbin);
+  zbin = _mm256_unpackhi_epi64(zbin, zbin);
+  cmp_mask = _mm256_permute4x64_epi64(temp0, 0xd8);
+  update_mask1_avx2(&cmp_mask, iscan, &is_found1, &mask1);
+  threshold[0] = threshold[1];
+  if (_mm256_movemask_epi8(cmp_mask) == 0) {
+    _mm256_store_si256((__m256i *)(qcoeff_ptr), zero);
+    _mm256_store_si256((__m256i *)(qcoeff_ptr + 8), zero);
+    _mm256_store_si256((__m256i *)(dqcoeff_ptr), zero);
+    _mm256_store_si256((__m256i *)(dqcoeff_ptr + 8), zero);
+    round = _mm256_unpackhi_epi64(round, round);
+    quant = _mm256_unpackhi_epi64(quant, quant);
+    shift = _mm256_unpackhi_epi64(shift, shift);
+    dequant = _mm256_unpackhi_epi64(dequant, dequant);
+  } else {
+    calculate_qcoeff_avx2(&qcoeff, &round, &quant, &shift);
+    round = _mm256_unpackhi_epi64(round, round);
+    quant = _mm256_unpackhi_epi64(quant, quant);
+    shift = _mm256_unpackhi_epi64(shift, shift);
+    // Reinsert signs
+    qcoeff = _mm256_sign_epi16(qcoeff, coeff);
+    // Mask out zbin threshold coeffs
+    qcoeff = _mm256_and_si256(qcoeff, temp0);
+    store_coefficients_avx2(qcoeff, qcoeff_ptr);
+    coeff = calculate_dqcoeff_avx2(qcoeff, dequant);
+    dequant = _mm256_unpackhi_epi64(dequant, dequant);
+    store_coefficients_avx2(coeff, dqcoeff_ptr);
+  }
+
+  // AC only loop.
+  while (index < n_coeffs) {
+    coeff = load_coefficients_avx2(coeff_ptr + index);
+    qcoeff = _mm256_abs_epi16(coeff);
+    update_mask0_avx2(&qcoeff, threshold, iscan + index, &is_found0, &mask0);
+    temp0 = _mm256_cmpgt_epi16(qcoeff, zbin);
+    cmp_mask = _mm256_permute4x64_epi64(temp0, 0xd8);
+    update_mask1_avx2(&cmp_mask, iscan + index, &is_found1, &mask1);
+    if (_mm256_movemask_epi8(cmp_mask) == 0) {
+      _mm256_store_si256((__m256i *)(qcoeff_ptr + index), zero);
+      _mm256_store_si256((__m256i *)(qcoeff_ptr + index + 8), zero);
+      _mm256_store_si256((__m256i *)(dqcoeff_ptr + index), zero);
+      _mm256_store_si256((__m256i *)(dqcoeff_ptr + index + 8), zero);
+      index += 16;
+      continue;
+    }
+    calculate_qcoeff_avx2(&qcoeff, &round, &quant, &shift);
+    qcoeff = _mm256_sign_epi16(qcoeff, coeff);
+    qcoeff = _mm256_and_si256(qcoeff, temp0);
+    store_coefficients_avx2(qcoeff, qcoeff_ptr + index);
+    coeff = calculate_dqcoeff_avx2(qcoeff, dequant);
+    store_coefficients_avx2(coeff, dqcoeff_ptr + index);
+    index += 16;
+  }
+  if (is_found0) {
+    temp_mask0 = _mm_max_epi16(_mm256_castsi256_si128(mask0),
+                               _mm256_extracti128_si256(mask0, 1));
+    non_zero_count = calculate_non_zero_count(temp_mask0);
+  }
+  if (is_found1) {
+    temp_mask1 = _mm_max_epi16(_mm256_castsi256_si128(mask1),
+                               _mm256_extracti128_si256(mask1, 1));
+    non_zero_count_prescan_add_zero = calculate_non_zero_count(temp_mask1);
+  }
+
+  for (int i = non_zero_count_prescan_add_zero - 1; i >= non_zero_count; i--) {
+    const int rc = scan[i];
+    qcoeff_ptr[rc] = 0;
+    dqcoeff_ptr[rc] = 0;
+  }
+
+  for (int i = non_zero_count - 1; i >= 0; i--) {
+    const int rc = scan[i];
+    if (qcoeff_ptr[rc]) {
+      eob = i;
+      break;
+    }
+  }
+
+  *eob_ptr = eob + 1;
+#if SKIP_EOB_FACTOR_ADJUST
+  // TODO(Aniket): Experiment the following loop with intrinsic by combining
+  // with the quantization loop above
+  for (int i = 0; i < non_zero_count; i++) {
+    const int rc = scan[i];
+    const int qcoeff0 = qcoeff_ptr[rc];
+    if (qcoeff0) {
+      first = i;
+      break;
+    }
+  }
+  if ((*eob_ptr - 1) >= 0 && first == (*eob_ptr - 1)) {
+    const int rc = scan[(*eob_ptr - 1)];
+    if (qcoeff_ptr[rc] == 1 || qcoeff_ptr[rc] == -1) {
+      const int coeff0 = coeff_ptr[rc] * wt;
+      const int coeff_sign = AOMSIGN(coeff0);
+      const int abs_coeff = (coeff0 ^ coeff_sign) - coeff_sign;
+      const int factor = EOB_FACTOR + SKIP_EOB_FACTOR_ADJUST;
+      const int prescan_add_val =
+          ROUND_POWER_OF_TWO(dequant_ptr[rc != 0] * factor, 7);
+      if (abs_coeff <
+          (zbin_ptr[rc != 0] * (1 << AOM_QM_BITS) + prescan_add_val)) {
+        qcoeff_ptr[rc] = 0;
+        dqcoeff_ptr[rc] = 0;
+        *eob_ptr = 0;
+      }
+    }
+  }
+#endif
+}
diff --git a/third_party/aom/aom_dsp/x86/adaptive_quantize_sse2.c b/third_party/aom/aom_dsp/x86/adaptive_quantize_sse2.c
new file mode 100644
index 0000000000..503b9b4682
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/adaptive_quantize_sse2.c
@@ -0,0 +1,633 @@
+/*
+ * Copyright (c) 2019, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <emmintrin.h>
+#include "config/aom_dsp_rtcd.h"
+#include "aom/aom_integer.h"
+#include "aom_dsp/quantize.h"
+#include "aom_dsp/x86/quantize_x86.h"
+
+void aom_quantize_b_adaptive_sse2(
+    const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr,
+    const int16_t *round_ptr, const int16_t *quant_ptr,
+    const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
+    tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
+    const int16_t *scan, const int16_t *iscan) {
+  int index = 16;
+  int non_zero_count = 0;
+  int non_zero_count_prescan_add_zero = 0;
+  int is_found0 = 0, is_found1 = 0;
+  int eob = -1;
+  const __m128i zero = _mm_setzero_si128();
+  __m128i zbin, round, quant, dequant, shift;
+  __m128i coeff0, coeff1, coeff0_sign, coeff1_sign;
+  __m128i qcoeff0, qcoeff1;
+  __m128i cmp_mask0, cmp_mask1;
+  __m128i all_zero;
+  __m128i mask0 = zero, mask1 = zero;
+
+  int prescan_add[2];
+  int thresh[4];
+  const qm_val_t wt = (1 << AOM_QM_BITS);
+  for (int i = 0; i < 2; ++i) {
+    prescan_add[i] = ROUND_POWER_OF_TWO(dequant_ptr[i] * EOB_FACTOR, 7);
+    thresh[i] = (zbin_ptr[i] * wt + prescan_add[i]) - 1;
+  }
+  thresh[2] = thresh[3] = thresh[1];
+  __m128i threshold[2];
+  threshold[0] = _mm_loadu_si128((__m128i *)&thresh[0]);
+  threshold[1] = _mm_unpackhi_epi64(threshold[0], threshold[0]);
+
+#if SKIP_EOB_FACTOR_ADJUST
+  int first = -1;
+#endif
+  // Setup global values.
+  load_b_values(zbin_ptr, &zbin, round_ptr, &round, quant_ptr, &quant,
+                dequant_ptr, &dequant, quant_shift_ptr, &shift);
+
+  // Do DC and first 15 AC.
+  coeff0 = load_coefficients(coeff_ptr);
+  coeff1 = load_coefficients(coeff_ptr + 8);
+
+  // Poor man's abs().
+  coeff0_sign = _mm_srai_epi16(coeff0, 15);
+  coeff1_sign = _mm_srai_epi16(coeff1, 15);
+  qcoeff0 = invert_sign_sse2(coeff0, coeff0_sign);
+  qcoeff1 = invert_sign_sse2(coeff1, coeff1_sign);
+
+  update_mask0(&qcoeff0, &qcoeff1, threshold, iscan, &is_found0, &mask0);
+
+  cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin);
+  zbin = _mm_unpackhi_epi64(zbin, zbin);  // Switch DC to AC
+  cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);
+
+  update_mask1(&cmp_mask0, &cmp_mask1, iscan, &is_found1, &mask1);
+
+  threshold[0] = threshold[1];
+  all_zero = _mm_or_si128(cmp_mask0, cmp_mask1);
+  if (_mm_movemask_epi8(all_zero) == 0) {
+    _mm_store_si128((__m128i *)(qcoeff_ptr), zero);
+    _mm_store_si128((__m128i *)(qcoeff_ptr + 4), zero);
+    _mm_store_si128((__m128i *)(qcoeff_ptr + 8), zero);
+    _mm_store_si128((__m128i *)(qcoeff_ptr + 12), zero);
+    _mm_store_si128((__m128i *)(dqcoeff_ptr), zero);
+    _mm_store_si128((__m128i *)(dqcoeff_ptr + 4), zero);
+    _mm_store_si128((__m128i *)(dqcoeff_ptr + 8), zero);
+    _mm_store_si128((__m128i *)(dqcoeff_ptr + 12), zero);
+    round = _mm_unpackhi_epi64(round, round);
+    quant = _mm_unpackhi_epi64(quant, quant);
+    shift = _mm_unpackhi_epi64(shift, shift);
+    dequant = _mm_unpackhi_epi64(dequant, dequant);
+  } else {
+    calculate_qcoeff(&qcoeff0, round, quant, shift);
+
+    round = _mm_unpackhi_epi64(round, round);
+    quant = _mm_unpackhi_epi64(quant, quant);
+    shift = _mm_unpackhi_epi64(shift, shift);
+
+    calculate_qcoeff(&qcoeff1, round, quant, shift);
+
+    // Reinsert signs
+    qcoeff0 = invert_sign_sse2(qcoeff0, coeff0_sign);
+    qcoeff1 = invert_sign_sse2(qcoeff1, coeff1_sign);
+
+    // Mask out zbin threshold coeffs
+    qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
+    qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
+
+    store_coefficients(qcoeff0, qcoeff_ptr);
+    store_coefficients(qcoeff1, qcoeff_ptr + 8);
+
+    coeff0 = calculate_dqcoeff(qcoeff0, dequant);
+    dequant = _mm_unpackhi_epi64(dequant, dequant);
+    coeff1 = calculate_dqcoeff(qcoeff1, dequant);
+
+    store_coefficients(coeff0, dqcoeff_ptr);
+    store_coefficients(coeff1, dqcoeff_ptr + 8);
+  }
+
+  // AC only loop.
+  while (index < n_coeffs) {
+    coeff0 = load_coefficients(coeff_ptr + index);
+    coeff1 = load_coefficients(coeff_ptr + index + 8);
+
+    coeff0_sign = _mm_srai_epi16(coeff0, 15);
+    coeff1_sign = _mm_srai_epi16(coeff1, 15);
+    qcoeff0 = invert_sign_sse2(coeff0, coeff0_sign);
+    qcoeff1 = invert_sign_sse2(coeff1, coeff1_sign);
+
+    update_mask0(&qcoeff0, &qcoeff1, threshold, iscan + index, &is_found0,
+                 &mask0);
+
+    cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin);
+    cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);
+
+    update_mask1(&cmp_mask0, &cmp_mask1, iscan + index, &is_found1, &mask1);
+
+    all_zero = _mm_or_si128(cmp_mask0, cmp_mask1);
+    if (_mm_movemask_epi8(all_zero) == 0) {
+      _mm_store_si128((__m128i *)(qcoeff_ptr + index), zero);
+      _mm_store_si128((__m128i *)(qcoeff_ptr + index + 4), zero);
+      _mm_store_si128((__m128i *)(qcoeff_ptr + index + 8), zero);
+      _mm_store_si128((__m128i *)(qcoeff_ptr + index + 12), zero);
+      _mm_store_si128((__m128i *)(dqcoeff_ptr + index), zero);
+      _mm_store_si128((__m128i *)(dqcoeff_ptr + index + 4), zero);
+      _mm_store_si128((__m128i *)(dqcoeff_ptr + index + 8), zero);
+      _mm_store_si128((__m128i *)(dqcoeff_ptr + index + 12), zero);
+      index += 16;
+      continue;
+    }
+    calculate_qcoeff(&qcoeff0, round, quant, shift);
+    calculate_qcoeff(&qcoeff1, round, quant, shift);
+
+    qcoeff0 = invert_sign_sse2(qcoeff0, coeff0_sign);
+    qcoeff1 = invert_sign_sse2(qcoeff1, coeff1_sign);
+
+    qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
+    qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
+
+    store_coefficients(qcoeff0, qcoeff_ptr + index);
+    store_coefficients(qcoeff1, qcoeff_ptr + index + 8);
+
+    coeff0 = calculate_dqcoeff(qcoeff0, dequant);
+    coeff1 = calculate_dqcoeff(qcoeff1, dequant);
+
+    store_coefficients(coeff0, dqcoeff_ptr + index);
+    store_coefficients(coeff1, dqcoeff_ptr + index + 8);
+
+    index += 16;
+  }
+  if (is_found0) non_zero_count = calculate_non_zero_count(mask0);
+  if (is_found1)
+    non_zero_count_prescan_add_zero = calculate_non_zero_count(mask1);
+
+  for (int i = non_zero_count_prescan_add_zero - 1; i >= non_zero_count; i--) {
+    const int rc = scan[i];
+    qcoeff_ptr[rc] = 0;
+    dqcoeff_ptr[rc] = 0;
+  }
+
+  for (int i = non_zero_count - 1; i >= 0; i--) {
+    const int rc = scan[i];
+    if (qcoeff_ptr[rc]) {
+      eob = i;
+      break;
+    }
+  }
+
+  *eob_ptr = eob + 1;
+#if SKIP_EOB_FACTOR_ADJUST
+  // TODO(Aniket): Experiment the following loop with intrinsic by combining
+  // with the quantization loop above
+  for (int i = 0; i < non_zero_count; i++) {
+    const int rc = scan[i];
+    const int qcoeff = qcoeff_ptr[rc];
+    if (qcoeff) {
+      first = i;
+      break;
+    }
+  }
+  if ((*eob_ptr - 1) >= 0 && first == (*eob_ptr - 1)) {
+    const int rc = scan[(*eob_ptr - 1)];
+    if (qcoeff_ptr[rc] == 1 || qcoeff_ptr[rc] == -1) {
+      const int coeff = coeff_ptr[rc] * wt;
+      const int coeff_sign = AOMSIGN(coeff);
+      const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+      const int factor = EOB_FACTOR + SKIP_EOB_FACTOR_ADJUST;
+      const int prescan_add_val =
+          ROUND_POWER_OF_TWO(dequant_ptr[rc != 0] * factor, 7);
+      if (abs_coeff <
+          (zbin_ptr[rc != 0] * (1 << AOM_QM_BITS) + prescan_add_val)) {
+        qcoeff_ptr[rc] = 0;
+        dqcoeff_ptr[rc] = 0;
+        *eob_ptr = 0;
+      }
+    }
+  }
+#endif
+}
+
+void aom_quantize_b_32x32_adaptive_sse2(
+    const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr,
+    const int16_t *round_ptr, const int16_t *quant_ptr,
+    const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
+    tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
+    const int16_t *scan, const int16_t *iscan) {
+  int index = 16;
+  const int log_scale = 1;
+  int non_zero_count = 0;
+  int non_zero_count_prescan_add_zero = 0;
+  int is_found0 = 0, is_found1 = 0;
+  int eob = -1;
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i one = _mm_set1_epi16(1);
+  const __m128i log_scale_vec = _mm_set1_epi16(log_scale);
+  __m128i zbin, round, quant, dequant, shift;
+  __m128i coeff0, coeff1, coeff0_sign, coeff1_sign;
+  __m128i qcoeff0, qcoeff1;
+  __m128i cmp_mask0, cmp_mask1;
+  __m128i all_zero;
+  __m128i mask0 = zero, mask1 = zero;
+
+  const int zbins[2] = { ROUND_POWER_OF_TWO(zbin_ptr[0], log_scale),
+                         ROUND_POWER_OF_TWO(zbin_ptr[1], log_scale) };
+  int prescan_add[2];
+  int thresh[4];
+  const qm_val_t wt = (1 << AOM_QM_BITS);
+  for (int i = 0; i < 2; ++i) {
+    prescan_add[i] = ROUND_POWER_OF_TWO(dequant_ptr[i] * EOB_FACTOR, 7);
+    thresh[i] = (zbins[i] * wt + prescan_add[i]) - 1;
+  }
+  thresh[2] = thresh[3] = thresh[1];
+  __m128i threshold[2];
+  threshold[0] = _mm_loadu_si128((__m128i *)&thresh[0]);
+  threshold[1] = _mm_unpackhi_epi64(threshold[0], threshold[0]);
+
+#if SKIP_EOB_FACTOR_ADJUST
+  int first = -1;
+#endif
+  // Setup global values.
+  zbin = _mm_load_si128((const __m128i *)zbin_ptr);
+  round = _mm_load_si128((const __m128i *)round_ptr);
+  quant = _mm_load_si128((const __m128i *)quant_ptr);
+  dequant = _mm_load_si128((const __m128i *)dequant_ptr);
+  shift = _mm_load_si128((const __m128i *)quant_shift_ptr);
+
+  // Shift with rounding.
+  zbin = _mm_add_epi16(zbin, log_scale_vec);
+  round = _mm_add_epi16(round, log_scale_vec);
+  zbin = _mm_srli_epi16(zbin, log_scale);
+  round = _mm_srli_epi16(round, log_scale);
+  zbin = _mm_sub_epi16(zbin, one);
+
+  // Do DC and first 15 AC.
+  coeff0 = load_coefficients(coeff_ptr);
+  coeff1 = load_coefficients(coeff_ptr + 8);
+
+  coeff0_sign = _mm_srai_epi16(coeff0, 15);
+  coeff1_sign = _mm_srai_epi16(coeff1, 15);
+  qcoeff0 = invert_sign_sse2(coeff0, coeff0_sign);
+  qcoeff1 = invert_sign_sse2(coeff1, coeff1_sign);
+
+  update_mask0(&qcoeff0, &qcoeff1, threshold, iscan, &is_found0, &mask0);
+
+  cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin);
+  zbin = _mm_unpackhi_epi64(zbin, zbin);  // Switch DC to AC
+  cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);
+
+  update_mask1(&cmp_mask0, &cmp_mask1, iscan, &is_found1, &mask1);
+
+  threshold[0] = threshold[1];
+  all_zero = _mm_or_si128(cmp_mask0, cmp_mask1);
+  if (_mm_movemask_epi8(all_zero) == 0) {
+    _mm_store_si128((__m128i *)(qcoeff_ptr), zero);
+    _mm_store_si128((__m128i *)(qcoeff_ptr + 4), zero);
+    _mm_store_si128((__m128i *)(qcoeff_ptr + 8), zero);
+    _mm_store_si128((__m128i *)(qcoeff_ptr + 12), zero);
+    _mm_store_si128((__m128i *)(dqcoeff_ptr), zero);
+    _mm_store_si128((__m128i *)(dqcoeff_ptr + 4), zero);
+    _mm_store_si128((__m128i *)(dqcoeff_ptr + 8), zero);
+    _mm_store_si128((__m128i *)(dqcoeff_ptr + 12), zero);
+    round = _mm_unpackhi_epi64(round, round);
+    quant = _mm_unpackhi_epi64(quant, quant);
+    shift = _mm_unpackhi_epi64(shift, shift);
+    dequant = _mm_unpackhi_epi64(dequant, dequant);
+  } else {
+    calculate_qcoeff_log_scale(&qcoeff0, round, quant, &shift, &log_scale);
+    round = _mm_unpackhi_epi64(round, round);
+    quant = _mm_unpackhi_epi64(quant, quant);
+    shift = _mm_unpackhi_epi64(shift, shift);
+    calculate_qcoeff_log_scale(&qcoeff1, round, quant, &shift, &log_scale);
+
+    // Reinsert signs
+    qcoeff0 = invert_sign_sse2(qcoeff0, coeff0_sign);
+    qcoeff1 = invert_sign_sse2(qcoeff1, coeff1_sign);
+
+    // Mask out zbin threshold coeffs
+    qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
+    qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
+
+    store_coefficients(qcoeff0, qcoeff_ptr);
+    store_coefficients(qcoeff1, qcoeff_ptr + 8);
+
+    calculate_dqcoeff_and_store_log_scale(qcoeff0, dequant, zero, dqcoeff_ptr,
+                                          &log_scale);
+    dequant = _mm_unpackhi_epi64(dequant, dequant);
+    calculate_dqcoeff_and_store_log_scale(qcoeff1, dequant, zero,
+                                          dqcoeff_ptr + 8, &log_scale);
+  }
+
+  // AC only loop.
+  while (index < n_coeffs) {
+    coeff0 = load_coefficients(coeff_ptr + index);
+    coeff1 = load_coefficients(coeff_ptr + index + 8);
+
+    coeff0_sign = _mm_srai_epi16(coeff0, 15);
+    coeff1_sign = _mm_srai_epi16(coeff1, 15);
+    qcoeff0 = invert_sign_sse2(coeff0, coeff0_sign);
+    qcoeff1 = invert_sign_sse2(coeff1, coeff1_sign);
+
+    update_mask0(&qcoeff0, &qcoeff1, threshold, iscan + index, &is_found0,
+                 &mask0);
+
+    cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin);
+    cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);
+
+    update_mask1(&cmp_mask0, &cmp_mask1, iscan + index, &is_found1, &mask1);
+
+    all_zero = _mm_or_si128(cmp_mask0, cmp_mask1);
+    if (_mm_movemask_epi8(all_zero) == 0) {
+      _mm_store_si128((__m128i *)(qcoeff_ptr + index), zero);
+      _mm_store_si128((__m128i *)(qcoeff_ptr + index + 4), zero);
+      _mm_store_si128((__m128i *)(qcoeff_ptr + index + 8), zero);
+      _mm_store_si128((__m128i *)(qcoeff_ptr + index + 12), zero);
+      _mm_store_si128((__m128i *)(dqcoeff_ptr + index), zero);
+      _mm_store_si128((__m128i *)(dqcoeff_ptr + index + 4), zero);
+      _mm_store_si128((__m128i *)(dqcoeff_ptr + index + 8), zero);
+      _mm_store_si128((__m128i *)(dqcoeff_ptr + index + 12), zero);
+      index += 16;
+      continue;
+    }
+    calculate_qcoeff_log_scale(&qcoeff0, round, quant, &shift, &log_scale);
+    calculate_qcoeff_log_scale(&qcoeff1, round, quant, &shift, &log_scale);
+
+    qcoeff0 = invert_sign_sse2(qcoeff0, coeff0_sign);
+    qcoeff1 = invert_sign_sse2(qcoeff1, coeff1_sign);
+
+    qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
+    qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
+
+    store_coefficients(qcoeff0, qcoeff_ptr + index);
+    store_coefficients(qcoeff1, qcoeff_ptr + index + 8);
+
+    calculate_dqcoeff_and_store_log_scale(qcoeff0, dequant, zero,
+                                          dqcoeff_ptr + index, &log_scale);
+    calculate_dqcoeff_and_store_log_scale(qcoeff1, dequant, zero,
+                                          dqcoeff_ptr + index + 8, &log_scale);
+    index += 16;
+  }
+  if (is_found0) non_zero_count = calculate_non_zero_count(mask0);
+  if (is_found1)
+    non_zero_count_prescan_add_zero = calculate_non_zero_count(mask1);
+
+  for (int i = non_zero_count_prescan_add_zero - 1; i >= non_zero_count; i--) {
+    const int rc = scan[i];
+    qcoeff_ptr[rc] = 0;
+    dqcoeff_ptr[rc] = 0;
+  }
+
+  for (int i = non_zero_count - 1; i >= 0; i--) {
+    const int rc = scan[i];
+    if (qcoeff_ptr[rc]) {
+      eob = i;
+      break;
+    }
+  }
+
+  *eob_ptr = eob + 1;
+#if SKIP_EOB_FACTOR_ADJUST
+  // TODO(Aniket): Experiment the following loop with intrinsic by combining
+  // with the quantization loop above
+  for (int i = 0; i < non_zero_count; i++) {
+    const int rc = scan[i];
+    const int qcoeff = qcoeff_ptr[rc];
+    if (qcoeff) {
+      first = i;
+      break;
+    }
+  }
+  if ((*eob_ptr - 1) >= 0 && first == (*eob_ptr - 1)) {
+    const int rc = scan[(*eob_ptr - 1)];
+    if (qcoeff_ptr[rc] == 1 || qcoeff_ptr[rc] == -1) {
+      const int coeff = coeff_ptr[rc] * wt;
+      const int coeff_sign = AOMSIGN(coeff);
+      const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+      const int factor = EOB_FACTOR + SKIP_EOB_FACTOR_ADJUST;
+      const int prescan_add_val =
+          ROUND_POWER_OF_TWO(dequant_ptr[rc != 0] * factor, 7);
+      if (abs_coeff < (zbins[rc != 0] * (1 << AOM_QM_BITS) + prescan_add_val)) {
+        qcoeff_ptr[rc] = 0;
+        dqcoeff_ptr[rc] = 0;
+        *eob_ptr = 0;
+      }
+    }
+  }
+#endif
+}
+
+void aom_quantize_b_64x64_adaptive_sse2(
+    const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr,
+    const int16_t *round_ptr, const int16_t *quant_ptr,
+    const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
+    tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
+    const int16_t *scan, const int16_t *iscan) {
+  int index = 16;
+  const int log_scale = 2;
+  int non_zero_count = 0;
+  int non_zero_count_prescan_add_zero = 0;
+  int is_found0 = 0, is_found1 = 0;
+  int eob = -1;
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i one = _mm_set1_epi16(1);
+  const __m128i log_scale_vec = _mm_set1_epi16(log_scale);
+  __m128i zbin, round, quant, dequant, shift;
+  __m128i coeff0, coeff1, coeff0_sign, coeff1_sign;
+  __m128i qcoeff0, qcoeff1;
+  __m128i cmp_mask0, cmp_mask1;
+  __m128i all_zero;
+  __m128i mask0 = zero, mask1 = zero;
+
+  const int zbins[2] = { ROUND_POWER_OF_TWO(zbin_ptr[0], log_scale),
+                         ROUND_POWER_OF_TWO(zbin_ptr[1], log_scale) };
+  int prescan_add[2];
+  int thresh[4];
+  const qm_val_t wt = (1 << AOM_QM_BITS);
+  for (int i = 0; i < 2; ++i) {
+    prescan_add[i] = ROUND_POWER_OF_TWO(dequant_ptr[i] * EOB_FACTOR, 7);
+    thresh[i] = (zbins[i] * wt + prescan_add[i]) - 1;
+  }
+  thresh[2] = thresh[3] = thresh[1];
+  __m128i threshold[2];
+  threshold[0] = _mm_loadu_si128((__m128i *)&thresh[0]);
+  threshold[1] = _mm_unpackhi_epi64(threshold[0], threshold[0]);
+
+#if SKIP_EOB_FACTOR_ADJUST
+  int first = -1;
+#endif
+  // Setup global values.
+  zbin = _mm_load_si128((const __m128i *)zbin_ptr);
+  round = _mm_load_si128((const __m128i *)round_ptr);
+  quant = _mm_load_si128((const __m128i *)quant_ptr);
+  dequant = _mm_load_si128((const __m128i *)dequant_ptr);
+  shift = _mm_load_si128((const __m128i *)quant_shift_ptr);
+
+  // Shift with rounding.
+  zbin = _mm_add_epi16(zbin, log_scale_vec);
+  round = _mm_add_epi16(round, log_scale_vec);
+  zbin = _mm_srli_epi16(zbin, log_scale);
+  round = _mm_srli_epi16(round, log_scale);
+  zbin = _mm_sub_epi16(zbin, one);
+
+  // Do DC and first 15 AC.
+  coeff0 = load_coefficients(coeff_ptr);
+  coeff1 = load_coefficients(coeff_ptr + 8);
+
+  coeff0_sign = _mm_srai_epi16(coeff0, 15);
+  coeff1_sign = _mm_srai_epi16(coeff1, 15);
+  qcoeff0 = invert_sign_sse2(coeff0, coeff0_sign);
+  qcoeff1 = invert_sign_sse2(coeff1, coeff1_sign);
+
+  update_mask0(&qcoeff0, &qcoeff1, threshold, iscan, &is_found0, &mask0);
+
+  cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin);
+  zbin = _mm_unpackhi_epi64(zbin, zbin);  // Switch DC to AC
+  cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);
+
+  update_mask1(&cmp_mask0, &cmp_mask1, iscan, &is_found1, &mask1);
+
+  threshold[0] = threshold[1];
+  all_zero = _mm_or_si128(cmp_mask0, cmp_mask1);
+  if (_mm_movemask_epi8(all_zero) == 0) {
+    _mm_store_si128((__m128i *)(qcoeff_ptr), zero);
+    _mm_store_si128((__m128i *)(qcoeff_ptr + 4), zero);
+    _mm_store_si128((__m128i *)(qcoeff_ptr + 8), zero);
+    _mm_store_si128((__m128i *)(qcoeff_ptr + 12), zero);
+    _mm_store_si128((__m128i *)(dqcoeff_ptr), zero);
+    _mm_store_si128((__m128i *)(dqcoeff_ptr + 4), zero);
+    _mm_store_si128((__m128i *)(dqcoeff_ptr + 8), zero);
+    _mm_store_si128((__m128i *)(dqcoeff_ptr + 12), zero);
+    round = _mm_unpackhi_epi64(round, round);
+    quant = _mm_unpackhi_epi64(quant, quant);
+    shift = _mm_unpackhi_epi64(shift, shift);
+    dequant = _mm_unpackhi_epi64(dequant, dequant);
+  } else {
+    calculate_qcoeff_log_scale(&qcoeff0, round, quant, &shift, &log_scale);
+    round = _mm_unpackhi_epi64(round, round);
+    quant = _mm_unpackhi_epi64(quant, quant);
+    shift = _mm_unpackhi_epi64(shift, shift);
+    calculate_qcoeff_log_scale(&qcoeff1, round, quant, &shift, &log_scale);
+
+    // Reinsert signs
+    qcoeff0 = invert_sign_sse2(qcoeff0, coeff0_sign);
+    qcoeff1 = invert_sign_sse2(qcoeff1, coeff1_sign);
+
+    // Mask out zbin threshold coeffs
+    qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
+    qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
+
+    store_coefficients(qcoeff0, qcoeff_ptr);
+    store_coefficients(qcoeff1, qcoeff_ptr + 8);
+
+    calculate_dqcoeff_and_store_log_scale(qcoeff0, dequant, zero, dqcoeff_ptr,
+                                          &log_scale);
+    dequant = _mm_unpackhi_epi64(dequant, dequant);
+    calculate_dqcoeff_and_store_log_scale(qcoeff1, dequant, zero,
+                                          dqcoeff_ptr + 8, &log_scale);
+  }
+
+  // AC only loop.
+  while (index < n_coeffs) {
+    coeff0 = load_coefficients(coeff_ptr + index);
+    coeff1 = load_coefficients(coeff_ptr + index + 8);
+
+    coeff0_sign = _mm_srai_epi16(coeff0, 15);
+    coeff1_sign = _mm_srai_epi16(coeff1, 15);
+    qcoeff0 = invert_sign_sse2(coeff0, coeff0_sign);
+    qcoeff1 = invert_sign_sse2(coeff1, coeff1_sign);
+
+    update_mask0(&qcoeff0, &qcoeff1, threshold, iscan + index, &is_found0,
+                 &mask0);
+
+    cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin);
+    cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);
+
+    update_mask1(&cmp_mask0, &cmp_mask1, iscan + index, &is_found1, &mask1);
+
+    all_zero = _mm_or_si128(cmp_mask0, cmp_mask1);
+    if (_mm_movemask_epi8(all_zero) == 0) {
+      _mm_store_si128((__m128i *)(qcoeff_ptr + index), zero);
+      _mm_store_si128((__m128i *)(qcoeff_ptr + index + 4), zero);
+      _mm_store_si128((__m128i *)(qcoeff_ptr + index + 8), zero);
+      _mm_store_si128((__m128i *)(qcoeff_ptr + index + 12), zero);
+      _mm_store_si128((__m128i *)(dqcoeff_ptr + index), zero);
+      _mm_store_si128((__m128i *)(dqcoeff_ptr + index + 4), zero);
+      _mm_store_si128((__m128i *)(dqcoeff_ptr + index + 8), zero);
+      _mm_store_si128((__m128i *)(dqcoeff_ptr + index + 12), zero);
+      index += 16;
+      continue;
+    }
+    calculate_qcoeff_log_scale(&qcoeff0, round, quant, &shift, &log_scale);
+    calculate_qcoeff_log_scale(&qcoeff1, round, quant, &shift, &log_scale);
+
+    qcoeff0 = invert_sign_sse2(qcoeff0, coeff0_sign);
+    qcoeff1 = invert_sign_sse2(qcoeff1, coeff1_sign);
+
+    qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
+    qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
+
+    store_coefficients(qcoeff0, qcoeff_ptr + index);
+    store_coefficients(qcoeff1, qcoeff_ptr + index + 8);
+
+    calculate_dqcoeff_and_store_log_scale(qcoeff0, dequant, zero,
+                                          dqcoeff_ptr + index, &log_scale);
+    calculate_dqcoeff_and_store_log_scale(qcoeff1, dequant, zero,
+                                          dqcoeff_ptr + index + 8, &log_scale);
+    index += 16;
+  }
+  if (is_found0) non_zero_count = calculate_non_zero_count(mask0);
+  if (is_found1)
+    non_zero_count_prescan_add_zero = calculate_non_zero_count(mask1);
+
+  for (int i = non_zero_count_prescan_add_zero - 1; i >= non_zero_count; i--) {
+    const int rc = scan[i];
+    qcoeff_ptr[rc] = 0;
+    dqcoeff_ptr[rc] = 0;
+  }
+
+  for (int i = non_zero_count - 1; i >= 0; i--) {
+    const int rc = scan[i];
+    if (qcoeff_ptr[rc]) {
+      eob = i;
+      break;
+    }
+  }
+
+  *eob_ptr = eob + 1;
+#if SKIP_EOB_FACTOR_ADJUST
+  // TODO(Aniket): Experiment the following loop with intrinsic by combining
+  // with the quantization loop above
+  for (int i = 0; i < non_zero_count; i++) {
+    const int rc = scan[i];
+    const int qcoeff = qcoeff_ptr[rc];
+    if (qcoeff) {
+      first = i;
+      break;
+    }
+  }
+  if ((*eob_ptr - 1) >= 0 && first == (*eob_ptr - 1)) {
+    const int rc = scan[(*eob_ptr - 1)];
+    if (qcoeff_ptr[rc] == 1 || qcoeff_ptr[rc] == -1) {
+      const int coeff = coeff_ptr[rc] * wt;
+      const int coeff_sign = AOMSIGN(coeff);
+      const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+      const int factor = EOB_FACTOR + SKIP_EOB_FACTOR_ADJUST;
+      const int prescan_add_val =
+          ROUND_POWER_OF_TWO(dequant_ptr[rc != 0] * factor, 7);
+      if (abs_coeff < (zbins[rc != 0] * (1 << AOM_QM_BITS) + prescan_add_val)) {
+        qcoeff_ptr[rc] = 0;
+        dqcoeff_ptr[rc] = 0;
+        *eob_ptr = 0;
+      }
+    }
+  }
+#endif
+}
diff --git a/third_party/aom/aom_dsp/x86/aom_asm_stubs.c b/third_party/aom/aom_dsp/x86/aom_asm_stubs.c
new file mode 100644
index 0000000000..b08ec2546b
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/aom_asm_stubs.c
@@ -0,0 +1,95 @@
+/*
+ * Copyright (c) 2016, 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 "config/aom_config.h"
+#include "config/aom_dsp_rtcd.h"
+
+#include "aom_dsp/x86/convolve.h"
+
+#if HAVE_SSE2
+filter8_1dfunction aom_filter_block1d16_v8_sse2;
+filter8_1dfunction aom_filter_block1d16_h8_sse2;
+filter8_1dfunction aom_filter_block1d8_v8_sse2;
+filter8_1dfunction aom_filter_block1d8_h8_sse2;
+filter8_1dfunction aom_filter_block1d4_v8_sse2;
+filter8_1dfunction aom_filter_block1d4_h8_sse2;
+filter8_1dfunction aom_filter_block1d16_v4_sse2;
+filter8_1dfunction aom_filter_block1d16_h4_sse2;
+
+filter8_1dfunction aom_filter_block1d8_h4_sse2;
+filter8_1dfunction aom_filter_block1d8_v4_sse2;
+filter8_1dfunction aom_filter_block1d4_h4_sse2;
+filter8_1dfunction aom_filter_block1d4_v4_sse2;
+
+filter8_1dfunction aom_filter_block1d16_v2_sse2;
+filter8_1dfunction aom_filter_block1d16_h2_sse2;
+filter8_1dfunction aom_filter_block1d8_v2_sse2;
+filter8_1dfunction aom_filter_block1d8_h2_sse2;
+filter8_1dfunction aom_filter_block1d4_v2_sse2;
+filter8_1dfunction aom_filter_block1d4_h2_sse2;
+
+// void aom_convolve8_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride,
+//                               uint8_t *dst, ptrdiff_t dst_stride,
+//                               const int16_t *filter_x, int x_step_q4,
+//                               const int16_t *filter_y, int y_step_q4,
+//                               int w, int h);
+// void aom_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride,
+//                              uint8_t *dst, ptrdiff_t dst_stride,
+//                              const int16_t *filter_x, int x_step_q4,
+//                              const int16_t *filter_y, int y_step_q4,
+//                              int w, int h);
+FUN_CONV_1D(horiz, x_step_q4, filter_x, h, src, , sse2)
+FUN_CONV_1D(vert, y_step_q4, filter_y, v, src - src_stride * 3, , sse2)
+
+#if CONFIG_AV1_HIGHBITDEPTH
+highbd_filter8_1dfunction aom_highbd_filter_block1d16_v8_sse2;
+highbd_filter8_1dfunction aom_highbd_filter_block1d16_h8_sse2;
+highbd_filter8_1dfunction aom_highbd_filter_block1d8_v8_sse2;
+highbd_filter8_1dfunction aom_highbd_filter_block1d8_h8_sse2;
+highbd_filter8_1dfunction aom_highbd_filter_block1d4_v8_sse2;
+highbd_filter8_1dfunction aom_highbd_filter_block1d4_h8_sse2;
+
+highbd_filter8_1dfunction aom_highbd_filter_block1d16_v4_sse2;
+highbd_filter8_1dfunction aom_highbd_filter_block1d16_h4_sse2;
+highbd_filter8_1dfunction aom_highbd_filter_block1d8_v4_sse2;
+highbd_filter8_1dfunction aom_highbd_filter_block1d8_h4_sse2;
+highbd_filter8_1dfunction aom_highbd_filter_block1d4_v4_sse2;
+highbd_filter8_1dfunction aom_highbd_filter_block1d4_h4_sse2;
+
+highbd_filter8_1dfunction aom_highbd_filter_block1d16_v2_sse2;
+highbd_filter8_1dfunction aom_highbd_filter_block1d16_h2_sse2;
+highbd_filter8_1dfunction aom_highbd_filter_block1d8_v2_sse2;
+highbd_filter8_1dfunction aom_highbd_filter_block1d8_h2_sse2;
+highbd_filter8_1dfunction aom_highbd_filter_block1d4_v2_sse2;
+highbd_filter8_1dfunction aom_highbd_filter_block1d4_h2_sse2;
+
+// void aom_highbd_convolve8_horiz_sse2(const uint8_t *src,
+//                                      ptrdiff_t src_stride,
+//                                      uint8_t *dst,
+//                                      ptrdiff_t dst_stride,
+//                                      const int16_t *filter_x,
+//                                      int x_step_q4,
+//                                      const int16_t *filter_y,
+//                                      int y_step_q4,
+//                                      int w, int h, int bd);
+// void aom_highbd_convolve8_vert_sse2(const uint8_t *src,
+//                                     ptrdiff_t src_stride,
+//                                     uint8_t *dst,
+//                                     ptrdiff_t dst_stride,
+//                                     const int16_t *filter_x,
+//                                     int x_step_q4,
+//                                     const int16_t *filter_y,
+//                                     int y_step_q4,
+//                                     int w, int h, int bd);
+HIGH_FUN_CONV_1D(horiz, x_step_q4, filter_x, h, src, , sse2)
+HIGH_FUN_CONV_1D(vert, y_step_q4, filter_y, v, src - src_stride * 3, , sse2)
+#endif
+#endif  // HAVE_SSE2
diff --git a/third_party/aom/aom_dsp/x86/aom_convolve_copy_avx2.c b/third_party/aom/aom_dsp/x86/aom_convolve_copy_avx2.c
new file mode 100644
index 0000000000..a1043828fe
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/aom_convolve_copy_avx2.c
@@ -0,0 +1,256 @@
+/*
+ *  Copyright (c) 2020, Alliance for Open Media. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <immintrin.h>
+
+#include "config/aom_dsp_rtcd.h"
+
+static INLINE void copy_128(const uint8_t *src, uint8_t *dst) {
+  __m256i s[4];
+  s[0] = _mm256_loadu_si256((__m256i *)(src + 0 * 32));
+  s[1] = _mm256_loadu_si256((__m256i *)(src + 1 * 32));
+  s[2] = _mm256_loadu_si256((__m256i *)(src + 2 * 32));
+  s[3] = _mm256_loadu_si256((__m256i *)(src + 3 * 32));
+  _mm256_storeu_si256((__m256i *)(dst + 0 * 32), s[0]);
+  _mm256_storeu_si256((__m256i *)(dst + 1 * 32), s[1]);
+  _mm256_storeu_si256((__m256i *)(dst + 2 * 32), s[2]);
+  _mm256_storeu_si256((__m256i *)(dst + 3 * 32), s[3]);
+}
+
+void aom_convolve_copy_avx2(const uint8_t *src, ptrdiff_t src_stride,
+                            uint8_t *dst, ptrdiff_t dst_stride, int w, int h) {
+  if (w >= 16) {
+    assert(!((intptr_t)dst % 16));
+    assert(!(dst_stride % 16));
+  }
+
+  if (w == 2) {
+    do {
+      memmove(dst, src, 2 * sizeof(*src));
+      src += src_stride;
+      dst += dst_stride;
+      memmove(dst, src, 2 * sizeof(*src));
+      src += src_stride;
+      dst += dst_stride;
+      h -= 2;
+    } while (h);
+  } else if (w == 4) {
+    do {
+      memmove(dst, src, 4 * sizeof(*src));
+      src += src_stride;
+      dst += dst_stride;
+      memmove(dst, src, 4 * sizeof(*src));
+      src += src_stride;
+      dst += dst_stride;
+      h -= 2;
+    } while (h);
+  } else if (w == 8) {
+    do {
+      __m128i s[2];
+      s[0] = _mm_loadl_epi64((__m128i *)src);
+      src += src_stride;
+      s[1] = _mm_loadl_epi64((__m128i *)src);
+      src += src_stride;
+      _mm_storel_epi64((__m128i *)dst, s[0]);
+      dst += dst_stride;
+      _mm_storel_epi64((__m128i *)dst, s[1]);
+      dst += dst_stride;
+      h -= 2;
+    } while (h);
+  } else if (w == 16) {
+    do {
+      __m128i s[2];
+      s[0] = _mm_loadu_si128((__m128i *)src);
+      src += src_stride;
+      s[1] = _mm_loadu_si128((__m128i *)src);
+      src += src_stride;
+      _mm_store_si128((__m128i *)dst, s[0]);
+      dst += dst_stride;
+      _mm_store_si128((__m128i *)dst, s[1]);
+      dst += dst_stride;
+      h -= 2;
+    } while (h);
+  } else if (w == 32) {
+    do {
+      __m256i s[2];
+      s[0] = _mm256_loadu_si256((__m256i *)src);
+      src += src_stride;
+      s[1] = _mm256_loadu_si256((__m256i *)src);
+      src += src_stride;
+      _mm256_storeu_si256((__m256i *)dst, s[0]);
+      dst += dst_stride;
+      _mm256_storeu_si256((__m256i *)dst, s[1]);
+      dst += dst_stride;
+      h -= 2;
+    } while (h);
+  } else if (w == 64) {
+    do {
+      __m256i s[4];
+      s[0] = _mm256_loadu_si256((__m256i *)(src + 0 * 32));
+      s[1] = _mm256_loadu_si256((__m256i *)(src + 1 * 32));
+      src += src_stride;
+      s[2] = _mm256_loadu_si256((__m256i *)(src + 0 * 32));
+      s[3] = _mm256_loadu_si256((__m256i *)(src + 1 * 32));
+      src += src_stride;
+      _mm256_storeu_si256((__m256i *)(dst + 0 * 32), s[0]);
+      _mm256_storeu_si256((__m256i *)(dst + 1 * 32), s[1]);
+      dst += dst_stride;
+      _mm256_storeu_si256((__m256i *)(dst + 0 * 32), s[2]);
+      _mm256_storeu_si256((__m256i *)(dst + 1 * 32), s[3]);
+      dst += dst_stride;
+      h -= 2;
+    } while (h);
+  } else {
+    do {
+      copy_128(src, dst);
+      src += src_stride;
+      dst += dst_stride;
+      copy_128(src, dst);
+      src += src_stride;
+      dst += dst_stride;
+      h -= 2;
+    } while (h);
+  }
+}
+
+#if CONFIG_AV1_HIGHBITDEPTH
+
+static INLINE void highbd_copy_64(const uint16_t *src, uint16_t *dst) {
+  __m256i s[4];
+  s[0] = _mm256_loadu_si256((__m256i *)(src + 0 * 16));
+  s[1] = _mm256_loadu_si256((__m256i *)(src + 1 * 16));
+  s[2] = _mm256_loadu_si256((__m256i *)(src + 2 * 16));
+  s[3] = _mm256_loadu_si256((__m256i *)(src + 3 * 16));
+  _mm256_storeu_si256((__m256i *)(dst + 0 * 16), s[0]);
+  _mm256_storeu_si256((__m256i *)(dst + 1 * 16), s[1]);
+  _mm256_storeu_si256((__m256i *)(dst + 2 * 16), s[2]);
+  _mm256_storeu_si256((__m256i *)(dst + 3 * 16), s[3]);
+}
+
+static INLINE void highbd_copy_128(const uint16_t *src, uint16_t *dst) {
+  __m256i s[8];
+  s[0] = _mm256_loadu_si256((__m256i *)(src + 0 * 16));
+  s[1] = _mm256_loadu_si256((__m256i *)(src + 1 * 16));
+  s[2] = _mm256_loadu_si256((__m256i *)(src + 2 * 16));
+  s[3] = _mm256_loadu_si256((__m256i *)(src + 3 * 16));
+  s[4] = _mm256_loadu_si256((__m256i *)(src + 4 * 16));
+  s[5] = _mm256_loadu_si256((__m256i *)(src + 5 * 16));
+  s[6] = _mm256_loadu_si256((__m256i *)(src + 6 * 16));
+  s[7] = _mm256_loadu_si256((__m256i *)(src + 7 * 16));
+
+  _mm256_storeu_si256((__m256i *)(dst + 0 * 16), s[0]);
+  _mm256_storeu_si256((__m256i *)(dst + 1 * 16), s[1]);
+  _mm256_storeu_si256((__m256i *)(dst + 2 * 16), s[2]);
+  _mm256_storeu_si256((__m256i *)(dst + 3 * 16), s[3]);
+  _mm256_storeu_si256((__m256i *)(dst + 4 * 16), s[4]);
+  _mm256_storeu_si256((__m256i *)(dst + 5 * 16), s[5]);
+  _mm256_storeu_si256((__m256i *)(dst + 6 * 16), s[6]);
+  _mm256_storeu_si256((__m256i *)(dst + 7 * 16), s[7]);
+}
+
+void aom_highbd_convolve_copy_avx2(const uint16_t *src, ptrdiff_t src_stride,
+                                   uint16_t *dst, ptrdiff_t dst_stride, int w,
+                                   int h) {
+  if (w >= 16) {
+    assert(!((intptr_t)dst % 16));
+    assert(!(dst_stride % 16));
+  }
+
+  if (w == 2) {
+    do {
+      memmove(dst, src, 2 * sizeof(*src));
+      src += src_stride;
+      dst += dst_stride;
+      memmove(dst, src, 2 * sizeof(*src));
+      src += src_stride;
+      dst += dst_stride;
+      h -= 2;
+    } while (h);
+  } else if (w == 4) {
+    do {
+      __m128i s[2];
+      s[0] = _mm_loadl_epi64((__m128i *)src);
+      src += src_stride;
+      s[1] = _mm_loadl_epi64((__m128i *)src);
+      src += src_stride;
+      _mm_storel_epi64((__m128i *)dst, s[0]);
+      dst += dst_stride;
+      _mm_storel_epi64((__m128i *)dst, s[1]);
+      dst += dst_stride;
+      h -= 2;
+    } while (h);
+  } else if (w == 8) {
+    do {
+      __m128i s[2];
+      s[0] = _mm_loadu_si128((__m128i *)src);
+      src += src_stride;
+      s[1] = _mm_loadu_si128((__m128i *)src);
+      src += src_stride;
+      _mm_store_si128((__m128i *)dst, s[0]);
+      dst += dst_stride;
+      _mm_store_si128((__m128i *)dst, s[1]);
+      dst += dst_stride;
+      h -= 2;
+    } while (h);
+  } else if (w == 16) {
+    do {
+      __m256i s[2];
+      s[0] = _mm256_loadu_si256((__m256i *)src);
+      src += src_stride;
+      s[1] = _mm256_loadu_si256((__m256i *)src);
+      src += src_stride;
+      _mm256_storeu_si256((__m256i *)dst, s[0]);
+      dst += dst_stride;
+      _mm256_storeu_si256((__m256i *)dst, s[1]);
+      dst += dst_stride;
+      h -= 2;
+    } while (h);
+  } else if (w == 32) {
+    do {
+      __m256i s[4];
+      s[0] = _mm256_loadu_si256((__m256i *)(src + 0 * 16));
+      s[1] = _mm256_loadu_si256((__m256i *)(src + 1 * 16));
+      src += src_stride;
+      s[2] = _mm256_loadu_si256((__m256i *)(src + 0 * 16));
+      s[3] = _mm256_loadu_si256((__m256i *)(src + 1 * 16));
+      src += src_stride;
+      _mm256_storeu_si256((__m256i *)(dst + 0 * 16), s[0]);
+      _mm256_storeu_si256((__m256i *)(dst + 1 * 16), s[1]);
+      dst += dst_stride;
+      _mm256_storeu_si256((__m256i *)(dst + 0 * 16), s[2]);
+      _mm256_storeu_si256((__m256i *)(dst + 1 * 16), s[3]);
+      dst += dst_stride;
+      h -= 2;
+    } while (h);
+  } else if (w == 64) {
+    do {
+      highbd_copy_64(src, dst);
+      src += src_stride;
+      dst += dst_stride;
+      highbd_copy_64(src, dst);
+      src += src_stride;
+      dst += dst_stride;
+      h -= 2;
+    } while (h);
+  } else {
+    assert(w == 128);
+    do {
+      highbd_copy_128(src, dst);
+      src += src_stride;
+      dst += dst_stride;
+      highbd_copy_128(src, dst);
+      src += src_stride;
+      dst += dst_stride;
+      h -= 2;
+    } while (h);
+  }
+}
+
+#endif  // CONFIG_AV1_HIGHBITDEPTH
diff --git a/third_party/aom/aom_dsp/x86/aom_convolve_copy_sse2.c b/third_party/aom/aom_dsp/x86/aom_convolve_copy_sse2.c
new file mode 100644
index 0000000000..e78845e97c
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/aom_convolve_copy_sse2.c
@@ -0,0 +1,308 @@
+/*
+ *  Copyright (c) 2020, Alliance for Open Media. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <immintrin.h>
+
+#include "config/aom_dsp_rtcd.h"
+
+static INLINE void copy_128(const uint8_t *src, uint8_t *dst) {
+  __m128i s[8];
+  s[0] = _mm_loadu_si128((__m128i *)(src + 0 * 16));
+  s[1] = _mm_loadu_si128((__m128i *)(src + 1 * 16));
+  s[2] = _mm_loadu_si128((__m128i *)(src + 2 * 16));
+  s[3] = _mm_loadu_si128((__m128i *)(src + 3 * 16));
+  s[4] = _mm_loadu_si128((__m128i *)(src + 4 * 16));
+  s[5] = _mm_loadu_si128((__m128i *)(src + 5 * 16));
+  s[6] = _mm_loadu_si128((__m128i *)(src + 6 * 16));
+  s[7] = _mm_loadu_si128((__m128i *)(src + 7 * 16));
+  _mm_store_si128((__m128i *)(dst + 0 * 16), s[0]);
+  _mm_store_si128((__m128i *)(dst + 1 * 16), s[1]);
+  _mm_store_si128((__m128i *)(dst + 2 * 16), s[2]);
+  _mm_store_si128((__m128i *)(dst + 3 * 16), s[3]);
+  _mm_store_si128((__m128i *)(dst + 4 * 16), s[4]);
+  _mm_store_si128((__m128i *)(dst + 5 * 16), s[5]);
+  _mm_store_si128((__m128i *)(dst + 6 * 16), s[6]);
+  _mm_store_si128((__m128i *)(dst + 7 * 16), s[7]);
+}
+
+void aom_convolve_copy_sse2(const uint8_t *src, ptrdiff_t src_stride,
+                            uint8_t *dst, ptrdiff_t dst_stride, int w, int h) {
+  if (w >= 16) {
+    assert(!((intptr_t)dst % 16));
+    assert(!(dst_stride % 16));
+  }
+
+  if (w == 2) {
+    do {
+      memmove(dst, src, 2 * sizeof(*src));
+      src += src_stride;
+      dst += dst_stride;
+      memmove(dst, src, 2 * sizeof(*src));
+      src += src_stride;
+      dst += dst_stride;
+      h -= 2;
+    } while (h);
+  } else if (w == 4) {
+    do {
+      memmove(dst, src, 4 * sizeof(*src));
+      src += src_stride;
+      dst += dst_stride;
+      memmove(dst, src, 4 * sizeof(*src));
+      src += src_stride;
+      dst += dst_stride;
+      h -= 2;
+    } while (h);
+  } else if (w == 8) {
+    do {
+      __m128i s[2];
+      s[0] = _mm_loadl_epi64((__m128i *)src);
+      src += src_stride;
+      s[1] = _mm_loadl_epi64((__m128i *)src);
+      src += src_stride;
+      _mm_storel_epi64((__m128i *)dst, s[0]);
+      dst += dst_stride;
+      _mm_storel_epi64((__m128i *)dst, s[1]);
+      dst += dst_stride;
+      h -= 2;
+    } while (h);
+  } else if (w == 16) {
+    do {
+      __m128i s[2];
+      s[0] = _mm_loadu_si128((__m128i *)src);
+      src += src_stride;
+      s[1] = _mm_loadu_si128((__m128i *)src);
+      src += src_stride;
+      _mm_store_si128((__m128i *)dst, s[0]);
+      dst += dst_stride;
+      _mm_store_si128((__m128i *)dst, s[1]);
+      dst += dst_stride;
+      h -= 2;
+    } while (h);
+  } else if (w == 32) {
+    do {
+      __m128i s[4];
+      s[0] = _mm_loadu_si128((__m128i *)(src + 0 * 16));
+      s[1] = _mm_loadu_si128((__m128i *)(src + 1 * 16));
+      src += src_stride;
+      s[2] = _mm_loadu_si128((__m128i *)(src + 0 * 16));
+      s[3] = _mm_loadu_si128((__m128i *)(src + 1 * 16));
+      src += src_stride;
+      _mm_store_si128((__m128i *)(dst + 0 * 16), s[0]);
+      _mm_store_si128((__m128i *)(dst + 1 * 16), s[1]);
+      dst += dst_stride;
+      _mm_store_si128((__m128i *)(dst + 0 * 16), s[2]);
+      _mm_store_si128((__m128i *)(dst + 1 * 16), s[3]);
+      dst += dst_stride;
+      h -= 2;
+    } while (h);
+  } else if (w == 64) {
+    do {
+      __m128i s[8];
+      s[0] = _mm_loadu_si128((__m128i *)(src + 0 * 16));
+      s[1] = _mm_loadu_si128((__m128i *)(src + 1 * 16));
+      s[2] = _mm_loadu_si128((__m128i *)(src + 2 * 16));
+      s[3] = _mm_loadu_si128((__m128i *)(src + 3 * 16));
+      src += src_stride;
+      s[4] = _mm_loadu_si128((__m128i *)(src + 0 * 16));
+      s[5] = _mm_loadu_si128((__m128i *)(src + 1 * 16));
+      s[6] = _mm_loadu_si128((__m128i *)(src + 2 * 16));
+      s[7] = _mm_loadu_si128((__m128i *)(src + 3 * 16));
+      src += src_stride;
+      _mm_store_si128((__m128i *)(dst + 0 * 16), s[0]);
+      _mm_store_si128((__m128i *)(dst + 1 * 16), s[1]);
+      _mm_store_si128((__m128i *)(dst + 2 * 16), s[2]);
+      _mm_store_si128((__m128i *)(dst + 3 * 16), s[3]);
+      dst += dst_stride;
+      _mm_store_si128((__m128i *)(dst + 0 * 16), s[4]);
+      _mm_store_si128((__m128i *)(dst + 1 * 16), s[5]);
+      _mm_store_si128((__m128i *)(dst + 2 * 16), s[6]);
+      _mm_store_si128((__m128i *)(dst + 3 * 16), s[7]);
+      dst += dst_stride;
+      h -= 2;
+    } while (h);
+  } else {
+    do {
+      copy_128(src, dst);
+      src += src_stride;
+      dst += dst_stride;
+      copy_128(src, dst);
+      src += src_stride;
+      dst += dst_stride;
+      h -= 2;
+    } while (h);
+  }
+}
+
+static INLINE void highbd_copy_64(const uint16_t *src, uint16_t *dst) {
+  __m128i s[8];
+  s[0] = _mm_loadu_si128((__m128i *)(src + 0 * 8));
+  s[1] = _mm_loadu_si128((__m128i *)(src + 1 * 8));
+  s[2] = _mm_loadu_si128((__m128i *)(src + 2 * 8));
+  s[3] = _mm_loadu_si128((__m128i *)(src + 3 * 8));
+  s[4] = _mm_loadu_si128((__m128i *)(src + 4 * 8));
+  s[5] = _mm_loadu_si128((__m128i *)(src + 5 * 8));
+  s[6] = _mm_loadu_si128((__m128i *)(src + 6 * 8));
+  s[7] = _mm_loadu_si128((__m128i *)(src + 7 * 8));
+  _mm_store_si128((__m128i *)(dst + 0 * 8), s[0]);
+  _mm_store_si128((__m128i *)(dst + 1 * 8), s[1]);
+  _mm_store_si128((__m128i *)(dst + 2 * 8), s[2]);
+  _mm_store_si128((__m128i *)(dst + 3 * 8), s[3]);
+  _mm_store_si128((__m128i *)(dst + 4 * 8), s[4]);
+  _mm_store_si128((__m128i *)(dst + 5 * 8), s[5]);
+  _mm_store_si128((__m128i *)(dst + 6 * 8), s[6]);
+  _mm_store_si128((__m128i *)(dst + 7 * 8), s[7]);
+}
+
+static INLINE void highbd_copy_128(const uint16_t *src, uint16_t *dst) {
+  __m128i s[16];
+  s[0] = _mm_loadu_si128((__m128i *)(src + 0 * 8));
+  s[1] = _mm_loadu_si128((__m128i *)(src + 1 * 8));
+  s[2] = _mm_loadu_si128((__m128i *)(src + 2 * 8));
+  s[3] = _mm_loadu_si128((__m128i *)(src + 3 * 8));
+  s[4] = _mm_loadu_si128((__m128i *)(src + 4 * 8));
+  s[5] = _mm_loadu_si128((__m128i *)(src + 5 * 8));
+  s[6] = _mm_loadu_si128((__m128i *)(src + 6 * 8));
+  s[7] = _mm_loadu_si128((__m128i *)(src + 7 * 8));
+  s[8] = _mm_loadu_si128((__m128i *)(src + 8 * 8));
+  s[9] = _mm_loadu_si128((__m128i *)(src + 9 * 8));
+  s[10] = _mm_loadu_si128((__m128i *)(src + 10 * 8));
+  s[11] = _mm_loadu_si128((__m128i *)(src + 11 * 8));
+  s[12] = _mm_loadu_si128((__m128i *)(src + 12 * 8));
+  s[13] = _mm_loadu_si128((__m128i *)(src + 13 * 8));
+  s[14] = _mm_loadu_si128((__m128i *)(src + 14 * 8));
+  s[15] = _mm_loadu_si128((__m128i *)(src + 15 * 8));
+  _mm_store_si128((__m128i *)(dst + 0 * 8), s[0]);
+  _mm_store_si128((__m128i *)(dst + 1 * 8), s[1]);
+  _mm_store_si128((__m128i *)(dst + 2 * 8), s[2]);
+  _mm_store_si128((__m128i *)(dst + 3 * 8), s[3]);
+  _mm_store_si128((__m128i *)(dst + 4 * 8), s[4]);
+  _mm_store_si128((__m128i *)(dst + 5 * 8), s[5]);
+  _mm_store_si128((__m128i *)(dst + 6 * 8), s[6]);
+  _mm_store_si128((__m128i *)(dst + 7 * 8), s[7]);
+  _mm_store_si128((__m128i *)(dst + 8 * 8), s[8]);
+  _mm_store_si128((__m128i *)(dst + 9 * 8), s[9]);
+  _mm_store_si128((__m128i *)(dst + 10 * 8), s[10]);
+  _mm_store_si128((__m128i *)(dst + 11 * 8), s[11]);
+  _mm_store_si128((__m128i *)(dst + 12 * 8), s[12]);
+  _mm_store_si128((__m128i *)(dst + 13 * 8), s[13]);
+  _mm_store_si128((__m128i *)(dst + 14 * 8), s[14]);
+  _mm_store_si128((__m128i *)(dst + 15 * 8), s[15]);
+}
+
+void aom_highbd_convolve_copy_sse2(const uint16_t *src, ptrdiff_t src_stride,
+                                   uint16_t *dst, ptrdiff_t dst_stride, int w,
+                                   int h) {
+  if (w >= 16) {
+    assert(!((intptr_t)dst % 16));
+    assert(!(dst_stride % 16));
+  }
+
+  if (w == 2) {
+    do {
+      __m128i s = _mm_loadl_epi64((__m128i *)src);
+      *(int *)dst = _mm_cvtsi128_si32(s);
+      src += src_stride;
+      dst += dst_stride;
+      s = _mm_loadl_epi64((__m128i *)src);
+      *(int *)dst = _mm_cvtsi128_si32(s);
+      src += src_stride;
+      dst += dst_stride;
+      h -= 2;
+    } while (h);
+  } else if (w == 4) {
+    do {
+      __m128i s[2];
+      s[0] = _mm_loadl_epi64((__m128i *)src);
+      src += src_stride;
+      s[1] = _mm_loadl_epi64((__m128i *)src);
+      src += src_stride;
+      _mm_storel_epi64((__m128i *)dst, s[0]);
+      dst += dst_stride;
+      _mm_storel_epi64((__m128i *)dst, s[1]);
+      dst += dst_stride;
+      h -= 2;
+    } while (h);
+  } else if (w == 8) {
+    do {
+      __m128i s[2];
+      s[0] = _mm_loadu_si128((__m128i *)src);
+      src += src_stride;
+      s[1] = _mm_loadu_si128((__m128i *)src);
+      src += src_stride;
+      _mm_store_si128((__m128i *)dst, s[0]);
+      dst += dst_stride;
+      _mm_store_si128((__m128i *)dst, s[1]);
+      dst += dst_stride;
+      h -= 2;
+    } while (h);
+  } else if (w == 16) {
+    do {
+      __m128i s[4];
+      s[0] = _mm_loadu_si128((__m128i *)(src + 0 * 8));
+      s[1] = _mm_loadu_si128((__m128i *)(src + 1 * 8));
+      src += src_stride;
+      s[2] = _mm_loadu_si128((__m128i *)(src + 0 * 8));
+      s[3] = _mm_loadu_si128((__m128i *)(src + 1 * 8));
+      src += src_stride;
+      _mm_store_si128((__m128i *)(dst + 0 * 8), s[0]);
+      _mm_store_si128((__m128i *)(dst + 1 * 8), s[1]);
+      dst += dst_stride;
+      _mm_store_si128((__m128i *)(dst + 0 * 8), s[2]);
+      _mm_store_si128((__m128i *)(dst + 1 * 8), s[3]);
+      dst += dst_stride;
+      h -= 2;
+    } while (h);
+  } else if (w == 32) {
+    do {
+      __m128i s[8];
+      s[0] = _mm_loadu_si128((__m128i *)(src + 0 * 8));
+      s[1] = _mm_loadu_si128((__m128i *)(src + 1 * 8));
+      s[2] = _mm_loadu_si128((__m128i *)(src + 2 * 8));
+      s[3] = _mm_loadu_si128((__m128i *)(src + 3 * 8));
+      src += src_stride;
+      s[4] = _mm_loadu_si128((__m128i *)(src + 0 * 8));
+      s[5] = _mm_loadu_si128((__m128i *)(src + 1 * 8));
+      s[6] = _mm_loadu_si128((__m128i *)(src + 2 * 8));
+      s[7] = _mm_loadu_si128((__m128i *)(src + 3 * 8));
+      src += src_stride;
+      _mm_store_si128((__m128i *)(dst + 0 * 8), s[0]);
+      _mm_store_si128((__m128i *)(dst + 1 * 8), s[1]);
+      _mm_store_si128((__m128i *)(dst + 2 * 8), s[2]);
+      _mm_store_si128((__m128i *)(dst + 3 * 8), s[3]);
+      dst += dst_stride;
+      _mm_store_si128((__m128i *)(dst + 0 * 8), s[4]);
+      _mm_store_si128((__m128i *)(dst + 1 * 8), s[5]);
+      _mm_store_si128((__m128i *)(dst + 2 * 8), s[6]);
+      _mm_store_si128((__m128i *)(dst + 3 * 8), s[7]);
+      dst += dst_stride;
+      h -= 2;
+    } while (h);
+  } else if (w == 64) {
+    do {
+      highbd_copy_64(src, dst);
+      src += src_stride;
+      dst += dst_stride;
+      highbd_copy_64(src, dst);
+      src += src_stride;
+      dst += dst_stride;
+      h -= 2;
+    } while (h);
+  } else {
+    do {
+      highbd_copy_128(src, dst);
+      src += src_stride;
+      dst += dst_stride;
+      highbd_copy_128(src, dst);
+      src += src_stride;
+      dst += dst_stride;
+      h -= 2;
+    } while (h);
+  }
+}
diff --git a/third_party/aom/aom_dsp/x86/aom_high_subpixel_8t_sse2.asm b/third_party/aom/aom_dsp/x86/aom_high_subpixel_8t_sse2.asm
new file mode 100644
index 0000000000..d392225906
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/aom_high_subpixel_8t_sse2.asm
@@ -0,0 +1,613 @@
+;
+; Copyright (c) 2016, 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 "aom_ports/x86_abi_support.asm"
+
+;Note: tap3 and tap4 have to be applied and added after other taps to avoid
+;overflow.
+
+%macro HIGH_GET_FILTERS_4 0
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rcx, 0x00000040
+
+    movdqa      xmm7, [rdx]                 ;load filters
+    pshuflw     xmm0, xmm7, 0b              ;k0
+    pshuflw     xmm1, xmm7, 01010101b       ;k1
+    pshuflw     xmm2, xmm7, 10101010b       ;k2
+    pshuflw     xmm3, xmm7, 11111111b       ;k3
+    psrldq      xmm7, 8
+    pshuflw     xmm4, xmm7, 0b              ;k4
+    pshuflw     xmm5, xmm7, 01010101b       ;k5
+    pshuflw     xmm6, xmm7, 10101010b       ;k6
+    pshuflw     xmm7, xmm7, 11111111b       ;k7
+
+    punpcklwd   xmm0, xmm6
+    punpcklwd   xmm2, xmm5
+    punpcklwd   xmm3, xmm4
+    punpcklwd   xmm1, xmm7
+
+    movdqa      k0k6, xmm0
+    movdqa      k2k5, xmm2
+    movdqa      k3k4, xmm3
+    movdqa      k1k7, xmm1
+
+    movq        xmm6, rcx
+    pshufd      xmm6, xmm6, 0
+    movdqa      krd, xmm6
+
+    ;Compute max and min values of a pixel
+    mov         rdx, 0x00010001
+    movsxd      rcx, DWORD PTR arg(6)      ;bps
+    movq        xmm0, rdx
+    movq        xmm1, rcx
+    pshufd      xmm0, xmm0, 0b
+    movdqa      xmm2, xmm0
+    psllw       xmm0, xmm1
+    psubw       xmm0, xmm2
+    pxor        xmm1, xmm1
+    movdqa      max, xmm0                  ;max value (for clamping)
+    movdqa      min, xmm1                  ;min value (for clamping)
+
+%endm
+
+%macro HIGH_APPLY_FILTER_4 1
+    punpcklwd   xmm0, xmm6                  ;two row in one register
+    punpcklwd   xmm1, xmm7
+    punpcklwd   xmm2, xmm5
+    punpcklwd   xmm3, xmm4
+
+    pmaddwd     xmm0, k0k6                  ;multiply the filter factors
+    pmaddwd     xmm1, k1k7
+    pmaddwd     xmm2, k2k5
+    pmaddwd     xmm3, k3k4
+
+    paddd       xmm0, xmm1                  ;sum
+    paddd       xmm0, xmm2
+    paddd       xmm0, xmm3
+
+    paddd       xmm0, krd                   ;rounding
+    psrad       xmm0, 7                     ;shift
+    packssdw    xmm0, xmm0                  ;pack to word
+
+    ;clamp the values
+    pminsw      xmm0, max
+    pmaxsw      xmm0, min
+
+%if %1
+    movq        xmm1, [rdi]
+    pavgw       xmm0, xmm1
+%endif
+    movq        [rdi], xmm0
+%endm
+
+%macro HIGH_GET_FILTERS 0
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+    mov         rcx, 0x00000040
+
+    movdqa      xmm7, [rdx]                 ;load filters
+    pshuflw     xmm0, xmm7, 0b              ;k0
+    pshuflw     xmm1, xmm7, 01010101b       ;k1
+    pshuflw     xmm2, xmm7, 10101010b       ;k2
+    pshuflw     xmm3, xmm7, 11111111b       ;k3
+    pshufhw     xmm4, xmm7, 0b              ;k4
+    pshufhw     xmm5, xmm7, 01010101b       ;k5
+    pshufhw     xmm6, xmm7, 10101010b       ;k6
+    pshufhw     xmm7, xmm7, 11111111b       ;k7
+    punpcklqdq  xmm2, xmm2
+    punpcklqdq  xmm3, xmm3
+    punpcklwd   xmm0, xmm1
+    punpckhwd   xmm6, xmm7
+    punpckhwd   xmm2, xmm5
+    punpckhwd   xmm3, xmm4
+
+    movdqa      k0k1, xmm0                  ;store filter factors on stack
+    movdqa      k6k7, xmm6
+    movdqa      k2k5, xmm2
+    movdqa      k3k4, xmm3
+
+    movq        xmm6, rcx
+    pshufd      xmm6, xmm6, 0
+    movdqa      krd, xmm6                   ;rounding
+
+    ;Compute max and min values of a pixel
+    mov         rdx, 0x00010001
+    movsxd      rcx, DWORD PTR arg(6)       ;bps
+    movq        xmm0, rdx
+    movq        xmm1, rcx
+    pshufd      xmm0, xmm0, 0b
+    movdqa      xmm2, xmm0
+    psllw       xmm0, xmm1
+    psubw       xmm0, xmm2
+    pxor        xmm1, xmm1
+    movdqa      max, xmm0                  ;max value (for clamping)
+    movdqa      min, xmm1                  ;min value (for clamping)
+%endm
+
+%macro LOAD_VERT_8 1
+    movdqu      xmm0, [rsi + %1]            ;0
+    movdqu      xmm1, [rsi + rax + %1]      ;1
+    movdqu      xmm6, [rsi + rdx * 2 + %1]  ;6
+    lea         rsi,  [rsi + rax]
+    movdqu      xmm7, [rsi + rdx * 2 + %1]  ;7
+    movdqu      xmm2, [rsi + rax + %1]      ;2
+    movdqu      xmm3, [rsi + rax * 2 + %1]  ;3
+    movdqu      xmm4, [rsi + rdx + %1]      ;4
+    movdqu      xmm5, [rsi + rax * 4 + %1]  ;5
+%endm
+
+%macro HIGH_APPLY_FILTER_8 2
+    movdqu      temp, xmm4
+    movdqa      xmm4, xmm0
+    punpcklwd   xmm0, xmm1
+    punpckhwd   xmm4, xmm1
+    movdqa      xmm1, xmm6
+    punpcklwd   xmm6, xmm7
+    punpckhwd   xmm1, xmm7
+    movdqa      xmm7, xmm2
+    punpcklwd   xmm2, xmm5
+    punpckhwd   xmm7, xmm5
+
+    movdqu      xmm5, temp
+    movdqu      temp, xmm4
+    movdqa      xmm4, xmm3
+    punpcklwd   xmm3, xmm5
+    punpckhwd   xmm4, xmm5
+    movdqu      xmm5, temp
+
+    pmaddwd     xmm0, k0k1
+    pmaddwd     xmm5, k0k1
+    pmaddwd     xmm6, k6k7
+    pmaddwd     xmm1, k6k7
+    pmaddwd     xmm2, k2k5
+    pmaddwd     xmm7, k2k5
+    pmaddwd     xmm3, k3k4
+    pmaddwd     xmm4, k3k4
+
+    paddd       xmm0, xmm6
+    paddd       xmm0, xmm2
+    paddd       xmm0, xmm3
+    paddd       xmm5, xmm1
+    paddd       xmm5, xmm7
+    paddd       xmm5, xmm4
+
+    paddd       xmm0, krd                   ;rounding
+    paddd       xmm5, krd
+    psrad       xmm0, 7                     ;shift
+    psrad       xmm5, 7
+    packssdw    xmm0, xmm5                  ;pack back to word
+
+    ;clamp the values
+    pminsw      xmm0, max
+    pmaxsw      xmm0, min
+
+%if %1
+    movdqu      xmm1, [rdi + %2]
+    pavgw       xmm0, xmm1
+%endif
+    movdqu      [rdi + %2], xmm0
+%endm
+
+SECTION .text
+
+;void aom_filter_block1d4_v8_sse2
+;(
+;    unsigned char *src_ptr,
+;    unsigned int   src_pitch,
+;    unsigned char *output_ptr,
+;    unsigned int   out_pitch,
+;    unsigned int   output_height,
+;    short *filter
+;)
+globalsym(aom_highbd_filter_block1d4_v8_sse2)
+sym(aom_highbd_filter_block1d4_v8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 7
+    %define k0k6 [rsp + 16 * 0]
+    %define k2k5 [rsp + 16 * 1]
+    %define k3k4 [rsp + 16 * 2]
+    %define k1k7 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define max [rsp + 16 * 5]
+    %define min [rsp + 16 * 6]
+
+    HIGH_GET_FILTERS_4
+
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rbx, DWORD PTR arg(3)       ;out_pitch
+    lea         rax, [rax + rax]            ;bytes per line
+    lea         rbx, [rbx + rbx]
+    lea         rdx, [rax + rax * 2]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movq        xmm0, [rsi]                 ;load src: row 0
+    movq        xmm1, [rsi + rax]           ;1
+    movq        xmm6, [rsi + rdx * 2]       ;6
+    lea         rsi,  [rsi + rax]
+    movq        xmm7, [rsi + rdx * 2]       ;7
+    movq        xmm2, [rsi + rax]           ;2
+    movq        xmm3, [rsi + rax * 2]       ;3
+    movq        xmm4, [rsi + rdx]           ;4
+    movq        xmm5, [rsi + rax * 4]       ;5
+
+    HIGH_APPLY_FILTER_4 0
+
+    lea         rdi, [rdi + rbx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 7
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void aom_filter_block1d8_v8_sse2
+;(
+;    unsigned char *src_ptr,
+;    unsigned int   src_pitch,
+;    unsigned char *output_ptr,
+;    unsigned int   out_pitch,
+;    unsigned int   output_height,
+;    short *filter
+;)
+globalsym(aom_highbd_filter_block1d8_v8_sse2)
+sym(aom_highbd_filter_block1d8_v8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 8
+    %define k0k1 [rsp + 16 * 0]
+    %define k6k7 [rsp + 16 * 1]
+    %define k2k5 [rsp + 16 * 2]
+    %define k3k4 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define temp [rsp + 16 * 5]
+    %define max [rsp + 16 * 6]
+    %define min [rsp + 16 * 7]
+
+    HIGH_GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rbx, DWORD PTR arg(3)       ;out_pitch
+    lea         rax, [rax + rax]            ;bytes per line
+    lea         rbx, [rbx + rbx]
+    lea         rdx, [rax + rax * 2]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    LOAD_VERT_8 0
+    HIGH_APPLY_FILTER_8 0, 0
+
+    lea         rdi, [rdi + rbx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 8
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void aom_filter_block1d16_v8_sse2
+;(
+;    unsigned char *src_ptr,
+;    unsigned int   src_pitch,
+;    unsigned char *output_ptr,
+;    unsigned int   out_pitch,
+;    unsigned int   output_height,
+;    short *filter
+;)
+globalsym(aom_highbd_filter_block1d16_v8_sse2)
+sym(aom_highbd_filter_block1d16_v8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 8
+    %define k0k1 [rsp + 16 * 0]
+    %define k6k7 [rsp + 16 * 1]
+    %define k2k5 [rsp + 16 * 2]
+    %define k3k4 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define temp [rsp + 16 * 5]
+    %define max [rsp + 16 * 6]
+    %define min [rsp + 16 * 7]
+
+    HIGH_GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rbx, DWORD PTR arg(3)       ;out_pitch
+    lea         rax, [rax + rax]            ;bytes per line
+    lea         rbx, [rbx + rbx]
+    lea         rdx, [rax + rax * 2]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    LOAD_VERT_8 0
+    HIGH_APPLY_FILTER_8 0, 0
+    sub         rsi, rax
+
+    LOAD_VERT_8 16
+    HIGH_APPLY_FILTER_8 0, 16
+    add         rdi, rbx
+
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 8
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void aom_filter_block1d4_h8_sse2
+;(
+;    unsigned char  *src_ptr,
+;    unsigned int    src_pixels_per_line,
+;    unsigned char  *output_ptr,
+;    unsigned int    output_pitch,
+;    unsigned int    output_height,
+;    short *filter
+;)
+globalsym(aom_highbd_filter_block1d4_h8_sse2)
+sym(aom_highbd_filter_block1d4_h8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 7
+    %define k0k6 [rsp + 16 * 0]
+    %define k2k5 [rsp + 16 * 1]
+    %define k3k4 [rsp + 16 * 2]
+    %define k1k7 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define max [rsp + 16 * 5]
+    %define min [rsp + 16 * 6]
+
+    HIGH_GET_FILTERS_4
+
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    lea         rax, [rax + rax]            ;bytes per line
+    lea         rdx, [rdx + rdx]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movdqu      xmm0,   [rsi - 6]           ;load src
+    movdqu      xmm4,   [rsi + 2]
+    movdqa      xmm1, xmm0
+    movdqa      xmm6, xmm4
+    movdqa      xmm7, xmm4
+    movdqa      xmm2, xmm0
+    movdqa      xmm3, xmm0
+    movdqa      xmm5, xmm4
+
+    psrldq      xmm1, 2
+    psrldq      xmm6, 4
+    psrldq      xmm7, 6
+    psrldq      xmm2, 4
+    psrldq      xmm3, 6
+    psrldq      xmm5, 2
+
+    HIGH_APPLY_FILTER_4 0
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 7
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void aom_filter_block1d8_h8_sse2
+;(
+;    unsigned char  *src_ptr,
+;    unsigned int    src_pixels_per_line,
+;    unsigned char  *output_ptr,
+;    unsigned int    output_pitch,
+;    unsigned int    output_height,
+;    short *filter
+;)
+globalsym(aom_highbd_filter_block1d8_h8_sse2)
+sym(aom_highbd_filter_block1d8_h8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 8
+    %define k0k1 [rsp + 16 * 0]
+    %define k6k7 [rsp + 16 * 1]
+    %define k2k5 [rsp + 16 * 2]
+    %define k3k4 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define temp [rsp + 16 * 5]
+    %define max [rsp + 16 * 6]
+    %define min [rsp + 16 * 7]
+
+    HIGH_GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    lea         rax, [rax + rax]            ;bytes per line
+    lea         rdx, [rdx + rdx]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movdqu      xmm0,   [rsi - 6]           ;load src
+    movdqu      xmm1,   [rsi - 4]
+    movdqu      xmm2,   [rsi - 2]
+    movdqu      xmm3,   [rsi]
+    movdqu      xmm4,   [rsi + 2]
+    movdqu      xmm5,   [rsi + 4]
+    movdqu      xmm6,   [rsi + 6]
+    movdqu      xmm7,   [rsi + 8]
+
+    HIGH_APPLY_FILTER_8 0, 0
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 8
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void aom_filter_block1d16_h8_sse2
+;(
+;    unsigned char  *src_ptr,
+;    unsigned int    src_pixels_per_line,
+;    unsigned char  *output_ptr,
+;    unsigned int    output_pitch,
+;    unsigned int    output_height,
+;    short *filter
+;)
+globalsym(aom_highbd_filter_block1d16_h8_sse2)
+sym(aom_highbd_filter_block1d16_h8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 8
+    %define k0k1 [rsp + 16 * 0]
+    %define k6k7 [rsp + 16 * 1]
+    %define k2k5 [rsp + 16 * 2]
+    %define k3k4 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define temp [rsp + 16 * 5]
+    %define max [rsp + 16 * 6]
+    %define min [rsp + 16 * 7]
+
+    HIGH_GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    lea         rax, [rax + rax]            ;bytes per line
+    lea         rdx, [rdx + rdx]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movdqu      xmm0,   [rsi - 6]           ;load src
+    movdqu      xmm1,   [rsi - 4]
+    movdqu      xmm2,   [rsi - 2]
+    movdqu      xmm3,   [rsi]
+    movdqu      xmm4,   [rsi + 2]
+    movdqu      xmm5,   [rsi + 4]
+    movdqu      xmm6,   [rsi + 6]
+    movdqu      xmm7,   [rsi + 8]
+
+    HIGH_APPLY_FILTER_8 0, 0
+
+    movdqu      xmm0,   [rsi + 10]           ;load src
+    movdqu      xmm1,   [rsi + 12]
+    movdqu      xmm2,   [rsi + 14]
+    movdqu      xmm3,   [rsi + 16]
+    movdqu      xmm4,   [rsi + 18]
+    movdqu      xmm5,   [rsi + 20]
+    movdqu      xmm6,   [rsi + 22]
+    movdqu      xmm7,   [rsi + 24]
+
+    HIGH_APPLY_FILTER_8 0, 16
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 8
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
diff --git a/third_party/aom/aom_dsp/x86/aom_high_subpixel_bilinear_sse2.asm b/third_party/aom/aom_dsp/x86/aom_high_subpixel_bilinear_sse2.asm
new file mode 100644
index 0000000000..db4cad9bcb
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/aom_high_subpixel_bilinear_sse2.asm
@@ -0,0 +1,367 @@
+;
+; Copyright (c) 2016, 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 "aom_ports/x86_abi_support.asm"
+
+%macro HIGH_GET_PARAM_4 0
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+    mov         rcx, 0x00000040
+
+    movdqa      xmm3, [rdx]                 ;load filters
+    pshuflw     xmm4, xmm3, 11111111b       ;k3
+    psrldq      xmm3, 8
+    pshuflw     xmm3, xmm3, 0b              ;k4
+    punpcklwd   xmm4, xmm3                  ;k3k4
+
+    movq        xmm3, rcx                   ;rounding
+    pshufd      xmm3, xmm3, 0
+
+    mov         rdx, 0x00010001
+    movsxd      rcx, DWORD PTR arg(6)       ;bps
+    movq        xmm5, rdx
+    movq        xmm2, rcx
+    pshufd      xmm5, xmm5, 0b
+    movdqa      xmm1, xmm5
+    psllw       xmm5, xmm2
+    psubw       xmm5, xmm1                  ;max value (for clamping)
+    pxor        xmm2, xmm2                  ;min value (for clamping)
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+%endm
+
+%macro HIGH_APPLY_FILTER_4 1
+
+    punpcklwd   xmm0, xmm1                  ;two row in one register
+    pmaddwd     xmm0, xmm4                  ;multiply the filter factors
+
+    paddd       xmm0, xmm3                  ;rounding
+    psrad       xmm0, 7                     ;shift
+    packssdw    xmm0, xmm0                  ;pack to word
+
+    ;clamp the values
+    pminsw      xmm0, xmm5
+    pmaxsw      xmm0, xmm2
+
+%if %1
+    movq        xmm1, [rdi]
+    pavgw       xmm0, xmm1
+%endif
+
+    movq        [rdi], xmm0
+    lea         rsi, [rsi + 2*rax]
+    lea         rdi, [rdi + 2*rdx]
+    dec         rcx
+%endm
+
+%macro HIGH_GET_PARAM 0
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+    mov         rcx, 0x00000040
+
+    movdqa      xmm6, [rdx]                 ;load filters
+
+    pshuflw     xmm7, xmm6, 11111111b       ;k3
+    pshufhw     xmm6, xmm6, 0b              ;k4
+    psrldq      xmm6, 8
+    punpcklwd   xmm7, xmm6                  ;k3k4k3k4k3k4k3k4
+
+    movq        xmm4, rcx                   ;rounding
+    pshufd      xmm4, xmm4, 0
+
+    mov         rdx, 0x00010001
+    movsxd      rcx, DWORD PTR arg(6)       ;bps
+    movq        xmm3, rdx
+    movq        xmm5, rcx
+    pshufd      xmm3, xmm3, 0b
+    movdqa      xmm1, xmm3
+    psllw       xmm3, xmm5
+    psubw       xmm3, xmm1                  ;max value (for clamping)
+    pxor        xmm5, xmm5                  ;min value (for clamping)
+
+    movdqa      max, xmm3
+    movdqa      min, xmm5
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+%endm
+
+%macro HIGH_APPLY_FILTER_8 1
+    movdqa      xmm6, xmm0
+    punpckhwd   xmm6, xmm1
+    punpcklwd   xmm0, xmm1
+    pmaddwd     xmm6, xmm7
+    pmaddwd     xmm0, xmm7
+
+    paddd       xmm6, xmm4                  ;rounding
+    paddd       xmm0, xmm4                  ;rounding
+    psrad       xmm6, 7                     ;shift
+    psrad       xmm0, 7                     ;shift
+    packssdw    xmm0, xmm6                  ;pack back to word
+
+    ;clamp the values
+    pminsw      xmm0, max
+    pmaxsw      xmm0, min
+
+%if %1
+    movdqu      xmm1, [rdi]
+    pavgw       xmm0, xmm1
+%endif
+    movdqu      [rdi], xmm0                 ;store the result
+
+    lea         rsi, [rsi + 2*rax]
+    lea         rdi, [rdi + 2*rdx]
+    dec         rcx
+%endm
+
+%macro HIGH_APPLY_FILTER_16 1
+    movdqa      xmm5, xmm0
+    movdqa      xmm6, xmm2
+    punpckhwd   xmm5, xmm1
+    punpckhwd   xmm6, xmm3
+    punpcklwd   xmm0, xmm1
+    punpcklwd   xmm2, xmm3
+
+    pmaddwd     xmm5, xmm7
+    pmaddwd     xmm6, xmm7
+    pmaddwd     xmm0, xmm7
+    pmaddwd     xmm2, xmm7
+
+    paddd       xmm5, xmm4                  ;rounding
+    paddd       xmm6, xmm4
+    paddd       xmm0, xmm4
+    paddd       xmm2, xmm4
+
+    psrad       xmm5, 7                     ;shift
+    psrad       xmm6, 7
+    psrad       xmm0, 7
+    psrad       xmm2, 7
+
+    packssdw    xmm0, xmm5                  ;pack back to word
+    packssdw    xmm2, xmm6                  ;pack back to word
+
+    ;clamp the values
+    pminsw      xmm0, max
+    pmaxsw      xmm0, min
+    pminsw      xmm2, max
+    pmaxsw      xmm2, min
+
+%if %1
+    movdqu      xmm1, [rdi]
+    movdqu      xmm3, [rdi + 16]
+    pavgw       xmm0, xmm1
+    pavgw       xmm2, xmm3
+%endif
+    movdqu      [rdi], xmm0               ;store the result
+    movdqu      [rdi + 16], xmm2          ;store the result
+
+    lea         rsi, [rsi + 2*rax]
+    lea         rdi, [rdi + 2*rdx]
+    dec         rcx
+%endm
+
+SECTION .text
+
+globalsym(aom_highbd_filter_block1d4_v2_sse2)
+sym(aom_highbd_filter_block1d4_v2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    HIGH_GET_PARAM_4
+.loop:
+    movq        xmm0, [rsi]                 ;load src
+    movq        xmm1, [rsi + 2*rax]
+
+    HIGH_APPLY_FILTER_4 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+globalsym(aom_highbd_filter_block1d8_v2_sse2)
+sym(aom_highbd_filter_block1d8_v2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 8
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 2
+    %define max [rsp + 16 * 0]
+    %define min [rsp + 16 * 1]
+
+    HIGH_GET_PARAM
+.loop:
+    movdqu      xmm0, [rsi]                 ;0
+    movdqu      xmm1, [rsi + 2*rax]         ;1
+
+    HIGH_APPLY_FILTER_8 0
+    jnz         .loop
+
+    add rsp, 16 * 2
+    pop rsp
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+globalsym(aom_highbd_filter_block1d16_v2_sse2)
+sym(aom_highbd_filter_block1d16_v2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 9
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 2
+    %define max [rsp + 16 * 0]
+    %define min [rsp + 16 * 1]
+
+    HIGH_GET_PARAM
+.loop:
+    movdqu        xmm0, [rsi]               ;0
+    movdqu        xmm2, [rsi + 16]
+    movdqu        xmm1, [rsi + 2*rax]       ;1
+    movdqu        xmm3, [rsi + 2*rax + 16]
+
+    HIGH_APPLY_FILTER_16 0
+    jnz         .loop
+
+    add rsp, 16 * 2
+    pop rsp
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+globalsym(aom_highbd_filter_block1d4_h2_sse2)
+sym(aom_highbd_filter_block1d4_h2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    HIGH_GET_PARAM_4
+.loop:
+    movdqu      xmm0, [rsi]                 ;load src
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 2
+
+    HIGH_APPLY_FILTER_4 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+globalsym(aom_highbd_filter_block1d8_h2_sse2)
+sym(aom_highbd_filter_block1d8_h2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 8
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 2
+    %define max [rsp + 16 * 0]
+    %define min [rsp + 16 * 1]
+
+    HIGH_GET_PARAM
+.loop:
+    movdqu      xmm0, [rsi]                 ;load src
+    movdqu      xmm1, [rsi + 2]
+
+    HIGH_APPLY_FILTER_8 0
+    jnz         .loop
+
+    add rsp, 16 * 2
+    pop rsp
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+globalsym(aom_highbd_filter_block1d16_h2_sse2)
+sym(aom_highbd_filter_block1d16_h2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 7
+    SAVE_XMM 9
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 2
+    %define max [rsp + 16 * 0]
+    %define min [rsp + 16 * 1]
+
+    HIGH_GET_PARAM
+.loop:
+    movdqu      xmm0,   [rsi]               ;load src
+    movdqu      xmm1,   [rsi + 2]
+    movdqu      xmm2,   [rsi + 16]
+    movdqu      xmm3,   [rsi + 18]
+
+    HIGH_APPLY_FILTER_16 0
+    jnz         .loop
+
+    add rsp, 16 * 2
+    pop rsp
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
diff --git a/third_party/aom/aom_dsp/x86/aom_quantize_avx.c b/third_party/aom/aom_dsp/x86/aom_quantize_avx.c
new file mode 100644
index 0000000000..b2d6d4b76d
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/aom_quantize_avx.c
@@ -0,0 +1,282 @@
+/*
+ * Copyright (c) 2020, 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"
+#include "aom/aom_integer.h"
+#include "aom_dsp/x86/bitdepth_conversion_sse2.h"
+#include "aom_dsp/x86/quantize_x86.h"
+
+static INLINE void calculate_dqcoeff_and_store(__m128i qcoeff, __m128i dequant,
+                                               tran_low_t *dqcoeff) {
+  const __m128i low = _mm_mullo_epi16(qcoeff, dequant);
+  const __m128i high = _mm_mulhi_epi16(qcoeff, dequant);
+
+  const __m128i dqcoeff32_0 = _mm_unpacklo_epi16(low, high);
+  const __m128i dqcoeff32_1 = _mm_unpackhi_epi16(low, high);
+
+  _mm_store_si128((__m128i *)(dqcoeff), dqcoeff32_0);
+  _mm_store_si128((__m128i *)(dqcoeff + 4), dqcoeff32_1);
+}
+
+void aom_quantize_b_avx(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                        const int16_t *zbin_ptr, const int16_t *round_ptr,
+                        const int16_t *quant_ptr,
+                        const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
+                        tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr,
+                        uint16_t *eob_ptr, const int16_t *scan,
+                        const int16_t *iscan) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m256i big_zero = _mm256_setzero_si256();
+  int index;
+
+  __m128i zbin, round, quant, dequant, shift;
+  __m128i coeff0, coeff1;
+  __m128i qcoeff0, qcoeff1;
+  __m128i cmp_mask0, cmp_mask1;
+  __m128i all_zero;
+  __m128i eob = zero, eob0;
+
+  (void)scan;
+
+  *eob_ptr = 0;
+
+  load_b_values(zbin_ptr, &zbin, round_ptr, &round, quant_ptr, &quant,
+                dequant_ptr, &dequant, quant_shift_ptr, &shift);
+
+  // Do DC and first 15 AC.
+  coeff0 = load_tran_low(coeff_ptr);
+  coeff1 = load_tran_low(coeff_ptr + 8);
+
+  qcoeff0 = _mm_abs_epi16(coeff0);
+  qcoeff1 = _mm_abs_epi16(coeff1);
+
+  cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin);
+  zbin = _mm_unpackhi_epi64(zbin, zbin);  // Switch DC to AC
+  cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);
+
+  all_zero = _mm_or_si128(cmp_mask0, cmp_mask1);
+  if (_mm_test_all_zeros(all_zero, all_zero)) {
+    _mm256_store_si256((__m256i *)(qcoeff_ptr), big_zero);
+    _mm256_store_si256((__m256i *)(dqcoeff_ptr), big_zero);
+    _mm256_store_si256((__m256i *)(qcoeff_ptr + 8), big_zero);
+    _mm256_store_si256((__m256i *)(dqcoeff_ptr + 8), big_zero);
+
+    if (n_coeffs == 16) return;
+
+    round = _mm_unpackhi_epi64(round, round);
+    quant = _mm_unpackhi_epi64(quant, quant);
+    shift = _mm_unpackhi_epi64(shift, shift);
+    dequant = _mm_unpackhi_epi64(dequant, dequant);
+  } else {
+    calculate_qcoeff(&qcoeff0, round, quant, shift);
+    round = _mm_unpackhi_epi64(round, round);
+    quant = _mm_unpackhi_epi64(quant, quant);
+    shift = _mm_unpackhi_epi64(shift, shift);
+    calculate_qcoeff(&qcoeff1, round, quant, shift);
+
+    // Reinsert signs
+    qcoeff0 = _mm_sign_epi16(qcoeff0, coeff0);
+    qcoeff1 = _mm_sign_epi16(qcoeff1, coeff1);
+
+    // Mask out zbin threshold coeffs
+    qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
+    qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
+
+    store_tran_low(qcoeff0, qcoeff_ptr);
+    store_tran_low(qcoeff1, qcoeff_ptr + 8);
+
+    calculate_dqcoeff_and_store(qcoeff0, dequant, dqcoeff_ptr);
+    dequant = _mm_unpackhi_epi64(dequant, dequant);
+    calculate_dqcoeff_and_store(qcoeff1, dequant, dqcoeff_ptr + 8);
+
+    eob =
+        scan_for_eob(&qcoeff0, &qcoeff1, cmp_mask0, cmp_mask1, iscan, 0, zero);
+  }
+
+  // AC only loop.
+  for (index = 16; index < n_coeffs; index += 16) {
+    coeff0 = load_tran_low(coeff_ptr + index);
+    coeff1 = load_tran_low(coeff_ptr + index + 8);
+
+    qcoeff0 = _mm_abs_epi16(coeff0);
+    qcoeff1 = _mm_abs_epi16(coeff1);
+
+    cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin);
+    cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);
+
+    all_zero = _mm_or_si128(cmp_mask0, cmp_mask1);
+    if (_mm_test_all_zeros(all_zero, all_zero)) {
+      _mm256_store_si256((__m256i *)(qcoeff_ptr + index), big_zero);
+      _mm256_store_si256((__m256i *)(dqcoeff_ptr + index), big_zero);
+      _mm256_store_si256((__m256i *)(qcoeff_ptr + index + 8), big_zero);
+      _mm256_store_si256((__m256i *)(dqcoeff_ptr + index + 8), big_zero);
+      continue;
+    }
+
+    calculate_qcoeff(&qcoeff0, round, quant, shift);
+    calculate_qcoeff(&qcoeff1, round, quant, shift);
+
+    qcoeff0 = _mm_sign_epi16(qcoeff0, coeff0);
+    qcoeff1 = _mm_sign_epi16(qcoeff1, coeff1);
+
+    qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
+    qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
+
+    store_tran_low(qcoeff0, qcoeff_ptr + index);
+    store_tran_low(qcoeff1, qcoeff_ptr + index + 8);
+
+    calculate_dqcoeff_and_store(qcoeff0, dequant, dqcoeff_ptr + index);
+    calculate_dqcoeff_and_store(qcoeff1, dequant, dqcoeff_ptr + index + 8);
+
+    eob0 = scan_for_eob(&qcoeff0, &qcoeff1, cmp_mask0, cmp_mask1, iscan, index,
+                        zero);
+    eob = _mm_max_epi16(eob, eob0);
+  }
+
+  *eob_ptr = accumulate_eob(eob);
+}
+
+void aom_quantize_b_32x32_avx(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                              const int16_t *zbin_ptr, const int16_t *round_ptr,
+                              const int16_t *quant_ptr,
+                              const int16_t *quant_shift_ptr,
+                              tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                              const int16_t *dequant_ptr, uint16_t *eob_ptr,
+                              const int16_t *scan, const int16_t *iscan) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i one = _mm_set1_epi16(1);
+  const __m256i big_zero = _mm256_setzero_si256();
+  int index;
+  const int log_scale = 1;
+
+  __m128i zbin, round, quant, dequant, shift;
+  __m128i coeff0, coeff1;
+  __m128i qcoeff0, qcoeff1;
+  __m128i cmp_mask0, cmp_mask1;
+  __m128i all_zero;
+  __m128i eob = zero, eob0;
+
+  (void)scan;
+
+  // Setup global values.
+  // The 32x32 halves zbin and round.
+  zbin = _mm_load_si128((const __m128i *)zbin_ptr);
+  // Shift with rounding.
+  zbin = _mm_add_epi16(zbin, one);
+  zbin = _mm_srli_epi16(zbin, 1);
+  // x86 has no "greater *or equal*" comparison. Subtract 1 from zbin so
+  // it is a strict "greater" comparison.
+  zbin = _mm_sub_epi16(zbin, one);
+
+  round = _mm_load_si128((const __m128i *)round_ptr);
+  round = _mm_add_epi16(round, one);
+  round = _mm_srli_epi16(round, 1);
+
+  quant = _mm_load_si128((const __m128i *)quant_ptr);
+  dequant = _mm_load_si128((const __m128i *)dequant_ptr);
+  shift = _mm_load_si128((const __m128i *)quant_shift_ptr);
+
+  // Do DC and first 15 AC.
+  coeff0 = load_tran_low(coeff_ptr);
+  coeff1 = load_tran_low(coeff_ptr + 8);
+
+  qcoeff0 = _mm_abs_epi16(coeff0);
+  qcoeff1 = _mm_abs_epi16(coeff1);
+
+  cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin);
+  zbin = _mm_unpackhi_epi64(zbin, zbin);  // Switch DC to AC.
+  cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);
+
+  all_zero = _mm_or_si128(cmp_mask0, cmp_mask1);
+  if (_mm_test_all_zeros(all_zero, all_zero)) {
+    _mm256_store_si256((__m256i *)(qcoeff_ptr), big_zero);
+    _mm256_store_si256((__m256i *)(dqcoeff_ptr), big_zero);
+    _mm256_store_si256((__m256i *)(qcoeff_ptr + 8), big_zero);
+    _mm256_store_si256((__m256i *)(dqcoeff_ptr + 8), big_zero);
+
+    round = _mm_unpackhi_epi64(round, round);
+    quant = _mm_unpackhi_epi64(quant, quant);
+    shift = _mm_unpackhi_epi64(shift, shift);
+    dequant = _mm_unpackhi_epi64(dequant, dequant);
+  } else {
+    calculate_qcoeff_log_scale(&qcoeff0, round, quant, &shift, &log_scale);
+    round = _mm_unpackhi_epi64(round, round);
+    quant = _mm_unpackhi_epi64(quant, quant);
+    shift = _mm_unpackhi_epi64(shift, shift);
+    calculate_qcoeff_log_scale(&qcoeff1, round, quant, &shift, &log_scale);
+
+    // Reinsert signs.
+    qcoeff0 = _mm_sign_epi16(qcoeff0, coeff0);
+    qcoeff1 = _mm_sign_epi16(qcoeff1, coeff1);
+
+    // Mask out zbin threshold coeffs.
+    qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
+    qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
+
+    store_tran_low(qcoeff0, qcoeff_ptr);
+    store_tran_low(qcoeff1, qcoeff_ptr + 8);
+
+    calculate_dqcoeff_and_store_log_scale(qcoeff0, dequant, zero, dqcoeff_ptr,
+                                          &log_scale);
+    dequant = _mm_unpackhi_epi64(dequant, dequant);
+    calculate_dqcoeff_and_store_log_scale(qcoeff1, dequant, zero,
+                                          dqcoeff_ptr + 8, &log_scale);
+
+    eob =
+        scan_for_eob(&qcoeff0, &qcoeff1, cmp_mask0, cmp_mask1, iscan, 0, zero);
+  }
+
+  // AC only loop.
+  for (index = 16; index < n_coeffs; index += 16) {
+    coeff0 = load_tran_low(coeff_ptr + index);
+    coeff1 = load_tran_low(coeff_ptr + index + 8);
+
+    qcoeff0 = _mm_abs_epi16(coeff0);
+    qcoeff1 = _mm_abs_epi16(coeff1);
+
+    cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin);
+    cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);
+
+    all_zero = _mm_or_si128(cmp_mask0, cmp_mask1);
+    if (_mm_test_all_zeros(all_zero, all_zero)) {
+      _mm256_store_si256((__m256i *)(qcoeff_ptr + index), big_zero);
+      _mm256_store_si256((__m256i *)(dqcoeff_ptr + index), big_zero);
+      _mm256_store_si256((__m256i *)(qcoeff_ptr + index + 8), big_zero);
+      _mm256_store_si256((__m256i *)(dqcoeff_ptr + index + 8), big_zero);
+      continue;
+    }
+
+    calculate_qcoeff_log_scale(&qcoeff0, round, quant, &shift, &log_scale);
+    calculate_qcoeff_log_scale(&qcoeff1, round, quant, &shift, &log_scale);
+
+    qcoeff0 = _mm_sign_epi16(qcoeff0, coeff0);
+    qcoeff1 = _mm_sign_epi16(qcoeff1, coeff1);
+
+    qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
+    qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
+
+    store_tran_low(qcoeff0, qcoeff_ptr + index);
+    store_tran_low(qcoeff1, qcoeff_ptr + index + 8);
+
+    calculate_dqcoeff_and_store_log_scale(qcoeff0, dequant, zero,
+                                          dqcoeff_ptr + index, &log_scale);
+    calculate_dqcoeff_and_store_log_scale(qcoeff1, dequant, zero,
+                                          dqcoeff_ptr + index + 8, &log_scale);
+
+    eob0 = scan_for_eob(&qcoeff0, &qcoeff1, cmp_mask0, cmp_mask1, iscan, index,
+                        zero);
+    eob = _mm_max_epi16(eob, eob0);
+  }
+
+  *eob_ptr = accumulate_eob(eob);
+}
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
new file mode 100644
index 0000000000..22f2e696d3
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/aom_subpixel_8t_intrin_avx2.c
@@ -0,0 +1,1441 @@
+/*
+ * Copyright (c) 2016, 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"
+
+#include "aom_dsp/x86/convolve.h"
+#include "aom_dsp/x86/convolve_avx2.h"
+#include "aom_ports/mem.h"
+
+#if defined(__clang__)
+#if (__clang_major__ > 0 && __clang_major__ < 3) ||            \
+    (__clang_major__ == 3 && __clang_minor__ <= 3) ||          \
+    (defined(__APPLE__) && defined(__apple_build_version__) && \
+     ((__clang_major__ == 4 && __clang_minor__ <= 2) ||        \
+      (__clang_major__ == 5 && __clang_minor__ == 0)))
+#define MM256_BROADCASTSI128_SI256(x) \
+  _mm_broadcastsi128_si256((__m128i const *)&(x))
+#else  // clang > 3.3, and not 5.0 on macosx.
+#define MM256_BROADCASTSI128_SI256(x) _mm256_broadcastsi128_si256(x)
+#endif  // clang <= 3.3
+#elif defined(__GNUC__)
+#if __GNUC__ < 4 || (__GNUC__ == 4 && __GNUC_MINOR__ <= 6)
+#define MM256_BROADCASTSI128_SI256(x) \
+  _mm_broadcastsi128_si256((__m128i const *)&(x))
+#elif __GNUC__ == 4 && __GNUC_MINOR__ == 7
+#define MM256_BROADCASTSI128_SI256(x) _mm_broadcastsi128_si256(x)
+#else  // gcc > 4.7
+#define MM256_BROADCASTSI128_SI256(x) _mm256_broadcastsi128_si256(x)
+#endif  // gcc <= 4.6
+#else   // !(gcc || clang)
+#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) {
+  *((int *)(output_ptr)) = _mm_cvtsi128_si32(_mm256_castsi256_si128(*a));
+  *((int *)(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_h4_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, firstFilters, srcReg32b1, srcRegFilt32b1_1;
+  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);
+
+  firstFilters =
+      _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi32(0x5040302u));
+  filt1Reg = _mm256_load_si256((__m256i const *)(filt4_d4_global_avx2));
+
+  // 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);
+
+    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;
+
+    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));
+
+    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
+    *((int *)(output_ptr)) = _mm_cvtsi128_si32(srcRegFilt1_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
+    *((int *)(output_ptr)) = _mm_cvtsi128_si32(srcRegFilt1_1);
+  }
+}
+
+static void aom_filter_block1d8_h4_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, filt2Reg, filt3Reg;
+  __m256i secondFilters, thirdFilters;
+  __m256i srcRegFilt32b1_1, srcRegFilt32b2, srcRegFilt32b3;
+  __m256i srcReg32b1, filtersReg32;
+  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
+  filtersReg32 = MM256_BROADCASTSI128_SI256(filtersReg);
+
+  // 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));
+
+  filt2Reg = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32));
+  filt3Reg = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32 * 2));
+
+  // multiply 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
+    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);
+
+    srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b3, srcRegFilt32b2);
+
+    // 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
+    srcRegFilt32b1_1 = _mm256_packus_epi16(srcRegFilt32b1_1, srcRegFilt32b1_1);
+
+    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
+    srcRegFilt2 = _mm_shuffle_epi8(srcReg1, _mm256_castsi256_si128(filt2Reg));
+    srcRegFilt3 = _mm_shuffle_epi8(srcReg1, _mm256_castsi256_si128(filt3Reg));
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt2 =
+        _mm_maddubs_epi16(srcRegFilt2, _mm256_castsi256_si128(secondFilters));
+    srcRegFilt3 =
+        _mm_maddubs_epi16(srcRegFilt3, _mm256_castsi256_si128(thirdFilters));
+
+    // add and saturate the results together
+    srcRegFilt1_1 = _mm_adds_epi16(srcRegFilt2, srcRegFilt3);
+
+    // 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
+    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_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_h4_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, filt2Reg, filt3Reg;
+  __m256i secondFilters, thirdFilters;
+  __m256i srcRegFilt32b1_1, srcRegFilt32b2_1, srcRegFilt32b2, srcRegFilt32b3;
+  __m256i srcReg32b1, srcReg32b2, filtersReg32;
+  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
+  filtersReg32 = MM256_BROADCASTSI128_SI256(filtersReg);
+
+  // 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));
+
+  filt2Reg = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32));
+  filt3Reg = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32 * 2));
+
+  // multiply 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
+    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);
+
+    srcRegFilt32b1_1 = _mm256_adds_epi16(srcRegFilt32b3, srcRegFilt32b2);
+
+    // reading 2 strides of the next 16 bytes
+    // (part of it was being read by earlier read)
+    srcReg32b2 =
+        xx_loadu2_mi128(src_ptr + src_pixels_per_line + 8, src_ptr + 8);
+
+    // filter the source buffer
+    srcRegFilt32b3 = _mm256_shuffle_epi8(srcReg32b2, filt2Reg);
+    srcRegFilt32b2 = _mm256_shuffle_epi8(srcReg32b2, filt3Reg);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt32b3 = _mm256_maddubs_epi16(srcRegFilt32b3, secondFilters);
+    srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, thirdFilters);
+
+    // add and saturate the results together
+    srcRegFilt32b2_1 = _mm256_adds_epi16(srcRegFilt32b3, srcRegFilt32b2);
+
+    // 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
+    srcRegFilt32b1_1 = _mm256_packus_epi16(srcRegFilt32b1_1, srcRegFilt32b2_1);
+
+    src_ptr += src_stride;
+
+    xx_store2_mi128(output_ptr, output_pitch, &srcRegFilt32b1_1);
+    output_ptr += dst_stride;
+  }
+
+  // if the number of strides is odd.
+  // process only 16 bytes
+  if (i > 0) {
+    __m256i srcReg1, srcReg12;
+    __m256i srcRegFilt2, srcRegFilt3, srcRegFilt1_1;
+
+    srcReg1 = _mm256_loadu_si256((const __m256i *)(src_ptr));
+    srcReg12 = _mm256_permute4x64_epi64(srcReg1, 0x94);
+
+    // filter the source buffer
+    srcRegFilt2 = _mm256_shuffle_epi8(srcReg12, filt2Reg);
+    srcRegFilt3 = _mm256_shuffle_epi8(srcReg12, filt3Reg);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt2 = _mm256_maddubs_epi16(srcRegFilt2, secondFilters);
+    srcRegFilt3 = _mm256_maddubs_epi16(srcRegFilt3, thirdFilters);
+
+    // add and saturate the results together
+    srcRegFilt1_1 = _mm256_adds_epi16(srcRegFilt2, srcRegFilt3);
+
+    // shift by 6 bit each 16 bit
+    srcRegFilt1_1 = _mm256_adds_epi16(srcRegFilt1_1, addFilterReg32);
+    srcRegFilt1_1 = _mm256_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 = _mm256_packus_epi16(srcRegFilt1_1, srcRegFilt1_1);
+    srcRegFilt1_1 = _mm256_permute4x64_epi64(srcRegFilt1_1, 0x8);
+
+    // save 16 bytes
+    _mm_store_si128((__m128i *)output_ptr,
+                    _mm256_castsi256_si128(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 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;
+  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
+  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);
+
+    // reading 2 strides of the next 16 bytes
+    // (part of it was being read by earlier read)
+    srcReg32b2 =
+        xx_loadu2_mi128(src_ptr + src_pixels_per_line + 8, src_ptr + 8);
+
+    // filter the source buffer
+    srcRegFilt32b2_1 = _mm256_shuffle_epi8(srcReg32b2, filt1Reg);
+    srcRegFilt32b2 = _mm256_shuffle_epi8(srcReg32b2, filt4Reg);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt32b2_1 = _mm256_maddubs_epi16(srcRegFilt32b2_1, firstFilters);
+    srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, forthFilters);
+
+    // add and saturate the results together
+    srcRegFilt32b2_1 = _mm256_adds_epi16(srcRegFilt32b2_1, srcRegFilt32b2);
+
+    // filter the source buffer
+    srcRegFilt32b3 = _mm256_shuffle_epi8(srcReg32b2, filt2Reg);
+    srcRegFilt32b2 = _mm256_shuffle_epi8(srcReg32b2, filt3Reg);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt32b3 = _mm256_maddubs_epi16(srcRegFilt32b3, secondFilters);
+    srcRegFilt32b2 = _mm256_maddubs_epi16(srcRegFilt32b2, thirdFilters);
+
+    // add and saturate the results together
+    srcRegFilt32b2_1 = _mm256_adds_epi16(
+        srcRegFilt32b2_1, _mm256_adds_epi16(srcRegFilt32b3, srcRegFilt32b2));
+
+    // 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
+    srcRegFilt32b1_1 = _mm256_packus_epi16(srcRegFilt32b1_1, srcRegFilt32b2_1);
+
+    src_ptr += src_stride;
+
+    xx_store2_mi128(output_ptr, output_pitch, &srcRegFilt32b1_1);
+    output_ptr += dst_stride;
+  }
+
+  // if the number of strides is odd.
+  // process only 16 bytes
+  if (i > 0) {
+    __m128i srcReg1, srcReg2, srcRegFilt1_1, srcRegFilt2_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));
+
+    // reading the next 16 bytes
+    // (part of it was being read by earlier read)
+    srcReg2 = _mm_loadu_si128((const __m128i *)(src_ptr + 8));
+
+    // filter the source buffer
+    srcRegFilt2_1 = _mm_shuffle_epi8(srcReg2, _mm256_castsi256_si128(filt1Reg));
+    srcRegFilt2 = _mm_shuffle_epi8(srcReg2, _mm256_castsi256_si128(filt4Reg));
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt2_1 =
+        _mm_maddubs_epi16(srcRegFilt2_1, _mm256_castsi256_si128(firstFilters));
+    srcRegFilt2 =
+        _mm_maddubs_epi16(srcRegFilt2, _mm256_castsi256_si128(forthFilters));
+
+    // add and saturate the results together
+    srcRegFilt2_1 = _mm_adds_epi16(srcRegFilt2_1, srcRegFilt2);
+
+    // filter the source buffer
+    srcRegFilt3 = _mm_shuffle_epi8(srcReg2, _mm256_castsi256_si128(filt2Reg));
+    srcRegFilt2 = _mm_shuffle_epi8(srcReg2, _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
+    srcRegFilt2_1 =
+        _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(addFilterReg32));
+    srcRegFilt1_1 = _mm_srai_epi16(srcRegFilt1_1, 6);
+
+    srcRegFilt2_1 =
+        _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
+    // result
+    srcRegFilt1_1 = _mm_packus_epi16(srcRegFilt1_1, srcRegFilt2_1);
+
+    // save 16 bytes
+    _mm_store_si128((__m128i *)output_ptr, srcRegFilt1_1);
+  }
+}
+
+static void aom_filter_block1d8_v4_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 filtersReg32, addFilterReg32;
+  __m256i srcReg23, srcReg4x, srcReg34, srcReg5x, srcReg45, srcReg6x, srcReg56;
+  __m256i srcReg23_34_lo, srcReg45_56_lo;
+  __m256i resReg23_34_lo, resReg45_56_lo;
+  __m256i resReglo, resReg;
+  __m256i secondFilters, thirdFilters;
+  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 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));
+
+  // multiple the size of the source and destination stride by two
+  src_stride = src_pitch << 1;
+  dst_stride = out_pitch << 1;
+
+  srcReg23 = xx_loadu2_epi64(src_ptr + src_pitch * 3, src_ptr + src_pitch * 2);
+  srcReg4x = _mm256_castsi128_si256(
+      _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 4)));
+
+  // have consecutive loads on the same 256 register
+  srcReg34 = _mm256_permute2x128_si256(srcReg23, srcReg4x, 0x21);
+
+  srcReg23_34_lo = _mm256_unpacklo_epi8(srcReg23, srcReg34);
+
+  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
+    srcReg5x = _mm256_castsi128_si256(
+        _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 5)));
+    srcReg45 =
+        _mm256_inserti128_si256(srcReg4x, _mm256_castsi256_si128(srcReg5x), 1);
+
+    srcReg6x = _mm256_castsi128_si256(
+        _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 6)));
+    srcReg56 =
+        _mm256_inserti128_si256(srcReg5x, _mm256_castsi256_si128(srcReg6x), 1);
+
+    // merge every two consecutive registers
+    srcReg45_56_lo = _mm256_unpacklo_epi8(srcReg45, srcReg56);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    resReg23_34_lo = _mm256_maddubs_epi16(srcReg23_34_lo, secondFilters);
+    resReg45_56_lo = _mm256_maddubs_epi16(srcReg45_56_lo, thirdFilters);
+
+    // add and saturate the results together
+    resReglo = _mm256_adds_epi16(resReg23_34_lo, resReg45_56_lo);
+
+    // shift by 6 bit each 16 bit
+    resReglo = _mm256_adds_epi16(resReglo, addFilterReg32);
+    resReglo = _mm256_srai_epi16(resReglo, 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
+    resReg = _mm256_packus_epi16(resReglo, resReglo);
+
+    src_ptr += src_stride;
+
+    xx_storeu2_epi64(output_ptr, out_pitch, &resReg);
+
+    output_ptr += dst_stride;
+
+    // save part of the registers for next strides
+    srcReg23_34_lo = srcReg45_56_lo;
+    srcReg4x = srcReg6x;
+  }
+}
+
+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_v4_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 filtersReg32, addFilterReg32;
+  __m256i srcReg23, srcReg4x, srcReg34, srcReg5x, srcReg45, srcReg6x, srcReg56;
+  __m256i srcReg23_34_lo, srcReg23_34_hi, srcReg45_56_lo, srcReg45_56_hi;
+  __m256i resReg23_34_lo, resReg23_34_hi, resReg45_56_lo, resReg45_56_hi;
+  __m256i resReglo, resReghi, resReg;
+  __m256i secondFilters, thirdFilters;
+  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 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));
+
+  // multiple the size of the source and destination stride by two
+  src_stride = src_pitch << 1;
+  dst_stride = out_pitch << 1;
+
+  srcReg23 = xx_loadu2_mi128(src_ptr + src_pitch * 3, src_ptr + src_pitch * 2);
+  srcReg4x = _mm256_castsi128_si256(
+      _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 4)));
+
+  // have consecutive loads on the same 256 register
+  srcReg34 = _mm256_permute2x128_si256(srcReg23, srcReg4x, 0x21);
+
+  srcReg23_34_lo = _mm256_unpacklo_epi8(srcReg23, srcReg34);
+  srcReg23_34_hi = _mm256_unpackhi_epi8(srcReg23, srcReg34);
+
+  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
+    srcReg5x = _mm256_castsi128_si256(
+        _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 5)));
+    srcReg45 =
+        _mm256_inserti128_si256(srcReg4x, _mm256_castsi256_si128(srcReg5x), 1);
+
+    srcReg6x = _mm256_castsi128_si256(
+        _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 6)));
+    srcReg56 =
+        _mm256_inserti128_si256(srcReg5x, _mm256_castsi256_si128(srcReg6x), 1);
+
+    // merge every two consecutive registers
+    srcReg45_56_lo = _mm256_unpacklo_epi8(srcReg45, srcReg56);
+    srcReg45_56_hi = _mm256_unpackhi_epi8(srcReg45, srcReg56);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    resReg23_34_lo = _mm256_maddubs_epi16(srcReg23_34_lo, secondFilters);
+    resReg45_56_lo = _mm256_maddubs_epi16(srcReg45_56_lo, thirdFilters);
+
+    // add and saturate the results together
+    resReglo = _mm256_adds_epi16(resReg23_34_lo, resReg45_56_lo);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    resReg23_34_hi = _mm256_maddubs_epi16(srcReg23_34_hi, secondFilters);
+    resReg45_56_hi = _mm256_maddubs_epi16(srcReg45_56_hi, thirdFilters);
+
+    // add and saturate the results together
+    resReghi = _mm256_adds_epi16(resReg23_34_hi, resReg45_56_hi);
+
+    // shift by 6 bit each 16 bit
+    resReglo = _mm256_adds_epi16(resReglo, addFilterReg32);
+    resReghi = _mm256_adds_epi16(resReghi, addFilterReg32);
+    resReglo = _mm256_srai_epi16(resReglo, 6);
+    resReghi = _mm256_srai_epi16(resReghi, 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
+    resReg = _mm256_packus_epi16(resReglo, resReghi);
+
+    src_ptr += src_stride;
+
+    xx_store2_mi128(output_ptr, out_pitch, &resReg);
+
+    output_ptr += dst_stride;
+
+    // save part of the registers for next strides
+    srcReg23_34_lo = srcReg45_56_lo;
+    srcReg23_34_hi = srcReg45_56_hi;
+    srcReg4x = srcReg6x;
+  }
+}
+
+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 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_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
+  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);
+  srcReg32b3 = _mm256_unpackhi_epi8(srcReg32b3, srcReg32b4);
+  srcReg32b2 = _mm256_unpacklo_epi8(srcReg32b5, srcReg32b6);
+  srcReg32b5 = _mm256_unpackhi_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_loadu_si128((const __m128i *)(src_ptr + src_pitch * 7)));
+    srcReg32b7 = _mm256_inserti128_si256(srcReg32b7,
+                                         _mm256_castsi256_si128(srcReg32b8), 1);
+    srcReg32b9 = _mm256_castsi128_si256(
+        _mm_loadu_si128((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);
+    srcReg32b7 = _mm256_unpackhi_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));
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcReg32b1 = _mm256_maddubs_epi16(srcReg32b1, firstFilters);
+    srcReg32b6 = _mm256_maddubs_epi16(srcReg32b7, forthFilters);
+
+    srcReg32b1 = _mm256_adds_epi16(srcReg32b1, srcReg32b6);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcReg32b8 = _mm256_maddubs_epi16(srcReg32b3, secondFilters);
+    srcReg32b12 = _mm256_maddubs_epi16(srcReg32b5, thirdFilters);
+
+    // add and saturate the results together
+    srcReg32b1 = _mm256_adds_epi16(srcReg32b1,
+                                   _mm256_adds_epi16(srcReg32b8, srcReg32b12));
+
+    // 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
+    // result
+    srcReg32b1 = _mm256_packus_epi16(srcReg32b10, srcReg32b1);
+
+    src_ptr += src_stride;
+
+    xx_store2_mi128(output_ptr, out_pitch, &srcReg32b1);
+
+    output_ptr += dst_stride;
+
+    // save part of the registers for next strides
+    srcReg32b10 = srcReg32b11;
+    srcReg32b1 = srcReg32b3;
+    srcReg32b11 = srcReg32b2;
+    srcReg32b3 = srcReg32b5;
+    srcReg32b2 = srcReg32b4;
+    srcReg32b5 = srcReg32b7;
+    srcReg32b7 = srcReg32b9;
+  }
+  if (i > 0) {
+    __m128i srcRegFilt1, srcRegFilt3, srcRegFilt4, srcRegFilt5;
+    __m128i srcRegFilt6, srcRegFilt7, srcRegFilt8;
+    // load the last 16 bytes
+    srcRegFilt8 = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 7));
+
+    // merge the last 2 results together
+    srcRegFilt4 =
+        _mm_unpacklo_epi8(_mm256_castsi256_si128(srcReg32b7), srcRegFilt8);
+    srcRegFilt7 =
+        _mm_unpackhi_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));
+    srcRegFilt3 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b1),
+                                    _mm256_castsi256_si128(firstFilters));
+    srcRegFilt7 =
+        _mm_maddubs_epi16(srcRegFilt7, _mm256_castsi256_si128(forthFilters));
+
+    // add and saturate the results together
+    srcRegFilt1 = _mm_adds_epi16(srcRegFilt1, srcRegFilt4);
+    srcRegFilt3 = _mm_adds_epi16(srcRegFilt3, srcRegFilt7);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt4 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b11),
+                                    _mm256_castsi256_si128(secondFilters));
+    srcRegFilt5 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b3),
+                                    _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));
+    srcRegFilt7 = _mm_maddubs_epi16(_mm256_castsi256_si128(srcReg32b5),
+                                    _mm256_castsi256_si128(thirdFilters));
+
+    // add and saturate the results together
+    srcRegFilt1 =
+        _mm_adds_epi16(srcRegFilt1, _mm_adds_epi16(srcRegFilt4, srcRegFilt6));
+    srcRegFilt3 =
+        _mm_adds_epi16(srcRegFilt3, _mm_adds_epi16(srcRegFilt5, srcRegFilt7));
+
+    // shift by 6 bit each 16 bit
+    srcRegFilt1 =
+        _mm_adds_epi16(srcRegFilt1, _mm256_castsi256_si128(addFilterReg32));
+    srcRegFilt3 =
+        _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
+    // result
+    srcRegFilt1 = _mm_packus_epi16(srcRegFilt1, srcRegFilt3);
+
+    // save 16 bytes
+    _mm_store_si128((__m128i *)output_ptr, srcRegFilt1);
+  }
+}
+
+static void aom_filter_block1d4_v4_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 filtersReg32, addFilterReg32;
+  __m256i srcReg23, srcReg4x, srcReg34, srcReg5x, srcReg45, srcReg6x, srcReg56;
+  __m256i srcReg23_34_lo, srcReg45_56_lo;
+  __m256i srcReg2345_3456_lo;
+  __m256i resReglo, resReg;
+  __m256i firstFilters;
+  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);
+
+  firstFilters =
+      _mm256_shuffle_epi8(filtersReg32, _mm256_set1_epi32(0x5040302u));
+
+  // multiple the size of the source and destination stride by two
+  src_stride = src_pitch << 1;
+  dst_stride = out_pitch << 1;
+
+  srcReg23 = xx_loadu2_epi64(src_ptr + src_pitch * 3, src_ptr + src_pitch * 2);
+  srcReg4x = _mm256_castsi128_si256(
+      _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 4)));
+
+  // have consecutive loads on the same 256 register
+  srcReg34 = _mm256_permute2x128_si256(srcReg23, srcReg4x, 0x21);
+
+  srcReg23_34_lo = _mm256_unpacklo_epi8(srcReg23, srcReg34);
+
+  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
+    srcReg5x = _mm256_castsi128_si256(
+        _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 5)));
+    srcReg45 =
+        _mm256_inserti128_si256(srcReg4x, _mm256_castsi256_si128(srcReg5x), 1);
+
+    srcReg6x = _mm256_castsi128_si256(
+        _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 6)));
+    srcReg56 =
+        _mm256_inserti128_si256(srcReg5x, _mm256_castsi256_si128(srcReg6x), 1);
+
+    // merge every two consecutive registers
+    srcReg45_56_lo = _mm256_unpacklo_epi8(srcReg45, srcReg56);
+
+    srcReg2345_3456_lo = _mm256_unpacklo_epi16(srcReg23_34_lo, srcReg45_56_lo);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    resReglo = _mm256_maddubs_epi16(srcReg2345_3456_lo, firstFilters);
+
+    resReglo = _mm256_hadds_epi16(resReglo, _mm256_setzero_si256());
+
+    // shift by 6 bit each 16 bit
+    resReglo = _mm256_adds_epi16(resReglo, addFilterReg32);
+    resReglo = _mm256_srai_epi16(resReglo, 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
+    resReg = _mm256_packus_epi16(resReglo, resReglo);
+
+    src_ptr += src_stride;
+
+    xx_storeu2_epi32(output_ptr, out_pitch, &resReg);
+
+    output_ptr += dst_stride;
+
+    // save part of the registers for next strides
+    srcReg23_34_lo = srcReg45_56_lo;
+    srcReg4x = srcReg6x;
+  }
+}
+
+#if HAVE_AVX2 && HAVE_SSSE3
+filter8_1dfunction aom_filter_block1d4_v8_ssse3;
+filter8_1dfunction aom_filter_block1d16_v2_ssse3;
+filter8_1dfunction aom_filter_block1d16_h2_ssse3;
+filter8_1dfunction aom_filter_block1d8_v2_ssse3;
+filter8_1dfunction aom_filter_block1d8_h2_ssse3;
+filter8_1dfunction aom_filter_block1d4_v2_ssse3;
+filter8_1dfunction aom_filter_block1d4_h2_ssse3;
+#define aom_filter_block1d4_v8_avx2 aom_filter_block1d4_v8_ssse3
+#define aom_filter_block1d16_v2_avx2 aom_filter_block1d16_v2_ssse3
+#define aom_filter_block1d16_h2_avx2 aom_filter_block1d16_h2_ssse3
+#define aom_filter_block1d8_v2_avx2 aom_filter_block1d8_v2_ssse3
+#define aom_filter_block1d8_h2_avx2 aom_filter_block1d8_h2_ssse3
+#define aom_filter_block1d4_v2_avx2 aom_filter_block1d4_v2_ssse3
+#define aom_filter_block1d4_h2_avx2 aom_filter_block1d4_h2_ssse3
+// void aom_convolve8_horiz_avx2(const uint8_t *src, ptrdiff_t src_stride,
+//                                uint8_t *dst, ptrdiff_t dst_stride,
+//                                const int16_t *filter_x, int x_step_q4,
+//                                const int16_t *filter_y, int y_step_q4,
+//                                int w, int h);
+// void aom_convolve8_vert_avx2(const uint8_t *src, ptrdiff_t src_stride,
+//                               uint8_t *dst, ptrdiff_t dst_stride,
+//                               const int16_t *filter_x, int x_step_q4,
+//                               const int16_t *filter_y, int y_step_q4,
+//                               int w, int h);
+FUN_CONV_1D(horiz, x_step_q4, filter_x, h, src, , avx2)
+FUN_CONV_1D(vert, y_step_q4, filter_y, v, src - src_stride * 3, , avx2)
+
+#endif  // HAVE_AX2 && HAVE_SSSE3
diff --git a/third_party/aom/aom_dsp/x86/aom_subpixel_8t_intrin_sse2.c b/third_party/aom/aom_dsp/x86/aom_subpixel_8t_intrin_sse2.c
new file mode 100644
index 0000000000..5c36b68727
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/aom_subpixel_8t_intrin_sse2.c
@@ -0,0 +1,569 @@
+/*
+ * 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 <emmintrin.h>  // SSE2
+
+#include "config/aom_dsp_rtcd.h"
+#include "aom_dsp/x86/convolve.h"
+#include "aom_ports/mem.h"
+
+void aom_filter_block1d16_h4_sse2(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;
+  __m128i addFilterReg32;
+  __m128i secondFilters, thirdFilters;
+  __m128i srcRegFilt32b1_1, srcRegFilt32b1_2, srcRegFilt32b2_1,
+      srcRegFilt32b2_2;
+  __m128i srcReg32b1, srcReg32b2;
+  unsigned int i;
+  src_ptr -= 3;
+  addFilterReg32 = _mm_set1_epi16(32);
+  filtersReg = _mm_loadu_si128((const __m128i *)filter);
+  filtersReg = _mm_srai_epi16(filtersReg, 1);
+
+  // coeffs 0 1 0 1 2 3 2 3
+  const __m128i tmp_0 = _mm_unpacklo_epi32(filtersReg, filtersReg);
+  // coeffs 4 5 4 5 6 7 6 7
+  const __m128i tmp_1 = _mm_unpackhi_epi32(filtersReg, filtersReg);
+
+  secondFilters = _mm_unpackhi_epi64(tmp_0, tmp_0);  // coeffs 2 3 2 3 2 3 2 3
+  thirdFilters = _mm_unpacklo_epi64(tmp_1, tmp_1);   // coeffs 4 5 4 5 4 5 4 5
+
+  for (i = output_height; i > 0; i -= 1) {
+    srcReg32b1 = _mm_loadu_si128((const __m128i *)src_ptr);
+
+    __m128i ss_2 = _mm_srli_si128(srcReg32b1, 2);
+    __m128i ss_4 = _mm_srli_si128(srcReg32b1, 4);
+    __m128i ss_1_1 = _mm_unpacklo_epi8(ss_2, _mm_setzero_si128());
+    __m128i ss_2_1 = _mm_unpacklo_epi8(ss_4, _mm_setzero_si128());
+    __m128i d1 = _mm_madd_epi16(ss_1_1, secondFilters);
+    __m128i d2 = _mm_madd_epi16(ss_2_1, thirdFilters);
+    srcRegFilt32b1_1 = _mm_add_epi32(d1, d2);
+
+    __m128i ss_1 = _mm_srli_si128(srcReg32b1, 3);
+    __m128i ss_3 = _mm_srli_si128(srcReg32b1, 5);
+    __m128i ss_1_2 = _mm_unpacklo_epi8(ss_1, _mm_setzero_si128());
+    __m128i ss_2_2 = _mm_unpacklo_epi8(ss_3, _mm_setzero_si128());
+    d1 = _mm_madd_epi16(ss_1_2, secondFilters);
+    d2 = _mm_madd_epi16(ss_2_2, thirdFilters);
+    srcRegFilt32b1_2 = _mm_add_epi32(d1, d2);
+
+    __m128i res_lo = _mm_unpacklo_epi32(srcRegFilt32b1_1, srcRegFilt32b1_2);
+    __m128i res_hi = _mm_unpackhi_epi32(srcRegFilt32b1_1, srcRegFilt32b1_2);
+    srcRegFilt32b1_1 = _mm_packs_epi32(res_lo, res_hi);
+
+    // reading stride of the next 16 bytes
+    // (part of it was being read by earlier read)
+    srcReg32b2 = _mm_loadu_si128((const __m128i *)(src_ptr + 8));
+
+    ss_2 = _mm_srli_si128(srcReg32b2, 2);
+    ss_4 = _mm_srli_si128(srcReg32b2, 4);
+    ss_1_1 = _mm_unpacklo_epi8(ss_2, _mm_setzero_si128());
+    ss_2_1 = _mm_unpacklo_epi8(ss_4, _mm_setzero_si128());
+    d1 = _mm_madd_epi16(ss_1_1, secondFilters);
+    d2 = _mm_madd_epi16(ss_2_1, thirdFilters);
+    srcRegFilt32b2_1 = _mm_add_epi32(d1, d2);
+
+    ss_1 = _mm_srli_si128(srcReg32b2, 3);
+    ss_3 = _mm_srli_si128(srcReg32b2, 5);
+    ss_1_2 = _mm_unpacklo_epi8(ss_1, _mm_setzero_si128());
+    ss_2_2 = _mm_unpacklo_epi8(ss_3, _mm_setzero_si128());
+    d1 = _mm_madd_epi16(ss_1_2, secondFilters);
+    d2 = _mm_madd_epi16(ss_2_2, thirdFilters);
+    srcRegFilt32b2_2 = _mm_add_epi32(d1, d2);
+
+    res_lo = _mm_unpacklo_epi32(srcRegFilt32b2_1, srcRegFilt32b2_2);
+    res_hi = _mm_unpackhi_epi32(srcRegFilt32b2_1, srcRegFilt32b2_2);
+    srcRegFilt32b2_1 = _mm_packs_epi32(res_lo, res_hi);
+
+    // shift by 6 bit each 16 bit
+    srcRegFilt32b1_1 = _mm_adds_epi16(srcRegFilt32b1_1, addFilterReg32);
+    srcRegFilt32b2_1 = _mm_adds_epi16(srcRegFilt32b2_1, addFilterReg32);
+    srcRegFilt32b1_1 = _mm_srai_epi16(srcRegFilt32b1_1, 6);
+    srcRegFilt32b2_1 = _mm_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
+    srcRegFilt32b1_1 = _mm_packus_epi16(srcRegFilt32b1_1, srcRegFilt32b2_1);
+
+    src_ptr += src_pixels_per_line;
+
+    _mm_store_si128((__m128i *)output_ptr, srcRegFilt32b1_1);
+
+    output_ptr += output_pitch;
+  }
+}
+
+void aom_filter_block1d16_v4_sse2(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;
+  __m128i srcReg2, srcReg3, srcReg4, srcReg5, srcReg6;
+  __m128i srcReg23_lo, srcReg23_hi, srcReg34_lo, srcReg34_hi;
+  __m128i srcReg45_lo, srcReg45_hi, srcReg56_lo, srcReg56_hi;
+  __m128i resReg23_lo, resReg34_lo, resReg45_lo, resReg56_lo;
+  __m128i resReg23_hi, resReg34_hi, resReg45_hi, resReg56_hi;
+  __m128i resReg23_45_lo, resReg34_56_lo, resReg23_45_hi, resReg34_56_hi;
+  __m128i resReg23_45, resReg34_56;
+  __m128i addFilterReg32, secondFilters, thirdFilters;
+  __m128i tmp_0, tmp_1;
+  unsigned int i;
+  ptrdiff_t src_stride, dst_stride;
+
+  addFilterReg32 = _mm_set1_epi16(32);
+  filtersReg = _mm_loadu_si128((const __m128i *)filter);
+  filtersReg = _mm_srai_epi16(filtersReg, 1);
+
+  // coeffs 0 1 0 1 2 3 2 3
+  const __m128i tmp0 = _mm_unpacklo_epi32(filtersReg, filtersReg);
+  // coeffs 4 5 4 5 6 7 6 7
+  const __m128i tmp1 = _mm_unpackhi_epi32(filtersReg, filtersReg);
+
+  secondFilters = _mm_unpackhi_epi64(tmp0, tmp0);  // coeffs 2 3 2 3 2 3 2 3
+  thirdFilters = _mm_unpacklo_epi64(tmp1, tmp1);   // coeffs 4 5 4 5 4 5 4 5
+
+  // multiply the size of the source and destination stride by two
+  src_stride = src_pitch << 1;
+  dst_stride = out_pitch << 1;
+
+  srcReg2 = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 2));
+  srcReg3 = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 3));
+  srcReg23_lo = _mm_unpacklo_epi8(srcReg2, srcReg3);
+  srcReg23_hi = _mm_unpackhi_epi8(srcReg2, srcReg3);
+  __m128i resReg23_lo_1 = _mm_unpacklo_epi8(srcReg23_lo, _mm_setzero_si128());
+  __m128i resReg23_lo_2 = _mm_unpackhi_epi8(srcReg23_lo, _mm_setzero_si128());
+  __m128i resReg23_hi_1 = _mm_unpacklo_epi8(srcReg23_hi, _mm_setzero_si128());
+  __m128i resReg23_hi_2 = _mm_unpackhi_epi8(srcReg23_hi, _mm_setzero_si128());
+
+  srcReg4 = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 4));
+  srcReg34_lo = _mm_unpacklo_epi8(srcReg3, srcReg4);
+  srcReg34_hi = _mm_unpackhi_epi8(srcReg3, srcReg4);
+  __m128i resReg34_lo_1 = _mm_unpacklo_epi8(srcReg34_lo, _mm_setzero_si128());
+  __m128i resReg34_lo_2 = _mm_unpackhi_epi8(srcReg34_lo, _mm_setzero_si128());
+  __m128i resReg34_hi_1 = _mm_unpacklo_epi8(srcReg34_hi, _mm_setzero_si128());
+  __m128i resReg34_hi_2 = _mm_unpackhi_epi8(srcReg34_hi, _mm_setzero_si128());
+
+  for (i = output_height; i > 1; i -= 2) {
+    srcReg5 = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 5));
+
+    srcReg45_lo = _mm_unpacklo_epi8(srcReg4, srcReg5);
+    srcReg45_hi = _mm_unpackhi_epi8(srcReg4, srcReg5);
+
+    srcReg6 = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 6));
+
+    srcReg56_lo = _mm_unpacklo_epi8(srcReg5, srcReg6);
+    srcReg56_hi = _mm_unpackhi_epi8(srcReg5, srcReg6);
+
+    // multiply 2 adjacent elements with the filter and add the result
+
+    tmp_0 = _mm_madd_epi16(resReg23_lo_1, secondFilters);
+    tmp_1 = _mm_madd_epi16(resReg23_lo_2, secondFilters);
+    resReg23_lo = _mm_packs_epi32(tmp_0, tmp_1);
+
+    tmp_0 = _mm_madd_epi16(resReg34_lo_1, secondFilters);
+    tmp_1 = _mm_madd_epi16(resReg34_lo_2, secondFilters);
+    resReg34_lo = _mm_packs_epi32(tmp_0, tmp_1);
+
+    __m128i resReg45_lo_1 = _mm_unpacklo_epi8(srcReg45_lo, _mm_setzero_si128());
+    __m128i resReg45_lo_2 = _mm_unpackhi_epi8(srcReg45_lo, _mm_setzero_si128());
+    tmp_0 = _mm_madd_epi16(resReg45_lo_1, thirdFilters);
+    tmp_1 = _mm_madd_epi16(resReg45_lo_2, thirdFilters);
+    resReg45_lo = _mm_packs_epi32(tmp_0, tmp_1);
+
+    __m128i resReg56_lo_1 = _mm_unpacklo_epi8(srcReg56_lo, _mm_setzero_si128());
+    __m128i resReg56_lo_2 = _mm_unpackhi_epi8(srcReg56_lo, _mm_setzero_si128());
+    tmp_0 = _mm_madd_epi16(resReg56_lo_1, thirdFilters);
+    tmp_1 = _mm_madd_epi16(resReg56_lo_2, thirdFilters);
+    resReg56_lo = _mm_packs_epi32(tmp_0, tmp_1);
+
+    // add and saturate the results together
+    resReg23_45_lo = _mm_adds_epi16(resReg23_lo, resReg45_lo);
+    resReg34_56_lo = _mm_adds_epi16(resReg34_lo, resReg56_lo);
+
+    // multiply 2 adjacent elements with the filter and add the result
+
+    tmp_0 = _mm_madd_epi16(resReg23_hi_1, secondFilters);
+    tmp_1 = _mm_madd_epi16(resReg23_hi_2, secondFilters);
+    resReg23_hi = _mm_packs_epi32(tmp_0, tmp_1);
+
+    tmp_0 = _mm_madd_epi16(resReg34_hi_1, secondFilters);
+    tmp_1 = _mm_madd_epi16(resReg34_hi_2, secondFilters);
+    resReg34_hi = _mm_packs_epi32(tmp_0, tmp_1);
+
+    __m128i resReg45_hi_1 = _mm_unpacklo_epi8(srcReg45_hi, _mm_setzero_si128());
+    __m128i resReg45_hi_2 = _mm_unpackhi_epi8(srcReg45_hi, _mm_setzero_si128());
+    tmp_0 = _mm_madd_epi16(resReg45_hi_1, thirdFilters);
+    tmp_1 = _mm_madd_epi16(resReg45_hi_2, thirdFilters);
+    resReg45_hi = _mm_packs_epi32(tmp_0, tmp_1);
+
+    __m128i resReg56_hi_1 = _mm_unpacklo_epi8(srcReg56_hi, _mm_setzero_si128());
+    __m128i resReg56_hi_2 = _mm_unpackhi_epi8(srcReg56_hi, _mm_setzero_si128());
+    tmp_0 = _mm_madd_epi16(resReg56_hi_1, thirdFilters);
+    tmp_1 = _mm_madd_epi16(resReg56_hi_2, thirdFilters);
+    resReg56_hi = _mm_packs_epi32(tmp_0, tmp_1);
+
+    // add and saturate the results together
+    resReg23_45_hi = _mm_adds_epi16(resReg23_hi, resReg45_hi);
+    resReg34_56_hi = _mm_adds_epi16(resReg34_hi, resReg56_hi);
+
+    // shift by 6 bit each 16 bit
+    resReg23_45_lo = _mm_adds_epi16(resReg23_45_lo, addFilterReg32);
+    resReg34_56_lo = _mm_adds_epi16(resReg34_56_lo, addFilterReg32);
+    resReg23_45_hi = _mm_adds_epi16(resReg23_45_hi, addFilterReg32);
+    resReg34_56_hi = _mm_adds_epi16(resReg34_56_hi, addFilterReg32);
+    resReg23_45_lo = _mm_srai_epi16(resReg23_45_lo, 6);
+    resReg34_56_lo = _mm_srai_epi16(resReg34_56_lo, 6);
+    resReg23_45_hi = _mm_srai_epi16(resReg23_45_hi, 6);
+    resReg34_56_hi = _mm_srai_epi16(resReg34_56_hi, 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
+    resReg23_45 = _mm_packus_epi16(resReg23_45_lo, resReg23_45_hi);
+    resReg34_56 = _mm_packus_epi16(resReg34_56_lo, resReg34_56_hi);
+
+    src_ptr += src_stride;
+
+    _mm_store_si128((__m128i *)output_ptr, (resReg23_45));
+    _mm_store_si128((__m128i *)(output_ptr + out_pitch), (resReg34_56));
+
+    output_ptr += dst_stride;
+
+    // save part of the registers for next strides
+    resReg23_lo_1 = resReg45_lo_1;
+    resReg23_lo_2 = resReg45_lo_2;
+    resReg23_hi_1 = resReg45_hi_1;
+    resReg23_hi_2 = resReg45_hi_2;
+    resReg34_lo_1 = resReg56_lo_1;
+    resReg34_lo_2 = resReg56_lo_2;
+    resReg34_hi_1 = resReg56_hi_1;
+    resReg34_hi_2 = resReg56_hi_2;
+    srcReg4 = srcReg6;
+  }
+}
+
+void aom_filter_block1d8_h4_sse2(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;
+  __m128i addFilterReg32;
+  __m128i secondFilters, thirdFilters;
+  __m128i srcRegFilt32b1_1, srcRegFilt32b1_2;
+  __m128i srcReg32b1;
+  unsigned int i;
+  src_ptr -= 3;
+  addFilterReg32 = _mm_set1_epi16(32);
+  filtersReg = _mm_loadu_si128((const __m128i *)filter);
+  filtersReg = _mm_srai_epi16(filtersReg, 1);
+
+  // coeffs 0 1 0 1 2 3 2 3
+  const __m128i tmp_0 = _mm_unpacklo_epi32(filtersReg, filtersReg);
+  // coeffs 4 5 4 5 6 7 6 7
+  const __m128i tmp_1 = _mm_unpackhi_epi32(filtersReg, filtersReg);
+
+  secondFilters = _mm_unpackhi_epi64(tmp_0, tmp_0);  // coeffs 2 3 2 3 2 3 2 3
+  thirdFilters = _mm_unpacklo_epi64(tmp_1, tmp_1);   // coeffs 4 5 4 5 4 5 4 5
+
+  for (i = output_height; i > 0; i -= 1) {
+    srcReg32b1 = _mm_loadu_si128((const __m128i *)src_ptr);
+
+    __m128i ss_2 = _mm_srli_si128(srcReg32b1, 2);
+    __m128i ss_4 = _mm_srli_si128(srcReg32b1, 4);
+    ss_2 = _mm_unpacklo_epi8(ss_2, _mm_setzero_si128());
+    ss_4 = _mm_unpacklo_epi8(ss_4, _mm_setzero_si128());
+    __m128i d1 = _mm_madd_epi16(ss_2, secondFilters);
+    __m128i d2 = _mm_madd_epi16(ss_4, thirdFilters);
+    srcRegFilt32b1_1 = _mm_add_epi32(d1, d2);
+
+    __m128i ss_3 = _mm_srli_si128(srcReg32b1, 3);
+    __m128i ss_5 = _mm_srli_si128(srcReg32b1, 5);
+    ss_3 = _mm_unpacklo_epi8(ss_3, _mm_setzero_si128());
+    ss_5 = _mm_unpacklo_epi8(ss_5, _mm_setzero_si128());
+    d1 = _mm_madd_epi16(ss_3, secondFilters);
+    d2 = _mm_madd_epi16(ss_5, thirdFilters);
+    srcRegFilt32b1_2 = _mm_add_epi32(d1, d2);
+
+    __m128i res_lo = _mm_unpacklo_epi32(srcRegFilt32b1_1, srcRegFilt32b1_2);
+    __m128i res_hi = _mm_unpackhi_epi32(srcRegFilt32b1_1, srcRegFilt32b1_2);
+    srcRegFilt32b1_1 = _mm_packs_epi32(res_lo, res_hi);
+
+    // shift by 6 bit each 16 bit
+    srcRegFilt32b1_1 = _mm_adds_epi16(srcRegFilt32b1_1, addFilterReg32);
+    srcRegFilt32b1_1 = _mm_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 = _mm_packus_epi16(srcRegFilt32b1_1, _mm_setzero_si128());
+
+    src_ptr += src_pixels_per_line;
+
+    _mm_storel_epi64((__m128i *)output_ptr, srcRegFilt32b1_1);
+
+    output_ptr += output_pitch;
+  }
+}
+
+void aom_filter_block1d8_v4_sse2(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;
+  __m128i srcReg2, srcReg3, srcReg4, srcReg5, srcReg6;
+  __m128i srcReg23_lo, srcReg34_lo;
+  __m128i srcReg45_lo, srcReg56_lo;
+  __m128i resReg23_lo, resReg34_lo, resReg45_lo, resReg56_lo;
+  __m128i resReg23_45_lo, resReg34_56_lo;
+  __m128i resReg23_45, resReg34_56;
+  __m128i addFilterReg32, secondFilters, thirdFilters;
+  __m128i tmp_0, tmp_1;
+  unsigned int i;
+  ptrdiff_t src_stride, dst_stride;
+
+  addFilterReg32 = _mm_set1_epi16(32);
+  filtersReg = _mm_loadu_si128((const __m128i *)filter);
+  filtersReg = _mm_srai_epi16(filtersReg, 1);
+
+  // coeffs 0 1 0 1 2 3 2 3
+  const __m128i tmp0 = _mm_unpacklo_epi32(filtersReg, filtersReg);
+  // coeffs 4 5 4 5 6 7 6 7
+  const __m128i tmp1 = _mm_unpackhi_epi32(filtersReg, filtersReg);
+
+  secondFilters = _mm_unpackhi_epi64(tmp0, tmp0);  // coeffs 2 3 2 3 2 3 2 3
+  thirdFilters = _mm_unpacklo_epi64(tmp1, tmp1);   // coeffs 4 5 4 5 4 5 4 5
+
+  // multiply the size of the source and destination stride by two
+  src_stride = src_pitch << 1;
+  dst_stride = out_pitch << 1;
+
+  srcReg2 = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 2));
+  srcReg3 = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 3));
+  srcReg23_lo = _mm_unpacklo_epi8(srcReg2, srcReg3);
+  __m128i resReg23_lo_1 = _mm_unpacklo_epi8(srcReg23_lo, _mm_setzero_si128());
+  __m128i resReg23_lo_2 = _mm_unpackhi_epi8(srcReg23_lo, _mm_setzero_si128());
+
+  srcReg4 = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 4));
+  srcReg34_lo = _mm_unpacklo_epi8(srcReg3, srcReg4);
+  __m128i resReg34_lo_1 = _mm_unpacklo_epi8(srcReg34_lo, _mm_setzero_si128());
+  __m128i resReg34_lo_2 = _mm_unpackhi_epi8(srcReg34_lo, _mm_setzero_si128());
+
+  for (i = output_height; i > 1; i -= 2) {
+    srcReg5 = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 5));
+    srcReg45_lo = _mm_unpacklo_epi8(srcReg4, srcReg5);
+
+    srcReg6 = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 6));
+    srcReg56_lo = _mm_unpacklo_epi8(srcReg5, srcReg6);
+
+    // multiply 2 adjacent elements with the filter and add the result
+
+    tmp_0 = _mm_madd_epi16(resReg23_lo_1, secondFilters);
+    tmp_1 = _mm_madd_epi16(resReg23_lo_2, secondFilters);
+    resReg23_lo = _mm_packs_epi32(tmp_0, tmp_1);
+
+    tmp_0 = _mm_madd_epi16(resReg34_lo_1, secondFilters);
+    tmp_1 = _mm_madd_epi16(resReg34_lo_2, secondFilters);
+    resReg34_lo = _mm_packs_epi32(tmp_0, tmp_1);
+
+    __m128i resReg45_lo_1 = _mm_unpacklo_epi8(srcReg45_lo, _mm_setzero_si128());
+    __m128i resReg45_lo_2 = _mm_unpackhi_epi8(srcReg45_lo, _mm_setzero_si128());
+    tmp_0 = _mm_madd_epi16(resReg45_lo_1, thirdFilters);
+    tmp_1 = _mm_madd_epi16(resReg45_lo_2, thirdFilters);
+    resReg45_lo = _mm_packs_epi32(tmp_0, tmp_1);
+
+    __m128i resReg56_lo_1 = _mm_unpacklo_epi8(srcReg56_lo, _mm_setzero_si128());
+    __m128i resReg56_lo_2 = _mm_unpackhi_epi8(srcReg56_lo, _mm_setzero_si128());
+    tmp_0 = _mm_madd_epi16(resReg56_lo_1, thirdFilters);
+    tmp_1 = _mm_madd_epi16(resReg56_lo_2, thirdFilters);
+    resReg56_lo = _mm_packs_epi32(tmp_0, tmp_1);
+
+    // add and saturate the results together
+    resReg23_45_lo = _mm_adds_epi16(resReg23_lo, resReg45_lo);
+    resReg34_56_lo = _mm_adds_epi16(resReg34_lo, resReg56_lo);
+
+    // shift by 6 bit each 16 bit
+    resReg23_45_lo = _mm_adds_epi16(resReg23_45_lo, addFilterReg32);
+    resReg34_56_lo = _mm_adds_epi16(resReg34_56_lo, addFilterReg32);
+    resReg23_45_lo = _mm_srai_epi16(resReg23_45_lo, 6);
+    resReg34_56_lo = _mm_srai_epi16(resReg34_56_lo, 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
+    resReg23_45 = _mm_packus_epi16(resReg23_45_lo, _mm_setzero_si128());
+    resReg34_56 = _mm_packus_epi16(resReg34_56_lo, _mm_setzero_si128());
+
+    src_ptr += src_stride;
+
+    _mm_storel_epi64((__m128i *)output_ptr, (resReg23_45));
+    _mm_storel_epi64((__m128i *)(output_ptr + out_pitch), (resReg34_56));
+
+    output_ptr += dst_stride;
+
+    // save part of the registers for next strides
+    resReg23_lo_1 = resReg45_lo_1;
+    resReg23_lo_2 = resReg45_lo_2;
+    resReg34_lo_1 = resReg56_lo_1;
+    resReg34_lo_2 = resReg56_lo_2;
+    srcReg4 = srcReg6;
+  }
+}
+
+void aom_filter_block1d4_h4_sse2(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;
+  __m128i addFilterReg32;
+  __m128i secondFilters, thirdFilters;
+  __m128i srcRegFilt32b1_1;
+  __m128i srcReg32b1;
+  unsigned int i;
+  src_ptr -= 3;
+  addFilterReg32 = _mm_set1_epi16(32);
+  filtersReg = _mm_loadu_si128((const __m128i *)filter);
+  filtersReg = _mm_srai_epi16(filtersReg, 1);
+
+  // coeffs 0 1 0 1 2 3 2 3
+  const __m128i tmp_0 = _mm_unpacklo_epi32(filtersReg, filtersReg);
+  // coeffs 4 5 4 5 6 7 6 7
+  const __m128i tmp_1 = _mm_unpackhi_epi32(filtersReg, filtersReg);
+
+  secondFilters = _mm_unpackhi_epi64(tmp_0, tmp_0);  // coeffs 2 3 2 3 2 3 2 3
+  thirdFilters = _mm_unpacklo_epi64(tmp_1, tmp_1);   // coeffs 4 5 4 5 4 5 4 5
+
+  for (i = output_height; i > 0; i -= 1) {
+    srcReg32b1 = _mm_loadu_si128((const __m128i *)src_ptr);
+
+    __m128i ss_2 = _mm_srli_si128(srcReg32b1, 2);
+    __m128i ss_3 = _mm_srli_si128(srcReg32b1, 3);
+    __m128i ss_4 = _mm_srli_si128(srcReg32b1, 4);
+    __m128i ss_5 = _mm_srli_si128(srcReg32b1, 5);
+
+    ss_2 = _mm_unpacklo_epi8(ss_2, _mm_setzero_si128());
+    ss_3 = _mm_unpacklo_epi8(ss_3, _mm_setzero_si128());
+    ss_4 = _mm_unpacklo_epi8(ss_4, _mm_setzero_si128());
+    ss_5 = _mm_unpacklo_epi8(ss_5, _mm_setzero_si128());
+
+    __m128i ss_1_1 = _mm_unpacklo_epi32(ss_2, ss_3);
+    __m128i ss_1_2 = _mm_unpacklo_epi32(ss_4, ss_5);
+
+    __m128i d1 = _mm_madd_epi16(ss_1_1, secondFilters);
+    __m128i d2 = _mm_madd_epi16(ss_1_2, thirdFilters);
+    srcRegFilt32b1_1 = _mm_add_epi32(d1, d2);
+
+    srcRegFilt32b1_1 = _mm_packs_epi32(srcRegFilt32b1_1, _mm_setzero_si128());
+
+    // shift by 6 bit each 16 bit
+    srcRegFilt32b1_1 = _mm_adds_epi16(srcRegFilt32b1_1, addFilterReg32);
+    srcRegFilt32b1_1 = _mm_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 = _mm_packus_epi16(srcRegFilt32b1_1, _mm_setzero_si128());
+
+    src_ptr += src_pixels_per_line;
+
+    *((int *)(output_ptr)) = _mm_cvtsi128_si32(srcRegFilt32b1_1);
+
+    output_ptr += output_pitch;
+  }
+}
+
+void aom_filter_block1d4_v4_sse2(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;
+  __m128i srcReg2, srcReg3, srcReg4, srcReg5, srcReg6;
+  __m128i srcReg23, srcReg34, srcReg45, srcReg56;
+  __m128i resReg23_34, resReg45_56;
+  __m128i resReg23_34_45_56;
+  __m128i addFilterReg32, secondFilters, thirdFilters;
+  __m128i tmp_0, tmp_1;
+  unsigned int i;
+  ptrdiff_t src_stride, dst_stride;
+
+  addFilterReg32 = _mm_set1_epi16(32);
+  filtersReg = _mm_loadu_si128((const __m128i *)filter);
+  filtersReg = _mm_srai_epi16(filtersReg, 1);
+
+  // coeffs 0 1 0 1 2 3 2 3
+  const __m128i tmp0 = _mm_unpacklo_epi32(filtersReg, filtersReg);
+  // coeffs 4 5 4 5 6 7 6 7
+  const __m128i tmp1 = _mm_unpackhi_epi32(filtersReg, filtersReg);
+
+  secondFilters = _mm_unpackhi_epi64(tmp0, tmp0);  // coeffs 2 3 2 3 2 3 2 3
+  thirdFilters = _mm_unpacklo_epi64(tmp1, tmp1);   // coeffs 4 5 4 5 4 5 4 5
+
+  // multiply the size of the source and destination stride by two
+  src_stride = src_pitch << 1;
+  dst_stride = out_pitch << 1;
+
+  srcReg2 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 2));
+  srcReg3 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 3));
+  srcReg23 = _mm_unpacklo_epi8(srcReg2, srcReg3);
+  __m128i resReg23 = _mm_unpacklo_epi8(srcReg23, _mm_setzero_si128());
+
+  srcReg4 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 4));
+  srcReg34 = _mm_unpacklo_epi8(srcReg3, srcReg4);
+  __m128i resReg34 = _mm_unpacklo_epi8(srcReg34, _mm_setzero_si128());
+
+  for (i = output_height; i > 1; i -= 2) {
+    srcReg5 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 5));
+    srcReg45 = _mm_unpacklo_epi8(srcReg4, srcReg5);
+    srcReg6 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 6));
+    srcReg56 = _mm_unpacklo_epi8(srcReg5, srcReg6);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    tmp_0 = _mm_madd_epi16(resReg23, secondFilters);
+    tmp_1 = _mm_madd_epi16(resReg34, secondFilters);
+    resReg23_34 = _mm_packs_epi32(tmp_0, tmp_1);
+
+    __m128i resReg45 = _mm_unpacklo_epi8(srcReg45, _mm_setzero_si128());
+    __m128i resReg56 = _mm_unpacklo_epi8(srcReg56, _mm_setzero_si128());
+
+    tmp_0 = _mm_madd_epi16(resReg45, thirdFilters);
+    tmp_1 = _mm_madd_epi16(resReg56, thirdFilters);
+    resReg45_56 = _mm_packs_epi32(tmp_0, tmp_1);
+
+    // add and saturate the results together
+    resReg23_34_45_56 = _mm_adds_epi16(resReg23_34, resReg45_56);
+
+    // shift by 6 bit each 16 bit
+    resReg23_34_45_56 = _mm_adds_epi16(resReg23_34_45_56, addFilterReg32);
+    resReg23_34_45_56 = _mm_srai_epi16(resReg23_34_45_56, 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
+    resReg23_34_45_56 =
+        _mm_packus_epi16(resReg23_34_45_56, _mm_setzero_si128());
+
+    src_ptr += src_stride;
+
+    *((int *)(output_ptr)) = _mm_cvtsi128_si32(resReg23_34_45_56);
+    *((int *)(output_ptr + out_pitch)) =
+        _mm_cvtsi128_si32(_mm_srli_si128(resReg23_34_45_56, 4));
+
+    output_ptr += dst_stride;
+
+    // save part of the registers for next strides
+    resReg23 = resReg45;
+    resReg34 = resReg56;
+    srcReg4 = srcReg6;
+  }
+}
diff --git a/third_party/aom/aom_dsp/x86/aom_subpixel_8t_intrin_ssse3.c b/third_party/aom/aom_dsp/x86/aom_subpixel_8t_intrin_ssse3.c
new file mode 100644
index 0000000000..245fda1e94
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/aom_subpixel_8t_intrin_ssse3.c
@@ -0,0 +1,847 @@
+/*
+ * Copyright (c) 2016, 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_dsp_rtcd.h"
+
+#include "aom_dsp/aom_filter.h"
+#include "aom_dsp/x86/convolve.h"
+#include "aom_dsp/x86/convolve_sse2.h"
+#include "aom_dsp/x86/convolve_ssse3.h"
+#include "aom_dsp/x86/mem_sse2.h"
+#include "aom_dsp/x86/transpose_sse2.h"
+#include "aom_mem/aom_mem.h"
+#include "aom_ports/mem.h"
+#include "aom_ports/emmintrin_compat.h"
+
+DECLARE_ALIGNED(32, static const uint8_t, filt_h4[]) = {
+  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, filtd4[]) = {
+  2, 3, 4, 5, 3, 4, 5, 6, 4, 5, 6, 7, 5, 6, 7, 8,
+  2, 3, 4, 5, 3, 4, 5, 6, 4, 5, 6, 7, 5, 6, 7, 8,
+};
+
+static void aom_filter_block1d4_h4_ssse3(
+    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;
+  __m128i addFilterReg32, filt1Reg, firstFilters, srcReg32b1, srcRegFilt32b1_1;
+  unsigned int i;
+  src_ptr -= 3;
+  addFilterReg32 = _mm_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);
+
+  firstFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi32(0x5040302u));
+  filt1Reg = _mm_load_si128((__m128i const *)(filtd4));
+
+  for (i = output_height; i > 0; i -= 1) {
+    // load the 2 strides of source
+    srcReg32b1 = _mm_loadu_si128((const __m128i *)src_ptr);
+
+    // filter the source buffer
+    srcRegFilt32b1_1 = _mm_shuffle_epi8(srcReg32b1, filt1Reg);
+
+    // multiply 4 adjacent elements with the filter and add the result
+    srcRegFilt32b1_1 = _mm_maddubs_epi16(srcRegFilt32b1_1, firstFilters);
+
+    srcRegFilt32b1_1 = _mm_hadds_epi16(srcRegFilt32b1_1, _mm_setzero_si128());
+
+    // shift by 6 bit each 16 bit
+    srcRegFilt32b1_1 = _mm_adds_epi16(srcRegFilt32b1_1, addFilterReg32);
+    srcRegFilt32b1_1 = _mm_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 = _mm_packus_epi16(srcRegFilt32b1_1, _mm_setzero_si128());
+
+    src_ptr += src_pixels_per_line;
+
+    *((int *)(output_ptr)) = _mm_cvtsi128_si32(srcRegFilt32b1_1);
+    output_ptr += output_pitch;
+  }
+}
+
+static void aom_filter_block1d4_v4_ssse3(
+    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;
+  __m128i addFilterReg32;
+  __m128i srcReg2, srcReg3, srcReg23, srcReg4, srcReg34, srcReg5, srcReg45,
+      srcReg6, srcReg56;
+  __m128i srcReg23_34_lo, srcReg45_56_lo;
+  __m128i srcReg2345_3456_lo, srcReg2345_3456_hi;
+  __m128i resReglo, resReghi;
+  __m128i firstFilters;
+  unsigned int i;
+  ptrdiff_t src_stride, dst_stride;
+
+  addFilterReg32 = _mm_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);
+
+  firstFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi32(0x5040302u));
+
+  // multiple the size of the source and destination stride by two
+  src_stride = src_pitch << 1;
+  dst_stride = out_pitch << 1;
+
+  srcReg2 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 2));
+  srcReg3 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 3));
+  srcReg23 = _mm_unpacklo_epi32(srcReg2, srcReg3);
+
+  srcReg4 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 4));
+
+  // have consecutive loads on the same 256 register
+  srcReg34 = _mm_unpacklo_epi32(srcReg3, srcReg4);
+
+  srcReg23_34_lo = _mm_unpacklo_epi8(srcReg23, srcReg34);
+
+  for (i = output_height; i > 1; i -= 2) {
+    srcReg5 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 5));
+    srcReg45 = _mm_unpacklo_epi32(srcReg4, srcReg5);
+
+    srcReg6 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 6));
+    srcReg56 = _mm_unpacklo_epi32(srcReg5, srcReg6);
+
+    // merge every two consecutive registers
+    srcReg45_56_lo = _mm_unpacklo_epi8(srcReg45, srcReg56);
+
+    srcReg2345_3456_lo = _mm_unpacklo_epi16(srcReg23_34_lo, srcReg45_56_lo);
+    srcReg2345_3456_hi = _mm_unpackhi_epi16(srcReg23_34_lo, srcReg45_56_lo);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    resReglo = _mm_maddubs_epi16(srcReg2345_3456_lo, firstFilters);
+    resReghi = _mm_maddubs_epi16(srcReg2345_3456_hi, firstFilters);
+
+    resReglo = _mm_hadds_epi16(resReglo, _mm_setzero_si128());
+    resReghi = _mm_hadds_epi16(resReghi, _mm_setzero_si128());
+
+    // shift by 6 bit each 16 bit
+    resReglo = _mm_adds_epi16(resReglo, addFilterReg32);
+    resReghi = _mm_adds_epi16(resReghi, addFilterReg32);
+    resReglo = _mm_srai_epi16(resReglo, 6);
+    resReghi = _mm_srai_epi16(resReghi, 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
+    resReglo = _mm_packus_epi16(resReglo, resReglo);
+    resReghi = _mm_packus_epi16(resReghi, resReghi);
+
+    src_ptr += src_stride;
+
+    *((int *)(output_ptr)) = _mm_cvtsi128_si32(resReglo);
+    *((int *)(output_ptr + out_pitch)) = _mm_cvtsi128_si32(resReghi);
+
+    output_ptr += dst_stride;
+
+    // save part of the registers for next strides
+    srcReg23_34_lo = srcReg45_56_lo;
+    srcReg4 = srcReg6;
+  }
+}
+
+static void aom_filter_block1d8_h4_ssse3(
+    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;
+  __m128i addFilterReg32, filt2Reg, filt3Reg;
+  __m128i secondFilters, thirdFilters;
+  __m128i srcRegFilt32b1_1, srcRegFilt32b2, srcRegFilt32b3;
+  __m128i srcReg32b1;
+  unsigned int i;
+  src_ptr -= 3;
+  addFilterReg32 = _mm_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);
+
+  // duplicate only the second 16 bits (third and forth byte)
+  // across 256 bit register
+  secondFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x302u));
+  // duplicate only the third 16 bits (fifth and sixth byte)
+  // across 256 bit register
+  thirdFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x504u));
+
+  filt2Reg = _mm_load_si128((__m128i const *)(filt_h4 + 32));
+  filt3Reg = _mm_load_si128((__m128i const *)(filt_h4 + 32 * 2));
+
+  for (i = output_height; i > 0; i -= 1) {
+    srcReg32b1 = _mm_loadu_si128((const __m128i *)src_ptr);
+
+    // filter the source buffer
+    srcRegFilt32b3 = _mm_shuffle_epi8(srcReg32b1, filt2Reg);
+    srcRegFilt32b2 = _mm_shuffle_epi8(srcReg32b1, filt3Reg);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt32b3 = _mm_maddubs_epi16(srcRegFilt32b3, secondFilters);
+    srcRegFilt32b2 = _mm_maddubs_epi16(srcRegFilt32b2, thirdFilters);
+
+    srcRegFilt32b1_1 = _mm_adds_epi16(srcRegFilt32b3, srcRegFilt32b2);
+
+    // shift by 6 bit each 16 bit
+    srcRegFilt32b1_1 = _mm_adds_epi16(srcRegFilt32b1_1, addFilterReg32);
+    srcRegFilt32b1_1 = _mm_srai_epi16(srcRegFilt32b1_1, 6);
+
+    // shrink to 8 bit each 16 bits
+    srcRegFilt32b1_1 = _mm_packus_epi16(srcRegFilt32b1_1, _mm_setzero_si128());
+
+    src_ptr += src_pixels_per_line;
+
+    _mm_storel_epi64((__m128i *)output_ptr, srcRegFilt32b1_1);
+
+    output_ptr += output_pitch;
+  }
+}
+
+static void aom_filter_block1d8_v4_ssse3(
+    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;
+  __m128i srcReg2, srcReg3, srcReg4, srcReg5, srcReg6;
+  __m128i srcReg23, srcReg34, srcReg45, srcReg56;
+  __m128i resReg23, resReg34, resReg45, resReg56;
+  __m128i resReg23_45, resReg34_56;
+  __m128i addFilterReg32, secondFilters, thirdFilters;
+  unsigned int i;
+  ptrdiff_t src_stride, dst_stride;
+
+  addFilterReg32 = _mm_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);
+
+  // duplicate only the second 16 bits (third and forth byte)
+  // across 128 bit register
+  secondFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x302u));
+  // duplicate only the third 16 bits (fifth and sixth byte)
+  // across 128 bit register
+  thirdFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x504u));
+
+  // multiple the size of the source and destination stride by two
+  src_stride = src_pitch << 1;
+  dst_stride = out_pitch << 1;
+
+  srcReg2 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 2));
+  srcReg3 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 3));
+  srcReg23 = _mm_unpacklo_epi8(srcReg2, srcReg3);
+
+  srcReg4 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 4));
+
+  // have consecutive loads on the same 256 register
+  srcReg34 = _mm_unpacklo_epi8(srcReg3, srcReg4);
+
+  for (i = output_height; i > 1; i -= 2) {
+    srcReg5 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 5));
+
+    srcReg45 = _mm_unpacklo_epi8(srcReg4, srcReg5);
+
+    srcReg6 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 6));
+
+    srcReg56 = _mm_unpacklo_epi8(srcReg5, srcReg6);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    resReg23 = _mm_maddubs_epi16(srcReg23, secondFilters);
+    resReg34 = _mm_maddubs_epi16(srcReg34, secondFilters);
+    resReg45 = _mm_maddubs_epi16(srcReg45, thirdFilters);
+    resReg56 = _mm_maddubs_epi16(srcReg56, thirdFilters);
+
+    // add and saturate the results together
+    resReg23_45 = _mm_adds_epi16(resReg23, resReg45);
+    resReg34_56 = _mm_adds_epi16(resReg34, resReg56);
+
+    // shift by 6 bit each 16 bit
+    resReg23_45 = _mm_adds_epi16(resReg23_45, addFilterReg32);
+    resReg34_56 = _mm_adds_epi16(resReg34_56, addFilterReg32);
+    resReg23_45 = _mm_srai_epi16(resReg23_45, 6);
+    resReg34_56 = _mm_srai_epi16(resReg34_56, 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
+    resReg23_45 = _mm_packus_epi16(resReg23_45, _mm_setzero_si128());
+    resReg34_56 = _mm_packus_epi16(resReg34_56, _mm_setzero_si128());
+
+    src_ptr += src_stride;
+
+    _mm_storel_epi64((__m128i *)output_ptr, (resReg23_45));
+    _mm_storel_epi64((__m128i *)(output_ptr + out_pitch), (resReg34_56));
+
+    output_ptr += dst_stride;
+
+    // save part of the registers for next strides
+    srcReg23 = srcReg45;
+    srcReg34 = srcReg56;
+    srcReg4 = srcReg6;
+  }
+}
+
+static void aom_filter_block1d16_h4_ssse3(
+    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;
+  __m128i addFilterReg32, filt2Reg, filt3Reg;
+  __m128i secondFilters, thirdFilters;
+  __m128i srcRegFilt32b1_1, srcRegFilt32b2_1, srcRegFilt32b2, srcRegFilt32b3;
+  __m128i srcReg32b1, srcReg32b2;
+  unsigned int i;
+  src_ptr -= 3;
+  addFilterReg32 = _mm_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);
+
+  // duplicate only the second 16 bits (third and forth byte)
+  // across 256 bit register
+  secondFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x302u));
+  // duplicate only the third 16 bits (fifth and sixth byte)
+  // across 256 bit register
+  thirdFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x504u));
+
+  filt2Reg = _mm_load_si128((__m128i const *)(filt_h4 + 32));
+  filt3Reg = _mm_load_si128((__m128i const *)(filt_h4 + 32 * 2));
+
+  for (i = output_height; i > 0; i -= 1) {
+    srcReg32b1 = _mm_loadu_si128((const __m128i *)src_ptr);
+
+    // filter the source buffer
+    srcRegFilt32b3 = _mm_shuffle_epi8(srcReg32b1, filt2Reg);
+    srcRegFilt32b2 = _mm_shuffle_epi8(srcReg32b1, filt3Reg);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt32b3 = _mm_maddubs_epi16(srcRegFilt32b3, secondFilters);
+    srcRegFilt32b2 = _mm_maddubs_epi16(srcRegFilt32b2, thirdFilters);
+
+    srcRegFilt32b1_1 = _mm_adds_epi16(srcRegFilt32b3, srcRegFilt32b2);
+
+    // reading stride of the next 16 bytes
+    // (part of it was being read by earlier read)
+    srcReg32b2 = _mm_loadu_si128((const __m128i *)(src_ptr + 8));
+
+    // filter the source buffer
+    srcRegFilt32b3 = _mm_shuffle_epi8(srcReg32b2, filt2Reg);
+    srcRegFilt32b2 = _mm_shuffle_epi8(srcReg32b2, filt3Reg);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    srcRegFilt32b3 = _mm_maddubs_epi16(srcRegFilt32b3, secondFilters);
+    srcRegFilt32b2 = _mm_maddubs_epi16(srcRegFilt32b2, thirdFilters);
+
+    // add and saturate the results together
+    srcRegFilt32b2_1 = _mm_adds_epi16(srcRegFilt32b3, srcRegFilt32b2);
+
+    // shift by 6 bit each 16 bit
+    srcRegFilt32b1_1 = _mm_adds_epi16(srcRegFilt32b1_1, addFilterReg32);
+    srcRegFilt32b2_1 = _mm_adds_epi16(srcRegFilt32b2_1, addFilterReg32);
+    srcRegFilt32b1_1 = _mm_srai_epi16(srcRegFilt32b1_1, 6);
+    srcRegFilt32b2_1 = _mm_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
+    srcRegFilt32b1_1 = _mm_packus_epi16(srcRegFilt32b1_1, srcRegFilt32b2_1);
+
+    src_ptr += src_pixels_per_line;
+
+    _mm_store_si128((__m128i *)output_ptr, srcRegFilt32b1_1);
+
+    output_ptr += output_pitch;
+  }
+}
+
+static void aom_filter_block1d16_v4_ssse3(
+    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;
+  __m128i srcReg2, srcReg3, srcReg4, srcReg5, srcReg6;
+  __m128i srcReg23_lo, srcReg23_hi, srcReg34_lo, srcReg34_hi;
+  __m128i srcReg45_lo, srcReg45_hi, srcReg56_lo, srcReg56_hi;
+  __m128i resReg23_lo, resReg34_lo, resReg45_lo, resReg56_lo;
+  __m128i resReg23_hi, resReg34_hi, resReg45_hi, resReg56_hi;
+  __m128i resReg23_45_lo, resReg34_56_lo, resReg23_45_hi, resReg34_56_hi;
+  __m128i resReg23_45, resReg34_56;
+  __m128i addFilterReg32, secondFilters, thirdFilters;
+  unsigned int i;
+  ptrdiff_t src_stride, dst_stride;
+
+  addFilterReg32 = _mm_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);
+
+  // duplicate only the second 16 bits (third and forth byte)
+  // across 128 bit register
+  secondFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x302u));
+  // duplicate only the third 16 bits (fifth and sixth byte)
+  // across 128 bit register
+  thirdFilters = _mm_shuffle_epi8(filtersReg, _mm_set1_epi16(0x504u));
+
+  // multiple the size of the source and destination stride by two
+  src_stride = src_pitch << 1;
+  dst_stride = out_pitch << 1;
+
+  srcReg2 = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 2));
+  srcReg3 = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 3));
+  srcReg23_lo = _mm_unpacklo_epi8(srcReg2, srcReg3);
+  srcReg23_hi = _mm_unpackhi_epi8(srcReg2, srcReg3);
+
+  srcReg4 = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 4));
+
+  // have consecutive loads on the same 256 register
+  srcReg34_lo = _mm_unpacklo_epi8(srcReg3, srcReg4);
+  srcReg34_hi = _mm_unpackhi_epi8(srcReg3, srcReg4);
+
+  for (i = output_height; i > 1; i -= 2) {
+    srcReg5 = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 5));
+
+    srcReg45_lo = _mm_unpacklo_epi8(srcReg4, srcReg5);
+    srcReg45_hi = _mm_unpackhi_epi8(srcReg4, srcReg5);
+
+    srcReg6 = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 6));
+
+    srcReg56_lo = _mm_unpacklo_epi8(srcReg5, srcReg6);
+    srcReg56_hi = _mm_unpackhi_epi8(srcReg5, srcReg6);
+
+    // multiply 2 adjacent elements with the filter and add the result
+    resReg23_lo = _mm_maddubs_epi16(srcReg23_lo, secondFilters);
+    resReg34_lo = _mm_maddubs_epi16(srcReg34_lo, secondFilters);
+    resReg45_lo = _mm_maddubs_epi16(srcReg45_lo, thirdFilters);
+    resReg56_lo = _mm_maddubs_epi16(srcReg56_lo, thirdFilters);
+
+    // add and saturate the results together
+    resReg23_45_lo = _mm_adds_epi16(resReg23_lo, resReg45_lo);
+    resReg34_56_lo = _mm_adds_epi16(resReg34_lo, resReg56_lo);
+
+    // multiply 2 adjacent elements with the filter and add the result
+
+    resReg23_hi = _mm_maddubs_epi16(srcReg23_hi, secondFilters);
+    resReg34_hi = _mm_maddubs_epi16(srcReg34_hi, secondFilters);
+    resReg45_hi = _mm_maddubs_epi16(srcReg45_hi, thirdFilters);
+    resReg56_hi = _mm_maddubs_epi16(srcReg56_hi, thirdFilters);
+
+    // add and saturate the results together
+    resReg23_45_hi = _mm_adds_epi16(resReg23_hi, resReg45_hi);
+    resReg34_56_hi = _mm_adds_epi16(resReg34_hi, resReg56_hi);
+
+    // shift by 6 bit each 16 bit
+    resReg23_45_lo = _mm_adds_epi16(resReg23_45_lo, addFilterReg32);
+    resReg34_56_lo = _mm_adds_epi16(resReg34_56_lo, addFilterReg32);
+    resReg23_45_hi = _mm_adds_epi16(resReg23_45_hi, addFilterReg32);
+    resReg34_56_hi = _mm_adds_epi16(resReg34_56_hi, addFilterReg32);
+    resReg23_45_lo = _mm_srai_epi16(resReg23_45_lo, 6);
+    resReg34_56_lo = _mm_srai_epi16(resReg34_56_lo, 6);
+    resReg23_45_hi = _mm_srai_epi16(resReg23_45_hi, 6);
+    resReg34_56_hi = _mm_srai_epi16(resReg34_56_hi, 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
+    resReg23_45 = _mm_packus_epi16(resReg23_45_lo, resReg23_45_hi);
+    resReg34_56 = _mm_packus_epi16(resReg34_56_lo, resReg34_56_hi);
+
+    src_ptr += src_stride;
+
+    _mm_store_si128((__m128i *)output_ptr, (resReg23_45));
+    _mm_store_si128((__m128i *)(output_ptr + out_pitch), (resReg34_56));
+
+    output_ptr += dst_stride;
+
+    // save part of the registers for next strides
+    srcReg23_lo = srcReg45_lo;
+    srcReg34_lo = srcReg56_lo;
+    srcReg23_hi = srcReg45_hi;
+    srcReg34_hi = srcReg56_hi;
+    srcReg4 = srcReg6;
+  }
+}
+
+static INLINE __m128i shuffle_filter_convolve8_8_ssse3(
+    const __m128i *const s, const int16_t *const filter) {
+  __m128i f[4];
+  shuffle_filter_ssse3(filter, f);
+  return convolve8_8_ssse3(s, f);
+}
+
+static void filter_horiz_w8_ssse3(const uint8_t *const src,
+                                  const ptrdiff_t src_stride,
+                                  uint8_t *const dst,
+                                  const int16_t *const x_filter) {
+  __m128i s[8], ss[4], temp;
+
+  load_8bit_8x8(src, src_stride, s);
+  // 00 01 10 11 20 21 30 31  40 41 50 51 60 61 70 71
+  // 02 03 12 13 22 23 32 33  42 43 52 53 62 63 72 73
+  // 04 05 14 15 24 25 34 35  44 45 54 55 64 65 74 75
+  // 06 07 16 17 26 27 36 37  46 47 56 57 66 67 76 77
+  transpose_16bit_4x8(s, ss);
+  temp = shuffle_filter_convolve8_8_ssse3(ss, x_filter);
+  // shrink to 8 bit each 16 bits
+  temp = _mm_packus_epi16(temp, temp);
+  // save only 8 bytes convolve result
+  _mm_storel_epi64((__m128i *)dst, temp);
+}
+
+static void transpose8x8_to_dst(const uint8_t *const src,
+                                const ptrdiff_t src_stride, uint8_t *const dst,
+                                const ptrdiff_t dst_stride) {
+  __m128i s[8];
+
+  load_8bit_8x8(src, src_stride, s);
+  transpose_8bit_8x8(s, s);
+  store_8bit_8x8(s, dst, dst_stride);
+}
+
+static void scaledconvolve_horiz_w8(const uint8_t *src,
+                                    const ptrdiff_t src_stride, uint8_t *dst,
+                                    const ptrdiff_t dst_stride,
+                                    const InterpKernel *const x_filters,
+                                    const int x0_q4, const int x_step_q4,
+                                    const int w, const int h) {
+  DECLARE_ALIGNED(16, uint8_t, temp[8 * 8]);
+  int x, y, z;
+  src -= SUBPEL_TAPS / 2 - 1;
+
+  // This function processes 8x8 areas. The intermediate height is not always
+  // a multiple of 8, so force it to be a multiple of 8 here.
+  y = h + (8 - (h & 0x7));
+
+  do {
+    int x_q4 = x0_q4;
+    for (x = 0; x < w; x += 8) {
+      // process 8 src_x steps
+      for (z = 0; z < 8; ++z) {
+        const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
+        const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
+        if (x_q4 & SUBPEL_MASK) {
+          filter_horiz_w8_ssse3(src_x, src_stride, temp + (z * 8), x_filter);
+        } else {
+          int i;
+          for (i = 0; i < 8; ++i) {
+            temp[z * 8 + i] = src_x[i * src_stride + 3];
+          }
+        }
+        x_q4 += x_step_q4;
+      }
+
+      // transpose the 8x8 filters values back to dst
+      transpose8x8_to_dst(temp, 8, dst + x, dst_stride);
+    }
+
+    src += src_stride * 8;
+    dst += dst_stride * 8;
+  } while (y -= 8);
+}
+
+static void filter_horiz_w4_ssse3(const uint8_t *const src,
+                                  const ptrdiff_t src_stride,
+                                  uint8_t *const dst,
+                                  const int16_t *const filter) {
+  __m128i s[4];
+  __m128i temp;
+
+  load_8bit_8x4(src, src_stride, s);
+  transpose_16bit_4x4(s, s);
+
+  temp = shuffle_filter_convolve8_8_ssse3(s, filter);
+  // shrink to 8 bit each 16 bits
+  temp = _mm_packus_epi16(temp, temp);
+  // save only 4 bytes
+  *(int *)dst = _mm_cvtsi128_si32(temp);
+}
+
+static void transpose4x4_to_dst(const uint8_t *const src,
+                                const ptrdiff_t src_stride, uint8_t *const dst,
+                                const ptrdiff_t dst_stride) {
+  __m128i s[4];
+
+  load_8bit_4x4(src, src_stride, s);
+  s[0] = transpose_8bit_4x4(s);
+  s[1] = _mm_srli_si128(s[0], 4);
+  s[2] = _mm_srli_si128(s[0], 8);
+  s[3] = _mm_srli_si128(s[0], 12);
+  store_8bit_4x4(s, dst, dst_stride);
+}
+
+static void scaledconvolve_horiz_w4(const uint8_t *src,
+                                    const ptrdiff_t src_stride, uint8_t *dst,
+                                    const ptrdiff_t dst_stride,
+                                    const InterpKernel *const x_filters,
+                                    const int x0_q4, const int x_step_q4,
+                                    const int w, const int h) {
+  DECLARE_ALIGNED(16, uint8_t, temp[4 * 4]);
+  int x, y, z;
+  src -= SUBPEL_TAPS / 2 - 1;
+
+  for (y = 0; y < h; y += 4) {
+    int x_q4 = x0_q4;
+    for (x = 0; x < w; x += 4) {
+      // process 4 src_x steps
+      for (z = 0; z < 4; ++z) {
+        const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
+        const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
+        if (x_q4 & SUBPEL_MASK) {
+          filter_horiz_w4_ssse3(src_x, src_stride, temp + (z * 4), x_filter);
+        } else {
+          int i;
+          for (i = 0; i < 4; ++i) {
+            temp[z * 4 + i] = src_x[i * src_stride + 3];
+          }
+        }
+        x_q4 += x_step_q4;
+      }
+
+      // transpose the 4x4 filters values back to dst
+      transpose4x4_to_dst(temp, 4, dst + x, dst_stride);
+    }
+
+    src += src_stride * 4;
+    dst += dst_stride * 4;
+  }
+}
+
+static __m128i filter_vert_kernel(const __m128i *const s,
+                                  const int16_t *const filter) {
+  __m128i ss[4];
+  __m128i temp;
+
+  // 00 10 01 11 02 12 03 13
+  ss[0] = _mm_unpacklo_epi8(s[0], s[1]);
+  // 20 30 21 31 22 32 23 33
+  ss[1] = _mm_unpacklo_epi8(s[2], s[3]);
+  // 40 50 41 51 42 52 43 53
+  ss[2] = _mm_unpacklo_epi8(s[4], s[5]);
+  // 60 70 61 71 62 72 63 73
+  ss[3] = _mm_unpacklo_epi8(s[6], s[7]);
+
+  temp = shuffle_filter_convolve8_8_ssse3(ss, filter);
+  // shrink to 8 bit each 16 bits
+  return _mm_packus_epi16(temp, temp);
+}
+
+static void filter_vert_w4_ssse3(const uint8_t *const src,
+                                 const ptrdiff_t src_stride, uint8_t *const dst,
+                                 const int16_t *const filter) {
+  __m128i s[8];
+  __m128i temp;
+
+  load_8bit_4x8(src, src_stride, s);
+  temp = filter_vert_kernel(s, filter);
+  // save only 4 bytes
+  *(int *)dst = _mm_cvtsi128_si32(temp);
+}
+
+static void scaledconvolve_vert_w4(
+    const uint8_t *src, const ptrdiff_t src_stride, uint8_t *const dst,
+    const ptrdiff_t dst_stride, const InterpKernel *const y_filters,
+    const int y0_q4, const int y_step_q4, const int w, const int h) {
+  int y;
+  int y_q4 = y0_q4;
+
+  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+  for (y = 0; y < h; ++y) {
+    const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+    const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
+
+    if (y_q4 & SUBPEL_MASK) {
+      filter_vert_w4_ssse3(src_y, src_stride, &dst[y * dst_stride], y_filter);
+    } else {
+      memcpy(&dst[y * dst_stride], &src_y[3 * src_stride], w);
+    }
+
+    y_q4 += y_step_q4;
+  }
+}
+
+static void filter_vert_w8_ssse3(const uint8_t *const src,
+                                 const ptrdiff_t src_stride, uint8_t *const dst,
+                                 const int16_t *const filter) {
+  __m128i s[8], temp;
+
+  load_8bit_8x8(src, src_stride, s);
+  temp = filter_vert_kernel(s, filter);
+  // save only 8 bytes convolve result
+  _mm_storel_epi64((__m128i *)dst, temp);
+}
+
+static void scaledconvolve_vert_w8(
+    const uint8_t *src, const ptrdiff_t src_stride, uint8_t *const dst,
+    const ptrdiff_t dst_stride, const InterpKernel *const y_filters,
+    const int y0_q4, const int y_step_q4, const int w, const int h) {
+  int y;
+  int y_q4 = y0_q4;
+
+  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+  for (y = 0; y < h; ++y) {
+    const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+    const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
+    if (y_q4 & SUBPEL_MASK) {
+      filter_vert_w8_ssse3(src_y, src_stride, &dst[y * dst_stride], y_filter);
+    } else {
+      memcpy(&dst[y * dst_stride], &src_y[3 * src_stride], w);
+    }
+    y_q4 += y_step_q4;
+  }
+}
+
+static void filter_vert_w16_ssse3(const uint8_t *src,
+                                  const ptrdiff_t src_stride,
+                                  uint8_t *const dst,
+                                  const int16_t *const filter, const int w) {
+  int i;
+  __m128i f[4];
+  shuffle_filter_ssse3(filter, f);
+
+  for (i = 0; i < w; i += 16) {
+    __m128i s[8], s_lo[4], s_hi[4], temp_lo, temp_hi;
+
+    loadu_8bit_16x8(src, src_stride, s);
+
+    // merge the result together
+    s_lo[0] = _mm_unpacklo_epi8(s[0], s[1]);
+    s_hi[0] = _mm_unpackhi_epi8(s[0], s[1]);
+    s_lo[1] = _mm_unpacklo_epi8(s[2], s[3]);
+    s_hi[1] = _mm_unpackhi_epi8(s[2], s[3]);
+    s_lo[2] = _mm_unpacklo_epi8(s[4], s[5]);
+    s_hi[2] = _mm_unpackhi_epi8(s[4], s[5]);
+    s_lo[3] = _mm_unpacklo_epi8(s[6], s[7]);
+    s_hi[3] = _mm_unpackhi_epi8(s[6], s[7]);
+    temp_lo = convolve8_8_ssse3(s_lo, f);
+    temp_hi = convolve8_8_ssse3(s_hi, f);
+
+    // shrink to 8 bit each 16 bits, the first lane contain the first convolve
+    // result and the second lane contain the second convolve result
+    temp_hi = _mm_packus_epi16(temp_lo, temp_hi);
+    src += 16;
+    // save 16 bytes convolve result
+    _mm_store_si128((__m128i *)&dst[i], temp_hi);
+  }
+}
+
+static void scaledconvolve_vert_w16(
+    const uint8_t *src, const ptrdiff_t src_stride, uint8_t *const dst,
+    const ptrdiff_t dst_stride, const InterpKernel *const y_filters,
+    const int y0_q4, const int y_step_q4, const int w, const int h) {
+  int y;
+  int y_q4 = y0_q4;
+
+  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+  for (y = 0; y < h; ++y) {
+    const unsigned char *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+    const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
+    if (y_q4 & SUBPEL_MASK) {
+      filter_vert_w16_ssse3(src_y, src_stride, &dst[y * dst_stride], y_filter,
+                            w);
+    } else {
+      memcpy(&dst[y * dst_stride], &src_y[3 * src_stride], w);
+    }
+    y_q4 += y_step_q4;
+  }
+}
+
+void aom_scaled_2d_ssse3(const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,
+                         ptrdiff_t dst_stride, const InterpKernel *filter,
+                         int x0_q4, int x_step_q4, int y0_q4, int y_step_q4,
+                         int w, int h) {
+  // Note: Fixed size intermediate buffer, temp, places limits on parameters.
+  // 2d filtering proceeds in 2 steps:
+  //   (1) Interpolate horizontally into an intermediate buffer, temp.
+  //   (2) Interpolate temp vertically to derive the sub-pixel result.
+  // Deriving the maximum number of rows in the temp buffer (135):
+  // --Smallest scaling factor is x1/2 ==> y_step_q4 = 32 (Normative).
+  // --Largest block size is 64x64 pixels.
+  // --64 rows in the downscaled frame span a distance of (64 - 1) * 32 in the
+  //   original frame (in 1/16th pixel units).
+  // --Must round-up because block may be located at sub-pixel position.
+  // --Require an additional SUBPEL_TAPS rows for the 8-tap filter tails.
+  // --((64 - 1) * 32 + 15) >> 4 + 8 = 135.
+  // --Require an additional 8 rows for the horiz_w8 transpose tail.
+  // When calling in frame scaling function, the smallest scaling factor is x1/4
+  // ==> y_step_q4 = 64. Since w and h are at most 16, the temp buffer is still
+  // big enough.
+  DECLARE_ALIGNED(16, uint8_t, temp[(135 + 8) * 64]);
+  const int intermediate_height =
+      (((h - 1) * y_step_q4 + y0_q4) >> SUBPEL_BITS) + SUBPEL_TAPS;
+
+  assert(w <= 64);
+  assert(h <= 64);
+  assert(y_step_q4 <= 32 || (y_step_q4 <= 64 && h <= 32));
+  assert(x_step_q4 <= 64);
+
+  if (w >= 8) {
+    scaledconvolve_horiz_w8(src - src_stride * (SUBPEL_TAPS / 2 - 1),
+                            src_stride, temp, 64, filter, x0_q4, x_step_q4, w,
+                            intermediate_height);
+  } else {
+    scaledconvolve_horiz_w4(src - src_stride * (SUBPEL_TAPS / 2 - 1),
+                            src_stride, temp, 64, filter, x0_q4, x_step_q4, w,
+                            intermediate_height);
+  }
+
+  if (w >= 16) {
+    scaledconvolve_vert_w16(temp + 64 * (SUBPEL_TAPS / 2 - 1), 64, dst,
+                            dst_stride, filter, y0_q4, y_step_q4, w, h);
+  } else if (w == 8) {
+    scaledconvolve_vert_w8(temp + 64 * (SUBPEL_TAPS / 2 - 1), 64, dst,
+                           dst_stride, filter, y0_q4, y_step_q4, w, h);
+  } else {
+    scaledconvolve_vert_w4(temp + 64 * (SUBPEL_TAPS / 2 - 1), 64, dst,
+                           dst_stride, filter, y0_q4, y_step_q4, w, h);
+  }
+}
+
+filter8_1dfunction aom_filter_block1d16_v8_ssse3;
+filter8_1dfunction aom_filter_block1d16_h8_ssse3;
+filter8_1dfunction aom_filter_block1d8_v8_ssse3;
+filter8_1dfunction aom_filter_block1d8_h8_ssse3;
+filter8_1dfunction aom_filter_block1d4_v8_ssse3;
+filter8_1dfunction aom_filter_block1d4_h8_ssse3;
+
+filter8_1dfunction aom_filter_block1d16_v2_ssse3;
+filter8_1dfunction aom_filter_block1d16_h2_ssse3;
+filter8_1dfunction aom_filter_block1d8_v2_ssse3;
+filter8_1dfunction aom_filter_block1d8_h2_ssse3;
+filter8_1dfunction aom_filter_block1d4_v2_ssse3;
+filter8_1dfunction aom_filter_block1d4_h2_ssse3;
+
+// void aom_convolve8_horiz_ssse3(const uint8_t *src, ptrdiff_t src_stride,
+//                                uint8_t *dst, ptrdiff_t dst_stride,
+//                                const int16_t *filter_x, int x_step_q4,
+//                                const int16_t *filter_y, int y_step_q4,
+//                                int w, int h);
+// void aom_convolve8_vert_ssse3(const uint8_t *src, ptrdiff_t src_stride,
+//                               uint8_t *dst, ptrdiff_t dst_stride,
+//                               const int16_t *filter_x, int x_step_q4,
+//                               const int16_t *filter_y, int y_step_q4,
+//                               int w, int h);
+FUN_CONV_1D(horiz, x_step_q4, filter_x, h, src, , ssse3)
+FUN_CONV_1D(vert, y_step_q4, filter_y, v, src - src_stride * 3, , ssse3)
diff --git a/third_party/aom/aom_dsp/x86/aom_subpixel_8t_sse2.asm b/third_party/aom/aom_dsp/x86/aom_subpixel_8t_sse2.asm
new file mode 100644
index 0000000000..640c5b2416
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/aom_subpixel_8t_sse2.asm
@@ -0,0 +1,615 @@
+;
+; Copyright (c) 2016, 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 "aom_ports/x86_abi_support.asm"
+
+;Note: tap3 and tap4 have to be applied and added after other taps to avoid
+;overflow.
+
+%macro GET_FILTERS_4 0
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rcx, 0x0400040
+
+    movdqa      xmm7, [rdx]                 ;load filters
+    pshuflw     xmm0, xmm7, 0b              ;k0
+    pshuflw     xmm1, xmm7, 01010101b       ;k1
+    pshuflw     xmm2, xmm7, 10101010b       ;k2
+    pshuflw     xmm3, xmm7, 11111111b       ;k3
+    psrldq      xmm7, 8
+    pshuflw     xmm4, xmm7, 0b              ;k4
+    pshuflw     xmm5, xmm7, 01010101b       ;k5
+    pshuflw     xmm6, xmm7, 10101010b       ;k6
+    pshuflw     xmm7, xmm7, 11111111b       ;k7
+
+    punpcklqdq  xmm0, xmm1
+    punpcklqdq  xmm2, xmm3
+    punpcklqdq  xmm5, xmm4
+    punpcklqdq  xmm6, xmm7
+
+    movdqa      k0k1, xmm0
+    movdqa      k2k3, xmm2
+    movdqa      k5k4, xmm5
+    movdqa      k6k7, xmm6
+
+    movq        xmm6, rcx
+    pshufd      xmm6, xmm6, 0
+    movdqa      krd, xmm6
+
+    pxor        xmm7, xmm7
+    movdqa      zero, xmm7
+%endm
+
+%macro APPLY_FILTER_4 1
+    punpckldq   xmm0, xmm1                  ;two row in one register
+    punpckldq   xmm6, xmm7
+    punpckldq   xmm2, xmm3
+    punpckldq   xmm5, xmm4
+
+    punpcklbw   xmm0, zero                  ;unpack to word
+    punpcklbw   xmm6, zero
+    punpcklbw   xmm2, zero
+    punpcklbw   xmm5, zero
+
+    pmullw      xmm0, k0k1                  ;multiply the filter factors
+    pmullw      xmm6, k6k7
+    pmullw      xmm2, k2k3
+    pmullw      xmm5, k5k4
+
+    paddsw      xmm0, xmm6                  ;sum
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 8
+    paddsw      xmm0, xmm1
+    paddsw      xmm0, xmm2
+    psrldq      xmm2, 8
+    paddsw      xmm0, xmm5
+    psrldq      xmm5, 8
+    paddsw      xmm0, xmm2
+    paddsw      xmm0, xmm5
+
+    paddsw      xmm0, krd                   ;rounding
+    psraw       xmm0, 7                     ;shift
+    packuswb    xmm0, xmm0                  ;pack to byte
+
+%if %1
+    movd        xmm1, [rdi]
+    pavgb       xmm0, xmm1
+%endif
+    movd        [rdi], xmm0
+%endm
+
+%macro GET_FILTERS 0
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+    mov         rcx, 0x0400040
+
+    movdqa      xmm7, [rdx]                 ;load filters
+    pshuflw     xmm0, xmm7, 0b              ;k0
+    pshuflw     xmm1, xmm7, 01010101b       ;k1
+    pshuflw     xmm2, xmm7, 10101010b       ;k2
+    pshuflw     xmm3, xmm7, 11111111b       ;k3
+    pshufhw     xmm4, xmm7, 0b              ;k4
+    pshufhw     xmm5, xmm7, 01010101b       ;k5
+    pshufhw     xmm6, xmm7, 10101010b       ;k6
+    pshufhw     xmm7, xmm7, 11111111b       ;k7
+
+    punpcklwd   xmm0, xmm0
+    punpcklwd   xmm1, xmm1
+    punpcklwd   xmm2, xmm2
+    punpcklwd   xmm3, xmm3
+    punpckhwd   xmm4, xmm4
+    punpckhwd   xmm5, xmm5
+    punpckhwd   xmm6, xmm6
+    punpckhwd   xmm7, xmm7
+
+    movdqa      k0,   xmm0                  ;store filter factors on stack
+    movdqa      k1,   xmm1
+    movdqa      k2,   xmm2
+    movdqa      k3,   xmm3
+    movdqa      k4,   xmm4
+    movdqa      k5,   xmm5
+    movdqa      k6,   xmm6
+    movdqa      k7,   xmm7
+
+    movq        xmm6, rcx
+    pshufd      xmm6, xmm6, 0
+    movdqa      krd, xmm6                   ;rounding
+
+    pxor        xmm7, xmm7
+    movdqa      zero, xmm7
+%endm
+
+%macro LOAD_VERT_8 1
+    movq        xmm0, [rsi + %1]            ;0
+    movq        xmm1, [rsi + rax + %1]      ;1
+    movq        xmm6, [rsi + rdx * 2 + %1]  ;6
+    lea         rsi,  [rsi + rax]
+    movq        xmm7, [rsi + rdx * 2 + %1]  ;7
+    movq        xmm2, [rsi + rax + %1]      ;2
+    movq        xmm3, [rsi + rax * 2 + %1]  ;3
+    movq        xmm4, [rsi + rdx + %1]      ;4
+    movq        xmm5, [rsi + rax * 4 + %1]  ;5
+%endm
+
+%macro APPLY_FILTER_8 2
+    punpcklbw   xmm0, zero
+    punpcklbw   xmm1, zero
+    punpcklbw   xmm6, zero
+    punpcklbw   xmm7, zero
+    punpcklbw   xmm2, zero
+    punpcklbw   xmm5, zero
+    punpcklbw   xmm3, zero
+    punpcklbw   xmm4, zero
+
+    pmullw      xmm0, k0
+    pmullw      xmm1, k1
+    pmullw      xmm6, k6
+    pmullw      xmm7, k7
+    pmullw      xmm2, k2
+    pmullw      xmm5, k5
+    pmullw      xmm3, k3
+    pmullw      xmm4, k4
+
+    paddsw      xmm0, xmm1
+    paddsw      xmm0, xmm6
+    paddsw      xmm0, xmm7
+    paddsw      xmm0, xmm2
+    paddsw      xmm0, xmm5
+    paddsw      xmm0, xmm3
+    paddsw      xmm0, xmm4
+
+    paddsw      xmm0, krd                   ;rounding
+    psraw       xmm0, 7                     ;shift
+    packuswb    xmm0, xmm0                  ;pack back to byte
+%if %1
+    movq        xmm1, [rdi + %2]
+    pavgb       xmm0, xmm1
+%endif
+    movq        [rdi + %2], xmm0
+%endm
+
+SECTION .text
+
+;void aom_filter_block1d4_v8_sse2
+;(
+;    unsigned char *src_ptr,
+;    unsigned int   src_pitch,
+;    unsigned char *output_ptr,
+;    unsigned int   out_pitch,
+;    unsigned int   output_height,
+;    short *filter
+;)
+globalsym(aom_filter_block1d4_v8_sse2)
+sym(aom_filter_block1d4_v8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 6
+    %define k0k1 [rsp + 16 * 0]
+    %define k2k3 [rsp + 16 * 1]
+    %define k5k4 [rsp + 16 * 2]
+    %define k6k7 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define zero [rsp + 16 * 5]
+
+    GET_FILTERS_4
+
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rbx, DWORD PTR arg(3)       ;out_pitch
+    lea         rdx, [rax + rax * 2]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movd        xmm0, [rsi]                 ;load src: row 0
+    movd        xmm1, [rsi + rax]           ;1
+    movd        xmm6, [rsi + rdx * 2]       ;6
+    lea         rsi,  [rsi + rax]
+    movd        xmm7, [rsi + rdx * 2]       ;7
+    movd        xmm2, [rsi + rax]           ;2
+    movd        xmm3, [rsi + rax * 2]       ;3
+    movd        xmm4, [rsi + rdx]           ;4
+    movd        xmm5, [rsi + rax * 4]       ;5
+
+    APPLY_FILTER_4 0
+
+    lea         rdi, [rdi + rbx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 6
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void aom_filter_block1d8_v8_sse2
+;(
+;    unsigned char *src_ptr,
+;    unsigned int   src_pitch,
+;    unsigned char *output_ptr,
+;    unsigned int   out_pitch,
+;    unsigned int   output_height,
+;    short *filter
+;)
+globalsym(aom_filter_block1d8_v8_sse2)
+sym(aom_filter_block1d8_v8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 10
+    %define k0 [rsp + 16 * 0]
+    %define k1 [rsp + 16 * 1]
+    %define k2 [rsp + 16 * 2]
+    %define k3 [rsp + 16 * 3]
+    %define k4 [rsp + 16 * 4]
+    %define k5 [rsp + 16 * 5]
+    %define k6 [rsp + 16 * 6]
+    %define k7 [rsp + 16 * 7]
+    %define krd [rsp + 16 * 8]
+    %define zero [rsp + 16 * 9]
+
+    GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rbx, DWORD PTR arg(3)       ;out_pitch
+    lea         rdx, [rax + rax * 2]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    LOAD_VERT_8 0
+    APPLY_FILTER_8 0, 0
+
+    lea         rdi, [rdi + rbx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 10
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void aom_filter_block1d16_v8_sse2
+;(
+;    unsigned char *src_ptr,
+;    unsigned int   src_pitch,
+;    unsigned char *output_ptr,
+;    unsigned int   out_pitch,
+;    unsigned int   output_height,
+;    short *filter
+;)
+globalsym(aom_filter_block1d16_v8_sse2)
+sym(aom_filter_block1d16_v8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    push        rbx
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 10
+    %define k0 [rsp + 16 * 0]
+    %define k1 [rsp + 16 * 1]
+    %define k2 [rsp + 16 * 2]
+    %define k3 [rsp + 16 * 3]
+    %define k4 [rsp + 16 * 4]
+    %define k5 [rsp + 16 * 5]
+    %define k6 [rsp + 16 * 6]
+    %define k7 [rsp + 16 * 7]
+    %define krd [rsp + 16 * 8]
+    %define zero [rsp + 16 * 9]
+
+    GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rbx, DWORD PTR arg(3)       ;out_pitch
+    lea         rdx, [rax + rax * 2]
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    LOAD_VERT_8 0
+    APPLY_FILTER_8 0, 0
+    sub         rsi, rax
+
+    LOAD_VERT_8 8
+    APPLY_FILTER_8 0, 8
+    add         rdi, rbx
+
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 10
+    pop rsp
+    pop rbx
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void aom_filter_block1d4_h8_sse2
+;(
+;    unsigned char  *src_ptr,
+;    unsigned int    src_pixels_per_line,
+;    unsigned char  *output_ptr,
+;    unsigned int    output_pitch,
+;    unsigned int    output_height,
+;    short *filter
+;)
+globalsym(aom_filter_block1d4_h8_sse2)
+sym(aom_filter_block1d4_h8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 6
+    %define k0k1 [rsp + 16 * 0]
+    %define k2k3 [rsp + 16 * 1]
+    %define k5k4 [rsp + 16 * 2]
+    %define k6k7 [rsp + 16 * 3]
+    %define krd [rsp + 16 * 4]
+    %define zero [rsp + 16 * 5]
+
+    GET_FILTERS_4
+
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movdqu      xmm0,   [rsi - 3]           ;load src
+
+    movdqa      xmm1, xmm0
+    movdqa      xmm6, xmm0
+    movdqa      xmm7, xmm0
+    movdqa      xmm2, xmm0
+    movdqa      xmm3, xmm0
+    movdqa      xmm5, xmm0
+    movdqa      xmm4, xmm0
+
+    psrldq      xmm1, 1
+    psrldq      xmm6, 6
+    psrldq      xmm7, 7
+    psrldq      xmm2, 2
+    psrldq      xmm3, 3
+    psrldq      xmm5, 5
+    psrldq      xmm4, 4
+
+    APPLY_FILTER_4 0
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 6
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void aom_filter_block1d8_h8_sse2
+;(
+;    unsigned char  *src_ptr,
+;    unsigned int    src_pixels_per_line,
+;    unsigned char  *output_ptr,
+;    unsigned int    output_pitch,
+;    unsigned int    output_height,
+;    short *filter
+;)
+globalsym(aom_filter_block1d8_h8_sse2)
+sym(aom_filter_block1d8_h8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 10
+    %define k0 [rsp + 16 * 0]
+    %define k1 [rsp + 16 * 1]
+    %define k2 [rsp + 16 * 2]
+    %define k3 [rsp + 16 * 3]
+    %define k4 [rsp + 16 * 4]
+    %define k5 [rsp + 16 * 5]
+    %define k6 [rsp + 16 * 6]
+    %define k7 [rsp + 16 * 7]
+    %define krd [rsp + 16 * 8]
+    %define zero [rsp + 16 * 9]
+
+    GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movdqu      xmm0,   [rsi - 3]           ;load src
+
+    movdqa      xmm1, xmm0
+    movdqa      xmm6, xmm0
+    movdqa      xmm7, xmm0
+    movdqa      xmm2, xmm0
+    movdqa      xmm5, xmm0
+    movdqa      xmm3, xmm0
+    movdqa      xmm4, xmm0
+
+    psrldq      xmm1, 1
+    psrldq      xmm6, 6
+    psrldq      xmm7, 7
+    psrldq      xmm2, 2
+    psrldq      xmm5, 5
+    psrldq      xmm3, 3
+    psrldq      xmm4, 4
+
+    APPLY_FILTER_8 0, 0
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 10
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void aom_filter_block1d16_h8_sse2
+;(
+;    unsigned char  *src_ptr,
+;    unsigned int    src_pixels_per_line,
+;    unsigned char  *output_ptr,
+;    unsigned int    output_pitch,
+;    unsigned int    output_height,
+;    short *filter
+;)
+globalsym(aom_filter_block1d16_h8_sse2)
+sym(aom_filter_block1d16_h8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    ALIGN_STACK 16, rax
+    sub         rsp, 16 * 10
+    %define k0 [rsp + 16 * 0]
+    %define k1 [rsp + 16 * 1]
+    %define k2 [rsp + 16 * 2]
+    %define k3 [rsp + 16 * 3]
+    %define k4 [rsp + 16 * 4]
+    %define k5 [rsp + 16 * 5]
+    %define k6 [rsp + 16 * 6]
+    %define k7 [rsp + 16 * 7]
+    %define krd [rsp + 16 * 8]
+    %define zero [rsp + 16 * 9]
+
+    GET_FILTERS
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+
+.loop:
+    movdqu      xmm0,   [rsi - 3]           ;load src
+
+    movdqa      xmm1, xmm0
+    movdqa      xmm6, xmm0
+    movdqa      xmm7, xmm0
+    movdqa      xmm2, xmm0
+    movdqa      xmm5, xmm0
+    movdqa      xmm3, xmm0
+    movdqa      xmm4, xmm0
+
+    psrldq      xmm1, 1
+    psrldq      xmm6, 6
+    psrldq      xmm7, 7
+    psrldq      xmm2, 2
+    psrldq      xmm5, 5
+    psrldq      xmm3, 3
+    psrldq      xmm4, 4
+
+    APPLY_FILTER_8 0, 0
+
+    movdqu      xmm0,   [rsi + 5]           ;load src
+
+    movdqa      xmm1, xmm0
+    movdqa      xmm6, xmm0
+    movdqa      xmm7, xmm0
+    movdqa      xmm2, xmm0
+    movdqa      xmm5, xmm0
+    movdqa      xmm3, xmm0
+    movdqa      xmm4, xmm0
+
+    psrldq      xmm1, 1
+    psrldq      xmm6, 6
+    psrldq      xmm7, 7
+    psrldq      xmm2, 2
+    psrldq      xmm5, 5
+    psrldq      xmm3, 3
+    psrldq      xmm4, 4
+
+    APPLY_FILTER_8 0, 8
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+    jnz         .loop
+
+    add rsp, 16 * 10
+    pop rsp
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
diff --git a/third_party/aom/aom_dsp/x86/aom_subpixel_8t_ssse3.asm b/third_party/aom/aom_dsp/x86/aom_subpixel_8t_ssse3.asm
new file mode 100644
index 0000000000..e5fafb0302
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/aom_subpixel_8t_ssse3.asm
@@ -0,0 +1,870 @@
+;
+; Copyright (c) 2016, 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 "third_party/x86inc/x86inc.asm"
+
+SECTION_RODATA
+pw_64:    times 8 dw 64
+even_byte_mask: times 8 dw 0x00ff
+
+; %define USE_PMULHRSW
+; NOTE: pmulhrsw has a latency of 5 cycles.  Tests showed a performance loss
+; when using this instruction.
+;
+; The add order below (based on ffav1) must be followed to prevent outranges.
+; x = k0k1 + k4k5
+; y = k2k3 + k6k7
+; z = signed SAT(x + y)
+
+SECTION .text
+%define LOCAL_VARS_SIZE 16*6
+
+%macro SETUP_LOCAL_VARS 0
+    ; TODO(slavarnway): using xmm registers for these on AOM_ARCH_X86_64 +
+    ; pmaddubsw has a higher latency on some platforms, this might be eased by
+    ; interleaving the instructions.
+    %define    k0k1  [rsp + 16*0]
+    %define    k2k3  [rsp + 16*1]
+    %define    k4k5  [rsp + 16*2]
+    %define    k6k7  [rsp + 16*3]
+    packsswb     m4, m4
+    ; TODO(slavarnway): multiple pshufb instructions had a higher latency on
+    ; some platforms.
+    pshuflw      m0, m4, 0b              ;k0_k1
+    pshuflw      m1, m4, 01010101b       ;k2_k3
+    pshuflw      m2, m4, 10101010b       ;k4_k5
+    pshuflw      m3, m4, 11111111b       ;k6_k7
+    punpcklqdq   m0, m0
+    punpcklqdq   m1, m1
+    punpcklqdq   m2, m2
+    punpcklqdq   m3, m3
+    mova       k0k1, m0
+    mova       k2k3, m1
+    mova       k4k5, m2
+    mova       k6k7, m3
+%if AOM_ARCH_X86_64
+    %define     krd  m12
+    %define    tmp0  [rsp + 16*4]
+    %define    tmp1  [rsp + 16*5]
+    mova        krd, [GLOBAL(pw_64)]
+%else
+    %define     krd  [rsp + 16*4]
+%if CONFIG_PIC=0
+    mova         m6, [GLOBAL(pw_64)]
+%else
+    ; build constants without accessing global memory
+    pcmpeqb      m6, m6                  ;all ones
+    psrlw        m6, 15
+    psllw        m6, 6                   ;aka pw_64
+%endif
+    mova        krd, m6
+%endif
+%endm
+
+;-------------------------------------------------------------------------------
+%if AOM_ARCH_X86_64
+  %define LOCAL_VARS_SIZE_H4 0
+%else
+  %define LOCAL_VARS_SIZE_H4 16*4
+%endif
+
+%macro SUBPIX_HFILTER4 1
+cglobal filter_block1d4_%1, 6, 6, 11, LOCAL_VARS_SIZE_H4, \
+                            src, sstride, dst, dstride, height, filter
+    mova                m4, [filterq]
+    packsswb            m4, m4
+%if AOM_ARCH_X86_64
+    %define       k0k1k4k5  m8
+    %define       k2k3k6k7  m9
+    %define            krd  m10
+    mova               krd, [GLOBAL(pw_64)]
+    pshuflw       k0k1k4k5, m4, 0b              ;k0_k1
+    pshufhw       k0k1k4k5, k0k1k4k5, 10101010b ;k0_k1_k4_k5
+    pshuflw       k2k3k6k7, m4, 01010101b       ;k2_k3
+    pshufhw       k2k3k6k7, k2k3k6k7, 11111111b ;k2_k3_k6_k7
+%else
+    %define       k0k1k4k5  [rsp + 16*0]
+    %define       k2k3k6k7  [rsp + 16*1]
+    %define            krd  [rsp + 16*2]
+    pshuflw             m6, m4, 0b              ;k0_k1
+    pshufhw             m6, m6, 10101010b       ;k0_k1_k4_k5
+    pshuflw             m7, m4, 01010101b       ;k2_k3
+    pshufhw             m7, m7, 11111111b       ;k2_k3_k6_k7
+%if CONFIG_PIC=0
+    mova                m1, [GLOBAL(pw_64)]
+%else
+    ; build constants without accessing global memory
+    pcmpeqb             m1, m1                  ;all ones
+    psrlw               m1, 15
+    psllw               m1, 6                   ;aka pw_64
+%endif
+    mova          k0k1k4k5, m6
+    mova          k2k3k6k7, m7
+    mova               krd, m1
+%endif
+    dec            heightd
+
+.loop:
+    ;Do two rows at once
+    movu                m4, [srcq - 3]
+    movu                m5, [srcq + sstrideq - 3]
+    punpckhbw           m1, m4, m4
+    punpcklbw           m4, m4
+    punpckhbw           m3, m5, m5
+    punpcklbw           m5, m5
+    palignr             m0, m1, m4, 1
+    pmaddubsw           m0, k0k1k4k5
+    palignr             m1, m4, 5
+    pmaddubsw           m1, k2k3k6k7
+    palignr             m2, m3, m5, 1
+    pmaddubsw           m2, k0k1k4k5
+    palignr             m3, m5, 5
+    pmaddubsw           m3, k2k3k6k7
+    punpckhqdq          m4, m0, m2
+    punpcklqdq          m0, m2
+    punpckhqdq          m5, m1, m3
+    punpcklqdq          m1, m3
+    paddsw              m0, m4
+    paddsw              m1, m5
+%ifidn %1, h8_avg
+    movd                m4, [dstq]
+    movd                m5, [dstq + dstrideq]
+%endif
+    paddsw              m0, m1
+    paddsw              m0, krd
+    psraw               m0, 7
+%ifidn %1, h8_add_src
+    pxor                 m3, m3
+    movu                 m4, [srcq]
+    movu                 m5, [srcq + sstrideq]
+    punpckldq            m4, m5 ; Bytes 0,1,2,3 from row 0, then 0,1,2,3 from row 2
+    punpcklbw            m4, m3
+    paddsw               m0, m4
+%endif
+    packuswb            m0, m0
+    psrldq              m1, m0, 4
+
+%ifidn %1, h8_avg
+    pavgb               m0, m4
+    pavgb               m1, m5
+%endif
+    movd            [dstq], m0
+    movd [dstq + dstrideq], m1
+
+    lea               srcq, [srcq + sstrideq        ]
+    prefetcht0              [srcq + 4 * sstrideq - 3]
+    lea               srcq, [srcq + sstrideq        ]
+    lea               dstq, [dstq + 2 * dstrideq    ]
+    prefetcht0              [srcq + 2 * sstrideq - 3]
+
+    sub            heightd, 2
+    jg               .loop
+
+    ; Do last row if output_height is odd
+    jne              .done
+
+    movu                m4, [srcq - 3]
+    punpckhbw           m1, m4, m4
+    punpcklbw           m4, m4
+    palignr             m0, m1, m4, 1
+    palignr             m1, m4, 5
+    pmaddubsw           m0, k0k1k4k5
+    pmaddubsw           m1, k2k3k6k7
+    psrldq              m2, m0, 8
+    psrldq              m3, m1, 8
+    paddsw              m0, m2
+    paddsw              m1, m3
+    paddsw              m0, m1
+    paddsw              m0, krd
+    psraw               m0, 7
+%ifidn %1, h8_add_src
+    pxor                m3, m3
+    movu                m4, [srcq]
+    punpcklbw           m4, m3
+    paddsw              m0, m4
+%endif
+    packuswb            m0, m0
+%ifidn %1, h8_avg
+    movd                m4, [dstq]
+    pavgb               m0, m4
+%endif
+    movd            [dstq], m0
+.done:
+    REP_RET
+%endm
+
+;-------------------------------------------------------------------------------
+%macro SUBPIX_HFILTER8 1
+cglobal filter_block1d8_%1, 6, 6, 14, LOCAL_VARS_SIZE, \
+                            src, sstride, dst, dstride, height, filter
+    mova                 m4, [filterq]
+    SETUP_LOCAL_VARS
+    dec             heightd
+
+.loop:
+    ;Do two rows at once
+    movu                 m0, [srcq - 3]
+    movu                 m4, [srcq + sstrideq - 3]
+    punpckhbw            m1, m0, m0
+    punpcklbw            m0, m0
+    palignr              m5, m1, m0, 13
+    pmaddubsw            m5, k6k7
+    palignr              m2, m1, m0, 5
+    palignr              m3, m1, m0, 9
+    palignr              m1, m0, 1
+    pmaddubsw            m1, k0k1
+    punpckhbw            m6, m4, m4
+    punpcklbw            m4, m4
+    pmaddubsw            m2, k2k3
+    pmaddubsw            m3, k4k5
+
+    palignr              m7, m6, m4, 13
+    palignr              m0, m6, m4, 5
+    pmaddubsw            m7, k6k7
+    paddsw               m1, m3
+    paddsw               m2, m5
+    paddsw               m1, m2
+%ifidn %1, h8_avg
+    movh                 m2, [dstq]
+    movhps               m2, [dstq + dstrideq]
+%endif
+    palignr              m5, m6, m4, 9
+    palignr              m6, m4, 1
+    pmaddubsw            m0, k2k3
+    pmaddubsw            m6, k0k1
+    paddsw               m1, krd
+    pmaddubsw            m5, k4k5
+    psraw                m1, 7
+    paddsw               m0, m7
+    paddsw               m6, m5
+    paddsw               m6, m0
+    paddsw               m6, krd
+    psraw                m6, 7
+%ifidn %1, h8_add_src
+    pxor                 m3, m3
+    movu                 m4, [srcq]
+    movu                 m5, [srcq + sstrideq]
+    punpcklbw            m4, m3
+    punpcklbw            m5, m3
+    paddsw               m1, m4
+    paddsw               m6, m5
+%endif
+    packuswb             m1, m6
+%ifidn %1, h8_avg
+    pavgb                m1, m2
+%endif
+    movh              [dstq], m1
+    movhps [dstq + dstrideq], m1
+
+    lea                srcq, [srcq + sstrideq        ]
+    prefetcht0               [srcq + 4 * sstrideq - 3]
+    lea                srcq, [srcq + sstrideq        ]
+    lea                dstq, [dstq + 2 * dstrideq    ]
+    prefetcht0               [srcq + 2 * sstrideq - 3]
+    sub             heightd, 2
+    jg                .loop
+
+    ; Do last row if output_height is odd
+    jne               .done
+
+    movu                 m0, [srcq - 3]
+    punpckhbw            m3, m0, m0
+    punpcklbw            m0, m0
+    palignr              m1, m3, m0, 1
+    palignr              m2, m3, m0, 5
+    palignr              m4, m3, m0, 13
+    palignr              m3, m0, 9
+    pmaddubsw            m1, k0k1
+    pmaddubsw            m2, k2k3
+    pmaddubsw            m3, k4k5
+    pmaddubsw            m4, k6k7
+    paddsw               m1, m3
+    paddsw               m4, m2
+    paddsw               m1, m4
+    paddsw               m1, krd
+    psraw                m1, 7
+%ifidn %1, h8_add_src
+    pxor                 m6, m6
+    movu                 m5, [srcq]
+    punpcklbw            m5, m6
+    paddsw               m1, m5
+%endif
+    packuswb             m1, m1
+%ifidn %1, h8_avg
+    movh                 m0, [dstq]
+    pavgb                m1, m0
+%endif
+    movh             [dstq], m1
+.done:
+    REP_RET
+%endm
+
+;-------------------------------------------------------------------------------
+%macro SUBPIX_HFILTER16 1
+cglobal filter_block1d16_%1, 6, 6, 14, LOCAL_VARS_SIZE, \
+                             src, sstride, dst, dstride, height, filter
+    mova          m4, [filterq]
+    SETUP_LOCAL_VARS
+
+.loop:
+    prefetcht0        [srcq + 2 * sstrideq -3]
+
+    movu          m0, [srcq - 3]
+    movu          m4, [srcq - 2]
+    pmaddubsw     m0, k0k1
+    pmaddubsw     m4, k0k1
+    movu          m1, [srcq - 1]
+    movu          m5, [srcq + 0]
+    pmaddubsw     m1, k2k3
+    pmaddubsw     m5, k2k3
+    movu          m2, [srcq + 1]
+    movu          m6, [srcq + 2]
+    pmaddubsw     m2, k4k5
+    pmaddubsw     m6, k4k5
+    movu          m3, [srcq + 3]
+    movu          m7, [srcq + 4]
+    pmaddubsw     m3, k6k7
+    pmaddubsw     m7, k6k7
+    paddsw        m0, m2
+    paddsw        m1, m3
+    paddsw        m0, m1
+    paddsw        m4, m6
+    paddsw        m5, m7
+    paddsw        m4, m5
+    paddsw        m0, krd
+    paddsw        m4, krd
+    psraw         m0, 7
+    psraw         m4, 7
+%ifidn %1, h8_add_src
+%if AOM_ARCH_X86=1 && CONFIG_PIC=1
+    pcmpeqb       m2, m2                  ;all ones
+    psrlw         m2, 8                   ;even_byte_mask
+%else
+    mova          m2, [GLOBAL(even_byte_mask)]
+%endif
+    movu          m5, [srcq]
+    mova          m7, m5
+    pand          m5, m2
+    psrlw         m7, 8
+    paddsw        m0, m5
+    paddsw        m4, m7
+%endif
+    packuswb      m0, m0
+    packuswb      m4, m4
+    punpcklbw     m0, m4
+%ifidn %1, h8_avg
+    pavgb         m0, [dstq]
+%endif
+    lea         srcq, [srcq + sstrideq]
+    mova      [dstq], m0
+    lea         dstq, [dstq + dstrideq]
+    dec      heightd
+    jnz        .loop
+    REP_RET
+%endm
+
+INIT_XMM ssse3
+SUBPIX_HFILTER16 h8
+SUBPIX_HFILTER8  h8
+SUBPIX_HFILTER4  h8
+
+;-------------------------------------------------------------------------------
+
+; TODO(Linfeng): Detect cpu type and choose the code with better performance.
+%define X86_SUBPIX_VFILTER_PREFER_SLOW_CELERON 1
+
+%if AOM_ARCH_X86_64 && X86_SUBPIX_VFILTER_PREFER_SLOW_CELERON
+    %define NUM_GENERAL_REG_USED 9
+%else
+    %define NUM_GENERAL_REG_USED 6
+%endif
+
+%macro SUBPIX_VFILTER 2
+cglobal filter_block1d%2_%1, 6, NUM_GENERAL_REG_USED, 15, LOCAL_VARS_SIZE, \
+                             src, sstride, dst, dstride, height, filter
+    mova          m4, [filterq]
+    SETUP_LOCAL_VARS
+
+%ifidn %2, 8
+    %define                movx  movh
+%else
+    %define                movx  movd
+%endif
+
+    dec                 heightd
+
+%if AOM_ARCH_X86 || X86_SUBPIX_VFILTER_PREFER_SLOW_CELERON
+
+%if AOM_ARCH_X86_64
+    %define               src1q  r7
+    %define           sstride6q  r8
+    %define          dst_stride  dstrideq
+%else
+    %define               src1q  filterq
+    %define           sstride6q  dstrideq
+    %define          dst_stride  dstridemp
+%endif
+    mov                   src1q, srcq
+    add                   src1q, sstrideq
+    lea               sstride6q, [sstrideq + sstrideq * 4]
+    add               sstride6q, sstrideq                   ;pitch * 6
+
+.loop:
+    ;Do two rows at once
+    movx                     m0, [srcq                ]     ;A
+    movx                     m1, [src1q               ]     ;B
+    punpcklbw                m0, m1                         ;A B
+    movx                     m2, [srcq + sstrideq * 2 ]     ;C
+    pmaddubsw                m0, k0k1
+    mova                     m6, m2
+    movx                     m3, [src1q + sstrideq * 2]     ;D
+    punpcklbw                m2, m3                         ;C D
+    pmaddubsw                m2, k2k3
+    movx                     m4, [srcq + sstrideq * 4 ]     ;E
+    mova                     m7, m4
+    movx                     m5, [src1q + sstrideq * 4]     ;F
+    punpcklbw                m4, m5                         ;E F
+    pmaddubsw                m4, k4k5
+    punpcklbw                m1, m6                         ;A B next iter
+    movx                     m6, [srcq + sstride6q    ]     ;G
+    punpcklbw                m5, m6                         ;E F next iter
+    punpcklbw                m3, m7                         ;C D next iter
+    pmaddubsw                m5, k4k5
+    movx                     m7, [src1q + sstride6q   ]     ;H
+    punpcklbw                m6, m7                         ;G H
+    pmaddubsw                m6, k6k7
+    pmaddubsw                m3, k2k3
+    pmaddubsw                m1, k0k1
+    paddsw                   m0, m4
+    paddsw                   m2, m6
+    movx                     m6, [srcq + sstrideq * 8 ]     ;H next iter
+    punpcklbw                m7, m6
+    pmaddubsw                m7, k6k7
+    paddsw                   m0, m2
+    paddsw                   m0, krd
+    psraw                    m0, 7
+    paddsw                   m1, m5
+%ifidn %1, v8_add_src
+    pxor                     m6, m6
+    movu                     m4, [srcq]
+    punpcklbw                m4, m6
+    paddsw                   m0, m4
+%endif
+    packuswb                 m0, m0
+
+    paddsw                   m3, m7
+    paddsw                   m1, m3
+    paddsw                   m1, krd
+    psraw                    m1, 7
+%ifidn %1, v8_add_src
+    movu                     m4, [src1q]
+    punpcklbw                m4, m6
+    paddsw                   m1, m4
+%endif
+    lea                    srcq, [srcq + sstrideq * 2 ]
+    lea                   src1q, [src1q + sstrideq * 2]
+    packuswb                 m1, m1
+
+%ifidn %1, v8_avg
+    movx                     m2, [dstq]
+    pavgb                    m0, m2
+%endif
+    movx                 [dstq], m0
+    add                    dstq, dst_stride
+%ifidn %1, v8_avg
+    movx                     m3, [dstq]
+    pavgb                    m1, m3
+%endif
+    movx                 [dstq], m1
+    add                    dstq, dst_stride
+    sub                 heightd, 2
+    jg                    .loop
+
+    ; Do last row if output_height is odd
+    jne                   .done
+
+    movx                     m0, [srcq                ]     ;A
+    movx                     m1, [srcq + sstrideq     ]     ;B
+    movx                     m6, [srcq + sstride6q    ]     ;G
+    punpcklbw                m0, m1                         ;A B
+    movx                     m7, [src1q + sstride6q   ]     ;H
+    pmaddubsw                m0, k0k1
+    movx                     m2, [srcq + sstrideq * 2 ]     ;C
+    punpcklbw                m6, m7                         ;G H
+    movx                     m3, [src1q + sstrideq * 2]     ;D
+    pmaddubsw                m6, k6k7
+    movx                     m4, [srcq + sstrideq * 4 ]     ;E
+    punpcklbw                m2, m3                         ;C D
+    movx                     m5, [src1q + sstrideq * 4]     ;F
+    punpcklbw                m4, m5                         ;E F
+    pmaddubsw                m2, k2k3
+    pmaddubsw                m4, k4k5
+    paddsw                   m2, m6
+    paddsw                   m0, m4
+    paddsw                   m0, m2
+    paddsw                   m0, krd
+    psraw                    m0, 7
+%ifidn %1, v8_add_src
+    pxor                     m6, m6
+    movu                     m4, [srcq]
+    punpcklbw                m4, m6
+    paddsw                   m0, m4
+%endif
+    packuswb                 m0, m0
+%ifidn %1, v8_avg
+    movx                     m1, [dstq]
+    pavgb                    m0, m1
+%endif
+    movx                 [dstq], m0
+
+%else
+    ; AOM_ARCH_X86_64
+
+    movx                     m0, [srcq                ]     ;A
+    movx                     m1, [srcq + sstrideq     ]     ;B
+    lea                    srcq, [srcq + sstrideq * 2 ]
+    movx                     m2, [srcq]                     ;C
+    movx                     m3, [srcq + sstrideq]          ;D
+    lea                    srcq, [srcq + sstrideq * 2 ]
+    movx                     m4, [srcq]                     ;E
+    movx                     m5, [srcq + sstrideq]          ;F
+    lea                    srcq, [srcq + sstrideq * 2 ]
+    movx                     m6, [srcq]                     ;G
+    punpcklbw                m0, m1                         ;A B
+    punpcklbw                m1, m2                         ;A B next iter
+    punpcklbw                m2, m3                         ;C D
+    punpcklbw                m3, m4                         ;C D next iter
+    punpcklbw                m4, m5                         ;E F
+    punpcklbw                m5, m6                         ;E F next iter
+
+.loop:
+    ;Do two rows at once
+    movx                     m7, [srcq + sstrideq]          ;H
+    lea                    srcq, [srcq + sstrideq * 2 ]
+    movx                    m14, [srcq]                     ;H next iter
+    punpcklbw                m6, m7                         ;G H
+    punpcklbw                m7, m14                        ;G H next iter
+    pmaddubsw                m8, m0, k0k1
+    pmaddubsw                m9, m1, k0k1
+    mova                     m0, m2
+    mova                     m1, m3
+    pmaddubsw               m10, m2, k2k3
+    pmaddubsw               m11, m3, k2k3
+    mova                     m2, m4
+    mova                     m3, m5
+    pmaddubsw                m4, k4k5
+    pmaddubsw                m5, k4k5
+    paddsw                   m8, m4
+    paddsw                   m9, m5
+    mova                     m4, m6
+    mova                     m5, m7
+    pmaddubsw                m6, k6k7
+    pmaddubsw                m7, k6k7
+    paddsw                  m10, m6
+    paddsw                  m11, m7
+    paddsw                   m8, m10
+    paddsw                   m9, m11
+    mova                     m6, m14
+    paddsw                   m8, krd
+    paddsw                   m9, krd
+    psraw                    m8, 7
+    psraw                    m9, 7
+%ifidn %2, 4
+    packuswb                 m8, m8
+    packuswb                 m9, m9
+%else
+    packuswb                 m8, m9
+%endif
+
+%ifidn %1, v8_avg
+    movx                     m7, [dstq]
+%ifidn %2, 4
+    movx                    m10, [dstq + dstrideq]
+    pavgb                    m9, m10
+%else
+    movhpd                   m7, [dstq + dstrideq]
+%endif
+    pavgb                    m8, m7
+%endif
+    movx                 [dstq], m8
+%ifidn %2, 4
+    movx      [dstq + dstrideq], m9
+%else
+    movhpd    [dstq + dstrideq], m8
+%endif
+
+    lea                    dstq, [dstq + dstrideq * 2 ]
+    sub                 heightd, 2
+    jg                    .loop
+
+    ; Do last row if output_height is odd
+    jne                   .done
+
+    movx                     m7, [srcq + sstrideq]          ;H
+    punpcklbw                m6, m7                         ;G H
+    pmaddubsw                m0, k0k1
+    pmaddubsw                m2, k2k3
+    pmaddubsw                m4, k4k5
+    pmaddubsw                m6, k6k7
+    paddsw                   m0, m4
+    paddsw                   m2, m6
+    paddsw                   m0, m2
+    paddsw                   m0, krd
+    psraw                    m0, 7
+    packuswb                 m0, m0
+%ifidn %1, v8_avg
+    movx                     m1, [dstq]
+    pavgb                    m0, m1
+%endif
+    movx                 [dstq], m0
+
+%endif ; AOM_ARCH_X86_64
+
+.done:
+    REP_RET
+
+%endm
+
+;-------------------------------------------------------------------------------
+%macro SUBPIX_VFILTER16 1
+cglobal filter_block1d16_%1, 6, NUM_GENERAL_REG_USED, 16, LOCAL_VARS_SIZE, \
+                             src, sstride, dst, dstride, height, filter
+    mova                     m4, [filterq]
+    SETUP_LOCAL_VARS
+
+%if AOM_ARCH_X86 || X86_SUBPIX_VFILTER_PREFER_SLOW_CELERON
+
+%if AOM_ARCH_X86_64
+    %define               src1q  r7
+    %define           sstride6q  r8
+    %define          dst_stride  dstrideq
+%else
+    %define               src1q  filterq
+    %define           sstride6q  dstrideq
+    %define          dst_stride  dstridemp
+%endif
+    lea                   src1q, [srcq + sstrideq]
+    lea               sstride6q, [sstrideq + sstrideq * 4]
+    add               sstride6q, sstrideq                   ;pitch * 6
+
+.loop:
+    movh                     m0, [srcq                ]     ;A
+    movh                     m1, [src1q               ]     ;B
+    movh                     m2, [srcq + sstrideq * 2 ]     ;C
+    movh                     m3, [src1q + sstrideq * 2]     ;D
+    movh                     m4, [srcq + sstrideq * 4 ]     ;E
+    movh                     m5, [src1q + sstrideq * 4]     ;F
+
+    punpcklbw                m0, m1                         ;A B
+    movh                     m6, [srcq + sstride6q]         ;G
+    punpcklbw                m2, m3                         ;C D
+    movh                     m7, [src1q + sstride6q]        ;H
+    punpcklbw                m4, m5                         ;E F
+    pmaddubsw                m0, k0k1
+    movh                     m3, [srcq + 8]                 ;A
+    pmaddubsw                m2, k2k3
+    punpcklbw                m6, m7                         ;G H
+    movh                     m5, [srcq + sstrideq + 8]      ;B
+    pmaddubsw                m4, k4k5
+    punpcklbw                m3, m5                         ;A B
+    movh                     m7, [srcq + sstrideq * 2 + 8]  ;C
+    pmaddubsw                m6, k6k7
+    movh                     m5, [src1q + sstrideq * 2 + 8] ;D
+    punpcklbw                m7, m5                         ;C D
+    paddsw                   m2, m6
+    pmaddubsw                m3, k0k1
+    movh                     m1, [srcq + sstrideq * 4 + 8]  ;E
+    paddsw                   m0, m4
+    pmaddubsw                m7, k2k3
+    movh                     m6, [src1q + sstrideq * 4 + 8] ;F
+    punpcklbw                m1, m6                         ;E F
+    paddsw                   m0, m2
+    paddsw                   m0, krd
+    movh                     m2, [srcq + sstride6q + 8]     ;G
+    pmaddubsw                m1, k4k5
+    movh                     m5, [src1q + sstride6q + 8]    ;H
+    psraw                    m0, 7
+    punpcklbw                m2, m5                         ;G H
+    pmaddubsw                m2, k6k7
+    paddsw                   m7, m2
+    paddsw                   m3, m1
+    paddsw                   m3, m7
+    paddsw                   m3, krd
+    psraw                    m3, 7
+%ifidn %1, v8_add_src
+    pxor                     m6, m6
+    movu                     m4, [src1q + 2 * sstrideq] ; Fetch from 3 rows down
+    mova                     m5, m4
+    punpcklbw                m4, m6
+    punpckhbw                m5, m6
+    paddsw                   m0, m4
+    paddsw                   m3, m5
+%endif
+    packuswb                 m0, m3
+
+    add                    srcq, sstrideq
+    add                   src1q, sstrideq
+%ifidn %1, v8_avg
+    pavgb                    m0, [dstq]
+%endif
+    mova                 [dstq], m0
+    add                    dstq, dst_stride
+    dec                 heightd
+    jnz                   .loop
+    REP_RET
+
+%else
+    ; AOM_ARCH_X86_64
+    dec                 heightd
+
+    movu                     m1, [srcq                ]     ;A
+    movu                     m3, [srcq + sstrideq     ]     ;B
+    lea                    srcq, [srcq + sstrideq * 2]
+    punpcklbw                m0, m1, m3                     ;A B
+    punpckhbw                m1, m3                         ;A B
+    movu                     m5, [srcq]                     ;C
+    punpcklbw                m2, m3, m5                     ;A B next iter
+    punpckhbw                m3, m5                         ;A B next iter
+    mova                   tmp0, m2                         ;store to stack
+    mova                   tmp1, m3                         ;store to stack
+    movu                     m7, [srcq + sstrideq]          ;D
+    lea                    srcq, [srcq + sstrideq * 2]
+    punpcklbw                m4, m5, m7                     ;C D
+    punpckhbw                m5, m7                         ;C D
+    movu                     m9, [srcq]                     ;E
+    punpcklbw                m6, m7, m9                     ;C D next iter
+    punpckhbw                m7, m9                         ;C D next iter
+    movu                    m11, [srcq + sstrideq]          ;F
+    lea                    srcq, [srcq + sstrideq * 2]
+    punpcklbw                m8, m9, m11                    ;E F
+    punpckhbw                m9, m11                        ;E F
+    movu                     m2, [srcq]                     ;G
+    punpcklbw               m10, m11, m2                    ;E F next iter
+    punpckhbw               m11, m2                         ;E F next iter
+
+.loop:
+    ;Do two rows at once
+    pmaddubsw               m13, m0, k0k1
+    mova                     m0, m4
+    pmaddubsw               m14, m8, k4k5
+    pmaddubsw               m15, m4, k2k3
+    mova                     m4, m8
+    paddsw                  m13, m14
+    movu                     m3, [srcq + sstrideq]          ;H
+    lea                    srcq, [srcq + sstrideq * 2]
+    punpcklbw               m14, m2, m3                     ;G H
+    mova                     m8, m14
+    pmaddubsw               m14, k6k7
+    paddsw                  m15, m14
+    paddsw                  m13, m15
+    paddsw                  m13, krd
+    psraw                   m13, 7
+
+    pmaddubsw               m14, m1, k0k1
+    pmaddubsw                m1, m9, k4k5
+    pmaddubsw               m15, m5, k2k3
+    paddsw                  m14, m1
+    mova                     m1, m5
+    mova                     m5, m9
+    punpckhbw                m2, m3                         ;G H
+    mova                     m9, m2
+    pmaddubsw                m2, k6k7
+    paddsw                  m15, m2
+    paddsw                  m14, m15
+    paddsw                  m14, krd
+    psraw                   m14, 7
+    packuswb                m13, m14
+%ifidn %1, v8_avg
+    pavgb                   m13, [dstq]
+%endif
+    mova                 [dstq], m13
+
+    ; next iter
+    pmaddubsw               m15, tmp0, k0k1
+    pmaddubsw               m14, m10, k4k5
+    pmaddubsw               m13, m6, k2k3
+    paddsw                  m15, m14
+    mova                   tmp0, m6
+    mova                     m6, m10
+    movu                     m2, [srcq]                     ;G next iter
+    punpcklbw               m14, m3, m2                     ;G H next iter
+    mova                    m10, m14
+    pmaddubsw               m14, k6k7
+    paddsw                  m13, m14
+    paddsw                  m15, m13
+    paddsw                  m15, krd
+    psraw                   m15, 7
+
+    pmaddubsw               m14, tmp1, k0k1
+    mova                   tmp1, m7
+    pmaddubsw               m13, m7, k2k3
+    mova                     m7, m11
+    pmaddubsw               m11, k4k5
+    paddsw                  m14, m11
+    punpckhbw                m3, m2                         ;G H next iter
+    mova                    m11, m3
+    pmaddubsw                m3, k6k7
+    paddsw                  m13, m3
+    paddsw                  m14, m13
+    paddsw                  m14, krd
+    psraw                   m14, 7
+    packuswb                m15, m14
+%ifidn %1, v8_avg
+    pavgb                   m15, [dstq + dstrideq]
+%endif
+    mova      [dstq + dstrideq], m15
+    lea                    dstq, [dstq + dstrideq * 2]
+    sub                 heightd, 2
+    jg                    .loop
+
+    ; Do last row if output_height is odd
+    jne                   .done
+
+    movu                     m3, [srcq + sstrideq]          ;H
+    punpcklbw                m6, m2, m3                     ;G H
+    punpckhbw                m2, m3                         ;G H
+    pmaddubsw                m0, k0k1
+    pmaddubsw                m1, k0k1
+    pmaddubsw                m4, k2k3
+    pmaddubsw                m5, k2k3
+    pmaddubsw                m8, k4k5
+    pmaddubsw                m9, k4k5
+    pmaddubsw                m6, k6k7
+    pmaddubsw                m2, k6k7
+    paddsw                   m0, m8
+    paddsw                   m1, m9
+    paddsw                   m4, m6
+    paddsw                   m5, m2
+    paddsw                   m0, m4
+    paddsw                   m1, m5
+    paddsw                   m0, krd
+    paddsw                   m1, krd
+    psraw                    m0, 7
+    psraw                    m1, 7
+    packuswb                 m0, m1
+%ifidn %1, v8_avg
+    pavgb                    m0, [dstq]
+%endif
+    mova                 [dstq], m0
+
+.done:
+    REP_RET
+
+%endif ; AOM_ARCH_X86_64
+
+%endm
+
+INIT_XMM ssse3
+SUBPIX_VFILTER16     v8
+SUBPIX_VFILTER       v8, 8
+SUBPIX_VFILTER       v8, 4
diff --git a/third_party/aom/aom_dsp/x86/aom_subpixel_bilinear_sse2.asm b/third_party/aom/aom_dsp/x86/aom_subpixel_bilinear_sse2.asm
new file mode 100644
index 0000000000..90dd55a4be
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/aom_subpixel_bilinear_sse2.asm
@@ -0,0 +1,295 @@
+;
+; Copyright (c) 2016, 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 "aom_ports/x86_abi_support.asm"
+
+%macro GET_PARAM_4 0
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+    mov         rcx, 0x0400040
+
+    movdqa      xmm3, [rdx]                 ;load filters
+    pshuflw     xmm4, xmm3, 11111111b       ;k3
+    psrldq      xmm3, 8
+    pshuflw     xmm3, xmm3, 0b              ;k4
+    punpcklqdq  xmm4, xmm3                  ;k3k4
+
+    movq        xmm3, rcx                   ;rounding
+    pshufd      xmm3, xmm3, 0
+
+    pxor        xmm2, xmm2
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+%endm
+
+%macro APPLY_FILTER_4 1
+
+    punpckldq   xmm0, xmm1                  ;two row in one register
+    punpcklbw   xmm0, xmm2                  ;unpack to word
+    pmullw      xmm0, xmm4                  ;multiply the filter factors
+
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 8
+    paddsw      xmm0, xmm1
+
+    paddsw      xmm0, xmm3                  ;rounding
+    psraw       xmm0, 7                     ;shift
+    packuswb    xmm0, xmm0                  ;pack to byte
+
+%if %1
+    movd        xmm1, [rdi]
+    pavgb       xmm0, xmm1
+%endif
+
+    movd        [rdi], xmm0
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+%endm
+
+%macro GET_PARAM 0
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+    mov         rcx, 0x0400040
+
+    movdqa      xmm7, [rdx]                 ;load filters
+
+    pshuflw     xmm6, xmm7, 11111111b       ;k3
+    pshufhw     xmm7, xmm7, 0b              ;k4
+    punpcklwd   xmm6, xmm6
+    punpckhwd   xmm7, xmm7
+
+    movq        xmm4, rcx                   ;rounding
+    pshufd      xmm4, xmm4, 0
+
+    pxor        xmm5, xmm5
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+%endm
+
+%macro APPLY_FILTER_8 1
+    punpcklbw   xmm0, xmm5
+    punpcklbw   xmm1, xmm5
+
+    pmullw      xmm0, xmm6
+    pmullw      xmm1, xmm7
+    paddsw      xmm0, xmm1
+    paddsw      xmm0, xmm4                  ;rounding
+    psraw       xmm0, 7                     ;shift
+    packuswb    xmm0, xmm0                  ;pack back to byte
+%if %1
+    movq        xmm1, [rdi]
+    pavgb       xmm0, xmm1
+%endif
+    movq        [rdi], xmm0                 ;store the result
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+%endm
+
+%macro APPLY_FILTER_16 1
+    punpcklbw   xmm0, xmm5
+    punpcklbw   xmm1, xmm5
+    punpckhbw   xmm2, xmm5
+    punpckhbw   xmm3, xmm5
+
+    pmullw      xmm0, xmm6
+    pmullw      xmm1, xmm7
+    pmullw      xmm2, xmm6
+    pmullw      xmm3, xmm7
+
+    paddsw      xmm0, xmm1
+    paddsw      xmm2, xmm3
+
+    paddsw      xmm0, xmm4                  ;rounding
+    paddsw      xmm2, xmm4
+    psraw       xmm0, 7                     ;shift
+    psraw       xmm2, 7
+    packuswb    xmm0, xmm2                  ;pack back to byte
+%if %1
+    movdqu      xmm1, [rdi]
+    pavgb       xmm0, xmm1
+%endif
+    movdqu      [rdi], xmm0                 ;store the result
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+%endm
+
+SECTION .text
+
+globalsym(aom_filter_block1d4_v2_sse2)
+sym(aom_filter_block1d4_v2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM_4
+.loop:
+    movd        xmm0, [rsi]                 ;load src
+    movd        xmm1, [rsi + rax]
+
+    APPLY_FILTER_4 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+globalsym(aom_filter_block1d8_v2_sse2)
+sym(aom_filter_block1d8_v2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movq        xmm0, [rsi]                 ;0
+    movq        xmm1, [rsi + rax]           ;1
+
+    APPLY_FILTER_8 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+globalsym(aom_filter_block1d16_v2_sse2)
+sym(aom_filter_block1d16_v2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movdqu        xmm0, [rsi]               ;0
+    movdqu        xmm1, [rsi + rax]         ;1
+    movdqa        xmm2, xmm0
+    movdqa        xmm3, xmm1
+
+    APPLY_FILTER_16 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+globalsym(aom_filter_block1d4_h2_sse2)
+sym(aom_filter_block1d4_h2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM_4
+.loop:
+    movdqu      xmm0, [rsi]                 ;load src
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 1
+
+    APPLY_FILTER_4 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+globalsym(aom_filter_block1d8_h2_sse2)
+sym(aom_filter_block1d8_h2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movdqu      xmm0, [rsi]                 ;load src
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 1
+
+    APPLY_FILTER_8 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+globalsym(aom_filter_block1d16_h2_sse2)
+sym(aom_filter_block1d16_h2_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movdqu      xmm0,   [rsi]               ;load src
+    movdqu      xmm1,   [rsi + 1]
+    movdqa      xmm2, xmm0
+    movdqa      xmm3, xmm1
+
+    APPLY_FILTER_16 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
diff --git a/third_party/aom/aom_dsp/x86/aom_subpixel_bilinear_ssse3.asm b/third_party/aom/aom_dsp/x86/aom_subpixel_bilinear_ssse3.asm
new file mode 100644
index 0000000000..253bc26d38
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/aom_subpixel_bilinear_ssse3.asm
@@ -0,0 +1,267 @@
+;
+; Copyright (c) 2016, 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 "aom_ports/x86_abi_support.asm"
+
+%macro GET_PARAM_4 0
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+    mov         ecx, 0x01000100
+
+    movdqa      xmm3, [rdx]                 ;load filters
+    psrldq      xmm3, 6
+    packsswb    xmm3, xmm3
+    pshuflw     xmm3, xmm3, 0b              ;k3_k4
+
+    movd        xmm2, ecx                   ;rounding_shift
+    pshufd      xmm2, xmm2, 0
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+%endm
+
+%macro APPLY_FILTER_4 1
+    punpcklbw   xmm0, xmm1
+    pmaddubsw   xmm0, xmm3
+
+    pmulhrsw    xmm0, xmm2                  ;rounding(+64)+shift(>>7)
+    packuswb    xmm0, xmm0                  ;pack to byte
+
+%if %1
+    movd        xmm1, [rdi]
+    pavgb       xmm0, xmm1
+%endif
+    movd        [rdi], xmm0
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+%endm
+
+%macro GET_PARAM 0
+    mov         rdx, arg(5)                 ;filter ptr
+    mov         rsi, arg(0)                 ;src_ptr
+    mov         rdi, arg(2)                 ;output_ptr
+    mov         ecx, 0x01000100
+
+    movdqa      xmm7, [rdx]                 ;load filters
+    psrldq      xmm7, 6
+    packsswb    xmm7, xmm7
+    pshuflw     xmm7, xmm7, 0b              ;k3_k4
+    punpcklwd   xmm7, xmm7
+
+    movd        xmm6, ecx                   ;rounding_shift
+    pshufd      xmm6, xmm6, 0
+
+    movsxd      rax, DWORD PTR arg(1)       ;pixels_per_line
+    movsxd      rdx, DWORD PTR arg(3)       ;out_pitch
+    movsxd      rcx, DWORD PTR arg(4)       ;output_height
+%endm
+
+%macro APPLY_FILTER_8 1
+    punpcklbw   xmm0, xmm1
+    pmaddubsw   xmm0, xmm7
+
+    pmulhrsw    xmm0, xmm6                  ;rounding(+64)+shift(>>7)
+    packuswb    xmm0, xmm0                  ;pack back to byte
+
+%if %1
+    movq        xmm1, [rdi]
+    pavgb       xmm0, xmm1
+%endif
+    movq        [rdi], xmm0                 ;store the result
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+%endm
+
+%macro APPLY_FILTER_16 1
+    punpcklbw   xmm0, xmm1
+    punpckhbw   xmm2, xmm1
+    pmaddubsw   xmm0, xmm7
+    pmaddubsw   xmm2, xmm7
+
+    pmulhrsw    xmm0, xmm6                  ;rounding(+64)+shift(>>7)
+    pmulhrsw    xmm2, xmm6
+    packuswb    xmm0, xmm2                  ;pack back to byte
+
+%if %1
+    movdqu      xmm1, [rdi]
+    pavgb       xmm0, xmm1
+%endif
+    movdqu      [rdi], xmm0                 ;store the result
+
+    lea         rsi, [rsi + rax]
+    lea         rdi, [rdi + rdx]
+    dec         rcx
+%endm
+
+SECTION .text
+
+globalsym(aom_filter_block1d4_v2_ssse3)
+sym(aom_filter_block1d4_v2_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM_4
+.loop:
+    movd        xmm0, [rsi]                 ;load src
+    movd        xmm1, [rsi + rax]
+
+    APPLY_FILTER_4 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+globalsym(aom_filter_block1d8_v2_ssse3)
+sym(aom_filter_block1d8_v2_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movq        xmm0, [rsi]                 ;0
+    movq        xmm1, [rsi + rax]           ;1
+
+    APPLY_FILTER_8 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+globalsym(aom_filter_block1d16_v2_ssse3)
+sym(aom_filter_block1d16_v2_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movdqu        xmm0, [rsi]               ;0
+    movdqu        xmm1, [rsi + rax]         ;1
+    movdqa        xmm2, xmm0
+
+    APPLY_FILTER_16 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+globalsym(aom_filter_block1d4_h2_ssse3)
+sym(aom_filter_block1d4_h2_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM_4
+.loop:
+    movdqu      xmm0, [rsi]                 ;load src
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 1
+
+    APPLY_FILTER_4 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+globalsym(aom_filter_block1d8_h2_ssse3)
+sym(aom_filter_block1d8_h2_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movdqu      xmm0, [rsi]                 ;load src
+    movdqa      xmm1, xmm0
+    psrldq      xmm1, 1
+
+    APPLY_FILTER_8 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+globalsym(aom_filter_block1d16_h2_ssse3)
+sym(aom_filter_block1d16_h2_ssse3):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    GET_PARAM
+.loop:
+    movdqu      xmm0,   [rsi]               ;load src
+    movdqu      xmm1,   [rsi + 1]
+    movdqa      xmm2, xmm0
+
+    APPLY_FILTER_16 0
+    jnz         .loop
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
diff --git a/third_party/aom/aom_dsp/x86/avg_intrin_avx2.c b/third_party/aom/aom_dsp/x86/avg_intrin_avx2.c
new file mode 100644
index 0000000000..49fcd72098
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/avg_intrin_avx2.c
@@ -0,0 +1,897 @@
+/*
+ * Copyright (c) 2016, 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"
+#include "aom/aom_integer.h"
+#include "aom_dsp/x86/bitdepth_conversion_avx2.h"
+#include "aom_ports/mem.h"
+
+static INLINE void sign_extend_16bit_to_32bit_avx2(__m256i in, __m256i zero,
+                                                   __m256i *out_lo,
+                                                   __m256i *out_hi) {
+  const __m256i sign_bits = _mm256_cmpgt_epi16(zero, in);
+  *out_lo = _mm256_unpacklo_epi16(in, sign_bits);
+  *out_hi = _mm256_unpackhi_epi16(in, sign_bits);
+}
+
+static void hadamard_col8x2_avx2(__m256i *in, int iter) {
+  __m256i a0 = in[0];
+  __m256i a1 = in[1];
+  __m256i a2 = in[2];
+  __m256i a3 = in[3];
+  __m256i a4 = in[4];
+  __m256i a5 = in[5];
+  __m256i a6 = in[6];
+  __m256i a7 = in[7];
+
+  __m256i b0 = _mm256_add_epi16(a0, a1);
+  __m256i b1 = _mm256_sub_epi16(a0, a1);
+  __m256i b2 = _mm256_add_epi16(a2, a3);
+  __m256i b3 = _mm256_sub_epi16(a2, a3);
+  __m256i b4 = _mm256_add_epi16(a4, a5);
+  __m256i b5 = _mm256_sub_epi16(a4, a5);
+  __m256i b6 = _mm256_add_epi16(a6, a7);
+  __m256i b7 = _mm256_sub_epi16(a6, a7);
+
+  a0 = _mm256_add_epi16(b0, b2);
+  a1 = _mm256_add_epi16(b1, b3);
+  a2 = _mm256_sub_epi16(b0, b2);
+  a3 = _mm256_sub_epi16(b1, b3);
+  a4 = _mm256_add_epi16(b4, b6);
+  a5 = _mm256_add_epi16(b5, b7);
+  a6 = _mm256_sub_epi16(b4, b6);
+  a7 = _mm256_sub_epi16(b5, b7);
+
+  if (iter == 0) {
+    b0 = _mm256_add_epi16(a0, a4);
+    b7 = _mm256_add_epi16(a1, a5);
+    b3 = _mm256_add_epi16(a2, a6);
+    b4 = _mm256_add_epi16(a3, a7);
+    b2 = _mm256_sub_epi16(a0, a4);
+    b6 = _mm256_sub_epi16(a1, a5);
+    b1 = _mm256_sub_epi16(a2, a6);
+    b5 = _mm256_sub_epi16(a3, a7);
+
+    a0 = _mm256_unpacklo_epi16(b0, b1);
+    a1 = _mm256_unpacklo_epi16(b2, b3);
+    a2 = _mm256_unpackhi_epi16(b0, b1);
+    a3 = _mm256_unpackhi_epi16(b2, b3);
+    a4 = _mm256_unpacklo_epi16(b4, b5);
+    a5 = _mm256_unpacklo_epi16(b6, b7);
+    a6 = _mm256_unpackhi_epi16(b4, b5);
+    a7 = _mm256_unpackhi_epi16(b6, b7);
+
+    b0 = _mm256_unpacklo_epi32(a0, a1);
+    b1 = _mm256_unpacklo_epi32(a4, a5);
+    b2 = _mm256_unpackhi_epi32(a0, a1);
+    b3 = _mm256_unpackhi_epi32(a4, a5);
+    b4 = _mm256_unpacklo_epi32(a2, a3);
+    b5 = _mm256_unpacklo_epi32(a6, a7);
+    b6 = _mm256_unpackhi_epi32(a2, a3);
+    b7 = _mm256_unpackhi_epi32(a6, a7);
+
+    in[0] = _mm256_unpacklo_epi64(b0, b1);
+    in[1] = _mm256_unpackhi_epi64(b0, b1);
+    in[2] = _mm256_unpacklo_epi64(b2, b3);
+    in[3] = _mm256_unpackhi_epi64(b2, b3);
+    in[4] = _mm256_unpacklo_epi64(b4, b5);
+    in[5] = _mm256_unpackhi_epi64(b4, b5);
+    in[6] = _mm256_unpacklo_epi64(b6, b7);
+    in[7] = _mm256_unpackhi_epi64(b6, b7);
+  } else {
+    in[0] = _mm256_add_epi16(a0, a4);
+    in[7] = _mm256_add_epi16(a1, a5);
+    in[3] = _mm256_add_epi16(a2, a6);
+    in[4] = _mm256_add_epi16(a3, a7);
+    in[2] = _mm256_sub_epi16(a0, a4);
+    in[6] = _mm256_sub_epi16(a1, a5);
+    in[1] = _mm256_sub_epi16(a2, a6);
+    in[5] = _mm256_sub_epi16(a3, a7);
+  }
+}
+
+void aom_hadamard_lp_8x8_dual_avx2(const int16_t *src_diff,
+                                   ptrdiff_t src_stride, int16_t *coeff) {
+  __m256i src[8];
+  src[0] = _mm256_loadu_si256((const __m256i *)src_diff);
+  src[1] = _mm256_loadu_si256((const __m256i *)(src_diff += src_stride));
+  src[2] = _mm256_loadu_si256((const __m256i *)(src_diff += src_stride));
+  src[3] = _mm256_loadu_si256((const __m256i *)(src_diff += src_stride));
+  src[4] = _mm256_loadu_si256((const __m256i *)(src_diff += src_stride));
+  src[5] = _mm256_loadu_si256((const __m256i *)(src_diff += src_stride));
+  src[6] = _mm256_loadu_si256((const __m256i *)(src_diff += src_stride));
+  src[7] = _mm256_loadu_si256((const __m256i *)(src_diff + src_stride));
+
+  hadamard_col8x2_avx2(src, 0);
+  hadamard_col8x2_avx2(src, 1);
+
+  _mm256_storeu_si256((__m256i *)coeff,
+                      _mm256_permute2x128_si256(src[0], src[1], 0x20));
+  coeff += 16;
+  _mm256_storeu_si256((__m256i *)coeff,
+                      _mm256_permute2x128_si256(src[2], src[3], 0x20));
+  coeff += 16;
+  _mm256_storeu_si256((__m256i *)coeff,
+                      _mm256_permute2x128_si256(src[4], src[5], 0x20));
+  coeff += 16;
+  _mm256_storeu_si256((__m256i *)coeff,
+                      _mm256_permute2x128_si256(src[6], src[7], 0x20));
+  coeff += 16;
+  _mm256_storeu_si256((__m256i *)coeff,
+                      _mm256_permute2x128_si256(src[0], src[1], 0x31));
+  coeff += 16;
+  _mm256_storeu_si256((__m256i *)coeff,
+                      _mm256_permute2x128_si256(src[2], src[3], 0x31));
+  coeff += 16;
+  _mm256_storeu_si256((__m256i *)coeff,
+                      _mm256_permute2x128_si256(src[4], src[5], 0x31));
+  coeff += 16;
+  _mm256_storeu_si256((__m256i *)coeff,
+                      _mm256_permute2x128_si256(src[6], src[7], 0x31));
+}
+
+static INLINE void hadamard_16x16_avx2(const int16_t *src_diff,
+                                       ptrdiff_t src_stride, tran_low_t *coeff,
+                                       int is_final) {
+  DECLARE_ALIGNED(32, int16_t, temp_coeff[16 * 16]);
+  int16_t *t_coeff = temp_coeff;
+  int16_t *coeff16 = (int16_t *)coeff;
+  int idx;
+  for (idx = 0; idx < 2; ++idx) {
+    const int16_t *src_ptr = src_diff + idx * 8 * src_stride;
+    aom_hadamard_lp_8x8_dual_avx2(src_ptr, src_stride,
+                                  t_coeff + (idx * 64 * 2));
+  }
+
+  for (idx = 0; idx < 64; idx += 16) {
+    const __m256i coeff0 = _mm256_loadu_si256((const __m256i *)t_coeff);
+    const __m256i coeff1 = _mm256_loadu_si256((const __m256i *)(t_coeff + 64));
+    const __m256i coeff2 = _mm256_loadu_si256((const __m256i *)(t_coeff + 128));
+    const __m256i coeff3 = _mm256_loadu_si256((const __m256i *)(t_coeff + 192));
+
+    __m256i b0 = _mm256_add_epi16(coeff0, coeff1);
+    __m256i b1 = _mm256_sub_epi16(coeff0, coeff1);
+    __m256i b2 = _mm256_add_epi16(coeff2, coeff3);
+    __m256i b3 = _mm256_sub_epi16(coeff2, coeff3);
+
+    b0 = _mm256_srai_epi16(b0, 1);
+    b1 = _mm256_srai_epi16(b1, 1);
+    b2 = _mm256_srai_epi16(b2, 1);
+    b3 = _mm256_srai_epi16(b3, 1);
+    if (is_final) {
+      store_tran_low(_mm256_add_epi16(b0, b2), coeff);
+      store_tran_low(_mm256_add_epi16(b1, b3), coeff + 64);
+      store_tran_low(_mm256_sub_epi16(b0, b2), coeff + 128);
+      store_tran_low(_mm256_sub_epi16(b1, b3), coeff + 192);
+      coeff += 16;
+    } else {
+      _mm256_storeu_si256((__m256i *)coeff16, _mm256_add_epi16(b0, b2));
+      _mm256_storeu_si256((__m256i *)(coeff16 + 64), _mm256_add_epi16(b1, b3));
+      _mm256_storeu_si256((__m256i *)(coeff16 + 128), _mm256_sub_epi16(b0, b2));
+      _mm256_storeu_si256((__m256i *)(coeff16 + 192), _mm256_sub_epi16(b1, b3));
+      coeff16 += 16;
+    }
+    t_coeff += 16;
+  }
+}
+
+void aom_hadamard_16x16_avx2(const int16_t *src_diff, ptrdiff_t src_stride,
+                             tran_low_t *coeff) {
+  hadamard_16x16_avx2(src_diff, src_stride, coeff, 1);
+}
+
+void aom_hadamard_lp_16x16_avx2(const int16_t *src_diff, ptrdiff_t src_stride,
+                                int16_t *coeff) {
+  int16_t *t_coeff = coeff;
+  for (int idx = 0; idx < 2; ++idx) {
+    const int16_t *src_ptr = src_diff + idx * 8 * src_stride;
+    aom_hadamard_lp_8x8_dual_avx2(src_ptr, src_stride,
+                                  t_coeff + (idx * 64 * 2));
+  }
+
+  for (int idx = 0; idx < 64; idx += 16) {
+    const __m256i coeff0 = _mm256_loadu_si256((const __m256i *)t_coeff);
+    const __m256i coeff1 = _mm256_loadu_si256((const __m256i *)(t_coeff + 64));
+    const __m256i coeff2 = _mm256_loadu_si256((const __m256i *)(t_coeff + 128));
+    const __m256i coeff3 = _mm256_loadu_si256((const __m256i *)(t_coeff + 192));
+
+    __m256i b0 = _mm256_add_epi16(coeff0, coeff1);
+    __m256i b1 = _mm256_sub_epi16(coeff0, coeff1);
+    __m256i b2 = _mm256_add_epi16(coeff2, coeff3);
+    __m256i b3 = _mm256_sub_epi16(coeff2, coeff3);
+
+    b0 = _mm256_srai_epi16(b0, 1);
+    b1 = _mm256_srai_epi16(b1, 1);
+    b2 = _mm256_srai_epi16(b2, 1);
+    b3 = _mm256_srai_epi16(b3, 1);
+    _mm256_storeu_si256((__m256i *)coeff, _mm256_add_epi16(b0, b2));
+    _mm256_storeu_si256((__m256i *)(coeff + 64), _mm256_add_epi16(b1, b3));
+    _mm256_storeu_si256((__m256i *)(coeff + 128), _mm256_sub_epi16(b0, b2));
+    _mm256_storeu_si256((__m256i *)(coeff + 192), _mm256_sub_epi16(b1, b3));
+    coeff += 16;
+    t_coeff += 16;
+  }
+}
+
+void aom_hadamard_32x32_avx2(const int16_t *src_diff, ptrdiff_t src_stride,
+                             tran_low_t *coeff) {
+  // For high bitdepths, it is unnecessary to store_tran_low
+  // (mult/unpack/store), then load_tran_low (load/pack) the same memory in the
+  // next stage.  Output to an intermediate buffer first, then store_tran_low()
+  // in the final stage.
+  DECLARE_ALIGNED(32, int16_t, temp_coeff[32 * 32]);
+  int16_t *t_coeff = temp_coeff;
+  int idx;
+  __m256i coeff0_lo, coeff1_lo, coeff2_lo, coeff3_lo, b0_lo, b1_lo, b2_lo,
+      b3_lo;
+  __m256i coeff0_hi, coeff1_hi, coeff2_hi, coeff3_hi, b0_hi, b1_hi, b2_hi,
+      b3_hi;
+  __m256i b0, b1, b2, b3;
+  const __m256i zero = _mm256_setzero_si256();
+  for (idx = 0; idx < 4; ++idx) {
+    // src_diff: 9 bit, dynamic range [-255, 255]
+    const int16_t *src_ptr =
+        src_diff + (idx >> 1) * 16 * src_stride + (idx & 0x01) * 16;
+    hadamard_16x16_avx2(src_ptr, src_stride,
+                        (tran_low_t *)(t_coeff + idx * 256), 0);
+  }
+
+  for (idx = 0; idx < 256; idx += 16) {
+    const __m256i coeff0 = _mm256_loadu_si256((const __m256i *)t_coeff);
+    const __m256i coeff1 = _mm256_loadu_si256((const __m256i *)(t_coeff + 256));
+    const __m256i coeff2 = _mm256_loadu_si256((const __m256i *)(t_coeff + 512));
+    const __m256i coeff3 = _mm256_loadu_si256((const __m256i *)(t_coeff + 768));
+
+    // Sign extend 16 bit to 32 bit.
+    sign_extend_16bit_to_32bit_avx2(coeff0, zero, &coeff0_lo, &coeff0_hi);
+    sign_extend_16bit_to_32bit_avx2(coeff1, zero, &coeff1_lo, &coeff1_hi);
+    sign_extend_16bit_to_32bit_avx2(coeff2, zero, &coeff2_lo, &coeff2_hi);
+    sign_extend_16bit_to_32bit_avx2(coeff3, zero, &coeff3_lo, &coeff3_hi);
+
+    b0_lo = _mm256_add_epi32(coeff0_lo, coeff1_lo);
+    b0_hi = _mm256_add_epi32(coeff0_hi, coeff1_hi);
+
+    b1_lo = _mm256_sub_epi32(coeff0_lo, coeff1_lo);
+    b1_hi = _mm256_sub_epi32(coeff0_hi, coeff1_hi);
+
+    b2_lo = _mm256_add_epi32(coeff2_lo, coeff3_lo);
+    b2_hi = _mm256_add_epi32(coeff2_hi, coeff3_hi);
+
+    b3_lo = _mm256_sub_epi32(coeff2_lo, coeff3_lo);
+    b3_hi = _mm256_sub_epi32(coeff2_hi, coeff3_hi);
+
+    b0_lo = _mm256_srai_epi32(b0_lo, 2);
+    b1_lo = _mm256_srai_epi32(b1_lo, 2);
+    b2_lo = _mm256_srai_epi32(b2_lo, 2);
+    b3_lo = _mm256_srai_epi32(b3_lo, 2);
+
+    b0_hi = _mm256_srai_epi32(b0_hi, 2);
+    b1_hi = _mm256_srai_epi32(b1_hi, 2);
+    b2_hi = _mm256_srai_epi32(b2_hi, 2);
+    b3_hi = _mm256_srai_epi32(b3_hi, 2);
+
+    b0 = _mm256_packs_epi32(b0_lo, b0_hi);
+    b1 = _mm256_packs_epi32(b1_lo, b1_hi);
+    b2 = _mm256_packs_epi32(b2_lo, b2_hi);
+    b3 = _mm256_packs_epi32(b3_lo, b3_hi);
+
+    store_tran_low(_mm256_add_epi16(b0, b2), coeff);
+    store_tran_low(_mm256_add_epi16(b1, b3), coeff + 256);
+    store_tran_low(_mm256_sub_epi16(b0, b2), coeff + 512);
+    store_tran_low(_mm256_sub_epi16(b1, b3), coeff + 768);
+
+    coeff += 16;
+    t_coeff += 16;
+  }
+}
+
+#if CONFIG_AV1_HIGHBITDEPTH
+static void highbd_hadamard_col8_avx2(__m256i *in, int iter) {
+  __m256i a0 = in[0];
+  __m256i a1 = in[1];
+  __m256i a2 = in[2];
+  __m256i a3 = in[3];
+  __m256i a4 = in[4];
+  __m256i a5 = in[5];
+  __m256i a6 = in[6];
+  __m256i a7 = in[7];
+
+  __m256i b0 = _mm256_add_epi32(a0, a1);
+  __m256i b1 = _mm256_sub_epi32(a0, a1);
+  __m256i b2 = _mm256_add_epi32(a2, a3);
+  __m256i b3 = _mm256_sub_epi32(a2, a3);
+  __m256i b4 = _mm256_add_epi32(a4, a5);
+  __m256i b5 = _mm256_sub_epi32(a4, a5);
+  __m256i b6 = _mm256_add_epi32(a6, a7);
+  __m256i b7 = _mm256_sub_epi32(a6, a7);
+
+  a0 = _mm256_add_epi32(b0, b2);
+  a1 = _mm256_add_epi32(b1, b3);
+  a2 = _mm256_sub_epi32(b0, b2);
+  a3 = _mm256_sub_epi32(b1, b3);
+  a4 = _mm256_add_epi32(b4, b6);
+  a5 = _mm256_add_epi32(b5, b7);
+  a6 = _mm256_sub_epi32(b4, b6);
+  a7 = _mm256_sub_epi32(b5, b7);
+
+  if (iter == 0) {
+    b0 = _mm256_add_epi32(a0, a4);
+    b7 = _mm256_add_epi32(a1, a5);
+    b3 = _mm256_add_epi32(a2, a6);
+    b4 = _mm256_add_epi32(a3, a7);
+    b2 = _mm256_sub_epi32(a0, a4);
+    b6 = _mm256_sub_epi32(a1, a5);
+    b1 = _mm256_sub_epi32(a2, a6);
+    b5 = _mm256_sub_epi32(a3, a7);
+
+    a0 = _mm256_unpacklo_epi32(b0, b1);
+    a1 = _mm256_unpacklo_epi32(b2, b3);
+    a2 = _mm256_unpackhi_epi32(b0, b1);
+    a3 = _mm256_unpackhi_epi32(b2, b3);
+    a4 = _mm256_unpacklo_epi32(b4, b5);
+    a5 = _mm256_unpacklo_epi32(b6, b7);
+    a6 = _mm256_unpackhi_epi32(b4, b5);
+    a7 = _mm256_unpackhi_epi32(b6, b7);
+
+    b0 = _mm256_unpacklo_epi64(a0, a1);
+    b1 = _mm256_unpacklo_epi64(a4, a5);
+    b2 = _mm256_unpackhi_epi64(a0, a1);
+    b3 = _mm256_unpackhi_epi64(a4, a5);
+    b4 = _mm256_unpacklo_epi64(a2, a3);
+    b5 = _mm256_unpacklo_epi64(a6, a7);
+    b6 = _mm256_unpackhi_epi64(a2, a3);
+    b7 = _mm256_unpackhi_epi64(a6, a7);
+
+    in[0] = _mm256_permute2x128_si256(b0, b1, 0x20);
+    in[1] = _mm256_permute2x128_si256(b0, b1, 0x31);
+    in[2] = _mm256_permute2x128_si256(b2, b3, 0x20);
+    in[3] = _mm256_permute2x128_si256(b2, b3, 0x31);
+    in[4] = _mm256_permute2x128_si256(b4, b5, 0x20);
+    in[5] = _mm256_permute2x128_si256(b4, b5, 0x31);
+    in[6] = _mm256_permute2x128_si256(b6, b7, 0x20);
+    in[7] = _mm256_permute2x128_si256(b6, b7, 0x31);
+  } else {
+    in[0] = _mm256_add_epi32(a0, a4);
+    in[7] = _mm256_add_epi32(a1, a5);
+    in[3] = _mm256_add_epi32(a2, a6);
+    in[4] = _mm256_add_epi32(a3, a7);
+    in[2] = _mm256_sub_epi32(a0, a4);
+    in[6] = _mm256_sub_epi32(a1, a5);
+    in[1] = _mm256_sub_epi32(a2, a6);
+    in[5] = _mm256_sub_epi32(a3, a7);
+  }
+}
+
+void aom_highbd_hadamard_8x8_avx2(const int16_t *src_diff, ptrdiff_t src_stride,
+                                  tran_low_t *coeff) {
+  __m128i src16[8];
+  __m256i src32[8];
+
+  src16[0] = _mm_loadu_si128((const __m128i *)src_diff);
+  src16[1] = _mm_loadu_si128((const __m128i *)(src_diff += src_stride));
+  src16[2] = _mm_loadu_si128((const __m128i *)(src_diff += src_stride));
+  src16[3] = _mm_loadu_si128((const __m128i *)(src_diff += src_stride));
+  src16[4] = _mm_loadu_si128((const __m128i *)(src_diff += src_stride));
+  src16[5] = _mm_loadu_si128((const __m128i *)(src_diff += src_stride));
+  src16[6] = _mm_loadu_si128((const __m128i *)(src_diff += src_stride));
+  src16[7] = _mm_loadu_si128((const __m128i *)(src_diff + src_stride));
+
+  src32[0] = _mm256_cvtepi16_epi32(src16[0]);
+  src32[1] = _mm256_cvtepi16_epi32(src16[1]);
+  src32[2] = _mm256_cvtepi16_epi32(src16[2]);
+  src32[3] = _mm256_cvtepi16_epi32(src16[3]);
+  src32[4] = _mm256_cvtepi16_epi32(src16[4]);
+  src32[5] = _mm256_cvtepi16_epi32(src16[5]);
+  src32[6] = _mm256_cvtepi16_epi32(src16[6]);
+  src32[7] = _mm256_cvtepi16_epi32(src16[7]);
+
+  highbd_hadamard_col8_avx2(src32, 0);
+  highbd_hadamard_col8_avx2(src32, 1);
+
+  _mm256_storeu_si256((__m256i *)coeff, src32[0]);
+  coeff += 8;
+  _mm256_storeu_si256((__m256i *)coeff, src32[1]);
+  coeff += 8;
+  _mm256_storeu_si256((__m256i *)coeff, src32[2]);
+  coeff += 8;
+  _mm256_storeu_si256((__m256i *)coeff, src32[3]);
+  coeff += 8;
+  _mm256_storeu_si256((__m256i *)coeff, src32[4]);
+  coeff += 8;
+  _mm256_storeu_si256((__m256i *)coeff, src32[5]);
+  coeff += 8;
+  _mm256_storeu_si256((__m256i *)coeff, src32[6]);
+  coeff += 8;
+  _mm256_storeu_si256((__m256i *)coeff, src32[7]);
+}
+
+void aom_highbd_hadamard_16x16_avx2(const int16_t *src_diff,
+                                    ptrdiff_t src_stride, tran_low_t *coeff) {
+  int idx;
+  tran_low_t *t_coeff = coeff;
+  for (idx = 0; idx < 4; ++idx) {
+    const int16_t *src_ptr =
+        src_diff + (idx >> 1) * 8 * src_stride + (idx & 0x01) * 8;
+    aom_highbd_hadamard_8x8_avx2(src_ptr, src_stride, t_coeff + idx * 64);
+  }
+
+  for (idx = 0; idx < 64; idx += 8) {
+    __m256i coeff0 = _mm256_loadu_si256((const __m256i *)t_coeff);
+    __m256i coeff1 = _mm256_loadu_si256((const __m256i *)(t_coeff + 64));
+    __m256i coeff2 = _mm256_loadu_si256((const __m256i *)(t_coeff + 128));
+    __m256i coeff3 = _mm256_loadu_si256((const __m256i *)(t_coeff + 192));
+
+    __m256i b0 = _mm256_add_epi32(coeff0, coeff1);
+    __m256i b1 = _mm256_sub_epi32(coeff0, coeff1);
+    __m256i b2 = _mm256_add_epi32(coeff2, coeff3);
+    __m256i b3 = _mm256_sub_epi32(coeff2, coeff3);
+
+    b0 = _mm256_srai_epi32(b0, 1);
+    b1 = _mm256_srai_epi32(b1, 1);
+    b2 = _mm256_srai_epi32(b2, 1);
+    b3 = _mm256_srai_epi32(b3, 1);
+
+    coeff0 = _mm256_add_epi32(b0, b2);
+    coeff1 = _mm256_add_epi32(b1, b3);
+    coeff2 = _mm256_sub_epi32(b0, b2);
+    coeff3 = _mm256_sub_epi32(b1, b3);
+
+    _mm256_storeu_si256((__m256i *)coeff, coeff0);
+    _mm256_storeu_si256((__m256i *)(coeff + 64), coeff1);
+    _mm256_storeu_si256((__m256i *)(coeff + 128), coeff2);
+    _mm256_storeu_si256((__m256i *)(coeff + 192), coeff3);
+
+    coeff += 8;
+    t_coeff += 8;
+  }
+}
+
+void aom_highbd_hadamard_32x32_avx2(const int16_t *src_diff,
+                                    ptrdiff_t src_stride, tran_low_t *coeff) {
+  int idx;
+  tran_low_t *t_coeff = coeff;
+  for (idx = 0; idx < 4; ++idx) {
+    const int16_t *src_ptr =
+        src_diff + (idx >> 1) * 16 * src_stride + (idx & 0x01) * 16;
+    aom_highbd_hadamard_16x16_avx2(src_ptr, src_stride, t_coeff + idx * 256);
+  }
+
+  for (idx = 0; idx < 256; idx += 8) {
+    __m256i coeff0 = _mm256_loadu_si256((const __m256i *)t_coeff);
+    __m256i coeff1 = _mm256_loadu_si256((const __m256i *)(t_coeff + 256));
+    __m256i coeff2 = _mm256_loadu_si256((const __m256i *)(t_coeff + 512));
+    __m256i coeff3 = _mm256_loadu_si256((const __m256i *)(t_coeff + 768));
+
+    __m256i b0 = _mm256_add_epi32(coeff0, coeff1);
+    __m256i b1 = _mm256_sub_epi32(coeff0, coeff1);
+    __m256i b2 = _mm256_add_epi32(coeff2, coeff3);
+    __m256i b3 = _mm256_sub_epi32(coeff2, coeff3);
+
+    b0 = _mm256_srai_epi32(b0, 2);
+    b1 = _mm256_srai_epi32(b1, 2);
+    b2 = _mm256_srai_epi32(b2, 2);
+    b3 = _mm256_srai_epi32(b3, 2);
+
+    coeff0 = _mm256_add_epi32(b0, b2);
+    coeff1 = _mm256_add_epi32(b1, b3);
+    coeff2 = _mm256_sub_epi32(b0, b2);
+    coeff3 = _mm256_sub_epi32(b1, b3);
+
+    _mm256_storeu_si256((__m256i *)coeff, coeff0);
+    _mm256_storeu_si256((__m256i *)(coeff + 256), coeff1);
+    _mm256_storeu_si256((__m256i *)(coeff + 512), coeff2);
+    _mm256_storeu_si256((__m256i *)(coeff + 768), coeff3);
+
+    coeff += 8;
+    t_coeff += 8;
+  }
+}
+#endif  // CONFIG_AV1_HIGHBITDEPTH
+
+int aom_satd_avx2(const tran_low_t *coeff, int length) {
+  __m256i accum = _mm256_setzero_si256();
+  int i;
+
+  for (i = 0; i < length; i += 8, coeff += 8) {
+    const __m256i src_line = _mm256_loadu_si256((const __m256i *)coeff);
+    const __m256i abs = _mm256_abs_epi32(src_line);
+    accum = _mm256_add_epi32(accum, abs);
+  }
+
+  {  // 32 bit horizontal add
+    const __m256i a = _mm256_srli_si256(accum, 8);
+    const __m256i b = _mm256_add_epi32(accum, a);
+    const __m256i c = _mm256_srli_epi64(b, 32);
+    const __m256i d = _mm256_add_epi32(b, c);
+    const __m128i accum_128 = _mm_add_epi32(_mm256_castsi256_si128(d),
+                                            _mm256_extractf128_si256(d, 1));
+    return _mm_cvtsi128_si32(accum_128);
+  }
+}
+
+int aom_satd_lp_avx2(const int16_t *coeff, int length) {
+  const __m256i one = _mm256_set1_epi16(1);
+  __m256i accum = _mm256_setzero_si256();
+
+  for (int i = 0; i < length; i += 16) {
+    const __m256i src_line = _mm256_loadu_si256((const __m256i *)coeff);
+    const __m256i abs = _mm256_abs_epi16(src_line);
+    const __m256i sum = _mm256_madd_epi16(abs, one);
+    accum = _mm256_add_epi32(accum, sum);
+    coeff += 16;
+  }
+
+  {  // 32 bit horizontal add
+    const __m256i a = _mm256_srli_si256(accum, 8);
+    const __m256i b = _mm256_add_epi32(accum, a);
+    const __m256i c = _mm256_srli_epi64(b, 32);
+    const __m256i d = _mm256_add_epi32(b, c);
+    const __m128i accum_128 = _mm_add_epi32(_mm256_castsi256_si128(d),
+                                            _mm256_extractf128_si256(d, 1));
+    return _mm_cvtsi128_si32(accum_128);
+  }
+}
+
+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;
+}
+
+void aom_avg_8x8_quad_avx2(const uint8_t *s, int p, int x16_idx, int y16_idx,
+                           int *avg) {
+  const uint8_t *s_y0 = s + y16_idx * p + x16_idx;
+  const uint8_t *s_y1 = s_y0 + 8 * p;
+  __m256i sum0, sum1, s0, s1, s2, s3, u0;
+  u0 = _mm256_setzero_si256();
+  s0 = _mm256_sad_epu8(xx_loadu2_mi128(s_y1, s_y0), u0);
+  s1 = _mm256_sad_epu8(xx_loadu2_mi128(s_y1 + p, s_y0 + p), u0);
+  s2 = _mm256_sad_epu8(xx_loadu2_mi128(s_y1 + 2 * p, s_y0 + 2 * p), u0);
+  s3 = _mm256_sad_epu8(xx_loadu2_mi128(s_y1 + 3 * p, s_y0 + 3 * p), u0);
+  sum0 = _mm256_add_epi16(s0, s1);
+  sum1 = _mm256_add_epi16(s2, s3);
+  s0 = _mm256_sad_epu8(xx_loadu2_mi128(s_y1 + 4 * p, s_y0 + 4 * p), u0);
+  s1 = _mm256_sad_epu8(xx_loadu2_mi128(s_y1 + 5 * p, s_y0 + 5 * p), u0);
+  s2 = _mm256_sad_epu8(xx_loadu2_mi128(s_y1 + 6 * p, s_y0 + 6 * p), u0);
+  s3 = _mm256_sad_epu8(xx_loadu2_mi128(s_y1 + 7 * p, s_y0 + 7 * p), u0);
+  sum0 = _mm256_add_epi16(sum0, _mm256_add_epi16(s0, s1));
+  sum1 = _mm256_add_epi16(sum1, _mm256_add_epi16(s2, s3));
+  sum0 = _mm256_add_epi16(sum0, sum1);
+
+  // (avg + 32) >> 6
+  __m256i rounding = _mm256_set1_epi32(32);
+  sum0 = _mm256_add_epi32(sum0, rounding);
+  sum0 = _mm256_srli_epi32(sum0, 6);
+  __m128i lo = _mm256_castsi256_si128(sum0);
+  __m128i hi = _mm256_extracti128_si256(sum0, 1);
+  avg[0] = _mm_cvtsi128_si32(lo);
+  avg[1] = _mm_extract_epi32(lo, 2);
+  avg[2] = _mm_cvtsi128_si32(hi);
+  avg[3] = _mm_extract_epi32(hi, 2);
+}
+
+void aom_int_pro_row_avx2(int16_t *hbuf, const uint8_t *ref,
+                          const int ref_stride, const int width,
+                          const int height, int norm_factor) {
+  // SIMD implementation assumes width and height to be multiple of 16 and 2
+  // respectively. For any odd width or height, SIMD support needs to be added.
+  assert(width % 16 == 0 && height % 2 == 0);
+
+  if (width % 32 == 0) {
+    const __m256i zero = _mm256_setzero_si256();
+    for (int wd = 0; wd < width; wd += 32) {
+      const uint8_t *ref_tmp = ref + wd;
+      int16_t *hbuf_tmp = hbuf + wd;
+      __m256i s0 = zero;
+      __m256i s1 = zero;
+      int idx = 0;
+      do {
+        __m256i src_line = _mm256_loadu_si256((const __m256i *)ref_tmp);
+        __m256i t0 = _mm256_unpacklo_epi8(src_line, zero);
+        __m256i t1 = _mm256_unpackhi_epi8(src_line, zero);
+        s0 = _mm256_add_epi16(s0, t0);
+        s1 = _mm256_add_epi16(s1, t1);
+        ref_tmp += ref_stride;
+
+        src_line = _mm256_loadu_si256((const __m256i *)ref_tmp);
+        t0 = _mm256_unpacklo_epi8(src_line, zero);
+        t1 = _mm256_unpackhi_epi8(src_line, zero);
+        s0 = _mm256_add_epi16(s0, t0);
+        s1 = _mm256_add_epi16(s1, t1);
+        ref_tmp += ref_stride;
+        idx += 2;
+      } while (idx < height);
+      s0 = _mm256_srai_epi16(s0, norm_factor);
+      s1 = _mm256_srai_epi16(s1, norm_factor);
+      _mm_storeu_si128((__m128i *)(hbuf_tmp), _mm256_castsi256_si128(s0));
+      _mm_storeu_si128((__m128i *)(hbuf_tmp + 8), _mm256_castsi256_si128(s1));
+      _mm_storeu_si128((__m128i *)(hbuf_tmp + 16),
+                       _mm256_extractf128_si256(s0, 1));
+      _mm_storeu_si128((__m128i *)(hbuf_tmp + 24),
+                       _mm256_extractf128_si256(s1, 1));
+    }
+  } else if (width % 16 == 0) {
+    aom_int_pro_row_sse2(hbuf, ref, ref_stride, width, height, norm_factor);
+  }
+}
+
+static INLINE void load_from_src_buf(const uint8_t *ref1, __m256i *src,
+                                     const int stride) {
+  src[0] = _mm256_loadu_si256((const __m256i *)ref1);
+  src[1] = _mm256_loadu_si256((const __m256i *)(ref1 + stride));
+  src[2] = _mm256_loadu_si256((const __m256i *)(ref1 + (2 * stride)));
+  src[3] = _mm256_loadu_si256((const __m256i *)(ref1 + (3 * stride)));
+}
+
+#define CALC_TOT_SAD_AND_STORE                                                \
+  /* r00 r10 x x r01 r11 x x | r02 r12 x x r03 r13 x x */                     \
+  const __m256i r01 = _mm256_add_epi16(_mm256_slli_si256(r1, 2), r0);         \
+  /* r00 r10 r20 x r01 r11 r21 x | r02 r12 r22 x r03 r13 r23 x */             \
+  const __m256i r012 = _mm256_add_epi16(_mm256_slli_si256(r2, 4), r01);       \
+  /* r00 r10 r20 r30 r01 r11 r21 r31 | r02 r12 r22 r32 r03 r13 r23 r33 */     \
+  const __m256i result0 = _mm256_add_epi16(_mm256_slli_si256(r3, 6), r012);   \
+                                                                              \
+  const __m128i results0 = _mm_add_epi16(                                     \
+      _mm256_castsi256_si128(result0), _mm256_extractf128_si256(result0, 1)); \
+  const __m128i results1 =                                                    \
+      _mm_add_epi16(results0, _mm_srli_si128(results0, 8));                   \
+  _mm_storel_epi64((__m128i *)vbuf, _mm_srli_epi16(results1, norm_factor));
+
+static INLINE void aom_int_pro_col_16wd_avx2(int16_t *vbuf, const uint8_t *ref,
+                                             const int ref_stride,
+                                             const int height,
+                                             int norm_factor) {
+  const __m256i zero = _mm256_setzero_si256();
+  int ht = 0;
+  // Post sad operation, the data is present in lower 16-bit of each 64-bit lane
+  // and higher 16-bits are Zero. Here, we are processing 8 rows at a time to
+  // utilize the higher 16-bits efficiently.
+  do {
+    __m256i src_00 =
+        _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(ref)));
+    src_00 = _mm256_inserti128_si256(
+        src_00, _mm_loadu_si128((const __m128i *)(ref + ref_stride * 4)), 1);
+    __m256i src_01 = _mm256_castsi128_si256(
+        _mm_loadu_si128((const __m128i *)(ref + ref_stride * 1)));
+    src_01 = _mm256_inserti128_si256(
+        src_01, _mm_loadu_si128((const __m128i *)(ref + ref_stride * 5)), 1);
+    __m256i src_10 = _mm256_castsi128_si256(
+        _mm_loadu_si128((const __m128i *)(ref + ref_stride * 2)));
+    src_10 = _mm256_inserti128_si256(
+        src_10, _mm_loadu_si128((const __m128i *)(ref + ref_stride * 6)), 1);
+    __m256i src_11 = _mm256_castsi128_si256(
+        _mm_loadu_si128((const __m128i *)(ref + ref_stride * 3)));
+    src_11 = _mm256_inserti128_si256(
+        src_11, _mm_loadu_si128((const __m128i *)(ref + ref_stride * 7)), 1);
+
+    // s00 x x x s01 x x x | s40 x x x s41 x x x
+    const __m256i s0 = _mm256_sad_epu8(src_00, zero);
+    // s10 x x x s11 x x x | s50 x x x s51 x x x
+    const __m256i s1 = _mm256_sad_epu8(src_01, zero);
+    // s20 x x x s21 x x x | s60 x x x s61 x x x
+    const __m256i s2 = _mm256_sad_epu8(src_10, zero);
+    // s30 x x x s31 x x x | s70 x x x s71 x x x
+    const __m256i s3 = _mm256_sad_epu8(src_11, zero);
+
+    // s00 s10 x x x x x x | s40 s50 x x x x x x
+    const __m256i s0_lo = _mm256_unpacklo_epi16(s0, s1);
+    // s01 s11 x x x x x x | s41 s51 x x x x x x
+    const __m256i s0_hi = _mm256_unpackhi_epi16(s0, s1);
+    // s20 s30 x x x x x x | s60 s70 x x x x x x
+    const __m256i s1_lo = _mm256_unpacklo_epi16(s2, s3);
+    // s21 s31 x x x x x x | s61 s71 x x x x x x
+    const __m256i s1_hi = _mm256_unpackhi_epi16(s2, s3);
+
+    // s0 s1 x x x x x x | s4 s5 x x x x x x
+    const __m256i s0_add = _mm256_add_epi16(s0_lo, s0_hi);
+    // s2 s3 x x x x x x | s6 s7 x x x x x x
+    const __m256i s1_add = _mm256_add_epi16(s1_lo, s1_hi);
+
+    // s1 s1 s2 s3 s4 s5 s6 s7
+    const __m128i results = _mm256_castsi256_si128(
+        _mm256_permute4x64_epi64(_mm256_unpacklo_epi32(s0_add, s1_add), 0x08));
+    _mm_storeu_si128((__m128i *)vbuf, _mm_srli_epi16(results, norm_factor));
+    vbuf += 8;
+    ref += (ref_stride << 3);
+    ht += 8;
+  } while (ht < height);
+}
+
+void aom_int_pro_col_avx2(int16_t *vbuf, const uint8_t *ref,
+                          const int ref_stride, const int width,
+                          const int height, int norm_factor) {
+  assert(width % 16 == 0);
+  if (width == 128) {
+    const __m256i zero = _mm256_setzero_si256();
+    for (int ht = 0; ht < height; ht += 4) {
+      __m256i src[16];
+      // Load source data.
+      load_from_src_buf(ref, &src[0], ref_stride);
+      load_from_src_buf(ref + 32, &src[4], ref_stride);
+      load_from_src_buf(ref + 64, &src[8], ref_stride);
+      load_from_src_buf(ref + 96, &src[12], ref_stride);
+
+      // Row0 output: r00 x x x r01 x x x | r02 x x x r03 x x x
+      const __m256i s0 = _mm256_add_epi16(_mm256_sad_epu8(src[0], zero),
+                                          _mm256_sad_epu8(src[4], zero));
+      const __m256i s1 = _mm256_add_epi16(_mm256_sad_epu8(src[8], zero),
+                                          _mm256_sad_epu8(src[12], zero));
+      const __m256i r0 = _mm256_add_epi16(s0, s1);
+      // Row1 output: r10 x x x r11 x x x | r12 x x x r13 x x x
+      const __m256i s2 = _mm256_add_epi16(_mm256_sad_epu8(src[1], zero),
+                                          _mm256_sad_epu8(src[5], zero));
+      const __m256i s3 = _mm256_add_epi16(_mm256_sad_epu8(src[9], zero),
+                                          _mm256_sad_epu8(src[13], zero));
+      const __m256i r1 = _mm256_add_epi16(s2, s3);
+      // Row2 output: r20 x x x r21 x x x | r22 x x x r23 x x x
+      const __m256i s4 = _mm256_add_epi16(_mm256_sad_epu8(src[2], zero),
+                                          _mm256_sad_epu8(src[6], zero));
+      const __m256i s5 = _mm256_add_epi16(_mm256_sad_epu8(src[10], zero),
+                                          _mm256_sad_epu8(src[14], zero));
+      const __m256i r2 = _mm256_add_epi16(s4, s5);
+      // Row3 output: r30 x x x r31 x x x | r32 x x x r33 x x x
+      const __m256i s6 = _mm256_add_epi16(_mm256_sad_epu8(src[3], zero),
+                                          _mm256_sad_epu8(src[7], zero));
+      const __m256i s7 = _mm256_add_epi16(_mm256_sad_epu8(src[11], zero),
+                                          _mm256_sad_epu8(src[15], zero));
+      const __m256i r3 = _mm256_add_epi16(s6, s7);
+
+      CALC_TOT_SAD_AND_STORE
+      vbuf += 4;
+      ref += ref_stride << 2;
+    }
+  } else if (width == 64) {
+    const __m256i zero = _mm256_setzero_si256();
+    for (int ht = 0; ht < height; ht += 4) {
+      __m256i src[8];
+      // Load source data.
+      load_from_src_buf(ref, &src[0], ref_stride);
+      load_from_src_buf(ref + 32, &src[4], ref_stride);
+
+      // Row0 output: r00 x x x r01 x x x | r02 x x x r03 x x x
+      const __m256i s0 = _mm256_sad_epu8(src[0], zero);
+      const __m256i s1 = _mm256_sad_epu8(src[4], zero);
+      const __m256i r0 = _mm256_add_epi16(s0, s1);
+      // Row1 output: r10 x x x r11 x x x | r12 x x x r13 x x x
+      const __m256i s2 = _mm256_sad_epu8(src[1], zero);
+      const __m256i s3 = _mm256_sad_epu8(src[5], zero);
+      const __m256i r1 = _mm256_add_epi16(s2, s3);
+      // Row2 output: r20 x x x r21 x x x | r22 x x x r23 x x x
+      const __m256i s4 = _mm256_sad_epu8(src[2], zero);
+      const __m256i s5 = _mm256_sad_epu8(src[6], zero);
+      const __m256i r2 = _mm256_add_epi16(s4, s5);
+      // Row3 output: r30 x x x r31 x x x | r32 x x x r33 x x x
+      const __m256i s6 = _mm256_sad_epu8(src[3], zero);
+      const __m256i s7 = _mm256_sad_epu8(src[7], zero);
+      const __m256i r3 = _mm256_add_epi16(s6, s7);
+
+      CALC_TOT_SAD_AND_STORE
+      vbuf += 4;
+      ref += ref_stride << 2;
+    }
+  } else if (width == 32) {
+    assert(height % 2 == 0);
+    const __m256i zero = _mm256_setzero_si256();
+    for (int ht = 0; ht < height; ht += 4) {
+      __m256i src[4];
+      // Load source data.
+      load_from_src_buf(ref, &src[0], ref_stride);
+
+      // s00 x x x s01 x x x s02 x x x s03 x x x
+      const __m256i r0 = _mm256_sad_epu8(src[0], zero);
+      // s10 x x x s11 x x x s12 x x x s13 x x x
+      const __m256i r1 = _mm256_sad_epu8(src[1], zero);
+      // s20 x x x s21 x x x s22 x x x s23 x x x
+      const __m256i r2 = _mm256_sad_epu8(src[2], zero);
+      // s30 x x x s31 x x x s32 x x x s33 x x x
+      const __m256i r3 = _mm256_sad_epu8(src[3], zero);
+
+      CALC_TOT_SAD_AND_STORE
+      vbuf += 4;
+      ref += ref_stride << 2;
+    }
+  } else if (width == 16) {
+    aom_int_pro_col_16wd_avx2(vbuf, ref, ref_stride, height, norm_factor);
+  }
+}
+
+static inline void calc_vector_mean_sse_64wd(const int16_t *ref,
+                                             const int16_t *src, __m256i *mean,
+                                             __m256i *sse) {
+  const __m256i src_line0 = _mm256_loadu_si256((const __m256i *)src);
+  const __m256i src_line1 = _mm256_loadu_si256((const __m256i *)(src + 16));
+  const __m256i src_line2 = _mm256_loadu_si256((const __m256i *)(src + 32));
+  const __m256i src_line3 = _mm256_loadu_si256((const __m256i *)(src + 48));
+  const __m256i ref_line0 = _mm256_loadu_si256((const __m256i *)ref);
+  const __m256i ref_line1 = _mm256_loadu_si256((const __m256i *)(ref + 16));
+  const __m256i ref_line2 = _mm256_loadu_si256((const __m256i *)(ref + 32));
+  const __m256i ref_line3 = _mm256_loadu_si256((const __m256i *)(ref + 48));
+
+  const __m256i diff0 = _mm256_sub_epi16(ref_line0, src_line0);
+  const __m256i diff1 = _mm256_sub_epi16(ref_line1, src_line1);
+  const __m256i diff2 = _mm256_sub_epi16(ref_line2, src_line2);
+  const __m256i diff3 = _mm256_sub_epi16(ref_line3, src_line3);
+  const __m256i diff_sqr0 = _mm256_madd_epi16(diff0, diff0);
+  const __m256i diff_sqr1 = _mm256_madd_epi16(diff1, diff1);
+  const __m256i diff_sqr2 = _mm256_madd_epi16(diff2, diff2);
+  const __m256i diff_sqr3 = _mm256_madd_epi16(diff3, diff3);
+
+  *mean = _mm256_add_epi16(*mean, _mm256_add_epi16(diff0, diff1));
+  *mean = _mm256_add_epi16(*mean, diff2);
+  *mean = _mm256_add_epi16(*mean, diff3);
+  *sse = _mm256_add_epi32(*sse, _mm256_add_epi32(diff_sqr0, diff_sqr1));
+  *sse = _mm256_add_epi32(*sse, diff_sqr2);
+  *sse = _mm256_add_epi32(*sse, diff_sqr3);
+}
+
+#define CALC_VAR_FROM_MEAN_SSE(mean, sse)                                    \
+  {                                                                          \
+    mean = _mm256_madd_epi16(mean, _mm256_set1_epi16(1));                    \
+    mean = _mm256_hadd_epi32(mean, sse);                                     \
+    mean = _mm256_add_epi32(mean, _mm256_bsrli_epi128(mean, 4));             \
+    const __m128i result = _mm_add_epi32(_mm256_castsi256_si128(mean),       \
+                                         _mm256_extractf128_si256(mean, 1)); \
+    /*(mean * mean): dynamic range 31 bits.*/                                \
+    const int mean_int = _mm_extract_epi32(result, 0);                       \
+    const int sse_int = _mm_extract_epi32(result, 2);                        \
+    const unsigned int mean_abs = abs(mean_int);                             \
+    var = sse_int - ((mean_abs * mean_abs) >> (bwl + 2));                    \
+  }
+
+// ref: [0 - 510]
+// src: [0 - 510]
+// bwl: {2, 3, 4, 5}
+int aom_vector_var_avx2(const int16_t *ref, const int16_t *src, int bwl) {
+  const int width = 4 << bwl;
+  assert(width % 16 == 0 && width <= 128);
+  int var = 0;
+
+  // Instead of having a loop over width 16, considered loop unrolling to avoid
+  // some addition operations.
+  if (width == 128) {
+    __m256i mean = _mm256_setzero_si256();
+    __m256i sse = _mm256_setzero_si256();
+
+    calc_vector_mean_sse_64wd(src, ref, &mean, &sse);
+    calc_vector_mean_sse_64wd(src + 64, ref + 64, &mean, &sse);
+    CALC_VAR_FROM_MEAN_SSE(mean, sse)
+  } else if (width == 64) {
+    __m256i mean = _mm256_setzero_si256();
+    __m256i sse = _mm256_setzero_si256();
+
+    calc_vector_mean_sse_64wd(src, ref, &mean, &sse);
+    CALC_VAR_FROM_MEAN_SSE(mean, sse)
+  } else if (width == 32) {
+    const __m256i src_line0 = _mm256_loadu_si256((const __m256i *)src);
+    const __m256i ref_line0 = _mm256_loadu_si256((const __m256i *)ref);
+    const __m256i src_line1 = _mm256_loadu_si256((const __m256i *)(src + 16));
+    const __m256i ref_line1 = _mm256_loadu_si256((const __m256i *)(ref + 16));
+
+    const __m256i diff0 = _mm256_sub_epi16(ref_line0, src_line0);
+    const __m256i diff1 = _mm256_sub_epi16(ref_line1, src_line1);
+    const __m256i diff_sqr0 = _mm256_madd_epi16(diff0, diff0);
+    const __m256i diff_sqr1 = _mm256_madd_epi16(diff1, diff1);
+    const __m256i sse = _mm256_add_epi32(diff_sqr0, diff_sqr1);
+    __m256i mean = _mm256_add_epi16(diff0, diff1);
+
+    CALC_VAR_FROM_MEAN_SSE(mean, sse)
+  } else if (width == 16) {
+    const __m256i src_line = _mm256_loadu_si256((const __m256i *)src);
+    const __m256i ref_line = _mm256_loadu_si256((const __m256i *)ref);
+    __m256i mean = _mm256_sub_epi16(ref_line, src_line);
+    const __m256i sse = _mm256_madd_epi16(mean, mean);
+
+    CALC_VAR_FROM_MEAN_SSE(mean, sse)
+  }
+  return var;
+}
diff --git a/third_party/aom/aom_dsp/x86/avg_intrin_sse2.c b/third_party/aom/aom_dsp/x86/avg_intrin_sse2.c
new file mode 100644
index 0000000000..9ab9143eee
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/avg_intrin_sse2.c
@@ -0,0 +1,700 @@
+/*
+ * Copyright (c) 2016, 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"
+#include "aom/aom_integer.h"
+#include "aom_dsp/x86/bitdepth_conversion_sse2.h"
+#include "aom_dsp/x86/mem_sse2.h"
+#include "aom_ports/mem.h"
+
+static INLINE void sign_extend_16bit_to_32bit_sse2(__m128i in, __m128i zero,
+                                                   __m128i *out_lo,
+                                                   __m128i *out_hi) {
+  const __m128i sign_bits = _mm_cmplt_epi16(in, zero);
+  *out_lo = _mm_unpacklo_epi16(in, sign_bits);
+  *out_hi = _mm_unpackhi_epi16(in, sign_bits);
+}
+
+static INLINE __m128i invert_sign_32_sse2(__m128i a, __m128i sign) {
+  a = _mm_xor_si128(a, sign);
+  return _mm_sub_epi32(a, sign);
+}
+
+void aom_minmax_8x8_sse2(const uint8_t *s, int p, const uint8_t *d, int dp,
+                         int *min, int *max) {
+  __m128i u0, s0, d0, diff, maxabsdiff, minabsdiff, negdiff, absdiff0, absdiff;
+  u0 = _mm_setzero_si128();
+  // Row 0
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s)), u0);
+  d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d)), u0);
+  diff = _mm_subs_epi16(s0, d0);
+  negdiff = _mm_subs_epi16(u0, diff);
+  absdiff0 = _mm_max_epi16(diff, negdiff);
+  // Row 1
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + p)), u0);
+  d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + dp)), u0);
+  diff = _mm_subs_epi16(s0, d0);
+  negdiff = _mm_subs_epi16(u0, diff);
+  absdiff = _mm_max_epi16(diff, negdiff);
+  maxabsdiff = _mm_max_epi16(absdiff0, absdiff);
+  minabsdiff = _mm_min_epi16(absdiff0, absdiff);
+  // Row 2
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 2 * p)), u0);
+  d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 2 * dp)), u0);
+  diff = _mm_subs_epi16(s0, d0);
+  negdiff = _mm_subs_epi16(u0, diff);
+  absdiff = _mm_max_epi16(diff, negdiff);
+  maxabsdiff = _mm_max_epi16(maxabsdiff, absdiff);
+  minabsdiff = _mm_min_epi16(minabsdiff, absdiff);
+  // Row 3
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 3 * p)), u0);
+  d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 3 * dp)), u0);
+  diff = _mm_subs_epi16(s0, d0);
+  negdiff = _mm_subs_epi16(u0, diff);
+  absdiff = _mm_max_epi16(diff, negdiff);
+  maxabsdiff = _mm_max_epi16(maxabsdiff, absdiff);
+  minabsdiff = _mm_min_epi16(minabsdiff, absdiff);
+  // Row 4
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 4 * p)), u0);
+  d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 4 * dp)), u0);
+  diff = _mm_subs_epi16(s0, d0);
+  negdiff = _mm_subs_epi16(u0, diff);
+  absdiff = _mm_max_epi16(diff, negdiff);
+  maxabsdiff = _mm_max_epi16(maxabsdiff, absdiff);
+  minabsdiff = _mm_min_epi16(minabsdiff, absdiff);
+  // Row 5
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 5 * p)), u0);
+  d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 5 * dp)), u0);
+  diff = _mm_subs_epi16(s0, d0);
+  negdiff = _mm_subs_epi16(u0, diff);
+  absdiff = _mm_max_epi16(diff, negdiff);
+  maxabsdiff = _mm_max_epi16(maxabsdiff, absdiff);
+  minabsdiff = _mm_min_epi16(minabsdiff, absdiff);
+  // Row 6
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 6 * p)), u0);
+  d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 6 * dp)), u0);
+  diff = _mm_subs_epi16(s0, d0);
+  negdiff = _mm_subs_epi16(u0, diff);
+  absdiff = _mm_max_epi16(diff, negdiff);
+  maxabsdiff = _mm_max_epi16(maxabsdiff, absdiff);
+  minabsdiff = _mm_min_epi16(minabsdiff, absdiff);
+  // Row 7
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 7 * p)), u0);
+  d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 7 * dp)), u0);
+  diff = _mm_subs_epi16(s0, d0);
+  negdiff = _mm_subs_epi16(u0, diff);
+  absdiff = _mm_max_epi16(diff, negdiff);
+  maxabsdiff = _mm_max_epi16(maxabsdiff, absdiff);
+  minabsdiff = _mm_min_epi16(minabsdiff, absdiff);
+
+  maxabsdiff = _mm_max_epi16(maxabsdiff, _mm_srli_si128(maxabsdiff, 8));
+  maxabsdiff = _mm_max_epi16(maxabsdiff, _mm_srli_epi64(maxabsdiff, 32));
+  maxabsdiff = _mm_max_epi16(maxabsdiff, _mm_srli_epi64(maxabsdiff, 16));
+  *max = _mm_extract_epi16(maxabsdiff, 0);
+
+  minabsdiff = _mm_min_epi16(minabsdiff, _mm_srli_si128(minabsdiff, 8));
+  minabsdiff = _mm_min_epi16(minabsdiff, _mm_srli_epi64(minabsdiff, 32));
+  minabsdiff = _mm_min_epi16(minabsdiff, _mm_srli_epi64(minabsdiff, 16));
+  *min = _mm_extract_epi16(minabsdiff, 0);
+}
+
+unsigned int aom_avg_8x8_sse2(const uint8_t *s, int p) {
+  __m128i sum0, sum1, s0, s1, s2, s3, u0;
+  unsigned int avg = 0;
+  u0 = _mm_setzero_si128();
+  s0 = loadh_epi64((const __m128i *)(s + p),
+                   _mm_loadl_epi64((const __m128i *)(s)));
+  s1 = loadh_epi64((const __m128i *)(s + 3 * p),
+                   _mm_loadl_epi64((const __m128i *)(s + 2 * p)));
+  s2 = loadh_epi64((const __m128i *)(s + 5 * p),
+                   _mm_loadl_epi64((const __m128i *)(s + 4 * p)));
+  s3 = loadh_epi64((const __m128i *)(s + 7 * p),
+                   _mm_loadl_epi64((const __m128i *)(s + 6 * p)));
+  s0 = _mm_sad_epu8(s0, u0);
+  s1 = _mm_sad_epu8(s1, u0);
+  s2 = _mm_sad_epu8(s2, u0);
+  s3 = _mm_sad_epu8(s3, u0);
+
+  sum0 = _mm_add_epi16(s0, s1);
+  sum1 = _mm_add_epi16(s2, s3);
+  sum0 = _mm_add_epi16(sum0, sum1);
+  sum0 = _mm_add_epi16(sum0, _mm_srli_si128(sum0, 8));
+  avg = _mm_cvtsi128_si32(sum0);
+  return (avg + 32) >> 6;
+}
+
+void calc_avg_8x8_dual_sse2(const uint8_t *s, int p, int *avg) {
+  __m128i sum0, sum1, s0, s1, s2, s3, u0;
+  u0 = _mm_setzero_si128();
+  s0 = _mm_sad_epu8(_mm_loadu_si128((const __m128i *)(s)), u0);
+  s1 = _mm_sad_epu8(_mm_loadu_si128((const __m128i *)(s + p)), u0);
+  s2 = _mm_sad_epu8(_mm_loadu_si128((const __m128i *)(s + 2 * p)), u0);
+  s3 = _mm_sad_epu8(_mm_loadu_si128((const __m128i *)(s + 3 * p)), u0);
+  sum0 = _mm_add_epi16(s0, s1);
+  sum1 = _mm_add_epi16(s2, s3);
+  s0 = _mm_sad_epu8(_mm_loadu_si128((const __m128i *)(s + 4 * p)), u0);
+  s1 = _mm_sad_epu8(_mm_loadu_si128((const __m128i *)(s + 5 * p)), u0);
+  s2 = _mm_sad_epu8(_mm_loadu_si128((const __m128i *)(s + 6 * p)), u0);
+  s3 = _mm_sad_epu8(_mm_loadu_si128((const __m128i *)(s + 7 * p)), u0);
+  sum0 = _mm_add_epi16(sum0, _mm_add_epi16(s0, s1));
+  sum1 = _mm_add_epi16(sum1, _mm_add_epi16(s2, s3));
+  sum0 = _mm_add_epi16(sum0, sum1);
+
+  // (avg + 32) >> 6
+  __m128i rounding = _mm_set1_epi32(32);
+  sum0 = _mm_add_epi32(sum0, rounding);
+  sum0 = _mm_srli_epi32(sum0, 6);
+  avg[0] = _mm_cvtsi128_si32(sum0);
+  avg[1] = _mm_extract_epi16(sum0, 4);
+}
+
+void aom_avg_8x8_quad_sse2(const uint8_t *s, int p, int x16_idx, int y16_idx,
+                           int *avg) {
+  const uint8_t *s_ptr = s + y16_idx * p + x16_idx;
+  for (int k = 0; k < 2; k++) {
+    calc_avg_8x8_dual_sse2(s_ptr, p, avg + k * 2);
+    s_ptr += 8 * p;
+  }
+}
+
+unsigned int aom_avg_4x4_sse2(const uint8_t *s, int p) {
+  __m128i s0, s1, u0;
+  unsigned int avg = 0;
+  u0 = _mm_setzero_si128();
+  s0 = _mm_unpacklo_epi32(_mm_cvtsi32_si128(*(const int *)(s)),
+                          _mm_cvtsi32_si128(*(const int *)(s + p)));
+  s1 = _mm_unpacklo_epi32(_mm_cvtsi32_si128(*(const int *)(s + p * 2)),
+                          _mm_cvtsi32_si128(*(const int *)(s + p * 3)));
+  s0 = _mm_sad_epu8(s0, u0);
+  s1 = _mm_sad_epu8(s1, u0);
+  s0 = _mm_add_epi16(s0, s1);
+  avg = _mm_cvtsi128_si32(s0);
+  return (avg + 8) >> 4;
+}
+
+static INLINE void hadamard_col4_sse2(__m128i *in, int iter) {
+  const __m128i a0 = in[0];
+  const __m128i a1 = in[1];
+  const __m128i a2 = in[2];
+  const __m128i a3 = in[3];
+  const __m128i b0 = _mm_srai_epi16(_mm_add_epi16(a0, a1), 1);
+  const __m128i b1 = _mm_srai_epi16(_mm_sub_epi16(a0, a1), 1);
+  const __m128i b2 = _mm_srai_epi16(_mm_add_epi16(a2, a3), 1);
+  const __m128i b3 = _mm_srai_epi16(_mm_sub_epi16(a2, a3), 1);
+  in[0] = _mm_add_epi16(b0, b2);
+  in[1] = _mm_add_epi16(b1, b3);
+  in[2] = _mm_sub_epi16(b0, b2);
+  in[3] = _mm_sub_epi16(b1, b3);
+
+  if (iter == 0) {
+    const __m128i ba = _mm_unpacklo_epi16(in[0], in[1]);
+    const __m128i dc = _mm_unpacklo_epi16(in[2], in[3]);
+    const __m128i dcba_lo = _mm_unpacklo_epi32(ba, dc);
+    const __m128i dcba_hi = _mm_unpackhi_epi32(ba, dc);
+    in[0] = dcba_lo;
+    in[1] = _mm_srli_si128(dcba_lo, 8);
+    in[2] = dcba_hi;
+    in[3] = _mm_srli_si128(dcba_hi, 8);
+  }
+}
+
+void aom_hadamard_4x4_sse2(const int16_t *src_diff, ptrdiff_t src_stride,
+                           tran_low_t *coeff) {
+  __m128i src[4];
+  src[0] = _mm_loadl_epi64((const __m128i *)src_diff);
+  src[1] = _mm_loadl_epi64((const __m128i *)(src_diff += src_stride));
+  src[2] = _mm_loadl_epi64((const __m128i *)(src_diff += src_stride));
+  src[3] = _mm_loadl_epi64((const __m128i *)(src_diff + src_stride));
+
+  hadamard_col4_sse2(src, 0);
+  hadamard_col4_sse2(src, 1);
+
+  store_tran_low(_mm_unpacklo_epi64(src[0], src[1]), coeff);
+  coeff += 8;
+  store_tran_low(_mm_unpacklo_epi64(src[2], src[3]), coeff);
+}
+
+static INLINE void hadamard_col8_sse2(__m128i *in, int iter) {
+  __m128i a0 = in[0];
+  __m128i a1 = in[1];
+  __m128i a2 = in[2];
+  __m128i a3 = in[3];
+  __m128i a4 = in[4];
+  __m128i a5 = in[5];
+  __m128i a6 = in[6];
+  __m128i a7 = in[7];
+
+  __m128i b0 = _mm_add_epi16(a0, a1);
+  __m128i b1 = _mm_sub_epi16(a0, a1);
+  __m128i b2 = _mm_add_epi16(a2, a3);
+  __m128i b3 = _mm_sub_epi16(a2, a3);
+  __m128i b4 = _mm_add_epi16(a4, a5);
+  __m128i b5 = _mm_sub_epi16(a4, a5);
+  __m128i b6 = _mm_add_epi16(a6, a7);
+  __m128i b7 = _mm_sub_epi16(a6, a7);
+
+  a0 = _mm_add_epi16(b0, b2);
+  a1 = _mm_add_epi16(b1, b3);
+  a2 = _mm_sub_epi16(b0, b2);
+  a3 = _mm_sub_epi16(b1, b3);
+  a4 = _mm_add_epi16(b4, b6);
+  a5 = _mm_add_epi16(b5, b7);
+  a6 = _mm_sub_epi16(b4, b6);
+  a7 = _mm_sub_epi16(b5, b7);
+
+  if (iter == 0) {
+    b0 = _mm_add_epi16(a0, a4);
+    b7 = _mm_add_epi16(a1, a5);
+    b3 = _mm_add_epi16(a2, a6);
+    b4 = _mm_add_epi16(a3, a7);
+    b2 = _mm_sub_epi16(a0, a4);
+    b6 = _mm_sub_epi16(a1, a5);
+    b1 = _mm_sub_epi16(a2, a6);
+    b5 = _mm_sub_epi16(a3, a7);
+
+    a0 = _mm_unpacklo_epi16(b0, b1);
+    a1 = _mm_unpacklo_epi16(b2, b3);
+    a2 = _mm_unpackhi_epi16(b0, b1);
+    a3 = _mm_unpackhi_epi16(b2, b3);
+    a4 = _mm_unpacklo_epi16(b4, b5);
+    a5 = _mm_unpacklo_epi16(b6, b7);
+    a6 = _mm_unpackhi_epi16(b4, b5);
+    a7 = _mm_unpackhi_epi16(b6, b7);
+
+    b0 = _mm_unpacklo_epi32(a0, a1);
+    b1 = _mm_unpacklo_epi32(a4, a5);
+    b2 = _mm_unpackhi_epi32(a0, a1);
+    b3 = _mm_unpackhi_epi32(a4, a5);
+    b4 = _mm_unpacklo_epi32(a2, a3);
+    b5 = _mm_unpacklo_epi32(a6, a7);
+    b6 = _mm_unpackhi_epi32(a2, a3);
+    b7 = _mm_unpackhi_epi32(a6, a7);
+
+    in[0] = _mm_unpacklo_epi64(b0, b1);
+    in[1] = _mm_unpackhi_epi64(b0, b1);
+    in[2] = _mm_unpacklo_epi64(b2, b3);
+    in[3] = _mm_unpackhi_epi64(b2, b3);
+    in[4] = _mm_unpacklo_epi64(b4, b5);
+    in[5] = _mm_unpackhi_epi64(b4, b5);
+    in[6] = _mm_unpacklo_epi64(b6, b7);
+    in[7] = _mm_unpackhi_epi64(b6, b7);
+  } else {
+    in[0] = _mm_add_epi16(a0, a4);
+    in[7] = _mm_add_epi16(a1, a5);
+    in[3] = _mm_add_epi16(a2, a6);
+    in[4] = _mm_add_epi16(a3, a7);
+    in[2] = _mm_sub_epi16(a0, a4);
+    in[6] = _mm_sub_epi16(a1, a5);
+    in[1] = _mm_sub_epi16(a2, a6);
+    in[5] = _mm_sub_epi16(a3, a7);
+  }
+}
+
+static INLINE void hadamard_8x8_sse2(const int16_t *src_diff,
+                                     ptrdiff_t src_stride, tran_low_t *coeff,
+                                     int is_final) {
+  __m128i src[8];
+  src[0] = _mm_load_si128((const __m128i *)src_diff);
+  src[1] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
+  src[2] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
+  src[3] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
+  src[4] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
+  src[5] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
+  src[6] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
+  src[7] = _mm_load_si128((const __m128i *)(src_diff + src_stride));
+
+  hadamard_col8_sse2(src, 0);
+  hadamard_col8_sse2(src, 1);
+
+  if (is_final) {
+    store_tran_low(src[0], coeff);
+    coeff += 8;
+    store_tran_low(src[1], coeff);
+    coeff += 8;
+    store_tran_low(src[2], coeff);
+    coeff += 8;
+    store_tran_low(src[3], coeff);
+    coeff += 8;
+    store_tran_low(src[4], coeff);
+    coeff += 8;
+    store_tran_low(src[5], coeff);
+    coeff += 8;
+    store_tran_low(src[6], coeff);
+    coeff += 8;
+    store_tran_low(src[7], coeff);
+  } else {
+    int16_t *coeff16 = (int16_t *)coeff;
+    _mm_store_si128((__m128i *)coeff16, src[0]);
+    coeff16 += 8;
+    _mm_store_si128((__m128i *)coeff16, src[1]);
+    coeff16 += 8;
+    _mm_store_si128((__m128i *)coeff16, src[2]);
+    coeff16 += 8;
+    _mm_store_si128((__m128i *)coeff16, src[3]);
+    coeff16 += 8;
+    _mm_store_si128((__m128i *)coeff16, src[4]);
+    coeff16 += 8;
+    _mm_store_si128((__m128i *)coeff16, src[5]);
+    coeff16 += 8;
+    _mm_store_si128((__m128i *)coeff16, src[6]);
+    coeff16 += 8;
+    _mm_store_si128((__m128i *)coeff16, src[7]);
+  }
+}
+
+void aom_hadamard_8x8_sse2(const int16_t *src_diff, ptrdiff_t src_stride,
+                           tran_low_t *coeff) {
+  hadamard_8x8_sse2(src_diff, src_stride, coeff, 1);
+}
+
+static INLINE void hadamard_lp_8x8_sse2(const int16_t *src_diff,
+                                        ptrdiff_t src_stride, int16_t *coeff) {
+  __m128i src[8];
+  src[0] = _mm_load_si128((const __m128i *)src_diff);
+  src[1] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
+  src[2] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
+  src[3] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
+  src[4] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
+  src[5] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
+  src[6] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
+  src[7] = _mm_load_si128((const __m128i *)(src_diff + src_stride));
+
+  hadamard_col8_sse2(src, 0);
+  hadamard_col8_sse2(src, 1);
+
+  _mm_store_si128((__m128i *)coeff, src[0]);
+  coeff += 8;
+  _mm_store_si128((__m128i *)coeff, src[1]);
+  coeff += 8;
+  _mm_store_si128((__m128i *)coeff, src[2]);
+  coeff += 8;
+  _mm_store_si128((__m128i *)coeff, src[3]);
+  coeff += 8;
+  _mm_store_si128((__m128i *)coeff, src[4]);
+  coeff += 8;
+  _mm_store_si128((__m128i *)coeff, src[5]);
+  coeff += 8;
+  _mm_store_si128((__m128i *)coeff, src[6]);
+  coeff += 8;
+  _mm_store_si128((__m128i *)coeff, src[7]);
+}
+
+void aom_hadamard_lp_8x8_sse2(const int16_t *src_diff, ptrdiff_t src_stride,
+                              int16_t *coeff) {
+  hadamard_lp_8x8_sse2(src_diff, src_stride, coeff);
+}
+
+void aom_hadamard_lp_8x8_dual_sse2(const int16_t *src_diff,
+                                   ptrdiff_t src_stride, int16_t *coeff) {
+  for (int i = 0; i < 2; i++) {
+    hadamard_lp_8x8_sse2(src_diff + (i * 8), src_stride, coeff + (i * 64));
+  }
+}
+
+void aom_hadamard_lp_16x16_sse2(const int16_t *src_diff, ptrdiff_t src_stride,
+                                int16_t *coeff) {
+  for (int idx = 0; idx < 4; ++idx) {
+    const int16_t *src_ptr =
+        src_diff + (idx >> 1) * 8 * src_stride + (idx & 0x01) * 8;
+    hadamard_lp_8x8_sse2(src_ptr, src_stride, coeff + idx * 64);
+  }
+
+  int16_t *t_coeff = coeff;
+  for (int idx = 0; idx < 64; idx += 8) {
+    __m128i coeff0 = _mm_load_si128((const __m128i *)t_coeff);
+    __m128i coeff1 = _mm_load_si128((const __m128i *)(t_coeff + 64));
+    __m128i coeff2 = _mm_load_si128((const __m128i *)(t_coeff + 128));
+    __m128i coeff3 = _mm_load_si128((const __m128i *)(t_coeff + 192));
+
+    __m128i b0 = _mm_add_epi16(coeff0, coeff1);
+    __m128i b1 = _mm_sub_epi16(coeff0, coeff1);
+    __m128i b2 = _mm_add_epi16(coeff2, coeff3);
+    __m128i b3 = _mm_sub_epi16(coeff2, coeff3);
+
+    b0 = _mm_srai_epi16(b0, 1);
+    b1 = _mm_srai_epi16(b1, 1);
+    b2 = _mm_srai_epi16(b2, 1);
+    b3 = _mm_srai_epi16(b3, 1);
+
+    coeff0 = _mm_add_epi16(b0, b2);
+    coeff1 = _mm_add_epi16(b1, b3);
+    coeff2 = _mm_sub_epi16(b0, b2);
+    coeff3 = _mm_sub_epi16(b1, b3);
+
+    _mm_store_si128((__m128i *)t_coeff, coeff0);
+    _mm_store_si128((__m128i *)(t_coeff + 64), coeff1);
+    _mm_store_si128((__m128i *)(t_coeff + 128), coeff2);
+    _mm_store_si128((__m128i *)(t_coeff + 192), coeff3);
+
+    t_coeff += 8;
+  }
+}
+
+static INLINE void hadamard_16x16_sse2(const int16_t *src_diff,
+                                       ptrdiff_t src_stride, tran_low_t *coeff,
+                                       int is_final) {
+  // For high bitdepths, it is unnecessary to store_tran_low
+  // (mult/unpack/store), then load_tran_low (load/pack) the same memory in the
+  // next stage.  Output to an intermediate buffer first, then store_tran_low()
+  // in the final stage.
+  DECLARE_ALIGNED(32, int16_t, temp_coeff[16 * 16]);
+  int16_t *t_coeff = temp_coeff;
+  int16_t *coeff16 = (int16_t *)coeff;
+  int idx;
+  for (idx = 0; idx < 4; ++idx) {
+    const int16_t *src_ptr =
+        src_diff + (idx >> 1) * 8 * src_stride + (idx & 0x01) * 8;
+    hadamard_8x8_sse2(src_ptr, src_stride, (tran_low_t *)(t_coeff + idx * 64),
+                      0);
+  }
+
+  for (idx = 0; idx < 64; idx += 8) {
+    __m128i coeff0 = _mm_load_si128((const __m128i *)t_coeff);
+    __m128i coeff1 = _mm_load_si128((const __m128i *)(t_coeff + 64));
+    __m128i coeff2 = _mm_load_si128((const __m128i *)(t_coeff + 128));
+    __m128i coeff3 = _mm_load_si128((const __m128i *)(t_coeff + 192));
+
+    __m128i b0 = _mm_add_epi16(coeff0, coeff1);
+    __m128i b1 = _mm_sub_epi16(coeff0, coeff1);
+    __m128i b2 = _mm_add_epi16(coeff2, coeff3);
+    __m128i b3 = _mm_sub_epi16(coeff2, coeff3);
+
+    b0 = _mm_srai_epi16(b0, 1);
+    b1 = _mm_srai_epi16(b1, 1);
+    b2 = _mm_srai_epi16(b2, 1);
+    b3 = _mm_srai_epi16(b3, 1);
+
+    coeff0 = _mm_add_epi16(b0, b2);
+    coeff1 = _mm_add_epi16(b1, b3);
+    coeff2 = _mm_sub_epi16(b0, b2);
+    coeff3 = _mm_sub_epi16(b1, b3);
+
+    if (is_final) {
+      store_tran_low_offset_4(coeff0, coeff);
+      store_tran_low_offset_4(coeff1, coeff + 64);
+      store_tran_low_offset_4(coeff2, coeff + 128);
+      store_tran_low_offset_4(coeff3, coeff + 192);
+      coeff += 4;
+    } else {
+      _mm_store_si128((__m128i *)coeff16, coeff0);
+      _mm_store_si128((__m128i *)(coeff16 + 64), coeff1);
+      _mm_store_si128((__m128i *)(coeff16 + 128), coeff2);
+      _mm_store_si128((__m128i *)(coeff16 + 192), coeff3);
+      coeff16 += 8;
+    }
+
+    t_coeff += 8;
+    // Increment the pointer additionally by 0 and 8 in alternate
+    // iterations(instead of 8) to ensure the coherency with the implementation
+    // of store_tran_low_offset_4()
+    coeff += (((idx >> 3) & 1) << 3);
+  }
+}
+
+void aom_hadamard_16x16_sse2(const int16_t *src_diff, ptrdiff_t src_stride,
+                             tran_low_t *coeff) {
+  hadamard_16x16_sse2(src_diff, src_stride, coeff, 1);
+}
+
+void aom_hadamard_32x32_sse2(const int16_t *src_diff, ptrdiff_t src_stride,
+                             tran_low_t *coeff) {
+  // For high bitdepths, it is unnecessary to store_tran_low
+  // (mult/unpack/store), then load_tran_low (load/pack) the same memory in the
+  // next stage.  Output to an intermediate buffer first, then store_tran_low()
+  // in the final stage.
+  DECLARE_ALIGNED(32, int16_t, temp_coeff[32 * 32]);
+  int16_t *t_coeff = temp_coeff;
+  int idx;
+  __m128i coeff0_lo, coeff1_lo, coeff2_lo, coeff3_lo, b0_lo, b1_lo, b2_lo,
+      b3_lo;
+  __m128i coeff0_hi, coeff1_hi, coeff2_hi, coeff3_hi, b0_hi, b1_hi, b2_hi,
+      b3_hi;
+  __m128i b0, b1, b2, b3;
+  const __m128i zero = _mm_setzero_si128();
+  for (idx = 0; idx < 4; ++idx) {
+    const int16_t *src_ptr =
+        src_diff + (idx >> 1) * 16 * src_stride + (idx & 0x01) * 16;
+    hadamard_16x16_sse2(src_ptr, src_stride,
+                        (tran_low_t *)(t_coeff + idx * 256), 0);
+  }
+
+  for (idx = 0; idx < 256; idx += 8) {
+    __m128i coeff0 = _mm_load_si128((const __m128i *)t_coeff);
+    __m128i coeff1 = _mm_load_si128((const __m128i *)(t_coeff + 256));
+    __m128i coeff2 = _mm_load_si128((const __m128i *)(t_coeff + 512));
+    __m128i coeff3 = _mm_load_si128((const __m128i *)(t_coeff + 768));
+
+    // Sign extend 16 bit to 32 bit.
+    sign_extend_16bit_to_32bit_sse2(coeff0, zero, &coeff0_lo, &coeff0_hi);
+    sign_extend_16bit_to_32bit_sse2(coeff1, zero, &coeff1_lo, &coeff1_hi);
+    sign_extend_16bit_to_32bit_sse2(coeff2, zero, &coeff2_lo, &coeff2_hi);
+    sign_extend_16bit_to_32bit_sse2(coeff3, zero, &coeff3_lo, &coeff3_hi);
+
+    b0_lo = _mm_add_epi32(coeff0_lo, coeff1_lo);
+    b0_hi = _mm_add_epi32(coeff0_hi, coeff1_hi);
+
+    b1_lo = _mm_sub_epi32(coeff0_lo, coeff1_lo);
+    b1_hi = _mm_sub_epi32(coeff0_hi, coeff1_hi);
+
+    b2_lo = _mm_add_epi32(coeff2_lo, coeff3_lo);
+    b2_hi = _mm_add_epi32(coeff2_hi, coeff3_hi);
+
+    b3_lo = _mm_sub_epi32(coeff2_lo, coeff3_lo);
+    b3_hi = _mm_sub_epi32(coeff2_hi, coeff3_hi);
+
+    b0_lo = _mm_srai_epi32(b0_lo, 2);
+    b1_lo = _mm_srai_epi32(b1_lo, 2);
+    b2_lo = _mm_srai_epi32(b2_lo, 2);
+    b3_lo = _mm_srai_epi32(b3_lo, 2);
+
+    b0_hi = _mm_srai_epi32(b0_hi, 2);
+    b1_hi = _mm_srai_epi32(b1_hi, 2);
+    b2_hi = _mm_srai_epi32(b2_hi, 2);
+    b3_hi = _mm_srai_epi32(b3_hi, 2);
+
+    b0 = _mm_packs_epi32(b0_lo, b0_hi);
+    b1 = _mm_packs_epi32(b1_lo, b1_hi);
+    b2 = _mm_packs_epi32(b2_lo, b2_hi);
+    b3 = _mm_packs_epi32(b3_lo, b3_hi);
+
+    coeff0 = _mm_add_epi16(b0, b2);
+    coeff1 = _mm_add_epi16(b1, b3);
+    store_tran_low_offset_4(coeff0, coeff);
+    store_tran_low_offset_4(coeff1, coeff + 256);
+
+    coeff2 = _mm_sub_epi16(b0, b2);
+    coeff3 = _mm_sub_epi16(b1, b3);
+    store_tran_low_offset_4(coeff2, coeff + 512);
+    store_tran_low_offset_4(coeff3, coeff + 768);
+
+    // Increment the pointer by 4 and 12 in alternate iterations(instead of 8)
+    // to ensure the coherency with the implementation of
+    // store_tran_low_offset_4()
+    coeff += (4 + (((idx >> 3) & 1) << 3));
+    t_coeff += 8;
+  }
+}
+
+int aom_satd_sse2(const tran_low_t *coeff, int length) {
+  int i;
+  const __m128i zero = _mm_setzero_si128();
+  __m128i accum = zero;
+
+  for (i = 0; i < length; i += 4) {
+    const __m128i src_line = _mm_load_si128((const __m128i *)coeff);
+    const __m128i coeff_sign = _mm_srai_epi32(src_line, 31);
+    const __m128i abs_coeff = invert_sign_32_sse2(src_line, coeff_sign);
+    accum = _mm_add_epi32(accum, abs_coeff);
+    coeff += 4;
+  }
+
+  {  // cascading summation of accum
+    __m128i hi = _mm_srli_si128(accum, 8);
+    accum = _mm_add_epi32(accum, hi);
+    hi = _mm_srli_epi64(accum, 32);
+    accum = _mm_add_epi32(accum, hi);
+  }
+
+  return _mm_cvtsi128_si32(accum);
+}
+
+int aom_satd_lp_sse2(const int16_t *coeff, int length) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i one = _mm_set1_epi16(1);
+  __m128i accum = zero;
+
+  for (int i = 0; i < length; i += 16) {
+    const __m128i src_line0 = _mm_loadu_si128((const __m128i *)coeff);
+    const __m128i src_line1 = _mm_loadu_si128((const __m128i *)(coeff + 8));
+    const __m128i inv0 = _mm_sub_epi16(zero, src_line0);
+    const __m128i inv1 = _mm_sub_epi16(zero, src_line1);
+    const __m128i abs0 = _mm_max_epi16(src_line0, inv0);  // abs(src_line)
+    const __m128i abs1 = _mm_max_epi16(src_line1, inv1);  // abs(src_line)
+    const __m128i sum0 = _mm_madd_epi16(abs0, one);
+    const __m128i sum1 = _mm_madd_epi16(abs1, one);
+    accum = _mm_add_epi32(accum, sum0);
+    accum = _mm_add_epi32(accum, sum1);
+    coeff += 16;
+  }
+
+  {  // cascading summation of accum
+    __m128i hi = _mm_srli_si128(accum, 8);
+    accum = _mm_add_epi32(accum, hi);
+    hi = _mm_srli_epi64(accum, 32);
+    accum = _mm_add_epi32(accum, hi);
+  }
+
+  return _mm_cvtsi128_si32(accum);
+}
+
+void aom_int_pro_row_sse2(int16_t *hbuf, const uint8_t *ref,
+                          const int ref_stride, const int width,
+                          const int height, int norm_factor) {
+  // SIMD implementation assumes width and height to be multiple of 16 and 2
+  // respectively. For any odd width or height, SIMD support needs to be added.
+  assert(width % 16 == 0 && height % 2 == 0);
+  __m128i zero = _mm_setzero_si128();
+
+  for (int wd = 0; wd < width; wd += 16) {
+    const uint8_t *ref_tmp = ref + wd;
+    int16_t *hbuf_tmp = hbuf + wd;
+    __m128i s0 = zero;
+    __m128i s1 = zero;
+    int idx = 0;
+    do {
+      __m128i src_line = _mm_loadu_si128((const __m128i *)ref_tmp);
+      __m128i t0 = _mm_unpacklo_epi8(src_line, zero);
+      __m128i t1 = _mm_unpackhi_epi8(src_line, zero);
+      s0 = _mm_add_epi16(s0, t0);
+      s1 = _mm_add_epi16(s1, t1);
+      ref_tmp += ref_stride;
+
+      src_line = _mm_loadu_si128((const __m128i *)ref_tmp);
+      t0 = _mm_unpacklo_epi8(src_line, zero);
+      t1 = _mm_unpackhi_epi8(src_line, zero);
+      s0 = _mm_add_epi16(s0, t0);
+      s1 = _mm_add_epi16(s1, t1);
+      ref_tmp += ref_stride;
+      idx += 2;
+    } while (idx < height);
+
+    s0 = _mm_srai_epi16(s0, norm_factor);
+    s1 = _mm_srai_epi16(s1, norm_factor);
+    _mm_storeu_si128((__m128i *)(hbuf_tmp), s0);
+    _mm_storeu_si128((__m128i *)(hbuf_tmp + 8), s1);
+  }
+}
+
+void aom_int_pro_col_sse2(int16_t *vbuf, const uint8_t *ref,
+                          const int ref_stride, const int width,
+                          const int height, int norm_factor) {
+  // SIMD implementation assumes width to be multiple of 16.
+  assert(width % 16 == 0);
+
+  for (int ht = 0; ht < height; ht++) {
+    const uint8_t *ref_tmp = ref + (ht * ref_stride);
+    __m128i zero = _mm_setzero_si128();
+    __m128i s0 = zero;
+    __m128i s1, src_line;
+    for (int i = 0; i < width; i += 16) {
+      src_line = _mm_loadu_si128((const __m128i *)ref_tmp);
+      s1 = _mm_sad_epu8(src_line, zero);
+      s0 = _mm_add_epi16(s0, s1);
+      ref_tmp += 16;
+    }
+
+    s1 = _mm_srli_si128(s0, 8);
+    s0 = _mm_add_epi16(s0, s1);
+    vbuf[ht] = _mm_cvtsi128_si32(s0) >> norm_factor;
+  }
+}
diff --git a/third_party/aom/aom_dsp/x86/avg_intrin_sse4.c b/third_party/aom/aom_dsp/x86/avg_intrin_sse4.c
new file mode 100644
index 0000000000..b83b43122a
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/avg_intrin_sse4.c
@@ -0,0 +1,59 @@
+/*
+ * Copyright (c) 2022, 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 <smmintrin.h>
+
+#include "config/aom_dsp_rtcd.h"
+
+// ref: [0 - 510]
+// src: [0 - 510]
+// bwl: {2, 3, 4, 5}
+int aom_vector_var_sse4_1(const int16_t *ref, const int16_t *src, int bwl) {
+  const int width = 4 << bwl;
+  assert(width % 16 == 0);
+
+  const __m128i k_one_epi16 = _mm_set1_epi16((int16_t)1);
+  __m128i mean = _mm_setzero_si128();
+  __m128i sse = _mm_setzero_si128();
+
+  for (int i = 0; i < width; i += 16) {
+    const __m128i src_line = _mm_loadu_si128((const __m128i *)src);
+    const __m128i ref_line = _mm_loadu_si128((const __m128i *)ref);
+    const __m128i src_line2 = _mm_loadu_si128((const __m128i *)(src + 8));
+    const __m128i ref_line2 = _mm_loadu_si128((const __m128i *)(ref + 8));
+    __m128i diff = _mm_sub_epi16(ref_line, src_line);
+    const __m128i diff2 = _mm_sub_epi16(ref_line2, src_line2);
+    __m128i diff_sqr = _mm_madd_epi16(diff, diff);
+    const __m128i diff_sqr2 = _mm_madd_epi16(diff2, diff2);
+
+    diff = _mm_add_epi16(diff, diff2);
+    diff_sqr = _mm_add_epi32(diff_sqr, diff_sqr2);
+    sse = _mm_add_epi32(sse, diff_sqr);
+    mean = _mm_add_epi16(mean, diff);
+
+    src += 16;
+    ref += 16;
+  }
+
+  // m0 m1 m2 m3
+  mean = _mm_madd_epi16(mean, k_one_epi16);
+  // m0+m1 m2+m3 s0+s1 s2+s3
+  __m128i result = _mm_hadd_epi32(mean, sse);
+  // m0+m1+m2+m3 s0+s1+s2+s3 x x
+  result = _mm_add_epi32(result, _mm_bsrli_si128(result, 4));
+
+  // (mean * mean): dynamic range 31 bits.
+  const int mean_int = _mm_extract_epi32(result, 0);
+  const int sse_int = _mm_extract_epi32(result, 2);
+  const unsigned int mean_abs = abs(mean_int);
+  const int var = sse_int - ((mean_abs * mean_abs) >> (bwl + 2));
+  return var;
+}
diff --git a/third_party/aom/aom_dsp/x86/bitdepth_conversion_avx2.h b/third_party/aom/aom_dsp/x86/bitdepth_conversion_avx2.h
new file mode 100644
index 0000000000..85896e2768
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/bitdepth_conversion_avx2.h
@@ -0,0 +1,32 @@
+/*
+ * Copyright (c) 2016, 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_config.h"
+#include "aom/aom_integer.h"
+#include "aom_dsp/aom_dsp_common.h"
+
+static INLINE __m256i load_tran_low(const tran_low_t *a) {
+  const __m256i a_low = _mm256_loadu_si256((const __m256i *)a);
+  const __m256i a_high = _mm256_loadu_si256((const __m256i *)(a + 8));
+  return _mm256_packs_epi32(a_low, a_high);
+}
+
+static INLINE void store_tran_low(__m256i a, tran_low_t *b) {
+  const __m256i one = _mm256_set1_epi16(1);
+  const __m256i a_hi = _mm256_mulhi_epi16(a, one);
+  const __m256i a_lo = _mm256_mullo_epi16(a, one);
+  const __m256i a_1 = _mm256_unpacklo_epi16(a_lo, a_hi);
+  const __m256i a_2 = _mm256_unpackhi_epi16(a_lo, a_hi);
+  _mm256_storeu_si256((__m256i *)b, a_1);
+  _mm256_storeu_si256((__m256i *)(b + 8), a_2);
+}
diff --git a/third_party/aom/aom_dsp/x86/bitdepth_conversion_sse2.h b/third_party/aom/aom_dsp/x86/bitdepth_conversion_sse2.h
new file mode 100644
index 0000000000..ff77760b6f
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/bitdepth_conversion_sse2.h
@@ -0,0 +1,49 @@
+/*
+ * Copyright (c) 2016, 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 <xmmintrin.h>
+
+#include "config/aom_config.h"
+#include "aom/aom_integer.h"
+#include "aom_dsp/aom_dsp_common.h"
+
+// Load 8 16 bit values. If the source is 32 bits then pack down with
+// saturation.
+static INLINE __m128i load_tran_low(const tran_low_t *a) {
+  const __m128i a_low = _mm_load_si128((const __m128i *)a);
+  return _mm_packs_epi32(a_low, *(const __m128i *)(a + 4));
+}
+
+static INLINE void unpack_trans(__m128i a, __m128i *a_1, __m128i *a_2) {
+  const __m128i one = _mm_set1_epi16(1);
+  const __m128i a_hi = _mm_mulhi_epi16(a, one);
+  const __m128i a_lo = _mm_mullo_epi16(a, one);
+  *a_1 = _mm_unpacklo_epi16(a_lo, a_hi);
+  *a_2 = _mm_unpackhi_epi16(a_lo, a_hi);
+}
+
+// Store 8 16 bit values. If the destination is 32 bits then sign extend the
+// values by multiplying by 1.
+static INLINE void store_tran_low(__m128i a, tran_low_t *b) {
+  __m128i a_1, a_2;
+  unpack_trans(a, &a_1, &a_2);
+  _mm_store_si128((__m128i *)(b), a_1);
+  _mm_store_si128((__m128i *)(b + 4), a_2);
+}
+// Stores the second result at an offset of 8 (instead of 4) to match the output
+// with that of AVX2 implementation and the function is similar to
+// store_tran_low().
+static INLINE void store_tran_low_offset_4(__m128i a, tran_low_t *b) {
+  __m128i a_1, a_2;
+  unpack_trans(a, &a_1, &a_2);
+  _mm_store_si128((__m128i *)(b), a_1);
+  _mm_store_si128((__m128i *)(b + 8), a_2);
+}
diff --git a/third_party/aom/aom_dsp/x86/blend_a64_hmask_sse4.c b/third_party/aom/aom_dsp/x86/blend_a64_hmask_sse4.c
new file mode 100644
index 0000000000..e0289abe12
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/blend_a64_hmask_sse4.c
@@ -0,0 +1,36 @@
+/*
+ * Copyright (c) 2016, 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 "aom/aom_integer.h"
+
+#include "config/aom_dsp_rtcd.h"
+
+// To start out, just dispatch to the function using the 2D mask and
+// pass mask stride as 0. This can be improved upon if necessary.
+
+void aom_blend_a64_hmask_sse4_1(uint8_t *dst, uint32_t dst_stride,
+                                const uint8_t *src0, uint32_t src0_stride,
+                                const uint8_t *src1, uint32_t src1_stride,
+                                const uint8_t *mask, int w, int h) {
+  aom_blend_a64_mask_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                            src1_stride, mask, 0, w, h, 0, 0);
+}
+
+#if CONFIG_AV1_HIGHBITDEPTH
+void aom_highbd_blend_a64_hmask_sse4_1(
+    uint8_t *dst_8, uint32_t dst_stride, const uint8_t *src0_8,
+    uint32_t src0_stride, const uint8_t *src1_8, uint32_t src1_stride,
+    const uint8_t *mask, int w, int h, int bd) {
+  aom_highbd_blend_a64_mask_sse4_1(dst_8, dst_stride, src0_8, src0_stride,
+                                   src1_8, src1_stride, mask, 0, w, h, 0, 0,
+                                   bd);
+}
+#endif
diff --git a/third_party/aom/aom_dsp/x86/blend_a64_mask_avx2.c b/third_party/aom/aom_dsp/x86/blend_a64_mask_avx2.c
new file mode 100644
index 0000000000..dfbab324d0
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/blend_a64_mask_avx2.c
@@ -0,0 +1,1374 @@
+/*
+ * 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 <smmintrin.h>  // SSE4.1
+#include <immintrin.h>  // AVX2
+
+#include <assert.h>
+
+#include "aom/aom_integer.h"
+#include "aom_ports/mem.h"
+#include "aom_dsp/aom_dsp_common.h"
+
+#include "aom_dsp/x86/synonyms.h"
+#include "aom_dsp/x86/synonyms_avx2.h"
+#include "aom_dsp/x86/blend_sse4.h"
+#include "aom_dsp/x86/blend_mask_sse4.h"
+
+#include "config/aom_dsp_rtcd.h"
+
+static INLINE void blend_a64_d16_mask_w16_avx2(
+    uint8_t *dst, const CONV_BUF_TYPE *src0, const CONV_BUF_TYPE *src1,
+    const __m256i *m0, const __m256i *v_round_offset, const __m256i *v_maxval,
+    int shift) {
+  const __m256i max_minus_m0 = _mm256_sub_epi16(*v_maxval, *m0);
+  const __m256i s0_0 = yy_loadu_256(src0);
+  const __m256i s1_0 = yy_loadu_256(src1);
+  __m256i res0_lo = _mm256_madd_epi16(_mm256_unpacklo_epi16(s0_0, s1_0),
+                                      _mm256_unpacklo_epi16(*m0, max_minus_m0));
+  __m256i res0_hi = _mm256_madd_epi16(_mm256_unpackhi_epi16(s0_0, s1_0),
+                                      _mm256_unpackhi_epi16(*m0, max_minus_m0));
+  res0_lo =
+      _mm256_srai_epi32(_mm256_sub_epi32(res0_lo, *v_round_offset), shift);
+  res0_hi =
+      _mm256_srai_epi32(_mm256_sub_epi32(res0_hi, *v_round_offset), shift);
+  const __m256i res0 = _mm256_packs_epi32(res0_lo, res0_hi);
+  __m256i res = _mm256_packus_epi16(res0, res0);
+  res = _mm256_permute4x64_epi64(res, 0xd8);
+  _mm_storeu_si128((__m128i *)(dst), _mm256_castsi256_si128(res));
+}
+
+static INLINE void blend_a64_d16_mask_w32_avx2(
+    uint8_t *dst, const CONV_BUF_TYPE *src0, const CONV_BUF_TYPE *src1,
+    const __m256i *m0, const __m256i *m1, const __m256i *v_round_offset,
+    const __m256i *v_maxval, int shift) {
+  const __m256i max_minus_m0 = _mm256_sub_epi16(*v_maxval, *m0);
+  const __m256i max_minus_m1 = _mm256_sub_epi16(*v_maxval, *m1);
+  const __m256i s0_0 = yy_loadu_256(src0);
+  const __m256i s0_1 = yy_loadu_256(src0 + 16);
+  const __m256i s1_0 = yy_loadu_256(src1);
+  const __m256i s1_1 = yy_loadu_256(src1 + 16);
+  __m256i res0_lo = _mm256_madd_epi16(_mm256_unpacklo_epi16(s0_0, s1_0),
+                                      _mm256_unpacklo_epi16(*m0, max_minus_m0));
+  __m256i res0_hi = _mm256_madd_epi16(_mm256_unpackhi_epi16(s0_0, s1_0),
+                                      _mm256_unpackhi_epi16(*m0, max_minus_m0));
+  __m256i res1_lo = _mm256_madd_epi16(_mm256_unpacklo_epi16(s0_1, s1_1),
+                                      _mm256_unpacklo_epi16(*m1, max_minus_m1));
+  __m256i res1_hi = _mm256_madd_epi16(_mm256_unpackhi_epi16(s0_1, s1_1),
+                                      _mm256_unpackhi_epi16(*m1, max_minus_m1));
+  res0_lo =
+      _mm256_srai_epi32(_mm256_sub_epi32(res0_lo, *v_round_offset), shift);
+  res0_hi =
+      _mm256_srai_epi32(_mm256_sub_epi32(res0_hi, *v_round_offset), shift);
+  res1_lo =
+      _mm256_srai_epi32(_mm256_sub_epi32(res1_lo, *v_round_offset), shift);
+  res1_hi =
+      _mm256_srai_epi32(_mm256_sub_epi32(res1_hi, *v_round_offset), shift);
+  const __m256i res0 = _mm256_packs_epi32(res0_lo, res0_hi);
+  const __m256i res1 = _mm256_packs_epi32(res1_lo, res1_hi);
+  __m256i res = _mm256_packus_epi16(res0, res1);
+  res = _mm256_permute4x64_epi64(res, 0xd8);
+  _mm256_storeu_si256((__m256i *)(dst), res);
+}
+
+static INLINE void lowbd_blend_a64_d16_mask_subw0_subh0_w16_avx2(
+    uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0,
+    uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h,
+    const __m256i *round_offset, int shift) {
+  const __m256i v_maxval = _mm256_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+  for (int i = 0; i < h; ++i) {
+    const __m128i m = xx_loadu_128(mask);
+    const __m256i m0 = _mm256_cvtepu8_epi16(m);
+
+    blend_a64_d16_mask_w16_avx2(dst, src0, src1, &m0, round_offset, &v_maxval,
+                                shift);
+    mask += mask_stride;
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+  }
+}
+
+static INLINE void lowbd_blend_a64_d16_mask_subw0_subh0_w32_avx2(
+    uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0,
+    uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, int w,
+    const __m256i *round_offset, int shift) {
+  const __m256i v_maxval = _mm256_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+  for (int i = 0; i < h; ++i) {
+    for (int j = 0; j < w; j += 32) {
+      const __m256i m = yy_loadu_256(mask + j);
+      const __m256i m0 = _mm256_cvtepu8_epi16(_mm256_castsi256_si128(m));
+      const __m256i m1 = _mm256_cvtepu8_epi16(_mm256_extracti128_si256(m, 1));
+
+      blend_a64_d16_mask_w32_avx2(dst + j, src0 + j, src1 + j, &m0, &m1,
+                                  round_offset, &v_maxval, shift);
+    }
+    mask += mask_stride;
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+  }
+}
+
+static INLINE void lowbd_blend_a64_d16_mask_subw1_subh1_w16_avx2(
+    uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0,
+    uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h,
+    const __m256i *round_offset, int shift) {
+  const __m256i v_maxval = _mm256_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+  const __m256i one_b = _mm256_set1_epi8(1);
+  const __m256i two_w = _mm256_set1_epi16(2);
+  for (int i = 0; i < h; ++i) {
+    const __m256i m_i00 = yy_loadu_256(mask);
+    const __m256i m_i10 = yy_loadu_256(mask + mask_stride);
+
+    const __m256i m0_ac = _mm256_adds_epu8(m_i00, m_i10);
+    const __m256i m0_acbd = _mm256_maddubs_epi16(m0_ac, one_b);
+    const __m256i m0 = _mm256_srli_epi16(_mm256_add_epi16(m0_acbd, two_w), 2);
+
+    blend_a64_d16_mask_w16_avx2(dst, src0, src1, &m0, round_offset, &v_maxval,
+                                shift);
+    mask += mask_stride << 1;
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+  }
+}
+
+static INLINE void lowbd_blend_a64_d16_mask_subw1_subh1_w32_avx2(
+    uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0,
+    uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, int w,
+    const __m256i *round_offset, int shift) {
+  const __m256i v_maxval = _mm256_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+  const __m256i one_b = _mm256_set1_epi8(1);
+  const __m256i two_w = _mm256_set1_epi16(2);
+  for (int i = 0; i < h; ++i) {
+    for (int j = 0; j < w; j += 32) {
+      const __m256i m_i00 = yy_loadu_256(mask + 2 * j);
+      const __m256i m_i01 = yy_loadu_256(mask + 2 * j + 32);
+      const __m256i m_i10 = yy_loadu_256(mask + mask_stride + 2 * j);
+      const __m256i m_i11 = yy_loadu_256(mask + mask_stride + 2 * j + 32);
+
+      const __m256i m0_ac = _mm256_adds_epu8(m_i00, m_i10);
+      const __m256i m1_ac = _mm256_adds_epu8(m_i01, m_i11);
+      const __m256i m0_acbd = _mm256_maddubs_epi16(m0_ac, one_b);
+      const __m256i m1_acbd = _mm256_maddubs_epi16(m1_ac, one_b);
+      const __m256i m0 = _mm256_srli_epi16(_mm256_add_epi16(m0_acbd, two_w), 2);
+      const __m256i m1 = _mm256_srli_epi16(_mm256_add_epi16(m1_acbd, two_w), 2);
+
+      blend_a64_d16_mask_w32_avx2(dst + j, src0 + j, src1 + j, &m0, &m1,
+                                  round_offset, &v_maxval, shift);
+    }
+    mask += mask_stride << 1;
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+  }
+}
+
+static INLINE void lowbd_blend_a64_d16_mask_subw1_subh0_w16_avx2(
+    uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0,
+    uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, int w,
+    const __m256i *round_offset, int shift) {
+  const __m256i v_maxval = _mm256_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+  const __m256i one_b = _mm256_set1_epi8(1);
+  const __m256i zeros = _mm256_setzero_si256();
+  for (int i = 0; i < h; ++i) {
+    for (int j = 0; j < w; j += 16) {
+      const __m256i m_i00 = yy_loadu_256(mask + 2 * j);
+      const __m256i m0_ac = _mm256_maddubs_epi16(m_i00, one_b);
+      const __m256i m0 = _mm256_avg_epu16(m0_ac, zeros);
+
+      blend_a64_d16_mask_w16_avx2(dst + j, src0 + j, src1 + j, &m0,
+                                  round_offset, &v_maxval, shift);
+    }
+    mask += mask_stride;
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+  }
+}
+
+static INLINE void lowbd_blend_a64_d16_mask_subw1_subh0_w32_avx2(
+    uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0,
+    uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, int w,
+    const __m256i *round_offset, int shift) {
+  const __m256i v_maxval = _mm256_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+  const __m256i one_b = _mm256_set1_epi8(1);
+  const __m256i zeros = _mm256_setzero_si256();
+  for (int i = 0; i < h; ++i) {
+    for (int j = 0; j < w; j += 32) {
+      const __m256i m_i00 = yy_loadu_256(mask + 2 * j);
+      const __m256i m_i01 = yy_loadu_256(mask + 2 * j + 32);
+      const __m256i m0_ac = _mm256_maddubs_epi16(m_i00, one_b);
+      const __m256i m1_ac = _mm256_maddubs_epi16(m_i01, one_b);
+      const __m256i m0 = _mm256_avg_epu16(m0_ac, zeros);
+      const __m256i m1 = _mm256_avg_epu16(m1_ac, zeros);
+
+      blend_a64_d16_mask_w32_avx2(dst + j, src0 + j, src1 + j, &m0, &m1,
+                                  round_offset, &v_maxval, shift);
+    }
+    mask += mask_stride;
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+  }
+}
+
+static INLINE void lowbd_blend_a64_d16_mask_subw0_subh1_w16_avx2(
+    uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0,
+    uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, int w,
+    const __m256i *round_offset, int shift) {
+  const __m256i v_maxval = _mm256_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+  const __m128i zeros = _mm_setzero_si128();
+  for (int i = 0; i < h; ++i) {
+    for (int j = 0; j < w; j += 16) {
+      const __m128i m_i00 = xx_loadu_128(mask + j);
+      const __m128i m_i10 = xx_loadu_128(mask + mask_stride + j);
+
+      const __m128i m_ac = _mm_avg_epu8(_mm_adds_epu8(m_i00, m_i10), zeros);
+      const __m256i m0 = _mm256_cvtepu8_epi16(m_ac);
+
+      blend_a64_d16_mask_w16_avx2(dst + j, src0 + j, src1 + j, &m0,
+                                  round_offset, &v_maxval, shift);
+    }
+    mask += mask_stride << 1;
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+  }
+}
+
+static INLINE void lowbd_blend_a64_d16_mask_subw0_subh1_w32_avx2(
+    uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0,
+    uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, int w,
+    const __m256i *round_offset, int shift) {
+  const __m256i v_maxval = _mm256_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+  const __m256i zeros = _mm256_setzero_si256();
+  for (int i = 0; i < h; ++i) {
+    for (int j = 0; j < w; j += 32) {
+      const __m256i m_i00 = yy_loadu_256(mask + j);
+      const __m256i m_i10 = yy_loadu_256(mask + mask_stride + j);
+
+      const __m256i m_ac =
+          _mm256_avg_epu8(_mm256_adds_epu8(m_i00, m_i10), zeros);
+      const __m256i m0 = _mm256_cvtepu8_epi16(_mm256_castsi256_si128(m_ac));
+      const __m256i m1 =
+          _mm256_cvtepu8_epi16(_mm256_extracti128_si256(m_ac, 1));
+
+      blend_a64_d16_mask_w32_avx2(dst + j, src0 + j, src1 + j, &m0, &m1,
+                                  round_offset, &v_maxval, shift);
+    }
+    mask += mask_stride << 1;
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+  }
+}
+
+void aom_lowbd_blend_a64_d16_mask_avx2(
+    uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0,
+    uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh,
+    ConvolveParams *conv_params) {
+  const int bd = 8;
+  const int round_bits =
+      2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+
+  const int round_offset =
+      ((1 << (round_bits + bd)) + (1 << (round_bits + bd - 1)) -
+       (1 << (round_bits - 1)))
+      << AOM_BLEND_A64_ROUND_BITS;
+
+  const int shift = round_bits + AOM_BLEND_A64_ROUND_BITS;
+  assert(IMPLIES((void *)src0 == dst, src0_stride == dst_stride));
+  assert(IMPLIES((void *)src1 == dst, src1_stride == dst_stride));
+
+  assert(h >= 4);
+  assert(w >= 4);
+  assert(IS_POWER_OF_TWO(h));
+  assert(IS_POWER_OF_TWO(w));
+  const __m128i v_round_offset = _mm_set1_epi32(round_offset);
+  const __m256i y_round_offset = _mm256_set1_epi32(round_offset);
+
+  if (subw == 0 && subh == 0) {
+    switch (w) {
+      case 4:
+        aom_lowbd_blend_a64_d16_mask_subw0_subh0_w4_sse4_1(
+            dst, dst_stride, src0, src0_stride, src1, src1_stride, mask,
+            mask_stride, h, &v_round_offset, shift);
+        break;
+      case 8:
+        aom_lowbd_blend_a64_d16_mask_subw0_subh0_w8_sse4_1(
+            dst, dst_stride, src0, src0_stride, src1, src1_stride, mask,
+            mask_stride, h, &v_round_offset, shift);
+        break;
+      case 16:
+        lowbd_blend_a64_d16_mask_subw0_subh0_w16_avx2(
+            dst, dst_stride, src0, src0_stride, src1, src1_stride, mask,
+            mask_stride, h, &y_round_offset, shift);
+        break;
+      default:
+        lowbd_blend_a64_d16_mask_subw0_subh0_w32_avx2(
+            dst, dst_stride, src0, src0_stride, src1, src1_stride, mask,
+            mask_stride, h, w, &y_round_offset, shift);
+        break;
+    }
+  } else if (subw == 1 && subh == 1) {
+    switch (w) {
+      case 4:
+        aom_lowbd_blend_a64_d16_mask_subw1_subh1_w4_sse4_1(
+            dst, dst_stride, src0, src0_stride, src1, src1_stride, mask,
+            mask_stride, h, &v_round_offset, shift);
+        break;
+      case 8:
+        aom_lowbd_blend_a64_d16_mask_subw1_subh1_w8_sse4_1(
+            dst, dst_stride, src0, src0_stride, src1, src1_stride, mask,
+            mask_stride, h, &v_round_offset, shift);
+        break;
+      case 16:
+        lowbd_blend_a64_d16_mask_subw1_subh1_w16_avx2(
+            dst, dst_stride, src0, src0_stride, src1, src1_stride, mask,
+            mask_stride, h, &y_round_offset, shift);
+        break;
+      default:
+        lowbd_blend_a64_d16_mask_subw1_subh1_w32_avx2(
+            dst, dst_stride, src0, src0_stride, src1, src1_stride, mask,
+            mask_stride, h, w, &y_round_offset, shift);
+        break;
+    }
+  } else if (subw == 1 && subh == 0) {
+    switch (w) {
+      case 4:
+        aom_lowbd_blend_a64_d16_mask_subw1_subh0_w4_sse4_1(
+            dst, dst_stride, src0, src0_stride, src1, src1_stride, mask,
+            mask_stride, h, &v_round_offset, shift);
+        break;
+      case 8:
+        aom_lowbd_blend_a64_d16_mask_subw1_subh0_w8_sse4_1(
+            dst, dst_stride, src0, src0_stride, src1, src1_stride, mask,
+            mask_stride, h, &v_round_offset, shift);
+        break;
+      case 16:
+        lowbd_blend_a64_d16_mask_subw1_subh0_w16_avx2(
+            dst, dst_stride, src0, src0_stride, src1, src1_stride, mask,
+            mask_stride, h, w, &y_round_offset, shift);
+        break;
+      default:
+        lowbd_blend_a64_d16_mask_subw1_subh0_w32_avx2(
+            dst, dst_stride, src0, src0_stride, src1, src1_stride, mask,
+            mask_stride, h, w, &y_round_offset, shift);
+        break;
+    }
+  } else {
+    switch (w) {
+      case 4:
+        aom_lowbd_blend_a64_d16_mask_subw0_subh1_w4_sse4_1(
+            dst, dst_stride, src0, src0_stride, src1, src1_stride, mask,
+            mask_stride, h, &v_round_offset, shift);
+        break;
+      case 8:
+        aom_lowbd_blend_a64_d16_mask_subw0_subh1_w8_sse4_1(
+            dst, dst_stride, src0, src0_stride, src1, src1_stride, mask,
+            mask_stride, h, &v_round_offset, shift);
+        break;
+      case 16:
+        lowbd_blend_a64_d16_mask_subw0_subh1_w16_avx2(
+            dst, dst_stride, src0, src0_stride, src1, src1_stride, mask,
+            mask_stride, h, w, &y_round_offset, shift);
+        break;
+      default:
+        lowbd_blend_a64_d16_mask_subw0_subh1_w32_avx2(
+            dst, dst_stride, src0, src0_stride, src1, src1_stride, mask,
+            mask_stride, h, w, &y_round_offset, shift);
+        break;
+    }
+  }
+}
+
+static INLINE __m256i blend_16_u8_avx2(const uint8_t *src0, const uint8_t *src1,
+                                       const __m256i *v_m0_b,
+                                       const __m256i *v_m1_b,
+                                       const int32_t bits) {
+  const __m256i v_s0_b = _mm256_castsi128_si256(xx_loadu_128(src0));
+  const __m256i v_s1_b = _mm256_castsi128_si256(xx_loadu_128(src1));
+  const __m256i v_s0_s_b = _mm256_permute4x64_epi64(v_s0_b, 0xd8);
+  const __m256i v_s1_s_b = _mm256_permute4x64_epi64(v_s1_b, 0xd8);
+
+  const __m256i v_p0_w =
+      _mm256_maddubs_epi16(_mm256_unpacklo_epi8(v_s0_s_b, v_s1_s_b),
+                           _mm256_unpacklo_epi8(*v_m0_b, *v_m1_b));
+
+  const __m256i v_res0_w = yy_roundn_epu16(v_p0_w, bits);
+  const __m256i v_res_b = _mm256_packus_epi16(v_res0_w, v_res0_w);
+  const __m256i v_res = _mm256_permute4x64_epi64(v_res_b, 0xd8);
+  return v_res;
+}
+
+static INLINE __m256i blend_32_u8_avx2(const uint8_t *src0, const uint8_t *src1,
+                                       const __m256i *v_m0_b,
+                                       const __m256i *v_m1_b,
+                                       const int32_t bits) {
+  const __m256i v_s0_b = yy_loadu_256(src0);
+  const __m256i v_s1_b = yy_loadu_256(src1);
+
+  const __m256i v_p0_w =
+      _mm256_maddubs_epi16(_mm256_unpacklo_epi8(v_s0_b, v_s1_b),
+                           _mm256_unpacklo_epi8(*v_m0_b, *v_m1_b));
+  const __m256i v_p1_w =
+      _mm256_maddubs_epi16(_mm256_unpackhi_epi8(v_s0_b, v_s1_b),
+                           _mm256_unpackhi_epi8(*v_m0_b, *v_m1_b));
+
+  const __m256i v_res0_w = yy_roundn_epu16(v_p0_w, bits);
+  const __m256i v_res1_w = yy_roundn_epu16(v_p1_w, bits);
+  const __m256i v_res = _mm256_packus_epi16(v_res0_w, v_res1_w);
+  return v_res;
+}
+
+static INLINE void blend_a64_mask_sx_sy_w16_avx2(
+    uint8_t *dst, uint32_t dst_stride, const uint8_t *src0,
+    uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h) {
+  const __m256i v_zmask_b = _mm256_set1_epi16(0xFF);
+  const __m256i v_maxval_b = _mm256_set1_epi8(AOM_BLEND_A64_MAX_ALPHA);
+  do {
+    const __m256i v_ral_b = yy_loadu_256(mask);
+    const __m256i v_rbl_b = yy_loadu_256(mask + mask_stride);
+    const __m256i v_rvsl_b = _mm256_add_epi8(v_ral_b, v_rbl_b);
+    const __m256i v_rvsal_w = _mm256_and_si256(v_rvsl_b, v_zmask_b);
+    const __m256i v_rvsbl_w =
+        _mm256_and_si256(_mm256_srli_si256(v_rvsl_b, 1), v_zmask_b);
+    const __m256i v_rsl_w = _mm256_add_epi16(v_rvsal_w, v_rvsbl_w);
+
+    const __m256i v_m0_w = yy_roundn_epu16(v_rsl_w, 2);
+    const __m256i v_m0_b = _mm256_packus_epi16(v_m0_w, v_m0_w);
+    const __m256i v_m1_b = _mm256_sub_epi8(v_maxval_b, v_m0_b);
+
+    const __m256i y_res_b = blend_16_u8_avx2(src0, src1, &v_m0_b, &v_m1_b,
+                                             AOM_BLEND_A64_ROUND_BITS);
+
+    xx_storeu_128(dst, _mm256_castsi256_si128(y_res_b));
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += 2 * mask_stride;
+  } while (--h);
+}
+
+static INLINE void blend_a64_mask_sx_sy_w32n_avx2(
+    uint8_t *dst, uint32_t dst_stride, const uint8_t *src0,
+    uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int w, int h) {
+  const __m256i v_maxval_b = _mm256_set1_epi8(AOM_BLEND_A64_MAX_ALPHA);
+  const __m256i v_zmask_b = _mm256_set1_epi16(0xFF);
+  do {
+    int c;
+    for (c = 0; c < w; c += 32) {
+      const __m256i v_ral_b = yy_loadu_256(mask + 2 * c);
+      const __m256i v_rah_b = yy_loadu_256(mask + 2 * c + 32);
+      const __m256i v_rbl_b = yy_loadu_256(mask + mask_stride + 2 * c);
+      const __m256i v_rbh_b = yy_loadu_256(mask + mask_stride + 2 * c + 32);
+      const __m256i v_rvsl_b = _mm256_add_epi8(v_ral_b, v_rbl_b);
+      const __m256i v_rvsh_b = _mm256_add_epi8(v_rah_b, v_rbh_b);
+      const __m256i v_rvsal_w = _mm256_and_si256(v_rvsl_b, v_zmask_b);
+      const __m256i v_rvsah_w = _mm256_and_si256(v_rvsh_b, v_zmask_b);
+      const __m256i v_rvsbl_w =
+          _mm256_and_si256(_mm256_srli_si256(v_rvsl_b, 1), v_zmask_b);
+      const __m256i v_rvsbh_w =
+          _mm256_and_si256(_mm256_srli_si256(v_rvsh_b, 1), v_zmask_b);
+      const __m256i v_rsl_w = _mm256_add_epi16(v_rvsal_w, v_rvsbl_w);
+      const __m256i v_rsh_w = _mm256_add_epi16(v_rvsah_w, v_rvsbh_w);
+
+      const __m256i v_m0l_w = yy_roundn_epu16(v_rsl_w, 2);
+      const __m256i v_m0h_w = yy_roundn_epu16(v_rsh_w, 2);
+      const __m256i v_m0_b =
+          _mm256_permute4x64_epi64(_mm256_packus_epi16(v_m0l_w, v_m0h_w), 0xd8);
+      const __m256i v_m1_b = _mm256_sub_epi8(v_maxval_b, v_m0_b);
+
+      const __m256i v_res_b = blend_32_u8_avx2(
+          src0 + c, src1 + c, &v_m0_b, &v_m1_b, AOM_BLEND_A64_ROUND_BITS);
+
+      yy_storeu_256(dst + c, v_res_b);
+    }
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += 2 * mask_stride;
+  } while (--h);
+}
+
+static INLINE void blend_a64_mask_sx_sy_avx2(
+    uint8_t *dst, uint32_t dst_stride, const uint8_t *src0,
+    uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int w, int h) {
+  const __m128i v_shuffle_b = xx_loadu_128(g_blend_a64_mask_shuffle);
+  const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA);
+  const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS));
+  switch (w) {
+    case 4:
+      do {
+        const __m128i v_ra_b = xx_loadl_64(mask);
+        const __m128i v_rb_b = xx_loadl_64(mask + mask_stride);
+        const __m128i v_rvs_b = _mm_add_epi8(v_ra_b, v_rb_b);
+        const __m128i v_r_s_b = _mm_shuffle_epi8(v_rvs_b, v_shuffle_b);
+        const __m128i v_r0_s_w = _mm_cvtepu8_epi16(v_r_s_b);
+        const __m128i v_r1_s_w = _mm_cvtepu8_epi16(_mm_srli_si128(v_r_s_b, 8));
+        const __m128i v_rs_w = _mm_add_epi16(v_r0_s_w, v_r1_s_w);
+        const __m128i v_m0_w = xx_roundn_epu16(v_rs_w, 2);
+        const __m128i v_m0_b = _mm_packus_epi16(v_m0_w, v_m0_w);
+        const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b);
+
+        const __m128i v_res_b = blend_4_u8(src0, src1, &v_m0_b, &v_m1_b, &_r);
+
+        xx_storel_32(dst, v_res_b);
+
+        dst += dst_stride;
+        src0 += src0_stride;
+        src1 += src1_stride;
+        mask += 2 * mask_stride;
+      } while (--h);
+      break;
+    case 8:
+      do {
+        const __m128i v_ra_b = xx_loadu_128(mask);
+        const __m128i v_rb_b = xx_loadu_128(mask + mask_stride);
+        const __m128i v_rvs_b = _mm_add_epi8(v_ra_b, v_rb_b);
+        const __m128i v_r_s_b = _mm_shuffle_epi8(v_rvs_b, v_shuffle_b);
+        const __m128i v_r0_s_w = _mm_cvtepu8_epi16(v_r_s_b);
+        const __m128i v_r1_s_w = _mm_cvtepu8_epi16(_mm_srli_si128(v_r_s_b, 8));
+        const __m128i v_rs_w = _mm_add_epi16(v_r0_s_w, v_r1_s_w);
+        const __m128i v_m0_w = xx_roundn_epu16(v_rs_w, 2);
+        const __m128i v_m0_b = _mm_packus_epi16(v_m0_w, v_m0_w);
+        const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b);
+
+        const __m128i v_res_b = blend_8_u8(src0, src1, &v_m0_b, &v_m1_b, &_r);
+
+        xx_storel_64(dst, v_res_b);
+
+        dst += dst_stride;
+        src0 += src0_stride;
+        src1 += src1_stride;
+        mask += 2 * mask_stride;
+      } while (--h);
+      break;
+    case 16:
+      blend_a64_mask_sx_sy_w16_avx2(dst, dst_stride, src0, src0_stride, src1,
+                                    src1_stride, mask, mask_stride, h);
+      break;
+    default:
+      blend_a64_mask_sx_sy_w32n_avx2(dst, dst_stride, src0, src0_stride, src1,
+                                     src1_stride, mask, mask_stride, w, h);
+      break;
+  }
+}
+
+static INLINE void blend_a64_mask_sx_w16_avx2(
+    uint8_t *dst, uint32_t dst_stride, const uint8_t *src0,
+    uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h) {
+  const __m256i v_maxval_b = _mm256_set1_epi8(AOM_BLEND_A64_MAX_ALPHA);
+  const __m256i v_zmask_b = _mm256_set1_epi16(0xff);
+  do {
+    const __m256i v_rl_b = yy_loadu_256(mask);
+    const __m256i v_al_b =
+        _mm256_avg_epu8(v_rl_b, _mm256_srli_si256(v_rl_b, 1));
+
+    const __m256i v_m0_w = _mm256_and_si256(v_al_b, v_zmask_b);
+    const __m256i v_m0_b = _mm256_packus_epi16(v_m0_w, _mm256_setzero_si256());
+    const __m256i v_m1_b = _mm256_sub_epi8(v_maxval_b, v_m0_b);
+
+    const __m256i v_res_b = blend_16_u8_avx2(src0, src1, &v_m0_b, &v_m1_b,
+                                             AOM_BLEND_A64_ROUND_BITS);
+
+    xx_storeu_128(dst, _mm256_castsi256_si128(v_res_b));
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += mask_stride;
+  } while (--h);
+}
+
+static INLINE void blend_a64_mask_sx_w32n_avx2(
+    uint8_t *dst, uint32_t dst_stride, const uint8_t *src0,
+    uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int w, int h) {
+  const __m256i v_shuffle_b = yy_loadu_256(g_blend_a64_mask_shuffle);
+  const __m256i v_maxval_b = _mm256_set1_epi8(AOM_BLEND_A64_MAX_ALPHA);
+  do {
+    int c;
+    for (c = 0; c < w; c += 32) {
+      const __m256i v_r0_b = yy_loadu_256(mask + 2 * c);
+      const __m256i v_r1_b = yy_loadu_256(mask + 2 * c + 32);
+      const __m256i v_r0_s_b = _mm256_shuffle_epi8(v_r0_b, v_shuffle_b);
+      const __m256i v_r1_s_b = _mm256_shuffle_epi8(v_r1_b, v_shuffle_b);
+      const __m256i v_al_b =
+          _mm256_avg_epu8(v_r0_s_b, _mm256_srli_si256(v_r0_s_b, 8));
+      const __m256i v_ah_b =
+          _mm256_avg_epu8(v_r1_s_b, _mm256_srli_si256(v_r1_s_b, 8));
+
+      const __m256i v_m0_b =
+          _mm256_permute4x64_epi64(_mm256_unpacklo_epi64(v_al_b, v_ah_b), 0xd8);
+      const __m256i v_m1_b = _mm256_sub_epi8(v_maxval_b, v_m0_b);
+
+      const __m256i v_res_b = blend_32_u8_avx2(
+          src0 + c, src1 + c, &v_m0_b, &v_m1_b, AOM_BLEND_A64_ROUND_BITS);
+
+      yy_storeu_256(dst + c, v_res_b);
+    }
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += mask_stride;
+  } while (--h);
+}
+
+static INLINE void blend_a64_mask_sx_avx2(
+    uint8_t *dst, uint32_t dst_stride, const uint8_t *src0,
+    uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int w, int h) {
+  const __m128i v_shuffle_b = xx_loadu_128(g_blend_a64_mask_shuffle);
+  const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA);
+  const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS));
+  switch (w) {
+    case 4:
+      do {
+        const __m128i v_r_b = xx_loadl_64(mask);
+        const __m128i v_r0_s_b = _mm_shuffle_epi8(v_r_b, v_shuffle_b);
+        const __m128i v_r_lo_b = _mm_unpacklo_epi64(v_r0_s_b, v_r0_s_b);
+        const __m128i v_r_hi_b = _mm_unpackhi_epi64(v_r0_s_b, v_r0_s_b);
+        const __m128i v_m0_b = _mm_avg_epu8(v_r_lo_b, v_r_hi_b);
+        const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b);
+
+        const __m128i v_res_b = blend_4_u8(src0, src1, &v_m0_b, &v_m1_b, &_r);
+
+        xx_storel_32(dst, v_res_b);
+
+        dst += dst_stride;
+        src0 += src0_stride;
+        src1 += src1_stride;
+        mask += mask_stride;
+      } while (--h);
+      break;
+    case 8:
+      do {
+        const __m128i v_r_b = xx_loadu_128(mask);
+        const __m128i v_r0_s_b = _mm_shuffle_epi8(v_r_b, v_shuffle_b);
+        const __m128i v_r_lo_b = _mm_unpacklo_epi64(v_r0_s_b, v_r0_s_b);
+        const __m128i v_r_hi_b = _mm_unpackhi_epi64(v_r0_s_b, v_r0_s_b);
+        const __m128i v_m0_b = _mm_avg_epu8(v_r_lo_b, v_r_hi_b);
+        const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b);
+
+        const __m128i v_res_b = blend_8_u8(src0, src1, &v_m0_b, &v_m1_b, &_r);
+
+        xx_storel_64(dst, v_res_b);
+
+        dst += dst_stride;
+        src0 += src0_stride;
+        src1 += src1_stride;
+        mask += mask_stride;
+      } while (--h);
+      break;
+    case 16:
+      blend_a64_mask_sx_w16_avx2(dst, dst_stride, src0, src0_stride, src1,
+                                 src1_stride, mask, mask_stride, h);
+      break;
+    default:
+      blend_a64_mask_sx_w32n_avx2(dst, dst_stride, src0, src0_stride, src1,
+                                  src1_stride, mask, mask_stride, w, h);
+      break;
+  }
+}
+
+static INLINE void blend_a64_mask_sy_w16_avx2(
+    uint8_t *dst, uint32_t dst_stride, const uint8_t *src0,
+    uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h) {
+  const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS));
+  const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA);
+  do {
+    const __m128i v_ra_b = xx_loadu_128(mask);
+    const __m128i v_rb_b = xx_loadu_128(mask + mask_stride);
+    const __m128i v_m0_b = _mm_avg_epu8(v_ra_b, v_rb_b);
+
+    const __m128i v_m1_b = _mm_sub_epi16(v_maxval_b, v_m0_b);
+    const __m128i v_res_b = blend_16_u8(src0, src1, &v_m0_b, &v_m1_b, &_r);
+
+    xx_storeu_128(dst, v_res_b);
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += 2 * mask_stride;
+  } while (--h);
+}
+
+static INLINE void blend_a64_mask_sy_w32n_avx2(
+    uint8_t *dst, uint32_t dst_stride, const uint8_t *src0,
+    uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int w, int h) {
+  const __m256i v_maxval_b = _mm256_set1_epi8(AOM_BLEND_A64_MAX_ALPHA);
+  do {
+    int c;
+    for (c = 0; c < w; c += 32) {
+      const __m256i v_ra_b = yy_loadu_256(mask + c);
+      const __m256i v_rb_b = yy_loadu_256(mask + c + mask_stride);
+      const __m256i v_m0_b = _mm256_avg_epu8(v_ra_b, v_rb_b);
+      const __m256i v_m1_b = _mm256_sub_epi8(v_maxval_b, v_m0_b);
+      const __m256i v_res_b = blend_32_u8_avx2(
+          src0 + c, src1 + c, &v_m0_b, &v_m1_b, AOM_BLEND_A64_ROUND_BITS);
+
+      yy_storeu_256(dst + c, v_res_b);
+    }
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += 2 * mask_stride;
+  } while (--h);
+}
+
+static INLINE void blend_a64_mask_sy_avx2(
+    uint8_t *dst, uint32_t dst_stride, const uint8_t *src0,
+    uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int w, int h) {
+  const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS));
+  const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA);
+  switch (w) {
+    case 4:
+      do {
+        const __m128i v_ra_b = xx_loadl_32(mask);
+        const __m128i v_rb_b = xx_loadl_32(mask + mask_stride);
+        const __m128i v_m0_b = _mm_avg_epu8(v_ra_b, v_rb_b);
+        const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b);
+        const __m128i v_res_b = blend_4_u8(src0, src1, &v_m0_b, &v_m1_b, &_r);
+
+        xx_storel_32(dst, v_res_b);
+
+        dst += dst_stride;
+        src0 += src0_stride;
+        src1 += src1_stride;
+        mask += 2 * mask_stride;
+      } while (--h);
+      break;
+    case 8:
+      do {
+        const __m128i v_ra_b = xx_loadl_64(mask);
+        const __m128i v_rb_b = xx_loadl_64(mask + mask_stride);
+        const __m128i v_m0_b = _mm_avg_epu8(v_ra_b, v_rb_b);
+        const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b);
+        const __m128i v_res_b = blend_8_u8(src0, src1, &v_m0_b, &v_m1_b, &_r);
+
+        xx_storel_64(dst, v_res_b);
+
+        dst += dst_stride;
+        src0 += src0_stride;
+        src1 += src1_stride;
+        mask += 2 * mask_stride;
+      } while (--h);
+      break;
+    case 16:
+      blend_a64_mask_sy_w16_avx2(dst, dst_stride, src0, src0_stride, src1,
+                                 src1_stride, mask, mask_stride, h);
+      break;
+    default:
+      blend_a64_mask_sy_w32n_avx2(dst, dst_stride, src0, src0_stride, src1,
+                                  src1_stride, mask, mask_stride, w, h);
+  }
+}
+
+static INLINE void blend_a64_mask_w32n_avx2(
+    uint8_t *dst, uint32_t dst_stride, const uint8_t *src0,
+    uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int w, int h) {
+  const __m256i v_maxval_b = _mm256_set1_epi8(AOM_BLEND_A64_MAX_ALPHA);
+  do {
+    int c;
+    for (c = 0; c < w; c += 32) {
+      const __m256i v_m0_b = yy_loadu_256(mask + c);
+      const __m256i v_m1_b = _mm256_sub_epi8(v_maxval_b, v_m0_b);
+
+      const __m256i v_res_b = blend_32_u8_avx2(
+          src0 + c, src1 + c, &v_m0_b, &v_m1_b, AOM_BLEND_A64_ROUND_BITS);
+
+      yy_storeu_256(dst + c, v_res_b);
+    }
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += mask_stride;
+  } while (--h);
+}
+
+static INLINE void blend_a64_mask_avx2(
+    uint8_t *dst, uint32_t dst_stride, const uint8_t *src0,
+    uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int w, int h) {
+  const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA);
+  const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS));
+  switch (w) {
+    case 4:
+      do {
+        const __m128i v_m0_b = xx_loadl_32(mask);
+        const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b);
+        const __m128i v_res_b = blend_4_u8(src0, src1, &v_m0_b, &v_m1_b, &_r);
+
+        xx_storel_32(dst, v_res_b);
+
+        dst += dst_stride;
+        src0 += src0_stride;
+        src1 += src1_stride;
+        mask += mask_stride;
+      } while (--h);
+      break;
+    case 8:
+      do {
+        const __m128i v_m0_b = xx_loadl_64(mask);
+        const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b);
+        const __m128i v_res_b = blend_8_u8(src0, src1, &v_m0_b, &v_m1_b, &_r);
+
+        xx_storel_64(dst, v_res_b);
+
+        dst += dst_stride;
+        src0 += src0_stride;
+        src1 += src1_stride;
+        mask += mask_stride;
+      } while (--h);
+      break;
+    case 16:
+      do {
+        const __m128i v_m0_b = xx_loadu_128(mask);
+        const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b);
+        const __m128i v_res_b = blend_16_u8(src0, src1, &v_m0_b, &v_m1_b, &_r);
+
+        xx_storeu_128(dst, v_res_b);
+        dst += dst_stride;
+        src0 += src0_stride;
+        src1 += src1_stride;
+        mask += mask_stride;
+      } while (--h);
+      break;
+    default:
+      blend_a64_mask_w32n_avx2(dst, dst_stride, src0, src0_stride, src1,
+                               src1_stride, mask, mask_stride, w, h);
+  }
+}
+
+void aom_blend_a64_mask_avx2(uint8_t *dst, uint32_t dst_stride,
+                             const uint8_t *src0, uint32_t src0_stride,
+                             const uint8_t *src1, uint32_t src1_stride,
+                             const uint8_t *mask, uint32_t mask_stride, int w,
+                             int h, int subw, int subh) {
+  assert(IMPLIES(src0 == dst, src0_stride == dst_stride));
+  assert(IMPLIES(src1 == dst, src1_stride == dst_stride));
+
+  assert(h >= 1);
+  assert(w >= 1);
+  assert(IS_POWER_OF_TWO(h));
+  assert(IS_POWER_OF_TWO(w));
+
+  if (UNLIKELY((h | w) & 3)) {  // if (w <= 2 || h <= 2)
+    aom_blend_a64_mask_c(dst, dst_stride, src0, src0_stride, src1, src1_stride,
+                         mask, mask_stride, w, h, subw, subh);
+  } else {
+    if (subw & subh) {
+      blend_a64_mask_sx_sy_avx2(dst, dst_stride, src0, src0_stride, src1,
+                                src1_stride, mask, mask_stride, w, h);
+    } else if (subw) {
+      blend_a64_mask_sx_avx2(dst, dst_stride, src0, src0_stride, src1,
+                             src1_stride, mask, mask_stride, w, h);
+    } else if (subh) {
+      blend_a64_mask_sy_avx2(dst, dst_stride, src0, src0_stride, src1,
+                             src1_stride, mask, mask_stride, w, h);
+    } else {
+      blend_a64_mask_avx2(dst, dst_stride, src0, src0_stride, src1, src1_stride,
+                          mask, mask_stride, w, h);
+    }
+  }
+}
+
+#if CONFIG_AV1_HIGHBITDEPTH
+//////////////////////////////////////////////////////////////////////////////
+// aom_highbd_blend_a64_d16_mask_avx2()
+//////////////////////////////////////////////////////////////////////////////
+
+static INLINE void highbd_blend_a64_d16_mask_w4_avx2(
+    uint16_t *dst, int dst_stride, const CONV_BUF_TYPE *src0, int src0_stride,
+    const CONV_BUF_TYPE *src1, int src1_stride, const __m256i *mask0,
+    const __m256i *round_offset, int shift, const __m256i *clip_low,
+    const __m256i *clip_high, const __m256i *mask_max) {
+  // Load 4x u16 pixels from each of 4 rows from each source
+  const __m256i s0 = _mm256_set_epi64x(*(int64_t *)(src0 + 3 * src0_stride),
+                                       *(int64_t *)(src0 + 2 * src0_stride),
+                                       *(int64_t *)(src0 + 1 * src0_stride),
+                                       *(int64_t *)(src0 + 0 * src0_stride));
+  const __m256i s1 = _mm256_set_epi64x(*(int64_t *)(src1 + 3 * src1_stride),
+                                       *(int64_t *)(src1 + 2 * src1_stride),
+                                       *(int64_t *)(src1 + 1 * src1_stride),
+                                       *(int64_t *)(src1 + 0 * src1_stride));
+  // Generate the inverse mask
+  const __m256i mask1 = _mm256_sub_epi16(*mask_max, *mask0);
+
+  // Multiply each mask by the respective source
+  const __m256i mul0_highs = _mm256_mulhi_epu16(*mask0, s0);
+  const __m256i mul0_lows = _mm256_mullo_epi16(*mask0, s0);
+  const __m256i mul0h = _mm256_unpackhi_epi16(mul0_lows, mul0_highs);
+  const __m256i mul0l = _mm256_unpacklo_epi16(mul0_lows, mul0_highs);
+  // Note that AVX2 unpack orders 64-bit words as [3 1] [2 0] to keep within
+  // lanes Later, packs does the same again which cancels this out with no need
+  // for a permute.  The intermediate values being reordered makes no difference
+
+  const __m256i mul1_highs = _mm256_mulhi_epu16(mask1, s1);
+  const __m256i mul1_lows = _mm256_mullo_epi16(mask1, s1);
+  const __m256i mul1h = _mm256_unpackhi_epi16(mul1_lows, mul1_highs);
+  const __m256i mul1l = _mm256_unpacklo_epi16(mul1_lows, mul1_highs);
+
+  const __m256i sumh = _mm256_add_epi32(mul0h, mul1h);
+  const __m256i suml = _mm256_add_epi32(mul0l, mul1l);
+
+  const __m256i roundh =
+      _mm256_srai_epi32(_mm256_sub_epi32(sumh, *round_offset), shift);
+  const __m256i roundl =
+      _mm256_srai_epi32(_mm256_sub_epi32(suml, *round_offset), shift);
+
+  const __m256i pack = _mm256_packs_epi32(roundl, roundh);
+  const __m256i clip =
+      _mm256_min_epi16(_mm256_max_epi16(pack, *clip_low), *clip_high);
+
+  // _mm256_extract_epi64 doesn't exist on x86, so do it the old-fashioned way:
+  const __m128i cliph = _mm256_extracti128_si256(clip, 1);
+  xx_storel_64(dst + 3 * dst_stride, _mm_srli_si128(cliph, 8));
+  xx_storel_64(dst + 2 * dst_stride, cliph);
+  const __m128i clipl = _mm256_castsi256_si128(clip);
+  xx_storel_64(dst + 1 * dst_stride, _mm_srli_si128(clipl, 8));
+  xx_storel_64(dst + 0 * dst_stride, clipl);
+}
+
+static INLINE void highbd_blend_a64_d16_mask_subw0_subh0_w4_avx2(
+    uint16_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0,
+    uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h,
+    const __m256i *round_offset, int shift, const __m256i *clip_low,
+    const __m256i *clip_high, const __m256i *mask_max) {
+  do {
+    // Load 8x u8 pixels from each of 4 rows of the mask, pad each to u16
+    const __m128i mask08 = _mm_set_epi32(*(int32_t *)(mask + 3 * mask_stride),
+                                         *(int32_t *)(mask + 2 * mask_stride),
+                                         *(int32_t *)(mask + 1 * mask_stride),
+                                         *(int32_t *)(mask + 0 * mask_stride));
+    const __m256i mask0 = _mm256_cvtepu8_epi16(mask08);
+
+    highbd_blend_a64_d16_mask_w4_avx2(dst, dst_stride, src0, src0_stride, src1,
+                                      src1_stride, &mask0, round_offset, shift,
+                                      clip_low, clip_high, mask_max);
+
+    dst += dst_stride * 4;
+    src0 += src0_stride * 4;
+    src1 += src1_stride * 4;
+    mask += mask_stride * 4;
+  } while (h -= 4);
+}
+
+static INLINE void highbd_blend_a64_d16_mask_subw1_subh1_w4_avx2(
+    uint16_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0,
+    uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h,
+    const __m256i *round_offset, int shift, const __m256i *clip_low,
+    const __m256i *clip_high, const __m256i *mask_max) {
+  const __m256i one_b = _mm256_set1_epi8(1);
+  const __m256i two_w = _mm256_set1_epi16(2);
+  do {
+    // Load 8 pixels from each of 8 rows of mask,
+    // (saturating) add together rows then use madd to add adjacent pixels
+    // Finally, divide each value by 4 (with rounding)
+    const __m256i m0246 =
+        _mm256_set_epi64x(*(int64_t *)(mask + 6 * mask_stride),
+                          *(int64_t *)(mask + 4 * mask_stride),
+                          *(int64_t *)(mask + 2 * mask_stride),
+                          *(int64_t *)(mask + 0 * mask_stride));
+    const __m256i m1357 =
+        _mm256_set_epi64x(*(int64_t *)(mask + 7 * mask_stride),
+                          *(int64_t *)(mask + 5 * mask_stride),
+                          *(int64_t *)(mask + 3 * mask_stride),
+                          *(int64_t *)(mask + 1 * mask_stride));
+    const __m256i addrows = _mm256_adds_epu8(m0246, m1357);
+    const __m256i adjacent = _mm256_maddubs_epi16(addrows, one_b);
+    const __m256i mask0 =
+        _mm256_srli_epi16(_mm256_add_epi16(adjacent, two_w), 2);
+
+    highbd_blend_a64_d16_mask_w4_avx2(dst, dst_stride, src0, src0_stride, src1,
+                                      src1_stride, &mask0, round_offset, shift,
+                                      clip_low, clip_high, mask_max);
+
+    dst += dst_stride * 4;
+    src0 += src0_stride * 4;
+    src1 += src1_stride * 4;
+    mask += mask_stride * 8;
+  } while (h -= 4);
+}
+
+static INLINE void highbd_blend_a64_d16_mask_w8_avx2(
+    uint16_t *dst, int dst_stride, const CONV_BUF_TYPE *src0, int src0_stride,
+    const CONV_BUF_TYPE *src1, int src1_stride, const __m256i *mask0a,
+    const __m256i *mask0b, const __m256i *round_offset, int shift,
+    const __m256i *clip_low, const __m256i *clip_high,
+    const __m256i *mask_max) {
+  // Load 8x u16 pixels from each of 4 rows from each source
+  const __m256i s0a =
+      yy_loadu2_128(src0 + 0 * src0_stride, src0 + 1 * src0_stride);
+  const __m256i s0b =
+      yy_loadu2_128(src0 + 2 * src0_stride, src0 + 3 * src0_stride);
+  const __m256i s1a =
+      yy_loadu2_128(src1 + 0 * src1_stride, src1 + 1 * src1_stride);
+  const __m256i s1b =
+      yy_loadu2_128(src1 + 2 * src1_stride, src1 + 3 * src1_stride);
+
+  // Generate inverse masks
+  const __m256i mask1a = _mm256_sub_epi16(*mask_max, *mask0a);
+  const __m256i mask1b = _mm256_sub_epi16(*mask_max, *mask0b);
+
+  // Multiply sources by respective masks
+  const __m256i mul0a_highs = _mm256_mulhi_epu16(*mask0a, s0a);
+  const __m256i mul0a_lows = _mm256_mullo_epi16(*mask0a, s0a);
+  const __m256i mul0ah = _mm256_unpackhi_epi16(mul0a_lows, mul0a_highs);
+  const __m256i mul0al = _mm256_unpacklo_epi16(mul0a_lows, mul0a_highs);
+  // Note that AVX2 unpack orders 64-bit words as [3 1] [2 0] to keep within
+  // lanes Later, packs does the same again which cancels this out with no need
+  // for a permute.  The intermediate values being reordered makes no difference
+
+  const __m256i mul1a_highs = _mm256_mulhi_epu16(mask1a, s1a);
+  const __m256i mul1a_lows = _mm256_mullo_epi16(mask1a, s1a);
+  const __m256i mul1ah = _mm256_unpackhi_epi16(mul1a_lows, mul1a_highs);
+  const __m256i mul1al = _mm256_unpacklo_epi16(mul1a_lows, mul1a_highs);
+
+  const __m256i sumah = _mm256_add_epi32(mul0ah, mul1ah);
+  const __m256i sumal = _mm256_add_epi32(mul0al, mul1al);
+
+  const __m256i mul0b_highs = _mm256_mulhi_epu16(*mask0b, s0b);
+  const __m256i mul0b_lows = _mm256_mullo_epi16(*mask0b, s0b);
+  const __m256i mul0bh = _mm256_unpackhi_epi16(mul0b_lows, mul0b_highs);
+  const __m256i mul0bl = _mm256_unpacklo_epi16(mul0b_lows, mul0b_highs);
+
+  const __m256i mul1b_highs = _mm256_mulhi_epu16(mask1b, s1b);
+  const __m256i mul1b_lows = _mm256_mullo_epi16(mask1b, s1b);
+  const __m256i mul1bh = _mm256_unpackhi_epi16(mul1b_lows, mul1b_highs);
+  const __m256i mul1bl = _mm256_unpacklo_epi16(mul1b_lows, mul1b_highs);
+
+  const __m256i sumbh = _mm256_add_epi32(mul0bh, mul1bh);
+  const __m256i sumbl = _mm256_add_epi32(mul0bl, mul1bl);
+
+  // Divide down each result, with rounding
+  const __m256i roundah =
+      _mm256_srai_epi32(_mm256_sub_epi32(sumah, *round_offset), shift);
+  const __m256i roundal =
+      _mm256_srai_epi32(_mm256_sub_epi32(sumal, *round_offset), shift);
+  const __m256i roundbh =
+      _mm256_srai_epi32(_mm256_sub_epi32(sumbh, *round_offset), shift);
+  const __m256i roundbl =
+      _mm256_srai_epi32(_mm256_sub_epi32(sumbl, *round_offset), shift);
+
+  // Pack each i32 down to an i16 with saturation, then clip to valid range
+  const __m256i packa = _mm256_packs_epi32(roundal, roundah);
+  const __m256i clipa =
+      _mm256_min_epi16(_mm256_max_epi16(packa, *clip_low), *clip_high);
+  const __m256i packb = _mm256_packs_epi32(roundbl, roundbh);
+  const __m256i clipb =
+      _mm256_min_epi16(_mm256_max_epi16(packb, *clip_low), *clip_high);
+
+  // Store 8x u16 pixels to each of 4 rows in the destination
+  yy_storeu2_128(dst + 0 * dst_stride, dst + 1 * dst_stride, clipa);
+  yy_storeu2_128(dst + 2 * dst_stride, dst + 3 * dst_stride, clipb);
+}
+
+static INLINE void highbd_blend_a64_d16_mask_subw0_subh0_w8_avx2(
+    uint16_t *dst, int dst_stride, const CONV_BUF_TYPE *src0, int src0_stride,
+    const CONV_BUF_TYPE *src1, int src1_stride, const uint8_t *mask,
+    int mask_stride, int h, const __m256i *round_offset, int shift,
+    const __m256i *clip_low, const __m256i *clip_high,
+    const __m256i *mask_max) {
+  do {
+    // Load 8x u8 pixels from each of 4 rows in the mask
+    const __m128i mask0a8 =
+        _mm_set_epi64x(*(int64_t *)mask, *(uint64_t *)(mask + mask_stride));
+    const __m128i mask0b8 =
+        _mm_set_epi64x(*(int64_t *)(mask + 2 * mask_stride),
+                       *(int64_t *)(mask + 3 * mask_stride));
+    const __m256i mask0a = _mm256_cvtepu8_epi16(mask0a8);
+    const __m256i mask0b = _mm256_cvtepu8_epi16(mask0b8);
+
+    highbd_blend_a64_d16_mask_w8_avx2(
+        dst, dst_stride, src0, src0_stride, src1, src1_stride, &mask0a, &mask0b,
+        round_offset, shift, clip_low, clip_high, mask_max);
+
+    dst += dst_stride * 4;
+    src0 += src0_stride * 4;
+    src1 += src1_stride * 4;
+    mask += mask_stride * 4;
+  } while (h -= 4);
+}
+
+static INLINE void highbd_blend_a64_d16_mask_subw1_subh1_w8_avx2(
+    uint16_t *dst, int dst_stride, const CONV_BUF_TYPE *src0, int src0_stride,
+    const CONV_BUF_TYPE *src1, int src1_stride, const uint8_t *mask,
+    int mask_stride, int h, const __m256i *round_offset, int shift,
+    const __m256i *clip_low, const __m256i *clip_high,
+    const __m256i *mask_max) {
+  const __m256i one_b = _mm256_set1_epi8(1);
+  const __m256i two_w = _mm256_set1_epi16(2);
+  do {
+    // Load 16x u8 pixels from each of 8 rows in the mask,
+    // (saturating) add together rows then use madd to add adjacent pixels
+    // Finally, divide each value by 4 (with rounding)
+    const __m256i m02 =
+        yy_loadu2_128(mask + 0 * mask_stride, mask + 2 * mask_stride);
+    const __m256i m13 =
+        yy_loadu2_128(mask + 1 * mask_stride, mask + 3 * mask_stride);
+    const __m256i m0123 =
+        _mm256_maddubs_epi16(_mm256_adds_epu8(m02, m13), one_b);
+    const __m256i mask_0a =
+        _mm256_srli_epi16(_mm256_add_epi16(m0123, two_w), 2);
+    const __m256i m46 =
+        yy_loadu2_128(mask + 4 * mask_stride, mask + 6 * mask_stride);
+    const __m256i m57 =
+        yy_loadu2_128(mask + 5 * mask_stride, mask + 7 * mask_stride);
+    const __m256i m4567 =
+        _mm256_maddubs_epi16(_mm256_adds_epu8(m46, m57), one_b);
+    const __m256i mask_0b =
+        _mm256_srli_epi16(_mm256_add_epi16(m4567, two_w), 2);
+
+    highbd_blend_a64_d16_mask_w8_avx2(
+        dst, dst_stride, src0, src0_stride, src1, src1_stride, &mask_0a,
+        &mask_0b, round_offset, shift, clip_low, clip_high, mask_max);
+
+    dst += dst_stride * 4;
+    src0 += src0_stride * 4;
+    src1 += src1_stride * 4;
+    mask += mask_stride * 8;
+  } while (h -= 4);
+}
+
+static INLINE void highbd_blend_a64_d16_mask_w16_avx2(
+    uint16_t *dst, int dst_stride, const CONV_BUF_TYPE *src0, int src0_stride,
+    const CONV_BUF_TYPE *src1, int src1_stride, const __m256i *mask0a,
+    const __m256i *mask0b, const __m256i *round_offset, int shift,
+    const __m256i *clip_low, const __m256i *clip_high,
+    const __m256i *mask_max) {
+  // Load 16x pixels from each of 2 rows from each source
+  const __m256i s0a = yy_loadu_256(src0);
+  const __m256i s0b = yy_loadu_256(src0 + src0_stride);
+  const __m256i s1a = yy_loadu_256(src1);
+  const __m256i s1b = yy_loadu_256(src1 + src1_stride);
+
+  // Calculate inverse masks
+  const __m256i mask1a = _mm256_sub_epi16(*mask_max, *mask0a);
+  const __m256i mask1b = _mm256_sub_epi16(*mask_max, *mask0b);
+
+  // Multiply each source by appropriate mask
+  const __m256i mul0a_highs = _mm256_mulhi_epu16(*mask0a, s0a);
+  const __m256i mul0a_lows = _mm256_mullo_epi16(*mask0a, s0a);
+  const __m256i mul0ah = _mm256_unpackhi_epi16(mul0a_lows, mul0a_highs);
+  const __m256i mul0al = _mm256_unpacklo_epi16(mul0a_lows, mul0a_highs);
+  // Note that AVX2 unpack orders 64-bit words as [3 1] [2 0] to keep within
+  // lanes Later, packs does the same again which cancels this out with no need
+  // for a permute.  The intermediate values being reordered makes no difference
+
+  const __m256i mul1a_highs = _mm256_mulhi_epu16(mask1a, s1a);
+  const __m256i mul1a_lows = _mm256_mullo_epi16(mask1a, s1a);
+  const __m256i mul1ah = _mm256_unpackhi_epi16(mul1a_lows, mul1a_highs);
+  const __m256i mul1al = _mm256_unpacklo_epi16(mul1a_lows, mul1a_highs);
+
+  const __m256i mulah = _mm256_add_epi32(mul0ah, mul1ah);
+  const __m256i mulal = _mm256_add_epi32(mul0al, mul1al);
+
+  const __m256i mul0b_highs = _mm256_mulhi_epu16(*mask0b, s0b);
+  const __m256i mul0b_lows = _mm256_mullo_epi16(*mask0b, s0b);
+  const __m256i mul0bh = _mm256_unpackhi_epi16(mul0b_lows, mul0b_highs);
+  const __m256i mul0bl = _mm256_unpacklo_epi16(mul0b_lows, mul0b_highs);
+
+  const __m256i mul1b_highs = _mm256_mulhi_epu16(mask1b, s1b);
+  const __m256i mul1b_lows = _mm256_mullo_epi16(mask1b, s1b);
+  const __m256i mul1bh = _mm256_unpackhi_epi16(mul1b_lows, mul1b_highs);
+  const __m256i mul1bl = _mm256_unpacklo_epi16(mul1b_lows, mul1b_highs);
+
+  const __m256i mulbh = _mm256_add_epi32(mul0bh, mul1bh);
+  const __m256i mulbl = _mm256_add_epi32(mul0bl, mul1bl);
+
+  const __m256i resah =
+      _mm256_srai_epi32(_mm256_sub_epi32(mulah, *round_offset), shift);
+  const __m256i resal =
+      _mm256_srai_epi32(_mm256_sub_epi32(mulal, *round_offset), shift);
+  const __m256i resbh =
+      _mm256_srai_epi32(_mm256_sub_epi32(mulbh, *round_offset), shift);
+  const __m256i resbl =
+      _mm256_srai_epi32(_mm256_sub_epi32(mulbl, *round_offset), shift);
+
+  // Signed saturating pack from i32 to i16:
+  const __m256i packa = _mm256_packs_epi32(resal, resah);
+  const __m256i packb = _mm256_packs_epi32(resbl, resbh);
+
+  // Clip the values to the valid range
+  const __m256i clipa =
+      _mm256_min_epi16(_mm256_max_epi16(packa, *clip_low), *clip_high);
+  const __m256i clipb =
+      _mm256_min_epi16(_mm256_max_epi16(packb, *clip_low), *clip_high);
+
+  // Store 16 pixels
+  yy_storeu_256(dst, clipa);
+  yy_storeu_256(dst + dst_stride, clipb);
+}
+
+static INLINE void highbd_blend_a64_d16_mask_subw0_subh0_w16_avx2(
+    uint16_t *dst, int dst_stride, const CONV_BUF_TYPE *src0, int src0_stride,
+    const CONV_BUF_TYPE *src1, int src1_stride, const uint8_t *mask,
+    int mask_stride, int h, int w, const __m256i *round_offset, int shift,
+    const __m256i *clip_low, const __m256i *clip_high,
+    const __m256i *mask_max) {
+  for (int i = 0; i < h; i += 2) {
+    for (int j = 0; j < w; j += 16) {
+      // Load 16x u8 alpha-mask values from each of two rows and pad to u16
+      const __m128i masks_a8 = xx_loadu_128(mask + j);
+      const __m128i masks_b8 = xx_loadu_128(mask + mask_stride + j);
+      const __m256i mask0a = _mm256_cvtepu8_epi16(masks_a8);
+      const __m256i mask0b = _mm256_cvtepu8_epi16(masks_b8);
+
+      highbd_blend_a64_d16_mask_w16_avx2(
+          dst + j, dst_stride, src0 + j, src0_stride, src1 + j, src1_stride,
+          &mask0a, &mask0b, round_offset, shift, clip_low, clip_high, mask_max);
+    }
+    dst += dst_stride * 2;
+    src0 += src0_stride * 2;
+    src1 += src1_stride * 2;
+    mask += mask_stride * 2;
+  }
+}
+
+static INLINE void highbd_blend_a64_d16_mask_subw1_subh1_w16_avx2(
+    uint16_t *dst, int dst_stride, const CONV_BUF_TYPE *src0, int src0_stride,
+    const CONV_BUF_TYPE *src1, int src1_stride, const uint8_t *mask,
+    int mask_stride, int h, int w, const __m256i *round_offset, int shift,
+    const __m256i *clip_low, const __m256i *clip_high,
+    const __m256i *mask_max) {
+  const __m256i one_b = _mm256_set1_epi8(1);
+  const __m256i two_w = _mm256_set1_epi16(2);
+  for (int i = 0; i < h; i += 2) {
+    for (int j = 0; j < w; j += 16) {
+      // Load 32x u8 alpha-mask values from each of four rows
+      // (saturating) add pairs of rows, then use madd to add adjacent values
+      // Finally, divide down each result with rounding
+      const __m256i m0 = yy_loadu_256(mask + 0 * mask_stride + 2 * j);
+      const __m256i m1 = yy_loadu_256(mask + 1 * mask_stride + 2 * j);
+      const __m256i m2 = yy_loadu_256(mask + 2 * mask_stride + 2 * j);
+      const __m256i m3 = yy_loadu_256(mask + 3 * mask_stride + 2 * j);
+
+      const __m256i m01_8 = _mm256_adds_epu8(m0, m1);
+      const __m256i m23_8 = _mm256_adds_epu8(m2, m3);
+
+      const __m256i m01 = _mm256_maddubs_epi16(m01_8, one_b);
+      const __m256i m23 = _mm256_maddubs_epi16(m23_8, one_b);
+
+      const __m256i mask0a = _mm256_srli_epi16(_mm256_add_epi16(m01, two_w), 2);
+      const __m256i mask0b = _mm256_srli_epi16(_mm256_add_epi16(m23, two_w), 2);
+
+      highbd_blend_a64_d16_mask_w16_avx2(
+          dst + j, dst_stride, src0 + j, src0_stride, src1 + j, src1_stride,
+          &mask0a, &mask0b, round_offset, shift, clip_low, clip_high, mask_max);
+    }
+    dst += dst_stride * 2;
+    src0 += src0_stride * 2;
+    src1 += src1_stride * 2;
+    mask += mask_stride * 4;
+  }
+}
+
+void aom_highbd_blend_a64_d16_mask_avx2(
+    uint8_t *dst8, uint32_t dst_stride, const CONV_BUF_TYPE *src0,
+    uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh,
+    ConvolveParams *conv_params, const int bd) {
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  const int round_bits =
+      2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+  const int32_t round_offset =
+      ((1 << (round_bits + bd)) + (1 << (round_bits + bd - 1)) -
+       (1 << (round_bits - 1)))
+      << AOM_BLEND_A64_ROUND_BITS;
+  const __m256i v_round_offset = _mm256_set1_epi32(round_offset);
+  const int shift = round_bits + AOM_BLEND_A64_ROUND_BITS;
+
+  const __m256i clip_low = _mm256_setzero_si256();
+  const __m256i clip_high = _mm256_set1_epi16((1 << bd) - 1);
+  const __m256i mask_max = _mm256_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  assert(IMPLIES((void *)src0 == dst, src0_stride == dst_stride));
+  assert(IMPLIES((void *)src1 == dst, src1_stride == dst_stride));
+
+  assert(h >= 4);
+  assert(w >= 4);
+  assert(IS_POWER_OF_TWO(h));
+  assert(IS_POWER_OF_TWO(w));
+
+  if (subw == 0 && subh == 0) {
+    switch (w) {
+      case 4:
+        highbd_blend_a64_d16_mask_subw0_subh0_w4_avx2(
+            dst, dst_stride, src0, src0_stride, src1, src1_stride, mask,
+            mask_stride, h, &v_round_offset, shift, &clip_low, &clip_high,
+            &mask_max);
+        break;
+      case 8:
+        highbd_blend_a64_d16_mask_subw0_subh0_w8_avx2(
+            dst, dst_stride, src0, src0_stride, src1, src1_stride, mask,
+            mask_stride, h, &v_round_offset, shift, &clip_low, &clip_high,
+            &mask_max);
+        break;
+      default:  // >= 16
+        highbd_blend_a64_d16_mask_subw0_subh0_w16_avx2(
+            dst, dst_stride, src0, src0_stride, src1, src1_stride, mask,
+            mask_stride, h, w, &v_round_offset, shift, &clip_low, &clip_high,
+            &mask_max);
+        break;
+    }
+
+  } else if (subw == 1 && subh == 1) {
+    switch (w) {
+      case 4:
+        highbd_blend_a64_d16_mask_subw1_subh1_w4_avx2(
+            dst, dst_stride, src0, src0_stride, src1, src1_stride, mask,
+            mask_stride, h, &v_round_offset, shift, &clip_low, &clip_high,
+            &mask_max);
+        break;
+      case 8:
+        highbd_blend_a64_d16_mask_subw1_subh1_w8_avx2(
+            dst, dst_stride, src0, src0_stride, src1, src1_stride, mask,
+            mask_stride, h, &v_round_offset, shift, &clip_low, &clip_high,
+            &mask_max);
+        break;
+      default:  // >= 16
+        highbd_blend_a64_d16_mask_subw1_subh1_w16_avx2(
+            dst, dst_stride, src0, src0_stride, src1, src1_stride, mask,
+            mask_stride, h, w, &v_round_offset, shift, &clip_low, &clip_high,
+            &mask_max);
+        break;
+    }
+  } else {
+    // Sub-sampling in only one axis doesn't seem to happen very much, so fall
+    // back to the vanilla C implementation instead of having all the optimised
+    // code for these.
+    aom_highbd_blend_a64_d16_mask_c(dst8, dst_stride, src0, src0_stride, src1,
+                                    src1_stride, mask, mask_stride, w, h, subw,
+                                    subh, conv_params, bd);
+  }
+}
+#endif  // CONFIG_AV1_HIGHBITDEPTH
diff --git a/third_party/aom/aom_dsp/x86/blend_a64_mask_sse4.c b/third_party/aom/aom_dsp/x86/blend_a64_mask_sse4.c
new file mode 100644
index 0000000000..58a7345ec2
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/blend_a64_mask_sse4.c
@@ -0,0 +1,1560 @@
+/*
+ * Copyright (c) 2016, 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 <smmintrin.h>  // SSE4.1
+
+#include <assert.h>
+
+#include "aom/aom_integer.h"
+#include "aom_ports/mem.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/blend.h"
+
+#include "aom_dsp/x86/synonyms.h"
+#include "aom_dsp/x86/blend_sse4.h"
+#include "aom_dsp/x86/blend_mask_sse4.h"
+
+#include "config/aom_dsp_rtcd.h"
+
+//////////////////////////////////////////////////////////////////////////////
+// No sub-sampling
+//////////////////////////////////////////////////////////////////////////////
+
+static void blend_a64_mask_w4_sse4_1(uint8_t *dst, uint32_t dst_stride,
+                                     const uint8_t *src0, uint32_t src0_stride,
+                                     const uint8_t *src1, uint32_t src1_stride,
+                                     const uint8_t *mask, uint32_t mask_stride,
+                                     int w, int h) {
+  (void)w;
+  const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA);
+  const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS));
+  do {
+    const __m128i v_m0_b = xx_loadl_32(mask);
+    const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b);
+    const __m128i v_res_b = blend_4_u8(src0, src1, &v_m0_b, &v_m1_b, &_r);
+    xx_storel_32(dst, v_res_b);
+
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += mask_stride;
+  } while (--h);
+}
+
+static void blend_a64_mask_w8_sse4_1(uint8_t *dst, uint32_t dst_stride,
+                                     const uint8_t *src0, uint32_t src0_stride,
+                                     const uint8_t *src1, uint32_t src1_stride,
+                                     const uint8_t *mask, uint32_t mask_stride,
+                                     int w, int h) {
+  (void)w;
+  const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA);
+  const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS));
+  do {
+    const __m128i v_m0_b = xx_loadl_64(mask);
+    const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b);
+    const __m128i v_res_b = blend_8_u8(src0, src1, &v_m0_b, &v_m1_b, &_r);
+    xx_storel_64(dst, v_res_b);
+
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += mask_stride;
+  } while (--h);
+}
+
+static void blend_a64_mask_w16n_sse4_1(
+    uint8_t *dst, uint32_t dst_stride, const uint8_t *src0,
+    uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int w, int h) {
+  const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA);
+  const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS));
+
+  do {
+    int c;
+    for (c = 0; c < w; c += 16) {
+      const __m128i v_m0_b = xx_loadu_128(mask + c);
+      const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b);
+
+      const __m128i v_res_b =
+          blend_16_u8(src0 + c, src1 + c, &v_m0_b, &v_m1_b, &_r);
+
+      xx_storeu_128(dst + c, v_res_b);
+    }
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += mask_stride;
+  } while (--h);
+}
+
+//////////////////////////////////////////////////////////////////////////////
+// Horizontal sub-sampling
+//////////////////////////////////////////////////////////////////////////////
+
+static void blend_a64_mask_sx_w4_sse4_1(
+    uint8_t *dst, uint32_t dst_stride, const uint8_t *src0,
+    uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int w, int h) {
+  (void)w;
+
+  const __m128i v_shuffle_b = xx_loadu_128(g_blend_a64_mask_shuffle);
+  const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA);
+  const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS));
+  do {
+    const __m128i v_r_b = xx_loadl_64(mask);
+    const __m128i v_r0_s_b = _mm_shuffle_epi8(v_r_b, v_shuffle_b);
+    const __m128i v_r_lo_b = _mm_unpacklo_epi64(v_r0_s_b, v_r0_s_b);
+    const __m128i v_r_hi_b = _mm_unpackhi_epi64(v_r0_s_b, v_r0_s_b);
+    const __m128i v_m0_b = _mm_avg_epu8(v_r_lo_b, v_r_hi_b);
+    const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b);
+
+    const __m128i v_res_b = blend_4_u8(src0, src1, &v_m0_b, &v_m1_b, &_r);
+    xx_storel_32(dst, v_res_b);
+
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += mask_stride;
+  } while (--h);
+}
+
+static void blend_a64_mask_sx_w8_sse4_1(
+    uint8_t *dst, uint32_t dst_stride, const uint8_t *src0,
+    uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int w, int h) {
+  (void)w;
+
+  const __m128i v_shuffle_b = xx_loadu_128(g_blend_a64_mask_shuffle);
+  const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA);
+  const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS));
+  do {
+    const __m128i v_r_b = xx_loadu_128(mask);
+    const __m128i v_r0_s_b = _mm_shuffle_epi8(v_r_b, v_shuffle_b);
+    const __m128i v_r_lo_b = _mm_unpacklo_epi64(v_r0_s_b, v_r0_s_b);
+    const __m128i v_r_hi_b = _mm_unpackhi_epi64(v_r0_s_b, v_r0_s_b);
+    const __m128i v_m0_b = _mm_avg_epu8(v_r_lo_b, v_r_hi_b);
+    const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b);
+
+    const __m128i v_res_b = blend_8_u8(src0, src1, &v_m0_b, &v_m1_b, &_r);
+
+    xx_storel_64(dst, v_res_b);
+
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += mask_stride;
+  } while (--h);
+}
+
+static void blend_a64_mask_sx_w16n_sse4_1(
+    uint8_t *dst, uint32_t dst_stride, const uint8_t *src0,
+    uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int w, int h) {
+  const __m128i v_shuffle_b = xx_loadu_128(g_blend_a64_mask_shuffle);
+  const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA);
+  const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS));
+
+  do {
+    int c;
+    for (c = 0; c < w; c += 16) {
+      const __m128i v_r0_b = xx_loadu_128(mask + 2 * c);
+      const __m128i v_r1_b = xx_loadu_128(mask + 2 * c + 16);
+      const __m128i v_r0_s_b = _mm_shuffle_epi8(v_r0_b, v_shuffle_b);
+      const __m128i v_r1_s_b = _mm_shuffle_epi8(v_r1_b, v_shuffle_b);
+      const __m128i v_r_lo_b = _mm_unpacklo_epi64(v_r0_s_b, v_r1_s_b);
+      const __m128i v_r_hi_b = _mm_unpackhi_epi64(v_r0_s_b, v_r1_s_b);
+      const __m128i v_m0_b = _mm_avg_epu8(v_r_lo_b, v_r_hi_b);
+      const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b);
+
+      const __m128i v_res_b =
+          blend_16_u8(src0 + c, src1 + c, &v_m0_b, &v_m1_b, &_r);
+
+      xx_storeu_128(dst + c, v_res_b);
+    }
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += mask_stride;
+  } while (--h);
+}
+
+//////////////////////////////////////////////////////////////////////////////
+// Vertical sub-sampling
+//////////////////////////////////////////////////////////////////////////////
+
+static void blend_a64_mask_sy_w4_sse4_1(
+    uint8_t *dst, uint32_t dst_stride, const uint8_t *src0,
+    uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int w, int h) {
+  (void)w;
+
+  const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA);
+  const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS));
+
+  do {
+    const __m128i v_ra_b = xx_loadl_32(mask);
+    const __m128i v_rb_b = xx_loadl_32(mask + mask_stride);
+    const __m128i v_m0_b = _mm_avg_epu8(v_ra_b, v_rb_b);
+    const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b);
+
+    const __m128i v_res_b = blend_4_u8(src0, src1, &v_m0_b, &v_m1_b, &_r);
+
+    xx_storel_32(dst, v_res_b);
+
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += 2 * mask_stride;
+  } while (--h);
+}
+
+static void blend_a64_mask_sy_w8_sse4_1(
+    uint8_t *dst, uint32_t dst_stride, const uint8_t *src0,
+    uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int w, int h) {
+  (void)w;
+
+  const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA);
+  const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS));
+  do {
+    const __m128i v_ra_b = xx_loadl_64(mask);
+    const __m128i v_rb_b = xx_loadl_64(mask + mask_stride);
+    const __m128i v_m0_b = _mm_avg_epu8(v_ra_b, v_rb_b);
+    const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b);
+    const __m128i v_res_b = blend_8_u8(src0, src1, &v_m0_b, &v_m1_b, &_r);
+
+    xx_storel_64(dst, v_res_b);
+
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += 2 * mask_stride;
+  } while (--h);
+}
+
+static void blend_a64_mask_sy_w16n_sse4_1(
+    uint8_t *dst, uint32_t dst_stride, const uint8_t *src0,
+    uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int w, int h) {
+  const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA);
+  const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS));
+  do {
+    int c;
+    for (c = 0; c < w; c += 16) {
+      const __m128i v_ra_b = xx_loadu_128(mask + c);
+      const __m128i v_rb_b = xx_loadu_128(mask + c + mask_stride);
+      const __m128i v_m0_b = _mm_avg_epu8(v_ra_b, v_rb_b);
+      const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b);
+
+      const __m128i v_res_b =
+          blend_16_u8(src0 + c, src1 + c, &v_m0_b, &v_m1_b, &_r);
+
+      xx_storeu_128(dst + c, v_res_b);
+    }
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += 2 * mask_stride;
+  } while (--h);
+}
+
+//////////////////////////////////////////////////////////////////////////////
+// Horizontal and Vertical sub-sampling
+//////////////////////////////////////////////////////////////////////////////
+
+static void blend_a64_mask_sx_sy_w4_sse4_1(
+    uint8_t *dst, uint32_t dst_stride, const uint8_t *src0,
+    uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int w, int h) {
+  const __m128i v_shuffle_b = xx_loadu_128(g_blend_a64_mask_shuffle);
+  const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA);
+  const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS));
+  (void)w;
+
+  do {
+    const __m128i v_ra_b = xx_loadl_64(mask);
+    const __m128i v_rb_b = xx_loadl_64(mask + mask_stride);
+    const __m128i v_rvs_b = _mm_add_epi8(v_ra_b, v_rb_b);
+    const __m128i v_r_s_b = _mm_shuffle_epi8(v_rvs_b, v_shuffle_b);
+    const __m128i v_r0_s_w = _mm_cvtepu8_epi16(v_r_s_b);
+    const __m128i v_r1_s_w = _mm_cvtepu8_epi16(_mm_srli_si128(v_r_s_b, 8));
+    const __m128i v_rs_w = _mm_add_epi16(v_r0_s_w, v_r1_s_w);
+    const __m128i v_m0_w = xx_roundn_epu16(v_rs_w, 2);
+    const __m128i v_m0_b = _mm_packus_epi16(v_m0_w, v_m0_w);
+    const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b);
+
+    const __m128i v_res_b = blend_4_u8(src0, src1, &v_m0_b, &v_m1_b, &_r);
+
+    xx_storel_32(dst, v_res_b);
+
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += 2 * mask_stride;
+  } while (--h);
+}
+
+static void blend_a64_mask_sx_sy_w8_sse4_1(
+    uint8_t *dst, uint32_t dst_stride, const uint8_t *src0,
+    uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int w, int h) {
+  const __m128i v_shuffle_b = xx_loadu_128(g_blend_a64_mask_shuffle);
+  const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA);
+  const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS));
+  (void)w;
+
+  do {
+    const __m128i v_ra_b = xx_loadu_128(mask);
+    const __m128i v_rb_b = xx_loadu_128(mask + mask_stride);
+
+    const __m128i v_rvs_b = _mm_add_epi8(v_ra_b, v_rb_b);
+    const __m128i v_r_s_b = _mm_shuffle_epi8(v_rvs_b, v_shuffle_b);
+    const __m128i v_r0_s_w = _mm_cvtepu8_epi16(v_r_s_b);
+    const __m128i v_r1_s_w = _mm_cvtepu8_epi16(_mm_srli_si128(v_r_s_b, 8));
+    const __m128i v_rs_w = _mm_add_epi16(v_r0_s_w, v_r1_s_w);
+    const __m128i v_m0_w = xx_roundn_epu16(v_rs_w, 2);
+    const __m128i v_m0_b = _mm_packus_epi16(v_m0_w, v_m0_w);
+    const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b);
+
+    const __m128i v_res_b = blend_8_u8(src0, src1, &v_m0_b, &v_m1_b, &_r);
+
+    xx_storel_64(dst, v_res_b);
+
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += 2 * mask_stride;
+  } while (--h);
+}
+
+static void blend_a64_mask_sx_sy_w16n_sse4_1(
+    uint8_t *dst, uint32_t dst_stride, const uint8_t *src0,
+    uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int w, int h) {
+  const __m128i v_zmask_b =
+      _mm_set_epi8(0, -1, 0, -1, 0, -1, 0, -1, 0, -1, 0, -1, 0, -1, 0, -1);
+  const __m128i v_maxval_b = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA);
+  const __m128i _r = _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS));
+  do {
+    int c;
+    for (c = 0; c < w; c += 16) {
+      const __m128i v_ral_b = xx_loadu_128(mask + 2 * c);
+      const __m128i v_rah_b = xx_loadu_128(mask + 2 * c + 16);
+      const __m128i v_rbl_b = xx_loadu_128(mask + mask_stride + 2 * c);
+      const __m128i v_rbh_b = xx_loadu_128(mask + mask_stride + 2 * c + 16);
+      const __m128i v_rvsl_b = _mm_add_epi8(v_ral_b, v_rbl_b);
+      const __m128i v_rvsh_b = _mm_add_epi8(v_rah_b, v_rbh_b);
+      const __m128i v_rvsal_w = _mm_and_si128(v_rvsl_b, v_zmask_b);
+      const __m128i v_rvsah_w = _mm_and_si128(v_rvsh_b, v_zmask_b);
+      const __m128i v_rvsbl_w =
+          _mm_and_si128(_mm_srli_si128(v_rvsl_b, 1), v_zmask_b);
+      const __m128i v_rvsbh_w =
+          _mm_and_si128(_mm_srli_si128(v_rvsh_b, 1), v_zmask_b);
+      const __m128i v_rsl_w = _mm_add_epi16(v_rvsal_w, v_rvsbl_w);
+      const __m128i v_rsh_w = _mm_add_epi16(v_rvsah_w, v_rvsbh_w);
+
+      const __m128i v_m0l_w = xx_roundn_epu16(v_rsl_w, 2);
+      const __m128i v_m0h_w = xx_roundn_epu16(v_rsh_w, 2);
+      const __m128i v_m0_b = _mm_packus_epi16(v_m0l_w, v_m0h_w);
+      const __m128i v_m1_b = _mm_sub_epi8(v_maxval_b, v_m0_b);
+
+      const __m128i v_res_b =
+          blend_16_u8(src0 + c, src1 + c, &v_m0_b, &v_m1_b, &_r);
+
+      xx_storeu_128(dst + c, v_res_b);
+    }
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += 2 * mask_stride;
+  } while (--h);
+}
+
+//////////////////////////////////////////////////////////////////////////////
+// Dispatch
+//////////////////////////////////////////////////////////////////////////////
+
+void aom_blend_a64_mask_sse4_1(uint8_t *dst, uint32_t dst_stride,
+                               const uint8_t *src0, uint32_t src0_stride,
+                               const uint8_t *src1, uint32_t src1_stride,
+                               const uint8_t *mask, uint32_t mask_stride, int w,
+                               int h, int subw, int subh) {
+  typedef void (*blend_fn)(
+      uint8_t * dst, uint32_t dst_stride, const uint8_t *src0,
+      uint32_t src0_stride, const uint8_t *src1, uint32_t src1_stride,
+      const uint8_t *mask, uint32_t mask_stride, int w, int h);
+
+  // Dimensions are: width_index X subx X suby
+  static const blend_fn blend[3][2][2] = {
+    { // w % 16 == 0
+      { blend_a64_mask_w16n_sse4_1, blend_a64_mask_sy_w16n_sse4_1 },
+      { blend_a64_mask_sx_w16n_sse4_1, blend_a64_mask_sx_sy_w16n_sse4_1 } },
+    { // w == 4
+      { blend_a64_mask_w4_sse4_1, blend_a64_mask_sy_w4_sse4_1 },
+      { blend_a64_mask_sx_w4_sse4_1, blend_a64_mask_sx_sy_w4_sse4_1 } },
+    { // w == 8
+      { blend_a64_mask_w8_sse4_1, blend_a64_mask_sy_w8_sse4_1 },
+      { blend_a64_mask_sx_w8_sse4_1, blend_a64_mask_sx_sy_w8_sse4_1 } }
+  };
+
+  assert(IMPLIES(src0 == dst, src0_stride == dst_stride));
+  assert(IMPLIES(src1 == dst, src1_stride == dst_stride));
+
+  assert(h >= 1);
+  assert(w >= 1);
+  assert(IS_POWER_OF_TWO(h));
+  assert(IS_POWER_OF_TWO(w));
+
+  if (UNLIKELY((h | w) & 3)) {  // if (w <= 2 || h <= 2)
+    aom_blend_a64_mask_c(dst, dst_stride, src0, src0_stride, src1, src1_stride,
+                         mask, mask_stride, w, h, subw, subh);
+  } else {
+    blend[(w >> 2) & 3][subw != 0][subh != 0](dst, dst_stride, src0,
+                                              src0_stride, src1, src1_stride,
+                                              mask, mask_stride, w, h);
+  }
+}
+
+#if CONFIG_AV1_HIGHBITDEPTH
+//////////////////////////////////////////////////////////////////////////////
+// No sub-sampling
+//////////////////////////////////////////////////////////////////////////////
+
+static INLINE void blend_a64_mask_bn_w4_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, blend_unit_fn blend) {
+  const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  do {
+    const __m128i v_m0_b = xx_loadl_32(mask);
+    const __m128i v_m0_w = _mm_cvtepu8_epi16(v_m0_b);
+    const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w);
+
+    const __m128i v_res_w = blend(src0, src1, v_m0_w, v_m1_w);
+
+    xx_storel_64(dst, v_res_w);
+
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += mask_stride;
+  } while (--h);
+}
+
+static void blend_a64_mask_b10_w4_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int w, int h) {
+  (void)w;
+  blend_a64_mask_bn_w4_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                              src1_stride, mask, mask_stride, h, blend_4_b10);
+}
+
+static void blend_a64_mask_b12_w4_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int w, int h) {
+  (void)w;
+  blend_a64_mask_bn_w4_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                              src1_stride, mask, mask_stride, h, blend_4_b12);
+}
+
+static INLINE void blend_a64_mask_bn_w8n_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int w, int h,
+    blend_unit_fn blend) {
+  const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  do {
+    int c;
+    for (c = 0; c < w; c += 8) {
+      const __m128i v_m0_b = xx_loadl_64(mask + c);
+      const __m128i v_m0_w = _mm_cvtepu8_epi16(v_m0_b);
+      const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w);
+
+      const __m128i v_res_w = blend(src0 + c, src1 + c, v_m0_w, v_m1_w);
+
+      xx_storeu_128(dst + c, v_res_w);
+    }
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += mask_stride;
+  } while (--h);
+}
+
+static void blend_a64_mask_b10_w8n_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int w, int h) {
+  blend_a64_mask_bn_w8n_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                               src1_stride, mask, mask_stride, w, h,
+                               blend_8_b10);
+}
+
+static void blend_a64_mask_b12_w8n_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int w, int h) {
+  blend_a64_mask_bn_w8n_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                               src1_stride, mask, mask_stride, w, h,
+                               blend_8_b12);
+}
+
+//////////////////////////////////////////////////////////////////////////////
+// Horizontal sub-sampling
+//////////////////////////////////////////////////////////////////////////////
+
+static INLINE void blend_a64_mask_bn_sx_w4_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, blend_unit_fn blend) {
+  const __m128i v_zmask_b =
+      _mm_set_epi8(0, -1, 0, -1, 0, -1, 0, -1, 0, -1, 0, -1, 0, -1, 0, -1);
+  const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  do {
+    const __m128i v_r_b = xx_loadl_64(mask);
+    const __m128i v_a_b = _mm_avg_epu8(v_r_b, _mm_srli_si128(v_r_b, 1));
+
+    const __m128i v_m0_w = _mm_and_si128(v_a_b, v_zmask_b);
+    const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w);
+
+    const __m128i v_res_w = blend(src0, src1, v_m0_w, v_m1_w);
+
+    xx_storel_64(dst, v_res_w);
+
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += mask_stride;
+  } while (--h);
+}
+
+static void blend_a64_mask_b10_sx_w4_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int w, int h) {
+  (void)w;
+  blend_a64_mask_bn_sx_w4_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                                 src1_stride, mask, mask_stride, h,
+                                 blend_4_b10);
+}
+
+static void blend_a64_mask_b12_sx_w4_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int w, int h) {
+  (void)w;
+  blend_a64_mask_bn_sx_w4_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                                 src1_stride, mask, mask_stride, h,
+                                 blend_4_b12);
+}
+
+static INLINE void blend_a64_mask_bn_sx_w8n_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int w, int h,
+    blend_unit_fn blend) {
+  const __m128i v_zmask_b =
+      _mm_set_epi8(0, -1, 0, -1, 0, -1, 0, -1, 0, -1, 0, -1, 0, -1, 0, -1);
+  const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  do {
+    int c;
+    for (c = 0; c < w; c += 8) {
+      const __m128i v_r_b = xx_loadu_128(mask + 2 * c);
+      const __m128i v_a_b = _mm_avg_epu8(v_r_b, _mm_srli_si128(v_r_b, 1));
+
+      const __m128i v_m0_w = _mm_and_si128(v_a_b, v_zmask_b);
+      const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w);
+
+      const __m128i v_res_w = blend(src0 + c, src1 + c, v_m0_w, v_m1_w);
+
+      xx_storeu_128(dst + c, v_res_w);
+    }
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += mask_stride;
+  } while (--h);
+}
+
+static void blend_a64_mask_b10_sx_w8n_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int w, int h) {
+  blend_a64_mask_bn_sx_w8n_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                                  src1_stride, mask, mask_stride, w, h,
+                                  blend_8_b10);
+}
+
+static void blend_a64_mask_b12_sx_w8n_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int w, int h) {
+  blend_a64_mask_bn_sx_w8n_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                                  src1_stride, mask, mask_stride, w, h,
+                                  blend_8_b12);
+}
+
+//////////////////////////////////////////////////////////////////////////////
+// Vertical sub-sampling
+//////////////////////////////////////////////////////////////////////////////
+
+static INLINE void blend_a64_mask_bn_sy_w4_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, blend_unit_fn blend) {
+  const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  do {
+    const __m128i v_ra_b = xx_loadl_32(mask);
+    const __m128i v_rb_b = xx_loadl_32(mask + mask_stride);
+    const __m128i v_a_b = _mm_avg_epu8(v_ra_b, v_rb_b);
+
+    const __m128i v_m0_w = _mm_cvtepu8_epi16(v_a_b);
+    const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w);
+
+    const __m128i v_res_w = blend(src0, src1, v_m0_w, v_m1_w);
+
+    xx_storel_64(dst, v_res_w);
+
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += 2 * mask_stride;
+  } while (--h);
+}
+
+static void blend_a64_mask_b10_sy_w4_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int w, int h) {
+  (void)w;
+  blend_a64_mask_bn_sy_w4_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                                 src1_stride, mask, mask_stride, h,
+                                 blend_4_b10);
+}
+
+static void blend_a64_mask_b12_sy_w4_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int w, int h) {
+  (void)w;
+  blend_a64_mask_bn_sy_w4_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                                 src1_stride, mask, mask_stride, h,
+                                 blend_4_b12);
+}
+
+static INLINE void blend_a64_mask_bn_sy_w8n_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int w, int h,
+    blend_unit_fn blend) {
+  const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  do {
+    int c;
+    for (c = 0; c < w; c += 8) {
+      const __m128i v_ra_b = xx_loadl_64(mask + c);
+      const __m128i v_rb_b = xx_loadl_64(mask + c + mask_stride);
+      const __m128i v_a_b = _mm_avg_epu8(v_ra_b, v_rb_b);
+
+      const __m128i v_m0_w = _mm_cvtepu8_epi16(v_a_b);
+      const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w);
+
+      const __m128i v_res_w = blend(src0 + c, src1 + c, v_m0_w, v_m1_w);
+
+      xx_storeu_128(dst + c, v_res_w);
+    }
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += 2 * mask_stride;
+  } while (--h);
+}
+
+static void blend_a64_mask_b10_sy_w8n_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int w, int h) {
+  blend_a64_mask_bn_sy_w8n_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                                  src1_stride, mask, mask_stride, w, h,
+                                  blend_8_b10);
+}
+
+static void blend_a64_mask_b12_sy_w8n_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int w, int h) {
+  blend_a64_mask_bn_sy_w8n_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                                  src1_stride, mask, mask_stride, w, h,
+                                  blend_8_b12);
+}
+
+//////////////////////////////////////////////////////////////////////////////
+// Horizontal and Vertical sub-sampling
+//////////////////////////////////////////////////////////////////////////////
+
+static INLINE void blend_a64_mask_bn_sx_sy_w4_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, blend_unit_fn blend) {
+  const __m128i v_zmask_b =
+      _mm_set_epi8(0, -1, 0, -1, 0, -1, 0, -1, 0, -1, 0, -1, 0, -1, 0, -1);
+  const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  do {
+    const __m128i v_ra_b = xx_loadl_64(mask);
+    const __m128i v_rb_b = xx_loadl_64(mask + mask_stride);
+    const __m128i v_rvs_b = _mm_add_epi8(v_ra_b, v_rb_b);
+    const __m128i v_rvsa_w = _mm_and_si128(v_rvs_b, v_zmask_b);
+    const __m128i v_rvsb_w =
+        _mm_and_si128(_mm_srli_si128(v_rvs_b, 1), v_zmask_b);
+    const __m128i v_rs_w = _mm_add_epi16(v_rvsa_w, v_rvsb_w);
+
+    const __m128i v_m0_w = xx_roundn_epu16(v_rs_w, 2);
+    const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w);
+
+    const __m128i v_res_w = blend(src0, src1, v_m0_w, v_m1_w);
+
+    xx_storel_64(dst, v_res_w);
+
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += 2 * mask_stride;
+  } while (--h);
+}
+
+static void blend_a64_mask_b10_sx_sy_w4_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int w, int h) {
+  (void)w;
+  blend_a64_mask_bn_sx_sy_w4_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                                    src1_stride, mask, mask_stride, h,
+                                    blend_4_b10);
+}
+
+static void blend_a64_mask_b12_sx_sy_w4_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int w, int h) {
+  (void)w;
+  blend_a64_mask_bn_sx_sy_w4_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                                    src1_stride, mask, mask_stride, h,
+                                    blend_4_b12);
+}
+
+static INLINE void blend_a64_mask_bn_sx_sy_w8n_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int w, int h,
+    blend_unit_fn blend) {
+  const __m128i v_zmask_b =
+      _mm_set_epi8(0, -1, 0, -1, 0, -1, 0, -1, 0, -1, 0, -1, 0, -1, 0, -1);
+  const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  do {
+    int c;
+    for (c = 0; c < w; c += 8) {
+      const __m128i v_ra_b = xx_loadu_128(mask + 2 * c);
+      const __m128i v_rb_b = xx_loadu_128(mask + 2 * c + mask_stride);
+      const __m128i v_rvs_b = _mm_add_epi8(v_ra_b, v_rb_b);
+      const __m128i v_rvsa_w = _mm_and_si128(v_rvs_b, v_zmask_b);
+      const __m128i v_rvsb_w =
+          _mm_and_si128(_mm_srli_si128(v_rvs_b, 1), v_zmask_b);
+      const __m128i v_rs_w = _mm_add_epi16(v_rvsa_w, v_rvsb_w);
+
+      const __m128i v_m0_w = xx_roundn_epu16(v_rs_w, 2);
+      const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w);
+
+      const __m128i v_res_w = blend(src0 + c, src1 + c, v_m0_w, v_m1_w);
+
+      xx_storeu_128(dst + c, v_res_w);
+    }
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += 2 * mask_stride;
+  } while (--h);
+}
+
+static void blend_a64_mask_b10_sx_sy_w8n_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int w, int h) {
+  blend_a64_mask_bn_sx_sy_w8n_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                                     src1_stride, mask, mask_stride, w, h,
+                                     blend_8_b10);
+}
+
+static void blend_a64_mask_b12_sx_sy_w8n_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int w, int h) {
+  blend_a64_mask_bn_sx_sy_w8n_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                                     src1_stride, mask, mask_stride, w, h,
+                                     blend_8_b12);
+}
+
+//////////////////////////////////////////////////////////////////////////////
+// Dispatch
+//////////////////////////////////////////////////////////////////////////////
+void aom_highbd_blend_a64_mask_sse4_1(uint8_t *dst_8, uint32_t dst_stride,
+                                      const uint8_t *src0_8,
+                                      uint32_t src0_stride,
+                                      const uint8_t *src1_8,
+                                      uint32_t src1_stride, const uint8_t *mask,
+                                      uint32_t mask_stride, int w, int h,
+                                      int subw, int subh, int bd) {
+  typedef void (*blend_fn)(
+      uint16_t * dst, uint32_t dst_stride, const uint16_t *src0,
+      uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+      const uint8_t *mask, uint32_t mask_stride, int w, int h);
+
+  // Dimensions are: bd_index X width_index X subw X subh
+  static const blend_fn blend[2][2][2][2] = {
+    {   // bd == 8 or 10
+      { // w % 8 == 0
+        { blend_a64_mask_b10_w8n_sse4_1, blend_a64_mask_b10_sy_w8n_sse4_1 },
+        { blend_a64_mask_b10_sx_w8n_sse4_1,
+          blend_a64_mask_b10_sx_sy_w8n_sse4_1 } },
+      { // w == 4
+        { blend_a64_mask_b10_w4_sse4_1, blend_a64_mask_b10_sy_w4_sse4_1 },
+        { blend_a64_mask_b10_sx_w4_sse4_1,
+          blend_a64_mask_b10_sx_sy_w4_sse4_1 } } },
+    {   // bd == 12
+      { // w % 8 == 0
+        { blend_a64_mask_b12_w8n_sse4_1, blend_a64_mask_b12_sy_w8n_sse4_1 },
+        { blend_a64_mask_b12_sx_w8n_sse4_1,
+          blend_a64_mask_b12_sx_sy_w8n_sse4_1 } },
+      { // w == 4
+        { blend_a64_mask_b12_w4_sse4_1, blend_a64_mask_b12_sy_w4_sse4_1 },
+        { blend_a64_mask_b12_sx_w4_sse4_1,
+          blend_a64_mask_b12_sx_sy_w4_sse4_1 } } }
+  };
+
+  assert(IMPLIES(src0_8 == dst_8, src0_stride == dst_stride));
+  assert(IMPLIES(src1_8 == dst_8, src1_stride == dst_stride));
+
+  assert(h >= 1);
+  assert(w >= 1);
+  assert(IS_POWER_OF_TWO(h));
+  assert(IS_POWER_OF_TWO(w));
+
+  assert(bd == 8 || bd == 10 || bd == 12);
+  if (UNLIKELY((h | w) & 3)) {  // if (w <= 2 || h <= 2)
+    aom_highbd_blend_a64_mask_c(dst_8, dst_stride, src0_8, src0_stride, src1_8,
+                                src1_stride, mask, mask_stride, w, h, subw,
+                                subh, bd);
+  } else {
+    uint16_t *const dst = CONVERT_TO_SHORTPTR(dst_8);
+    const uint16_t *const src0 = CONVERT_TO_SHORTPTR(src0_8);
+    const uint16_t *const src1 = CONVERT_TO_SHORTPTR(src1_8);
+
+    blend[bd == 12][(w >> 2) & 1][subw != 0][subh != 0](
+        dst, dst_stride, src0, src0_stride, src1, src1_stride, mask,
+        mask_stride, w, h);
+  }
+}
+#endif  // CONFIG_AV1_HIGHBITDEPTH
+
+static INLINE void blend_a64_d16_mask_w16_sse41(
+    uint8_t *dst, const CONV_BUF_TYPE *src0, const CONV_BUF_TYPE *src1,
+    const __m128i *m0, const __m128i *m1, const __m128i *v_round_offset,
+    const __m128i *v_maxval, int shift) {
+  const __m128i max_minus_m0 = _mm_sub_epi16(*v_maxval, *m0);
+  const __m128i max_minus_m1 = _mm_sub_epi16(*v_maxval, *m1);
+  const __m128i s0_0 = xx_loadu_128(src0);
+  const __m128i s0_1 = xx_loadu_128(src0 + 8);
+  const __m128i s1_0 = xx_loadu_128(src1);
+  const __m128i s1_1 = xx_loadu_128(src1 + 8);
+  __m128i res0_lo = _mm_madd_epi16(_mm_unpacklo_epi16(s0_0, s1_0),
+                                   _mm_unpacklo_epi16(*m0, max_minus_m0));
+  __m128i res0_hi = _mm_madd_epi16(_mm_unpackhi_epi16(s0_0, s1_0),
+                                   _mm_unpackhi_epi16(*m0, max_minus_m0));
+  __m128i res1_lo = _mm_madd_epi16(_mm_unpacklo_epi16(s0_1, s1_1),
+                                   _mm_unpacklo_epi16(*m1, max_minus_m1));
+  __m128i res1_hi = _mm_madd_epi16(_mm_unpackhi_epi16(s0_1, s1_1),
+                                   _mm_unpackhi_epi16(*m1, max_minus_m1));
+  res0_lo = _mm_srai_epi32(_mm_sub_epi32(res0_lo, *v_round_offset), shift);
+  res0_hi = _mm_srai_epi32(_mm_sub_epi32(res0_hi, *v_round_offset), shift);
+  res1_lo = _mm_srai_epi32(_mm_sub_epi32(res1_lo, *v_round_offset), shift);
+  res1_hi = _mm_srai_epi32(_mm_sub_epi32(res1_hi, *v_round_offset), shift);
+  const __m128i res0 = _mm_packs_epi32(res0_lo, res0_hi);
+  const __m128i res1 = _mm_packs_epi32(res1_lo, res1_hi);
+  const __m128i res = _mm_packus_epi16(res0, res1);
+
+  _mm_storeu_si128((__m128i *)(dst), res);
+}
+
+static INLINE void lowbd_blend_a64_d16_mask_subw0_subh0_w16_sse4_1(
+    uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0,
+    uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, int w,
+    const __m128i *round_offset, int shift) {
+  const __m128i v_maxval = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+  for (int i = 0; i < h; ++i) {
+    for (int j = 0; j < w; j += 16) {
+      const __m128i m = xx_loadu_128(mask + j);
+      const __m128i m0 = _mm_cvtepu8_epi16(m);
+      const __m128i m1 = _mm_cvtepu8_epi16(_mm_srli_si128(m, 8));
+
+      blend_a64_d16_mask_w16_sse41(dst + j, src0 + j, src1 + j, &m0, &m1,
+                                   round_offset, &v_maxval, shift);
+    }
+    mask += mask_stride;
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+  }
+}
+
+static INLINE void lowbd_blend_a64_d16_mask_subw1_subh1_w16_sse4_1(
+    uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0,
+    uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, int w,
+    const __m128i *round_offset, int shift) {
+  const __m128i v_maxval = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+  const __m128i one_b = _mm_set1_epi8(1);
+  const __m128i two_w = _mm_set1_epi16(2);
+  for (int i = 0; i < h; ++i) {
+    for (int j = 0; j < w; j += 16) {
+      const __m128i m_i00 = xx_loadu_128(mask + 2 * j);
+      const __m128i m_i01 = xx_loadu_128(mask + 2 * j + 16);
+      const __m128i m_i10 = xx_loadu_128(mask + mask_stride + 2 * j);
+      const __m128i m_i11 = xx_loadu_128(mask + mask_stride + 2 * j + 16);
+
+      const __m128i m0_ac = _mm_adds_epu8(m_i00, m_i10);
+      const __m128i m1_ac = _mm_adds_epu8(m_i01, m_i11);
+      const __m128i m0_acbd = _mm_maddubs_epi16(m0_ac, one_b);
+      const __m128i m1_acbd = _mm_maddubs_epi16(m1_ac, one_b);
+      const __m128i m0 = _mm_srli_epi16(_mm_add_epi16(m0_acbd, two_w), 2);
+      const __m128i m1 = _mm_srli_epi16(_mm_add_epi16(m1_acbd, two_w), 2);
+
+      blend_a64_d16_mask_w16_sse41(dst + j, src0 + j, src1 + j, &m0, &m1,
+                                   round_offset, &v_maxval, shift);
+    }
+    mask += mask_stride << 1;
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+  }
+}
+
+static INLINE void lowbd_blend_a64_d16_mask_subw1_subh0_w16_sse4_1(
+    uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0,
+    uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, int w,
+    const __m128i *round_offset, int shift) {
+  const __m128i v_maxval = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+  const __m128i one_b = _mm_set1_epi8(1);
+  const __m128i zeros = _mm_setzero_si128();
+  for (int i = 0; i < h; ++i) {
+    for (int j = 0; j < w; j += 16) {
+      const __m128i m_i00 = xx_loadu_128(mask + 2 * j);
+      const __m128i m_i01 = xx_loadu_128(mask + 2 * j + 16);
+      const __m128i m0_ac = _mm_maddubs_epi16(m_i00, one_b);
+      const __m128i m1_ac = _mm_maddubs_epi16(m_i01, one_b);
+      const __m128i m0 = _mm_avg_epu16(m0_ac, zeros);
+      const __m128i m1 = _mm_avg_epu16(m1_ac, zeros);
+
+      blend_a64_d16_mask_w16_sse41(dst + j, src0 + j, src1 + j, &m0, &m1,
+                                   round_offset, &v_maxval, shift);
+    }
+    mask += mask_stride;
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+  }
+}
+
+static INLINE void lowbd_blend_a64_d16_mask_subw0_subh1_w16_sse4_1(
+    uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0,
+    uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, int w,
+    const __m128i *round_offset, int shift) {
+  const __m128i v_maxval = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+  const __m128i zeros = _mm_setzero_si128();
+  for (int i = 0; i < h; ++i) {
+    for (int j = 0; j < w; j += 16) {
+      const __m128i m_i00 = xx_loadu_128(mask + j);
+      const __m128i m_i10 = xx_loadu_128(mask + mask_stride + j);
+
+      const __m128i m_ac = _mm_avg_epu8(_mm_adds_epu8(m_i00, m_i10), zeros);
+      const __m128i m0 = _mm_cvtepu8_epi16(m_ac);
+      const __m128i m1 = _mm_cvtepu8_epi16(_mm_srli_si128(m_ac, 8));
+
+      blend_a64_d16_mask_w16_sse41(dst + j, src0 + j, src1 + j, &m0, &m1,
+                                   round_offset, &v_maxval, shift);
+    }
+    mask += mask_stride << 1;
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+  }
+}
+
+void aom_lowbd_blend_a64_d16_mask_sse4_1(
+    uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0,
+    uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh,
+    ConvolveParams *conv_params) {
+  const int bd = 8;
+  const int round_bits =
+      2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+
+  const int round_offset =
+      ((1 << (round_bits + bd)) + (1 << (round_bits + bd - 1)) -
+       (1 << (round_bits - 1)))
+      << AOM_BLEND_A64_ROUND_BITS;
+
+  const int shift = round_bits + AOM_BLEND_A64_ROUND_BITS;
+  assert(IMPLIES((void *)src0 == dst, src0_stride == dst_stride));
+  assert(IMPLIES((void *)src1 == dst, src1_stride == dst_stride));
+
+  assert(h >= 4);
+  assert(w >= 4);
+  assert(IS_POWER_OF_TWO(h));
+  assert(IS_POWER_OF_TWO(w));
+
+  const __m128i v_round_offset = _mm_set1_epi32(round_offset);
+
+  if (subw == 0 && subh == 0) {
+    switch (w) {
+      case 4:
+        aom_lowbd_blend_a64_d16_mask_subw0_subh0_w4_sse4_1(
+            dst, dst_stride, src0, src0_stride, src1, src1_stride, mask,
+            mask_stride, h, &v_round_offset, shift);
+        break;
+      case 8:
+        aom_lowbd_blend_a64_d16_mask_subw0_subh0_w8_sse4_1(
+            dst, dst_stride, src0, src0_stride, src1, src1_stride, mask,
+            mask_stride, h, &v_round_offset, shift);
+        break;
+      default:
+        lowbd_blend_a64_d16_mask_subw0_subh0_w16_sse4_1(
+            dst, dst_stride, src0, src0_stride, src1, src1_stride, mask,
+            mask_stride, h, w, &v_round_offset, shift);
+        break;
+    }
+
+  } else if (subw == 1 && subh == 1) {
+    switch (w) {
+      case 4:
+        aom_lowbd_blend_a64_d16_mask_subw1_subh1_w4_sse4_1(
+            dst, dst_stride, src0, src0_stride, src1, src1_stride, mask,
+            mask_stride, h, &v_round_offset, shift);
+        break;
+      case 8:
+        aom_lowbd_blend_a64_d16_mask_subw1_subh1_w8_sse4_1(
+            dst, dst_stride, src0, src0_stride, src1, src1_stride, mask,
+            mask_stride, h, &v_round_offset, shift);
+        break;
+      default:
+        lowbd_blend_a64_d16_mask_subw1_subh1_w16_sse4_1(
+            dst, dst_stride, src0, src0_stride, src1, src1_stride, mask,
+            mask_stride, h, w, &v_round_offset, shift);
+        break;
+    }
+  } else if (subw == 1 && subh == 0) {
+    switch (w) {
+      case 4:
+        aom_lowbd_blend_a64_d16_mask_subw1_subh0_w4_sse4_1(
+            dst, dst_stride, src0, src0_stride, src1, src1_stride, mask,
+            mask_stride, h, &v_round_offset, shift);
+        break;
+      case 8:
+        aom_lowbd_blend_a64_d16_mask_subw1_subh0_w8_sse4_1(
+            dst, dst_stride, src0, src0_stride, src1, src1_stride, mask,
+            mask_stride, h, &v_round_offset, shift);
+        break;
+      default:
+        lowbd_blend_a64_d16_mask_subw1_subh0_w16_sse4_1(
+            dst, dst_stride, src0, src0_stride, src1, src1_stride, mask,
+            mask_stride, h, w, &v_round_offset, shift);
+        break;
+    }
+  } else {
+    switch (w) {
+      case 4:
+        aom_lowbd_blend_a64_d16_mask_subw0_subh1_w4_sse4_1(
+            dst, dst_stride, src0, src0_stride, src1, src1_stride, mask,
+            mask_stride, h, &v_round_offset, shift);
+        break;
+      case 8:
+        aom_lowbd_blend_a64_d16_mask_subw0_subh1_w8_sse4_1(
+            dst, dst_stride, src0, src0_stride, src1, src1_stride, mask,
+            mask_stride, h, &v_round_offset, shift);
+        break;
+      default:
+        lowbd_blend_a64_d16_mask_subw0_subh1_w16_sse4_1(
+            dst, dst_stride, src0, src0_stride, src1, src1_stride, mask,
+            mask_stride, h, w, &v_round_offset, shift);
+        break;
+    }
+  }
+}
+
+//////////////////////////////////////////////////////////////////////////////
+// aom_highbd_blend_a64_d16_mask_sse4_1()
+//////////////////////////////////////////////////////////////////////////////
+#if CONFIG_AV1_HIGHBITDEPTH
+static INLINE void highbd_blend_a64_d16_mask_w4_sse4_1(
+    uint16_t *dst, int dst_stride, const CONV_BUF_TYPE *src0, int src0_stride,
+    const CONV_BUF_TYPE *src1, int src1_stride, const __m128i *mask0a,
+    const __m128i *mask0b, const __m128i *round_offset, int shift,
+    const __m128i *clip_low, const __m128i *clip_high,
+    const __m128i *mask_max) {
+  // Load 4 pixels from each of 4 rows from each source
+  const __m128i s0a =
+      _mm_set_epi64x(*(int64_t *)src0, *(int64_t *)(src0 + src0_stride));
+  const __m128i s0b = _mm_set_epi64x(*(int64_t *)(src0 + 2 * src0_stride),
+                                     *(int64_t *)(src0 + 3 * src0_stride));
+  const __m128i s1a =
+      _mm_set_epi64x(*(int64_t *)(src1), *(int64_t *)(src1 + src1_stride));
+  const __m128i s1b = _mm_set_epi64x(*(int64_t *)(src1 + 2 * src1_stride),
+                                     *(int64_t *)(src1 + 3 * src1_stride));
+
+  // Generate the inverse masks
+  const __m128i mask1a = _mm_sub_epi16(*mask_max, *mask0a);
+  const __m128i mask1b = _mm_sub_epi16(*mask_max, *mask0b);
+
+  // Multiply each mask by the respective source
+  const __m128i mul0a_highs = _mm_mulhi_epu16(*mask0a, s0a);
+  const __m128i mul0a_lows = _mm_mullo_epi16(*mask0a, s0a);
+  const __m128i mul0ah = _mm_unpackhi_epi16(mul0a_lows, mul0a_highs);
+  const __m128i mul0al = _mm_unpacklo_epi16(mul0a_lows, mul0a_highs);
+  const __m128i mul1a_highs = _mm_mulhi_epu16(mask1a, s1a);
+  const __m128i mul1a_lows = _mm_mullo_epi16(mask1a, s1a);
+  const __m128i mul1ah = _mm_unpackhi_epi16(mul1a_lows, mul1a_highs);
+  const __m128i mul1al = _mm_unpacklo_epi16(mul1a_lows, mul1a_highs);
+
+  const __m128i mul0b_highs = _mm_mulhi_epu16(*mask0b, s0b);
+  const __m128i mul0b_lows = _mm_mullo_epi16(*mask0b, s0b);
+  const __m128i mul0bh = _mm_unpackhi_epi16(mul0b_lows, mul0b_highs);
+  const __m128i mul0bl = _mm_unpacklo_epi16(mul0b_lows, mul0b_highs);
+  const __m128i mul1b_highs = _mm_mulhi_epu16(mask1b, s1b);
+  const __m128i mul1b_lows = _mm_mullo_epi16(mask1b, s1b);
+  const __m128i mul1bh = _mm_unpackhi_epi16(mul1b_lows, mul1b_highs);
+  const __m128i mul1bl = _mm_unpacklo_epi16(mul1b_lows, mul1b_highs);
+
+  const __m128i sumah = _mm_add_epi32(mul0ah, mul1ah);
+  const __m128i sumal = _mm_add_epi32(mul0al, mul1al);
+  const __m128i sumbh = _mm_add_epi32(mul0bh, mul1bh);
+  const __m128i sumbl = _mm_add_epi32(mul0bl, mul1bl);
+
+  const __m128i roundah =
+      _mm_srai_epi32(_mm_sub_epi32(sumah, *round_offset), shift);
+  const __m128i roundbh =
+      _mm_srai_epi32(_mm_sub_epi32(sumbh, *round_offset), shift);
+  const __m128i roundal =
+      _mm_srai_epi32(_mm_sub_epi32(sumal, *round_offset), shift);
+  const __m128i roundbl =
+      _mm_srai_epi32(_mm_sub_epi32(sumbl, *round_offset), shift);
+
+  const __m128i packa = _mm_packs_epi32(roundal, roundah);
+  const __m128i packb = _mm_packs_epi32(roundbl, roundbh);
+
+  const __m128i clipa =
+      _mm_min_epi16(_mm_max_epi16(packa, *clip_low), *clip_high);
+  const __m128i clipb =
+      _mm_min_epi16(_mm_max_epi16(packb, *clip_low), *clip_high);
+
+  xx_storel_64(dst, _mm_srli_si128(clipa, 8));
+  xx_storel_64(dst + dst_stride, clipa);
+  xx_storel_64(dst + 2 * dst_stride, _mm_srli_si128(clipb, 8));
+  xx_storel_64(dst + 3 * dst_stride, clipb);
+}
+
+static INLINE void highbd_blend_a64_d16_mask_subw0_subh0_w4_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0,
+    uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h,
+    const __m128i *round_offset, int shift, const __m128i *clip_low,
+    const __m128i *clip_high, const __m128i *mask_max) {
+  do {
+    const __m128i mask0a8 =
+        _mm_set_epi32(0, 0, *(int32_t *)mask, *(int32_t *)(mask + mask_stride));
+    const __m128i mask0b8 =
+        _mm_set_epi32(0, 0, *(int32_t *)(mask + 2 * mask_stride),
+                      *(int32_t *)(mask + 3 * mask_stride));
+    const __m128i mask0a = _mm_cvtepu8_epi16(mask0a8);
+    const __m128i mask0b = _mm_cvtepu8_epi16(mask0b8);
+
+    highbd_blend_a64_d16_mask_w4_sse4_1(
+        dst, dst_stride, src0, src0_stride, src1, src1_stride, &mask0a, &mask0b,
+        round_offset, shift, clip_low, clip_high, mask_max);
+
+    dst += dst_stride * 4;
+    src0 += src0_stride * 4;
+    src1 += src1_stride * 4;
+    mask += mask_stride * 4;
+  } while (h -= 4);
+}
+
+static INLINE void highbd_blend_a64_d16_mask_subw1_subh1_w4_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0,
+    uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h,
+    const __m128i *round_offset, int shift, const __m128i *clip_low,
+    const __m128i *clip_high, const __m128i *mask_max) {
+  const __m128i one_b = _mm_set1_epi8(1);
+  const __m128i two_w = _mm_set1_epi16(2);
+  do {
+    // Load 8 pixels from each of 8 rows of mask,
+    // (saturating) add together rows then use madd to add adjacent pixels
+    // Finally, divide each value by 4 (with rounding)
+    const __m128i m02 = _mm_set_epi64x(*(int64_t *)(mask),
+                                       *(int64_t *)(mask + 2 * mask_stride));
+    const __m128i m13 = _mm_set_epi64x(*(int64_t *)(mask + mask_stride),
+                                       *(int64_t *)(mask + 3 * mask_stride));
+    const __m128i m0123 = _mm_maddubs_epi16(_mm_adds_epu8(m02, m13), one_b);
+    const __m128i mask_0a = _mm_srli_epi16(_mm_add_epi16(m0123, two_w), 2);
+    const __m128i m46 = _mm_set_epi64x(*(int64_t *)(mask + 4 * mask_stride),
+                                       *(int64_t *)(mask + 6 * mask_stride));
+    const __m128i m57 = _mm_set_epi64x(*(int64_t *)(mask + 5 * mask_stride),
+                                       *(int64_t *)(mask + 7 * mask_stride));
+    const __m128i m4567 = _mm_maddubs_epi16(_mm_adds_epu8(m46, m57), one_b);
+    const __m128i mask_0b = _mm_srli_epi16(_mm_add_epi16(m4567, two_w), 2);
+
+    highbd_blend_a64_d16_mask_w4_sse4_1(
+        dst, dst_stride, src0, src0_stride, src1, src1_stride, &mask_0a,
+        &mask_0b, round_offset, shift, clip_low, clip_high, mask_max);
+
+    dst += dst_stride * 4;
+    src0 += src0_stride * 4;
+    src1 += src1_stride * 4;
+    mask += mask_stride * 8;
+  } while (h -= 4);
+}
+
+static INLINE void highbd_blend_a64_d16_mask_w8_sse4_1(
+    uint16_t *dst, int dst_stride, const CONV_BUF_TYPE *src0, int src0_stride,
+    const CONV_BUF_TYPE *src1, int src1_stride, const __m128i *mask0a,
+    const __m128i *mask0b, const __m128i *round_offset, int shift,
+    const __m128i *clip_low, const __m128i *clip_high,
+    const __m128i *max_mask) {
+  // Load 8x pixels from each of 2 rows from each source
+  const __m128i s0a = xx_loadu_128(src0);
+  const __m128i s0b = xx_loadu_128(src0 + src0_stride);
+  const __m128i s1a = xx_loadu_128(src1);
+  const __m128i s1b = xx_loadu_128(src1 + src1_stride);
+
+  // Generate inverse masks
+  const __m128i mask1a = _mm_sub_epi16(*max_mask, *mask0a);
+  const __m128i mask1b = _mm_sub_epi16(*max_mask, *mask0b);
+
+  // Multiply sources by respective masks
+  const __m128i mul0a_highs = _mm_mulhi_epu16(*mask0a, s0a);
+  const __m128i mul0a_lows = _mm_mullo_epi16(*mask0a, s0a);
+  const __m128i mul0ah = _mm_unpackhi_epi16(mul0a_lows, mul0a_highs);
+  const __m128i mul0al = _mm_unpacklo_epi16(mul0a_lows, mul0a_highs);
+
+  const __m128i mul1a_highs = _mm_mulhi_epu16(mask1a, s1a);
+  const __m128i mul1a_lows = _mm_mullo_epi16(mask1a, s1a);
+  const __m128i mul1ah = _mm_unpackhi_epi16(mul1a_lows, mul1a_highs);
+  const __m128i mul1al = _mm_unpacklo_epi16(mul1a_lows, mul1a_highs);
+
+  const __m128i sumah = _mm_add_epi32(mul0ah, mul1ah);
+  const __m128i sumal = _mm_add_epi32(mul0al, mul1al);
+
+  const __m128i mul0b_highs = _mm_mulhi_epu16(*mask0b, s0b);
+  const __m128i mul0b_lows = _mm_mullo_epi16(*mask0b, s0b);
+  const __m128i mul0bh = _mm_unpackhi_epi16(mul0b_lows, mul0b_highs);
+  const __m128i mul0bl = _mm_unpacklo_epi16(mul0b_lows, mul0b_highs);
+
+  const __m128i mul1b_highs = _mm_mulhi_epu16(mask1b, s1b);
+  const __m128i mul1b_lows = _mm_mullo_epi16(mask1b, s1b);
+  const __m128i mul1bh = _mm_unpackhi_epi16(mul1b_lows, mul1b_highs);
+  const __m128i mul1bl = _mm_unpacklo_epi16(mul1b_lows, mul1b_highs);
+
+  const __m128i sumbh = _mm_add_epi32(mul0bh, mul1bh);
+  const __m128i sumbl = _mm_add_epi32(mul0bl, mul1bl);
+
+  const __m128i roundah =
+      _mm_srai_epi32(_mm_sub_epi32(sumah, *round_offset), shift);
+  const __m128i roundal =
+      _mm_srai_epi32(_mm_sub_epi32(sumal, *round_offset), shift);
+  const __m128i roundbh =
+      _mm_srai_epi32(_mm_sub_epi32(sumbh, *round_offset), shift);
+  const __m128i roundbl =
+      _mm_srai_epi32(_mm_sub_epi32(sumbl, *round_offset), shift);
+
+  const __m128i packa = _mm_packs_epi32(roundal, roundah);
+  const __m128i clipa =
+      _mm_min_epi16(_mm_max_epi16(packa, *clip_low), *clip_high);
+  const __m128i packb = _mm_packs_epi32(roundbl, roundbh);
+  const __m128i clipb =
+      _mm_min_epi16(_mm_max_epi16(packb, *clip_low), *clip_high);
+
+  xx_storeu_128(dst, clipa);
+  xx_storeu_128(dst + dst_stride, clipb);
+}
+
+static INLINE void highbd_blend_a64_d16_mask_subw0_subh0_w8_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0,
+    uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h,
+    const __m128i *round_offset, int shift, const __m128i *clip_low,
+    const __m128i *clip_high, const __m128i *max_mask) {
+  do {
+    const __m128i mask0a = _mm_cvtepu8_epi16(xx_loadl_64(mask));
+    const __m128i mask0b = _mm_cvtepu8_epi16(xx_loadl_64(mask + mask_stride));
+    highbd_blend_a64_d16_mask_w8_sse4_1(
+        dst, dst_stride, src0, src0_stride, src1, src1_stride, &mask0a, &mask0b,
+        round_offset, shift, clip_low, clip_high, max_mask);
+
+    dst += dst_stride * 2;
+    src0 += src0_stride * 2;
+    src1 += src1_stride * 2;
+    mask += mask_stride * 2;
+  } while (h -= 2);
+}
+
+static INLINE void highbd_blend_a64_d16_mask_subw1_subh1_w8_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0,
+    uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h,
+    const __m128i *round_offset, int shift, const __m128i *clip_low,
+    const __m128i *clip_high, const __m128i *max_mask) {
+  const __m128i one_b = _mm_set1_epi8(1);
+  const __m128i two_w = _mm_set1_epi16(2);
+  do {
+    const __m128i mask_thisrowa = xx_loadu_128(mask);
+    const __m128i mask_nextrowa = xx_loadu_128(mask + mask_stride);
+    const __m128i mask_thisrowb = xx_loadu_128(mask + 2 * mask_stride);
+    const __m128i mask_nextrowb = xx_loadu_128(mask + 3 * mask_stride);
+    const __m128i mask_bothrowsa = _mm_adds_epu8(mask_thisrowa, mask_nextrowa);
+    const __m128i mask_bothrowsb = _mm_adds_epu8(mask_thisrowb, mask_nextrowb);
+    const __m128i mask_16a = _mm_maddubs_epi16(mask_bothrowsa, one_b);
+    const __m128i mask_16b = _mm_maddubs_epi16(mask_bothrowsb, one_b);
+    const __m128i mask_sa = _mm_srli_epi16(_mm_add_epi16(mask_16a, two_w), 2);
+    const __m128i mask_sb = _mm_srli_epi16(_mm_add_epi16(mask_16b, two_w), 2);
+
+    highbd_blend_a64_d16_mask_w8_sse4_1(
+        dst, dst_stride, src0, src0_stride, src1, src1_stride, &mask_sa,
+        &mask_sb, round_offset, shift, clip_low, clip_high, max_mask);
+
+    dst += dst_stride * 2;
+    src0 += src0_stride * 2;
+    src1 += src1_stride * 2;
+    mask += mask_stride * 4;
+  } while (h -= 2);
+}
+
+static INLINE void highbd_blend_a64_d16_mask_w16_sse4_1(
+    uint16_t *dst, const CONV_BUF_TYPE *src0, const CONV_BUF_TYPE *src1,
+    const __m128i *round_offset, int shift, const __m128i *mask0l,
+    const __m128i *mask0h, const __m128i *clip_low, const __m128i *clip_high,
+    const __m128i *mask_max) {
+  // Load 16x u16 pixels for this row from each src
+  const __m128i s0l = xx_loadu_128(src0);
+  const __m128i s0h = xx_loadu_128(src0 + 8);
+  const __m128i s1l = xx_loadu_128(src1);
+  const __m128i s1h = xx_loadu_128(src1 + 8);
+
+  // Calculate inverse masks
+  const __m128i mask1h = _mm_sub_epi16(*mask_max, *mask0h);
+  const __m128i mask1l = _mm_sub_epi16(*mask_max, *mask0l);
+
+  const __m128i mul0_highs = _mm_mulhi_epu16(*mask0h, s0h);
+  const __m128i mul0_lows = _mm_mullo_epi16(*mask0h, s0h);
+  const __m128i mul0h = _mm_unpackhi_epi16(mul0_lows, mul0_highs);
+  const __m128i mul0l = _mm_unpacklo_epi16(mul0_lows, mul0_highs);
+
+  const __m128i mul1_highs = _mm_mulhi_epu16(mask1h, s1h);
+  const __m128i mul1_lows = _mm_mullo_epi16(mask1h, s1h);
+  const __m128i mul1h = _mm_unpackhi_epi16(mul1_lows, mul1_highs);
+  const __m128i mul1l = _mm_unpacklo_epi16(mul1_lows, mul1_highs);
+
+  const __m128i mulhh = _mm_add_epi32(mul0h, mul1h);
+  const __m128i mulhl = _mm_add_epi32(mul0l, mul1l);
+
+  const __m128i mul2_highs = _mm_mulhi_epu16(*mask0l, s0l);
+  const __m128i mul2_lows = _mm_mullo_epi16(*mask0l, s0l);
+  const __m128i mul2h = _mm_unpackhi_epi16(mul2_lows, mul2_highs);
+  const __m128i mul2l = _mm_unpacklo_epi16(mul2_lows, mul2_highs);
+
+  const __m128i mul3_highs = _mm_mulhi_epu16(mask1l, s1l);
+  const __m128i mul3_lows = _mm_mullo_epi16(mask1l, s1l);
+  const __m128i mul3h = _mm_unpackhi_epi16(mul3_lows, mul3_highs);
+  const __m128i mul3l = _mm_unpacklo_epi16(mul3_lows, mul3_highs);
+
+  const __m128i mullh = _mm_add_epi32(mul2h, mul3h);
+  const __m128i mulll = _mm_add_epi32(mul2l, mul3l);
+
+  const __m128i reshh =
+      _mm_srai_epi32(_mm_sub_epi32(mulhh, *round_offset), shift);
+  const __m128i reshl =
+      _mm_srai_epi32(_mm_sub_epi32(mulhl, *round_offset), shift);
+  const __m128i reslh =
+      _mm_srai_epi32(_mm_sub_epi32(mullh, *round_offset), shift);
+  const __m128i resll =
+      _mm_srai_epi32(_mm_sub_epi32(mulll, *round_offset), shift);
+
+  // Signed saturating pack from i32 to i16:
+  const __m128i packh = _mm_packs_epi32(reshl, reshh);
+  const __m128i packl = _mm_packs_epi32(resll, reslh);
+
+  // Clip the values to the valid range
+  const __m128i cliph =
+      _mm_min_epi16(_mm_max_epi16(packh, *clip_low), *clip_high);
+  const __m128i clipl =
+      _mm_min_epi16(_mm_max_epi16(packl, *clip_low), *clip_high);
+
+  // Store 16 pixels
+  xx_storeu_128(dst, clipl);
+  xx_storeu_128(dst + 8, cliph);
+}
+
+static INLINE void highbd_blend_a64_d16_mask_subw0_subh0_w16_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0,
+    uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, int w,
+    const __m128i *round_offset, int shift, const __m128i *clip_low,
+    const __m128i *clip_high, const __m128i *mask_max) {
+  for (int i = 0; i < h; i++) {
+    for (int j = 0; j < w; j += 16) {
+      // Load 16x u8 alpha-mask values and pad to u16
+      const __m128i masks_u8 = xx_loadu_128(mask + j);
+      const __m128i mask0l = _mm_cvtepu8_epi16(masks_u8);
+      const __m128i mask0h = _mm_cvtepu8_epi16(_mm_srli_si128(masks_u8, 8));
+
+      highbd_blend_a64_d16_mask_w16_sse4_1(
+          dst + j, src0 + j, src1 + j, round_offset, shift, &mask0l, &mask0h,
+          clip_low, clip_high, mask_max);
+    }
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += mask_stride;
+  }
+}
+
+static INLINE void highbd_blend_a64_d16_mask_subw1_subh1_w16_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0,
+    uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h, int w,
+    const __m128i *round_offset, int shift, const __m128i *clip_low,
+    const __m128i *clip_high, const __m128i *mask_max) {
+  const __m128i one_b = _mm_set1_epi8(1);
+  const __m128i two_w = _mm_set1_epi16(2);
+  for (int i = 0; i < h; i++) {
+    for (int j = 0; j < w; j += 16) {
+      const __m128i m_i00 = xx_loadu_128(mask + 2 * j);
+      const __m128i m_i01 = xx_loadu_128(mask + 2 * j + 16);
+      const __m128i m_i10 = xx_loadu_128(mask + mask_stride + 2 * j);
+      const __m128i m_i11 = xx_loadu_128(mask + mask_stride + 2 * j + 16);
+
+      const __m128i m0_ac = _mm_adds_epu8(m_i00, m_i10);
+      const __m128i m1_ac = _mm_adds_epu8(m_i01, m_i11);
+      const __m128i m0_acbd = _mm_maddubs_epi16(m0_ac, one_b);
+      const __m128i m1_acbd = _mm_maddubs_epi16(m1_ac, one_b);
+      const __m128i mask_l = _mm_srli_epi16(_mm_add_epi16(m0_acbd, two_w), 2);
+      const __m128i mask_h = _mm_srli_epi16(_mm_add_epi16(m1_acbd, two_w), 2);
+
+      highbd_blend_a64_d16_mask_w16_sse4_1(
+          dst + j, src0 + j, src1 + j, round_offset, shift, &mask_l, &mask_h,
+          clip_low, clip_high, mask_max);
+    }
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += mask_stride * 2;
+  }
+}
+
+void aom_highbd_blend_a64_d16_mask_sse4_1(
+    uint8_t *dst8, uint32_t dst_stride, const CONV_BUF_TYPE *src0,
+    uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int w, int h, int subw, int subh,
+    ConvolveParams *conv_params, const int bd) {
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  const int round_bits =
+      2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+  const int32_t round_offset =
+      ((1 << (round_bits + bd)) + (1 << (round_bits + bd - 1)) -
+       (1 << (round_bits - 1)))
+      << AOM_BLEND_A64_ROUND_BITS;
+  const __m128i v_round_offset = _mm_set1_epi32(round_offset);
+  const int shift = round_bits + AOM_BLEND_A64_ROUND_BITS;
+
+  const __m128i clip_low = _mm_setzero_si128();
+  const __m128i clip_high = _mm_set1_epi16((1 << bd) - 1);
+  const __m128i mask_max = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  assert(IMPLIES((void *)src0 == dst, src0_stride == dst_stride));
+  assert(IMPLIES((void *)src1 == dst, src1_stride == dst_stride));
+
+  assert(h >= 4);
+  assert(w >= 4);
+  assert(IS_POWER_OF_TWO(h));
+  assert(IS_POWER_OF_TWO(w));
+
+  if (subw == 0 && subh == 0) {
+    switch (w) {
+      case 4:
+        highbd_blend_a64_d16_mask_subw0_subh0_w4_sse4_1(
+            dst, dst_stride, src0, src0_stride, src1, src1_stride, mask,
+            mask_stride, h, &v_round_offset, shift, &clip_low, &clip_high,
+            &mask_max);
+        break;
+      case 8:
+        highbd_blend_a64_d16_mask_subw0_subh0_w8_sse4_1(
+            dst, dst_stride, src0, src0_stride, src1, src1_stride, mask,
+            mask_stride, h, &v_round_offset, shift, &clip_low, &clip_high,
+            &mask_max);
+        break;
+      default:  // >=16
+        highbd_blend_a64_d16_mask_subw0_subh0_w16_sse4_1(
+            dst, dst_stride, src0, src0_stride, src1, src1_stride, mask,
+            mask_stride, h, w, &v_round_offset, shift, &clip_low, &clip_high,
+            &mask_max);
+        break;
+    }
+
+  } else if (subw == 1 && subh == 1) {
+    switch (w) {
+      case 4:
+        highbd_blend_a64_d16_mask_subw1_subh1_w4_sse4_1(
+            dst, dst_stride, src0, src0_stride, src1, src1_stride, mask,
+            mask_stride, h, &v_round_offset, shift, &clip_low, &clip_high,
+            &mask_max);
+        break;
+      case 8:
+        highbd_blend_a64_d16_mask_subw1_subh1_w8_sse4_1(
+            dst, dst_stride, src0, src0_stride, src1, src1_stride, mask,
+            mask_stride, h, &v_round_offset, shift, &clip_low, &clip_high,
+            &mask_max);
+        break;
+      default:  // >=16
+        highbd_blend_a64_d16_mask_subw1_subh1_w16_sse4_1(
+            dst, dst_stride, src0, src0_stride, src1, src1_stride, mask,
+            mask_stride, h, w, &v_round_offset, shift, &clip_low, &clip_high,
+            &mask_max);
+        break;
+    }
+  } else {
+    // Sub-sampling in only one axis doesn't seem to happen very much, so fall
+    // back to the vanilla C implementation instead of having all the optimised
+    // code for these.
+    aom_highbd_blend_a64_d16_mask_c(dst8, dst_stride, src0, src0_stride, src1,
+                                    src1_stride, mask, mask_stride, w, h, subw,
+                                    subh, conv_params, bd);
+  }
+}
+#endif  // CONFIG_AV1_HIGHBITDEPTH
diff --git a/third_party/aom/aom_dsp/x86/blend_a64_vmask_sse4.c b/third_party/aom/aom_dsp/x86/blend_a64_vmask_sse4.c
new file mode 100644
index 0000000000..75fb1c5a94
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/blend_a64_vmask_sse4.c
@@ -0,0 +1,285 @@
+/*
+ * Copyright (c) 2016, 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 <smmintrin.h>  // SSE4.1
+
+#include <assert.h>
+
+#include "aom/aom_integer.h"
+#include "aom_ports/mem.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/blend.h"
+
+#include "aom_dsp/x86/synonyms.h"
+#include "aom_dsp/x86/blend_sse4.h"
+
+#include "config/aom_dsp_rtcd.h"
+
+//////////////////////////////////////////////////////////////////////////////
+// Implementation - No sub-sampling
+//////////////////////////////////////////////////////////////////////////////
+
+static void blend_a64_vmask_w4_sse4_1(uint8_t *dst, uint32_t dst_stride,
+                                      const uint8_t *src0, uint32_t src0_stride,
+                                      const uint8_t *src1, uint32_t src1_stride,
+                                      const uint8_t *mask, int w, int h) {
+  const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  (void)w;
+
+  do {
+    const __m128i v_m0_w = _mm_set1_epi16(*mask);
+    const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w);
+
+    const __m128i v_res_w = blend_4(src0, src1, &v_m0_w, &v_m1_w);
+
+    const __m128i v_res_b = _mm_packus_epi16(v_res_w, v_res_w);
+
+    xx_storel_32(dst, v_res_b);
+
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += 1;
+  } while (--h);
+}
+
+static void blend_a64_vmask_w8_sse4_1(uint8_t *dst, uint32_t dst_stride,
+                                      const uint8_t *src0, uint32_t src0_stride,
+                                      const uint8_t *src1, uint32_t src1_stride,
+                                      const uint8_t *mask, int w, int h) {
+  const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  (void)w;
+
+  do {
+    const __m128i v_m0_w = _mm_set1_epi16(*mask);
+    const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w);
+
+    const __m128i v_res_w = blend_8(src0, src1, &v_m0_w, &v_m1_w);
+
+    const __m128i v_res_b = _mm_packus_epi16(v_res_w, v_res_w);
+
+    xx_storel_64(dst, v_res_b);
+
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += 1;
+  } while (--h);
+}
+
+static void blend_a64_vmask_w16n_sse4_1(uint8_t *dst, uint32_t dst_stride,
+                                        const uint8_t *src0,
+                                        uint32_t src0_stride,
+                                        const uint8_t *src1,
+                                        uint32_t src1_stride,
+                                        const uint8_t *mask, int w, int h) {
+  const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  do {
+    int c;
+    const __m128i v_m0_w = _mm_set1_epi16(*mask);
+    const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w);
+    for (c = 0; c < w; c += 16) {
+      const __m128i v_resl_w = blend_8(src0 + c, src1 + c, &v_m0_w, &v_m1_w);
+      const __m128i v_resh_w =
+          blend_8(src0 + c + 8, src1 + c + 8, &v_m0_w, &v_m1_w);
+
+      const __m128i v_res_b = _mm_packus_epi16(v_resl_w, v_resh_w);
+
+      xx_storeu_128(dst + c, v_res_b);
+    }
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += 1;
+  } while (--h);
+}
+
+//////////////////////////////////////////////////////////////////////////////
+// Dispatch
+//////////////////////////////////////////////////////////////////////////////
+
+void aom_blend_a64_vmask_sse4_1(uint8_t *dst, uint32_t dst_stride,
+                                const uint8_t *src0, uint32_t src0_stride,
+                                const uint8_t *src1, uint32_t src1_stride,
+                                const uint8_t *mask, int w, int h) {
+  typedef void (*blend_fn)(uint8_t * dst, uint32_t dst_stride,
+                           const uint8_t *src0, uint32_t src0_stride,
+                           const uint8_t *src1, uint32_t src1_stride,
+                           const uint8_t *mask, int w, int h);
+
+  // Dimension: width_index
+  static const blend_fn blend[9] = {
+    blend_a64_vmask_w16n_sse4_1,  // w % 16 == 0
+    aom_blend_a64_vmask_c,        // w == 1
+    aom_blend_a64_vmask_c,        // w == 2
+    NULL,                         // INVALID
+    blend_a64_vmask_w4_sse4_1,    // w == 4
+    NULL,                         // INVALID
+    NULL,                         // INVALID
+    NULL,                         // INVALID
+    blend_a64_vmask_w8_sse4_1,    // w == 8
+  };
+
+  assert(IMPLIES(src0 == dst, src0_stride == dst_stride));
+  assert(IMPLIES(src1 == dst, src1_stride == dst_stride));
+
+  assert(h >= 1);
+  assert(w >= 1);
+  assert(IS_POWER_OF_TWO(h));
+  assert(IS_POWER_OF_TWO(w));
+
+  blend[w & 0xf](dst, dst_stride, src0, src0_stride, src1, src1_stride, mask, w,
+                 h);
+}
+
+#if CONFIG_AV1_HIGHBITDEPTH
+//////////////////////////////////////////////////////////////////////////////
+// Implementation - No sub-sampling
+//////////////////////////////////////////////////////////////////////////////
+
+static INLINE void blend_a64_vmask_bn_w4_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, int h, blend_unit_fn blend) {
+  const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  do {
+    const __m128i v_m0_w = _mm_set1_epi16(*mask);
+    const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w);
+
+    const __m128i v_res_w = blend(src0, src1, v_m0_w, v_m1_w);
+
+    xx_storel_64(dst, v_res_w);
+
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += 1;
+  } while (--h);
+}
+
+static void blend_a64_vmask_b10_w4_sse4_1(uint16_t *dst, uint32_t dst_stride,
+                                          const uint16_t *src0,
+                                          uint32_t src0_stride,
+                                          const uint16_t *src1,
+                                          uint32_t src1_stride,
+                                          const uint8_t *mask, int w, int h) {
+  (void)w;
+  blend_a64_vmask_bn_w4_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                               src1_stride, mask, h, blend_4_b10);
+}
+
+static void blend_a64_vmask_b12_w4_sse4_1(uint16_t *dst, uint32_t dst_stride,
+                                          const uint16_t *src0,
+                                          uint32_t src0_stride,
+                                          const uint16_t *src1,
+                                          uint32_t src1_stride,
+                                          const uint8_t *mask, int w, int h) {
+  (void)w;
+  blend_a64_vmask_bn_w4_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                               src1_stride, mask, h, blend_4_b12);
+}
+
+static INLINE void blend_a64_vmask_bn_w8n_sse4_1(
+    uint16_t *dst, uint32_t dst_stride, const uint16_t *src0,
+    uint32_t src0_stride, const uint16_t *src1, uint32_t src1_stride,
+    const uint8_t *mask, int w, int h, blend_unit_fn blend) {
+  const __m128i v_maxval_w = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+
+  do {
+    int c;
+    const __m128i v_m0_w = _mm_set1_epi16(*mask);
+    const __m128i v_m1_w = _mm_sub_epi16(v_maxval_w, v_m0_w);
+    for (c = 0; c < w; c += 8) {
+      const __m128i v_res_w = blend(src0 + c, src1 + c, v_m0_w, v_m1_w);
+
+      xx_storeu_128(dst + c, v_res_w);
+    }
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+    mask += 1;
+  } while (--h);
+}
+
+static void blend_a64_vmask_b10_w8n_sse4_1(uint16_t *dst, uint32_t dst_stride,
+                                           const uint16_t *src0,
+                                           uint32_t src0_stride,
+                                           const uint16_t *src1,
+                                           uint32_t src1_stride,
+                                           const uint8_t *mask, int w, int h) {
+  blend_a64_vmask_bn_w8n_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                                src1_stride, mask, w, h, blend_8_b10);
+}
+
+static void blend_a64_vmask_b12_w8n_sse4_1(uint16_t *dst, uint32_t dst_stride,
+                                           const uint16_t *src0,
+                                           uint32_t src0_stride,
+                                           const uint16_t *src1,
+                                           uint32_t src1_stride,
+                                           const uint8_t *mask, int w, int h) {
+  blend_a64_vmask_bn_w8n_sse4_1(dst, dst_stride, src0, src0_stride, src1,
+                                src1_stride, mask, w, h, blend_8_b12);
+}
+
+//////////////////////////////////////////////////////////////////////////////
+// Dispatch
+//////////////////////////////////////////////////////////////////////////////
+
+void aom_highbd_blend_a64_vmask_sse4_1(
+    uint8_t *dst_8, uint32_t dst_stride, const uint8_t *src0_8,
+    uint32_t src0_stride, const uint8_t *src1_8, uint32_t src1_stride,
+    const uint8_t *mask, int w, int h, int bd) {
+  typedef void (*blend_fn)(uint16_t * dst, uint32_t dst_stride,
+                           const uint16_t *src0, uint32_t src0_stride,
+                           const uint16_t *src1, uint32_t src1_stride,
+                           const uint8_t *mask, int w, int h);
+
+  // Dimensions are: bd_index X width_index
+  static const blend_fn blend[2][2] = {
+    {
+        // bd == 8 or 10
+        blend_a64_vmask_b10_w8n_sse4_1,  // w % 8 == 0
+        blend_a64_vmask_b10_w4_sse4_1,   // w == 4
+    },
+    {
+        // bd == 12
+        blend_a64_vmask_b12_w8n_sse4_1,  // w % 8 == 0
+        blend_a64_vmask_b12_w4_sse4_1,   // w == 4
+    }
+  };
+
+  assert(IMPLIES(src0_8 == dst_8, src0_stride == dst_stride));
+  assert(IMPLIES(src1_8 == dst_8, src1_stride == dst_stride));
+
+  assert(h >= 1);
+  assert(w >= 1);
+  assert(IS_POWER_OF_TWO(h));
+  assert(IS_POWER_OF_TWO(w));
+
+  assert(bd == 8 || bd == 10 || bd == 12);
+
+  if (UNLIKELY((h | w) & 3)) {  // if (w <= 2 || h <= 2)
+    aom_highbd_blend_a64_vmask_c(dst_8, dst_stride, src0_8, src0_stride, src1_8,
+                                 src1_stride, mask, w, h, bd);
+  } else {
+    uint16_t *const dst = CONVERT_TO_SHORTPTR(dst_8);
+    const uint16_t *const src0 = CONVERT_TO_SHORTPTR(src0_8);
+    const uint16_t *const src1 = CONVERT_TO_SHORTPTR(src1_8);
+
+    blend[bd == 12][(w >> 2) & 1](dst, dst_stride, src0, src0_stride, src1,
+                                  src1_stride, mask, w, h);
+  }
+}
+#endif  // CONFIG_AV1_HIGHBITDEPTH
diff --git a/third_party/aom/aom_dsp/x86/blend_mask_sse4.h b/third_party/aom/aom_dsp/x86/blend_mask_sse4.h
new file mode 100644
index 0000000000..c071fdcfc4
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/blend_mask_sse4.h
@@ -0,0 +1,237 @@
+/*
+ * 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_AOM_DSP_X86_BLEND_MASK_SSE4_H_
+#define AOM_AOM_DSP_X86_BLEND_MASK_SSE4_H_
+#include <smmintrin.h>  // SSE4.1
+
+#include <assert.h>
+
+#include "aom/aom_integer.h"
+#include "aom_ports/mem.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/blend.h"
+
+#include "aom_dsp/x86/synonyms.h"
+
+#include "config/aom_dsp_rtcd.h"
+
+static INLINE void blend_a64_d16_mask_w4_sse41(
+    uint8_t *dst, const CONV_BUF_TYPE *src0, const CONV_BUF_TYPE *src1,
+    const __m128i *m, const __m128i *v_round_offset, const __m128i *v_maxval,
+    int shift) {
+  const __m128i max_minus_m = _mm_sub_epi16(*v_maxval, *m);
+  const __m128i s0 = xx_loadl_64(src0);
+  const __m128i s1 = xx_loadl_64(src1);
+  const __m128i s0_s1 = _mm_unpacklo_epi16(s0, s1);
+  const __m128i m_max_minus_m = _mm_unpacklo_epi16(*m, max_minus_m);
+  const __m128i res_a = _mm_madd_epi16(s0_s1, m_max_minus_m);
+  const __m128i res_c = _mm_sub_epi32(res_a, *v_round_offset);
+  const __m128i res_d = _mm_srai_epi32(res_c, shift);
+  const __m128i res_e = _mm_packs_epi32(res_d, res_d);
+  const __m128i res = _mm_packus_epi16(res_e, res_e);
+
+  xx_storel_32(dst, res);
+}
+
+static INLINE void blend_a64_d16_mask_w8_sse41(
+    uint8_t *dst, const CONV_BUF_TYPE *src0, const CONV_BUF_TYPE *src1,
+    const __m128i *m, const __m128i *v_round_offset, const __m128i *v_maxval,
+    int shift) {
+  const __m128i max_minus_m = _mm_sub_epi16(*v_maxval, *m);
+  const __m128i s0 = xx_loadu_128(src0);
+  const __m128i s1 = xx_loadu_128(src1);
+  __m128i res_lo = _mm_madd_epi16(_mm_unpacklo_epi16(s0, s1),
+                                  _mm_unpacklo_epi16(*m, max_minus_m));
+  __m128i res_hi = _mm_madd_epi16(_mm_unpackhi_epi16(s0, s1),
+                                  _mm_unpackhi_epi16(*m, max_minus_m));
+  res_lo = _mm_srai_epi32(_mm_sub_epi32(res_lo, *v_round_offset), shift);
+  res_hi = _mm_srai_epi32(_mm_sub_epi32(res_hi, *v_round_offset), shift);
+  const __m128i res_e = _mm_packs_epi32(res_lo, res_hi);
+  const __m128i res = _mm_packus_epi16(res_e, res_e);
+
+  _mm_storel_epi64((__m128i *)(dst), res);
+}
+
+static INLINE void aom_lowbd_blend_a64_d16_mask_subw0_subh0_w4_sse4_1(
+    uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0,
+    uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h,
+    const __m128i *round_offset, int shift) {
+  const __m128i v_maxval = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+  for (int i = 0; i < h; ++i) {
+    const __m128i m0 = xx_loadl_32(mask);
+    const __m128i m = _mm_cvtepu8_epi16(m0);
+
+    blend_a64_d16_mask_w4_sse41(dst, src0, src1, &m, round_offset, &v_maxval,
+                                shift);
+    mask += mask_stride;
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+  }
+}
+
+static INLINE void aom_lowbd_blend_a64_d16_mask_subw0_subh0_w8_sse4_1(
+    uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0,
+    uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h,
+    const __m128i *round_offset, int shift) {
+  const __m128i v_maxval = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+  for (int i = 0; i < h; ++i) {
+    const __m128i m0 = xx_loadl_64(mask);
+    const __m128i m = _mm_cvtepu8_epi16(m0);
+    blend_a64_d16_mask_w8_sse41(dst, src0, src1, &m, round_offset, &v_maxval,
+                                shift);
+    mask += mask_stride;
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+  }
+}
+
+static INLINE void aom_lowbd_blend_a64_d16_mask_subw1_subh1_w4_sse4_1(
+    uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0,
+    uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h,
+    const __m128i *round_offset, int shift) {
+  const __m128i v_maxval = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+  const __m128i one_b = _mm_set1_epi8(1);
+  const __m128i two_w = _mm_set1_epi16(2);
+  for (int i = 0; i < h; ++i) {
+    const __m128i m_i0 = xx_loadl_64(mask);
+    const __m128i m_i1 = xx_loadl_64(mask + mask_stride);
+    const __m128i m_ac = _mm_adds_epu8(m_i0, m_i1);
+    const __m128i m_acbd = _mm_maddubs_epi16(m_ac, one_b);
+    const __m128i m_acbd_2 = _mm_add_epi16(m_acbd, two_w);
+    const __m128i m = _mm_srli_epi16(m_acbd_2, 2);
+
+    blend_a64_d16_mask_w4_sse41(dst, src0, src1, &m, round_offset, &v_maxval,
+                                shift);
+    mask += mask_stride << 1;
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+  }
+}
+
+static INLINE void aom_lowbd_blend_a64_d16_mask_subw1_subh1_w8_sse4_1(
+    uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0,
+    uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h,
+    const __m128i *round_offset, int shift) {
+  const __m128i v_maxval = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+  const __m128i one_b = _mm_set1_epi8(1);
+  const __m128i two_w = _mm_set1_epi16(2);
+  for (int i = 0; i < h; ++i) {
+    const __m128i m_i0 = xx_loadu_128(mask);
+    const __m128i m_i1 = xx_loadu_128(mask + mask_stride);
+    const __m128i m_ac = _mm_adds_epu8(m_i0, m_i1);
+    const __m128i m_acbd = _mm_maddubs_epi16(m_ac, one_b);
+    const __m128i m_acbd_2 = _mm_add_epi16(m_acbd, two_w);
+    const __m128i m = _mm_srli_epi16(m_acbd_2, 2);
+
+    blend_a64_d16_mask_w8_sse41(dst, src0, src1, &m, round_offset, &v_maxval,
+                                shift);
+    mask += mask_stride << 1;
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+  }
+}
+
+static INLINE void aom_lowbd_blend_a64_d16_mask_subw1_subh0_w4_sse4_1(
+    uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0,
+    uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h,
+    const __m128i *round_offset, int shift) {
+  const __m128i v_maxval = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+  const __m128i one_b = _mm_set1_epi8(1);
+  const __m128i zeros = _mm_setzero_si128();
+  for (int i = 0; i < h; ++i) {
+    const __m128i m_i0 = xx_loadl_64(mask);
+    const __m128i m_ac = _mm_maddubs_epi16(m_i0, one_b);
+    const __m128i m = _mm_avg_epu16(m_ac, zeros);
+
+    blend_a64_d16_mask_w4_sse41(dst, src0, src1, &m, round_offset, &v_maxval,
+                                shift);
+    mask += mask_stride;
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+  }
+}
+
+static INLINE void aom_lowbd_blend_a64_d16_mask_subw1_subh0_w8_sse4_1(
+    uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0,
+    uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h,
+    const __m128i *round_offset, int shift) {
+  const __m128i v_maxval = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+  const __m128i one_b = _mm_set1_epi8(1);
+  const __m128i zeros = _mm_setzero_si128();
+  for (int i = 0; i < h; ++i) {
+    const __m128i m_i0 = xx_loadu_128(mask);
+    const __m128i m_ac = _mm_maddubs_epi16(m_i0, one_b);
+    const __m128i m = _mm_avg_epu16(m_ac, zeros);
+
+    blend_a64_d16_mask_w8_sse41(dst, src0, src1, &m, round_offset, &v_maxval,
+                                shift);
+    mask += mask_stride;
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+  }
+}
+static INLINE void aom_lowbd_blend_a64_d16_mask_subw0_subh1_w4_sse4_1(
+    uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0,
+    uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h,
+    const __m128i *round_offset, int shift) {
+  const __m128i v_maxval = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+  const __m128i zeros = _mm_setzero_si128();
+  for (int i = 0; i < h; ++i) {
+    const __m128i m_i0 = xx_loadl_64(mask);
+    const __m128i m_i1 = xx_loadl_64(mask + mask_stride);
+    const __m128i m_ac = _mm_adds_epu8(m_i0, m_i1);
+    const __m128i m = _mm_cvtepu8_epi16(_mm_avg_epu8(m_ac, zeros));
+
+    blend_a64_d16_mask_w4_sse41(dst, src0, src1, &m, round_offset, &v_maxval,
+                                shift);
+    mask += mask_stride << 1;
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+  }
+}
+
+static INLINE void aom_lowbd_blend_a64_d16_mask_subw0_subh1_w8_sse4_1(
+    uint8_t *dst, uint32_t dst_stride, const CONV_BUF_TYPE *src0,
+    uint32_t src0_stride, const CONV_BUF_TYPE *src1, uint32_t src1_stride,
+    const uint8_t *mask, uint32_t mask_stride, int h,
+    const __m128i *round_offset, int shift) {
+  const __m128i v_maxval = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+  const __m128i zeros = _mm_setzero_si128();
+  for (int i = 0; i < h; ++i) {
+    const __m128i m_i0 = xx_loadl_64(mask);
+    const __m128i m_i1 = xx_loadl_64(mask + mask_stride);
+    const __m128i m_ac = _mm_adds_epu8(m_i0, m_i1);
+    const __m128i m = _mm_cvtepu8_epi16(_mm_avg_epu8(m_ac, zeros));
+
+    blend_a64_d16_mask_w8_sse41(dst, src0, src1, &m, round_offset, &v_maxval,
+                                shift);
+    mask += mask_stride << 1;
+    dst += dst_stride;
+    src0 += src0_stride;
+    src1 += src1_stride;
+  }
+}
+#endif  // AOM_AOM_DSP_X86_BLEND_MASK_SSE4_H_
diff --git a/third_party/aom/aom_dsp/x86/blend_sse4.h b/third_party/aom/aom_dsp/x86/blend_sse4.h
new file mode 100644
index 0000000000..8d9b325101
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/blend_sse4.h
@@ -0,0 +1,191 @@
+/*
+ * Copyright (c) 2016, 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_AOM_DSP_X86_BLEND_SSE4_H_
+#define AOM_AOM_DSP_X86_BLEND_SSE4_H_
+
+#include "aom_dsp/blend.h"
+#include "aom_dsp/x86/synonyms.h"
+static const uint8_t g_blend_a64_mask_shuffle[32] = {
+  0, 2, 4, 6, 8, 10, 12, 14, 1, 3, 5, 7, 9, 11, 13, 15,
+  0, 2, 4, 6, 8, 10, 12, 14, 1, 3, 5, 7, 9, 11, 13, 15,
+};
+
+//////////////////////////////////////////////////////////////////////////////
+// Common kernels
+//////////////////////////////////////////////////////////////////////////////
+
+static INLINE __m128i blend_4(const uint8_t *src0, const uint8_t *src1,
+                              const __m128i *v_m0_w, const __m128i *v_m1_w) {
+  const __m128i v_s0_b = xx_loadl_32(src0);
+  const __m128i v_s1_b = xx_loadl_32(src1);
+  const __m128i v_s0_w = _mm_cvtepu8_epi16(v_s0_b);
+  const __m128i v_s1_w = _mm_cvtepu8_epi16(v_s1_b);
+
+  const __m128i v_p0_w = _mm_mullo_epi16(v_s0_w, *v_m0_w);
+  const __m128i v_p1_w = _mm_mullo_epi16(v_s1_w, *v_m1_w);
+  const __m128i v_sum_w = _mm_add_epi16(v_p0_w, v_p1_w);
+  const __m128i v_res_w = xx_roundn_epu16(v_sum_w, AOM_BLEND_A64_ROUND_BITS);
+
+  return v_res_w;
+}
+
+static INLINE __m128i blend_8(const uint8_t *src0, const uint8_t *src1,
+                              const __m128i *v_m0_w, const __m128i *v_m1_w) {
+  const __m128i v_s0_b = xx_loadl_64(src0);
+  const __m128i v_s1_b = xx_loadl_64(src1);
+  const __m128i v_s0_w = _mm_cvtepu8_epi16(v_s0_b);
+  const __m128i v_s1_w = _mm_cvtepu8_epi16(v_s1_b);
+
+  const __m128i v_p0_w = _mm_mullo_epi16(v_s0_w, *v_m0_w);
+  const __m128i v_p1_w = _mm_mullo_epi16(v_s1_w, *v_m1_w);
+
+  const __m128i v_sum_w = _mm_add_epi16(v_p0_w, v_p1_w);
+
+  const __m128i v_res_w = xx_roundn_epu16(v_sum_w, AOM_BLEND_A64_ROUND_BITS);
+
+  return v_res_w;
+}
+
+static INLINE __m128i blend_4_u8(const uint8_t *src0, const uint8_t *src1,
+                                 const __m128i *v_m0_b, const __m128i *v_m1_b,
+                                 const __m128i *rounding) {
+  const __m128i v_s0_b = xx_loadl_32(src0);
+  const __m128i v_s1_b = xx_loadl_32(src1);
+
+  const __m128i v_p0_w = _mm_maddubs_epi16(_mm_unpacklo_epi8(v_s0_b, v_s1_b),
+                                           _mm_unpacklo_epi8(*v_m0_b, *v_m1_b));
+
+  const __m128i v_res_w = _mm_mulhrs_epi16(v_p0_w, *rounding);
+  const __m128i v_res = _mm_packus_epi16(v_res_w, v_res_w);
+  return v_res;
+}
+
+static INLINE __m128i blend_8_u8(const uint8_t *src0, const uint8_t *src1,
+                                 const __m128i *v_m0_b, const __m128i *v_m1_b,
+                                 const __m128i *rounding) {
+  const __m128i v_s0_b = xx_loadl_64(src0);
+  const __m128i v_s1_b = xx_loadl_64(src1);
+
+  const __m128i v_p0_w = _mm_maddubs_epi16(_mm_unpacklo_epi8(v_s0_b, v_s1_b),
+                                           _mm_unpacklo_epi8(*v_m0_b, *v_m1_b));
+
+  const __m128i v_res_w = _mm_mulhrs_epi16(v_p0_w, *rounding);
+  const __m128i v_res = _mm_packus_epi16(v_res_w, v_res_w);
+  return v_res;
+}
+
+static INLINE __m128i blend_16_u8(const uint8_t *src0, const uint8_t *src1,
+                                  const __m128i *v_m0_b, const __m128i *v_m1_b,
+                                  const __m128i *rounding) {
+  const __m128i v_s0_b = xx_loadu_128(src0);
+  const __m128i v_s1_b = xx_loadu_128(src1);
+
+  const __m128i v_p0_w = _mm_maddubs_epi16(_mm_unpacklo_epi8(v_s0_b, v_s1_b),
+                                           _mm_unpacklo_epi8(*v_m0_b, *v_m1_b));
+  const __m128i v_p1_w = _mm_maddubs_epi16(_mm_unpackhi_epi8(v_s0_b, v_s1_b),
+                                           _mm_unpackhi_epi8(*v_m0_b, *v_m1_b));
+
+  const __m128i v_res0_w = _mm_mulhrs_epi16(v_p0_w, *rounding);
+  const __m128i v_res1_w = _mm_mulhrs_epi16(v_p1_w, *rounding);
+  const __m128i v_res = _mm_packus_epi16(v_res0_w, v_res1_w);
+  return v_res;
+}
+
+typedef __m128i (*blend_unit_fn)(const uint16_t *src0, const uint16_t *src1,
+                                 const __m128i v_m0_w, const __m128i v_m1_w);
+
+static INLINE __m128i blend_4_b10(const uint16_t *src0, const uint16_t *src1,
+                                  const __m128i v_m0_w, const __m128i v_m1_w) {
+  const __m128i v_s0_w = xx_loadl_64(src0);
+  const __m128i v_s1_w = xx_loadl_64(src1);
+
+  const __m128i v_p0_w = _mm_mullo_epi16(v_s0_w, v_m0_w);
+  const __m128i v_p1_w = _mm_mullo_epi16(v_s1_w, v_m1_w);
+
+  const __m128i v_sum_w = _mm_add_epi16(v_p0_w, v_p1_w);
+
+  const __m128i v_res_w = xx_roundn_epu16(v_sum_w, AOM_BLEND_A64_ROUND_BITS);
+
+  return v_res_w;
+}
+
+static INLINE __m128i blend_8_b10(const uint16_t *src0, const uint16_t *src1,
+                                  const __m128i v_m0_w, const __m128i v_m1_w) {
+  const __m128i v_s0_w = xx_loadu_128(src0);
+  const __m128i v_s1_w = xx_loadu_128(src1);
+
+  const __m128i v_p0_w = _mm_mullo_epi16(v_s0_w, v_m0_w);
+  const __m128i v_p1_w = _mm_mullo_epi16(v_s1_w, v_m1_w);
+
+  const __m128i v_sum_w = _mm_add_epi16(v_p0_w, v_p1_w);
+
+  const __m128i v_res_w = xx_roundn_epu16(v_sum_w, AOM_BLEND_A64_ROUND_BITS);
+
+  return v_res_w;
+}
+
+static INLINE __m128i blend_4_b12(const uint16_t *src0, const uint16_t *src1,
+                                  const __m128i v_m0_w, const __m128i v_m1_w) {
+  const __m128i v_s0_w = xx_loadl_64(src0);
+  const __m128i v_s1_w = xx_loadl_64(src1);
+
+  // Interleave
+  const __m128i v_m01_w = _mm_unpacklo_epi16(v_m0_w, v_m1_w);
+  const __m128i v_s01_w = _mm_unpacklo_epi16(v_s0_w, v_s1_w);
+
+  // Multiply-Add
+  const __m128i v_sum_d = _mm_madd_epi16(v_s01_w, v_m01_w);
+
+  // Scale
+  const __m128i v_ssum_d =
+      _mm_srli_epi32(v_sum_d, AOM_BLEND_A64_ROUND_BITS - 1);
+
+  // Pack
+  const __m128i v_pssum_d = _mm_packs_epi32(v_ssum_d, v_ssum_d);
+
+  // Round
+  const __m128i v_res_w = xx_round_epu16(v_pssum_d);
+
+  return v_res_w;
+}
+
+static INLINE __m128i blend_8_b12(const uint16_t *src0, const uint16_t *src1,
+                                  const __m128i v_m0_w, const __m128i v_m1_w) {
+  const __m128i v_s0_w = xx_loadu_128(src0);
+  const __m128i v_s1_w = xx_loadu_128(src1);
+
+  // Interleave
+  const __m128i v_m01l_w = _mm_unpacklo_epi16(v_m0_w, v_m1_w);
+  const __m128i v_m01h_w = _mm_unpackhi_epi16(v_m0_w, v_m1_w);
+  const __m128i v_s01l_w = _mm_unpacklo_epi16(v_s0_w, v_s1_w);
+  const __m128i v_s01h_w = _mm_unpackhi_epi16(v_s0_w, v_s1_w);
+
+  // Multiply-Add
+  const __m128i v_suml_d = _mm_madd_epi16(v_s01l_w, v_m01l_w);
+  const __m128i v_sumh_d = _mm_madd_epi16(v_s01h_w, v_m01h_w);
+
+  // Scale
+  const __m128i v_ssuml_d =
+      _mm_srli_epi32(v_suml_d, AOM_BLEND_A64_ROUND_BITS - 1);
+  const __m128i v_ssumh_d =
+      _mm_srli_epi32(v_sumh_d, AOM_BLEND_A64_ROUND_BITS - 1);
+
+  // Pack
+  const __m128i v_pssum_d = _mm_packs_epi32(v_ssuml_d, v_ssumh_d);
+
+  // Round
+  const __m128i v_res_w = xx_round_epu16(v_pssum_d);
+
+  return v_res_w;
+}
+
+#endif  // AOM_AOM_DSP_X86_BLEND_SSE4_H_
diff --git a/third_party/aom/aom_dsp/x86/blk_sse_sum_avx2.c b/third_party/aom/aom_dsp/x86/blk_sse_sum_avx2.c
new file mode 100644
index 0000000000..fdf7de3f4c
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/blk_sse_sum_avx2.c
@@ -0,0 +1,185 @@
+/*
+ * Copyright (c) 2019, 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 accumulate_sse_sum(__m256i regx_sum, __m256i regx2_sum,
+                                      int *x_sum, int64_t *x2_sum) {
+  __m256i sum_buffer, sse_buffer;
+  __m128i out_buffer;
+
+  // Accumulate the various elements of register into first element.
+  sum_buffer = _mm256_permute2f128_si256(regx_sum, regx_sum, 1);
+  regx_sum = _mm256_add_epi32(sum_buffer, regx_sum);
+  regx_sum = _mm256_add_epi32(regx_sum, _mm256_srli_si256(regx_sum, 8));
+  regx_sum = _mm256_add_epi32(regx_sum, _mm256_srli_si256(regx_sum, 4));
+
+  sse_buffer = _mm256_permute2f128_si256(regx2_sum, regx2_sum, 1);
+  regx2_sum = _mm256_add_epi64(sse_buffer, regx2_sum);
+  regx2_sum = _mm256_add_epi64(regx2_sum, _mm256_srli_si256(regx2_sum, 8));
+
+  out_buffer = _mm256_castsi256_si128(regx_sum);
+  *x_sum += _mm_cvtsi128_si32(out_buffer);
+  out_buffer = _mm256_castsi256_si128(regx2_sum);
+#if AOM_ARCH_X86_64
+  *x2_sum += _mm_cvtsi128_si64(out_buffer);
+#else
+  {
+    int64_t tmp;
+    _mm_storel_epi64((__m128i *)&tmp, out_buffer);
+    *x2_sum += tmp;
+  }
+#endif
+}
+
+static INLINE void sse_sum_wd4_avx2(const int16_t *data, int stride, int bh,
+                                    int *x_sum, int64_t *x2_sum) {
+  __m128i row1, row2, row3;
+  __m256i regx_sum, regx2_sum, load_pixels, sum_buffer, sse_buffer,
+      temp_buffer1, temp_buffer2, row_sum_buffer, row_sse_buffer;
+  const int16_t *data_tmp = data;
+  __m256i one = _mm256_set1_epi16(1);
+  regx_sum = _mm256_setzero_si256();
+  regx2_sum = regx_sum;
+  sum_buffer = _mm256_setzero_si256();
+  sse_buffer = sum_buffer;
+
+  for (int j = 0; j < (bh >> 2); ++j) {
+    // Load 4 rows at a time.
+    row1 = _mm_loadl_epi64((__m128i const *)(data_tmp));
+    row2 = _mm_loadl_epi64((__m128i const *)(data_tmp + stride));
+    row1 = _mm_unpacklo_epi64(row1, row2);
+    row2 = _mm_loadl_epi64((__m128i const *)(data_tmp + 2 * stride));
+    row3 = _mm_loadl_epi64((__m128i const *)(data_tmp + 3 * stride));
+    row2 = _mm_unpacklo_epi64(row2, row3);
+    load_pixels =
+        _mm256_insertf128_si256(_mm256_castsi128_si256(row1), row2, 1);
+
+    row_sum_buffer = _mm256_madd_epi16(load_pixels, one);
+    row_sse_buffer = _mm256_madd_epi16(load_pixels, load_pixels);
+    sum_buffer = _mm256_add_epi32(row_sum_buffer, sum_buffer);
+    sse_buffer = _mm256_add_epi32(row_sse_buffer, sse_buffer);
+    data_tmp += 4 * stride;
+  }
+
+  // To prevent 32-bit variable overflow, unpack the elements to 64-bit.
+  temp_buffer1 = _mm256_unpacklo_epi32(sse_buffer, _mm256_setzero_si256());
+  temp_buffer2 = _mm256_unpackhi_epi32(sse_buffer, _mm256_setzero_si256());
+  sse_buffer = _mm256_add_epi64(temp_buffer1, temp_buffer2);
+  regx_sum = _mm256_add_epi32(sum_buffer, regx_sum);
+  regx2_sum = _mm256_add_epi64(sse_buffer, regx2_sum);
+
+  accumulate_sse_sum(regx_sum, regx2_sum, x_sum, x2_sum);
+}
+
+static INLINE void sse_sum_wd8_avx2(const int16_t *data, int stride, int bh,
+                                    int *x_sum, int64_t *x2_sum) {
+  __m128i load_128bit, load_next_128bit;
+  __m256i regx_sum, regx2_sum, load_pixels, sum_buffer, sse_buffer,
+      temp_buffer1, temp_buffer2, row_sum_buffer, row_sse_buffer;
+  const int16_t *data_tmp = data;
+  __m256i one = _mm256_set1_epi16(1);
+  regx_sum = _mm256_setzero_si256();
+  regx2_sum = regx_sum;
+  sum_buffer = _mm256_setzero_si256();
+  sse_buffer = sum_buffer;
+
+  for (int j = 0; j < (bh >> 1); ++j) {
+    // Load 2 rows at a time.
+    load_128bit = _mm_loadu_si128((__m128i const *)(data_tmp));
+    load_next_128bit = _mm_loadu_si128((__m128i const *)(data_tmp + stride));
+    load_pixels = _mm256_insertf128_si256(_mm256_castsi128_si256(load_128bit),
+                                          load_next_128bit, 1);
+
+    row_sum_buffer = _mm256_madd_epi16(load_pixels, one);
+    row_sse_buffer = _mm256_madd_epi16(load_pixels, load_pixels);
+    sum_buffer = _mm256_add_epi32(row_sum_buffer, sum_buffer);
+    sse_buffer = _mm256_add_epi32(row_sse_buffer, sse_buffer);
+    data_tmp += 2 * stride;
+  }
+
+  temp_buffer1 = _mm256_unpacklo_epi32(sse_buffer, _mm256_setzero_si256());
+  temp_buffer2 = _mm256_unpackhi_epi32(sse_buffer, _mm256_setzero_si256());
+  sse_buffer = _mm256_add_epi64(temp_buffer1, temp_buffer2);
+  regx_sum = _mm256_add_epi32(sum_buffer, regx_sum);
+  regx2_sum = _mm256_add_epi64(sse_buffer, regx2_sum);
+
+  accumulate_sse_sum(regx_sum, regx2_sum, x_sum, x2_sum);
+}
+
+static INLINE void sse_sum_wd16_avx2(const int16_t *data, int stride, int bh,
+                                     int *x_sum, int64_t *x2_sum,
+                                     int loop_count) {
+  __m256i regx_sum, regx2_sum, load_pixels, sum_buffer, sse_buffer,
+      temp_buffer1, temp_buffer2, row_sum_buffer, row_sse_buffer;
+  const int16_t *data_tmp = data;
+  __m256i one = _mm256_set1_epi16(1);
+  regx_sum = _mm256_setzero_si256();
+  regx2_sum = regx_sum;
+  sum_buffer = _mm256_setzero_si256();
+  sse_buffer = sum_buffer;
+
+  for (int i = 0; i < loop_count; ++i) {
+    data_tmp = data + 16 * i;
+    for (int j = 0; j < bh; ++j) {
+      load_pixels = _mm256_lddqu_si256((__m256i const *)(data_tmp));
+
+      row_sum_buffer = _mm256_madd_epi16(load_pixels, one);
+      row_sse_buffer = _mm256_madd_epi16(load_pixels, load_pixels);
+      sum_buffer = _mm256_add_epi32(row_sum_buffer, sum_buffer);
+      sse_buffer = _mm256_add_epi32(row_sse_buffer, sse_buffer);
+      data_tmp += stride;
+    }
+  }
+
+  temp_buffer1 = _mm256_unpacklo_epi32(sse_buffer, _mm256_setzero_si256());
+  temp_buffer2 = _mm256_unpackhi_epi32(sse_buffer, _mm256_setzero_si256());
+  sse_buffer = _mm256_add_epi64(temp_buffer1, temp_buffer2);
+  regx_sum = _mm256_add_epi32(sum_buffer, regx_sum);
+  regx2_sum = _mm256_add_epi64(sse_buffer, regx2_sum);
+
+  accumulate_sse_sum(regx_sum, regx2_sum, x_sum, x2_sum);
+}
+
+void aom_get_blk_sse_sum_avx2(const int16_t *data, int stride, int bw, int bh,
+                              int *x_sum, int64_t *x2_sum) {
+  *x_sum = 0;
+  *x2_sum = 0;
+
+  if ((bh & 3) == 0) {
+    switch (bw) {
+        // For smaller block widths, compute multiple rows simultaneously.
+      case 4: sse_sum_wd4_avx2(data, stride, bh, x_sum, x2_sum); break;
+      case 8: sse_sum_wd8_avx2(data, stride, bh, x_sum, x2_sum); break;
+      case 16:
+      case 32:
+        sse_sum_wd16_avx2(data, stride, bh, x_sum, x2_sum, bw >> 4);
+        break;
+      case 64:
+        // 32-bit variables will overflow for 64 rows at a single time, so
+        // compute 32 rows at a time.
+        if (bh <= 32) {
+          sse_sum_wd16_avx2(data, stride, bh, x_sum, x2_sum, bw >> 4);
+        } else {
+          sse_sum_wd16_avx2(data, stride, 32, x_sum, x2_sum, bw >> 4);
+          sse_sum_wd16_avx2(data + 32 * stride, stride, 32, x_sum, x2_sum,
+                            bw >> 4);
+        }
+        break;
+
+      default: aom_get_blk_sse_sum_c(data, stride, bw, bh, x_sum, x2_sum);
+    }
+  } else {
+    aom_get_blk_sse_sum_c(data, stride, bw, bh, x_sum, x2_sum);
+  }
+}
diff --git a/third_party/aom/aom_dsp/x86/blk_sse_sum_sse2.c b/third_party/aom/aom_dsp/x86/blk_sse_sum_sse2.c
new file mode 100644
index 0000000000..bf89427872
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/blk_sse_sum_sse2.c
@@ -0,0 +1,138 @@
+/*
+ * Copyright (c) 2019, 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 <emmintrin.h>
+
+#include "config/aom_dsp_rtcd.h"
+
+static INLINE void sse_sum_wd4_sse2(const int16_t *data, int stride, int bh,
+                                    int *x_sum, int64_t *x2_sum) {
+  const int16_t *data_tmp = data;
+  __m128i temp_buffer1, temp_buffer2;
+  __m128i load_pixels_low, load_pixels_hi, sum_buffer, sse_buffer;
+  __m128i one = _mm_set1_epi16(1);
+  __m128i regx_sum = _mm_setzero_si128();
+  __m128i regx2_sum = regx_sum;
+
+  for (int j = 0; j < (bh >> 1); ++j) {
+    // Load 2 rows (8 pixels) at a time.
+    load_pixels_low = _mm_loadl_epi64((__m128i const *)(data_tmp));
+    load_pixels_hi = _mm_loadl_epi64((__m128i const *)(data_tmp + stride));
+    load_pixels_low = _mm_unpacklo_epi64(load_pixels_low, load_pixels_hi);
+    sum_buffer = _mm_madd_epi16(load_pixels_low, one);
+    sse_buffer = _mm_madd_epi16(load_pixels_low, load_pixels_low);
+    regx_sum = _mm_add_epi32(sum_buffer, regx_sum);
+    regx2_sum = _mm_add_epi32(sse_buffer, regx2_sum);
+    data_tmp += 2 * stride;
+  }
+
+  regx_sum = _mm_add_epi32(regx_sum, _mm_srli_si128(regx_sum, 8));
+  regx_sum = _mm_add_epi32(regx_sum, _mm_srli_si128(regx_sum, 4));
+  *x_sum = _mm_cvtsi128_si32(regx_sum);
+  temp_buffer1 = _mm_unpacklo_epi32(regx2_sum, _mm_setzero_si128());
+  temp_buffer2 = _mm_unpackhi_epi32(regx2_sum, _mm_setzero_si128());
+  regx2_sum = _mm_add_epi64(temp_buffer1, temp_buffer2);
+  regx2_sum = _mm_add_epi64(regx2_sum, _mm_srli_si128(regx2_sum, 8));
+#if AOM_ARCH_X86_64
+  *x2_sum += _mm_cvtsi128_si64(regx2_sum);
+#else
+  {
+    int64_t tmp;
+    _mm_storel_epi64((__m128i *)&tmp, regx2_sum);
+    *x2_sum += tmp;
+  }
+#endif
+}
+
+static INLINE void sse_sum_wd8_sse2(const int16_t *data, int stride, int bh,
+                                    int *x_sum, int64_t *x2_sum,
+                                    int loop_cycles) {
+  const int16_t *data_tmp;
+  __m128i temp_buffer1, temp_buffer2;
+  __m128i one = _mm_set1_epi16(1);
+  __m128i regx_sum = _mm_setzero_si128();
+  __m128i regx2_sum = regx_sum;
+  __m128i load_pixels, sum_buffer, sse_buffer;
+
+  for (int i = 0; i < loop_cycles; ++i) {
+    data_tmp = data + (8 * i);
+    for (int j = 0; j < bh; ++j) {
+      // Load 1 row (8-pixels) at a time.
+      load_pixels = _mm_loadu_si128((__m128i const *)(data_tmp));
+      sum_buffer = _mm_madd_epi16(load_pixels, one);
+      sse_buffer = _mm_madd_epi16(load_pixels, load_pixels);
+      regx_sum = _mm_add_epi32(sum_buffer, regx_sum);
+      regx2_sum = _mm_add_epi32(sse_buffer, regx2_sum);
+      data_tmp += stride;
+    }
+  }
+
+  regx_sum = _mm_add_epi32(regx_sum, _mm_srli_si128(regx_sum, 8));
+  regx_sum = _mm_add_epi32(regx_sum, _mm_srli_si128(regx_sum, 4));
+  *x_sum += _mm_cvtsi128_si32(regx_sum);
+  temp_buffer1 = _mm_unpacklo_epi32(regx2_sum, _mm_setzero_si128());
+  temp_buffer2 = _mm_unpackhi_epi32(regx2_sum, _mm_setzero_si128());
+  regx2_sum = _mm_add_epi64(temp_buffer1, temp_buffer2);
+  regx2_sum = _mm_add_epi64(regx2_sum, _mm_srli_si128(regx2_sum, 8));
+#if AOM_ARCH_X86_64
+  *x2_sum += _mm_cvtsi128_si64(regx2_sum);
+#else
+  {
+    int64_t tmp;
+    _mm_storel_epi64((__m128i *)&tmp, regx2_sum);
+    *x2_sum += tmp;
+  }
+#endif
+}
+
+// This functions adds SSE2 Support for the functions 'get_blk_sse_sum_c'
+void aom_get_blk_sse_sum_sse2(const int16_t *data, int stride, int bw, int bh,
+                              int *x_sum, int64_t *x2_sum) {
+  *x_sum = 0;
+  *x2_sum = 0;
+
+  if ((bh & 3) == 0) {
+    switch (bw) {
+      case 4: sse_sum_wd4_sse2(data, stride, bh, x_sum, x2_sum); break;
+      case 8:
+      case 16:
+        sse_sum_wd8_sse2(data, stride, bh, x_sum, x2_sum, bw >> 3);
+        break;
+        // For widths 32 and 64, the registers may overflow. So compute
+        // partial widths at a time.
+      case 32:
+        if (bh <= 32) {
+          sse_sum_wd8_sse2(data, stride, bh, x_sum, x2_sum, bw >> 3);
+          break;
+        } else {
+          sse_sum_wd8_sse2(data, stride, 32, x_sum, x2_sum, bw >> 3);
+          sse_sum_wd8_sse2(data + 32 * stride, stride, 32, x_sum, x2_sum,
+                           bw >> 3);
+          break;
+        }
+
+      case 64:
+        if (bh <= 16) {
+          sse_sum_wd8_sse2(data, stride, bh, x_sum, x2_sum, bw >> 3);
+          break;
+        } else {
+          for (int i = 0; i < bh; i += 16)
+            sse_sum_wd8_sse2(data + i * stride, stride, 16, x_sum, x2_sum,
+                             bw >> 3);
+          break;
+        }
+
+      default: aom_get_blk_sse_sum_c(data, stride, bw, bh, x_sum, x2_sum);
+    }
+  } else {
+    aom_get_blk_sse_sum_c(data, stride, bw, bh, x_sum, x2_sum);
+  }
+}
diff --git a/third_party/aom/aom_dsp/x86/common_avx2.h b/third_party/aom/aom_dsp/x86/common_avx2.h
new file mode 100644
index 0000000000..96fe4ebb67
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/common_avx2.h
@@ -0,0 +1,147 @@
+/*
+ * Copyright (c) 2017, 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_AOM_DSP_X86_COMMON_AVX2_H_
+#define AOM_AOM_DSP_X86_COMMON_AVX2_H_
+
+#include <immintrin.h>
+
+#include "config/aom_config.h"
+
+// Note: in and out could have the same value
+static INLINE void mm256_transpose_16x16(const __m256i *in, __m256i *out) {
+  __m256i tr0_0 = _mm256_unpacklo_epi16(in[0], in[1]);
+  __m256i tr0_1 = _mm256_unpackhi_epi16(in[0], in[1]);
+  __m256i tr0_2 = _mm256_unpacklo_epi16(in[2], in[3]);
+  __m256i tr0_3 = _mm256_unpackhi_epi16(in[2], in[3]);
+  __m256i tr0_4 = _mm256_unpacklo_epi16(in[4], in[5]);
+  __m256i tr0_5 = _mm256_unpackhi_epi16(in[4], in[5]);
+  __m256i tr0_6 = _mm256_unpacklo_epi16(in[6], in[7]);
+  __m256i tr0_7 = _mm256_unpackhi_epi16(in[6], in[7]);
+
+  __m256i tr0_8 = _mm256_unpacklo_epi16(in[8], in[9]);
+  __m256i tr0_9 = _mm256_unpackhi_epi16(in[8], in[9]);
+  __m256i tr0_a = _mm256_unpacklo_epi16(in[10], in[11]);
+  __m256i tr0_b = _mm256_unpackhi_epi16(in[10], in[11]);
+  __m256i tr0_c = _mm256_unpacklo_epi16(in[12], in[13]);
+  __m256i tr0_d = _mm256_unpackhi_epi16(in[12], in[13]);
+  __m256i tr0_e = _mm256_unpacklo_epi16(in[14], in[15]);
+  __m256i tr0_f = _mm256_unpackhi_epi16(in[14], in[15]);
+
+  // 00 10 01 11 02 12 03 13  08 18 09 19 0a 1a 0b 1b
+  // 04 14 05 15 06 16 07 17  0c 1c 0d 1d 0e 1e 0f 1f
+  // 20 30 21 31 22 32 23 33  28 38 29 39 2a 3a 2b 3b
+  // 24 34 25 35 26 36 27 37  2c 3c 2d 3d 2e 3e 2f 3f
+  // 40 50 41 51 42 52 43 53  48 58 49 59 4a 5a 4b 5b
+  // 44 54 45 55 46 56 47 57  4c 5c 4d 5d 4e 5e 4f 5f
+  // 60 70 61 71 62 72 63 73  68 78 69 79 6a 7a 6b 7b
+  // 64 74 65 75 66 76 67 77  6c 7c 6d 7d 6e 7e 6f 7f
+
+  // 80 90 81 91 82 92 83 93  88 98 89 99 8a 9a 8b 9b
+  // 84 94 85 95 86 96 87 97  8c 9c 8d 9d 8e 9e 8f 9f
+  // a0 b0 a1 b1 a2 b2 a3 b3  a8 b8 a9 b9 aa ba ab bb
+  // a4 b4 a5 b5 a6 b6 a7 b7  ac bc ad bd ae be af bf
+  // c0 d0 c1 d1 c2 d2 c3 d3  c8 d8 c9 d9 ca da cb db
+  // c4 d4 c5 d5 c6 d6 c7 d7  cc dc cd dd ce de cf df
+  // e0 f0 e1 f1 e2 f2 e3 f3  e8 f8 e9 f9 ea fa eb fb
+  // e4 f4 e5 f5 e6 f6 e7 f7  ec fc ed fd ee fe ef ff
+
+  __m256i tr1_0 = _mm256_unpacklo_epi32(tr0_0, tr0_2);
+  __m256i tr1_1 = _mm256_unpackhi_epi32(tr0_0, tr0_2);
+  __m256i tr1_2 = _mm256_unpacklo_epi32(tr0_1, tr0_3);
+  __m256i tr1_3 = _mm256_unpackhi_epi32(tr0_1, tr0_3);
+  __m256i tr1_4 = _mm256_unpacklo_epi32(tr0_4, tr0_6);
+  __m256i tr1_5 = _mm256_unpackhi_epi32(tr0_4, tr0_6);
+  __m256i tr1_6 = _mm256_unpacklo_epi32(tr0_5, tr0_7);
+  __m256i tr1_7 = _mm256_unpackhi_epi32(tr0_5, tr0_7);
+
+  __m256i tr1_8 = _mm256_unpacklo_epi32(tr0_8, tr0_a);
+  __m256i tr1_9 = _mm256_unpackhi_epi32(tr0_8, tr0_a);
+  __m256i tr1_a = _mm256_unpacklo_epi32(tr0_9, tr0_b);
+  __m256i tr1_b = _mm256_unpackhi_epi32(tr0_9, tr0_b);
+  __m256i tr1_c = _mm256_unpacklo_epi32(tr0_c, tr0_e);
+  __m256i tr1_d = _mm256_unpackhi_epi32(tr0_c, tr0_e);
+  __m256i tr1_e = _mm256_unpacklo_epi32(tr0_d, tr0_f);
+  __m256i tr1_f = _mm256_unpackhi_epi32(tr0_d, tr0_f);
+
+  // 00 10 20 30 01 11 21 31  08 18 28 38 09 19 29 39
+  // 02 12 22 32 03 13 23 33  0a 1a 2a 3a 0b 1b 2b 3b
+  // 04 14 24 34 05 15 25 35  0c 1c 2c 3c 0d 1d 2d 3d
+  // 06 16 26 36 07 17 27 37  0e 1e 2e 3e 0f 1f 2f 3f
+  // 40 50 60 70 41 51 61 71  48 58 68 78 49 59 69 79
+  // 42 52 62 72 43 53 63 73  4a 5a 6a 7a 4b 5b 6b 7b
+  // 44 54 64 74 45 55 65 75  4c 5c 6c 7c 4d 5d 6d 7d
+  // 46 56 66 76 47 57 67 77  4e 5e 6e 7e 4f 5f 6f 7f
+
+  // 80 90 a0 b0 81 91 a1 b1  88 98 a8 b8 89 99 a9 b9
+  // 82 92 a2 b2 83 93 a3 b3  8a 9a aa ba 8b 9b ab bb
+  // 84 94 a4 b4 85 95 a5 b5  8c 9c ac bc 8d 9d ad bd
+  // 86 96 a6 b6 87 97 a7 b7  8e ae 9e be 8f 9f af bf
+  // c0 d0 e0 f0 c1 d1 e1 f1  c8 d8 e8 f8 c9 d9 e9 f9
+  // c2 d2 e2 f2 c3 d3 e3 f3  ca da ea fa cb db eb fb
+  // c4 d4 e4 f4 c5 d5 e5 f5  cc dc ef fc cd dd ed fd
+  // c6 d6 e6 f6 c7 d7 e7 f7  ce de ee fe cf df ef ff
+
+  tr0_0 = _mm256_unpacklo_epi64(tr1_0, tr1_4);
+  tr0_1 = _mm256_unpackhi_epi64(tr1_0, tr1_4);
+  tr0_2 = _mm256_unpacklo_epi64(tr1_1, tr1_5);
+  tr0_3 = _mm256_unpackhi_epi64(tr1_1, tr1_5);
+  tr0_4 = _mm256_unpacklo_epi64(tr1_2, tr1_6);
+  tr0_5 = _mm256_unpackhi_epi64(tr1_2, tr1_6);
+  tr0_6 = _mm256_unpacklo_epi64(tr1_3, tr1_7);
+  tr0_7 = _mm256_unpackhi_epi64(tr1_3, tr1_7);
+
+  tr0_8 = _mm256_unpacklo_epi64(tr1_8, tr1_c);
+  tr0_9 = _mm256_unpackhi_epi64(tr1_8, tr1_c);
+  tr0_a = _mm256_unpacklo_epi64(tr1_9, tr1_d);
+  tr0_b = _mm256_unpackhi_epi64(tr1_9, tr1_d);
+  tr0_c = _mm256_unpacklo_epi64(tr1_a, tr1_e);
+  tr0_d = _mm256_unpackhi_epi64(tr1_a, tr1_e);
+  tr0_e = _mm256_unpacklo_epi64(tr1_b, tr1_f);
+  tr0_f = _mm256_unpackhi_epi64(tr1_b, tr1_f);
+
+  // 00 10 20 30 40 50 60 70  08 18 28 38 48 58 68 78
+  // 01 11 21 31 41 51 61 71  09 19 29 39 49 59 69 79
+  // 02 12 22 32 42 52 62 72  0a 1a 2a 3a 4a 5a 6a 7a
+  // 03 13 23 33 43 53 63 73  0b 1b 2b 3b 4b 5b 6b 7b
+  // 04 14 24 34 44 54 64 74  0c 1c 2c 3c 4c 5c 6c 7c
+  // 05 15 25 35 45 55 65 75  0d 1d 2d 3d 4d 5d 6d 7d
+  // 06 16 26 36 46 56 66 76  0e 1e 2e 3e 4e 5e 6e 7e
+  // 07 17 27 37 47 57 67 77  0f 1f 2f 3f 4f 5f 6f 7f
+
+  // 80 90 a0 b0 c0 d0 e0 f0  88 98 a8 b8 c8 d8 e8 f8
+  // 81 91 a1 b1 c1 d1 e1 f1  89 99 a9 b9 c9 d9 e9 f9
+  // 82 92 a2 b2 c2 d2 e2 f2  8a 9a aa ba ca da ea fa
+  // 83 93 a3 b3 c3 d3 e3 f3  8b 9b ab bb cb db eb fb
+  // 84 94 a4 b4 c4 d4 e4 f4  8c 9c ac bc cc dc ef fc
+  // 85 95 a5 b5 c5 d5 e5 f5  8d 9d ad bd cd dd ed fd
+  // 86 96 a6 b6 c6 d6 e6 f6  8e ae 9e be ce de ee fe
+  // 87 97 a7 b7 c7 d7 e7 f7  8f 9f af bf cf df ef ff
+
+  out[0] = _mm256_permute2x128_si256(tr0_0, tr0_8, 0x20);  // 0010 0000
+  out[8] = _mm256_permute2x128_si256(tr0_0, tr0_8, 0x31);  // 0011 0001
+  out[1] = _mm256_permute2x128_si256(tr0_1, tr0_9, 0x20);
+  out[9] = _mm256_permute2x128_si256(tr0_1, tr0_9, 0x31);
+  out[2] = _mm256_permute2x128_si256(tr0_2, tr0_a, 0x20);
+  out[10] = _mm256_permute2x128_si256(tr0_2, tr0_a, 0x31);
+  out[3] = _mm256_permute2x128_si256(tr0_3, tr0_b, 0x20);
+  out[11] = _mm256_permute2x128_si256(tr0_3, tr0_b, 0x31);
+
+  out[4] = _mm256_permute2x128_si256(tr0_4, tr0_c, 0x20);
+  out[12] = _mm256_permute2x128_si256(tr0_4, tr0_c, 0x31);
+  out[5] = _mm256_permute2x128_si256(tr0_5, tr0_d, 0x20);
+  out[13] = _mm256_permute2x128_si256(tr0_5, tr0_d, 0x31);
+  out[6] = _mm256_permute2x128_si256(tr0_6, tr0_e, 0x20);
+  out[14] = _mm256_permute2x128_si256(tr0_6, tr0_e, 0x31);
+  out[7] = _mm256_permute2x128_si256(tr0_7, tr0_f, 0x20);
+  out[15] = _mm256_permute2x128_si256(tr0_7, tr0_f, 0x31);
+}
+#endif  // AOM_AOM_DSP_X86_COMMON_AVX2_H_
diff --git a/third_party/aom/aom_dsp/x86/convolve.h b/third_party/aom/aom_dsp/x86/convolve.h
new file mode 100644
index 0000000000..4ca214f469
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/convolve.h
@@ -0,0 +1,204 @@
+/*
+ * Copyright (c) 2016, 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_AOM_DSP_X86_CONVOLVE_H_
+#define AOM_AOM_DSP_X86_CONVOLVE_H_
+
+#include <assert.h>
+
+#include "config/aom_config.h"
+#include "config/aom_dsp_rtcd.h"
+
+#include "aom/aom_integer.h"
+#include "aom_ports/mem.h"
+
+typedef void filter8_1dfunction(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);
+
+#define FUN_CONV_1D(name, step_q4, filter, dir, src_start, avg, opt)         \
+  void aom_convolve8_##name##_##opt(                                         \
+      const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst,                \
+      ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4,          \
+      const int16_t *filter_y, int y_step_q4, int w, int h) {                \
+    (void)filter_x;                                                          \
+    (void)x_step_q4;                                                         \
+    (void)filter_y;                                                          \
+    (void)y_step_q4;                                                         \
+    assert((-128 <= filter[3]) && (filter[3] <= 127));                       \
+    assert(step_q4 == 16);                                                   \
+    if (((filter[0] | filter[1] | filter[6] | filter[7]) == 0) &&            \
+        (filter[2] | filter[5])) {                                           \
+      while (w >= 16) {                                                      \
+        aom_filter_block1d16_##dir##4_##avg##opt(src_start, src_stride, dst, \
+                                                 dst_stride, h, filter);     \
+        src += 16;                                                           \
+        dst += 16;                                                           \
+        w -= 16;                                                             \
+      }                                                                      \
+      while (w >= 8) {                                                       \
+        aom_filter_block1d8_##dir##4_##avg##opt(src_start, src_stride, dst,  \
+                                                dst_stride, h, filter);      \
+        src += 8;                                                            \
+        dst += 8;                                                            \
+        w -= 8;                                                              \
+      }                                                                      \
+      while (w >= 4) {                                                       \
+        aom_filter_block1d4_##dir##4_##avg##opt(src_start, src_stride, dst,  \
+                                                dst_stride, h, filter);      \
+        src += 4;                                                            \
+        dst += 4;                                                            \
+        w -= 4;                                                              \
+      }                                                                      \
+    } else if (filter[0] | filter[1] | filter[2]) {                          \
+      while (w >= 16) {                                                      \
+        aom_filter_block1d16_##dir##8_##avg##opt(src_start, src_stride, dst, \
+                                                 dst_stride, h, filter);     \
+        src += 16;                                                           \
+        dst += 16;                                                           \
+        w -= 16;                                                             \
+      }                                                                      \
+      while (w >= 8) {                                                       \
+        aom_filter_block1d8_##dir##8_##avg##opt(src_start, src_stride, dst,  \
+                                                dst_stride, h, filter);      \
+        src += 8;                                                            \
+        dst += 8;                                                            \
+        w -= 8;                                                              \
+      }                                                                      \
+      while (w >= 4) {                                                       \
+        aom_filter_block1d4_##dir##8_##avg##opt(src_start, src_stride, dst,  \
+                                                dst_stride, h, filter);      \
+        src += 4;                                                            \
+        dst += 4;                                                            \
+        w -= 4;                                                              \
+      }                                                                      \
+    } else {                                                                 \
+      while (w >= 16) {                                                      \
+        aom_filter_block1d16_##dir##2_##avg##opt(src, src_stride, dst,       \
+                                                 dst_stride, h, filter);     \
+        src += 16;                                                           \
+        dst += 16;                                                           \
+        w -= 16;                                                             \
+      }                                                                      \
+      while (w >= 8) {                                                       \
+        aom_filter_block1d8_##dir##2_##avg##opt(src, src_stride, dst,        \
+                                                dst_stride, h, filter);      \
+        src += 8;                                                            \
+        dst += 8;                                                            \
+        w -= 8;                                                              \
+      }                                                                      \
+      while (w >= 4) {                                                       \
+        aom_filter_block1d4_##dir##2_##avg##opt(src, src_stride, dst,        \
+                                                dst_stride, h, filter);      \
+        src += 4;                                                            \
+        dst += 4;                                                            \
+        w -= 4;                                                              \
+      }                                                                      \
+    }                                                                        \
+    if (w) {                                                                 \
+      aom_convolve8_##name##_c(src, src_stride, dst, dst_stride, filter_x,   \
+                               x_step_q4, filter_y, y_step_q4, w, h);        \
+    }                                                                        \
+  }
+
+#if CONFIG_AV1_HIGHBITDEPTH
+typedef void highbd_filter8_1dfunction(const uint16_t *src_ptr,
+                                       const ptrdiff_t src_pitch,
+                                       uint16_t *output_ptr,
+                                       ptrdiff_t out_pitch,
+                                       unsigned int output_height,
+                                       const int16_t *filter, int bd);
+
+#define HIGH_FUN_CONV_1D(name, step_q4, filter, dir, src_start, avg, opt)  \
+  void aom_highbd_convolve8_##name##_##opt(                                \
+      const uint8_t *src8, ptrdiff_t src_stride, uint8_t *dst8,            \
+      ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4,        \
+      const int16_t *filter_y, int y_step_q4, int w, int h, int bd) {      \
+    uint16_t *src = CONVERT_TO_SHORTPTR(src8);                             \
+    uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);                             \
+    if (step_q4 == 16 && filter[3] != 128) {                               \
+      if (((filter[0] | filter[1] | filter[6] | filter[7]) == 0) &&        \
+          (filter[2] | filter[5])) {                                       \
+        while (w >= 16) {                                                  \
+          aom_highbd_filter_block1d16_##dir##4_##avg##opt(                 \
+              src_start, src_stride, dst, dst_stride, h, filter, bd);      \
+          src += 16;                                                       \
+          dst += 16;                                                       \
+          w -= 16;                                                         \
+        }                                                                  \
+        while (w >= 8) {                                                   \
+          aom_highbd_filter_block1d8_##dir##4_##avg##opt(                  \
+              src_start, src_stride, dst, dst_stride, h, filter, bd);      \
+          src += 8;                                                        \
+          dst += 8;                                                        \
+          w -= 8;                                                          \
+        }                                                                  \
+        while (w >= 4) {                                                   \
+          aom_highbd_filter_block1d4_##dir##4_##avg##opt(                  \
+              src_start, src_stride, dst, dst_stride, h, filter, bd);      \
+          src += 4;                                                        \
+          dst += 4;                                                        \
+          w -= 4;                                                          \
+        }                                                                  \
+      } else if (filter[0] | filter[1] | filter[2]) {                      \
+        while (w >= 16) {                                                  \
+          aom_highbd_filter_block1d16_##dir##8_##avg##opt(                 \
+              src_start, src_stride, dst, dst_stride, h, filter, bd);      \
+          src += 16;                                                       \
+          dst += 16;                                                       \
+          w -= 16;                                                         \
+        }                                                                  \
+        while (w >= 8) {                                                   \
+          aom_highbd_filter_block1d8_##dir##8_##avg##opt(                  \
+              src_start, src_stride, dst, dst_stride, h, filter, bd);      \
+          src += 8;                                                        \
+          dst += 8;                                                        \
+          w -= 8;                                                          \
+        }                                                                  \
+        while (w >= 4) {                                                   \
+          aom_highbd_filter_block1d4_##dir##8_##avg##opt(                  \
+              src_start, src_stride, dst, dst_stride, h, filter, bd);      \
+          src += 4;                                                        \
+          dst += 4;                                                        \
+          w -= 4;                                                          \
+        }                                                                  \
+      } else {                                                             \
+        while (w >= 16) {                                                  \
+          aom_highbd_filter_block1d16_##dir##2_##avg##opt(                 \
+              src, src_stride, dst, dst_stride, h, filter, bd);            \
+          src += 16;                                                       \
+          dst += 16;                                                       \
+          w -= 16;                                                         \
+        }                                                                  \
+        while (w >= 8) {                                                   \
+          aom_highbd_filter_block1d8_##dir##2_##avg##opt(                  \
+              src, src_stride, dst, dst_stride, h, filter, bd);            \
+          src += 8;                                                        \
+          dst += 8;                                                        \
+          w -= 8;                                                          \
+        }                                                                  \
+        while (w >= 4) {                                                   \
+          aom_highbd_filter_block1d4_##dir##2_##avg##opt(                  \
+              src, src_stride, dst, dst_stride, h, filter, bd);            \
+          src += 4;                                                        \
+          dst += 4;                                                        \
+          w -= 4;                                                          \
+        }                                                                  \
+      }                                                                    \
+    }                                                                      \
+    if (w) {                                                               \
+      aom_highbd_convolve8_##name##_c(                                     \
+          CONVERT_TO_BYTEPTR(src), src_stride, CONVERT_TO_BYTEPTR(dst),    \
+          dst_stride, filter_x, x_step_q4, filter_y, y_step_q4, w, h, bd); \
+    }                                                                      \
+  }
+#endif  // CONFIG_AV1_HIGHBITDEPTH
+
+#endif  // AOM_AOM_DSP_X86_CONVOLVE_H_
diff --git a/third_party/aom/aom_dsp/x86/convolve_avx2.h b/third_party/aom/aom_dsp/x86/convolve_avx2.h
new file mode 100644
index 0000000000..f5a382ce4e
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/convolve_avx2.h
@@ -0,0 +1,922 @@
+/*
+ * 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_AOM_DSP_X86_CONVOLVE_AVX2_H_
+#define AOM_AOM_DSP_X86_CONVOLVE_AVX2_H_
+
+#include <immintrin.h>
+
+#include "aom_ports/mem.h"
+
+#include "av1/common/convolve.h"
+#include "av1/common/filter.h"
+
+// filters for 16
+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, 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,
+};
+
+DECLARE_ALIGNED(32, static const uint8_t, filt4_d4_global_avx2[]) = {
+  2, 3, 4, 5, 3, 4, 5, 6, 4, 5, 6, 7, 5, 6, 7, 8,
+  2, 3, 4, 5, 3, 4, 5, 6, 4, 5, 6, 7, 5, 6, 7, 8,
+};
+
+DECLARE_ALIGNED(32, static const uint8_t, filt_center_global_avx2[32]) = {
+  3, 255, 4, 255, 5, 255, 6, 255, 7, 255, 8, 255, 9, 255, 10, 255,
+  3, 255, 4, 255, 5, 255, 6, 255, 7, 255, 8, 255, 9, 255, 10, 255
+};
+
+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,
+                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
+};
+
+#define CONVOLVE_SR_HORIZONTAL_FILTER_4TAP                                     \
+  for (i = 0; i < (im_h - 2); i += 2) {                                        \
+    __m256i data = _mm256_castsi128_si256(                                     \
+        _mm_loadu_si128((__m128i *)&src_ptr[(i * src_stride) + j]));           \
+    data = _mm256_inserti128_si256(                                            \
+        data,                                                                  \
+        _mm_loadu_si128(                                                       \
+            (__m128i *)&src_ptr[(i * src_stride) + j + src_stride]),           \
+        1);                                                                    \
+    __m256i res = convolve_lowbd_x_4tap(data, coeffs_h + 1, filt);             \
+    res =                                                                      \
+        _mm256_sra_epi16(_mm256_add_epi16(res, round_const_h), round_shift_h); \
+    _mm256_store_si256((__m256i *)&im_block[i * im_stride], res);              \
+  }                                                                            \
+  __m256i data_1 = _mm256_castsi128_si256(                                     \
+      _mm_loadu_si128((__m128i *)&src_ptr[(i * src_stride) + j]));             \
+  __m256i res = convolve_lowbd_x_4tap(data_1, coeffs_h + 1, filt);             \
+  res = _mm256_sra_epi16(_mm256_add_epi16(res, round_const_h), round_shift_h); \
+  _mm256_store_si256((__m256i *)&im_block[i * im_stride], res);
+
+#define CONVOLVE_SR_VERTICAL_FILTER_4TAP                                      \
+  __m256i s[6];                                                               \
+  __m256i src_0 = _mm256_loadu_si256((__m256i *)(im_block + 0 * im_stride));  \
+  __m256i src_1 = _mm256_loadu_si256((__m256i *)(im_block + 1 * im_stride));  \
+  __m256i src_2 = _mm256_loadu_si256((__m256i *)(im_block + 2 * im_stride));  \
+  __m256i src_3 = _mm256_loadu_si256((__m256i *)(im_block + 3 * im_stride));  \
+                                                                              \
+  s[0] = _mm256_unpacklo_epi16(src_0, src_1);                                 \
+  s[1] = _mm256_unpacklo_epi16(src_2, src_3);                                 \
+  s[3] = _mm256_unpackhi_epi16(src_0, src_1);                                 \
+  s[4] = _mm256_unpackhi_epi16(src_2, src_3);                                 \
+                                                                              \
+  for (i = 0; i < h; i += 2) {                                                \
+    const int16_t *data = &im_block[i * im_stride];                           \
+    const __m256i s4 = _mm256_loadu_si256((__m256i *)(data + 4 * im_stride)); \
+    const __m256i s5 = _mm256_loadu_si256((__m256i *)(data + 5 * im_stride)); \
+    s[2] = _mm256_unpacklo_epi16(s4, s5);                                     \
+    s[5] = _mm256_unpackhi_epi16(s4, s5);                                     \
+                                                                              \
+    __m256i res_a = convolve_4tap(s, coeffs_v + 1);                           \
+    __m256i res_b = convolve_4tap(s + 3, coeffs_v + 1);                       \
+                                                                              \
+    res_a =                                                                   \
+        _mm256_sra_epi32(_mm256_add_epi32(res_a, sum_round_v), sum_shift_v);  \
+    res_b =                                                                   \
+        _mm256_sra_epi32(_mm256_add_epi32(res_b, sum_round_v), sum_shift_v);  \
+    const __m256i res_a_round = _mm256_sra_epi32(                             \
+        _mm256_add_epi32(res_a, round_const_v), round_shift_v);               \
+    const __m256i res_b_round = _mm256_sra_epi32(                             \
+        _mm256_add_epi32(res_b, round_const_v), round_shift_v);               \
+    const __m256i res_16bit = _mm256_packs_epi32(res_a_round, res_b_round);   \
+    const __m256i res_8b = _mm256_packus_epi16(res_16bit, res_16bit);         \
+    const __m128i res_0 = _mm256_castsi256_si128(res_8b);                     \
+    const __m128i res_1 = _mm256_extracti128_si256(res_8b, 1);                \
+                                                                              \
+    __m128i *const p_0 = (__m128i *)&dst[i * dst_stride + j];                 \
+    __m128i *const p_1 = (__m128i *)&dst[i * dst_stride + j + dst_stride];    \
+    if (w - j > 4) {                                                          \
+      _mm_storel_epi64(p_0, res_0);                                           \
+      _mm_storel_epi64(p_1, res_1);                                           \
+    } else if (w == 4) {                                                      \
+      xx_storel_32(p_0, res_0);                                               \
+      xx_storel_32(p_1, res_1);                                               \
+    } else {                                                                  \
+      *(uint16_t *)p_0 = (uint16_t)_mm_cvtsi128_si32(res_0);                  \
+      *(uint16_t *)p_1 = (uint16_t)_mm_cvtsi128_si32(res_1);                  \
+    }                                                                         \
+                                                                              \
+    s[0] = s[1];                                                              \
+    s[1] = s[2];                                                              \
+    s[3] = s[4];                                                              \
+    s[4] = s[5];                                                              \
+  }
+
+#define CONVOLVE_SR_HORIZONTAL_FILTER_6TAP                                     \
+  for (i = 0; i < (im_h - 2); i += 2) {                                        \
+    __m256i data = _mm256_castsi128_si256(                                     \
+        _mm_loadu_si128((__m128i *)&src_ptr[(i * src_stride) + j]));           \
+    data = _mm256_inserti128_si256(                                            \
+        data,                                                                  \
+        _mm_loadu_si128(                                                       \
+            (__m128i *)&src_ptr[(i * src_stride) + j + src_stride]),           \
+        1);                                                                    \
+                                                                               \
+    __m256i res = convolve_lowbd_x_6tap(data, coeffs_h, filt);                 \
+    res =                                                                      \
+        _mm256_sra_epi16(_mm256_add_epi16(res, round_const_h), round_shift_h); \
+    _mm256_store_si256((__m256i *)&im_block[i * im_stride], res);              \
+  }                                                                            \
+                                                                               \
+  __m256i data_1 = _mm256_castsi128_si256(                                     \
+      _mm_loadu_si128((__m128i *)&src_ptr[(i * src_stride) + j]));             \
+                                                                               \
+  __m256i res = convolve_lowbd_x_6tap(data_1, coeffs_h, filt);                 \
+                                                                               \
+  res = _mm256_sra_epi16(_mm256_add_epi16(res, round_const_h), round_shift_h); \
+                                                                               \
+  _mm256_store_si256((__m256i *)&im_block[i * im_stride], res);
+
+#define CONVOLVE_SR_VERTICAL_FILTER_6TAP                                      \
+  __m256i src_0 = _mm256_loadu_si256((__m256i *)(im_block + 0 * im_stride));  \
+  __m256i src_1 = _mm256_loadu_si256((__m256i *)(im_block + 1 * im_stride));  \
+  __m256i src_2 = _mm256_loadu_si256((__m256i *)(im_block + 2 * im_stride));  \
+  __m256i src_3 = _mm256_loadu_si256((__m256i *)(im_block + 3 * im_stride));  \
+                                                                              \
+  __m256i s[8];                                                               \
+  s[0] = _mm256_unpacklo_epi16(src_0, src_1);                                 \
+  s[1] = _mm256_unpacklo_epi16(src_2, src_3);                                 \
+                                                                              \
+  s[3] = _mm256_unpackhi_epi16(src_0, src_1);                                 \
+  s[4] = _mm256_unpackhi_epi16(src_2, src_3);                                 \
+                                                                              \
+  for (i = 0; i < h; i += 2) {                                                \
+    const int16_t *data = &im_block[i * im_stride];                           \
+                                                                              \
+    const __m256i s6 = _mm256_loadu_si256((__m256i *)(data + 4 * im_stride)); \
+    const __m256i s7 = _mm256_loadu_si256((__m256i *)(data + 5 * im_stride)); \
+                                                                              \
+    s[2] = _mm256_unpacklo_epi16(s6, s7);                                     \
+    s[5] = _mm256_unpackhi_epi16(s6, s7);                                     \
+                                                                              \
+    __m256i res_a = convolve_6tap(s, coeffs_v);                               \
+    __m256i res_b = convolve_6tap(s + 3, coeffs_v);                           \
+                                                                              \
+    res_a =                                                                   \
+        _mm256_sra_epi32(_mm256_add_epi32(res_a, sum_round_v), sum_shift_v);  \
+    res_b =                                                                   \
+        _mm256_sra_epi32(_mm256_add_epi32(res_b, sum_round_v), sum_shift_v);  \
+                                                                              \
+    const __m256i res_a_round = _mm256_sra_epi32(                             \
+        _mm256_add_epi32(res_a, round_const_v), round_shift_v);               \
+    const __m256i res_b_round = _mm256_sra_epi32(                             \
+        _mm256_add_epi32(res_b, round_const_v), round_shift_v);               \
+                                                                              \
+    const __m256i res_16bit = _mm256_packs_epi32(res_a_round, res_b_round);   \
+    const __m256i res_8b = _mm256_packus_epi16(res_16bit, res_16bit);         \
+                                                                              \
+    const __m128i res_0 = _mm256_castsi256_si128(res_8b);                     \
+    const __m128i res_1 = _mm256_extracti128_si256(res_8b, 1);                \
+                                                                              \
+    __m128i *const p_0 = (__m128i *)&dst[i * dst_stride + j];                 \
+    __m128i *const p_1 = (__m128i *)&dst[i * dst_stride + j + dst_stride];    \
+    if (w - j > 4) {                                                          \
+      _mm_storel_epi64(p_0, res_0);                                           \
+      _mm_storel_epi64(p_1, res_1);                                           \
+    } else if (w == 4) {                                                      \
+      xx_storel_32(p_0, res_0);                                               \
+      xx_storel_32(p_1, res_1);                                               \
+    } else {                                                                  \
+      *(uint16_t *)p_0 = (uint16_t)_mm_cvtsi128_si32(res_0);                  \
+      *(uint16_t *)p_1 = (uint16_t)_mm_cvtsi128_si32(res_1);                  \
+    }                                                                         \
+                                                                              \
+    s[0] = s[1];                                                              \
+    s[1] = s[2];                                                              \
+                                                                              \
+    s[3] = s[4];                                                              \
+    s[4] = s[5];                                                              \
+  }
+
+#define CONVOLVE_SR_HORIZONTAL_FILTER_8TAP                                     \
+  for (i = 0; i < (im_h - 2); i += 2) {                                        \
+    __m256i data = _mm256_castsi128_si256(                                     \
+        _mm_loadu_si128((__m128i *)&src_ptr[(i * src_stride) + j]));           \
+    data = _mm256_inserti128_si256(                                            \
+        data,                                                                  \
+        _mm_loadu_si128(                                                       \
+            (__m128i *)&src_ptr[(i * src_stride) + j + src_stride]),           \
+        1);                                                                    \
+                                                                               \
+    __m256i res = convolve_lowbd_x(data, coeffs_h, filt);                      \
+    res =                                                                      \
+        _mm256_sra_epi16(_mm256_add_epi16(res, round_const_h), round_shift_h); \
+    _mm256_store_si256((__m256i *)&im_block[i * im_stride], res);              \
+  }                                                                            \
+                                                                               \
+  __m256i data_1 = _mm256_castsi128_si256(                                     \
+      _mm_loadu_si128((__m128i *)&src_ptr[(i * src_stride) + j]));             \
+                                                                               \
+  __m256i res = convolve_lowbd_x(data_1, coeffs_h, filt);                      \
+                                                                               \
+  res = _mm256_sra_epi16(_mm256_add_epi16(res, round_const_h), round_shift_h); \
+                                                                               \
+  _mm256_store_si256((__m256i *)&im_block[i * im_stride], res);
+
+#define CONVOLVE_SR_VERTICAL_FILTER_8TAP                                      \
+  __m256i src_0 = _mm256_loadu_si256((__m256i *)(im_block + 0 * im_stride));  \
+  __m256i src_1 = _mm256_loadu_si256((__m256i *)(im_block + 1 * im_stride));  \
+  __m256i src_2 = _mm256_loadu_si256((__m256i *)(im_block + 2 * im_stride));  \
+  __m256i src_3 = _mm256_loadu_si256((__m256i *)(im_block + 3 * im_stride));  \
+  __m256i src_4 = _mm256_loadu_si256((__m256i *)(im_block + 4 * im_stride));  \
+  __m256i src_5 = _mm256_loadu_si256((__m256i *)(im_block + 5 * im_stride));  \
+                                                                              \
+  __m256i s[8];                                                               \
+  s[0] = _mm256_unpacklo_epi16(src_0, src_1);                                 \
+  s[1] = _mm256_unpacklo_epi16(src_2, src_3);                                 \
+  s[2] = _mm256_unpacklo_epi16(src_4, src_5);                                 \
+                                                                              \
+  s[4] = _mm256_unpackhi_epi16(src_0, src_1);                                 \
+  s[5] = _mm256_unpackhi_epi16(src_2, src_3);                                 \
+  s[6] = _mm256_unpackhi_epi16(src_4, src_5);                                 \
+                                                                              \
+  for (i = 0; i < h; i += 2) {                                                \
+    const int16_t *data = &im_block[i * im_stride];                           \
+                                                                              \
+    const __m256i s6 = _mm256_loadu_si256((__m256i *)(data + 6 * im_stride)); \
+    const __m256i s7 = _mm256_loadu_si256((__m256i *)(data + 7 * im_stride)); \
+                                                                              \
+    s[3] = _mm256_unpacklo_epi16(s6, s7);                                     \
+    s[7] = _mm256_unpackhi_epi16(s6, s7);                                     \
+                                                                              \
+    __m256i res_a = convolve(s, coeffs_v);                                    \
+    __m256i res_b = convolve(s + 4, coeffs_v);                                \
+                                                                              \
+    res_a =                                                                   \
+        _mm256_sra_epi32(_mm256_add_epi32(res_a, sum_round_v), sum_shift_v);  \
+    res_b =                                                                   \
+        _mm256_sra_epi32(_mm256_add_epi32(res_b, sum_round_v), sum_shift_v);  \
+                                                                              \
+    const __m256i res_a_round = _mm256_sra_epi32(                             \
+        _mm256_add_epi32(res_a, round_const_v), round_shift_v);               \
+    const __m256i res_b_round = _mm256_sra_epi32(                             \
+        _mm256_add_epi32(res_b, round_const_v), round_shift_v);               \
+                                                                              \
+    const __m256i res_16bit = _mm256_packs_epi32(res_a_round, res_b_round);   \
+    const __m256i res_8b = _mm256_packus_epi16(res_16bit, res_16bit);         \
+                                                                              \
+    const __m128i res_0 = _mm256_castsi256_si128(res_8b);                     \
+    const __m128i res_1 = _mm256_extracti128_si256(res_8b, 1);                \
+                                                                              \
+    __m128i *const p_0 = (__m128i *)&dst[i * dst_stride + j];                 \
+    __m128i *const p_1 = (__m128i *)&dst[i * dst_stride + j + dst_stride];    \
+    if (w - j > 4) {                                                          \
+      _mm_storel_epi64(p_0, res_0);                                           \
+      _mm_storel_epi64(p_1, res_1);                                           \
+    } else if (w == 4) {                                                      \
+      xx_storel_32(p_0, res_0);                                               \
+      xx_storel_32(p_1, res_1);                                               \
+    } else {                                                                  \
+      *(uint16_t *)p_0 = (uint16_t)_mm_cvtsi128_si32(res_0);                  \
+      *(uint16_t *)p_1 = (uint16_t)_mm_cvtsi128_si32(res_1);                  \
+    }                                                                         \
+                                                                              \
+    s[0] = s[1];                                                              \
+    s[1] = s[2];                                                              \
+    s[2] = s[3];                                                              \
+                                                                              \
+    s[4] = s[5];                                                              \
+    s[5] = s[6];                                                              \
+    s[6] = s[7];                                                              \
+  }
+
+#define CONVOLVE_SR_HORIZONTAL_FILTER_12TAP                                    \
+  const __m256i v_zero = _mm256_setzero_si256();                               \
+  __m256i s[12];                                                               \
+  if (w <= 4) {                                                                \
+    for (i = 0; i < im_h; i += 2) {                                            \
+      const __m256i data = _mm256_permute2x128_si256(                          \
+          _mm256_castsi128_si256(                                              \
+              _mm_loadu_si128((__m128i *)(&src_ptr[i * src_stride + j]))),     \
+          _mm256_castsi128_si256(_mm_loadu_si128(                              \
+              (__m128i *)(&src_ptr[i * src_stride + src_stride + j]))),        \
+          0x20);                                                               \
+      const __m256i s_16lo = _mm256_unpacklo_epi8(data, v_zero);               \
+      const __m256i s_16hi = _mm256_unpackhi_epi8(data, v_zero);               \
+      const __m256i s_lolo = _mm256_unpacklo_epi16(s_16lo, s_16lo);            \
+      const __m256i s_lohi = _mm256_unpackhi_epi16(s_16lo, s_16lo);            \
+                                                                               \
+      const __m256i s_hilo = _mm256_unpacklo_epi16(s_16hi, s_16hi);            \
+      const __m256i s_hihi = _mm256_unpackhi_epi16(s_16hi, s_16hi);            \
+                                                                               \
+      s[0] = _mm256_alignr_epi8(s_lohi, s_lolo, 2);                            \
+      s[1] = _mm256_alignr_epi8(s_lohi, s_lolo, 10);                           \
+      s[2] = _mm256_alignr_epi8(s_hilo, s_lohi, 2);                            \
+      s[3] = _mm256_alignr_epi8(s_hilo, s_lohi, 10);                           \
+      s[4] = _mm256_alignr_epi8(s_hihi, s_hilo, 2);                            \
+      s[5] = _mm256_alignr_epi8(s_hihi, s_hilo, 10);                           \
+                                                                               \
+      const __m256i res_lo = convolve_12taps(s, coeffs_h);                     \
+                                                                               \
+      __m256i res_32b_lo = _mm256_sra_epi32(                                   \
+          _mm256_add_epi32(res_lo, round_const_h12), round_shift_h12);         \
+      __m256i res_16b_lo = _mm256_packs_epi32(res_32b_lo, res_32b_lo);         \
+      const __m128i res_0 = _mm256_extracti128_si256(res_16b_lo, 0);           \
+      const __m128i res_1 = _mm256_extracti128_si256(res_16b_lo, 1);           \
+      if (w > 2) {                                                             \
+        _mm_storel_epi64((__m128i *)&im_block[i * im_stride], res_0);          \
+        _mm_storel_epi64((__m128i *)&im_block[i * im_stride + im_stride],      \
+                         res_1);                                               \
+      } else {                                                                 \
+        uint32_t horiz_2;                                                      \
+        horiz_2 = (uint32_t)_mm_cvtsi128_si32(res_0);                          \
+        im_block[i * im_stride] = (uint16_t)horiz_2;                           \
+        im_block[i * im_stride + 1] = (uint16_t)(horiz_2 >> 16);               \
+        horiz_2 = (uint32_t)_mm_cvtsi128_si32(res_1);                          \
+        im_block[i * im_stride + im_stride] = (uint16_t)horiz_2;               \
+        im_block[i * im_stride + im_stride + 1] = (uint16_t)(horiz_2 >> 16);   \
+      }                                                                        \
+    }                                                                          \
+  } else {                                                                     \
+    for (i = 0; i < im_h; i++) {                                               \
+      const __m256i data = _mm256_permute2x128_si256(                          \
+          _mm256_castsi128_si256(                                              \
+              _mm_loadu_si128((__m128i *)(&src_ptr[i * src_stride + j]))),     \
+          _mm256_castsi128_si256(                                              \
+              _mm_loadu_si128((__m128i *)(&src_ptr[i * src_stride + j + 4]))), \
+          0x20);                                                               \
+      const __m256i s_16lo = _mm256_unpacklo_epi8(data, v_zero);               \
+      const __m256i s_16hi = _mm256_unpackhi_epi8(data, v_zero);               \
+                                                                               \
+      const __m256i s_lolo = _mm256_unpacklo_epi16(s_16lo, s_16lo);            \
+      const __m256i s_lohi = _mm256_unpackhi_epi16(s_16lo, s_16lo);            \
+                                                                               \
+      const __m256i s_hilo = _mm256_unpacklo_epi16(s_16hi, s_16hi);            \
+      const __m256i s_hihi = _mm256_unpackhi_epi16(s_16hi, s_16hi);            \
+                                                                               \
+      s[0] = _mm256_alignr_epi8(s_lohi, s_lolo, 2);                            \
+      s[1] = _mm256_alignr_epi8(s_lohi, s_lolo, 10);                           \
+      s[2] = _mm256_alignr_epi8(s_hilo, s_lohi, 2);                            \
+      s[3] = _mm256_alignr_epi8(s_hilo, s_lohi, 10);                           \
+      s[4] = _mm256_alignr_epi8(s_hihi, s_hilo, 2);                            \
+      s[5] = _mm256_alignr_epi8(s_hihi, s_hilo, 10);                           \
+                                                                               \
+      const __m256i res_lo = convolve_12taps(s, coeffs_h);                     \
+                                                                               \
+      __m256i res_32b_lo = _mm256_sra_epi32(                                   \
+          _mm256_add_epi32(res_lo, round_const_h12), round_shift_h12);         \
+                                                                               \
+      __m256i res_16b_lo = _mm256_packs_epi32(res_32b_lo, res_32b_lo);         \
+      _mm_store_si128((__m128i *)&im_block[i * im_stride],                     \
+                      _mm256_extracti128_si256(                                \
+                          _mm256_permute4x64_epi64(res_16b_lo, 0x88), 0));     \
+    }                                                                          \
+  }
+
+#define CONVOLVE_SR_VERTICAL_FILTER_12TAP                                      \
+  __m256i src_0 = _mm256_loadu_si256((__m256i *)(im_block + 0 * im_stride));   \
+  __m256i src_1 = _mm256_loadu_si256((__m256i *)(im_block + 1 * im_stride));   \
+  __m256i src_2 = _mm256_loadu_si256((__m256i *)(im_block + 2 * im_stride));   \
+  __m256i src_3 = _mm256_loadu_si256((__m256i *)(im_block + 3 * im_stride));   \
+  __m256i src_4 = _mm256_loadu_si256((__m256i *)(im_block + 4 * im_stride));   \
+  __m256i src_5 = _mm256_loadu_si256((__m256i *)(im_block + 5 * im_stride));   \
+  __m256i src_6 = _mm256_loadu_si256((__m256i *)(im_block + 6 * im_stride));   \
+  __m256i src_7 = _mm256_loadu_si256((__m256i *)(im_block + 7 * im_stride));   \
+  __m256i src_8 = _mm256_loadu_si256((__m256i *)(im_block + 8 * im_stride));   \
+  __m256i src_9 = _mm256_loadu_si256((__m256i *)(im_block + 9 * im_stride));   \
+                                                                               \
+  s[0] = _mm256_unpacklo_epi16(src_0, src_1);                                  \
+  s[1] = _mm256_unpacklo_epi16(src_2, src_3);                                  \
+  s[2] = _mm256_unpacklo_epi16(src_4, src_5);                                  \
+  s[3] = _mm256_unpacklo_epi16(src_6, src_7);                                  \
+  s[4] = _mm256_unpacklo_epi16(src_8, src_9);                                  \
+                                                                               \
+  s[6] = _mm256_unpackhi_epi16(src_0, src_1);                                  \
+  s[7] = _mm256_unpackhi_epi16(src_2, src_3);                                  \
+  s[8] = _mm256_unpackhi_epi16(src_4, src_5);                                  \
+  s[9] = _mm256_unpackhi_epi16(src_6, src_7);                                  \
+  s[10] = _mm256_unpackhi_epi16(src_8, src_9);                                 \
+                                                                               \
+  for (i = 0; i < h; i += 2) {                                                 \
+    const int16_t *data = &im_block[i * im_stride];                            \
+                                                                               \
+    const __m256i s6 = _mm256_loadu_si256((__m256i *)(data + 10 * im_stride)); \
+    const __m256i s7 = _mm256_loadu_si256((__m256i *)(data + 11 * im_stride)); \
+                                                                               \
+    s[5] = _mm256_unpacklo_epi16(s6, s7);                                      \
+    s[11] = _mm256_unpackhi_epi16(s6, s7);                                     \
+                                                                               \
+    __m256i res_a = convolve_12taps(s, coeffs_v);                              \
+    __m256i res_b = convolve_12taps(s + 6, coeffs_v);                          \
+                                                                               \
+    res_a =                                                                    \
+        _mm256_sra_epi32(_mm256_add_epi32(res_a, sum_round_v), sum_shift_v);   \
+    res_b =                                                                    \
+        _mm256_sra_epi32(_mm256_add_epi32(res_b, sum_round_v), sum_shift_v);   \
+                                                                               \
+    const __m256i res_a_round = _mm256_sra_epi32(                              \
+        _mm256_add_epi32(res_a, round_const_v), round_shift_v);                \
+    const __m256i res_b_round = _mm256_sra_epi32(                              \
+        _mm256_add_epi32(res_b, round_const_v), round_shift_v);                \
+                                                                               \
+    const __m256i res_16bit = _mm256_packs_epi32(res_a_round, res_b_round);    \
+    const __m256i res_8b = _mm256_packus_epi16(res_16bit, res_16bit);          \
+                                                                               \
+    const __m128i res_0 = _mm256_castsi256_si128(res_8b);                      \
+    const __m128i res_1 = _mm256_extracti128_si256(res_8b, 1);                 \
+                                                                               \
+    __m128i *const p_0 = (__m128i *)&dst[i * dst_stride + j];                  \
+    __m128i *const p_1 = (__m128i *)&dst[i * dst_stride + j + dst_stride];     \
+    if (w - j > 4) {                                                           \
+      _mm_storel_epi64(p_0, res_0);                                            \
+      _mm_storel_epi64(p_1, res_1);                                            \
+    } else if (w == 4) {                                                       \
+      xx_storel_32(p_0, res_0);                                                \
+      xx_storel_32(p_1, res_1);                                                \
+    } else {                                                                   \
+      *(uint16_t *)p_0 = (uint16_t)_mm_cvtsi128_si32(res_0);                   \
+      *(uint16_t *)p_1 = (uint16_t)_mm_cvtsi128_si32(res_1);                   \
+    }                                                                          \
+                                                                               \
+    s[0] = s[1];                                                               \
+    s[1] = s[2];                                                               \
+    s[2] = s[3];                                                               \
+    s[3] = s[4];                                                               \
+    s[4] = s[5];                                                               \
+                                                                               \
+    s[6] = s[7];                                                               \
+    s[7] = s[8];                                                               \
+    s[8] = s[9];                                                               \
+    s[9] = s[10];                                                              \
+    s[10] = s[11];                                                             \
+  }
+
+#define DIST_WTD_CONVOLVE_HORIZONTAL_FILTER_8TAP                        \
+  do {                                                                  \
+    for (i = 0; i < im_h; i += 2) {                                     \
+      __m256i data =                                                    \
+          _mm256_castsi128_si256(_mm_loadu_si128((__m128i *)src_h));    \
+      if (i + 1 < im_h)                                                 \
+        data = _mm256_inserti128_si256(                                 \
+            data, _mm_loadu_si128((__m128i *)(src_h + src_stride)), 1); \
+      src_h += (src_stride << 1);                                       \
+      __m256i res = convolve_lowbd_x(data, coeffs_x, filt);             \
+                                                                        \
+      res = _mm256_sra_epi16(_mm256_add_epi16(res, round_const_h),      \
+                             round_shift_h);                            \
+                                                                        \
+      _mm256_store_si256((__m256i *)&im_block[i * im_stride], res);     \
+    }                                                                   \
+  } while (0)
+
+#define DIST_WTD_CONVOLVE_VERTICAL_FILTER_8TAP                                 \
+  do {                                                                         \
+    __m256i s[8];                                                              \
+    __m256i s0 = _mm256_loadu_si256((__m256i *)(im_block + 0 * im_stride));    \
+    __m256i s1 = _mm256_loadu_si256((__m256i *)(im_block + 1 * im_stride));    \
+    __m256i s2 = _mm256_loadu_si256((__m256i *)(im_block + 2 * im_stride));    \
+    __m256i s3 = _mm256_loadu_si256((__m256i *)(im_block + 3 * im_stride));    \
+    __m256i s4 = _mm256_loadu_si256((__m256i *)(im_block + 4 * im_stride));    \
+    __m256i s5 = _mm256_loadu_si256((__m256i *)(im_block + 5 * im_stride));    \
+                                                                               \
+    s[0] = _mm256_unpacklo_epi16(s0, s1);                                      \
+    s[1] = _mm256_unpacklo_epi16(s2, s3);                                      \
+    s[2] = _mm256_unpacklo_epi16(s4, s5);                                      \
+                                                                               \
+    s[4] = _mm256_unpackhi_epi16(s0, s1);                                      \
+    s[5] = _mm256_unpackhi_epi16(s2, s3);                                      \
+    s[6] = _mm256_unpackhi_epi16(s4, s5);                                      \
+                                                                               \
+    for (i = 0; i < h; i += 2) {                                               \
+      const int16_t *data = &im_block[i * im_stride];                          \
+                                                                               \
+      const __m256i s6 =                                                       \
+          _mm256_loadu_si256((__m256i *)(data + 6 * im_stride));               \
+      const __m256i s7 =                                                       \
+          _mm256_loadu_si256((__m256i *)(data + 7 * im_stride));               \
+                                                                               \
+      s[3] = _mm256_unpacklo_epi16(s6, s7);                                    \
+      s[7] = _mm256_unpackhi_epi16(s6, s7);                                    \
+                                                                               \
+      const __m256i res_a = convolve(s, coeffs_y);                             \
+      const __m256i res_a_round = _mm256_sra_epi32(                            \
+          _mm256_add_epi32(res_a, round_const_v), round_shift_v);              \
+                                                                               \
+      if (w - j > 4) {                                                         \
+        const __m256i res_b = convolve(s + 4, coeffs_y);                       \
+        const __m256i res_b_round = _mm256_sra_epi32(                          \
+            _mm256_add_epi32(res_b, round_const_v), round_shift_v);            \
+        const __m256i res_16b = _mm256_packs_epi32(res_a_round, res_b_round);  \
+        const __m256i res_unsigned = _mm256_add_epi16(res_16b, offset_const);  \
+                                                                               \
+        if (do_average) {                                                      \
+          const __m256i data_ref_0 =                                           \
+              load_line2_avx2(&dst[i * dst_stride + j],                        \
+                              &dst[i * dst_stride + j + dst_stride]);          \
+          const __m256i comp_avg_res = comp_avg(&data_ref_0, &res_unsigned,    \
+                                                &wt, use_dist_wtd_comp_avg);   \
+                                                                               \
+          const __m256i round_result = convolve_rounding(                      \
+              &comp_avg_res, &offset_const, &rounding_const, rounding_shift);  \
+                                                                               \
+          const __m256i res_8 =                                                \
+              _mm256_packus_epi16(round_result, round_result);                 \
+          const __m128i res_0 = _mm256_castsi256_si128(res_8);                 \
+          const __m128i res_1 = _mm256_extracti128_si256(res_8, 1);            \
+                                                                               \
+          _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_0);    \
+          _mm_storel_epi64(                                                    \
+              (__m128i *)((&dst0[i * dst_stride0 + j + dst_stride0])), res_1); \
+        } else {                                                               \
+          const __m128i res_0 = _mm256_castsi256_si128(res_unsigned);          \
+          _mm_store_si128((__m128i *)(&dst[i * dst_stride + j]), res_0);       \
+                                                                               \
+          const __m128i res_1 = _mm256_extracti128_si256(res_unsigned, 1);     \
+          _mm_store_si128((__m128i *)(&dst[i * dst_stride + j + dst_stride]),  \
+                          res_1);                                              \
+        }                                                                      \
+      } else {                                                                 \
+        const __m256i res_16b = _mm256_packs_epi32(res_a_round, res_a_round);  \
+        const __m256i res_unsigned = _mm256_add_epi16(res_16b, offset_const);  \
+                                                                               \
+        if (do_average) {                                                      \
+          const __m256i data_ref_0 =                                           \
+              load_line2_avx2(&dst[i * dst_stride + j],                        \
+                              &dst[i * dst_stride + j + dst_stride]);          \
+                                                                               \
+          const __m256i comp_avg_res = comp_avg(&data_ref_0, &res_unsigned,    \
+                                                &wt, use_dist_wtd_comp_avg);   \
+                                                                               \
+          const __m256i round_result = convolve_rounding(                      \
+              &comp_avg_res, &offset_const, &rounding_const, rounding_shift);  \
+                                                                               \
+          const __m256i res_8 =                                                \
+              _mm256_packus_epi16(round_result, round_result);                 \
+          const __m128i res_0 = _mm256_castsi256_si128(res_8);                 \
+          const __m128i res_1 = _mm256_extracti128_si256(res_8, 1);            \
+                                                                               \
+          *(int *)(&dst0[i * dst_stride0 + j]) = _mm_cvtsi128_si32(res_0);     \
+          *(int *)(&dst0[i * dst_stride0 + j + dst_stride0]) =                 \
+              _mm_cvtsi128_si32(res_1);                                        \
+                                                                               \
+        } else {                                                               \
+          const __m128i res_0 = _mm256_castsi256_si128(res_unsigned);          \
+          _mm_store_si128((__m128i *)(&dst[i * dst_stride + j]), res_0);       \
+                                                                               \
+          const __m128i res_1 = _mm256_extracti128_si256(res_unsigned, 1);     \
+          _mm_store_si128((__m128i *)(&dst[i * dst_stride + j + dst_stride]),  \
+                          res_1);                                              \
+        }                                                                      \
+      }                                                                        \
+                                                                               \
+      s[0] = s[1];                                                             \
+      s[1] = s[2];                                                             \
+      s[2] = s[3];                                                             \
+                                                                               \
+      s[4] = s[5];                                                             \
+      s[5] = s[6];                                                             \
+      s[6] = s[7];                                                             \
+    }                                                                          \
+  } while (0)
+
+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);
+  const __m128i coeffs_8 = _mm_loadu_si128((__m128i *)filter);
+  const __m256i filter_coeffs = _mm256_broadcastsi128_si256(coeffs_8);
+
+  // right shift all filter co-efficients by 1 to reduce the bits required.
+  // This extra right shift will be taken care of at the end while rounding
+  // the result.
+  // Since all filter co-efficients are even, this change will not affect the
+  // end result
+  assert(_mm_test_all_zeros(_mm_and_si128(coeffs_8, _mm_set1_epi16(1)),
+                            _mm_set1_epi16((short)0xffff)));
+
+  const __m256i coeffs_1 = _mm256_srai_epi16(filter_coeffs, 1);
+
+  // coeffs 0 1 0 1 0 1 0 1
+  coeffs[0] = _mm256_shuffle_epi8(coeffs_1, _mm256_set1_epi16(0x0200u));
+  // coeffs 2 3 2 3 2 3 2 3
+  coeffs[1] = _mm256_shuffle_epi8(coeffs_1, _mm256_set1_epi16(0x0604u));
+  // coeffs 4 5 4 5 4 5 4 5
+  coeffs[2] = _mm256_shuffle_epi8(coeffs_1, _mm256_set1_epi16(0x0a08u));
+  // coeffs 6 7 6 7 6 7 6 7
+  coeffs[3] = _mm256_shuffle_epi8(coeffs_1, _mm256_set1_epi16(0x0e0cu));
+}
+
+static INLINE void prepare_coeffs_6t_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);
+  const __m128i coeffs_8 = _mm_loadu_si128((__m128i *)filter);
+  const __m256i filter_coeffs = _mm256_broadcastsi128_si256(coeffs_8);
+
+  // right shift all filter co-efficients by 1 to reduce the bits required.
+  // This extra right shift will be taken care of at the end while rounding
+  // the result.
+  // Since all filter co-efficients are even, this change will not affect the
+  // end result
+  assert(_mm_test_all_zeros(_mm_and_si128(coeffs_8, _mm_set1_epi16(1)),
+                            _mm_set1_epi16((int16_t)0xffff)));
+
+  const __m256i coeffs_1 = _mm256_srai_epi16(filter_coeffs, 1);
+
+  // coeffs 1 2 1 2 1 2 1 2
+  coeffs[0] = _mm256_shuffle_epi8(coeffs_1, _mm256_set1_epi16(0x0402u));
+  // coeffs 3 4 3 4 3 4 3 4
+  coeffs[1] = _mm256_shuffle_epi8(coeffs_1, _mm256_set1_epi16(0x0806u));
+  // coeffs 5 6 5 6 5 6 5 6
+  coeffs[2] = _mm256_shuffle_epi8(coeffs_1, _mm256_set1_epi16(0x0c0au));
+}
+
+static INLINE void prepare_coeffs_6t(
+    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);
+
+  const __m128i coeff_8 = _mm_loadu_si128((__m128i *)(filter + 1));
+  const __m256i coeff = _mm256_broadcastsi128_si256(coeff_8);
+
+  // coeffs 1 2 1 2 1 2 1 2
+  coeffs[0] = _mm256_shuffle_epi32(coeff, 0x00);
+  // coeffs 3 4 3 4 3 4 3 4
+  coeffs[1] = _mm256_shuffle_epi32(coeff, 0x55);
+  // coeffs 5 6 5 6 5 6 5 6
+  coeffs[2] = _mm256_shuffle_epi32(coeff, 0xaa);
+}
+
+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);
+
+  const __m128i coeff_8 = _mm_loadu_si128((__m128i *)filter);
+  const __m256i coeff = _mm256_broadcastsi128_si256(coeff_8);
+
+  // coeffs 0 1 0 1 0 1 0 1
+  coeffs[0] = _mm256_shuffle_epi32(coeff, 0x00);
+  // coeffs 2 3 2 3 2 3 2 3
+  coeffs[1] = _mm256_shuffle_epi32(coeff, 0x55);
+  // coeffs 4 5 4 5 4 5 4 5
+  coeffs[2] = _mm256_shuffle_epi32(coeff, 0xaa);
+  // coeffs 6 7 6 7 6 7 6 7
+  coeffs[3] = _mm256_shuffle_epi32(coeff, 0xff);
+}
+
+static INLINE void prepare_coeffs_12taps(
+    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);
+
+  __m128i coeff_8 = _mm_loadu_si128((__m128i *)filter);
+  __m256i coeff = _mm256_broadcastsi128_si256(coeff_8);
+
+  // coeffs 0 1 0 1 0 1 0 1
+  coeffs[0] = _mm256_shuffle_epi32(coeff, 0x00);
+  // coeffs 2 3 2 3 2 3 2 3
+  coeffs[1] = _mm256_shuffle_epi32(coeff, 0x55);
+  // coeffs 4 5 4 5 4 5 4 5
+  coeffs[2] = _mm256_shuffle_epi32(coeff, 0xaa);
+  // coeffs 6 7 6 7 6 7 6 7
+  coeffs[3] = _mm256_shuffle_epi32(coeff, 0xff);
+  // coeffs 8 9 10 11 0 0 0 0
+  coeff_8 = _mm_loadl_epi64((__m128i *)(filter + 8));
+  coeff = _mm256_broadcastq_epi64(coeff_8);
+  coeffs[4] = _mm256_shuffle_epi32(coeff, 0x00);  // coeffs 8 9 8 9 8 9 8 9
+  coeffs[5] = _mm256_shuffle_epi32(coeff, 0x55);  // coeffs 10 11 10 11.. 10 11
+}
+
+static INLINE __m256i convolve_lowbd(const __m256i *const s,
+                                     const __m256i *const coeffs) {
+  const __m256i res_01 = _mm256_maddubs_epi16(s[0], coeffs[0]);
+  const __m256i res_23 = _mm256_maddubs_epi16(s[1], coeffs[1]);
+  const __m256i res_45 = _mm256_maddubs_epi16(s[2], coeffs[2]);
+  const __m256i res_67 = _mm256_maddubs_epi16(s[3], coeffs[3]);
+
+  // order: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
+  const __m256i res = _mm256_add_epi16(_mm256_add_epi16(res_01, res_45),
+                                       _mm256_add_epi16(res_23, res_67));
+
+  return res;
+}
+
+static INLINE __m256i convolve_lowbd_6tap(const __m256i *const s,
+                                          const __m256i *const coeffs) {
+  const __m256i res_01 = _mm256_maddubs_epi16(s[0], coeffs[0]);
+  const __m256i res_23 = _mm256_maddubs_epi16(s[1], coeffs[1]);
+  const __m256i res_45 = _mm256_maddubs_epi16(s[2], coeffs[2]);
+
+  // order: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
+  const __m256i res =
+      _mm256_add_epi16(_mm256_add_epi16(res_01, res_45), res_23);
+
+  return res;
+}
+
+static INLINE __m256i convolve_lowbd_4tap(const __m256i *const s,
+                                          const __m256i *const coeffs) {
+  const __m256i res_23 = _mm256_maddubs_epi16(s[0], coeffs[0]);
+  const __m256i res_45 = _mm256_maddubs_epi16(s[1], coeffs[1]);
+
+  // order: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
+  const __m256i res = _mm256_add_epi16(res_45, res_23);
+
+  return res;
+}
+
+static INLINE __m256i convolve_6tap(const __m256i *const s,
+                                    const __m256i *const coeffs) {
+  const __m256i res_0 = _mm256_madd_epi16(s[0], coeffs[0]);
+  const __m256i res_1 = _mm256_madd_epi16(s[1], coeffs[1]);
+  const __m256i res_2 = _mm256_madd_epi16(s[2], coeffs[2]);
+
+  const __m256i res = _mm256_add_epi32(_mm256_add_epi32(res_0, res_1), res_2);
+
+  return res;
+}
+
+static INLINE __m256i convolve_12taps(const __m256i *const s,
+                                      const __m256i *const coeffs) {
+  const __m256i res_0 = _mm256_madd_epi16(s[0], coeffs[0]);
+  const __m256i res_1 = _mm256_madd_epi16(s[1], coeffs[1]);
+  const __m256i res_2 = _mm256_madd_epi16(s[2], coeffs[2]);
+  const __m256i res_3 = _mm256_madd_epi16(s[3], coeffs[3]);
+  const __m256i res_4 = _mm256_madd_epi16(s[4], coeffs[4]);
+  const __m256i res_5 = _mm256_madd_epi16(s[5], coeffs[5]);
+
+  const __m256i res1 = _mm256_add_epi32(_mm256_add_epi32(res_0, res_1),
+                                        _mm256_add_epi32(res_2, res_3));
+  const __m256i res = _mm256_add_epi32(_mm256_add_epi32(res_4, res_5), res1);
+
+  return res;
+}
+
+static INLINE __m256i convolve(const __m256i *const s,
+                               const __m256i *const coeffs) {
+  const __m256i res_0 = _mm256_madd_epi16(s[0], coeffs[0]);
+  const __m256i res_1 = _mm256_madd_epi16(s[1], coeffs[1]);
+  const __m256i res_2 = _mm256_madd_epi16(s[2], coeffs[2]);
+  const __m256i res_3 = _mm256_madd_epi16(s[3], coeffs[3]);
+
+  const __m256i res = _mm256_add_epi32(_mm256_add_epi32(res_0, res_1),
+                                       _mm256_add_epi32(res_2, res_3));
+
+  return res;
+}
+
+static INLINE __m256i convolve_4tap(const __m256i *const s,
+                                    const __m256i *const coeffs) {
+  const __m256i res_1 = _mm256_madd_epi16(s[0], coeffs[0]);
+  const __m256i res_2 = _mm256_madd_epi16(s[1], coeffs[1]);
+
+  const __m256i res = _mm256_add_epi32(res_1, res_2);
+  return res;
+}
+
+static INLINE __m256i convolve_lowbd_x(const __m256i data,
+                                       const __m256i *const coeffs,
+                                       const __m256i *const filt) {
+  __m256i s[4];
+
+  s[0] = _mm256_shuffle_epi8(data, filt[0]);
+  s[1] = _mm256_shuffle_epi8(data, filt[1]);
+  s[2] = _mm256_shuffle_epi8(data, filt[2]);
+  s[3] = _mm256_shuffle_epi8(data, filt[3]);
+
+  return convolve_lowbd(s, coeffs);
+}
+
+static INLINE __m256i convolve_lowbd_x_6tap(const __m256i data,
+                                            const __m256i *const coeffs,
+                                            const __m256i *const filt) {
+  __m256i s[4];
+
+  s[0] = _mm256_shuffle_epi8(data, filt[0]);
+  s[1] = _mm256_shuffle_epi8(data, filt[1]);
+  s[2] = _mm256_shuffle_epi8(data, filt[2]);
+
+  return convolve_lowbd_6tap(s, coeffs);
+}
+
+static INLINE __m256i convolve_lowbd_x_4tap(const __m256i data,
+                                            const __m256i *const coeffs,
+                                            const __m256i *const filt) {
+  __m256i s[2];
+
+  s[0] = _mm256_shuffle_epi8(data, filt[0]);
+  s[1] = _mm256_shuffle_epi8(data, filt[1]);
+
+  return convolve_lowbd_4tap(s, coeffs);
+}
+
+static INLINE void add_store_aligned_256(CONV_BUF_TYPE *const dst,
+                                         const __m256i *const res,
+                                         const int do_average) {
+  __m256i d;
+  if (do_average) {
+    d = _mm256_load_si256((__m256i *)dst);
+    d = _mm256_add_epi32(d, *res);
+    d = _mm256_srai_epi32(d, 1);
+  } else {
+    d = *res;
+  }
+  _mm256_store_si256((__m256i *)dst, d);
+}
+
+static INLINE __m256i comp_avg(const __m256i *const data_ref_0,
+                               const __m256i *const res_unsigned,
+                               const __m256i *const wt,
+                               const int use_dist_wtd_comp_avg) {
+  __m256i res;
+  if (use_dist_wtd_comp_avg) {
+    const __m256i data_lo = _mm256_unpacklo_epi16(*data_ref_0, *res_unsigned);
+    const __m256i data_hi = _mm256_unpackhi_epi16(*data_ref_0, *res_unsigned);
+
+    const __m256i wt_res_lo = _mm256_madd_epi16(data_lo, *wt);
+    const __m256i wt_res_hi = _mm256_madd_epi16(data_hi, *wt);
+
+    const __m256i res_lo = _mm256_srai_epi32(wt_res_lo, DIST_PRECISION_BITS);
+    const __m256i res_hi = _mm256_srai_epi32(wt_res_hi, DIST_PRECISION_BITS);
+
+    res = _mm256_packs_epi32(res_lo, res_hi);
+  } else {
+    const __m256i wt_res = _mm256_add_epi16(*data_ref_0, *res_unsigned);
+    res = _mm256_srai_epi16(wt_res, 1);
+  }
+  return res;
+}
+
+static INLINE __m256i convolve_rounding(const __m256i *const res_unsigned,
+                                        const __m256i *const offset_const,
+                                        const __m256i *const round_const,
+                                        const int round_shift) {
+  const __m256i res_signed = _mm256_sub_epi16(*res_unsigned, *offset_const);
+  const __m256i res_round = _mm256_srai_epi16(
+      _mm256_add_epi16(res_signed, *round_const), round_shift);
+  return res_round;
+}
+
+static INLINE __m256i highbd_comp_avg(const __m256i *const data_ref_0,
+                                      const __m256i *const res_unsigned,
+                                      const __m256i *const wt0,
+                                      const __m256i *const wt1,
+                                      const int use_dist_wtd_comp_avg) {
+  __m256i res;
+  if (use_dist_wtd_comp_avg) {
+    const __m256i wt0_res = _mm256_mullo_epi32(*data_ref_0, *wt0);
+    const __m256i wt1_res = _mm256_mullo_epi32(*res_unsigned, *wt1);
+    const __m256i wt_res = _mm256_add_epi32(wt0_res, wt1_res);
+    res = _mm256_srai_epi32(wt_res, DIST_PRECISION_BITS);
+  } else {
+    const __m256i wt_res = _mm256_add_epi32(*data_ref_0, *res_unsigned);
+    res = _mm256_srai_epi32(wt_res, 1);
+  }
+  return res;
+}
+
+static INLINE __m256i highbd_convolve_rounding(
+    const __m256i *const res_unsigned, const __m256i *const offset_const,
+    const __m256i *const round_const, const int round_shift) {
+  const __m256i res_signed = _mm256_sub_epi32(*res_unsigned, *offset_const);
+  const __m256i res_round = _mm256_srai_epi32(
+      _mm256_add_epi32(res_signed, *round_const), round_shift);
+
+  return res_round;
+}
+
+#endif  // AOM_AOM_DSP_X86_CONVOLVE_AVX2_H_
diff --git a/third_party/aom/aom_dsp/x86/convolve_common_intrin.h b/third_party/aom/aom_dsp/x86/convolve_common_intrin.h
new file mode 100644
index 0000000000..9e8662af46
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/convolve_common_intrin.h
@@ -0,0 +1,102 @@
+/*
+ * 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_AOM_DSP_X86_CONVOLVE_COMMON_INTRIN_H_
+#define AOM_AOM_DSP_X86_CONVOLVE_COMMON_INTRIN_H_
+
+// Note:
+//  This header file should be put below any x86 intrinsics head file
+
+static INLINE void add_store(CONV_BUF_TYPE *const dst, const __m128i *const res,
+                             const int do_average) {
+  __m128i d;
+  if (do_average) {
+    d = _mm_load_si128((__m128i *)dst);
+    d = _mm_add_epi32(d, *res);
+    d = _mm_srai_epi32(d, 1);
+  } else {
+    d = *res;
+  }
+  _mm_store_si128((__m128i *)dst, d);
+}
+
+static INLINE void prepare_coeffs_12tap(const InterpFilterParams *filter_params,
+                                        int subpel_q4,
+                                        __m128i *coeffs /* [6] */) {
+  const int16_t *const y_filter = av1_get_interp_filter_subpel_kernel(
+      filter_params, subpel_q4 & SUBPEL_MASK);
+
+  __m128i coeffs_y = _mm_loadu_si128((__m128i *)y_filter);
+
+  coeffs[0] = _mm_shuffle_epi32(coeffs_y, 0);    // coeffs 0 1 0 1 0 1 0 1
+  coeffs[1] = _mm_shuffle_epi32(coeffs_y, 85);   // coeffs 2 3 2 3 2 3 2 3
+  coeffs[2] = _mm_shuffle_epi32(coeffs_y, 170);  // coeffs 4 5 4 5 4 5 4 5
+  coeffs[3] = _mm_shuffle_epi32(coeffs_y, 255);  // coeffs 6 7 6 7 6 7 6 7
+
+  coeffs_y = _mm_loadl_epi64((__m128i *)(y_filter + 8));
+
+  coeffs[4] = _mm_shuffle_epi32(coeffs_y, 0);  // coeffs 8 9 8 9 8 9 8 9
+  coeffs[5] =
+      _mm_shuffle_epi32(coeffs_y, 85);  // coeffs 10 11 10 11 10 11 10 11
+}
+
+static INLINE __m128i convolve_12tap(const __m128i *s, const __m128i *coeffs) {
+  const __m128i d0 = _mm_madd_epi16(s[0], coeffs[0]);
+  const __m128i d1 = _mm_madd_epi16(s[1], coeffs[1]);
+  const __m128i d2 = _mm_madd_epi16(s[2], coeffs[2]);
+  const __m128i d3 = _mm_madd_epi16(s[3], coeffs[3]);
+  const __m128i d4 = _mm_madd_epi16(s[4], coeffs[4]);
+  const __m128i d5 = _mm_madd_epi16(s[5], coeffs[5]);
+  const __m128i d_0123 =
+      _mm_add_epi32(_mm_add_epi32(d0, d1), _mm_add_epi32(d2, d3));
+  const __m128i d = _mm_add_epi32(_mm_add_epi32(d4, d5), d_0123);
+  return d;
+}
+
+static INLINE __m128i convolve_lo_x_12tap(const __m128i *s,
+                                          const __m128i *coeffs,
+                                          const __m128i zero) {
+  __m128i ss[6];
+  ss[0] = _mm_unpacklo_epi8(s[0], zero);  //  0  1  1  2  2  3  3  4
+  ss[1] = _mm_unpacklo_epi8(s[1], zero);  //  2  3  3  4  4  5  5  6
+  ss[2] = _mm_unpacklo_epi8(s[2], zero);  //  4  5  5  6  6  7  7  8
+  ss[3] = _mm_unpacklo_epi8(s[3], zero);  //  6  7  7  8  8  9  9 10
+  ss[4] = _mm_unpackhi_epi8(s[2], zero);  //  8  9  9 10 10 11 11 12
+  ss[5] = _mm_unpackhi_epi8(s[3], zero);  // 10 11 11 12 12 13 13 14
+  return convolve_12tap(ss, coeffs);
+}
+
+static INLINE __m128i convolve_lo_y_12tap(const __m128i *s,
+                                          const __m128i *coeffs) {
+  __m128i ss[6];
+  const __m128i zero = _mm_setzero_si128();
+  ss[0] = _mm_unpacklo_epi8(s[0], zero);
+  ss[1] = _mm_unpacklo_epi8(s[2], zero);
+  ss[2] = _mm_unpacklo_epi8(s[4], zero);
+  ss[3] = _mm_unpacklo_epi8(s[6], zero);
+  ss[4] = _mm_unpacklo_epi8(s[8], zero);
+  ss[5] = _mm_unpacklo_epi8(s[10], zero);
+  return convolve_12tap(ss, coeffs);
+}
+
+static INLINE __m128i convolve_hi_y_12tap(const __m128i *s,
+                                          const __m128i *coeffs) {
+  __m128i ss[6];
+  const __m128i zero = _mm_setzero_si128();
+  ss[0] = _mm_unpackhi_epi8(s[0], zero);
+  ss[1] = _mm_unpackhi_epi8(s[2], zero);
+  ss[2] = _mm_unpackhi_epi8(s[4], zero);
+  ss[3] = _mm_unpackhi_epi8(s[6], zero);
+  ss[4] = _mm_unpackhi_epi8(s[8], zero);
+  ss[5] = _mm_unpackhi_epi8(s[10], zero);
+  return convolve_12tap(ss, coeffs);
+}
+#endif  // AOM_AOM_DSP_X86_CONVOLVE_COMMON_INTRIN_H_
diff --git a/third_party/aom/aom_dsp/x86/convolve_sse2.h b/third_party/aom/aom_dsp/x86/convolve_sse2.h
new file mode 100644
index 0000000000..36b7d62b98
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/convolve_sse2.h
@@ -0,0 +1,122 @@
+/*
+ * 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_AOM_DSP_X86_CONVOLVE_SSE2_H_
+#define AOM_AOM_DSP_X86_CONVOLVE_SSE2_H_
+
+#include "config/aom_scale_rtcd.h"
+
+// Note:
+//  This header file should be put below any x86 intrinsics head file
+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);
+  const __m128i coeff = _mm_loadu_si128((__m128i *)filter);
+
+  // coeffs 0 1 0 1 0 1 0 1
+  coeffs[0] = _mm_shuffle_epi32(coeff, 0x00);
+  // coeffs 2 3 2 3 2 3 2 3
+  coeffs[1] = _mm_shuffle_epi32(coeff, 0x55);
+  // coeffs 4 5 4 5 4 5 4 5
+  coeffs[2] = _mm_shuffle_epi32(coeff, 0xaa);
+  // coeffs 6 7 6 7 6 7 6 7
+  coeffs[3] = _mm_shuffle_epi32(coeff, 0xff);
+}
+
+static INLINE __m128i convolve(const __m128i *const s,
+                               const __m128i *const coeffs) {
+  const __m128i res_0 = _mm_madd_epi16(s[0], coeffs[0]);
+  const __m128i res_1 = _mm_madd_epi16(s[1], coeffs[1]);
+  const __m128i res_2 = _mm_madd_epi16(s[2], coeffs[2]);
+  const __m128i res_3 = _mm_madd_epi16(s[3], coeffs[3]);
+
+  const __m128i res =
+      _mm_add_epi32(_mm_add_epi32(res_0, res_1), _mm_add_epi32(res_2, res_3));
+
+  return res;
+}
+
+static INLINE __m128i convolve_lo_x(const __m128i *const s,
+                                    const __m128i *const coeffs) {
+  __m128i ss[4];
+  ss[0] = _mm_unpacklo_epi8(s[0], _mm_setzero_si128());
+  ss[1] = _mm_unpacklo_epi8(s[1], _mm_setzero_si128());
+  ss[2] = _mm_unpacklo_epi8(s[2], _mm_setzero_si128());
+  ss[3] = _mm_unpacklo_epi8(s[3], _mm_setzero_si128());
+  return convolve(ss, coeffs);
+}
+
+static INLINE __m128i convolve_lo_y(const __m128i *const s,
+                                    const __m128i *const coeffs) {
+  __m128i ss[4];
+  ss[0] = _mm_unpacklo_epi8(s[0], _mm_setzero_si128());
+  ss[1] = _mm_unpacklo_epi8(s[2], _mm_setzero_si128());
+  ss[2] = _mm_unpacklo_epi8(s[4], _mm_setzero_si128());
+  ss[3] = _mm_unpacklo_epi8(s[6], _mm_setzero_si128());
+  return convolve(ss, coeffs);
+}
+
+static INLINE __m128i convolve_hi_y(const __m128i *const s,
+                                    const __m128i *const coeffs) {
+  __m128i ss[4];
+  ss[0] = _mm_unpackhi_epi8(s[0], _mm_setzero_si128());
+  ss[1] = _mm_unpackhi_epi8(s[2], _mm_setzero_si128());
+  ss[2] = _mm_unpackhi_epi8(s[4], _mm_setzero_si128());
+  ss[3] = _mm_unpackhi_epi8(s[6], _mm_setzero_si128());
+  return convolve(ss, coeffs);
+}
+
+static INLINE __m128i comp_avg(const __m128i *const data_ref_0,
+                               const __m128i *const res_unsigned,
+                               const __m128i *const wt,
+                               const int use_dist_wtd_avg) {
+  __m128i res;
+  if (use_dist_wtd_avg) {
+    const __m128i data_lo = _mm_unpacklo_epi16(*data_ref_0, *res_unsigned);
+    const __m128i data_hi = _mm_unpackhi_epi16(*data_ref_0, *res_unsigned);
+
+    const __m128i wt_res_lo = _mm_madd_epi16(data_lo, *wt);
+    const __m128i wt_res_hi = _mm_madd_epi16(data_hi, *wt);
+
+    const __m128i res_lo = _mm_srai_epi32(wt_res_lo, DIST_PRECISION_BITS);
+    const __m128i res_hi = _mm_srai_epi32(wt_res_hi, DIST_PRECISION_BITS);
+
+    res = _mm_packs_epi32(res_lo, res_hi);
+  } else {
+    const __m128i wt_res = _mm_add_epi16(*data_ref_0, *res_unsigned);
+    res = _mm_srai_epi16(wt_res, 1);
+  }
+  return res;
+}
+
+static INLINE __m128i convolve_rounding(const __m128i *const res_unsigned,
+                                        const __m128i *const offset_const,
+                                        const __m128i *const round_const,
+                                        const int round_shift) {
+  const __m128i res_signed = _mm_sub_epi16(*res_unsigned, *offset_const);
+  const __m128i res_round =
+      _mm_srai_epi16(_mm_add_epi16(res_signed, *round_const), round_shift);
+  return res_round;
+}
+
+static INLINE __m128i highbd_convolve_rounding_sse2(
+    const __m128i *const res_unsigned, const __m128i *const offset_const,
+    const __m128i *const round_const, const int round_shift) {
+  const __m128i res_signed = _mm_sub_epi32(*res_unsigned, *offset_const);
+  const __m128i res_round =
+      _mm_srai_epi32(_mm_add_epi32(res_signed, *round_const), round_shift);
+
+  return res_round;
+}
+
+#endif  // AOM_AOM_DSP_X86_CONVOLVE_SSE2_H_
diff --git a/third_party/aom/aom_dsp/x86/convolve_sse4_1.h b/third_party/aom/aom_dsp/x86/convolve_sse4_1.h
new file mode 100644
index 0000000000..b1a3bb4664
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/convolve_sse4_1.h
@@ -0,0 +1,53 @@
+/*
+ * 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_AOM_DSP_X86_CONVOLVE_SSE4_1_H_
+#define AOM_AOM_DSP_X86_CONVOLVE_SSE4_1_H_
+
+// Note:
+//  This header file should be put below any x86 intrinsics head file
+
+static INLINE void mult_add_store(CONV_BUF_TYPE *const dst,
+                                  const __m128i *const res,
+                                  const __m128i *const wt0,
+                                  const __m128i *const wt1,
+                                  const int do_average) {
+  __m128i d;
+  if (do_average) {
+    d = _mm_load_si128((__m128i *)dst);
+    d = _mm_add_epi32(_mm_mullo_epi32(d, *wt0), _mm_mullo_epi32(*res, *wt1));
+    d = _mm_srai_epi32(d, DIST_PRECISION_BITS);
+  } else {
+    d = *res;
+  }
+  _mm_store_si128((__m128i *)dst, d);
+}
+
+static INLINE __m128i highbd_comp_avg_sse4_1(const __m128i *const data_ref_0,
+                                             const __m128i *const res_unsigned,
+                                             const __m128i *const wt0,
+                                             const __m128i *const wt1,
+                                             const int use_dist_wtd_avg) {
+  __m128i res;
+  if (use_dist_wtd_avg) {
+    const __m128i wt0_res = _mm_mullo_epi32(*data_ref_0, *wt0);
+    const __m128i wt1_res = _mm_mullo_epi32(*res_unsigned, *wt1);
+
+    const __m128i wt_res = _mm_add_epi32(wt0_res, wt1_res);
+    res = _mm_srai_epi32(wt_res, DIST_PRECISION_BITS);
+  } else {
+    const __m128i wt_res = _mm_add_epi32(*data_ref_0, *res_unsigned);
+    res = _mm_srai_epi32(wt_res, 1);
+  }
+  return res;
+}
+
+#endif  // AOM_AOM_DSP_X86_CONVOLVE_SSE4_1_H_
diff --git a/third_party/aom/aom_dsp/x86/convolve_ssse3.h b/third_party/aom/aom_dsp/x86/convolve_ssse3.h
new file mode 100644
index 0000000000..b1abead146
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/convolve_ssse3.h
@@ -0,0 +1,50 @@
+/*
+ * Copyright (c) 2021, 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_AOM_DSP_X86_CONVOLVE_SSSE3_H_
+#define AOM_AOM_DSP_X86_CONVOLVE_SSSE3_H_
+
+#include <tmmintrin.h>  // SSSE3
+
+static INLINE void shuffle_filter_ssse3(const int16_t *const filter,
+                                        __m128i *const f) {
+  const __m128i f_values = _mm_load_si128((const __m128i *)filter);
+  // pack and duplicate the filter values
+  f[0] = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0200u));
+  f[1] = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0604u));
+  f[2] = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0a08u));
+  f[3] = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0e0cu));
+}
+
+static INLINE __m128i convolve8_8_ssse3(const __m128i *const s,
+                                        const __m128i *const f) {
+  // multiply 2 adjacent elements with the filter and add the result
+  const __m128i k_64 = _mm_set1_epi16(1 << 6);
+  const __m128i x0 = _mm_maddubs_epi16(s[0], f[0]);
+  const __m128i x1 = _mm_maddubs_epi16(s[1], f[1]);
+  const __m128i x2 = _mm_maddubs_epi16(s[2], f[2]);
+  const __m128i x3 = _mm_maddubs_epi16(s[3], f[3]);
+  __m128i sum1, sum2;
+
+  // sum the results together, saturating only on the final step
+  // adding x0 with x2 and x1 with x3 is the only order that prevents
+  // outranges for all filters
+  sum1 = _mm_add_epi16(x0, x2);
+  sum2 = _mm_add_epi16(x1, x3);
+  // add the rounding offset early to avoid another saturated add
+  sum1 = _mm_add_epi16(sum1, k_64);
+  sum1 = _mm_adds_epi16(sum1, sum2);
+  // shift by 7 bit each 16 bit
+  sum1 = _mm_srai_epi16(sum1, 7);
+  return sum1;
+}
+
+#endif  // AOM_AOM_DSP_X86_CONVOLVE_SSSE3_H_
diff --git a/third_party/aom/aom_dsp/x86/fft_avx2.c b/third_party/aom/aom_dsp/x86/fft_avx2.c
new file mode 100644
index 0000000000..3f5a9bbeff
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/fft_avx2.c
@@ -0,0 +1,74 @@
+/*
+ * 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"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/fft_common.h"
+
+extern void aom_transpose_float_sse2(const float *A, float *B, int n);
+extern void aom_fft_unpack_2d_output_sse2(const float *col_fft, float *output,
+                                          int n);
+
+// Generate the 1d forward transforms for float using _mm256
+GEN_FFT_8(static INLINE void, avx2, float, __m256, _mm256_load_ps,
+          _mm256_store_ps, _mm256_set1_ps, _mm256_add_ps, _mm256_sub_ps,
+          _mm256_mul_ps)
+GEN_FFT_16(static INLINE void, avx2, float, __m256, _mm256_load_ps,
+           _mm256_store_ps, _mm256_set1_ps, _mm256_add_ps, _mm256_sub_ps,
+           _mm256_mul_ps)
+GEN_FFT_32(static INLINE void, avx2, float, __m256, _mm256_load_ps,
+           _mm256_store_ps, _mm256_set1_ps, _mm256_add_ps, _mm256_sub_ps,
+           _mm256_mul_ps)
+
+void aom_fft8x8_float_avx2(const float *input, float *temp, float *output) {
+  aom_fft_2d_gen(input, temp, output, 8, aom_fft1d_8_avx2,
+                 aom_transpose_float_sse2, aom_fft_unpack_2d_output_sse2, 8);
+}
+
+void aom_fft16x16_float_avx2(const float *input, float *temp, float *output) {
+  aom_fft_2d_gen(input, temp, output, 16, aom_fft1d_16_avx2,
+                 aom_transpose_float_sse2, aom_fft_unpack_2d_output_sse2, 8);
+}
+
+void aom_fft32x32_float_avx2(const float *input, float *temp, float *output) {
+  aom_fft_2d_gen(input, temp, output, 32, aom_fft1d_32_avx2,
+                 aom_transpose_float_sse2, aom_fft_unpack_2d_output_sse2, 8);
+}
+
+// Generate the 1d inverse transforms for float using _mm256
+GEN_IFFT_8(static INLINE void, avx2, float, __m256, _mm256_load_ps,
+           _mm256_store_ps, _mm256_set1_ps, _mm256_add_ps, _mm256_sub_ps,
+           _mm256_mul_ps)
+GEN_IFFT_16(static INLINE void, avx2, float, __m256, _mm256_load_ps,
+            _mm256_store_ps, _mm256_set1_ps, _mm256_add_ps, _mm256_sub_ps,
+            _mm256_mul_ps)
+GEN_IFFT_32(static INLINE void, avx2, float, __m256, _mm256_load_ps,
+            _mm256_store_ps, _mm256_set1_ps, _mm256_add_ps, _mm256_sub_ps,
+            _mm256_mul_ps)
+
+void aom_ifft8x8_float_avx2(const float *input, float *temp, float *output) {
+  aom_ifft_2d_gen(input, temp, output, 8, aom_fft1d_8_float, aom_fft1d_8_avx2,
+                  aom_ifft1d_8_avx2, aom_transpose_float_sse2, 8);
+}
+
+void aom_ifft16x16_float_avx2(const float *input, float *temp, float *output) {
+  aom_ifft_2d_gen(input, temp, output, 16, aom_fft1d_16_float,
+                  aom_fft1d_16_avx2, aom_ifft1d_16_avx2,
+                  aom_transpose_float_sse2, 8);
+}
+
+void aom_ifft32x32_float_avx2(const float *input, float *temp, float *output) {
+  aom_ifft_2d_gen(input, temp, output, 32, aom_fft1d_32_float,
+                  aom_fft1d_32_avx2, aom_ifft1d_32_avx2,
+                  aom_transpose_float_sse2, 8);
+}
diff --git a/third_party/aom/aom_dsp/x86/fft_sse2.c b/third_party/aom/aom_dsp/x86/fft_sse2.c
new file mode 100644
index 0000000000..bdd235bcd3
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/fft_sse2.c
@@ -0,0 +1,173 @@
+/*
+ * 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
+s * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <xmmintrin.h>
+
+#include "config/aom_dsp_rtcd.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/fft_common.h"
+
+static INLINE void transpose4x4(const float *A, float *B, const int lda,
+                                const int ldb) {
+  __m128 row1 = _mm_load_ps(&A[0 * lda]);
+  __m128 row2 = _mm_load_ps(&A[1 * lda]);
+  __m128 row3 = _mm_load_ps(&A[2 * lda]);
+  __m128 row4 = _mm_load_ps(&A[3 * lda]);
+  _MM_TRANSPOSE4_PS(row1, row2, row3, row4);
+  _mm_store_ps(&B[0 * ldb], row1);
+  _mm_store_ps(&B[1 * ldb], row2);
+  _mm_store_ps(&B[2 * ldb], row3);
+  _mm_store_ps(&B[3 * ldb], row4);
+}
+
+// Referenced by fft_avx2.c.
+void aom_transpose_float_sse2(const float *A, float *B, int n);
+
+void aom_transpose_float_sse2(const float *A, float *B, int n) {
+  for (int y = 0; y < n; y += 4) {
+    for (int x = 0; x < n; x += 4) {
+      transpose4x4(A + y * n + x, B + x * n + y, n, n);
+    }
+  }
+}
+
+// Referenced by fft_avx2.c.
+void aom_fft_unpack_2d_output_sse2(const float *packed, float *output, int n);
+
+void aom_fft_unpack_2d_output_sse2(const float *packed, float *output, int n) {
+  const int n2 = n / 2;
+  output[0] = packed[0];
+  output[1] = 0;
+  output[2 * (n2 * n)] = packed[n2 * n];
+  output[2 * (n2 * n) + 1] = 0;
+
+  output[2 * n2] = packed[n2];
+  output[2 * n2 + 1] = 0;
+  output[2 * (n2 * n + n2)] = packed[n2 * n + n2];
+  output[2 * (n2 * n + n2) + 1] = 0;
+
+  for (int c = 1; c < n2; ++c) {
+    output[2 * (0 * n + c)] = packed[c];
+    output[2 * (0 * n + c) + 1] = packed[c + n2];
+    output[2 * (n2 * n + c) + 0] = packed[n2 * n + c];
+    output[2 * (n2 * n + c) + 1] = packed[n2 * n + c + n2];
+  }
+  for (int r = 1; r < n2; ++r) {
+    output[2 * (r * n + 0)] = packed[r * n];
+    output[2 * (r * n + 0) + 1] = packed[(r + n2) * n];
+    output[2 * (r * n + n2) + 0] = packed[r * n + n2];
+    output[2 * (r * n + n2) + 1] = packed[(r + n2) * n + n2];
+
+    for (int c = 1; c < AOMMIN(n2, 4); ++c) {
+      output[2 * (r * n + c)] =
+          packed[r * n + c] - packed[(r + n2) * n + c + n2];
+      output[2 * (r * n + c) + 1] =
+          packed[(r + n2) * n + c] + packed[r * n + c + n2];
+    }
+
+    for (int c = 4; c < n2; c += 4) {
+      __m128 real1 = _mm_load_ps(packed + r * n + c);
+      __m128 real2 = _mm_load_ps(packed + (r + n2) * n + c + n2);
+      __m128 imag1 = _mm_load_ps(packed + (r + n2) * n + c);
+      __m128 imag2 = _mm_load_ps(packed + r * n + c + n2);
+      real1 = _mm_sub_ps(real1, real2);
+      imag1 = _mm_add_ps(imag1, imag2);
+      _mm_store_ps(output + 2 * (r * n + c), _mm_unpacklo_ps(real1, imag1));
+      _mm_store_ps(output + 2 * (r * n + c + 2), _mm_unpackhi_ps(real1, imag1));
+    }
+
+    int r2 = r + n2;
+    int r3 = n - r2;
+    output[2 * (r2 * n + 0)] = packed[r3 * n];
+    output[2 * (r2 * n + 0) + 1] = -packed[(r3 + n2) * n];
+    output[2 * (r2 * n + n2)] = packed[r3 * n + n2];
+    output[2 * (r2 * n + n2) + 1] = -packed[(r3 + n2) * n + n2];
+    for (int c = 1; c < AOMMIN(4, n2); ++c) {
+      output[2 * (r2 * n + c)] =
+          packed[r3 * n + c] + packed[(r3 + n2) * n + c + n2];
+      output[2 * (r2 * n + c) + 1] =
+          -packed[(r3 + n2) * n + c] + packed[r3 * n + c + n2];
+    }
+    for (int c = 4; c < n2; c += 4) {
+      __m128 real1 = _mm_load_ps(packed + r3 * n + c);
+      __m128 real2 = _mm_load_ps(packed + (r3 + n2) * n + c + n2);
+      __m128 imag1 = _mm_load_ps(packed + (r3 + n2) * n + c);
+      __m128 imag2 = _mm_load_ps(packed + r3 * n + c + n2);
+      real1 = _mm_add_ps(real1, real2);
+      imag1 = _mm_sub_ps(imag2, imag1);
+      _mm_store_ps(output + 2 * (r2 * n + c), _mm_unpacklo_ps(real1, imag1));
+      _mm_store_ps(output + 2 * (r2 * n + c + 2),
+                   _mm_unpackhi_ps(real1, imag1));
+    }
+  }
+}
+
+// Generate definitions for 1d transforms using float and __mm128
+GEN_FFT_4(static INLINE void, sse2, float, __m128, _mm_load_ps, _mm_store_ps,
+          _mm_set1_ps, _mm_add_ps, _mm_sub_ps)
+GEN_FFT_8(static INLINE void, sse2, float, __m128, _mm_load_ps, _mm_store_ps,
+          _mm_set1_ps, _mm_add_ps, _mm_sub_ps, _mm_mul_ps)
+GEN_FFT_16(static INLINE void, sse2, float, __m128, _mm_load_ps, _mm_store_ps,
+           _mm_set1_ps, _mm_add_ps, _mm_sub_ps, _mm_mul_ps)
+GEN_FFT_32(static INLINE void, sse2, float, __m128, _mm_load_ps, _mm_store_ps,
+           _mm_set1_ps, _mm_add_ps, _mm_sub_ps, _mm_mul_ps)
+
+void aom_fft4x4_float_sse2(const float *input, float *temp, float *output) {
+  aom_fft_2d_gen(input, temp, output, 4, aom_fft1d_4_sse2,
+                 aom_transpose_float_sse2, aom_fft_unpack_2d_output_sse2, 4);
+}
+
+void aom_fft8x8_float_sse2(const float *input, float *temp, float *output) {
+  aom_fft_2d_gen(input, temp, output, 8, aom_fft1d_8_sse2,
+                 aom_transpose_float_sse2, aom_fft_unpack_2d_output_sse2, 4);
+}
+
+void aom_fft16x16_float_sse2(const float *input, float *temp, float *output) {
+  aom_fft_2d_gen(input, temp, output, 16, aom_fft1d_16_sse2,
+                 aom_transpose_float_sse2, aom_fft_unpack_2d_output_sse2, 4);
+}
+
+void aom_fft32x32_float_sse2(const float *input, float *temp, float *output) {
+  aom_fft_2d_gen(input, temp, output, 32, aom_fft1d_32_sse2,
+                 aom_transpose_float_sse2, aom_fft_unpack_2d_output_sse2, 4);
+}
+
+// Generate definitions for 1d inverse transforms using float and mm128
+GEN_IFFT_4(static INLINE void, sse2, float, __m128, _mm_load_ps, _mm_store_ps,
+           _mm_set1_ps, _mm_add_ps, _mm_sub_ps)
+GEN_IFFT_8(static INLINE void, sse2, float, __m128, _mm_load_ps, _mm_store_ps,
+           _mm_set1_ps, _mm_add_ps, _mm_sub_ps, _mm_mul_ps)
+GEN_IFFT_16(static INLINE void, sse2, float, __m128, _mm_load_ps, _mm_store_ps,
+            _mm_set1_ps, _mm_add_ps, _mm_sub_ps, _mm_mul_ps)
+GEN_IFFT_32(static INLINE void, sse2, float, __m128, _mm_load_ps, _mm_store_ps,
+            _mm_set1_ps, _mm_add_ps, _mm_sub_ps, _mm_mul_ps)
+
+void aom_ifft4x4_float_sse2(const float *input, float *temp, float *output) {
+  aom_ifft_2d_gen(input, temp, output, 4, aom_fft1d_4_float, aom_fft1d_4_sse2,
+                  aom_ifft1d_4_sse2, aom_transpose_float_sse2, 4);
+}
+
+void aom_ifft8x8_float_sse2(const float *input, float *temp, float *output) {
+  aom_ifft_2d_gen(input, temp, output, 8, aom_fft1d_8_float, aom_fft1d_8_sse2,
+                  aom_ifft1d_8_sse2, aom_transpose_float_sse2, 4);
+}
+
+void aom_ifft16x16_float_sse2(const float *input, float *temp, float *output) {
+  aom_ifft_2d_gen(input, temp, output, 16, aom_fft1d_16_float,
+                  aom_fft1d_16_sse2, aom_ifft1d_16_sse2,
+                  aom_transpose_float_sse2, 4);
+}
+
+void aom_ifft32x32_float_sse2(const float *input, float *temp, float *output) {
+  aom_ifft_2d_gen(input, temp, output, 32, aom_fft1d_32_float,
+                  aom_fft1d_32_sse2, aom_ifft1d_32_sse2,
+                  aom_transpose_float_sse2, 4);
+}
diff --git a/third_party/aom/aom_dsp/x86/fwd_txfm_impl_sse2.h b/third_party/aom/aom_dsp/x86/fwd_txfm_impl_sse2.h
new file mode 100644
index 0000000000..7ee8ba330e
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/fwd_txfm_impl_sse2.h
@@ -0,0 +1,529 @@
+/*
+ * Copyright (c) 2016, 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 <emmintrin.h>  // SSE2
+
+#include "config/aom_dsp_rtcd.h"
+
+#include "aom_dsp/txfm_common.h"
+#include "aom_dsp/x86/fwd_txfm_sse2.h"
+#include "aom_dsp/x86/txfm_common_sse2.h"
+#include "aom_ports/mem.h"
+
+// TODO(jingning) The high bit-depth functions need rework for performance.
+// After we properly fix the high bit-depth function implementations, this
+// file's dependency should be substantially simplified.
+#if DCT_HIGH_BIT_DEPTH
+#define ADD_EPI16 _mm_adds_epi16
+#define SUB_EPI16 _mm_subs_epi16
+
+#else
+#define ADD_EPI16 _mm_add_epi16
+#define SUB_EPI16 _mm_sub_epi16
+#endif
+
+static void FDCT4x4_2D_HELPER(const int16_t *input, int stride, __m128i *in0,
+                              __m128i *in1) {
+  // Constants
+  // These are the coefficients used for the multiplies.
+  // In the comments, pN means cos(N pi /64) and mN is -cos(N pi /64),
+  // where cospi_N_64 = cos(N pi /64)
+  const __m128i k__cospi_A =
+      octa_set_epi16(cospi_16_64, cospi_16_64, cospi_16_64, cospi_16_64,
+                     cospi_16_64, -cospi_16_64, cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_B =
+      octa_set_epi16(cospi_16_64, -cospi_16_64, cospi_16_64, -cospi_16_64,
+                     cospi_16_64, cospi_16_64, cospi_16_64, cospi_16_64);
+  const __m128i k__cospi_C =
+      octa_set_epi16(cospi_8_64, cospi_24_64, cospi_8_64, cospi_24_64,
+                     cospi_24_64, -cospi_8_64, cospi_24_64, -cospi_8_64);
+  const __m128i k__cospi_D =
+      octa_set_epi16(cospi_24_64, -cospi_8_64, cospi_24_64, -cospi_8_64,
+                     cospi_8_64, cospi_24_64, cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_E =
+      octa_set_epi16(cospi_16_64, cospi_16_64, cospi_16_64, cospi_16_64,
+                     cospi_16_64, cospi_16_64, cospi_16_64, cospi_16_64);
+  const __m128i k__cospi_F =
+      octa_set_epi16(cospi_16_64, -cospi_16_64, cospi_16_64, -cospi_16_64,
+                     cospi_16_64, -cospi_16_64, cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_G =
+      octa_set_epi16(cospi_8_64, cospi_24_64, cospi_8_64, cospi_24_64,
+                     -cospi_8_64, -cospi_24_64, -cospi_8_64, -cospi_24_64);
+  const __m128i k__cospi_H =
+      octa_set_epi16(cospi_24_64, -cospi_8_64, cospi_24_64, -cospi_8_64,
+                     -cospi_24_64, cospi_8_64, -cospi_24_64, cospi_8_64);
+
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  // This second rounding constant saves doing some extra adds at the end
+  const __m128i k__DCT_CONST_ROUNDING2 =
+      _mm_set1_epi32(DCT_CONST_ROUNDING + (DCT_CONST_ROUNDING << 1));
+  const int DCT_CONST_BITS2 = DCT_CONST_BITS + 2;
+  const __m128i k__nonzero_bias_a = _mm_setr_epi16(0, 1, 1, 1, 1, 1, 1, 1);
+  const __m128i k__nonzero_bias_b = _mm_setr_epi16(1, 0, 0, 0, 0, 0, 0, 0);
+
+  // Load inputs.
+  *in0 = _mm_loadl_epi64((const __m128i *)(input + 0 * stride));
+  *in1 = _mm_loadl_epi64((const __m128i *)(input + 1 * stride));
+  *in1 = _mm_unpacklo_epi64(
+      *in1, _mm_loadl_epi64((const __m128i *)(input + 2 * stride)));
+  *in0 = _mm_unpacklo_epi64(
+      *in0, _mm_loadl_epi64((const __m128i *)(input + 3 * stride)));
+  // in0 = [i0 i1 i2 i3 iC iD iE iF]
+  // in1 = [i4 i5 i6 i7 i8 i9 iA iB]
+  // multiply by 16 to give some extra precision
+  *in0 = _mm_slli_epi16(*in0, 4);
+  *in1 = _mm_slli_epi16(*in1, 4);
+  // if (i == 0 && input[0]) input[0] += 1;
+  // add 1 to the upper left pixel if it is non-zero, which helps reduce
+  // the round-trip error
+  {
+    // The mask will only contain whether the first value is zero, all
+    // other comparison will fail as something shifted by 4 (above << 4)
+    // can never be equal to one. To increment in the non-zero case, we
+    // add the mask and one for the first element:
+    //   - if zero, mask = -1, v = v - 1 + 1 = v
+    //   - if non-zero, mask = 0, v = v + 0 + 1 = v + 1
+    __m128i mask = _mm_cmpeq_epi16(*in0, k__nonzero_bias_a);
+    *in0 = _mm_add_epi16(*in0, mask);
+    *in0 = _mm_add_epi16(*in0, k__nonzero_bias_b);
+  }
+  // There are 4 total stages, alternating between an add/subtract stage
+  // followed by an multiply-and-add stage.
+  {
+    // Stage 1: Add/subtract
+
+    // in0 = [i0 i1 i2 i3 iC iD iE iF]
+    // in1 = [i4 i5 i6 i7 i8 i9 iA iB]
+    const __m128i r0 = _mm_unpacklo_epi16(*in0, *in1);
+    const __m128i r1 = _mm_unpackhi_epi16(*in0, *in1);
+    // r0 = [i0 i4 i1 i5 i2 i6 i3 i7]
+    // r1 = [iC i8 iD i9 iE iA iF iB]
+    const __m128i r2 = _mm_shuffle_epi32(r0, 0xB4);
+    const __m128i r3 = _mm_shuffle_epi32(r1, 0xB4);
+    // r2 = [i0 i4 i1 i5 i3 i7 i2 i6]
+    // r3 = [iC i8 iD i9 iF iB iE iA]
+
+    const __m128i t0 = _mm_add_epi16(r2, r3);
+    const __m128i t1 = _mm_sub_epi16(r2, r3);
+    // t0 = [a0 a4 a1 a5 a3 a7 a2 a6]
+    // t1 = [aC a8 aD a9 aF aB aE aA]
+
+    // Stage 2: multiply by constants (which gets us into 32 bits).
+    // The constants needed here are:
+    // k__cospi_A = [p16 p16 p16 p16 p16 m16 p16 m16]
+    // k__cospi_B = [p16 m16 p16 m16 p16 p16 p16 p16]
+    // k__cospi_C = [p08 p24 p08 p24 p24 m08 p24 m08]
+    // k__cospi_D = [p24 m08 p24 m08 p08 p24 p08 p24]
+    const __m128i u0 = _mm_madd_epi16(t0, k__cospi_A);
+    const __m128i u2 = _mm_madd_epi16(t0, k__cospi_B);
+    const __m128i u1 = _mm_madd_epi16(t1, k__cospi_C);
+    const __m128i u3 = _mm_madd_epi16(t1, k__cospi_D);
+    // Then add and right-shift to get back to 16-bit range
+    const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
+    const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
+    const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
+    const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
+    const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+    const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+    const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+    const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+    // w0 = [b0 b1 b7 b6]
+    // w1 = [b8 b9 bF bE]
+    // w2 = [b4 b5 b3 b2]
+    // w3 = [bC bD bB bA]
+    const __m128i x0 = _mm_packs_epi32(w0, w1);
+    const __m128i x1 = _mm_packs_epi32(w2, w3);
+
+    // x0 = [b0 b1 b7 b6 b8 b9 bF bE]
+    // x1 = [b4 b5 b3 b2 bC bD bB bA]
+    *in0 = _mm_shuffle_epi32(x0, 0xD8);
+    *in1 = _mm_shuffle_epi32(x1, 0x8D);
+    // in0 = [b0 b1 b8 b9 b7 b6 bF bE]
+    // in1 = [b3 b2 bB bA b4 b5 bC bD]
+  }
+  {
+    // vertical DCTs finished. Now we do the horizontal DCTs.
+    // Stage 3: Add/subtract
+
+    const __m128i t0 = ADD_EPI16(*in0, *in1);
+    const __m128i t1 = SUB_EPI16(*in0, *in1);
+
+    // Stage 4: multiply by constants (which gets us into 32 bits).
+    {
+      // The constants needed here are:
+      // k__cospi_E = [p16 p16 p16 p16 p16 p16 p16 p16]
+      // k__cospi_F = [p16 m16 p16 m16 p16 m16 p16 m16]
+      // k__cospi_G = [p08 p24 p08 p24 m08 m24 m08 m24]
+      // k__cospi_H = [p24 m08 p24 m08 m24 p08 m24 p08]
+      const __m128i u0 = _mm_madd_epi16(t0, k__cospi_E);
+      const __m128i u1 = _mm_madd_epi16(t0, k__cospi_F);
+      const __m128i u2 = _mm_madd_epi16(t1, k__cospi_G);
+      const __m128i u3 = _mm_madd_epi16(t1, k__cospi_H);
+      // Then add and right-shift to get back to 16-bit range
+      // but this combines the final right-shift as well to save operations
+      // This unusual rounding operations is to maintain bit-accurate
+      // compatibility with the c version of this function which has two
+      // rounding steps in a row.
+      const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING2);
+      const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING2);
+      const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING2);
+      const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING2);
+      const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS2);
+      const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS2);
+      const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS2);
+      const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS2);
+      *in0 = _mm_packs_epi32(w0, w2);
+      *in1 = _mm_packs_epi32(w1, w3);
+    }
+  }
+}
+
+void FDCT4x4_2D(const int16_t *input, tran_low_t *output, int stride) {
+  // This 2D transform implements 4 vertical 1D transforms followed
+  // by 4 horizontal 1D transforms.  The multiplies and adds are as given
+  // by Chen, Smith and Fralick ('77).  The commands for moving the data
+  // around have been minimized by hand.
+  // For the purposes of the comments, the 16 inputs are referred to at i0
+  // through iF (in raster order), intermediate variables are a0, b0, c0
+  // through f, and correspond to the in-place computations mapped to input
+  // locations.  The outputs, o0 through oF are labeled according to the
+  // output locations.
+  __m128i in0, in1;
+  FDCT4x4_2D_HELPER(input, stride, &in0, &in1);
+
+  // Post-condition (v + 1) >> 2 is now incorporated into previous
+  // add and right-shift commands.  Only 2 store instructions needed
+  // because we are using the fact that 1/3 are stored just after 0/2.
+  storeu_output(&in0, output + 0 * 4);
+  storeu_output(&in1, output + 2 * 4);
+}
+
+void FDCT4x4_2D_LP(const int16_t *input, int16_t *output, int stride) {
+  __m128i in0, in1;
+  FDCT4x4_2D_HELPER(input, stride, &in0, &in1);
+  _mm_storeu_si128((__m128i *)(output + 0 * 4), in0);
+  _mm_storeu_si128((__m128i *)(output + 2 * 4), in1);
+}
+
+#if CONFIG_INTERNAL_STATS
+void FDCT8x8_2D(const int16_t *input, tran_low_t *output, int stride) {
+  int pass;
+  // Constants
+  //    When we use them, in one case, they are all the same. In all others
+  //    it's a pair of them that we need to repeat four times. This is done
+  //    by constructing the 32 bit constant corresponding to that pair.
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64);
+  const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64);
+  const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+#if DCT_HIGH_BIT_DEPTH
+  int overflow;
+#endif
+  // Load input
+  __m128i in0 = _mm_load_si128((const __m128i *)(input + 0 * stride));
+  __m128i in1 = _mm_load_si128((const __m128i *)(input + 1 * stride));
+  __m128i in2 = _mm_load_si128((const __m128i *)(input + 2 * stride));
+  __m128i in3 = _mm_load_si128((const __m128i *)(input + 3 * stride));
+  __m128i in4 = _mm_load_si128((const __m128i *)(input + 4 * stride));
+  __m128i in5 = _mm_load_si128((const __m128i *)(input + 5 * stride));
+  __m128i in6 = _mm_load_si128((const __m128i *)(input + 6 * stride));
+  __m128i in7 = _mm_load_si128((const __m128i *)(input + 7 * stride));
+  // Pre-condition input (shift by two)
+  in0 = _mm_slli_epi16(in0, 2);
+  in1 = _mm_slli_epi16(in1, 2);
+  in2 = _mm_slli_epi16(in2, 2);
+  in3 = _mm_slli_epi16(in3, 2);
+  in4 = _mm_slli_epi16(in4, 2);
+  in5 = _mm_slli_epi16(in5, 2);
+  in6 = _mm_slli_epi16(in6, 2);
+  in7 = _mm_slli_epi16(in7, 2);
+
+  // We do two passes, first the columns, then the rows. The results of the
+  // first pass are transposed so that the same column code can be reused. The
+  // results of the second pass are also transposed so that the rows (processed
+  // as columns) are put back in row positions.
+  for (pass = 0; pass < 2; pass++) {
+    // To store results of each pass before the transpose.
+    __m128i res0, res1, res2, res3, res4, res5, res6, res7;
+    // Add/subtract
+    const __m128i q0 = ADD_EPI16(in0, in7);
+    const __m128i q1 = ADD_EPI16(in1, in6);
+    const __m128i q2 = ADD_EPI16(in2, in5);
+    const __m128i q3 = ADD_EPI16(in3, in4);
+    const __m128i q4 = SUB_EPI16(in3, in4);
+    const __m128i q5 = SUB_EPI16(in2, in5);
+    const __m128i q6 = SUB_EPI16(in1, in6);
+    const __m128i q7 = SUB_EPI16(in0, in7);
+#if DCT_HIGH_BIT_DEPTH
+    if (pass == 1) {
+      overflow =
+          check_epi16_overflow_x8(&q0, &q1, &q2, &q3, &q4, &q5, &q6, &q7);
+      if (overflow) {
+        aom_highbd_fdct8x8_c(input, output, stride);
+        return;
+      }
+    }
+#endif  // DCT_HIGH_BIT_DEPTH
+    // Work on first four results
+    {
+      // Add/subtract
+      const __m128i r0 = ADD_EPI16(q0, q3);
+      const __m128i r1 = ADD_EPI16(q1, q2);
+      const __m128i r2 = SUB_EPI16(q1, q2);
+      const __m128i r3 = SUB_EPI16(q0, q3);
+#if DCT_HIGH_BIT_DEPTH
+      overflow = check_epi16_overflow_x4(&r0, &r1, &r2, &r3);
+      if (overflow) {
+        aom_highbd_fdct8x8_c(input, output, stride);
+        return;
+      }
+#endif  // DCT_HIGH_BIT_DEPTH
+      // Interleave to do the multiply by constants which gets us into 32bits
+      {
+        const __m128i t0 = _mm_unpacklo_epi16(r0, r1);
+        const __m128i t1 = _mm_unpackhi_epi16(r0, r1);
+        const __m128i t2 = _mm_unpacklo_epi16(r2, r3);
+        const __m128i t3 = _mm_unpackhi_epi16(r2, r3);
+        const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16);
+        const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_p16);
+        const __m128i u2 = _mm_madd_epi16(t0, k__cospi_p16_m16);
+        const __m128i u3 = _mm_madd_epi16(t1, k__cospi_p16_m16);
+        const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p24_p08);
+        const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p24_p08);
+        const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m08_p24);
+        const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m08_p24);
+        // dct_const_round_shift
+        const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
+        const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
+        const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
+        const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
+        const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
+        const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
+        const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
+        const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
+        const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+        const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+        const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+        const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+        const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
+        const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
+        const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
+        const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
+        // Combine
+        res0 = _mm_packs_epi32(w0, w1);
+        res4 = _mm_packs_epi32(w2, w3);
+        res2 = _mm_packs_epi32(w4, w5);
+        res6 = _mm_packs_epi32(w6, w7);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x4(&res0, &res4, &res2, &res6);
+        if (overflow) {
+          aom_highbd_fdct8x8_c(input, output, stride);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+    }
+    // Work on next four results
+    {
+      // Interleave to do the multiply by constants which gets us into 32bits
+      const __m128i d0 = _mm_unpacklo_epi16(q6, q5);
+      const __m128i d1 = _mm_unpackhi_epi16(q6, q5);
+      const __m128i e0 = _mm_madd_epi16(d0, k__cospi_p16_m16);
+      const __m128i e1 = _mm_madd_epi16(d1, k__cospi_p16_m16);
+      const __m128i e2 = _mm_madd_epi16(d0, k__cospi_p16_p16);
+      const __m128i e3 = _mm_madd_epi16(d1, k__cospi_p16_p16);
+      // dct_const_round_shift
+      const __m128i f0 = _mm_add_epi32(e0, k__DCT_CONST_ROUNDING);
+      const __m128i f1 = _mm_add_epi32(e1, k__DCT_CONST_ROUNDING);
+      const __m128i f2 = _mm_add_epi32(e2, k__DCT_CONST_ROUNDING);
+      const __m128i f3 = _mm_add_epi32(e3, k__DCT_CONST_ROUNDING);
+      const __m128i s0 = _mm_srai_epi32(f0, DCT_CONST_BITS);
+      const __m128i s1 = _mm_srai_epi32(f1, DCT_CONST_BITS);
+      const __m128i s2 = _mm_srai_epi32(f2, DCT_CONST_BITS);
+      const __m128i s3 = _mm_srai_epi32(f3, DCT_CONST_BITS);
+      // Combine
+      const __m128i r0 = _mm_packs_epi32(s0, s1);
+      const __m128i r1 = _mm_packs_epi32(s2, s3);
+#if DCT_HIGH_BIT_DEPTH
+      overflow = check_epi16_overflow_x2(&r0, &r1);
+      if (overflow) {
+        aom_highbd_fdct8x8_c(input, output, stride);
+        return;
+      }
+#endif  // DCT_HIGH_BIT_DEPTH
+      {
+        // Add/subtract
+        const __m128i x0 = ADD_EPI16(q4, r0);
+        const __m128i x1 = SUB_EPI16(q4, r0);
+        const __m128i x2 = SUB_EPI16(q7, r1);
+        const __m128i x3 = ADD_EPI16(q7, r1);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x4(&x0, &x1, &x2, &x3);
+        if (overflow) {
+          aom_highbd_fdct8x8_c(input, output, stride);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+        // Interleave to do the multiply by constants which gets us into 32bits
+        {
+          const __m128i t0 = _mm_unpacklo_epi16(x0, x3);
+          const __m128i t1 = _mm_unpackhi_epi16(x0, x3);
+          const __m128i t2 = _mm_unpacklo_epi16(x1, x2);
+          const __m128i t3 = _mm_unpackhi_epi16(x1, x2);
+          const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p28_p04);
+          const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p28_p04);
+          const __m128i u2 = _mm_madd_epi16(t0, k__cospi_m04_p28);
+          const __m128i u3 = _mm_madd_epi16(t1, k__cospi_m04_p28);
+          const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p12_p20);
+          const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p12_p20);
+          const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m20_p12);
+          const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m20_p12);
+          // dct_const_round_shift
+          const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
+          const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
+          const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
+          const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
+          const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
+          const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
+          const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
+          const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
+          const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+          const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+          const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+          const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+          const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
+          const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
+          const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
+          const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
+          // Combine
+          res1 = _mm_packs_epi32(w0, w1);
+          res7 = _mm_packs_epi32(w2, w3);
+          res5 = _mm_packs_epi32(w4, w5);
+          res3 = _mm_packs_epi32(w6, w7);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x4(&res1, &res7, &res5, &res3);
+          if (overflow) {
+            aom_highbd_fdct8x8_c(input, output, stride);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+      }
+    }
+    // Transpose the 8x8.
+    {
+      // 00 01 02 03 04 05 06 07
+      // 10 11 12 13 14 15 16 17
+      // 20 21 22 23 24 25 26 27
+      // 30 31 32 33 34 35 36 37
+      // 40 41 42 43 44 45 46 47
+      // 50 51 52 53 54 55 56 57
+      // 60 61 62 63 64 65 66 67
+      // 70 71 72 73 74 75 76 77
+      const __m128i tr0_0 = _mm_unpacklo_epi16(res0, res1);
+      const __m128i tr0_1 = _mm_unpacklo_epi16(res2, res3);
+      const __m128i tr0_2 = _mm_unpackhi_epi16(res0, res1);
+      const __m128i tr0_3 = _mm_unpackhi_epi16(res2, res3);
+      const __m128i tr0_4 = _mm_unpacklo_epi16(res4, res5);
+      const __m128i tr0_5 = _mm_unpacklo_epi16(res6, res7);
+      const __m128i tr0_6 = _mm_unpackhi_epi16(res4, res5);
+      const __m128i tr0_7 = _mm_unpackhi_epi16(res6, res7);
+      // 00 10 01 11 02 12 03 13
+      // 20 30 21 31 22 32 23 33
+      // 04 14 05 15 06 16 07 17
+      // 24 34 25 35 26 36 27 37
+      // 40 50 41 51 42 52 43 53
+      // 60 70 61 71 62 72 63 73
+      // 54 54 55 55 56 56 57 57
+      // 64 74 65 75 66 76 67 77
+      const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
+      const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3);
+      const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
+      const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3);
+      const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5);
+      const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
+      const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5);
+      const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
+      // 00 10 20 30 01 11 21 31
+      // 40 50 60 70 41 51 61 71
+      // 02 12 22 32 03 13 23 33
+      // 42 52 62 72 43 53 63 73
+      // 04 14 24 34 05 15 21 36
+      // 44 54 64 74 45 55 61 76
+      // 06 16 26 36 07 17 27 37
+      // 46 56 66 76 47 57 67 77
+      in0 = _mm_unpacklo_epi64(tr1_0, tr1_4);
+      in1 = _mm_unpackhi_epi64(tr1_0, tr1_4);
+      in2 = _mm_unpacklo_epi64(tr1_2, tr1_6);
+      in3 = _mm_unpackhi_epi64(tr1_2, tr1_6);
+      in4 = _mm_unpacklo_epi64(tr1_1, tr1_5);
+      in5 = _mm_unpackhi_epi64(tr1_1, tr1_5);
+      in6 = _mm_unpacklo_epi64(tr1_3, tr1_7);
+      in7 = _mm_unpackhi_epi64(tr1_3, tr1_7);
+      // 00 10 20 30 40 50 60 70
+      // 01 11 21 31 41 51 61 71
+      // 02 12 22 32 42 52 62 72
+      // 03 13 23 33 43 53 63 73
+      // 04 14 24 34 44 54 64 74
+      // 05 15 25 35 45 55 65 75
+      // 06 16 26 36 46 56 66 76
+      // 07 17 27 37 47 57 67 77
+    }
+  }
+  // Post-condition output and store it
+  {
+    // Post-condition (division by two)
+    //    division of two 16 bits signed numbers using shifts
+    //    n / 2 = (n - (n >> 15)) >> 1
+    const __m128i sign_in0 = _mm_srai_epi16(in0, 15);
+    const __m128i sign_in1 = _mm_srai_epi16(in1, 15);
+    const __m128i sign_in2 = _mm_srai_epi16(in2, 15);
+    const __m128i sign_in3 = _mm_srai_epi16(in3, 15);
+    const __m128i sign_in4 = _mm_srai_epi16(in4, 15);
+    const __m128i sign_in5 = _mm_srai_epi16(in5, 15);
+    const __m128i sign_in6 = _mm_srai_epi16(in6, 15);
+    const __m128i sign_in7 = _mm_srai_epi16(in7, 15);
+    in0 = _mm_sub_epi16(in0, sign_in0);
+    in1 = _mm_sub_epi16(in1, sign_in1);
+    in2 = _mm_sub_epi16(in2, sign_in2);
+    in3 = _mm_sub_epi16(in3, sign_in3);
+    in4 = _mm_sub_epi16(in4, sign_in4);
+    in5 = _mm_sub_epi16(in5, sign_in5);
+    in6 = _mm_sub_epi16(in6, sign_in6);
+    in7 = _mm_sub_epi16(in7, sign_in7);
+    in0 = _mm_srai_epi16(in0, 1);
+    in1 = _mm_srai_epi16(in1, 1);
+    in2 = _mm_srai_epi16(in2, 1);
+    in3 = _mm_srai_epi16(in3, 1);
+    in4 = _mm_srai_epi16(in4, 1);
+    in5 = _mm_srai_epi16(in5, 1);
+    in6 = _mm_srai_epi16(in6, 1);
+    in7 = _mm_srai_epi16(in7, 1);
+    // store results
+    store_output(&in0, (output + 0 * 8));
+    store_output(&in1, (output + 1 * 8));
+    store_output(&in2, (output + 2 * 8));
+    store_output(&in3, (output + 3 * 8));
+    store_output(&in4, (output + 4 * 8));
+    store_output(&in5, (output + 5 * 8));
+    store_output(&in6, (output + 6 * 8));
+    store_output(&in7, (output + 7 * 8));
+  }
+}
+#endif  // CONFIG_INTERNAL_STATS
+
+#undef ADD_EPI16
+#undef SUB_EPI16
diff --git a/third_party/aom/aom_dsp/x86/fwd_txfm_sse2.c b/third_party/aom/aom_dsp/x86/fwd_txfm_sse2.c
new file mode 100644
index 0000000000..0e4fb80468
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/fwd_txfm_sse2.c
@@ -0,0 +1,39 @@
+/*
+ * Copyright (c) 2016, 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 <emmintrin.h>  // SSE2
+
+#include "config/aom_config.h"
+#include "config/aom_dsp_rtcd.h"
+
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/x86/fwd_txfm_sse2.h"
+
+#define DCT_HIGH_BIT_DEPTH 0
+#define FDCT4x4_2D_HELPER fdct4x4_helper
+#define FDCT4x4_2D aom_fdct4x4_sse2
+#define FDCT4x4_2D_LP aom_fdct4x4_lp_sse2
+#define FDCT8x8_2D aom_fdct8x8_sse2
+#include "aom_dsp/x86/fwd_txfm_impl_sse2.h"
+#undef FDCT4x4_2D_HELPER
+#undef FDCT4x4_2D
+#undef FDCT4x4_2D_LP
+#undef FDCT8x8_2D
+
+#if CONFIG_AV1_HIGHBITDEPTH
+
+#undef DCT_HIGH_BIT_DEPTH
+#define DCT_HIGH_BIT_DEPTH 1
+#define FDCT8x8_2D aom_highbd_fdct8x8_sse2
+#include "aom_dsp/x86/fwd_txfm_impl_sse2.h"  // NOLINT
+#undef FDCT8x8_2D
+
+#endif
diff --git a/third_party/aom/aom_dsp/x86/fwd_txfm_sse2.h b/third_party/aom/aom_dsp/x86/fwd_txfm_sse2.h
new file mode 100644
index 0000000000..78ea98522e
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/fwd_txfm_sse2.h
@@ -0,0 +1,160 @@
+/*
+ * Copyright (c) 2016, 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_AOM_DSP_X86_FWD_TXFM_SSE2_H_
+#define AOM_AOM_DSP_X86_FWD_TXFM_SSE2_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+static INLINE __m128i k_madd_epi32(__m128i a, __m128i b) {
+  __m128i buf0, buf1;
+  buf0 = _mm_mul_epu32(a, b);
+  a = _mm_srli_epi64(a, 32);
+  b = _mm_srli_epi64(b, 32);
+  buf1 = _mm_mul_epu32(a, b);
+  return _mm_add_epi64(buf0, buf1);
+}
+
+static INLINE __m128i k_packs_epi64(__m128i a, __m128i b) {
+  __m128i buf0 = _mm_shuffle_epi32(a, _MM_SHUFFLE(0, 0, 2, 0));
+  __m128i buf1 = _mm_shuffle_epi32(b, _MM_SHUFFLE(0, 0, 2, 0));
+  return _mm_unpacklo_epi64(buf0, buf1);
+}
+
+static INLINE int check_epi16_overflow_x2(const __m128i *preg0,
+                                          const __m128i *preg1) {
+  const __m128i max_overflow = _mm_set1_epi16(0x7fff);
+  const __m128i min_overflow = _mm_set1_epi16((short)0x8000);
+  __m128i cmp0 = _mm_or_si128(_mm_cmpeq_epi16(*preg0, max_overflow),
+                              _mm_cmpeq_epi16(*preg0, min_overflow));
+  __m128i cmp1 = _mm_or_si128(_mm_cmpeq_epi16(*preg1, max_overflow),
+                              _mm_cmpeq_epi16(*preg1, min_overflow));
+  cmp0 = _mm_or_si128(cmp0, cmp1);
+  return _mm_movemask_epi8(cmp0);
+}
+
+static INLINE int check_epi16_overflow_x4(const __m128i *preg0,
+                                          const __m128i *preg1,
+                                          const __m128i *preg2,
+                                          const __m128i *preg3) {
+  const __m128i max_overflow = _mm_set1_epi16(0x7fff);
+  const __m128i min_overflow = _mm_set1_epi16((short)0x8000);
+  __m128i cmp0 = _mm_or_si128(_mm_cmpeq_epi16(*preg0, max_overflow),
+                              _mm_cmpeq_epi16(*preg0, min_overflow));
+  __m128i cmp1 = _mm_or_si128(_mm_cmpeq_epi16(*preg1, max_overflow),
+                              _mm_cmpeq_epi16(*preg1, min_overflow));
+  __m128i cmp2 = _mm_or_si128(_mm_cmpeq_epi16(*preg2, max_overflow),
+                              _mm_cmpeq_epi16(*preg2, min_overflow));
+  __m128i cmp3 = _mm_or_si128(_mm_cmpeq_epi16(*preg3, max_overflow),
+                              _mm_cmpeq_epi16(*preg3, min_overflow));
+  cmp0 = _mm_or_si128(_mm_or_si128(cmp0, cmp1), _mm_or_si128(cmp2, cmp3));
+  return _mm_movemask_epi8(cmp0);
+}
+
+static INLINE int check_epi16_overflow_x8(
+    const __m128i *preg0, const __m128i *preg1, const __m128i *preg2,
+    const __m128i *preg3, const __m128i *preg4, const __m128i *preg5,
+    const __m128i *preg6, const __m128i *preg7) {
+  int res0, res1;
+  res0 = check_epi16_overflow_x4(preg0, preg1, preg2, preg3);
+  res1 = check_epi16_overflow_x4(preg4, preg5, preg6, preg7);
+  return res0 + res1;
+}
+
+static INLINE int check_epi16_overflow_x12(
+    const __m128i *preg0, const __m128i *preg1, const __m128i *preg2,
+    const __m128i *preg3, const __m128i *preg4, const __m128i *preg5,
+    const __m128i *preg6, const __m128i *preg7, const __m128i *preg8,
+    const __m128i *preg9, const __m128i *preg10, const __m128i *preg11) {
+  int res0, res1;
+  res0 = check_epi16_overflow_x4(preg0, preg1, preg2, preg3);
+  res1 = check_epi16_overflow_x4(preg4, preg5, preg6, preg7);
+  if (!res0) res0 = check_epi16_overflow_x4(preg8, preg9, preg10, preg11);
+  return res0 + res1;
+}
+
+static INLINE int check_epi16_overflow_x16(
+    const __m128i *preg0, const __m128i *preg1, const __m128i *preg2,
+    const __m128i *preg3, const __m128i *preg4, const __m128i *preg5,
+    const __m128i *preg6, const __m128i *preg7, const __m128i *preg8,
+    const __m128i *preg9, const __m128i *preg10, const __m128i *preg11,
+    const __m128i *preg12, const __m128i *preg13, const __m128i *preg14,
+    const __m128i *preg15) {
+  int res0, res1;
+  res0 = check_epi16_overflow_x4(preg0, preg1, preg2, preg3);
+  res1 = check_epi16_overflow_x4(preg4, preg5, preg6, preg7);
+  if (!res0) {
+    res0 = check_epi16_overflow_x4(preg8, preg9, preg10, preg11);
+    if (!res1) res1 = check_epi16_overflow_x4(preg12, preg13, preg14, preg15);
+  }
+  return res0 + res1;
+}
+
+static INLINE int check_epi16_overflow_x32(
+    const __m128i *preg0, const __m128i *preg1, const __m128i *preg2,
+    const __m128i *preg3, const __m128i *preg4, const __m128i *preg5,
+    const __m128i *preg6, const __m128i *preg7, const __m128i *preg8,
+    const __m128i *preg9, const __m128i *preg10, const __m128i *preg11,
+    const __m128i *preg12, const __m128i *preg13, const __m128i *preg14,
+    const __m128i *preg15, const __m128i *preg16, const __m128i *preg17,
+    const __m128i *preg18, const __m128i *preg19, const __m128i *preg20,
+    const __m128i *preg21, const __m128i *preg22, const __m128i *preg23,
+    const __m128i *preg24, const __m128i *preg25, const __m128i *preg26,
+    const __m128i *preg27, const __m128i *preg28, const __m128i *preg29,
+    const __m128i *preg30, const __m128i *preg31) {
+  int res0, res1;
+  res0 = check_epi16_overflow_x4(preg0, preg1, preg2, preg3);
+  res1 = check_epi16_overflow_x4(preg4, preg5, preg6, preg7);
+  if (!res0) {
+    res0 = check_epi16_overflow_x4(preg8, preg9, preg10, preg11);
+    if (!res1) {
+      res1 = check_epi16_overflow_x4(preg12, preg13, preg14, preg15);
+      if (!res0) {
+        res0 = check_epi16_overflow_x4(preg16, preg17, preg18, preg19);
+        if (!res1) {
+          res1 = check_epi16_overflow_x4(preg20, preg21, preg22, preg23);
+          if (!res0) {
+            res0 = check_epi16_overflow_x4(preg24, preg25, preg26, preg27);
+            if (!res1)
+              res1 = check_epi16_overflow_x4(preg28, preg29, preg30, preg31);
+          }
+        }
+      }
+    }
+  }
+  return res0 + res1;
+}
+
+static INLINE void store_output(const __m128i *poutput, tran_low_t *dst_ptr) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i sign_bits = _mm_cmplt_epi16(*poutput, zero);
+  __m128i out0 = _mm_unpacklo_epi16(*poutput, sign_bits);
+  __m128i out1 = _mm_unpackhi_epi16(*poutput, sign_bits);
+  _mm_store_si128((__m128i *)(dst_ptr), out0);
+  _mm_store_si128((__m128i *)(dst_ptr + 4), out1);
+}
+
+static INLINE void storeu_output(const __m128i *poutput, tran_low_t *dst_ptr) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i sign_bits = _mm_cmplt_epi16(*poutput, zero);
+  __m128i out0 = _mm_unpacklo_epi16(*poutput, sign_bits);
+  __m128i out1 = _mm_unpackhi_epi16(*poutput, sign_bits);
+  _mm_storeu_si128((__m128i *)(dst_ptr), out0);
+  _mm_storeu_si128((__m128i *)(dst_ptr + 4), out1);
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_AOM_DSP_X86_FWD_TXFM_SSE2_H_
diff --git a/third_party/aom/aom_dsp/x86/fwd_txfm_ssse3_x86_64.asm b/third_party/aom/aom_dsp/x86/fwd_txfm_ssse3_x86_64.asm
new file mode 100644
index 0000000000..06879040b0
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/fwd_txfm_ssse3_x86_64.asm
@@ -0,0 +1,379 @@
+;
+; Copyright (c) 2016, 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 "third_party/x86inc/x86inc.asm"
+
+SECTION_RODATA
+
+pw_11585x2: times 8 dw 23170
+pd_8192:    times 4 dd 8192
+
+%macro TRANSFORM_COEFFS 2
+pw_%1_%2:   dw  %1,  %2,  %1,  %2,  %1,  %2,  %1,  %2
+pw_%2_m%1:  dw  %2, -%1,  %2, -%1,  %2, -%1,  %2, -%1
+%endmacro
+
+TRANSFORM_COEFFS 11585,  11585
+TRANSFORM_COEFFS 15137,   6270
+TRANSFORM_COEFFS 16069,   3196
+TRANSFORM_COEFFS  9102,  13623
+
+%macro STORE_OUTPUT 2 ; index, result
+  ; const __m128i sign_bits = _mm_cmplt_epi16(*poutput, zero);
+  ; __m128i out0 = _mm_unpacklo_epi16(*poutput, sign_bits);
+  ; __m128i out1 = _mm_unpackhi_epi16(*poutput, sign_bits);
+  ; _mm_store_si128((__m128i *)(dst_ptr), out0);
+  ; _mm_store_si128((__m128i *)(dst_ptr + 4), out1);
+  pxor               m11, m11
+  pcmpgtw            m11, m%2
+  movdqa             m12, m%2
+  punpcklwd          m%2, m11
+  punpckhwd          m12, m11
+  mova               [outputq + 4*%1 +  0], m%2
+  mova               [outputq + 4*%1 + 16], m12
+%endmacro
+
+SECTION .text
+
+%if AOM_ARCH_X86_64
+INIT_XMM ssse3
+cglobal fdct8x8, 3, 5, 13, input, output, stride
+
+  mova               m8, [GLOBAL(pd_8192)]
+  mova              m12, [GLOBAL(pw_11585x2)]
+
+  lea                r3, [2 * strideq]
+  lea                r4, [4 * strideq]
+  mova               m0, [inputq]
+  mova               m1, [inputq + r3]
+  lea                inputq, [inputq + r4]
+  mova               m2, [inputq]
+  mova               m3, [inputq + r3]
+  lea                inputq, [inputq + r4]
+  mova               m4, [inputq]
+  mova               m5, [inputq + r3]
+  lea                inputq, [inputq + r4]
+  mova               m6, [inputq]
+  mova               m7, [inputq + r3]
+
+  ; left shift by 2 to increase forward transformation precision
+  psllw              m0, 2
+  psllw              m1, 2
+  psllw              m2, 2
+  psllw              m3, 2
+  psllw              m4, 2
+  psllw              m5, 2
+  psllw              m6, 2
+  psllw              m7, 2
+
+  ; column transform
+  ; stage 1
+  paddw m10, m0, m7
+  psubw m0, m7
+
+  paddw m9, m1, m6
+  psubw m1, m6
+
+  paddw m7, m2, m5
+  psubw m2, m5
+
+  paddw m6, m3, m4
+  psubw m3, m4
+
+  ; stage 2
+  paddw m5, m9, m7
+  psubw m9, m7
+
+  paddw m4, m10, m6
+  psubw m10, m6
+
+  paddw m7, m1, m2
+  psubw m1, m2
+
+  ; stage 3
+  paddw m6, m4, m5
+  psubw m4, m5
+
+  pmulhrsw m1, m12
+  pmulhrsw m7, m12
+
+  ; sin(pi / 8), cos(pi / 8)
+  punpcklwd m2, m10, m9
+  punpckhwd m10, m9
+  pmaddwd m5, m2, [GLOBAL(pw_15137_6270)]
+  pmaddwd m2, [GLOBAL(pw_6270_m15137)]
+  pmaddwd m9, m10, [GLOBAL(pw_15137_6270)]
+  pmaddwd m10, [GLOBAL(pw_6270_m15137)]
+  paddd m5, m8
+  paddd m2, m8
+  paddd m9, m8
+  paddd m10, m8
+  psrad m5, 14
+  psrad m2, 14
+  psrad m9, 14
+  psrad m10, 14
+  packssdw m5, m9
+  packssdw m2, m10
+
+  pmulhrsw m6, m12
+  pmulhrsw m4, m12
+
+  paddw m9, m3, m1
+  psubw m3, m1
+
+  paddw m10, m0, m7
+  psubw m0, m7
+
+  ; stage 4
+  ; sin(pi / 16), cos(pi / 16)
+  punpcklwd m1, m10, m9
+  punpckhwd m10, m9
+  pmaddwd m7, m1, [GLOBAL(pw_16069_3196)]
+  pmaddwd m1, [GLOBAL(pw_3196_m16069)]
+  pmaddwd m9, m10, [GLOBAL(pw_16069_3196)]
+  pmaddwd m10, [GLOBAL(pw_3196_m16069)]
+  paddd m7, m8
+  paddd m1, m8
+  paddd m9, m8
+  paddd m10, m8
+  psrad m7, 14
+  psrad m1, 14
+  psrad m9, 14
+  psrad m10, 14
+  packssdw m7, m9
+  packssdw m1, m10
+
+  ; sin(3 * pi / 16), cos(3 * pi / 16)
+  punpcklwd m11, m0, m3
+  punpckhwd m0, m3
+  pmaddwd m9, m11, [GLOBAL(pw_9102_13623)]
+  pmaddwd m11, [GLOBAL(pw_13623_m9102)]
+  pmaddwd m3, m0, [GLOBAL(pw_9102_13623)]
+  pmaddwd m0, [GLOBAL(pw_13623_m9102)]
+  paddd m9, m8
+  paddd m11, m8
+  paddd m3, m8
+  paddd m0, m8
+  psrad m9, 14
+  psrad m11, 14
+  psrad m3, 14
+  psrad m0, 14
+  packssdw m9, m3
+  packssdw m11, m0
+
+  ; transpose
+  ; stage 1
+  punpcklwd m0, m6, m7
+  punpcklwd m3, m5, m11
+  punpckhwd m6, m7
+  punpckhwd m5, m11
+  punpcklwd m7, m4, m9
+  punpcklwd m10, m2, m1
+  punpckhwd m4, m9
+  punpckhwd m2, m1
+
+  ; stage 2
+  punpckldq m9, m0, m3
+  punpckldq m1, m6, m5
+  punpckhdq m0, m3
+  punpckhdq m6, m5
+  punpckldq m3, m7, m10
+  punpckldq m5, m4, m2
+  punpckhdq m7, m10
+  punpckhdq m4, m2
+
+  ; stage 3
+  punpcklqdq m10, m9, m3
+  punpckhqdq m9, m3
+  punpcklqdq m2, m0, m7
+  punpckhqdq m0, m7
+  punpcklqdq m3, m1, m5
+  punpckhqdq m1, m5
+  punpcklqdq m7, m6, m4
+  punpckhqdq m6, m4
+
+  ; row transform
+  ; stage 1
+  paddw m5, m10, m6
+  psubw m10, m6
+
+  paddw m4, m9, m7
+  psubw m9, m7
+
+  paddw m6, m2, m1
+  psubw m2, m1
+
+  paddw m7, m0, m3
+  psubw m0, m3
+
+  ;stage 2
+  paddw m1, m5, m7
+  psubw m5, m7
+
+  paddw m3, m4, m6
+  psubw m4, m6
+
+  paddw m7, m9, m2
+  psubw m9, m2
+
+  ; stage 3
+  punpcklwd m6, m1, m3
+  punpckhwd m1, m3
+  pmaddwd m2, m6, [GLOBAL(pw_11585_11585)]
+  pmaddwd m6, [GLOBAL(pw_11585_m11585)]
+  pmaddwd m3, m1, [GLOBAL(pw_11585_11585)]
+  pmaddwd m1, [GLOBAL(pw_11585_m11585)]
+  paddd m2, m8
+  paddd m6, m8
+  paddd m3, m8
+  paddd m1, m8
+  psrad m2, 14
+  psrad m6, 14
+  psrad m3, 14
+  psrad m1, 14
+  packssdw m2, m3
+  packssdw m6, m1
+
+  pmulhrsw m7, m12
+  pmulhrsw m9, m12
+
+  punpcklwd m3, m5, m4
+  punpckhwd m5, m4
+  pmaddwd m1, m3, [GLOBAL(pw_15137_6270)]
+  pmaddwd m3, [GLOBAL(pw_6270_m15137)]
+  pmaddwd m4, m5, [GLOBAL(pw_15137_6270)]
+  pmaddwd m5, [GLOBAL(pw_6270_m15137)]
+  paddd m1, m8
+  paddd m3, m8
+  paddd m4, m8
+  paddd m5, m8
+  psrad m1, 14
+  psrad m3, 14
+  psrad m4, 14
+  psrad m5, 14
+  packssdw m1, m4
+  packssdw m3, m5
+
+  paddw m4, m0, m9
+  psubw m0, m9
+
+  paddw m5, m10, m7
+  psubw m10, m7
+
+  ; stage 4
+  punpcklwd m9, m5, m4
+  punpckhwd m5, m4
+  pmaddwd m7, m9, [GLOBAL(pw_16069_3196)]
+  pmaddwd m9, [GLOBAL(pw_3196_m16069)]
+  pmaddwd m4, m5, [GLOBAL(pw_16069_3196)]
+  pmaddwd m5, [GLOBAL(pw_3196_m16069)]
+  paddd m7, m8
+  paddd m9, m8
+  paddd m4, m8
+  paddd m5, m8
+  psrad m7, 14
+  psrad m9, 14
+  psrad m4, 14
+  psrad m5, 14
+  packssdw m7, m4
+  packssdw m9, m5
+
+  punpcklwd m4, m10, m0
+  punpckhwd m10, m0
+  pmaddwd m5, m4, [GLOBAL(pw_9102_13623)]
+  pmaddwd m4, [GLOBAL(pw_13623_m9102)]
+  pmaddwd m0, m10, [GLOBAL(pw_9102_13623)]
+  pmaddwd m10, [GLOBAL(pw_13623_m9102)]
+  paddd m5, m8
+  paddd m4, m8
+  paddd m0, m8
+  paddd m10, m8
+  psrad m5, 14
+  psrad m4, 14
+  psrad m0, 14
+  psrad m10, 14
+  packssdw m5, m0
+  packssdw m4, m10
+
+  ; transpose
+  ; stage 1
+  punpcklwd m0, m2, m7
+  punpcklwd m10, m1, m4
+  punpckhwd m2, m7
+  punpckhwd m1, m4
+  punpcklwd m7, m6, m5
+  punpcklwd m4, m3, m9
+  punpckhwd m6, m5
+  punpckhwd m3, m9
+
+  ; stage 2
+  punpckldq m5, m0, m10
+  punpckldq m9, m2, m1
+  punpckhdq m0, m10
+  punpckhdq m2, m1
+  punpckldq m10, m7, m4
+  punpckldq m1, m6, m3
+  punpckhdq m7, m4
+  punpckhdq m6, m3
+
+  ; stage 3
+  punpcklqdq m4, m5, m10
+  punpckhqdq m5, m10
+  punpcklqdq m3, m0, m7
+  punpckhqdq m0, m7
+  punpcklqdq m10, m9, m1
+  punpckhqdq m9, m1
+  punpcklqdq m7, m2, m6
+  punpckhqdq m2, m6
+
+  psraw m1, m4, 15
+  psraw m6, m5, 15
+  psraw m8, m3, 15
+  psraw m11, m0, 15
+
+  psubw m4, m1
+  psubw m5, m6
+  psubw m3, m8
+  psubw m0, m11
+
+  psraw m4, 1
+  psraw m5, 1
+  psraw m3, 1
+  psraw m0, 1
+
+  psraw m1, m10, 15
+  psraw m6, m9, 15
+  psraw m8, m7, 15
+  psraw m11, m2, 15
+
+  psubw m10, m1
+  psubw m9, m6
+  psubw m7, m8
+  psubw m2, m11
+
+  psraw m10, 1
+  psraw m9, 1
+  psraw m7, 1
+  psraw m2, 1
+
+  STORE_OUTPUT  0,  4
+  STORE_OUTPUT  8,  5
+  STORE_OUTPUT 16,  3
+  STORE_OUTPUT 24,  0
+  STORE_OUTPUT 32, 10
+  STORE_OUTPUT 40,  9
+  STORE_OUTPUT 48,  7
+  STORE_OUTPUT 56,  2
+
+  RET
+%endif
diff --git a/third_party/aom/aom_dsp/x86/highbd_adaptive_quantize_avx2.c b/third_party/aom/aom_dsp/x86/highbd_adaptive_quantize_avx2.c
new file mode 100644
index 0000000000..05c87bcff9
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/highbd_adaptive_quantize_avx2.c
@@ -0,0 +1,456 @@
+/*
+ * Copyright (c) 2019, 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"
+
+#include "aom/aom_integer.h"
+#include "aom_dsp/quantize.h"
+#include "aom_dsp/x86/quantize_x86.h"
+
+static INLINE void highbd_load_b_values_avx2(
+    const int16_t *zbin_ptr, __m256i *zbin, const int16_t *round_ptr,
+    __m256i *round, const int16_t *quant_ptr, __m256i *quant,
+    const int16_t *dequant_ptr, __m256i *dequant, const int16_t *shift_ptr,
+    __m256i *shift) {
+  *zbin = _mm256_cvtepi16_epi32(_mm_load_si128((const __m128i *)zbin_ptr));
+  *zbin = _mm256_sub_epi32(*zbin, _mm256_set1_epi32(1));
+  *round = _mm256_cvtepi16_epi32(_mm_load_si128((const __m128i *)round_ptr));
+  *quant = _mm256_cvtepi16_epi32(_mm_load_si128((const __m128i *)quant_ptr));
+  *dequant =
+      _mm256_cvtepi16_epi32(_mm_load_si128((const __m128i *)dequant_ptr));
+  *shift = _mm256_cvtepi16_epi32(_mm_load_si128((const __m128i *)shift_ptr));
+}
+
+static INLINE void highbd_update_mask1_avx2(__m256i *cmp_mask,
+                                            const int16_t *iscan_ptr,
+                                            int *is_found, __m256i *mask) {
+  __m256i temp_mask = _mm256_setzero_si256();
+  if (_mm256_movemask_epi8(*cmp_mask)) {
+    __m256i iscan = _mm256_loadu_si256((const __m256i *)(iscan_ptr));
+    temp_mask = _mm256_and_si256(*cmp_mask, iscan);
+    *is_found = 1;
+  }
+  *mask = _mm256_max_epi16(temp_mask, *mask);
+}
+
+static INLINE void highbd_update_mask0_avx2(__m256i *qcoeff0, __m256i *qcoeff1,
+                                            __m256i *threshold,
+                                            const int16_t *iscan_ptr,
+                                            int *is_found, __m256i *mask) {
+  __m256i coeff[2], cmp_mask0, cmp_mask1;
+  coeff[0] = _mm256_slli_epi32(*qcoeff0, AOM_QM_BITS);
+  cmp_mask0 = _mm256_cmpgt_epi32(coeff[0], threshold[0]);
+  coeff[1] = _mm256_slli_epi32(*qcoeff1, AOM_QM_BITS);
+  cmp_mask1 = _mm256_cmpgt_epi32(coeff[1], threshold[1]);
+  cmp_mask0 =
+      _mm256_permute4x64_epi64(_mm256_packs_epi32(cmp_mask0, cmp_mask1), 0xd8);
+  highbd_update_mask1_avx2(&cmp_mask0, iscan_ptr, is_found, mask);
+}
+
+static INLINE void highbd_mul_shift_avx2(const __m256i *x, const __m256i *y,
+                                         __m256i *p, const int shift) {
+  __m256i prod_lo = _mm256_mul_epi32(*x, *y);
+  __m256i prod_hi = _mm256_srli_epi64(*x, 32);
+  const __m256i mult_hi = _mm256_srli_epi64(*y, 32);
+  prod_hi = _mm256_mul_epi32(prod_hi, mult_hi);
+
+  prod_lo = _mm256_srli_epi64(prod_lo, shift);
+  prod_hi = _mm256_srli_epi64(prod_hi, shift);
+
+  prod_hi = _mm256_slli_epi64(prod_hi, 32);
+  *p = _mm256_blend_epi32(prod_lo, prod_hi, 0xaa);
+}
+
+static INLINE void highbd_calculate_qcoeff_avx2(__m256i *coeff,
+                                                const __m256i *round,
+                                                const __m256i *quant,
+                                                const __m256i *shift,
+                                                const int *log_scale) {
+  __m256i tmp, qcoeff;
+  qcoeff = _mm256_add_epi32(*coeff, *round);
+  highbd_mul_shift_avx2(&qcoeff, quant, &tmp, 16);
+  qcoeff = _mm256_add_epi32(tmp, qcoeff);
+  highbd_mul_shift_avx2(&qcoeff, shift, coeff, 16 - *log_scale);
+}
+
+static INLINE __m256i highbd_calculate_dqcoeff_avx2(__m256i qcoeff,
+                                                    __m256i dequant) {
+  return _mm256_mullo_epi32(qcoeff, dequant);
+}
+
+static INLINE __m256i highbd_calculate_dqcoeff_log_scale_avx2(
+    __m256i qcoeff, __m256i dequant, const int log_scale) {
+  __m256i abs_coeff = _mm256_abs_epi32(qcoeff);
+  highbd_mul_shift_avx2(&abs_coeff, &dequant, &abs_coeff, log_scale);
+  return _mm256_sign_epi32(abs_coeff, qcoeff);
+}
+
+static INLINE void highbd_store_coefficients_avx2(__m256i coeff0,
+                                                  __m256i coeff1,
+                                                  tran_low_t *coeff_ptr) {
+  _mm256_store_si256((__m256i *)(coeff_ptr), coeff0);
+  _mm256_store_si256((__m256i *)(coeff_ptr + 8), coeff1);
+}
+
+void aom_highbd_quantize_b_adaptive_avx2(
+    const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr,
+    const int16_t *round_ptr, const int16_t *quant_ptr,
+    const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
+    tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
+    const int16_t *scan, const int16_t *iscan) {
+  int index = 16;
+  int non_zero_count = 0;
+  int non_zero_count_prescan_add_zero = 0;
+  int is_found0 = 0, is_found1 = 0;
+  int eob = -1;
+  const __m256i zero = _mm256_setzero_si256();
+  __m256i zbin, round, quant, dequant, shift;
+  __m256i coeff0, qcoeff0, coeff1, qcoeff1;
+  __m256i cmp_mask, mask0 = zero, mask1 = zero;
+  __m128i temp_mask0, temp_mask1;
+  int prescan_add[2];
+  int thresh[2];
+  const int log_scale = 0;
+  const qm_val_t wt = (1 << AOM_QM_BITS);
+  for (int i = 0; i < 2; ++i) {
+    prescan_add[i] = ROUND_POWER_OF_TWO(dequant_ptr[i] * EOB_FACTOR, 7);
+    thresh[i] = (zbin_ptr[i] * wt + prescan_add[i]) - 1;
+  }
+  __m256i threshold[2];
+  threshold[0] = _mm256_set1_epi32(thresh[0]);
+  threshold[1] = _mm256_set1_epi32(thresh[1]);
+  threshold[0] = _mm256_blend_epi32(threshold[0], threshold[1], 0xfe);
+
+#if SKIP_EOB_FACTOR_ADJUST
+  int first = -1;
+#endif
+
+  // Setup global values.
+  highbd_load_b_values_avx2(zbin_ptr, &zbin, round_ptr, &round, quant_ptr,
+                            &quant, dequant_ptr, &dequant, quant_shift_ptr,
+                            &shift);
+
+  // Do DC and first 15 AC.
+  coeff0 = _mm256_load_si256((__m256i *)(coeff_ptr));
+  qcoeff0 = _mm256_abs_epi32(coeff0);
+  coeff1 = _mm256_load_si256((__m256i *)(coeff_ptr + 8));
+  qcoeff1 = _mm256_abs_epi32(coeff1);
+  highbd_update_mask0_avx2(&qcoeff0, &qcoeff1, threshold, iscan, &is_found0,
+                           &mask0);
+  __m256i temp0 = _mm256_cmpgt_epi32(qcoeff0, zbin);
+  zbin = _mm256_unpackhi_epi64(zbin, zbin);
+  __m256i temp1 = _mm256_cmpgt_epi32(qcoeff1, zbin);
+  cmp_mask = _mm256_permute4x64_epi64(_mm256_packs_epi32(temp0, temp1), 0xd8);
+  highbd_update_mask1_avx2(&cmp_mask, iscan, &is_found1, &mask1);
+  threshold[0] = threshold[1];
+  if (_mm256_movemask_epi8(cmp_mask) == 0) {
+    _mm256_store_si256((__m256i *)(qcoeff_ptr), zero);
+    _mm256_store_si256((__m256i *)(qcoeff_ptr + 8), zero);
+    _mm256_store_si256((__m256i *)(dqcoeff_ptr), zero);
+    _mm256_store_si256((__m256i *)(dqcoeff_ptr + 8), zero);
+    round = _mm256_unpackhi_epi64(round, round);
+    quant = _mm256_unpackhi_epi64(quant, quant);
+    shift = _mm256_unpackhi_epi64(shift, shift);
+    dequant = _mm256_unpackhi_epi64(dequant, dequant);
+  } else {
+    highbd_calculate_qcoeff_avx2(&qcoeff0, &round, &quant, &shift, &log_scale);
+    round = _mm256_unpackhi_epi64(round, round);
+    quant = _mm256_unpackhi_epi64(quant, quant);
+    shift = _mm256_unpackhi_epi64(shift, shift);
+    highbd_calculate_qcoeff_avx2(&qcoeff1, &round, &quant, &shift, &log_scale);
+    // Reinsert signs
+    qcoeff0 = _mm256_sign_epi32(qcoeff0, coeff0);
+    qcoeff1 = _mm256_sign_epi32(qcoeff1, coeff1);
+    // Mask out zbin threshold coeffs
+    qcoeff0 = _mm256_and_si256(qcoeff0, temp0);
+    qcoeff1 = _mm256_and_si256(qcoeff1, temp1);
+    highbd_store_coefficients_avx2(qcoeff0, qcoeff1, qcoeff_ptr);
+    coeff0 = highbd_calculate_dqcoeff_avx2(qcoeff0, dequant);
+    dequant = _mm256_unpackhi_epi64(dequant, dequant);
+    coeff1 = highbd_calculate_dqcoeff_avx2(qcoeff1, dequant);
+    highbd_store_coefficients_avx2(coeff0, coeff1, dqcoeff_ptr);
+  }
+
+  // AC only loop.
+  while (index < n_coeffs) {
+    coeff0 = _mm256_load_si256((__m256i *)(coeff_ptr + index));
+    qcoeff0 = _mm256_abs_epi32(coeff0);
+    coeff1 = _mm256_load_si256((__m256i *)(coeff_ptr + index + 8));
+    qcoeff1 = _mm256_abs_epi32(coeff1);
+    highbd_update_mask0_avx2(&qcoeff0, &qcoeff1, threshold, iscan + index,
+                             &is_found0, &mask0);
+    temp0 = _mm256_cmpgt_epi32(qcoeff0, zbin);
+    temp1 = _mm256_cmpgt_epi32(qcoeff1, zbin);
+    cmp_mask = _mm256_permute4x64_epi64(_mm256_packs_epi32(temp0, temp1), 0xd8);
+    highbd_update_mask1_avx2(&cmp_mask, iscan + index, &is_found1, &mask1);
+    if (_mm256_movemask_epi8(cmp_mask) == 0) {
+      _mm256_store_si256((__m256i *)(qcoeff_ptr + index), zero);
+      _mm256_store_si256((__m256i *)(qcoeff_ptr + index + 8), zero);
+      _mm256_store_si256((__m256i *)(dqcoeff_ptr + index), zero);
+      _mm256_store_si256((__m256i *)(dqcoeff_ptr + index + 8), zero);
+      index += 16;
+      continue;
+    }
+    highbd_calculate_qcoeff_avx2(&qcoeff0, &round, &quant, &shift, &log_scale);
+    highbd_calculate_qcoeff_avx2(&qcoeff1, &round, &quant, &shift, &log_scale);
+    qcoeff0 = _mm256_sign_epi32(qcoeff0, coeff0);
+    qcoeff1 = _mm256_sign_epi32(qcoeff1, coeff1);
+    qcoeff0 = _mm256_and_si256(qcoeff0, temp0);
+    qcoeff1 = _mm256_and_si256(qcoeff1, temp1);
+    highbd_store_coefficients_avx2(qcoeff0, qcoeff1, qcoeff_ptr + index);
+    coeff0 = highbd_calculate_dqcoeff_avx2(qcoeff0, dequant);
+    coeff1 = highbd_calculate_dqcoeff_avx2(qcoeff1, dequant);
+    highbd_store_coefficients_avx2(coeff0, coeff1, dqcoeff_ptr + index);
+    index += 16;
+  }
+  if (is_found0) {
+    temp_mask0 = _mm_max_epi16(_mm256_castsi256_si128(mask0),
+                               _mm256_extracti128_si256(mask0, 1));
+    non_zero_count = calculate_non_zero_count(temp_mask0);
+  }
+  if (is_found1) {
+    temp_mask1 = _mm_max_epi16(_mm256_castsi256_si128(mask1),
+                               _mm256_extracti128_si256(mask1, 1));
+    non_zero_count_prescan_add_zero = calculate_non_zero_count(temp_mask1);
+  }
+
+  for (int i = non_zero_count_prescan_add_zero - 1; i >= non_zero_count; i--) {
+    const int rc = scan[i];
+    qcoeff_ptr[rc] = 0;
+    dqcoeff_ptr[rc] = 0;
+  }
+
+  for (int i = non_zero_count - 1; i >= 0; i--) {
+    const int rc = scan[i];
+    if (qcoeff_ptr[rc]) {
+      eob = i;
+      break;
+    }
+  }
+
+  *eob_ptr = eob + 1;
+#if SKIP_EOB_FACTOR_ADJUST
+  // TODO(Aniket): Experiment the following loop with intrinsic by combining
+  // with the quantization loop above
+  for (int i = 0; i < non_zero_count; i++) {
+    const int rc = scan[i];
+    const int qcoeff = qcoeff_ptr[rc];
+    if (qcoeff) {
+      first = i;
+      break;
+    }
+  }
+  if ((*eob_ptr - 1) >= 0 && first == (*eob_ptr - 1)) {
+    const int rc = scan[(*eob_ptr - 1)];
+    if (qcoeff_ptr[rc] == 1 || qcoeff_ptr[rc] == -1) {
+      const int coeff = coeff_ptr[rc] * wt;
+      const int coeff_sign = AOMSIGN(coeff);
+      const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+      const int factor = EOB_FACTOR + SKIP_EOB_FACTOR_ADJUST;
+      const int prescan_add_val =
+          ROUND_POWER_OF_TWO(dequant_ptr[rc != 0] * factor, 7);
+      if (abs_coeff <
+          (zbin_ptr[rc != 0] * (1 << AOM_QM_BITS) + prescan_add_val)) {
+        qcoeff_ptr[rc] = 0;
+        dqcoeff_ptr[rc] = 0;
+        *eob_ptr = 0;
+      }
+    }
+  }
+#endif
+}
+
+void aom_highbd_quantize_b_32x32_adaptive_avx2(
+    const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr,
+    const int16_t *round_ptr, const int16_t *quant_ptr,
+    const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
+    tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
+    const int16_t *scan, const int16_t *iscan) {
+  int index = 16;
+  int non_zero_count = 0;
+  int non_zero_count_prescan_add_zero = 0;
+  int is_found0 = 0, is_found1 = 0;
+  int eob = -1;
+  const int log_scale = 1;
+  const __m256i zero = _mm256_setzero_si256();
+  __m256i zbin, round, quant, dequant, shift;
+  __m256i coeff0, qcoeff0, coeff1, qcoeff1;
+  __m256i cmp_mask, mask0 = zero, mask1 = zero;
+  __m128i temp_mask0, temp_mask1;
+  const __m256i one = _mm256_set1_epi32(1);
+  const __m256i log_scale_vec = _mm256_set1_epi32(log_scale);
+  int prescan_add[2];
+  int thresh[2];
+  const int zbins[2] = { ROUND_POWER_OF_TWO(zbin_ptr[0], log_scale),
+                         ROUND_POWER_OF_TWO(zbin_ptr[1], log_scale) };
+  const qm_val_t wt = (1 << AOM_QM_BITS);
+  for (int i = 0; i < 2; ++i) {
+    prescan_add[i] = ROUND_POWER_OF_TWO(dequant_ptr[i] * EOB_FACTOR, 7);
+    thresh[i] = (zbins[i] * wt + prescan_add[i]) - 1;
+  }
+  __m256i threshold[2];
+  threshold[0] = _mm256_set1_epi32(thresh[0]);
+  threshold[1] = _mm256_set1_epi32(thresh[1]);
+  threshold[0] = _mm256_blend_epi32(threshold[0], threshold[1], 0xfe);
+
+#if SKIP_EOB_FACTOR_ADJUST
+  int first = -1;
+#endif
+
+  // Setup global values.
+  zbin = _mm256_cvtepi16_epi32(_mm_load_si128((const __m128i *)zbin_ptr));
+  round = _mm256_cvtepi16_epi32(_mm_load_si128((const __m128i *)round_ptr));
+  quant = _mm256_cvtepi16_epi32(_mm_load_si128((const __m128i *)quant_ptr));
+  dequant = _mm256_cvtepi16_epi32(_mm_load_si128((const __m128i *)dequant_ptr));
+  shift =
+      _mm256_cvtepi16_epi32(_mm_load_si128((const __m128i *)quant_shift_ptr));
+
+  // Shift with rounding.
+  zbin = _mm256_add_epi32(zbin, log_scale_vec);
+  round = _mm256_add_epi32(round, log_scale_vec);
+  zbin = _mm256_srli_epi32(zbin, log_scale);
+  round = _mm256_srli_epi32(round, log_scale);
+  zbin = _mm256_sub_epi32(zbin, one);
+
+  // Do DC and first 15 AC.
+  coeff0 = _mm256_load_si256((__m256i *)(coeff_ptr));
+  qcoeff0 = _mm256_abs_epi32(coeff0);
+  coeff1 = _mm256_load_si256((__m256i *)(coeff_ptr + 8));
+  qcoeff1 = _mm256_abs_epi32(coeff1);
+  highbd_update_mask0_avx2(&qcoeff0, &qcoeff1, threshold, iscan, &is_found0,
+                           &mask0);
+  __m256i temp0 = _mm256_cmpgt_epi32(qcoeff0, zbin);
+  zbin = _mm256_permute2x128_si256(zbin, zbin, 0x11);
+  __m256i temp1 = _mm256_cmpgt_epi32(qcoeff1, zbin);
+  cmp_mask = _mm256_permute4x64_epi64(_mm256_packs_epi32(temp0, temp1), 0xd8);
+  highbd_update_mask1_avx2(&cmp_mask, iscan, &is_found1, &mask1);
+  threshold[0] = threshold[1];
+  if (_mm256_movemask_epi8(cmp_mask) == 0) {
+    _mm256_store_si256((__m256i *)(qcoeff_ptr), zero);
+    _mm256_store_si256((__m256i *)(qcoeff_ptr + 8), zero);
+    _mm256_store_si256((__m256i *)(dqcoeff_ptr), zero);
+    _mm256_store_si256((__m256i *)(dqcoeff_ptr + 8), zero);
+    round = _mm256_permute2x128_si256(round, round, 0x11);
+    quant = _mm256_permute2x128_si256(quant, quant, 0x11);
+    shift = _mm256_permute2x128_si256(shift, shift, 0x11);
+    dequant = _mm256_permute2x128_si256(dequant, dequant, 0x11);
+  } else {
+    highbd_calculate_qcoeff_avx2(&qcoeff0, &round, &quant, &shift, &log_scale);
+    round = _mm256_permute2x128_si256(round, round, 0x11);
+    quant = _mm256_permute2x128_si256(quant, quant, 0x11);
+    shift = _mm256_permute2x128_si256(shift, shift, 0x11);
+    highbd_calculate_qcoeff_avx2(&qcoeff1, &round, &quant, &shift, &log_scale);
+    // Reinsert signs
+    qcoeff0 = _mm256_sign_epi32(qcoeff0, coeff0);
+    qcoeff1 = _mm256_sign_epi32(qcoeff1, coeff1);
+    // Mask out zbin threshold coeffs
+    qcoeff0 = _mm256_and_si256(qcoeff0, temp0);
+    qcoeff1 = _mm256_and_si256(qcoeff1, temp1);
+    highbd_store_coefficients_avx2(qcoeff0, qcoeff1, qcoeff_ptr);
+    coeff0 =
+        highbd_calculate_dqcoeff_log_scale_avx2(qcoeff0, dequant, log_scale);
+    dequant = _mm256_permute2x128_si256(dequant, dequant, 0x11);
+    coeff1 =
+        highbd_calculate_dqcoeff_log_scale_avx2(qcoeff1, dequant, log_scale);
+    highbd_store_coefficients_avx2(coeff0, coeff1, dqcoeff_ptr);
+  }
+
+  // AC only loop.
+  while (index < n_coeffs) {
+    coeff0 = _mm256_load_si256((__m256i *)(coeff_ptr + index));
+    qcoeff0 = _mm256_abs_epi32(coeff0);
+    coeff1 = _mm256_load_si256((__m256i *)(coeff_ptr + index + 8));
+    qcoeff1 = _mm256_abs_epi32(coeff1);
+    highbd_update_mask0_avx2(&qcoeff0, &qcoeff1, threshold, iscan + index,
+                             &is_found0, &mask0);
+    temp0 = _mm256_cmpgt_epi32(qcoeff0, zbin);
+    temp1 = _mm256_cmpgt_epi32(qcoeff1, zbin);
+    cmp_mask = _mm256_permute4x64_epi64(_mm256_packs_epi32(temp0, temp1), 0xd8);
+    highbd_update_mask1_avx2(&cmp_mask, iscan + index, &is_found1, &mask1);
+    if (_mm256_movemask_epi8(cmp_mask) == 0) {
+      _mm256_store_si256((__m256i *)(qcoeff_ptr + index), zero);
+      _mm256_store_si256((__m256i *)(qcoeff_ptr + index + 8), zero);
+      _mm256_store_si256((__m256i *)(dqcoeff_ptr + index), zero);
+      _mm256_store_si256((__m256i *)(dqcoeff_ptr + index + 8), zero);
+      index += 16;
+      continue;
+    }
+    highbd_calculate_qcoeff_avx2(&qcoeff0, &round, &quant, &shift, &log_scale);
+    highbd_calculate_qcoeff_avx2(&qcoeff1, &round, &quant, &shift, &log_scale);
+    qcoeff0 = _mm256_sign_epi32(qcoeff0, coeff0);
+    qcoeff1 = _mm256_sign_epi32(qcoeff1, coeff1);
+    qcoeff0 = _mm256_and_si256(qcoeff0, temp0);
+    qcoeff1 = _mm256_and_si256(qcoeff1, temp1);
+    highbd_store_coefficients_avx2(qcoeff0, qcoeff1, qcoeff_ptr + index);
+    coeff0 =
+        highbd_calculate_dqcoeff_log_scale_avx2(qcoeff0, dequant, log_scale);
+    coeff1 =
+        highbd_calculate_dqcoeff_log_scale_avx2(qcoeff1, dequant, log_scale);
+    highbd_store_coefficients_avx2(coeff0, coeff1, dqcoeff_ptr + index);
+    index += 16;
+  }
+  if (is_found0) {
+    temp_mask0 = _mm_max_epi16(_mm256_castsi256_si128(mask0),
+                               _mm256_extracti128_si256(mask0, 1));
+    non_zero_count = calculate_non_zero_count(temp_mask0);
+  }
+  if (is_found1) {
+    temp_mask1 = _mm_max_epi16(_mm256_castsi256_si128(mask1),
+                               _mm256_extracti128_si256(mask1, 1));
+    non_zero_count_prescan_add_zero = calculate_non_zero_count(temp_mask1);
+  }
+
+  for (int i = non_zero_count_prescan_add_zero - 1; i >= non_zero_count; i--) {
+    const int rc = scan[i];
+    qcoeff_ptr[rc] = 0;
+    dqcoeff_ptr[rc] = 0;
+  }
+
+  for (int i = non_zero_count - 1; i >= 0; i--) {
+    const int rc = scan[i];
+    if (qcoeff_ptr[rc]) {
+      eob = i;
+      break;
+    }
+  }
+
+  *eob_ptr = eob + 1;
+#if SKIP_EOB_FACTOR_ADJUST
+  // TODO(Aniket): Experiment the following loop with intrinsic by combining
+  // with the quantization loop above
+  for (int i = 0; i < non_zero_count; i++) {
+    const int rc = scan[i];
+    const int qcoeff = qcoeff_ptr[rc];
+    if (qcoeff) {
+      first = i;
+      break;
+    }
+  }
+  if ((*eob_ptr - 1) >= 0 && first == (*eob_ptr - 1)) {
+    const int rc = scan[(*eob_ptr - 1)];
+    if (qcoeff_ptr[rc] == 1 || qcoeff_ptr[rc] == -1) {
+      const int coeff = coeff_ptr[rc] * wt;
+      const int coeff_sign = AOMSIGN(coeff);
+      const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+      const int factor = EOB_FACTOR + SKIP_EOB_FACTOR_ADJUST;
+      const int prescan_add_val =
+          ROUND_POWER_OF_TWO(dequant_ptr[rc != 0] * factor, 7);
+      if (abs_coeff < (zbins[rc != 0] * (1 << AOM_QM_BITS) + prescan_add_val)) {
+        qcoeff_ptr[rc] = 0;
+        dqcoeff_ptr[rc] = 0;
+        *eob_ptr = 0;
+      }
+    }
+  }
+#endif
+}
diff --git a/third_party/aom/aom_dsp/x86/highbd_adaptive_quantize_sse2.c b/third_party/aom/aom_dsp/x86/highbd_adaptive_quantize_sse2.c
new file mode 100644
index 0000000000..ae31116e9d
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/highbd_adaptive_quantize_sse2.c
@@ -0,0 +1,732 @@
+/*
+ * Copyright (c) 2019, 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 <emmintrin.h>
+#include "config/aom_dsp_rtcd.h"
+
+#include "aom/aom_integer.h"
+#include "aom_dsp/quantize.h"
+#include "aom_dsp/x86/quantize_x86.h"
+
+static INLINE __m128i highbd_invert_sign_64bit_sse2(__m128i a, __m128i sign) {
+  a = _mm_xor_si128(a, sign);
+  return _mm_sub_epi64(a, sign);
+}
+
+static INLINE void highbd_mul_shift_sse2(const __m128i *x, const __m128i *y,
+                                         __m128i *p, const int shift) {
+  __m128i sign = _mm_srai_epi32(*y, 31);
+  __m128i sign_lo = _mm_unpacklo_epi32(sign, sign);
+  __m128i sign_hi = _mm_unpackhi_epi32(sign, sign);
+  __m128i abs_y = invert_sign_32_sse2(*y, sign);
+  __m128i prod_lo = _mm_mul_epu32(*x, abs_y);
+  __m128i prod_hi = _mm_srli_epi64(*x, 32);
+  const __m128i mult_hi = _mm_srli_epi64(abs_y, 32);
+  prod_hi = _mm_mul_epu32(prod_hi, mult_hi);
+  prod_lo = highbd_invert_sign_64bit_sse2(prod_lo, sign_lo);
+  prod_hi = highbd_invert_sign_64bit_sse2(prod_hi, sign_hi);
+
+  prod_lo = _mm_srli_epi64(prod_lo, shift);
+  const __m128i mask = _mm_set_epi32(0, -1, 0, -1);
+  prod_lo = _mm_and_si128(prod_lo, mask);
+  prod_hi = _mm_srli_epi64(prod_hi, shift);
+
+  prod_hi = _mm_slli_epi64(prod_hi, 32);
+  *p = _mm_or_si128(prod_lo, prod_hi);
+}
+
+static INLINE void highbd_calculate_qcoeff(__m128i *coeff, const __m128i *round,
+                                           const __m128i *quant,
+                                           const __m128i *shift,
+                                           const int *log_scale) {
+  __m128i tmp, qcoeff;
+  qcoeff = _mm_add_epi32(*coeff, *round);
+  highbd_mul_shift_sse2(&qcoeff, quant, &tmp, 16);
+  qcoeff = _mm_add_epi32(tmp, qcoeff);
+  highbd_mul_shift_sse2(&qcoeff, shift, coeff, 16 - *log_scale);
+}
+
+static INLINE void highbd_update_mask1(__m128i *cmp_mask0,
+                                       const int16_t *iscan_ptr, int *is_found,
+                                       __m128i *mask) {
+  __m128i temp_mask = _mm_setzero_si128();
+  if (_mm_movemask_epi8(*cmp_mask0)) {
+    __m128i iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr));
+    __m128i mask0 = _mm_and_si128(*cmp_mask0, iscan0);
+    temp_mask = mask0;
+    *is_found = 1;
+  }
+  *mask = _mm_max_epi16(temp_mask, *mask);
+}
+
+static INLINE void highbd_update_mask0(__m128i *qcoeff0, __m128i *qcoeff1,
+                                       __m128i *threshold,
+                                       const int16_t *iscan_ptr, int *is_found,
+                                       __m128i *mask) {
+  __m128i coeff[2], cmp_mask0, cmp_mask1;
+
+  coeff[0] = _mm_slli_epi32(*qcoeff0, AOM_QM_BITS);
+  cmp_mask0 = _mm_cmpgt_epi32(coeff[0], threshold[0]);
+  coeff[1] = _mm_slli_epi32(*qcoeff1, AOM_QM_BITS);
+  cmp_mask1 = _mm_cmpgt_epi32(coeff[1], threshold[1]);
+
+  cmp_mask0 = _mm_packs_epi32(cmp_mask0, cmp_mask1);
+
+  highbd_update_mask1(&cmp_mask0, iscan_ptr, is_found, mask);
+}
+
+static INLINE __m128i highbd_calculate_dqcoeff(__m128i qcoeff, __m128i dequant,
+                                               const int log_scale) {
+  __m128i coeff_sign = _mm_srai_epi32(qcoeff, 31);
+  __m128i abs_coeff = invert_sign_32_sse2(qcoeff, coeff_sign);
+  highbd_mul_shift_sse2(&abs_coeff, &dequant, &abs_coeff, log_scale);
+  return invert_sign_32_sse2(abs_coeff, coeff_sign);
+}
+
+void aom_highbd_quantize_b_adaptive_sse2(
+    const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr,
+    const int16_t *round_ptr, const int16_t *quant_ptr,
+    const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
+    tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
+    const int16_t *scan, const int16_t *iscan) {
+  int index = 8;
+  const int log_scale = 0;
+  int non_zero_count = 0;
+  int non_zero_count_prescan_add_zero = 0;
+  int is_found0 = 0, is_found1 = 0;
+  int eob = -1;
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i one = _mm_set1_epi32(1);
+  __m128i zbin, round, quant, dequant, shift;
+  __m128i coeff0, coeff1, coeff0_sign, coeff1_sign;
+  __m128i qcoeff0, qcoeff1;
+  __m128i cmp_mask0, cmp_mask1, cmp_mask;
+  __m128i all_zero;
+  __m128i mask0 = zero, mask1 = zero;
+
+  int prescan_add[2];
+  int thresh[4];
+  const qm_val_t wt = (1 << AOM_QM_BITS);
+  for (int i = 0; i < 2; ++i) {
+    prescan_add[i] = ROUND_POWER_OF_TWO(dequant_ptr[i] * EOB_FACTOR, 7);
+    thresh[i] = (zbin_ptr[i] * wt + prescan_add[i]) - 1;
+  }
+  thresh[2] = thresh[3] = thresh[1];
+  __m128i threshold[2];
+  threshold[0] = _mm_loadu_si128((__m128i *)&thresh[0]);
+  threshold[1] = _mm_unpackhi_epi64(threshold[0], threshold[0]);
+
+#if SKIP_EOB_FACTOR_ADJUST
+  int first = -1;
+#endif
+  // Setup global values.
+  zbin = _mm_load_si128((const __m128i *)zbin_ptr);
+  round = _mm_load_si128((const __m128i *)round_ptr);
+  quant = _mm_load_si128((const __m128i *)quant_ptr);
+  dequant = _mm_load_si128((const __m128i *)dequant_ptr);
+  shift = _mm_load_si128((const __m128i *)quant_shift_ptr);
+
+  __m128i zbin_sign = _mm_srai_epi16(zbin, 15);
+  __m128i round_sign = _mm_srai_epi16(round, 15);
+  __m128i quant_sign = _mm_srai_epi16(quant, 15);
+  __m128i dequant_sign = _mm_srai_epi16(dequant, 15);
+  __m128i shift_sign = _mm_srai_epi16(shift, 15);
+
+  zbin = _mm_unpacklo_epi16(zbin, zbin_sign);
+  round = _mm_unpacklo_epi16(round, round_sign);
+  quant = _mm_unpacklo_epi16(quant, quant_sign);
+  dequant = _mm_unpacklo_epi16(dequant, dequant_sign);
+  shift = _mm_unpacklo_epi16(shift, shift_sign);
+  zbin = _mm_sub_epi32(zbin, one);
+
+  // Do DC and first 15 AC.
+  coeff0 = _mm_load_si128((__m128i *)(coeff_ptr));
+  coeff1 = _mm_load_si128((__m128i *)(coeff_ptr + 4));
+
+  coeff0_sign = _mm_srai_epi32(coeff0, 31);
+  coeff1_sign = _mm_srai_epi32(coeff1, 31);
+  qcoeff0 = invert_sign_32_sse2(coeff0, coeff0_sign);
+  qcoeff1 = invert_sign_32_sse2(coeff1, coeff1_sign);
+
+  highbd_update_mask0(&qcoeff0, &qcoeff1, threshold, iscan, &is_found0, &mask0);
+
+  cmp_mask0 = _mm_cmpgt_epi32(qcoeff0, zbin);
+  zbin = _mm_unpackhi_epi64(zbin, zbin);  // Switch DC to AC
+  cmp_mask1 = _mm_cmpgt_epi32(qcoeff1, zbin);
+  cmp_mask = _mm_packs_epi32(cmp_mask0, cmp_mask1);
+  highbd_update_mask1(&cmp_mask, iscan, &is_found1, &mask1);
+
+  threshold[0] = threshold[1];
+  all_zero = _mm_or_si128(cmp_mask0, cmp_mask1);
+  if (_mm_movemask_epi8(all_zero) == 0) {
+    _mm_store_si128((__m128i *)(qcoeff_ptr), zero);
+    _mm_store_si128((__m128i *)(qcoeff_ptr + 4), zero);
+    _mm_store_si128((__m128i *)(dqcoeff_ptr), zero);
+    _mm_store_si128((__m128i *)(dqcoeff_ptr + 4), zero);
+
+    round = _mm_unpackhi_epi64(round, round);
+    quant = _mm_unpackhi_epi64(quant, quant);
+    shift = _mm_unpackhi_epi64(shift, shift);
+    dequant = _mm_unpackhi_epi64(dequant, dequant);
+  } else {
+    highbd_calculate_qcoeff(&qcoeff0, &round, &quant, &shift, &log_scale);
+
+    round = _mm_unpackhi_epi64(round, round);
+    quant = _mm_unpackhi_epi64(quant, quant);
+    shift = _mm_unpackhi_epi64(shift, shift);
+    highbd_calculate_qcoeff(&qcoeff1, &round, &quant, &shift, &log_scale);
+
+    // Reinsert signs
+    qcoeff0 = invert_sign_32_sse2(qcoeff0, coeff0_sign);
+    qcoeff1 = invert_sign_32_sse2(qcoeff1, coeff1_sign);
+
+    // Mask out zbin threshold coeffs
+    qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
+    qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
+
+    _mm_store_si128((__m128i *)(qcoeff_ptr), qcoeff0);
+    _mm_store_si128((__m128i *)(qcoeff_ptr + 4), qcoeff1);
+
+    coeff0 = highbd_calculate_dqcoeff(qcoeff0, dequant, log_scale);
+    dequant = _mm_unpackhi_epi64(dequant, dequant);
+    coeff1 = highbd_calculate_dqcoeff(qcoeff1, dequant, log_scale);
+    _mm_store_si128((__m128i *)(dqcoeff_ptr), coeff0);
+    _mm_store_si128((__m128i *)(dqcoeff_ptr + 4), coeff1);
+  }
+
+  // AC only loop.
+  while (index < n_coeffs) {
+    coeff0 = _mm_load_si128((__m128i *)(coeff_ptr + index));
+    coeff1 = _mm_load_si128((__m128i *)(coeff_ptr + index + 4));
+
+    coeff0_sign = _mm_srai_epi32(coeff0, 31);
+    coeff1_sign = _mm_srai_epi32(coeff1, 31);
+    qcoeff0 = invert_sign_32_sse2(coeff0, coeff0_sign);
+    qcoeff1 = invert_sign_32_sse2(coeff1, coeff1_sign);
+
+    highbd_update_mask0(&qcoeff0, &qcoeff1, threshold, iscan + index,
+                        &is_found0, &mask0);
+
+    cmp_mask0 = _mm_cmpgt_epi32(qcoeff0, zbin);
+    cmp_mask1 = _mm_cmpgt_epi32(qcoeff1, zbin);
+    cmp_mask = _mm_packs_epi32(cmp_mask0, cmp_mask1);
+    highbd_update_mask1(&cmp_mask, iscan + index, &is_found1, &mask1);
+
+    all_zero = _mm_or_si128(cmp_mask0, cmp_mask1);
+    if (_mm_movemask_epi8(all_zero) == 0) {
+      _mm_store_si128((__m128i *)(qcoeff_ptr + index), zero);
+      _mm_store_si128((__m128i *)(qcoeff_ptr + index + 4), zero);
+      _mm_store_si128((__m128i *)(dqcoeff_ptr + index), zero);
+      _mm_store_si128((__m128i *)(dqcoeff_ptr + index + 4), zero);
+      index += 8;
+      continue;
+    }
+    highbd_calculate_qcoeff(&qcoeff0, &round, &quant, &shift, &log_scale);
+    highbd_calculate_qcoeff(&qcoeff1, &round, &quant, &shift, &log_scale);
+
+    qcoeff0 = invert_sign_32_sse2(qcoeff0, coeff0_sign);
+    qcoeff1 = invert_sign_32_sse2(qcoeff1, coeff1_sign);
+
+    qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
+    qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
+
+    _mm_store_si128((__m128i *)(qcoeff_ptr + index), qcoeff0);
+    _mm_store_si128((__m128i *)(qcoeff_ptr + index + 4), qcoeff1);
+
+    coeff0 = highbd_calculate_dqcoeff(qcoeff0, dequant, log_scale);
+    coeff1 = highbd_calculate_dqcoeff(qcoeff1, dequant, log_scale);
+
+    _mm_store_si128((__m128i *)(dqcoeff_ptr + index), coeff0);
+    _mm_store_si128((__m128i *)(dqcoeff_ptr + index + 4), coeff1);
+
+    index += 8;
+  }
+  if (is_found0) non_zero_count = calculate_non_zero_count(mask0);
+  if (is_found1)
+    non_zero_count_prescan_add_zero = calculate_non_zero_count(mask1);
+
+  for (int i = non_zero_count_prescan_add_zero - 1; i >= non_zero_count; i--) {
+    const int rc = scan[i];
+    qcoeff_ptr[rc] = 0;
+    dqcoeff_ptr[rc] = 0;
+  }
+
+  for (int i = non_zero_count - 1; i >= 0; i--) {
+    const int rc = scan[i];
+    if (qcoeff_ptr[rc]) {
+      eob = i;
+      break;
+    }
+  }
+
+  *eob_ptr = eob + 1;
+#if SKIP_EOB_FACTOR_ADJUST
+  // TODO(Aniket): Experiment the following loop with intrinsic by combining
+  // with the quantization loop above
+  for (int i = 0; i < non_zero_count; i++) {
+    const int rc = scan[i];
+    const int qcoeff = qcoeff_ptr[rc];
+    if (qcoeff) {
+      first = i;
+      break;
+    }
+  }
+  if ((*eob_ptr - 1) >= 0 && first == (*eob_ptr - 1)) {
+    const int rc = scan[(*eob_ptr - 1)];
+    if (qcoeff_ptr[rc] == 1 || qcoeff_ptr[rc] == -1) {
+      const int coeff = coeff_ptr[rc] * wt;
+      const int coeff_sign = AOMSIGN(coeff);
+      const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+      const int factor = EOB_FACTOR + SKIP_EOB_FACTOR_ADJUST;
+      const int prescan_add_val =
+          ROUND_POWER_OF_TWO(dequant_ptr[rc != 0] * factor, 7);
+      if (abs_coeff <
+          (zbin_ptr[rc != 0] * (1 << AOM_QM_BITS) + prescan_add_val)) {
+        qcoeff_ptr[rc] = 0;
+        dqcoeff_ptr[rc] = 0;
+        *eob_ptr = 0;
+      }
+    }
+  }
+#endif
+}
+
+void aom_highbd_quantize_b_32x32_adaptive_sse2(
+    const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr,
+    const int16_t *round_ptr, const int16_t *quant_ptr,
+    const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
+    tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
+    const int16_t *scan, const int16_t *iscan) {
+  int index = 8;
+  const int log_scale = 1;
+  int non_zero_count = 0;
+  int non_zero_count_prescan_add_zero = 0;
+  int is_found0 = 0, is_found1 = 0;
+  int eob = -1;
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i one = _mm_set1_epi32(1);
+  const __m128i log_scale_vec = _mm_set1_epi32(log_scale);
+  __m128i zbin, round, quant, dequant, shift;
+  __m128i coeff0, coeff1, coeff0_sign, coeff1_sign;
+  __m128i qcoeff0, qcoeff1;
+  __m128i cmp_mask0, cmp_mask1, cmp_mask;
+  __m128i all_zero;
+  __m128i mask0 = zero, mask1 = zero;
+
+  const int zbins[2] = { ROUND_POWER_OF_TWO(zbin_ptr[0], log_scale),
+                         ROUND_POWER_OF_TWO(zbin_ptr[1], log_scale) };
+  int prescan_add[2];
+  int thresh[4];
+  const qm_val_t wt = (1 << AOM_QM_BITS);
+  for (int i = 0; i < 2; ++i) {
+    prescan_add[i] = ROUND_POWER_OF_TWO(dequant_ptr[i] * EOB_FACTOR, 7);
+    thresh[i] = (zbins[i] * wt + prescan_add[i]) - 1;
+  }
+  thresh[2] = thresh[3] = thresh[1];
+  __m128i threshold[2];
+  threshold[0] = _mm_loadu_si128((__m128i *)&thresh[0]);
+  threshold[1] = _mm_unpackhi_epi64(threshold[0], threshold[0]);
+
+#if SKIP_EOB_FACTOR_ADJUST
+  int first = -1;
+#endif
+  // Setup global values.
+  zbin = _mm_load_si128((const __m128i *)zbin_ptr);
+  round = _mm_load_si128((const __m128i *)round_ptr);
+  quant = _mm_load_si128((const __m128i *)quant_ptr);
+  dequant = _mm_load_si128((const __m128i *)dequant_ptr);
+  shift = _mm_load_si128((const __m128i *)quant_shift_ptr);
+
+  __m128i zbin_sign = _mm_srai_epi16(zbin, 15);
+  __m128i round_sign = _mm_srai_epi16(round, 15);
+  __m128i quant_sign = _mm_srai_epi16(quant, 15);
+  __m128i dequant_sign = _mm_srai_epi16(dequant, 15);
+  __m128i shift_sign = _mm_srai_epi16(shift, 15);
+
+  zbin = _mm_unpacklo_epi16(zbin, zbin_sign);
+  round = _mm_unpacklo_epi16(round, round_sign);
+  quant = _mm_unpacklo_epi16(quant, quant_sign);
+  dequant = _mm_unpacklo_epi16(dequant, dequant_sign);
+  shift = _mm_unpacklo_epi16(shift, shift_sign);
+
+  // Shift with rounding.
+  zbin = _mm_add_epi32(zbin, log_scale_vec);
+  round = _mm_add_epi32(round, log_scale_vec);
+  zbin = _mm_srli_epi32(zbin, log_scale);
+  round = _mm_srli_epi32(round, log_scale);
+  zbin = _mm_sub_epi32(zbin, one);
+
+  // Do DC and first 15 AC.
+  coeff0 = _mm_load_si128((__m128i *)(coeff_ptr));
+  coeff1 = _mm_load_si128((__m128i *)(coeff_ptr + 4));
+
+  coeff0_sign = _mm_srai_epi32(coeff0, 31);
+  coeff1_sign = _mm_srai_epi32(coeff1, 31);
+  qcoeff0 = invert_sign_32_sse2(coeff0, coeff0_sign);
+  qcoeff1 = invert_sign_32_sse2(coeff1, coeff1_sign);
+
+  highbd_update_mask0(&qcoeff0, &qcoeff1, threshold, iscan, &is_found0, &mask0);
+
+  cmp_mask0 = _mm_cmpgt_epi32(qcoeff0, zbin);
+  zbin = _mm_unpackhi_epi64(zbin, zbin);  // Switch DC to AC
+  cmp_mask1 = _mm_cmpgt_epi32(qcoeff1, zbin);
+  cmp_mask = _mm_packs_epi32(cmp_mask0, cmp_mask1);
+  highbd_update_mask1(&cmp_mask, iscan, &is_found1, &mask1);
+
+  threshold[0] = threshold[1];
+  all_zero = _mm_or_si128(cmp_mask0, cmp_mask1);
+  if (_mm_movemask_epi8(all_zero) == 0) {
+    _mm_store_si128((__m128i *)(qcoeff_ptr), zero);
+    _mm_store_si128((__m128i *)(qcoeff_ptr + 4), zero);
+    _mm_store_si128((__m128i *)(dqcoeff_ptr), zero);
+    _mm_store_si128((__m128i *)(dqcoeff_ptr + 4), zero);
+
+    round = _mm_unpackhi_epi64(round, round);
+    quant = _mm_unpackhi_epi64(quant, quant);
+    shift = _mm_unpackhi_epi64(shift, shift);
+    dequant = _mm_unpackhi_epi64(dequant, dequant);
+  } else {
+    highbd_calculate_qcoeff(&qcoeff0, &round, &quant, &shift, &log_scale);
+
+    round = _mm_unpackhi_epi64(round, round);
+    quant = _mm_unpackhi_epi64(quant, quant);
+    shift = _mm_unpackhi_epi64(shift, shift);
+    highbd_calculate_qcoeff(&qcoeff1, &round, &quant, &shift, &log_scale);
+
+    // Reinsert signs
+    qcoeff0 = invert_sign_32_sse2(qcoeff0, coeff0_sign);
+    qcoeff1 = invert_sign_32_sse2(qcoeff1, coeff1_sign);
+
+    // Mask out zbin threshold coeffs
+    qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
+    qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
+
+    _mm_store_si128((__m128i *)(qcoeff_ptr), qcoeff0);
+    _mm_store_si128((__m128i *)(qcoeff_ptr + 4), qcoeff1);
+
+    coeff0 = highbd_calculate_dqcoeff(qcoeff0, dequant, log_scale);
+    dequant = _mm_unpackhi_epi64(dequant, dequant);
+    coeff1 = highbd_calculate_dqcoeff(qcoeff1, dequant, log_scale);
+    _mm_store_si128((__m128i *)(dqcoeff_ptr), coeff0);
+    _mm_store_si128((__m128i *)(dqcoeff_ptr + 4), coeff1);
+  }
+
+  // AC only loop.
+  while (index < n_coeffs) {
+    coeff0 = _mm_load_si128((__m128i *)(coeff_ptr + index));
+    coeff1 = _mm_load_si128((__m128i *)(coeff_ptr + index + 4));
+
+    coeff0_sign = _mm_srai_epi32(coeff0, 31);
+    coeff1_sign = _mm_srai_epi32(coeff1, 31);
+    qcoeff0 = invert_sign_32_sse2(coeff0, coeff0_sign);
+    qcoeff1 = invert_sign_32_sse2(coeff1, coeff1_sign);
+
+    highbd_update_mask0(&qcoeff0, &qcoeff1, threshold, iscan + index,
+                        &is_found0, &mask0);
+
+    cmp_mask0 = _mm_cmpgt_epi32(qcoeff0, zbin);
+    cmp_mask1 = _mm_cmpgt_epi32(qcoeff1, zbin);
+    cmp_mask = _mm_packs_epi32(cmp_mask0, cmp_mask1);
+    highbd_update_mask1(&cmp_mask, iscan + index, &is_found1, &mask1);
+
+    all_zero = _mm_or_si128(cmp_mask0, cmp_mask1);
+    if (_mm_movemask_epi8(all_zero) == 0) {
+      _mm_store_si128((__m128i *)(qcoeff_ptr + index), zero);
+      _mm_store_si128((__m128i *)(qcoeff_ptr + index + 4), zero);
+      _mm_store_si128((__m128i *)(dqcoeff_ptr + index), zero);
+      _mm_store_si128((__m128i *)(dqcoeff_ptr + index + 4), zero);
+      index += 8;
+      continue;
+    }
+    highbd_calculate_qcoeff(&qcoeff0, &round, &quant, &shift, &log_scale);
+    highbd_calculate_qcoeff(&qcoeff1, &round, &quant, &shift, &log_scale);
+
+    qcoeff0 = invert_sign_32_sse2(qcoeff0, coeff0_sign);
+    qcoeff1 = invert_sign_32_sse2(qcoeff1, coeff1_sign);
+
+    qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
+    qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
+
+    _mm_store_si128((__m128i *)(qcoeff_ptr + index), qcoeff0);
+    _mm_store_si128((__m128i *)(qcoeff_ptr + index + 4), qcoeff1);
+
+    coeff0 = highbd_calculate_dqcoeff(qcoeff0, dequant, log_scale);
+    coeff1 = highbd_calculate_dqcoeff(qcoeff1, dequant, log_scale);
+
+    _mm_store_si128((__m128i *)(dqcoeff_ptr + index), coeff0);
+    _mm_store_si128((__m128i *)(dqcoeff_ptr + index + 4), coeff1);
+
+    index += 8;
+  }
+  if (is_found0) non_zero_count = calculate_non_zero_count(mask0);
+  if (is_found1)
+    non_zero_count_prescan_add_zero = calculate_non_zero_count(mask1);
+
+  for (int i = non_zero_count_prescan_add_zero - 1; i >= non_zero_count; i--) {
+    const int rc = scan[i];
+    qcoeff_ptr[rc] = 0;
+    dqcoeff_ptr[rc] = 0;
+  }
+
+  for (int i = non_zero_count - 1; i >= 0; i--) {
+    const int rc = scan[i];
+    if (qcoeff_ptr[rc]) {
+      eob = i;
+      break;
+    }
+  }
+
+  *eob_ptr = eob + 1;
+#if SKIP_EOB_FACTOR_ADJUST
+  // TODO(Aniket): Experiment the following loop with intrinsic by combining
+  // with the quantization loop above
+  for (int i = 0; i < non_zero_count; i++) {
+    const int rc = scan[i];
+    const int qcoeff = qcoeff_ptr[rc];
+    if (qcoeff) {
+      first = i;
+      break;
+    }
+  }
+  if ((*eob_ptr - 1) >= 0 && first == (*eob_ptr - 1)) {
+    const int rc = scan[(*eob_ptr - 1)];
+    if (qcoeff_ptr[rc] == 1 || qcoeff_ptr[rc] == -1) {
+      const int coeff = coeff_ptr[rc] * wt;
+      const int coeff_sign = AOMSIGN(coeff);
+      const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+      const int factor = EOB_FACTOR + SKIP_EOB_FACTOR_ADJUST;
+      const int prescan_add_val =
+          ROUND_POWER_OF_TWO(dequant_ptr[rc != 0] * factor, 7);
+      if (abs_coeff < (zbins[rc != 0] * (1 << AOM_QM_BITS) + prescan_add_val)) {
+        qcoeff_ptr[rc] = 0;
+        dqcoeff_ptr[rc] = 0;
+        *eob_ptr = 0;
+      }
+    }
+  }
+#endif
+}
+
+void aom_highbd_quantize_b_64x64_adaptive_sse2(
+    const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr,
+    const int16_t *round_ptr, const int16_t *quant_ptr,
+    const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
+    tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
+    const int16_t *scan, const int16_t *iscan) {
+  int index = 8;
+  const int log_scale = 2;
+  int non_zero_count = 0;
+  int non_zero_count_prescan_add_zero = 0;
+  int is_found0 = 0, is_found1 = 0;
+  int eob = -1;
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i one = _mm_set1_epi32(1);
+  const __m128i log_scale_vec = _mm_set1_epi32(log_scale);
+  __m128i zbin, round, quant, dequant, shift;
+  __m128i coeff0, coeff1, coeff0_sign, coeff1_sign;
+  __m128i qcoeff0, qcoeff1;
+  __m128i cmp_mask0, cmp_mask1, cmp_mask;
+  __m128i all_zero;
+  __m128i mask0 = zero, mask1 = zero;
+
+  const int zbins[2] = { ROUND_POWER_OF_TWO(zbin_ptr[0], log_scale),
+                         ROUND_POWER_OF_TWO(zbin_ptr[1], log_scale) };
+  int prescan_add[2];
+  int thresh[4];
+  const qm_val_t wt = (1 << AOM_QM_BITS);
+  for (int i = 0; i < 2; ++i) {
+    prescan_add[i] = ROUND_POWER_OF_TWO(dequant_ptr[i] * EOB_FACTOR, 7);
+    thresh[i] = (zbins[i] * wt + prescan_add[i]) - 1;
+  }
+  thresh[2] = thresh[3] = thresh[1];
+  __m128i threshold[2];
+  threshold[0] = _mm_loadu_si128((__m128i *)&thresh[0]);
+  threshold[1] = _mm_unpackhi_epi64(threshold[0], threshold[0]);
+
+#if SKIP_EOB_FACTOR_ADJUST
+  int first = -1;
+#endif
+  // Setup global values.
+  zbin = _mm_load_si128((const __m128i *)zbin_ptr);
+  round = _mm_load_si128((const __m128i *)round_ptr);
+  quant = _mm_load_si128((const __m128i *)quant_ptr);
+  dequant = _mm_load_si128((const __m128i *)dequant_ptr);
+  shift = _mm_load_si128((const __m128i *)quant_shift_ptr);
+
+  __m128i zbin_sign = _mm_srai_epi16(zbin, 15);
+  __m128i round_sign = _mm_srai_epi16(round, 15);
+  __m128i quant_sign = _mm_srai_epi16(quant, 15);
+  __m128i dequant_sign = _mm_srai_epi16(dequant, 15);
+  __m128i shift_sign = _mm_srai_epi16(shift, 15);
+
+  zbin = _mm_unpacklo_epi16(zbin, zbin_sign);
+  round = _mm_unpacklo_epi16(round, round_sign);
+  quant = _mm_unpacklo_epi16(quant, quant_sign);
+  dequant = _mm_unpacklo_epi16(dequant, dequant_sign);
+  shift = _mm_unpacklo_epi16(shift, shift_sign);
+
+  // Shift with rounding.
+  zbin = _mm_add_epi32(zbin, log_scale_vec);
+  round = _mm_add_epi32(round, log_scale_vec);
+  zbin = _mm_srli_epi32(zbin, log_scale);
+  round = _mm_srli_epi32(round, log_scale);
+  zbin = _mm_sub_epi32(zbin, one);
+
+  // Do DC and first 15 AC.
+  coeff0 = _mm_load_si128((__m128i *)(coeff_ptr));
+  coeff1 = _mm_load_si128((__m128i *)(coeff_ptr + 4));
+
+  coeff0_sign = _mm_srai_epi32(coeff0, 31);
+  coeff1_sign = _mm_srai_epi32(coeff1, 31);
+  qcoeff0 = invert_sign_32_sse2(coeff0, coeff0_sign);
+  qcoeff1 = invert_sign_32_sse2(coeff1, coeff1_sign);
+
+  highbd_update_mask0(&qcoeff0, &qcoeff1, threshold, iscan, &is_found0, &mask0);
+
+  cmp_mask0 = _mm_cmpgt_epi32(qcoeff0, zbin);
+  zbin = _mm_unpackhi_epi64(zbin, zbin);  // Switch DC to AC
+  cmp_mask1 = _mm_cmpgt_epi32(qcoeff1, zbin);
+  cmp_mask = _mm_packs_epi32(cmp_mask0, cmp_mask1);
+  highbd_update_mask1(&cmp_mask, iscan, &is_found1, &mask1);
+
+  threshold[0] = threshold[1];
+  all_zero = _mm_or_si128(cmp_mask0, cmp_mask1);
+  if (_mm_movemask_epi8(all_zero) == 0) {
+    _mm_store_si128((__m128i *)(qcoeff_ptr), zero);
+    _mm_store_si128((__m128i *)(qcoeff_ptr + 4), zero);
+    _mm_store_si128((__m128i *)(dqcoeff_ptr), zero);
+    _mm_store_si128((__m128i *)(dqcoeff_ptr + 4), zero);
+
+    round = _mm_unpackhi_epi64(round, round);
+    quant = _mm_unpackhi_epi64(quant, quant);
+    shift = _mm_unpackhi_epi64(shift, shift);
+    dequant = _mm_unpackhi_epi64(dequant, dequant);
+  } else {
+    highbd_calculate_qcoeff(&qcoeff0, &round, &quant, &shift, &log_scale);
+
+    round = _mm_unpackhi_epi64(round, round);
+    quant = _mm_unpackhi_epi64(quant, quant);
+    shift = _mm_unpackhi_epi64(shift, shift);
+    highbd_calculate_qcoeff(&qcoeff1, &round, &quant, &shift, &log_scale);
+
+    // Reinsert signs
+    qcoeff0 = invert_sign_32_sse2(qcoeff0, coeff0_sign);
+    qcoeff1 = invert_sign_32_sse2(qcoeff1, coeff1_sign);
+
+    // Mask out zbin threshold coeffs
+    qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
+    qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
+
+    _mm_store_si128((__m128i *)(qcoeff_ptr), qcoeff0);
+    _mm_store_si128((__m128i *)(qcoeff_ptr + 4), qcoeff1);
+
+    coeff0 = highbd_calculate_dqcoeff(qcoeff0, dequant, log_scale);
+    dequant = _mm_unpackhi_epi64(dequant, dequant);
+    coeff1 = highbd_calculate_dqcoeff(qcoeff1, dequant, log_scale);
+    _mm_store_si128((__m128i *)(dqcoeff_ptr), coeff0);
+    _mm_store_si128((__m128i *)(dqcoeff_ptr + 4), coeff1);
+  }
+
+  // AC only loop.
+  while (index < n_coeffs) {
+    coeff0 = _mm_load_si128((__m128i *)(coeff_ptr + index));
+    coeff1 = _mm_load_si128((__m128i *)(coeff_ptr + index + 4));
+
+    coeff0_sign = _mm_srai_epi32(coeff0, 31);
+    coeff1_sign = _mm_srai_epi32(coeff1, 31);
+    qcoeff0 = invert_sign_32_sse2(coeff0, coeff0_sign);
+    qcoeff1 = invert_sign_32_sse2(coeff1, coeff1_sign);
+
+    highbd_update_mask0(&qcoeff0, &qcoeff1, threshold, iscan + index,
+                        &is_found0, &mask0);
+
+    cmp_mask0 = _mm_cmpgt_epi32(qcoeff0, zbin);
+    cmp_mask1 = _mm_cmpgt_epi32(qcoeff1, zbin);
+    cmp_mask = _mm_packs_epi32(cmp_mask0, cmp_mask1);
+    highbd_update_mask1(&cmp_mask, iscan + index, &is_found1, &mask1);
+
+    all_zero = _mm_or_si128(cmp_mask0, cmp_mask1);
+    if (_mm_movemask_epi8(all_zero) == 0) {
+      _mm_store_si128((__m128i *)(qcoeff_ptr + index), zero);
+      _mm_store_si128((__m128i *)(qcoeff_ptr + index + 4), zero);
+      _mm_store_si128((__m128i *)(dqcoeff_ptr + index), zero);
+      _mm_store_si128((__m128i *)(dqcoeff_ptr + index + 4), zero);
+      index += 8;
+      continue;
+    }
+    highbd_calculate_qcoeff(&qcoeff0, &round, &quant, &shift, &log_scale);
+    highbd_calculate_qcoeff(&qcoeff1, &round, &quant, &shift, &log_scale);
+
+    qcoeff0 = invert_sign_32_sse2(qcoeff0, coeff0_sign);
+    qcoeff1 = invert_sign_32_sse2(qcoeff1, coeff1_sign);
+
+    qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
+    qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
+
+    _mm_store_si128((__m128i *)(qcoeff_ptr + index), qcoeff0);
+    _mm_store_si128((__m128i *)(qcoeff_ptr + index + 4), qcoeff1);
+
+    coeff0 = highbd_calculate_dqcoeff(qcoeff0, dequant, log_scale);
+    coeff1 = highbd_calculate_dqcoeff(qcoeff1, dequant, log_scale);
+
+    _mm_store_si128((__m128i *)(dqcoeff_ptr + index), coeff0);
+    _mm_store_si128((__m128i *)(dqcoeff_ptr + index + 4), coeff1);
+
+    index += 8;
+  }
+  if (is_found0) non_zero_count = calculate_non_zero_count(mask0);
+  if (is_found1)
+    non_zero_count_prescan_add_zero = calculate_non_zero_count(mask1);
+
+  for (int i = non_zero_count_prescan_add_zero - 1; i >= non_zero_count; i--) {
+    const int rc = scan[i];
+    qcoeff_ptr[rc] = 0;
+    dqcoeff_ptr[rc] = 0;
+  }
+
+  for (int i = non_zero_count - 1; i >= 0; i--) {
+    const int rc = scan[i];
+    if (qcoeff_ptr[rc]) {
+      eob = i;
+      break;
+    }
+  }
+
+  *eob_ptr = eob + 1;
+#if SKIP_EOB_FACTOR_ADJUST
+  // TODO(Aniket): Experiment the following loop with intrinsic by combining
+  // with the quantization loop above
+  for (int i = 0; i < non_zero_count; i++) {
+    const int rc = scan[i];
+    const int qcoeff = qcoeff_ptr[rc];
+    if (qcoeff) {
+      first = i;
+      break;
+    }
+  }
+  if ((*eob_ptr - 1) >= 0 && first == (*eob_ptr - 1)) {
+    const int rc = scan[(*eob_ptr - 1)];
+    if (qcoeff_ptr[rc] == 1 || qcoeff_ptr[rc] == -1) {
+      const int coeff = coeff_ptr[rc] * wt;
+      const int coeff_sign = AOMSIGN(coeff);
+      const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+      const int factor = EOB_FACTOR + SKIP_EOB_FACTOR_ADJUST;
+      const int prescan_add_val =
+          ROUND_POWER_OF_TWO(dequant_ptr[rc != 0] * factor, 7);
+      if (abs_coeff < (zbins[rc != 0] * (1 << AOM_QM_BITS) + prescan_add_val)) {
+        qcoeff_ptr[rc] = 0;
+        dqcoeff_ptr[rc] = 0;
+        *eob_ptr = 0;
+      }
+    }
+  }
+#endif
+}
diff --git a/third_party/aom/aom_dsp/x86/highbd_convolve_avx2.c b/third_party/aom/aom_dsp/x86/highbd_convolve_avx2.c
new file mode 100644
index 0000000000..11e45778c0
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/highbd_convolve_avx2.c
@@ -0,0 +1,1248 @@
+/*
+ * Copyright (c) 2017, 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 <string.h>
+
+#include "config/av1_rtcd.h"
+
+#include "aom_dsp/x86/convolve.h"
+#include "aom_dsp/x86/convolve_avx2.h"
+#include "aom_dsp/x86/synonyms.h"
+
+// -----------------------------------------------------------------------------
+// Copy and average
+
+static const uint8_t ip_shuffle_f2f3[32] = { 0, 1, 2, 3, 2, 3, 4, 5, 4, 5, 6,
+                                             7, 6, 7, 8, 9, 0, 1, 2, 3, 2, 3,
+                                             4, 5, 4, 5, 6, 7, 6, 7, 8, 9 };
+static const uint8_t ip_shuffle_f4f5[32] = { 4, 5, 6,  7,  6,  7,  8,  9,
+                                             8, 9, 10, 11, 10, 11, 12, 13,
+                                             4, 5, 6,  7,  6,  7,  8,  9,
+                                             8, 9, 10, 11, 10, 11, 12, 13 };
+
+void av1_highbd_convolve_x_sr_ssse3(const uint16_t *src, int src_stride,
+                                    uint16_t *dst, int dst_stride, int w, int h,
+                                    const InterpFilterParams *filter_params_x,
+                                    const int subpel_x_qn,
+                                    ConvolveParams *conv_params, int bd);
+void av1_highbd_convolve_y_sr_ssse3(const uint16_t *src, int src_stride,
+                                    uint16_t *dst, int dst_stride, int w, int h,
+                                    const InterpFilterParams *filter_params_y,
+                                    const int subpel_y_qn, int bd);
+
+void av1_highbd_convolve_y_sr_avx2(const uint16_t *src, int src_stride,
+                                   uint16_t *dst, int dst_stride, int w, int h,
+                                   const InterpFilterParams *filter_params_y,
+                                   const int subpel_y_qn, int bd) {
+  if (filter_params_y->taps == 12) {
+    av1_highbd_convolve_y_sr_ssse3(src, src_stride, dst, dst_stride, w, h,
+                                   filter_params_y, subpel_y_qn, bd);
+    return;
+  }
+  int i, j;
+  const int fo_vert = filter_params_y->taps / 2 - 1;
+  const uint16_t *const src_ptr = src - fo_vert * src_stride;
+
+  __m256i s[8], coeffs_y[4];
+
+  const int bits = FILTER_BITS;
+
+  const __m128i round_shift_bits = _mm_cvtsi32_si128(bits);
+  const __m256i round_const_bits = _mm256_set1_epi32((1 << bits) >> 1);
+  const __m256i clip_pixel =
+      _mm256_set1_epi16(bd == 10 ? 1023 : (bd == 12 ? 4095 : 255));
+  const __m256i zero = _mm256_setzero_si256();
+
+  prepare_coeffs(filter_params_y, subpel_y_qn, coeffs_y);
+
+  for (j = 0; j < w; j += 8) {
+    const uint16_t *data = &src_ptr[j];
+    /* Vertical filter */
+    {
+      __m256i src6;
+      __m256i s01 = _mm256_permute2x128_si256(
+          _mm256_castsi128_si256(
+              _mm_loadu_si128((__m128i *)(data + 0 * src_stride))),
+          _mm256_castsi128_si256(
+              _mm_loadu_si128((__m128i *)(data + 1 * src_stride))),
+          0x20);
+      __m256i s12 = _mm256_permute2x128_si256(
+          _mm256_castsi128_si256(
+              _mm_loadu_si128((__m128i *)(data + 1 * src_stride))),
+          _mm256_castsi128_si256(
+              _mm_loadu_si128((__m128i *)(data + 2 * src_stride))),
+          0x20);
+      __m256i s23 = _mm256_permute2x128_si256(
+          _mm256_castsi128_si256(
+              _mm_loadu_si128((__m128i *)(data + 2 * src_stride))),
+          _mm256_castsi128_si256(
+              _mm_loadu_si128((__m128i *)(data + 3 * src_stride))),
+          0x20);
+      __m256i s34 = _mm256_permute2x128_si256(
+          _mm256_castsi128_si256(
+              _mm_loadu_si128((__m128i *)(data + 3 * src_stride))),
+          _mm256_castsi128_si256(
+              _mm_loadu_si128((__m128i *)(data + 4 * src_stride))),
+          0x20);
+      __m256i s45 = _mm256_permute2x128_si256(
+          _mm256_castsi128_si256(
+              _mm_loadu_si128((__m128i *)(data + 4 * src_stride))),
+          _mm256_castsi128_si256(
+              _mm_loadu_si128((__m128i *)(data + 5 * src_stride))),
+          0x20);
+      src6 = _mm256_castsi128_si256(
+          _mm_loadu_si128((__m128i *)(data + 6 * src_stride)));
+      __m256i s56 = _mm256_permute2x128_si256(
+          _mm256_castsi128_si256(
+              _mm_loadu_si128((__m128i *)(data + 5 * src_stride))),
+          src6, 0x20);
+
+      s[0] = _mm256_unpacklo_epi16(s01, s12);
+      s[1] = _mm256_unpacklo_epi16(s23, s34);
+      s[2] = _mm256_unpacklo_epi16(s45, s56);
+
+      s[4] = _mm256_unpackhi_epi16(s01, s12);
+      s[5] = _mm256_unpackhi_epi16(s23, s34);
+      s[6] = _mm256_unpackhi_epi16(s45, s56);
+
+      for (i = 0; i < h; i += 2) {
+        data = &src_ptr[i * src_stride + j];
+
+        const __m256i s67 = _mm256_permute2x128_si256(
+            src6,
+            _mm256_castsi128_si256(
+                _mm_loadu_si128((__m128i *)(data + 7 * src_stride))),
+            0x20);
+
+        src6 = _mm256_castsi128_si256(
+            _mm_loadu_si128((__m128i *)(data + 8 * src_stride)));
+
+        const __m256i s78 = _mm256_permute2x128_si256(
+            _mm256_castsi128_si256(
+                _mm_loadu_si128((__m128i *)(data + 7 * src_stride))),
+            src6, 0x20);
+
+        s[3] = _mm256_unpacklo_epi16(s67, s78);
+        s[7] = _mm256_unpackhi_epi16(s67, s78);
+
+        const __m256i res_a = convolve(s, coeffs_y);
+
+        __m256i res_a_round = _mm256_sra_epi32(
+            _mm256_add_epi32(res_a, round_const_bits), round_shift_bits);
+
+        if (w - j > 4) {
+          const __m256i res_b = convolve(s + 4, coeffs_y);
+          __m256i res_b_round = _mm256_sra_epi32(
+              _mm256_add_epi32(res_b, round_const_bits), round_shift_bits);
+
+          __m256i res_16bit = _mm256_packs_epi32(res_a_round, res_b_round);
+          res_16bit = _mm256_min_epi16(res_16bit, clip_pixel);
+          res_16bit = _mm256_max_epi16(res_16bit, zero);
+
+          _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j],
+                           _mm256_castsi256_si128(res_16bit));
+          _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j + dst_stride],
+                           _mm256_extracti128_si256(res_16bit, 1));
+        } else if (w == 4) {
+          res_a_round = _mm256_packs_epi32(res_a_round, res_a_round);
+          res_a_round = _mm256_min_epi16(res_a_round, clip_pixel);
+          res_a_round = _mm256_max_epi16(res_a_round, zero);
+
+          _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j],
+                           _mm256_castsi256_si128(res_a_round));
+          _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j + dst_stride],
+                           _mm256_extracti128_si256(res_a_round, 1));
+        } else {
+          res_a_round = _mm256_packs_epi32(res_a_round, res_a_round);
+          res_a_round = _mm256_min_epi16(res_a_round, clip_pixel);
+          res_a_round = _mm256_max_epi16(res_a_round, zero);
+
+          xx_storel_32(&dst[i * dst_stride + j],
+                       _mm256_castsi256_si128(res_a_round));
+          xx_storel_32(&dst[i * dst_stride + j + dst_stride],
+                       _mm256_extracti128_si256(res_a_round, 1));
+        }
+
+        s[0] = s[1];
+        s[1] = s[2];
+        s[2] = s[3];
+
+        s[4] = s[5];
+        s[5] = s[6];
+        s[6] = s[7];
+      }
+    }
+  }
+}
+
+void av1_highbd_convolve_x_sr_avx2(const uint16_t *src, int src_stride,
+                                   uint16_t *dst, int dst_stride, int w, int h,
+                                   const InterpFilterParams *filter_params_x,
+                                   const int subpel_x_qn,
+                                   ConvolveParams *conv_params, int bd) {
+  if (filter_params_x->taps == 12) {
+    av1_highbd_convolve_x_sr_ssse3(src, src_stride, dst, dst_stride, w, h,
+                                   filter_params_x, subpel_x_qn, conv_params,
+                                   bd);
+    return;
+  }
+  int i, j;
+  const int fo_horiz = filter_params_x->taps / 2 - 1;
+  const uint16_t *const src_ptr = src - fo_horiz;
+
+  // Check that, even with 12-bit input, the intermediate values will fit
+  // into an unsigned 16-bit intermediate array.
+  assert(bd + FILTER_BITS + 2 - conv_params->round_0 <= 16);
+
+  __m256i s[4], coeffs_x[4];
+
+  const __m256i round_const_x =
+      _mm256_set1_epi32(((1 << conv_params->round_0) >> 1));
+  const __m128i round_shift_x = _mm_cvtsi32_si128(conv_params->round_0);
+
+  const int bits = FILTER_BITS - conv_params->round_0;
+  const __m128i round_shift_bits = _mm_cvtsi32_si128(bits);
+  const __m256i round_const_bits = _mm256_set1_epi32((1 << bits) >> 1);
+  const __m256i clip_pixel =
+      _mm256_set1_epi16(bd == 10 ? 1023 : (bd == 12 ? 4095 : 255));
+  const __m256i zero = _mm256_setzero_si256();
+
+  assert(bits >= 0);
+  assert((FILTER_BITS - conv_params->round_1) >= 0 ||
+         ((conv_params->round_0 + conv_params->round_1) == 2 * FILTER_BITS));
+
+  prepare_coeffs(filter_params_x, subpel_x_qn, coeffs_x);
+
+  for (j = 0; j < w; j += 8) {
+    /* Horizontal filter */
+    for (i = 0; i < h; i += 2) {
+      const __m256i row0 =
+          _mm256_loadu_si256((__m256i *)&src_ptr[i * src_stride + j]);
+      __m256i row1 =
+          _mm256_loadu_si256((__m256i *)&src_ptr[(i + 1) * src_stride + j]);
+
+      const __m256i r0 = _mm256_permute2x128_si256(row0, row1, 0x20);
+      const __m256i r1 = _mm256_permute2x128_si256(row0, row1, 0x31);
+
+      // even pixels
+      s[0] = _mm256_alignr_epi8(r1, r0, 0);
+      s[1] = _mm256_alignr_epi8(r1, r0, 4);
+      s[2] = _mm256_alignr_epi8(r1, r0, 8);
+      s[3] = _mm256_alignr_epi8(r1, r0, 12);
+
+      __m256i res_even = convolve(s, coeffs_x);
+      res_even = _mm256_sra_epi32(_mm256_add_epi32(res_even, round_const_x),
+                                  round_shift_x);
+
+      // odd pixels
+      s[0] = _mm256_alignr_epi8(r1, r0, 2);
+      s[1] = _mm256_alignr_epi8(r1, r0, 6);
+      s[2] = _mm256_alignr_epi8(r1, r0, 10);
+      s[3] = _mm256_alignr_epi8(r1, r0, 14);
+
+      __m256i res_odd = convolve(s, coeffs_x);
+      res_odd = _mm256_sra_epi32(_mm256_add_epi32(res_odd, round_const_x),
+                                 round_shift_x);
+
+      res_even = _mm256_sra_epi32(_mm256_add_epi32(res_even, round_const_bits),
+                                  round_shift_bits);
+      res_odd = _mm256_sra_epi32(_mm256_add_epi32(res_odd, round_const_bits),
+                                 round_shift_bits);
+
+      __m256i res_even1 = _mm256_packs_epi32(res_even, res_even);
+      __m256i res_odd1 = _mm256_packs_epi32(res_odd, res_odd);
+
+      __m256i res = _mm256_unpacklo_epi16(res_even1, res_odd1);
+      res = _mm256_min_epi16(res, clip_pixel);
+      res = _mm256_max_epi16(res, zero);
+
+      if (w - j > 4) {
+        _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j],
+                         _mm256_castsi256_si128(res));
+        _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j + dst_stride],
+                         _mm256_extracti128_si256(res, 1));
+      } else if (w == 4) {
+        _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j],
+                         _mm256_castsi256_si128(res));
+        _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j + dst_stride],
+                         _mm256_extracti128_si256(res, 1));
+      } else {
+        xx_storel_32(&dst[i * dst_stride + j], _mm256_castsi256_si128(res));
+        xx_storel_32(&dst[i * dst_stride + j + dst_stride],
+                     _mm256_extracti128_si256(res, 1));
+      }
+    }
+  }
+}
+
+#define CONV8_ROUNDING_BITS (7)
+
+// -----------------------------------------------------------------------------
+// Horizontal and vertical filtering
+
+static const uint8_t signal_pattern_0[32] = { 0, 1, 2, 3, 2, 3, 4, 5, 4, 5, 6,
+                                              7, 6, 7, 8, 9, 0, 1, 2, 3, 2, 3,
+                                              4, 5, 4, 5, 6, 7, 6, 7, 8, 9 };
+
+static const uint8_t signal_pattern_1[32] = { 4, 5, 6,  7,  6,  7,  8,  9,
+                                              8, 9, 10, 11, 10, 11, 12, 13,
+                                              4, 5, 6,  7,  6,  7,  8,  9,
+                                              8, 9, 10, 11, 10, 11, 12, 13 };
+
+static const uint8_t signal_pattern_2[32] = { 6,  7,  8,  9,  8,  9,  10, 11,
+                                              10, 11, 12, 13, 12, 13, 14, 15,
+                                              6,  7,  8,  9,  8,  9,  10, 11,
+                                              10, 11, 12, 13, 12, 13, 14, 15 };
+
+static const uint32_t signal_index[8] = { 2, 3, 4, 5, 2, 3, 4, 5 };
+
+// -----------------------------------------------------------------------------
+// Horizontal Filtering
+
+static INLINE void pack_pixels(const __m256i *s, __m256i *p /*p[4]*/) {
+  const __m256i idx = _mm256_loadu_si256((const __m256i *)signal_index);
+  const __m256i sf0 = _mm256_loadu_si256((const __m256i *)signal_pattern_0);
+  const __m256i sf1 = _mm256_loadu_si256((const __m256i *)signal_pattern_1);
+  const __m256i c = _mm256_permutevar8x32_epi32(*s, idx);
+
+  p[0] = _mm256_shuffle_epi8(*s, sf0);  // x0x6
+  p[1] = _mm256_shuffle_epi8(*s, sf1);  // x1x7
+  p[2] = _mm256_shuffle_epi8(c, sf0);   // x2x4
+  p[3] = _mm256_shuffle_epi8(c, sf1);   // x3x5
+}
+
+// Note:
+//  Shared by 8x2 and 16x1 block
+static INLINE void pack_16_pixels(const __m256i *s0, const __m256i *s1,
+                                  __m256i *x /*x[8]*/) {
+  __m256i pp[8];
+  pack_pixels(s0, pp);
+  pack_pixels(s1, &pp[4]);
+  x[0] = _mm256_permute2x128_si256(pp[0], pp[4], 0x20);
+  x[1] = _mm256_permute2x128_si256(pp[1], pp[5], 0x20);
+  x[2] = _mm256_permute2x128_si256(pp[2], pp[6], 0x20);
+  x[3] = _mm256_permute2x128_si256(pp[3], pp[7], 0x20);
+  x[4] = x[2];
+  x[5] = x[3];
+  x[6] = _mm256_permute2x128_si256(pp[0], pp[4], 0x31);
+  x[7] = _mm256_permute2x128_si256(pp[1], pp[5], 0x31);
+}
+
+static INLINE void pack_8x1_pixels(const uint16_t *src, __m256i *x) {
+  __m256i pp[8];
+  __m256i s0;
+  s0 = _mm256_loadu_si256((const __m256i *)src);
+  pack_pixels(&s0, pp);
+  x[0] = _mm256_permute2x128_si256(pp[0], pp[2], 0x30);
+  x[1] = _mm256_permute2x128_si256(pp[1], pp[3], 0x30);
+  x[2] = _mm256_permute2x128_si256(pp[2], pp[0], 0x30);
+  x[3] = _mm256_permute2x128_si256(pp[3], pp[1], 0x30);
+}
+
+static INLINE void pack_8x2_pixels(const uint16_t *src, ptrdiff_t stride,
+                                   __m256i *x) {
+  __m256i s0, s1;
+  s0 = _mm256_loadu_si256((const __m256i *)src);
+  s1 = _mm256_loadu_si256((const __m256i *)(src + stride));
+  pack_16_pixels(&s0, &s1, x);
+}
+
+static INLINE void pack_16x1_pixels(const uint16_t *src, __m256i *x) {
+  __m256i s0, s1;
+  s0 = _mm256_loadu_si256((const __m256i *)src);
+  s1 = _mm256_loadu_si256((const __m256i *)(src + 8));
+  pack_16_pixels(&s0, &s1, x);
+}
+
+// Note:
+//  Shared by horizontal and vertical filtering
+static INLINE void pack_filters(const int16_t *filter, __m256i *f /*f[4]*/) {
+  const __m128i h = _mm_loadu_si128((const __m128i *)filter);
+  const __m256i hh = _mm256_insertf128_si256(_mm256_castsi128_si256(h), h, 1);
+  const __m256i p0 = _mm256_set1_epi32(0x03020100);
+  const __m256i p1 = _mm256_set1_epi32(0x07060504);
+  const __m256i p2 = _mm256_set1_epi32(0x0b0a0908);
+  const __m256i p3 = _mm256_set1_epi32(0x0f0e0d0c);
+  f[0] = _mm256_shuffle_epi8(hh, p0);
+  f[1] = _mm256_shuffle_epi8(hh, p1);
+  f[2] = _mm256_shuffle_epi8(hh, p2);
+  f[3] = _mm256_shuffle_epi8(hh, p3);
+}
+
+static INLINE void pack_filters_4tap(const int16_t *filter,
+                                     __m256i *f /*f[4]*/) {
+  const __m128i h = _mm_loadu_si128((const __m128i *)filter);
+  const __m256i coeff = _mm256_broadcastsi128_si256(h);
+
+  // coeffs 2 3 2 3 2 3 2 3
+  f[0] = _mm256_shuffle_epi32(coeff, 0x55);
+  // coeffs 4 5 4 5 4 5 4 5
+  f[1] = _mm256_shuffle_epi32(coeff, 0xaa);
+}
+
+static INLINE void filter_8x1_pixels(const __m256i *sig /*sig[4]*/,
+                                     const __m256i *fil /*fil[4]*/,
+                                     __m256i *y) {
+  __m256i a, a0, a1;
+
+  a0 = _mm256_madd_epi16(fil[0], sig[0]);
+  a1 = _mm256_madd_epi16(fil[3], sig[3]);
+  a = _mm256_add_epi32(a0, a1);
+
+  a0 = _mm256_madd_epi16(fil[1], sig[1]);
+  a1 = _mm256_madd_epi16(fil[2], sig[2]);
+
+  {
+    const __m256i min = _mm256_min_epi32(a0, a1);
+    a = _mm256_add_epi32(a, min);
+  }
+  {
+    const __m256i max = _mm256_max_epi32(a0, a1);
+    a = _mm256_add_epi32(a, max);
+  }
+  {
+    const __m256i rounding = _mm256_set1_epi32(1 << (CONV8_ROUNDING_BITS - 1));
+    a = _mm256_add_epi32(a, rounding);
+    *y = _mm256_srai_epi32(a, CONV8_ROUNDING_BITS);
+  }
+}
+
+static INLINE void store_8x1_pixels(const __m256i *y, const __m256i *mask,
+                                    uint16_t *dst) {
+  const __m128i a0 = _mm256_castsi256_si128(*y);
+  const __m128i a1 = _mm256_extractf128_si256(*y, 1);
+  __m128i res = _mm_packus_epi32(a0, a1);
+  res = _mm_min_epi16(res, _mm256_castsi256_si128(*mask));
+  _mm_storeu_si128((__m128i *)dst, res);
+}
+
+static INLINE void store_8x2_pixels(const __m256i *y0, const __m256i *y1,
+                                    const __m256i *mask, uint16_t *dst,
+                                    ptrdiff_t pitch) {
+  __m256i a = _mm256_packus_epi32(*y0, *y1);
+  a = _mm256_min_epi16(a, *mask);
+  _mm_storeu_si128((__m128i *)dst, _mm256_castsi256_si128(a));
+  _mm_storeu_si128((__m128i *)(dst + pitch), _mm256_extractf128_si256(a, 1));
+}
+
+static INLINE void store_16x1_pixels(const __m256i *y0, const __m256i *y1,
+                                     const __m256i *mask, uint16_t *dst) {
+  __m256i a = _mm256_packus_epi32(*y0, *y1);
+  a = _mm256_min_epi16(a, *mask);
+  _mm256_storeu_si256((__m256i *)dst, a);
+}
+
+static void aom_highbd_filter_block1d8_h8_avx2(
+    const uint16_t *src_ptr, ptrdiff_t src_pitch, uint16_t *dst_ptr,
+    ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) {
+  __m256i signal[8], res0, res1;
+  const __m256i max = _mm256_set1_epi16((1 << bd) - 1);
+
+  __m256i ff[4];
+  pack_filters(filter, ff);
+
+  src_ptr -= 3;
+  do {
+    pack_8x2_pixels(src_ptr, src_pitch, signal);
+    filter_8x1_pixels(signal, ff, &res0);
+    filter_8x1_pixels(&signal[4], ff, &res1);
+    store_8x2_pixels(&res0, &res1, &max, dst_ptr, dst_pitch);
+    height -= 2;
+    src_ptr += src_pitch << 1;
+    dst_ptr += dst_pitch << 1;
+  } while (height > 1);
+
+  if (height > 0) {
+    pack_8x1_pixels(src_ptr, signal);
+    filter_8x1_pixels(signal, ff, &res0);
+    store_8x1_pixels(&res0, &max, dst_ptr);
+  }
+}
+
+static void aom_highbd_filter_block1d16_h8_avx2(
+    const uint16_t *src_ptr, ptrdiff_t src_pitch, uint16_t *dst_ptr,
+    ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) {
+  __m256i signal[8], res0, res1;
+  const __m256i max = _mm256_set1_epi16((1 << bd) - 1);
+
+  __m256i ff[4];
+  pack_filters(filter, ff);
+
+  src_ptr -= 3;
+  do {
+    pack_16x1_pixels(src_ptr, signal);
+    filter_8x1_pixels(signal, ff, &res0);
+    filter_8x1_pixels(&signal[4], ff, &res1);
+    store_16x1_pixels(&res0, &res1, &max, dst_ptr);
+    height -= 1;
+    src_ptr += src_pitch;
+    dst_ptr += dst_pitch;
+  } while (height > 0);
+}
+
+static void aom_highbd_filter_block1d4_h4_avx2(
+    const uint16_t *src_ptr, ptrdiff_t src_pitch, uint16_t *dst_ptr,
+    ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) {
+  const __m256i rounding = _mm256_set1_epi32(1 << (CONV8_ROUNDING_BITS - 1));
+  __m256i ff[2], s[2];
+  uint32_t i;
+  const __m256i clip_pixel =
+      _mm256_set1_epi16(bd == 10 ? 1023 : (bd == 12 ? 4095 : 255));
+  const __m256i zero = _mm256_setzero_si256();
+
+  static const uint8_t shuffle_mask[32] = { 0, 1, 2, 3, 2, 3, 4, 5, 4, 5, 6,
+                                            7, 6, 7, 8, 9, 0, 1, 2, 3, 2, 3,
+                                            4, 5, 4, 5, 6, 7, 6, 7, 8, 9 };
+
+  __m256i mask = _mm256_loadu_si256((__m256i *)shuffle_mask);
+  __m256i ip_mask_f2f3 = _mm256_loadu_si256((__m256i *)ip_shuffle_f2f3);
+  __m256i ip_mask_f4f5 = _mm256_loadu_si256((__m256i *)ip_shuffle_f4f5);
+
+  pack_filters_4tap(filter, ff);
+  src_ptr -= 3;
+  for (i = 0; i <= (height - 2); i += 2) {
+    __m256i row0 = _mm256_castsi128_si256(
+        _mm_loadu_si128((__m128i *)&src_ptr[i * src_pitch + 2]));
+    __m256i row1 = _mm256_castsi128_si256(
+        _mm_loadu_si128((__m128i *)&src_ptr[(i + 1) * src_pitch + 2]));
+
+    s[0] = _mm256_inserti128_si256(row0, _mm256_castsi256_si128(row1), 1);
+    s[1] = _mm256_alignr_epi8(s[0], s[0], 4);
+
+    s[0] = _mm256_shuffle_epi8(s[0], mask);
+    s[1] = _mm256_shuffle_epi8(s[1], mask);
+
+    __m256i res = convolve_4tap(s, ff);
+    res =
+        _mm256_srai_epi32(_mm256_add_epi32(res, rounding), CONV8_ROUNDING_BITS);
+
+    res = _mm256_packs_epi32(res, res);
+    res = _mm256_min_epi16(res, clip_pixel);
+    res = _mm256_max_epi16(res, zero);
+
+    _mm_storel_epi64((__m128i *)&dst_ptr[i * dst_pitch],
+                     _mm256_castsi256_si128(res));
+    _mm_storel_epi64((__m128i *)&dst_ptr[(i + 1) * dst_pitch],
+                     _mm256_extracti128_si256(res, 1));
+  }
+  if (height % 2 != 0) {
+    i = height - 1;
+    const __m256i row0_0 = _mm256_castsi128_si256(
+        _mm_loadu_si128((__m128i *)&src_ptr[i * src_pitch + 2]));
+    const __m256i row0_1 = _mm256_castsi128_si256(
+        _mm_loadu_si128((__m128i *)&src_ptr[i * src_pitch + 6]));
+
+    const __m256i r0 =
+        _mm256_inserti128_si256(row0_0, _mm256_castsi256_si128(row0_1), 1);
+
+    s[0] = _mm256_shuffle_epi8(r0, ip_mask_f2f3);
+    s[1] = _mm256_shuffle_epi8(r0, ip_mask_f4f5);
+
+    __m256i res = convolve_4tap(s, ff);
+    res =
+        _mm256_srai_epi32(_mm256_add_epi32(res, rounding), CONV8_ROUNDING_BITS);
+
+    res = _mm256_packs_epi32(res, res);
+    res = _mm256_min_epi16(res, clip_pixel);
+    res = _mm256_max_epi16(res, zero);
+
+    _mm_storel_epi64((__m128i *)&dst_ptr[i * dst_pitch],
+                     _mm256_castsi256_si128(res));
+  }
+}
+
+static void aom_highbd_filter_block1d8_h4_avx2(
+    const uint16_t *src_ptr, ptrdiff_t src_pitch, uint16_t *dst_ptr,
+    ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) {
+  const __m256i rounding = _mm256_set1_epi32(1 << (CONV8_ROUNDING_BITS - 1));
+  __m256i ff[2], s[2];
+  uint32_t i = 0;
+  const __m256i clip_pixel =
+      _mm256_set1_epi16(bd == 10 ? 1023 : (bd == 12 ? 4095 : 255));
+  const __m256i zero = _mm256_setzero_si256();
+
+  static const uint8_t shuffle_mask[32] = { 0, 1, 8,  9,  2, 3, 10, 11,
+                                            4, 5, 12, 13, 6, 7, 14, 15,
+                                            0, 1, 8,  9,  2, 3, 10, 11,
+                                            4, 5, 12, 13, 6, 7, 14, 15 };
+
+  __m256i mask = _mm256_loadu_si256((__m256i *)shuffle_mask);
+  __m256i ip_mask_f2f3 = _mm256_loadu_si256((__m256i *)ip_shuffle_f2f3);
+  __m256i ip_mask_f4f5 = _mm256_loadu_si256((__m256i *)ip_shuffle_f4f5);
+
+  pack_filters_4tap(filter, ff);
+  src_ptr -= 3;
+
+  /* Horizontal filter */
+
+  for (i = 0; i <= (height - 2); i += 2) {
+    const __m256i row0 =
+        _mm256_loadu_si256((__m256i *)&src_ptr[i * src_pitch + 2]);
+    __m256i row1 =
+        _mm256_loadu_si256((__m256i *)&src_ptr[(i + 1) * src_pitch + 2]);
+
+    const __m256i r0 =
+        _mm256_inserti128_si256(row0, _mm256_castsi256_si128(row1), 1);
+    const __m256i r1 = _mm256_permute2x128_si256(row0, row1, 0x31);
+
+    // even pixels
+    s[0] = r0;
+    s[1] = _mm256_alignr_epi8(r1, r0, 4);
+
+    __m256i res_even = convolve_4tap(s, ff);
+    res_even = _mm256_srai_epi32(_mm256_add_epi32(res_even, rounding),
+                                 CONV8_ROUNDING_BITS);
+
+    // odd pixels
+    s[0] = _mm256_alignr_epi8(r1, r0, 2);
+    s[1] = _mm256_alignr_epi8(r1, r0, 6);
+
+    __m256i res_odd = convolve_4tap(s, ff);
+    res_odd = _mm256_srai_epi32(_mm256_add_epi32(res_odd, rounding),
+                                CONV8_ROUNDING_BITS);
+
+    __m256i res = _mm256_packs_epi32(res_even, res_odd);
+    res = _mm256_shuffle_epi8(res, mask);
+
+    res = _mm256_min_epi16(res, clip_pixel);
+    res = _mm256_max_epi16(res, zero);
+
+    _mm_storeu_si128((__m128i *)&dst_ptr[i * dst_pitch],
+                     _mm256_castsi256_si128(res));
+    _mm_storeu_si128((__m128i *)&dst_ptr[i * dst_pitch + dst_pitch],
+                     _mm256_extracti128_si256(res, 1));
+  }
+
+  if (height % 2 != 0) {
+    i = height - 1;
+    const __m256i row0_0 =
+        _mm256_loadu_si256((__m256i *)&src_ptr[i * src_pitch + 2]);
+    const __m256i row0_1 =
+        _mm256_loadu_si256((__m256i *)&src_ptr[i * src_pitch + 6]);
+
+    const __m256i r0 =
+        _mm256_inserti128_si256(row0_0, _mm256_castsi256_si128(row0_1), 1);
+
+    s[0] = _mm256_shuffle_epi8(r0, ip_mask_f2f3);
+    s[1] = _mm256_shuffle_epi8(r0, ip_mask_f4f5);
+
+    __m256i res = convolve_4tap(s, ff);
+    res =
+        _mm256_srai_epi32(_mm256_add_epi32(res, rounding), CONV8_ROUNDING_BITS);
+
+    res = _mm256_packs_epi32(res, res);
+    res = _mm256_min_epi16(res, clip_pixel);
+    res = _mm256_max_epi16(res, zero);
+
+    _mm_storel_epi64((__m128i *)&dst_ptr[i * dst_pitch],
+                     _mm256_castsi256_si128(res));
+    _mm_storel_epi64((__m128i *)&dst_ptr[i * dst_pitch + 4],
+                     _mm256_extracti128_si256(res, 1));
+  }
+}
+
+static void aom_highbd_filter_block1d16_h4_avx2(
+    const uint16_t *src_ptr, ptrdiff_t src_pitch, uint16_t *dst_ptr,
+    ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) {
+  aom_highbd_filter_block1d8_h4_avx2(src_ptr, src_pitch, dst_ptr, dst_pitch,
+                                     height, filter, bd);
+  aom_highbd_filter_block1d8_h4_avx2(src_ptr + 8, src_pitch, dst_ptr + 8,
+                                     dst_pitch, height, filter, bd);
+}
+
+// -----------------------------------------------------------------------------
+// 2-tap horizontal filtering
+
+static INLINE void pack_2t_filter(const int16_t *filter, __m256i *f) {
+  const __m128i h = _mm_loadu_si128((const __m128i *)filter);
+  const __m256i hh = _mm256_insertf128_si256(_mm256_castsi128_si256(h), h, 1);
+  const __m256i p = _mm256_set1_epi32(0x09080706);
+  f[0] = _mm256_shuffle_epi8(hh, p);
+}
+
+// can be used by pack_8x2_2t_pixels() and pack_16x1_2t_pixels()
+// the difference is s0/s1 specifies first and second rows or,
+// first 16 samples and 8-sample shifted 16 samples
+static INLINE void pack_16_2t_pixels(const __m256i *s0, const __m256i *s1,
+                                     __m256i *sig) {
+  const __m256i idx = _mm256_loadu_si256((const __m256i *)signal_index);
+  const __m256i sf2 = _mm256_loadu_si256((const __m256i *)signal_pattern_2);
+  __m256i x0 = _mm256_shuffle_epi8(*s0, sf2);
+  __m256i x1 = _mm256_shuffle_epi8(*s1, sf2);
+  __m256i r0 = _mm256_permutevar8x32_epi32(*s0, idx);
+  __m256i r1 = _mm256_permutevar8x32_epi32(*s1, idx);
+  r0 = _mm256_shuffle_epi8(r0, sf2);
+  r1 = _mm256_shuffle_epi8(r1, sf2);
+  sig[0] = _mm256_permute2x128_si256(x0, x1, 0x20);
+  sig[1] = _mm256_permute2x128_si256(r0, r1, 0x20);
+}
+
+static INLINE void pack_8x2_2t_pixels(const uint16_t *src,
+                                      const ptrdiff_t pitch, __m256i *sig) {
+  const __m256i r0 = _mm256_loadu_si256((const __m256i *)src);
+  const __m256i r1 = _mm256_loadu_si256((const __m256i *)(src + pitch));
+  pack_16_2t_pixels(&r0, &r1, sig);
+}
+
+static INLINE void pack_16x1_2t_pixels(const uint16_t *src,
+                                       __m256i *sig /*sig[2]*/) {
+  const __m256i r0 = _mm256_loadu_si256((const __m256i *)src);
+  const __m256i r1 = _mm256_loadu_si256((const __m256i *)(src + 8));
+  pack_16_2t_pixels(&r0, &r1, sig);
+}
+
+static INLINE void pack_8x1_2t_pixels(const uint16_t *src,
+                                      __m256i *sig /*sig[2]*/) {
+  const __m256i idx = _mm256_loadu_si256((const __m256i *)signal_index);
+  const __m256i sf2 = _mm256_loadu_si256((const __m256i *)signal_pattern_2);
+  __m256i r0 = _mm256_loadu_si256((const __m256i *)src);
+  __m256i x0 = _mm256_shuffle_epi8(r0, sf2);
+  r0 = _mm256_permutevar8x32_epi32(r0, idx);
+  r0 = _mm256_shuffle_epi8(r0, sf2);
+  sig[0] = _mm256_permute2x128_si256(x0, r0, 0x20);
+}
+
+// can be used by filter_8x2_2t_pixels() and filter_16x1_2t_pixels()
+static INLINE void filter_16_2t_pixels(const __m256i *sig, const __m256i *f,
+                                       __m256i *y0, __m256i *y1) {
+  const __m256i rounding = _mm256_set1_epi32(1 << (CONV8_ROUNDING_BITS - 1));
+  __m256i x0 = _mm256_madd_epi16(sig[0], *f);
+  __m256i x1 = _mm256_madd_epi16(sig[1], *f);
+  x0 = _mm256_add_epi32(x0, rounding);
+  x1 = _mm256_add_epi32(x1, rounding);
+  *y0 = _mm256_srai_epi32(x0, CONV8_ROUNDING_BITS);
+  *y1 = _mm256_srai_epi32(x1, CONV8_ROUNDING_BITS);
+}
+
+static INLINE void filter_8x1_2t_pixels(const __m256i *sig, const __m256i *f,
+                                        __m256i *y0) {
+  const __m256i rounding = _mm256_set1_epi32(1 << (CONV8_ROUNDING_BITS - 1));
+  __m256i x0 = _mm256_madd_epi16(sig[0], *f);
+  x0 = _mm256_add_epi32(x0, rounding);
+  *y0 = _mm256_srai_epi32(x0, CONV8_ROUNDING_BITS);
+}
+
+static void aom_highbd_filter_block1d8_h2_avx2(
+    const uint16_t *src_ptr, ptrdiff_t src_pitch, uint16_t *dst_ptr,
+    ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) {
+  __m256i signal[2], res0, res1;
+  const __m256i max = _mm256_set1_epi16((1 << bd) - 1);
+
+  __m256i ff;
+  pack_2t_filter(filter, &ff);
+
+  src_ptr -= 3;
+  do {
+    pack_8x2_2t_pixels(src_ptr, src_pitch, signal);
+    filter_16_2t_pixels(signal, &ff, &res0, &res1);
+    store_8x2_pixels(&res0, &res1, &max, dst_ptr, dst_pitch);
+    height -= 2;
+    src_ptr += src_pitch << 1;
+    dst_ptr += dst_pitch << 1;
+  } while (height > 1);
+
+  if (height > 0) {
+    pack_8x1_2t_pixels(src_ptr, signal);
+    filter_8x1_2t_pixels(signal, &ff, &res0);
+    store_8x1_pixels(&res0, &max, dst_ptr);
+  }
+}
+
+static void aom_highbd_filter_block1d16_h2_avx2(
+    const uint16_t *src_ptr, ptrdiff_t src_pitch, uint16_t *dst_ptr,
+    ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) {
+  __m256i signal[2], res0, res1;
+  const __m256i max = _mm256_set1_epi16((1 << bd) - 1);
+
+  __m256i ff;
+  pack_2t_filter(filter, &ff);
+
+  src_ptr -= 3;
+  do {
+    pack_16x1_2t_pixels(src_ptr, signal);
+    filter_16_2t_pixels(signal, &ff, &res0, &res1);
+    store_16x1_pixels(&res0, &res1, &max, dst_ptr);
+    height -= 1;
+    src_ptr += src_pitch;
+    dst_ptr += dst_pitch;
+  } while (height > 0);
+}
+
+// -----------------------------------------------------------------------------
+// Vertical Filtering
+
+static void pack_8x9_init(const uint16_t *src, ptrdiff_t pitch, __m256i *sig) {
+  __m256i s0 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)src));
+  __m256i s1 =
+      _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(src + pitch)));
+  __m256i s2 = _mm256_castsi128_si256(
+      _mm_loadu_si128((const __m128i *)(src + 2 * pitch)));
+  __m256i s3 = _mm256_castsi128_si256(
+      _mm_loadu_si128((const __m128i *)(src + 3 * pitch)));
+  __m256i s4 = _mm256_castsi128_si256(
+      _mm_loadu_si128((const __m128i *)(src + 4 * pitch)));
+  __m256i s5 = _mm256_castsi128_si256(
+      _mm_loadu_si128((const __m128i *)(src + 5 * pitch)));
+  __m256i s6 = _mm256_castsi128_si256(
+      _mm_loadu_si128((const __m128i *)(src + 6 * pitch)));
+
+  s0 = _mm256_inserti128_si256(s0, _mm256_castsi256_si128(s1), 1);
+  s1 = _mm256_inserti128_si256(s1, _mm256_castsi256_si128(s2), 1);
+  s2 = _mm256_inserti128_si256(s2, _mm256_castsi256_si128(s3), 1);
+  s3 = _mm256_inserti128_si256(s3, _mm256_castsi256_si128(s4), 1);
+  s4 = _mm256_inserti128_si256(s4, _mm256_castsi256_si128(s5), 1);
+  s5 = _mm256_inserti128_si256(s5, _mm256_castsi256_si128(s6), 1);
+
+  sig[0] = _mm256_unpacklo_epi16(s0, s1);
+  sig[4] = _mm256_unpackhi_epi16(s0, s1);
+  sig[1] = _mm256_unpacklo_epi16(s2, s3);
+  sig[5] = _mm256_unpackhi_epi16(s2, s3);
+  sig[2] = _mm256_unpacklo_epi16(s4, s5);
+  sig[6] = _mm256_unpackhi_epi16(s4, s5);
+  sig[8] = s6;
+}
+
+static INLINE void pack_8x9_pixels(const uint16_t *src, ptrdiff_t pitch,
+                                   __m256i *sig) {
+  // base + 7th row
+  __m256i s0 = _mm256_castsi128_si256(
+      _mm_loadu_si128((const __m128i *)(src + 7 * pitch)));
+  // base + 8th row
+  __m256i s1 = _mm256_castsi128_si256(
+      _mm_loadu_si128((const __m128i *)(src + 8 * pitch)));
+  __m256i s2 = _mm256_inserti128_si256(sig[8], _mm256_castsi256_si128(s0), 1);
+  __m256i s3 = _mm256_inserti128_si256(s0, _mm256_castsi256_si128(s1), 1);
+  sig[3] = _mm256_unpacklo_epi16(s2, s3);
+  sig[7] = _mm256_unpackhi_epi16(s2, s3);
+  sig[8] = s1;
+}
+
+static INLINE void filter_8x9_pixels(const __m256i *sig, const __m256i *f,
+                                     __m256i *y0, __m256i *y1) {
+  filter_8x1_pixels(sig, f, y0);
+  filter_8x1_pixels(&sig[4], f, y1);
+}
+
+static INLINE void update_pixels(__m256i *sig) {
+  int i;
+  for (i = 0; i < 3; ++i) {
+    sig[i] = sig[i + 1];
+    sig[i + 4] = sig[i + 5];
+  }
+}
+
+static void aom_highbd_filter_block1d8_v8_avx2(
+    const uint16_t *src_ptr, ptrdiff_t src_pitch, uint16_t *dst_ptr,
+    ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) {
+  __m256i signal[9], res0, res1;
+  const __m256i max = _mm256_set1_epi16((1 << bd) - 1);
+
+  __m256i ff[4];
+  pack_filters(filter, ff);
+
+  pack_8x9_init(src_ptr, src_pitch, signal);
+
+  do {
+    pack_8x9_pixels(src_ptr, src_pitch, signal);
+
+    filter_8x9_pixels(signal, ff, &res0, &res1);
+    store_8x2_pixels(&res0, &res1, &max, dst_ptr, dst_pitch);
+    update_pixels(signal);
+
+    src_ptr += src_pitch << 1;
+    dst_ptr += dst_pitch << 1;
+    height -= 2;
+  } while (height > 0);
+}
+
+static void pack_16x9_init(const uint16_t *src, ptrdiff_t pitch, __m256i *sig) {
+  __m256i u0, u1, u2, u3;
+  // load 0-6 rows
+  const __m256i s0 = _mm256_loadu_si256((const __m256i *)src);
+  const __m256i s1 = _mm256_loadu_si256((const __m256i *)(src + pitch));
+  const __m256i s2 = _mm256_loadu_si256((const __m256i *)(src + 2 * pitch));
+  const __m256i s3 = _mm256_loadu_si256((const __m256i *)(src + 3 * pitch));
+  const __m256i s4 = _mm256_loadu_si256((const __m256i *)(src + 4 * pitch));
+  const __m256i s5 = _mm256_loadu_si256((const __m256i *)(src + 5 * pitch));
+  const __m256i s6 = _mm256_loadu_si256((const __m256i *)(src + 6 * pitch));
+
+  u0 = _mm256_permute2x128_si256(s0, s1, 0x20);  // 0, 1 low
+  u1 = _mm256_permute2x128_si256(s0, s1, 0x31);  // 0, 1 high
+
+  u2 = _mm256_permute2x128_si256(s1, s2, 0x20);  // 1, 2 low
+  u3 = _mm256_permute2x128_si256(s1, s2, 0x31);  // 1, 2 high
+
+  sig[0] = _mm256_unpacklo_epi16(u0, u2);
+  sig[4] = _mm256_unpackhi_epi16(u0, u2);
+
+  sig[8] = _mm256_unpacklo_epi16(u1, u3);
+  sig[12] = _mm256_unpackhi_epi16(u1, u3);
+
+  u0 = _mm256_permute2x128_si256(s2, s3, 0x20);
+  u1 = _mm256_permute2x128_si256(s2, s3, 0x31);
+
+  u2 = _mm256_permute2x128_si256(s3, s4, 0x20);
+  u3 = _mm256_permute2x128_si256(s3, s4, 0x31);
+
+  sig[1] = _mm256_unpacklo_epi16(u0, u2);
+  sig[5] = _mm256_unpackhi_epi16(u0, u2);
+
+  sig[9] = _mm256_unpacklo_epi16(u1, u3);
+  sig[13] = _mm256_unpackhi_epi16(u1, u3);
+
+  u0 = _mm256_permute2x128_si256(s4, s5, 0x20);
+  u1 = _mm256_permute2x128_si256(s4, s5, 0x31);
+
+  u2 = _mm256_permute2x128_si256(s5, s6, 0x20);
+  u3 = _mm256_permute2x128_si256(s5, s6, 0x31);
+
+  sig[2] = _mm256_unpacklo_epi16(u0, u2);
+  sig[6] = _mm256_unpackhi_epi16(u0, u2);
+
+  sig[10] = _mm256_unpacklo_epi16(u1, u3);
+  sig[14] = _mm256_unpackhi_epi16(u1, u3);
+
+  sig[16] = s6;
+}
+
+static void pack_16x9_pixels(const uint16_t *src, ptrdiff_t pitch,
+                             __m256i *sig) {
+  // base + 7th row
+  const __m256i s7 = _mm256_loadu_si256((const __m256i *)(src + 7 * pitch));
+  // base + 8th row
+  const __m256i s8 = _mm256_loadu_si256((const __m256i *)(src + 8 * pitch));
+
+  __m256i u0, u1, u2, u3;
+  u0 = _mm256_permute2x128_si256(sig[16], s7, 0x20);
+  u1 = _mm256_permute2x128_si256(sig[16], s7, 0x31);
+
+  u2 = _mm256_permute2x128_si256(s7, s8, 0x20);
+  u3 = _mm256_permute2x128_si256(s7, s8, 0x31);
+
+  sig[3] = _mm256_unpacklo_epi16(u0, u2);
+  sig[7] = _mm256_unpackhi_epi16(u0, u2);
+
+  sig[11] = _mm256_unpacklo_epi16(u1, u3);
+  sig[15] = _mm256_unpackhi_epi16(u1, u3);
+
+  sig[16] = s8;
+}
+
+static INLINE void filter_16x9_pixels(const __m256i *sig, const __m256i *f,
+                                      __m256i *y0, __m256i *y1) {
+  __m256i res[4];
+  int i;
+  for (i = 0; i < 4; ++i) {
+    filter_8x1_pixels(&sig[i << 2], f, &res[i]);
+  }
+
+  {
+    const __m256i l0l1 = _mm256_packus_epi32(res[0], res[1]);
+    const __m256i h0h1 = _mm256_packus_epi32(res[2], res[3]);
+    *y0 = _mm256_permute2x128_si256(l0l1, h0h1, 0x20);
+    *y1 = _mm256_permute2x128_si256(l0l1, h0h1, 0x31);
+  }
+}
+
+static INLINE void store_16x2_pixels(const __m256i *y0, const __m256i *y1,
+                                     const __m256i *mask, uint16_t *dst,
+                                     ptrdiff_t pitch) {
+  __m256i p = _mm256_min_epi16(*y0, *mask);
+  _mm256_storeu_si256((__m256i *)dst, p);
+  p = _mm256_min_epi16(*y1, *mask);
+  _mm256_storeu_si256((__m256i *)(dst + pitch), p);
+}
+
+static void update_16x9_pixels(__m256i *sig) {
+  update_pixels(&sig[0]);
+  update_pixels(&sig[8]);
+}
+
+static void aom_highbd_filter_block1d16_v8_avx2(
+    const uint16_t *src_ptr, ptrdiff_t src_pitch, uint16_t *dst_ptr,
+    ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) {
+  __m256i signal[17], res0, res1;
+  const __m256i max = _mm256_set1_epi16((1 << bd) - 1);
+
+  __m256i ff[4];
+  pack_filters(filter, ff);
+
+  pack_16x9_init(src_ptr, src_pitch, signal);
+
+  do {
+    pack_16x9_pixels(src_ptr, src_pitch, signal);
+    filter_16x9_pixels(signal, ff, &res0, &res1);
+    store_16x2_pixels(&res0, &res1, &max, dst_ptr, dst_pitch);
+    update_16x9_pixels(signal);
+
+    src_ptr += src_pitch << 1;
+    dst_ptr += dst_pitch << 1;
+    height -= 2;
+  } while (height > 0);
+}
+
+static void aom_highbd_filter_block1d4_v4_avx2(
+    const uint16_t *src_ptr, ptrdiff_t src_pitch, uint16_t *dst_ptr,
+    ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) {
+  const int bits = FILTER_BITS;
+
+  const __m128i round_shift_bits = _mm_cvtsi32_si128(bits);
+  const __m256i round_const_bits = _mm256_set1_epi32((1 << bits) >> 1);
+  const __m256i clip_pixel =
+      _mm256_set1_epi32(bd == 10 ? 1023 : (bd == 12 ? 4095 : 255));
+  const __m256i zero = _mm256_setzero_si256();
+  uint32_t i;
+  __m256i s[2], ff[2];
+
+  pack_filters_4tap(filter, ff);
+
+  const uint16_t *data = src_ptr;
+  /* Vertical filter */
+  {
+    __m128i s2 = _mm_loadl_epi64((__m128i *)(data + 2 * src_pitch));
+    __m128i s3 = _mm_loadl_epi64((__m128i *)(data + 3 * src_pitch));
+
+    __m256i s23 = _mm256_inserti128_si256(_mm256_castsi128_si256(s2), s3, 1);
+
+    __m128i s4 = _mm_loadl_epi64((__m128i *)(data + 4 * src_pitch));
+
+    __m256i s34 = _mm256_inserti128_si256(_mm256_castsi128_si256(s3), s4, 1);
+
+    s[0] = _mm256_unpacklo_epi16(s23, s34);
+
+    for (i = 0; i < height; i += 2) {
+      data = &src_ptr[i * src_pitch];
+
+      __m128i s5 = _mm_loadl_epi64((__m128i *)(data + 5 * src_pitch));
+      __m128i s6 = _mm_loadl_epi64((__m128i *)(data + 6 * src_pitch));
+
+      __m256i s45 = _mm256_inserti128_si256(_mm256_castsi128_si256(s4), s5, 1);
+      __m256i s56 = _mm256_inserti128_si256(_mm256_castsi128_si256(s5), s6, 1);
+
+      s[1] = _mm256_unpacklo_epi16(s45, s56);
+
+      const __m256i res_a = convolve_4tap(s, ff);
+
+      __m256i res_a_round = _mm256_sra_epi32(
+          _mm256_add_epi32(res_a, round_const_bits), round_shift_bits);
+
+      __m256i res_16bit = _mm256_min_epi32(res_a_round, clip_pixel);
+      res_16bit = _mm256_max_epi32(res_16bit, zero);
+      res_16bit = _mm256_packs_epi32(res_16bit, res_16bit);
+
+      _mm_storel_epi64((__m128i *)&dst_ptr[i * dst_pitch],
+                       _mm256_castsi256_si128(res_16bit));
+      _mm_storel_epi64((__m128i *)&dst_ptr[i * dst_pitch + dst_pitch],
+                       _mm256_extracti128_si256(res_16bit, 1));
+
+      s[0] = s[1];
+      s4 = s6;
+    }
+  }
+}
+
+static void aom_highbd_filter_block1d8_v4_avx2(
+    const uint16_t *src_ptr, ptrdiff_t src_pitch, uint16_t *dst_ptr,
+    ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) {
+  const int bits = FILTER_BITS;
+
+  const __m128i round_shift_bits = _mm_cvtsi32_si128(bits);
+  const __m256i round_const_bits = _mm256_set1_epi32((1 << bits) >> 1);
+  const __m256i clip_pixel =
+      _mm256_set1_epi16(bd == 10 ? 1023 : (bd == 12 ? 4095 : 255));
+  const __m256i zero = _mm256_setzero_si256();
+  __m256i s[4], ff[2];
+  uint32_t i;
+  pack_filters_4tap(filter, ff);
+
+  const uint16_t *data = src_ptr;
+  /* Vertical filter */
+  {
+    __m128i s2 = _mm_loadu_si128((__m128i *)(data + 2 * src_pitch));
+    __m128i s3 = _mm_loadu_si128((__m128i *)(data + 3 * src_pitch));
+
+    __m256i s23 = _mm256_inserti128_si256(_mm256_castsi128_si256(s2), s3, 1);
+
+    __m128i s4 = _mm_loadu_si128((__m128i *)(data + 4 * src_pitch));
+
+    __m256i s34 = _mm256_inserti128_si256(_mm256_castsi128_si256(s3), s4, 1);
+
+    s[0] = _mm256_unpacklo_epi16(s23, s34);
+    s[2] = _mm256_unpackhi_epi16(s23, s34);
+
+    for (i = 0; i < height; i += 2) {
+      data = &src_ptr[i * src_pitch];
+
+      __m128i s5 = _mm_loadu_si128((__m128i *)(data + 5 * src_pitch));
+      __m128i s6 = _mm_loadu_si128((__m128i *)(data + 6 * src_pitch));
+
+      __m256i s45 = _mm256_inserti128_si256(_mm256_castsi128_si256(s4), s5, 1);
+      __m256i s56 = _mm256_inserti128_si256(_mm256_castsi128_si256(s5), s6, 1);
+
+      s[1] = _mm256_unpacklo_epi16(s45, s56);
+      s[3] = _mm256_unpackhi_epi16(s45, s56);
+
+      const __m256i res_a = convolve_4tap(s, ff);
+
+      __m256i res_a_round = _mm256_sra_epi32(
+          _mm256_add_epi32(res_a, round_const_bits), round_shift_bits);
+
+      const __m256i res_b = convolve_4tap(s + 2, ff);
+      __m256i res_b_round = _mm256_sra_epi32(
+          _mm256_add_epi32(res_b, round_const_bits), round_shift_bits);
+
+      __m256i res_16bit = _mm256_packs_epi32(res_a_round, res_b_round);
+      res_16bit = _mm256_min_epi16(res_16bit, clip_pixel);
+      res_16bit = _mm256_max_epi16(res_16bit, zero);
+
+      _mm_storeu_si128((__m128i *)&dst_ptr[i * dst_pitch],
+                       _mm256_castsi256_si128(res_16bit));
+      _mm_storeu_si128((__m128i *)&dst_ptr[i * dst_pitch + dst_pitch],
+                       _mm256_extracti128_si256(res_16bit, 1));
+
+      s[0] = s[1];
+      s[2] = s[3];
+      s4 = s6;
+    }
+  }
+}
+
+static void aom_highbd_filter_block1d16_v4_avx2(
+    const uint16_t *src_ptr, ptrdiff_t src_pitch, uint16_t *dst_ptr,
+    ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) {
+  aom_highbd_filter_block1d8_v4_avx2(src_ptr, src_pitch, dst_ptr, dst_pitch,
+                                     height, filter, bd);
+
+  aom_highbd_filter_block1d8_v4_avx2(src_ptr + 8, src_pitch, dst_ptr + 8,
+                                     dst_pitch, height, filter, bd);
+}
+
+// -----------------------------------------------------------------------------
+// 2-tap vertical filtering
+
+static void pack_16x2_init(const uint16_t *src, __m256i *sig) {
+  sig[2] = _mm256_loadu_si256((const __m256i *)src);
+}
+
+static INLINE void pack_16x2_2t_pixels(const uint16_t *src, ptrdiff_t pitch,
+                                       __m256i *sig) {
+  // load the next row
+  const __m256i u = _mm256_loadu_si256((const __m256i *)(src + pitch));
+  sig[0] = _mm256_unpacklo_epi16(sig[2], u);
+  sig[1] = _mm256_unpackhi_epi16(sig[2], u);
+  sig[2] = u;
+}
+
+static INLINE void filter_16x2_2t_pixels(const __m256i *sig, const __m256i *f,
+                                         __m256i *y0, __m256i *y1) {
+  filter_16_2t_pixels(sig, f, y0, y1);
+}
+
+static void aom_highbd_filter_block1d16_v2_avx2(
+    const uint16_t *src_ptr, ptrdiff_t src_pitch, uint16_t *dst_ptr,
+    ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) {
+  __m256i signal[3], res0, res1;
+  const __m256i max = _mm256_set1_epi16((1 << bd) - 1);
+  __m256i ff;
+
+  pack_2t_filter(filter, &ff);
+  pack_16x2_init(src_ptr, signal);
+
+  do {
+    pack_16x2_2t_pixels(src_ptr, src_pitch, signal);
+    filter_16x2_2t_pixels(signal, &ff, &res0, &res1);
+    store_16x1_pixels(&res0, &res1, &max, dst_ptr);
+
+    src_ptr += src_pitch;
+    dst_ptr += dst_pitch;
+    height -= 1;
+  } while (height > 0);
+}
+
+static INLINE void pack_8x1_2t_filter(const int16_t *filter, __m128i *f) {
+  const __m128i h = _mm_loadu_si128((const __m128i *)filter);
+  const __m128i p = _mm_set1_epi32(0x09080706);
+  f[0] = _mm_shuffle_epi8(h, p);
+}
+
+static void pack_8x2_init(const uint16_t *src, __m128i *sig) {
+  sig[2] = _mm_loadu_si128((const __m128i *)src);
+}
+
+static INLINE void pack_8x2_2t_pixels_ver(const uint16_t *src, ptrdiff_t pitch,
+                                          __m128i *sig) {
+  // load the next row
+  const __m128i u = _mm_loadu_si128((const __m128i *)(src + pitch));
+  sig[0] = _mm_unpacklo_epi16(sig[2], u);
+  sig[1] = _mm_unpackhi_epi16(sig[2], u);
+  sig[2] = u;
+}
+
+static INLINE void filter_8_2t_pixels(const __m128i *sig, const __m128i *f,
+                                      __m128i *y0, __m128i *y1) {
+  const __m128i rounding = _mm_set1_epi32(1 << (CONV8_ROUNDING_BITS - 1));
+  __m128i x0 = _mm_madd_epi16(sig[0], *f);
+  __m128i x1 = _mm_madd_epi16(sig[1], *f);
+  x0 = _mm_add_epi32(x0, rounding);
+  x1 = _mm_add_epi32(x1, rounding);
+  *y0 = _mm_srai_epi32(x0, CONV8_ROUNDING_BITS);
+  *y1 = _mm_srai_epi32(x1, CONV8_ROUNDING_BITS);
+}
+
+static INLINE void store_8x1_2t_pixels_ver(const __m128i *y0, const __m128i *y1,
+                                           const __m128i *mask, uint16_t *dst) {
+  __m128i res = _mm_packus_epi32(*y0, *y1);
+  res = _mm_min_epi16(res, *mask);
+  _mm_storeu_si128((__m128i *)dst, res);
+}
+
+static void aom_highbd_filter_block1d8_v2_avx2(
+    const uint16_t *src_ptr, ptrdiff_t src_pitch, uint16_t *dst_ptr,
+    ptrdiff_t dst_pitch, uint32_t height, const int16_t *filter, int bd) {
+  __m128i signal[3], res0, res1;
+  const __m128i max = _mm_set1_epi16((1 << bd) - 1);
+  __m128i ff;
+
+  pack_8x1_2t_filter(filter, &ff);
+  pack_8x2_init(src_ptr, signal);
+
+  do {
+    pack_8x2_2t_pixels_ver(src_ptr, src_pitch, signal);
+    filter_8_2t_pixels(signal, &ff, &res0, &res1);
+    store_8x1_2t_pixels_ver(&res0, &res1, &max, dst_ptr);
+
+    src_ptr += src_pitch;
+    dst_ptr += dst_pitch;
+    height -= 1;
+  } while (height > 0);
+}
+
+void aom_highbd_filter_block1d4_h8_sse2(const uint16_t *, ptrdiff_t, uint16_t *,
+                                        ptrdiff_t, uint32_t, const int16_t *,
+                                        int);
+void aom_highbd_filter_block1d4_h2_sse2(const uint16_t *, ptrdiff_t, uint16_t *,
+                                        ptrdiff_t, uint32_t, const int16_t *,
+                                        int);
+void aom_highbd_filter_block1d4_v8_sse2(const uint16_t *, ptrdiff_t, uint16_t *,
+                                        ptrdiff_t, uint32_t, const int16_t *,
+                                        int);
+void aom_highbd_filter_block1d4_v2_sse2(const uint16_t *, ptrdiff_t, uint16_t *,
+                                        ptrdiff_t, uint32_t, const int16_t *,
+                                        int);
+#define aom_highbd_filter_block1d4_h8_avx2 aom_highbd_filter_block1d4_h8_sse2
+#define aom_highbd_filter_block1d4_h2_avx2 aom_highbd_filter_block1d4_h2_sse2
+#define aom_highbd_filter_block1d4_v8_avx2 aom_highbd_filter_block1d4_v8_sse2
+#define aom_highbd_filter_block1d4_v2_avx2 aom_highbd_filter_block1d4_v2_sse2
+
+HIGH_FUN_CONV_1D(horiz, x_step_q4, filter_x, h, src, , avx2)
+HIGH_FUN_CONV_1D(vert, y_step_q4, filter_y, v, src - src_stride * 3, , avx2)
+
+#undef HIGHBD_FUNC
diff --git a/third_party/aom/aom_dsp/x86/highbd_convolve_sse2.c b/third_party/aom/aom_dsp/x86/highbd_convolve_sse2.c
new file mode 100644
index 0000000000..a2bb283222
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/highbd_convolve_sse2.c
@@ -0,0 +1,351 @@
+/*
+ * 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 <emmintrin.h>
+
+#include "config/aom_dsp_rtcd.h"
+#include "aom_dsp/x86/convolve.h"
+
+// -----------------------------------------------------------------------------
+
+void aom_highbd_filter_block1d4_v4_sse2(const uint16_t *src_ptr,
+                                        ptrdiff_t src_pitch, uint16_t *dst_ptr,
+                                        ptrdiff_t dst_pitch, uint32_t height,
+                                        const int16_t *filter, int bd) {
+  __m128i filtersReg;
+  __m128i srcReg2, srcReg3, srcReg4, srcReg5, srcReg6;
+  __m128i srcReg23_lo, srcReg34_lo;
+  __m128i srcReg45_lo, srcReg56_lo;
+  __m128i resReg23_lo, resReg34_lo, resReg45_lo, resReg56_lo;
+  __m128i resReg23_45_lo, resReg34_56_lo;
+  __m128i resReg23_45, resReg34_56;
+  __m128i addFilterReg64, secondFilters, thirdFilters;
+  unsigned int i;
+  ptrdiff_t src_stride, dst_stride;
+
+  const __m128i max = _mm_set1_epi16((1 << bd) - 1);
+  addFilterReg64 = _mm_set1_epi32(64);
+  filtersReg = _mm_loadu_si128((const __m128i *)filter);
+
+  // coeffs 0 1 0 1 2 3 2 3
+  const __m128i tmp0 = _mm_unpacklo_epi32(filtersReg, filtersReg);
+  // coeffs 4 5 4 5 6 7 6 7
+  const __m128i tmp1 = _mm_unpackhi_epi32(filtersReg, filtersReg);
+
+  secondFilters = _mm_unpackhi_epi64(tmp0, tmp0);  // coeffs 2 3 2 3 2 3 2 3
+  thirdFilters = _mm_unpacklo_epi64(tmp1, tmp1);   // coeffs 4 5 4 5 4 5 4 5
+
+  // multiply the size of the source and destination stride by two
+  src_stride = src_pitch << 1;
+  dst_stride = dst_pitch << 1;
+
+  srcReg2 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 2));
+  srcReg3 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 3));
+  srcReg23_lo = _mm_unpacklo_epi16(srcReg2, srcReg3);
+
+  srcReg4 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 4));
+  srcReg34_lo = _mm_unpacklo_epi16(srcReg3, srcReg4);
+
+  for (i = height; i > 1; i -= 2) {
+    srcReg5 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 5));
+    srcReg45_lo = _mm_unpacklo_epi16(srcReg4, srcReg5);
+
+    srcReg6 = _mm_loadl_epi64((const __m128i *)(src_ptr + src_pitch * 6));
+    srcReg56_lo = _mm_unpacklo_epi16(srcReg5, srcReg6);
+
+    // multiply 2 adjacent elements with the filter and add the result
+
+    resReg23_lo = _mm_madd_epi16(srcReg23_lo, secondFilters);
+    resReg34_lo = _mm_madd_epi16(srcReg34_lo, secondFilters);
+    resReg45_lo = _mm_madd_epi16(srcReg45_lo, thirdFilters);
+    resReg56_lo = _mm_madd_epi16(srcReg56_lo, thirdFilters);
+
+    resReg23_45_lo = _mm_add_epi32(resReg23_lo, resReg45_lo);
+    resReg34_56_lo = _mm_add_epi32(resReg34_lo, resReg56_lo);
+
+    // shift by 7 bit each 32 bit
+    resReg23_45_lo = _mm_add_epi32(resReg23_45_lo, addFilterReg64);
+    resReg34_56_lo = _mm_add_epi32(resReg34_56_lo, addFilterReg64);
+    resReg23_45_lo = _mm_srai_epi32(resReg23_45_lo, 7);
+    resReg34_56_lo = _mm_srai_epi32(resReg34_56_lo, 7);
+
+    // shrink to 16 bit each 32 bits, the first lane contain the first
+    // convolve result and the second lane contain the second convolve
+    // result
+    resReg23_45 = _mm_packs_epi32(resReg23_45_lo, _mm_setzero_si128());
+    resReg34_56 = _mm_packs_epi32(resReg34_56_lo, _mm_setzero_si128());
+
+    resReg23_45 = _mm_max_epi16(resReg23_45, _mm_setzero_si128());
+    resReg23_45 = _mm_min_epi16(resReg23_45, max);
+    resReg34_56 = _mm_max_epi16(resReg34_56, _mm_setzero_si128());
+    resReg34_56 = _mm_min_epi16(resReg34_56, max);
+
+    src_ptr += src_stride;
+
+    _mm_storel_epi64((__m128i *)dst_ptr, (resReg23_45));
+    _mm_storel_epi64((__m128i *)(dst_ptr + dst_pitch), (resReg34_56));
+
+    dst_ptr += dst_stride;
+
+    // save part of the registers for next strides
+    srcReg23_lo = srcReg45_lo;
+    srcReg34_lo = srcReg56_lo;
+    srcReg4 = srcReg6;
+  }
+}
+
+void aom_highbd_filter_block1d4_h4_sse2(const uint16_t *src_ptr,
+                                        ptrdiff_t src_pitch, uint16_t *dst_ptr,
+                                        ptrdiff_t dst_pitch, uint32_t height,
+                                        const int16_t *filter, int bd) {
+  __m128i filtersReg;
+  __m128i addFilterReg64;
+  __m128i secondFilters, thirdFilters;
+  __m128i srcRegFilt32b1_1;
+  __m128i srcReg32b1;
+  unsigned int i;
+  src_ptr -= 3;
+  addFilterReg64 = _mm_set1_epi32(64);
+  filtersReg = _mm_loadu_si128((const __m128i *)filter);
+  const __m128i max = _mm_set1_epi16((1 << bd) - 1);
+
+  // coeffs 0 1 0 1 2 3 2 3
+  const __m128i tmp_0 = _mm_unpacklo_epi32(filtersReg, filtersReg);
+  // coeffs 4 5 4 5 6 7 6 7
+  const __m128i tmp_1 = _mm_unpackhi_epi32(filtersReg, filtersReg);
+
+  secondFilters = _mm_unpackhi_epi64(tmp_0, tmp_0);  // coeffs 2 3 2 3 2 3 2 3
+  thirdFilters = _mm_unpacklo_epi64(tmp_1, tmp_1);   // coeffs 4 5 4 5 4 5 4 5
+
+  for (i = height; i > 0; i -= 1) {
+    srcReg32b1 = _mm_loadu_si128((const __m128i *)(src_ptr + 2));
+
+    __m128i ss_3_1 = _mm_srli_si128(srcReg32b1, 2);
+    __m128i ss_4_1 = _mm_srli_si128(srcReg32b1, 4);
+    __m128i ss_5_1 = _mm_srli_si128(srcReg32b1, 6);
+    __m128i ss_23 = _mm_unpacklo_epi32(srcReg32b1, ss_3_1);
+    __m128i ss_45 = _mm_unpacklo_epi32(ss_4_1, ss_5_1);
+
+    ss_23 = _mm_madd_epi16(ss_23, secondFilters);
+    ss_45 = _mm_madd_epi16(ss_45, thirdFilters);
+    srcRegFilt32b1_1 = _mm_add_epi32(ss_23, ss_45);
+
+    // shift by 7 bit each 32 bit
+    srcRegFilt32b1_1 = _mm_add_epi32(srcRegFilt32b1_1, addFilterReg64);
+    srcRegFilt32b1_1 = _mm_srai_epi32(srcRegFilt32b1_1, 7);
+
+    srcRegFilt32b1_1 = _mm_packs_epi32(srcRegFilt32b1_1, _mm_setzero_si128());
+    srcRegFilt32b1_1 = _mm_max_epi16(srcRegFilt32b1_1, _mm_setzero_si128());
+    srcRegFilt32b1_1 = _mm_min_epi16(srcRegFilt32b1_1, max);
+
+    src_ptr += src_pitch;
+
+    _mm_storel_epi64((__m128i *)dst_ptr, srcRegFilt32b1_1);
+
+    dst_ptr += dst_pitch;
+  }
+}
+
+void aom_highbd_filter_block1d8_v4_sse2(const uint16_t *src_ptr,
+                                        ptrdiff_t src_pitch, uint16_t *dst_ptr,
+                                        ptrdiff_t dst_pitch, uint32_t height,
+                                        const int16_t *filter, int bd) {
+  __m128i filtersReg;
+  __m128i srcReg2, srcReg3, srcReg4, srcReg5, srcReg6;
+  __m128i srcReg23_lo, srcReg23_hi, srcReg34_lo, srcReg34_hi;
+  __m128i srcReg45_lo, srcReg45_hi, srcReg56_lo, srcReg56_hi;
+  __m128i resReg23_lo, resReg34_lo, resReg45_lo, resReg56_lo;
+  __m128i resReg23_hi, resReg34_hi, resReg45_hi, resReg56_hi;
+  __m128i resReg23_45_lo, resReg34_56_lo, resReg23_45_hi, resReg34_56_hi;
+  __m128i resReg23_45, resReg34_56;
+  __m128i addFilterReg64, secondFilters, thirdFilters;
+  unsigned int i;
+  ptrdiff_t src_stride, dst_stride;
+
+  const __m128i max = _mm_set1_epi16((1 << bd) - 1);
+  addFilterReg64 = _mm_set1_epi32(64);
+  filtersReg = _mm_loadu_si128((const __m128i *)filter);
+
+  // coeffs 0 1 0 1 2 3 2 3
+  const __m128i tmp0 = _mm_unpacklo_epi32(filtersReg, filtersReg);
+  // coeffs 4 5 4 5 6 7 6 7
+  const __m128i tmp1 = _mm_unpackhi_epi32(filtersReg, filtersReg);
+
+  secondFilters = _mm_unpackhi_epi64(tmp0, tmp0);  // coeffs 2 3 2 3 2 3 2 3
+  thirdFilters = _mm_unpacklo_epi64(tmp1, tmp1);   // coeffs 4 5 4 5 4 5 4 5
+
+  // multiple the size of the source and destination stride by two
+  src_stride = src_pitch << 1;
+  dst_stride = dst_pitch << 1;
+
+  srcReg2 = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 2));
+  srcReg3 = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 3));
+  srcReg23_lo = _mm_unpacklo_epi16(srcReg2, srcReg3);
+  srcReg23_hi = _mm_unpackhi_epi16(srcReg2, srcReg3);
+
+  srcReg4 = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 4));
+  srcReg34_lo = _mm_unpacklo_epi16(srcReg3, srcReg4);
+  srcReg34_hi = _mm_unpackhi_epi16(srcReg3, srcReg4);
+
+  for (i = height; i > 1; i -= 2) {
+    srcReg5 = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 5));
+
+    srcReg45_lo = _mm_unpacklo_epi16(srcReg4, srcReg5);
+    srcReg45_hi = _mm_unpackhi_epi16(srcReg4, srcReg5);
+
+    srcReg6 = _mm_loadu_si128((const __m128i *)(src_ptr + src_pitch * 6));
+
+    srcReg56_lo = _mm_unpacklo_epi16(srcReg5, srcReg6);
+    srcReg56_hi = _mm_unpackhi_epi16(srcReg5, srcReg6);
+
+    // multiply 2 adjacent elements with the filter and add the result
+
+    resReg23_lo = _mm_madd_epi16(srcReg23_lo, secondFilters);
+    resReg34_lo = _mm_madd_epi16(srcReg34_lo, secondFilters);
+    resReg45_lo = _mm_madd_epi16(srcReg45_lo, thirdFilters);
+    resReg56_lo = _mm_madd_epi16(srcReg56_lo, thirdFilters);
+
+    resReg23_45_lo = _mm_add_epi32(resReg23_lo, resReg45_lo);
+    resReg34_56_lo = _mm_add_epi32(resReg34_lo, resReg56_lo);
+
+    // multiply 2 adjacent elements with the filter and add the result
+
+    resReg23_hi = _mm_madd_epi16(srcReg23_hi, secondFilters);
+    resReg34_hi = _mm_madd_epi16(srcReg34_hi, secondFilters);
+    resReg45_hi = _mm_madd_epi16(srcReg45_hi, thirdFilters);
+    resReg56_hi = _mm_madd_epi16(srcReg56_hi, thirdFilters);
+
+    resReg23_45_hi = _mm_add_epi32(resReg23_hi, resReg45_hi);
+    resReg34_56_hi = _mm_add_epi32(resReg34_hi, resReg56_hi);
+
+    // shift by 7 bit each 32 bit
+    resReg23_45_lo = _mm_add_epi32(resReg23_45_lo, addFilterReg64);
+    resReg34_56_lo = _mm_add_epi32(resReg34_56_lo, addFilterReg64);
+    resReg23_45_hi = _mm_add_epi32(resReg23_45_hi, addFilterReg64);
+    resReg34_56_hi = _mm_add_epi32(resReg34_56_hi, addFilterReg64);
+    resReg23_45_lo = _mm_srai_epi32(resReg23_45_lo, 7);
+    resReg34_56_lo = _mm_srai_epi32(resReg34_56_lo, 7);
+    resReg23_45_hi = _mm_srai_epi32(resReg23_45_hi, 7);
+    resReg34_56_hi = _mm_srai_epi32(resReg34_56_hi, 7);
+
+    // shrink to 16 bit each 32 bits, the first lane contain the first
+    // convolve result and the second lane contain the second convolve
+    // result
+    resReg23_45 = _mm_packs_epi32(resReg23_45_lo, resReg23_45_hi);
+    resReg34_56 = _mm_packs_epi32(resReg34_56_lo, resReg34_56_hi);
+
+    resReg23_45 = _mm_max_epi16(resReg23_45, _mm_setzero_si128());
+    resReg23_45 = _mm_min_epi16(resReg23_45, max);
+    resReg34_56 = _mm_max_epi16(resReg34_56, _mm_setzero_si128());
+    resReg34_56 = _mm_min_epi16(resReg34_56, max);
+
+    src_ptr += src_stride;
+
+    _mm_store_si128((__m128i *)dst_ptr, (resReg23_45));
+    _mm_store_si128((__m128i *)(dst_ptr + dst_pitch), (resReg34_56));
+
+    dst_ptr += dst_stride;
+
+    // save part of the registers for next strides
+    srcReg23_lo = srcReg45_lo;
+    srcReg23_hi = srcReg45_hi;
+    srcReg34_lo = srcReg56_lo;
+    srcReg34_hi = srcReg56_hi;
+    srcReg4 = srcReg6;
+  }
+}
+
+void aom_highbd_filter_block1d8_h4_sse2(const uint16_t *src_ptr,
+                                        ptrdiff_t src_pitch, uint16_t *dst_ptr,
+                                        ptrdiff_t dst_pitch, uint32_t height,
+                                        const int16_t *filter, int bd) {
+  __m128i filtersReg;
+  __m128i addFilterReg64;
+  __m128i secondFilters, thirdFilters;
+  __m128i srcRegFilt32b1_1, srcRegFilt32b1_2;
+  __m128i srcReg32b1, srcReg32b2;
+  unsigned int i;
+  src_ptr -= 3;
+  addFilterReg64 = _mm_set1_epi32(64);
+  filtersReg = _mm_loadu_si128((const __m128i *)filter);
+  const __m128i max = _mm_set1_epi16((1 << bd) - 1);
+
+  // coeffs 0 1 0 1 2 3 2 3
+  const __m128i tmp_0 = _mm_unpacklo_epi32(filtersReg, filtersReg);
+  // coeffs 4 5 4 5 6 7 6 7
+  const __m128i tmp_1 = _mm_unpackhi_epi32(filtersReg, filtersReg);
+
+  secondFilters = _mm_unpackhi_epi64(tmp_0, tmp_0);  // coeffs 2 3 2 3 2 3 2 3
+  thirdFilters = _mm_unpacklo_epi64(tmp_1, tmp_1);   // coeffs 4 5 4 5 4 5 4 5
+
+  for (i = height; i > 0; i -= 1) {
+    srcReg32b1 = _mm_loadu_si128((const __m128i *)(src_ptr + 2));
+    srcReg32b2 = _mm_loadu_si128((const __m128i *)(src_ptr + 6));
+
+    __m128i ss_4_1 = _mm_srli_si128(srcReg32b1, 4);
+    __m128i ss_4_2 = _mm_srli_si128(srcReg32b2, 4);
+    __m128i ss_4 = _mm_unpacklo_epi64(ss_4_1, ss_4_2);
+
+    __m128i d1 = _mm_madd_epi16(srcReg32b1, secondFilters);
+    __m128i d2 = _mm_madd_epi16(ss_4, thirdFilters);
+    srcRegFilt32b1_1 = _mm_add_epi32(d1, d2);
+
+    __m128i ss_3_1 = _mm_srli_si128(srcReg32b1, 2);
+    __m128i ss_5_1 = _mm_srli_si128(srcReg32b1, 6);
+    __m128i ss_3_2 = _mm_srli_si128(srcReg32b2, 2);
+    __m128i ss_5_2 = _mm_srli_si128(srcReg32b2, 6);
+    __m128i ss_3 = _mm_unpacklo_epi64(ss_3_1, ss_3_2);
+    __m128i ss_5 = _mm_unpacklo_epi64(ss_5_1, ss_5_2);
+
+    d1 = _mm_madd_epi16(ss_3, secondFilters);
+    d2 = _mm_madd_epi16(ss_5, thirdFilters);
+    srcRegFilt32b1_2 = _mm_add_epi32(d1, d2);
+
+    __m128i res_lo_1 = _mm_unpacklo_epi32(srcRegFilt32b1_1, srcRegFilt32b1_2);
+    __m128i res_hi_1 = _mm_unpackhi_epi32(srcRegFilt32b1_1, srcRegFilt32b1_2);
+
+    // shift by 7 bit each 32 bit
+    res_lo_1 = _mm_add_epi32(res_lo_1, addFilterReg64);
+    res_hi_1 = _mm_add_epi32(res_hi_1, addFilterReg64);
+    res_lo_1 = _mm_srai_epi32(res_lo_1, 7);
+    res_hi_1 = _mm_srai_epi32(res_hi_1, 7);
+
+    srcRegFilt32b1_1 = _mm_packs_epi32(res_lo_1, res_hi_1);
+
+    srcRegFilt32b1_1 = _mm_max_epi16(srcRegFilt32b1_1, _mm_setzero_si128());
+    srcRegFilt32b1_1 = _mm_min_epi16(srcRegFilt32b1_1, max);
+
+    src_ptr += src_pitch;
+
+    _mm_store_si128((__m128i *)dst_ptr, srcRegFilt32b1_1);
+
+    dst_ptr += dst_pitch;
+  }
+}
+
+void aom_highbd_filter_block1d16_v4_sse2(const uint16_t *src_ptr,
+                                         ptrdiff_t src_pitch, uint16_t *dst_ptr,
+                                         ptrdiff_t dst_pitch, uint32_t height,
+                                         const int16_t *filter, int bd) {
+  aom_highbd_filter_block1d8_v4_sse2(src_ptr, src_pitch, dst_ptr, dst_pitch,
+                                     height, filter, bd);
+  aom_highbd_filter_block1d8_v4_sse2((src_ptr + 8), src_pitch, (dst_ptr + 8),
+                                     dst_pitch, height, filter, bd);
+}
+
+void aom_highbd_filter_block1d16_h4_sse2(const uint16_t *src_ptr,
+                                         ptrdiff_t src_pitch, uint16_t *dst_ptr,
+                                         ptrdiff_t dst_pitch, uint32_t height,
+                                         const int16_t *filter, int bd) {
+  aom_highbd_filter_block1d8_h4_sse2(src_ptr, src_pitch, dst_ptr, dst_pitch,
+                                     height, filter, bd);
+  aom_highbd_filter_block1d8_h4_sse2((src_ptr + 8), src_pitch, (dst_ptr + 8),
+                                     dst_pitch, height, filter, bd);
+}
diff --git a/third_party/aom/aom_dsp/x86/highbd_convolve_ssse3.c b/third_party/aom/aom_dsp/x86/highbd_convolve_ssse3.c
new file mode 100644
index 0000000000..31c3c31b3c
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/highbd_convolve_ssse3.c
@@ -0,0 +1,439 @@
+/*
+ * 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 <assert.h>
+
+#include "config/av1_rtcd.h"
+
+#include "aom_dsp/x86/convolve_sse2.h"
+#include "aom_dsp/x86/convolve_common_intrin.h"
+
+void av1_highbd_convolve_y_sr_ssse3(const uint16_t *src, int src_stride,
+                                    uint16_t *dst, int dst_stride, int w, int h,
+                                    const InterpFilterParams *filter_params_y,
+                                    const int subpel_y_qn, int bd) {
+  int i, j;
+  const int fo_vert = filter_params_y->taps / 2 - 1;
+  const uint16_t *const src_ptr = src - fo_vert * src_stride;
+  const int bits = FILTER_BITS;
+
+  const __m128i round_shift_bits = _mm_cvtsi32_si128(bits);
+  const __m128i round_const_bits = _mm_set1_epi32((1 << bits) >> 1);
+  const __m128i clip_pixel =
+      _mm_set1_epi16(bd == 10 ? 1023 : (bd == 12 ? 4095 : 255));
+  const __m128i zero = _mm_setzero_si128();
+  if (filter_params_y->taps == 12) {
+    __m128i s[24], coeffs_y[6];
+
+    prepare_coeffs_12tap(filter_params_y, subpel_y_qn, coeffs_y);
+
+    for (j = 0; j < w; j += 8) {
+      const uint16_t *data = &src_ptr[j];
+      /* Vertical filter */
+      __m128i s0 = _mm_loadu_si128((__m128i *)(data + 0 * src_stride));
+      __m128i s1 = _mm_loadu_si128((__m128i *)(data + 1 * src_stride));
+      __m128i s2 = _mm_loadu_si128((__m128i *)(data + 2 * src_stride));
+      __m128i s3 = _mm_loadu_si128((__m128i *)(data + 3 * src_stride));
+      __m128i s4 = _mm_loadu_si128((__m128i *)(data + 4 * src_stride));
+      __m128i s5 = _mm_loadu_si128((__m128i *)(data + 5 * src_stride));
+      __m128i s6 = _mm_loadu_si128((__m128i *)(data + 6 * src_stride));
+      __m128i s7 = _mm_loadu_si128((__m128i *)(data + 7 * src_stride));
+      __m128i s8 = _mm_loadu_si128((__m128i *)(data + 8 * src_stride));
+      __m128i s9 = _mm_loadu_si128((__m128i *)(data + 9 * src_stride));
+      __m128i s10 = _mm_loadu_si128((__m128i *)(data + 10 * src_stride));
+
+      s[0] = _mm_unpacklo_epi16(s0, s1);
+      s[1] = _mm_unpacklo_epi16(s2, s3);
+      s[2] = _mm_unpacklo_epi16(s4, s5);
+      s[3] = _mm_unpacklo_epi16(s6, s7);
+      s[4] = _mm_unpacklo_epi16(s8, s9);
+
+      s[6] = _mm_unpackhi_epi16(s0, s1);
+      s[7] = _mm_unpackhi_epi16(s2, s3);
+      s[8] = _mm_unpackhi_epi16(s4, s5);
+      s[9] = _mm_unpackhi_epi16(s6, s7);
+      s[10] = _mm_unpackhi_epi16(s8, s9);
+
+      s[12] = _mm_unpacklo_epi16(s1, s2);
+      s[13] = _mm_unpacklo_epi16(s3, s4);
+      s[14] = _mm_unpacklo_epi16(s5, s6);
+      s[15] = _mm_unpacklo_epi16(s7, s8);
+      s[16] = _mm_unpacklo_epi16(s9, s10);
+
+      s[18] = _mm_unpackhi_epi16(s1, s2);
+      s[19] = _mm_unpackhi_epi16(s3, s4);
+      s[20] = _mm_unpackhi_epi16(s5, s6);
+      s[21] = _mm_unpackhi_epi16(s7, s8);
+      s[22] = _mm_unpackhi_epi16(s9, s10);
+
+      for (i = 0; i < h; i += 2) {
+        data = &src_ptr[i * src_stride + j];
+
+        __m128i s11 = _mm_loadu_si128((__m128i *)(data + 11 * src_stride));
+        __m128i s12 = _mm_loadu_si128((__m128i *)(data + 12 * src_stride));
+
+        s[5] = _mm_unpacklo_epi16(s10, s11);
+        s[11] = _mm_unpackhi_epi16(s10, s11);
+
+        s[17] = _mm_unpacklo_epi16(s11, s12);
+        s[23] = _mm_unpackhi_epi16(s11, s12);
+
+        const __m128i res_a0 = convolve_12tap(s, coeffs_y);
+        __m128i res_a_round0 = _mm_sra_epi32(
+            _mm_add_epi32(res_a0, round_const_bits), round_shift_bits);
+
+        const __m128i res_a1 = convolve_12tap(s + 12, coeffs_y);
+        __m128i res_a_round1 = _mm_sra_epi32(
+            _mm_add_epi32(res_a1, round_const_bits), round_shift_bits);
+
+        if (w - j > 4) {
+          const __m128i res_b0 = convolve_12tap(s + 6, coeffs_y);
+          __m128i res_b_round0 = _mm_sra_epi32(
+              _mm_add_epi32(res_b0, round_const_bits), round_shift_bits);
+
+          const __m128i res_b1 = convolve_12tap(s + 18, coeffs_y);
+          __m128i res_b_round1 = _mm_sra_epi32(
+              _mm_add_epi32(res_b1, round_const_bits), round_shift_bits);
+
+          __m128i res_16bit0 = _mm_packs_epi32(res_a_round0, res_b_round0);
+          res_16bit0 = _mm_min_epi16(res_16bit0, clip_pixel);
+          res_16bit0 = _mm_max_epi16(res_16bit0, zero);
+
+          __m128i res_16bit1 = _mm_packs_epi32(res_a_round1, res_b_round1);
+          res_16bit1 = _mm_min_epi16(res_16bit1, clip_pixel);
+          res_16bit1 = _mm_max_epi16(res_16bit1, zero);
+
+          _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j], res_16bit0);
+          _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j + dst_stride],
+                           res_16bit1);
+        } else if (w == 4) {
+          res_a_round0 = _mm_packs_epi32(res_a_round0, res_a_round0);
+          res_a_round0 = _mm_min_epi16(res_a_round0, clip_pixel);
+          res_a_round0 = _mm_max_epi16(res_a_round0, zero);
+
+          res_a_round1 = _mm_packs_epi32(res_a_round1, res_a_round1);
+          res_a_round1 = _mm_min_epi16(res_a_round1, clip_pixel);
+          res_a_round1 = _mm_max_epi16(res_a_round1, zero);
+
+          _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j], res_a_round0);
+          _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j + dst_stride],
+                           res_a_round1);
+        } else {
+          res_a_round0 = _mm_packs_epi32(res_a_round0, res_a_round0);
+          res_a_round0 = _mm_min_epi16(res_a_round0, clip_pixel);
+          res_a_round0 = _mm_max_epi16(res_a_round0, zero);
+
+          res_a_round1 = _mm_packs_epi32(res_a_round1, res_a_round1);
+          res_a_round1 = _mm_min_epi16(res_a_round1, clip_pixel);
+          res_a_round1 = _mm_max_epi16(res_a_round1, zero);
+
+          *((int *)(&dst[i * dst_stride + j])) =
+              _mm_cvtsi128_si32(res_a_round0);
+
+          *((int *)(&dst[i * dst_stride + j + dst_stride])) =
+              _mm_cvtsi128_si32(res_a_round1);
+        }
+
+        s[0] = s[1];
+        s[1] = s[2];
+        s[2] = s[3];
+        s[3] = s[4];
+        s[4] = s[5];
+
+        s[6] = s[7];
+        s[7] = s[8];
+        s[8] = s[9];
+        s[9] = s[10];
+        s[10] = s[11];
+
+        s[12] = s[13];
+        s[13] = s[14];
+        s[14] = s[15];
+        s[15] = s[16];
+        s[16] = s[17];
+
+        s[18] = s[19];
+        s[19] = s[20];
+        s[20] = s[21];
+        s[21] = s[22];
+        s[22] = s[23];
+
+        s10 = s12;
+      }
+    }
+  } else {
+    __m128i s[16], coeffs_y[4];
+
+    prepare_coeffs(filter_params_y, subpel_y_qn, coeffs_y);
+
+    for (j = 0; j < w; j += 8) {
+      const uint16_t *data = &src_ptr[j];
+      /* Vertical filter */
+      {
+        __m128i s0 = _mm_loadu_si128((__m128i *)(data + 0 * src_stride));
+        __m128i s1 = _mm_loadu_si128((__m128i *)(data + 1 * src_stride));
+        __m128i s2 = _mm_loadu_si128((__m128i *)(data + 2 * src_stride));
+        __m128i s3 = _mm_loadu_si128((__m128i *)(data + 3 * src_stride));
+        __m128i s4 = _mm_loadu_si128((__m128i *)(data + 4 * src_stride));
+        __m128i s5 = _mm_loadu_si128((__m128i *)(data + 5 * src_stride));
+        __m128i s6 = _mm_loadu_si128((__m128i *)(data + 6 * src_stride));
+
+        s[0] = _mm_unpacklo_epi16(s0, s1);
+        s[1] = _mm_unpacklo_epi16(s2, s3);
+        s[2] = _mm_unpacklo_epi16(s4, s5);
+
+        s[4] = _mm_unpackhi_epi16(s0, s1);
+        s[5] = _mm_unpackhi_epi16(s2, s3);
+        s[6] = _mm_unpackhi_epi16(s4, s5);
+
+        s[0 + 8] = _mm_unpacklo_epi16(s1, s2);
+        s[1 + 8] = _mm_unpacklo_epi16(s3, s4);
+        s[2 + 8] = _mm_unpacklo_epi16(s5, s6);
+
+        s[4 + 8] = _mm_unpackhi_epi16(s1, s2);
+        s[5 + 8] = _mm_unpackhi_epi16(s3, s4);
+        s[6 + 8] = _mm_unpackhi_epi16(s5, s6);
+
+        for (i = 0; i < h; i += 2) {
+          data = &src_ptr[i * src_stride + j];
+
+          __m128i s7 = _mm_loadu_si128((__m128i *)(data + 7 * src_stride));
+          __m128i s8 = _mm_loadu_si128((__m128i *)(data + 8 * src_stride));
+
+          s[3] = _mm_unpacklo_epi16(s6, s7);
+          s[7] = _mm_unpackhi_epi16(s6, s7);
+
+          s[3 + 8] = _mm_unpacklo_epi16(s7, s8);
+          s[7 + 8] = _mm_unpackhi_epi16(s7, s8);
+
+          const __m128i res_a0 = convolve(s, coeffs_y);
+          __m128i res_a_round0 = _mm_sra_epi32(
+              _mm_add_epi32(res_a0, round_const_bits), round_shift_bits);
+
+          const __m128i res_a1 = convolve(s + 8, coeffs_y);
+          __m128i res_a_round1 = _mm_sra_epi32(
+              _mm_add_epi32(res_a1, round_const_bits), round_shift_bits);
+
+          if (w - j > 4) {
+            const __m128i res_b0 = convolve(s + 4, coeffs_y);
+            __m128i res_b_round0 = _mm_sra_epi32(
+                _mm_add_epi32(res_b0, round_const_bits), round_shift_bits);
+
+            const __m128i res_b1 = convolve(s + 4 + 8, coeffs_y);
+            __m128i res_b_round1 = _mm_sra_epi32(
+                _mm_add_epi32(res_b1, round_const_bits), round_shift_bits);
+
+            __m128i res_16bit0 = _mm_packs_epi32(res_a_round0, res_b_round0);
+            res_16bit0 = _mm_min_epi16(res_16bit0, clip_pixel);
+            res_16bit0 = _mm_max_epi16(res_16bit0, zero);
+
+            __m128i res_16bit1 = _mm_packs_epi32(res_a_round1, res_b_round1);
+            res_16bit1 = _mm_min_epi16(res_16bit1, clip_pixel);
+            res_16bit1 = _mm_max_epi16(res_16bit1, zero);
+
+            _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j], res_16bit0);
+            _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j + dst_stride],
+                             res_16bit1);
+          } else if (w == 4) {
+            res_a_round0 = _mm_packs_epi32(res_a_round0, res_a_round0);
+            res_a_round0 = _mm_min_epi16(res_a_round0, clip_pixel);
+            res_a_round0 = _mm_max_epi16(res_a_round0, zero);
+
+            res_a_round1 = _mm_packs_epi32(res_a_round1, res_a_round1);
+            res_a_round1 = _mm_min_epi16(res_a_round1, clip_pixel);
+            res_a_round1 = _mm_max_epi16(res_a_round1, zero);
+
+            _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j], res_a_round0);
+            _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j + dst_stride],
+                             res_a_round1);
+          } else {
+            res_a_round0 = _mm_packs_epi32(res_a_round0, res_a_round0);
+            res_a_round0 = _mm_min_epi16(res_a_round0, clip_pixel);
+            res_a_round0 = _mm_max_epi16(res_a_round0, zero);
+
+            res_a_round1 = _mm_packs_epi32(res_a_round1, res_a_round1);
+            res_a_round1 = _mm_min_epi16(res_a_round1, clip_pixel);
+            res_a_round1 = _mm_max_epi16(res_a_round1, zero);
+
+            *((int *)(&dst[i * dst_stride + j])) =
+                _mm_cvtsi128_si32(res_a_round0);
+
+            *((int *)(&dst[i * dst_stride + j + dst_stride])) =
+                _mm_cvtsi128_si32(res_a_round1);
+          }
+
+          s[0] = s[1];
+          s[1] = s[2];
+          s[2] = s[3];
+
+          s[4] = s[5];
+          s[5] = s[6];
+          s[6] = s[7];
+
+          s[0 + 8] = s[1 + 8];
+          s[1 + 8] = s[2 + 8];
+          s[2 + 8] = s[3 + 8];
+
+          s[4 + 8] = s[5 + 8];
+          s[5 + 8] = s[6 + 8];
+          s[6 + 8] = s[7 + 8];
+
+          s6 = s8;
+        }
+      }
+    }
+  }
+}
+
+void av1_highbd_convolve_x_sr_ssse3(const uint16_t *src, int src_stride,
+                                    uint16_t *dst, int dst_stride, int w, int h,
+                                    const InterpFilterParams *filter_params_x,
+                                    const int subpel_x_qn,
+                                    ConvolveParams *conv_params, int bd) {
+  int i, j;
+  const int fo_horiz = filter_params_x->taps / 2 - 1;
+  const uint16_t *const src_ptr = src - fo_horiz;
+
+  // Check that, even with 12-bit input, the intermediate values will fit
+  // into an unsigned 16-bit intermediate array.
+  assert(bd + FILTER_BITS + 2 - conv_params->round_0 <= 16);
+
+  const __m128i round_const_x =
+      _mm_set1_epi32(((1 << conv_params->round_0) >> 1));
+  const __m128i round_shift_x = _mm_cvtsi32_si128(conv_params->round_0);
+
+  const int bits = FILTER_BITS - conv_params->round_0;
+
+  const __m128i round_shift_bits = _mm_cvtsi32_si128(bits);
+  const __m128i round_const_bits = _mm_set1_epi32((1 << bits) >> 1);
+  const __m128i clip_pixel =
+      _mm_set1_epi16(bd == 10 ? 1023 : (bd == 12 ? 4095 : 255));
+  const __m128i zero = _mm_setzero_si128();
+
+  if (filter_params_x->taps == 12) {
+    __m128i s[6], coeffs_x[6];
+
+    prepare_coeffs_12tap(filter_params_x, subpel_x_qn, coeffs_x);
+
+    for (j = 0; j < w; j += 8) {
+      /* Horizontal filter */
+      {
+        for (i = 0; i < h; i += 1) {
+          const __m128i row00 =
+              _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + j]);
+          const __m128i row01 =
+              _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + (j + 8)]);
+          const __m128i row02 =
+              _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + (j + 16)]);
+
+          // even pixels
+          s[0] = _mm_alignr_epi8(row01, row00, 0);
+          s[1] = _mm_alignr_epi8(row01, row00, 4);
+          s[2] = _mm_alignr_epi8(row01, row00, 8);
+          s[3] = _mm_alignr_epi8(row01, row00, 12);
+          s[4] = _mm_alignr_epi8(row02, row01, 0);
+          s[5] = _mm_alignr_epi8(row02, row01, 4);
+
+          __m128i res_even = convolve_12tap(s, coeffs_x);
+          res_even = _mm_sra_epi32(_mm_add_epi32(res_even, round_const_x),
+                                   round_shift_x);
+          res_even = _mm_sra_epi32(_mm_add_epi32(res_even, round_const_bits),
+                                   round_shift_bits);
+
+          // odd pixels
+          s[0] = _mm_alignr_epi8(row01, row00, 2);
+          s[1] = _mm_alignr_epi8(row01, row00, 6);
+          s[2] = _mm_alignr_epi8(row01, row00, 10);
+          s[3] = _mm_alignr_epi8(row01, row00, 14);
+          s[4] = _mm_alignr_epi8(row02, row01, 2);
+          s[5] = _mm_alignr_epi8(row02, row01, 6);
+
+          __m128i res_odd = convolve_12tap(s, coeffs_x);
+          res_odd = _mm_sra_epi32(_mm_add_epi32(res_odd, round_const_x),
+                                  round_shift_x);
+          res_odd = _mm_sra_epi32(_mm_add_epi32(res_odd, round_const_bits),
+                                  round_shift_bits);
+
+          __m128i res_even1 = _mm_packs_epi32(res_even, res_even);
+          __m128i res_odd1 = _mm_packs_epi32(res_odd, res_odd);
+          __m128i res = _mm_unpacklo_epi16(res_even1, res_odd1);
+
+          res = _mm_min_epi16(res, clip_pixel);
+          res = _mm_max_epi16(res, zero);
+
+          if (w - j > 4) {
+            _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j], res);
+          } else if (w == 4) {
+            _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j], res);
+          } else {
+            *((int *)(&dst[i * dst_stride + j])) = _mm_cvtsi128_si32(res);
+          }
+        }
+      }
+    }
+  } else {
+    __m128i s[4], coeffs_x[4];
+    prepare_coeffs(filter_params_x, subpel_x_qn, coeffs_x);
+
+    for (j = 0; j < w; j += 8) {
+      /* Horizontal filter */
+      {
+        for (i = 0; i < h; i += 1) {
+          const __m128i row00 =
+              _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + j]);
+          const __m128i row01 =
+              _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + (j + 8)]);
+
+          // even pixels
+          s[0] = _mm_alignr_epi8(row01, row00, 0);
+          s[1] = _mm_alignr_epi8(row01, row00, 4);
+          s[2] = _mm_alignr_epi8(row01, row00, 8);
+          s[3] = _mm_alignr_epi8(row01, row00, 12);
+
+          __m128i res_even = convolve(s, coeffs_x);
+          res_even = _mm_sra_epi32(_mm_add_epi32(res_even, round_const_x),
+                                   round_shift_x);
+
+          // odd pixels
+          s[0] = _mm_alignr_epi8(row01, row00, 2);
+          s[1] = _mm_alignr_epi8(row01, row00, 6);
+          s[2] = _mm_alignr_epi8(row01, row00, 10);
+          s[3] = _mm_alignr_epi8(row01, row00, 14);
+
+          __m128i res_odd = convolve(s, coeffs_x);
+          res_odd = _mm_sra_epi32(_mm_add_epi32(res_odd, round_const_x),
+                                  round_shift_x);
+
+          res_even = _mm_sra_epi32(_mm_add_epi32(res_even, round_const_bits),
+                                   round_shift_bits);
+          res_odd = _mm_sra_epi32(_mm_add_epi32(res_odd, round_const_bits),
+                                  round_shift_bits);
+
+          __m128i res_even1 = _mm_packs_epi32(res_even, res_even);
+          __m128i res_odd1 = _mm_packs_epi32(res_odd, res_odd);
+          __m128i res = _mm_unpacklo_epi16(res_even1, res_odd1);
+
+          res = _mm_min_epi16(res, clip_pixel);
+          res = _mm_max_epi16(res, zero);
+
+          if (w - j > 4) {
+            _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j], res);
+          } else if (w == 4) {
+            _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j], res);
+          } else {
+            *((int *)(&dst[i * dst_stride + j])) = _mm_cvtsi128_si32(res);
+          }
+        }
+      }
+    }
+  }
+}
diff --git a/third_party/aom/aom_dsp/x86/highbd_intrapred_asm_sse2.asm b/third_party/aom/aom_dsp/x86/highbd_intrapred_asm_sse2.asm
new file mode 100644
index 0000000000..91b3d126ca
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/highbd_intrapred_asm_sse2.asm
@@ -0,0 +1,259 @@
+;
+; Copyright (c) 2016, 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 "third_party/x86inc/x86inc.asm"
+
+SECTION_RODATA
+pw_4:  times 8 dw 4
+pw_8:  times 8 dw 8
+pw_16: times 4 dd 16
+pw_32: times 4 dd 32
+
+SECTION .text
+INIT_XMM sse2
+cglobal highbd_dc_predictor_4x4, 4, 5, 4, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  movq                  m0, [aboveq]
+  movq                  m2, [leftq]
+  paddw                 m0, m2
+  pshuflw               m1, m0, 0xe
+  paddw                 m0, m1
+  pshuflw               m1, m0, 0x1
+  paddw                 m0, m1
+  paddw                 m0, [GLOBAL(pw_4)]
+  psraw                 m0, 3
+  pshuflw               m0, m0, 0x0
+  movq    [dstq          ], m0
+  movq    [dstq+strideq*2], m0
+  lea                 dstq, [dstq+strideq*4]
+  movq    [dstq          ], m0
+  movq    [dstq+strideq*2], m0
+
+  RESTORE_GOT
+  RET
+
+INIT_XMM sse2
+cglobal highbd_dc_predictor_8x8, 4, 5, 4, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  mova                  m0, [aboveq]
+  mova                  m2, [leftq]
+  DEFINE_ARGS dst, stride, stride3, one
+  mov                 oned, 0x00010001
+  lea             stride3q, [strideq*3]
+  movd                  m3, oned
+  pshufd                m3, m3, 0x0
+  paddw                 m0, m2
+  pmaddwd               m0, m3
+  packssdw              m0, m1
+  pmaddwd               m0, m3
+  packssdw              m0, m1
+  pmaddwd               m0, m3
+  paddw                 m0, [GLOBAL(pw_8)]
+  psrlw                 m0, 4
+  pshuflw               m0, m0, 0x0
+  punpcklqdq            m0, m0
+  mova   [dstq           ], m0
+  mova   [dstq+strideq*2 ], m0
+  mova   [dstq+strideq*4 ], m0
+  mova   [dstq+stride3q*2], m0
+  lea                 dstq, [dstq+strideq*8]
+  mova   [dstq           ], m0
+  mova   [dstq+strideq*2 ], m0
+  mova   [dstq+strideq*4 ], m0
+  mova   [dstq+stride3q*2], m0
+
+  RESTORE_GOT
+  RET
+
+INIT_XMM sse2
+cglobal highbd_dc_predictor_16x16, 4, 5, 5, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  mova                  m0, [aboveq]
+  mova                  m3, [aboveq+16]
+  mova                  m2, [leftq]
+  mova                  m4, [leftq+16]
+  DEFINE_ARGS dst, stride, stride3, lines4
+  lea             stride3q, [strideq*3]
+  mov              lines4d, 4
+  paddw                 m0, m2
+  paddw                 m0, m3
+  paddw                 m0, m4
+  movhlps               m2, m0
+  paddw                 m0, m2
+  punpcklwd             m0, m1
+  movhlps               m2, m0
+  paddd                 m0, m2
+  punpckldq             m0, m1
+  movhlps               m2, m0
+  paddd                 m0, m2
+  paddd                 m0, [GLOBAL(pw_16)]
+  psrad                 m0, 5
+  pshuflw               m0, m0, 0x0
+  punpcklqdq            m0, m0
+.loop:
+  mova   [dstq              ], m0
+  mova   [dstq           +16], m0
+  mova   [dstq+strideq*2    ], m0
+  mova   [dstq+strideq*2 +16], m0
+  mova   [dstq+strideq*4    ], m0
+  mova   [dstq+strideq*4 +16], m0
+  mova   [dstq+stride3q*2   ], m0
+  mova   [dstq+stride3q*2+16], m0
+  lea                 dstq, [dstq+strideq*8]
+  dec              lines4d
+  jnz .loop
+
+  RESTORE_GOT
+  REP_RET
+
+INIT_XMM sse2
+cglobal highbd_dc_predictor_32x32, 4, 5, 7, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  mova                  m0, [aboveq]
+  mova                  m2, [aboveq+16]
+  mova                  m3, [aboveq+32]
+  mova                  m4, [aboveq+48]
+  paddw                 m0, m2
+  paddw                 m3, m4
+  mova                  m2, [leftq]
+  mova                  m4, [leftq+16]
+  mova                  m5, [leftq+32]
+  mova                  m6, [leftq+48]
+  paddw                 m2, m4
+  paddw                 m5, m6
+  paddw                 m0, m3
+  paddw                 m2, m5
+  pxor                  m1, m1
+  paddw                 m0, m2
+  DEFINE_ARGS dst, stride, stride3, lines4
+  lea             stride3q, [strideq*3]
+  mov              lines4d, 8
+  movhlps               m2, m0
+  paddw                 m0, m2
+  punpcklwd             m0, m1
+  movhlps               m2, m0
+  paddd                 m0, m2
+  punpckldq             m0, m1
+  movhlps               m2, m0
+  paddd                 m0, m2
+  paddd                 m0, [GLOBAL(pw_32)]
+  psrad                 m0, 6
+  pshuflw               m0, m0, 0x0
+  punpcklqdq            m0, m0
+.loop:
+  mova [dstq               ], m0
+  mova [dstq          +16  ], m0
+  mova [dstq          +32  ], m0
+  mova [dstq          +48  ], m0
+  mova [dstq+strideq*2     ], m0
+  mova [dstq+strideq*2+16  ], m0
+  mova [dstq+strideq*2+32  ], m0
+  mova [dstq+strideq*2+48  ], m0
+  mova [dstq+strideq*4     ], m0
+  mova [dstq+strideq*4+16  ], m0
+  mova [dstq+strideq*4+32  ], m0
+  mova [dstq+strideq*4+48  ], m0
+  mova [dstq+stride3q*2    ], m0
+  mova [dstq+stride3q*2 +16], m0
+  mova [dstq+stride3q*2 +32], m0
+  mova [dstq+stride3q*2 +48], m0
+  lea                 dstq, [dstq+strideq*8]
+  dec              lines4d
+  jnz .loop
+
+  RESTORE_GOT
+  REP_RET
+
+INIT_XMM sse2
+cglobal highbd_v_predictor_4x4, 3, 3, 1, dst, stride, above
+  movq                  m0, [aboveq]
+  movq    [dstq          ], m0
+  movq    [dstq+strideq*2], m0
+  lea                 dstq, [dstq+strideq*4]
+  movq    [dstq          ], m0
+  movq    [dstq+strideq*2], m0
+  RET
+
+INIT_XMM sse2
+cglobal highbd_v_predictor_8x8, 3, 3, 1, dst, stride, above
+  mova                  m0, [aboveq]
+  DEFINE_ARGS dst, stride, stride3
+  lea             stride3q, [strideq*3]
+  mova   [dstq           ], m0
+  mova   [dstq+strideq*2 ], m0
+  mova   [dstq+strideq*4 ], m0
+  mova   [dstq+stride3q*2], m0
+  lea                 dstq, [dstq+strideq*8]
+  mova   [dstq           ], m0
+  mova   [dstq+strideq*2 ], m0
+  mova   [dstq+strideq*4 ], m0
+  mova   [dstq+stride3q*2], m0
+  RET
+
+INIT_XMM sse2
+cglobal highbd_v_predictor_16x16, 3, 4, 2, dst, stride, above
+  mova                  m0, [aboveq]
+  mova                  m1, [aboveq+16]
+  DEFINE_ARGS dst, stride, stride3, nlines4
+  lea             stride3q, [strideq*3]
+  mov              nlines4d, 4
+.loop:
+  mova    [dstq              ], m0
+  mova    [dstq           +16], m1
+  mova    [dstq+strideq*2    ], m0
+  mova    [dstq+strideq*2 +16], m1
+  mova    [dstq+strideq*4    ], m0
+  mova    [dstq+strideq*4 +16], m1
+  mova    [dstq+stride3q*2   ], m0
+  mova    [dstq+stride3q*2+16], m1
+  lea                 dstq, [dstq+strideq*8]
+  dec             nlines4d
+  jnz .loop
+  REP_RET
+
+INIT_XMM sse2
+cglobal highbd_v_predictor_32x32, 3, 4, 4, dst, stride, above
+  mova                  m0, [aboveq]
+  mova                  m1, [aboveq+16]
+  mova                  m2, [aboveq+32]
+  mova                  m3, [aboveq+48]
+  DEFINE_ARGS dst, stride, stride3, nlines4
+  lea             stride3q, [strideq*3]
+  mov              nlines4d, 8
+.loop:
+  mova [dstq               ], m0
+  mova [dstq            +16], m1
+  mova [dstq            +32], m2
+  mova [dstq            +48], m3
+  mova [dstq+strideq*2     ], m0
+  mova [dstq+strideq*2  +16], m1
+  mova [dstq+strideq*2  +32], m2
+  mova [dstq+strideq*2  +48], m3
+  mova [dstq+strideq*4     ], m0
+  mova [dstq+strideq*4  +16], m1
+  mova [dstq+strideq*4  +32], m2
+  mova [dstq+strideq*4  +48], m3
+  mova [dstq+stride3q*2    ], m0
+  mova [dstq+stride3q*2 +16], m1
+  mova [dstq+stride3q*2 +32], m2
+  mova [dstq+stride3q*2 +48], m3
+  lea                 dstq, [dstq+strideq*8]
+  dec             nlines4d
+  jnz .loop
+  REP_RET
diff --git a/third_party/aom/aom_dsp/x86/highbd_intrapred_sse2.c b/third_party/aom/aom_dsp/x86/highbd_intrapred_sse2.c
new file mode 100644
index 0000000000..6a2e915ed7
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/highbd_intrapred_sse2.c
@@ -0,0 +1,984 @@
+/*
+ * Copyright (c) 2017, 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 <emmintrin.h>
+
+#include "config/aom_dsp_rtcd.h"
+
+// -----------------------------------------------------------------------------
+// H_PRED
+
+void aom_highbd_h_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t stride,
+                                     const uint16_t *above,
+                                     const uint16_t *left, int bd) {
+  const __m128i left_u16 = _mm_loadl_epi64((const __m128i *)left);
+  const __m128i row0 = _mm_shufflelo_epi16(left_u16, 0x0);
+  const __m128i row1 = _mm_shufflelo_epi16(left_u16, 0x55);
+  const __m128i row2 = _mm_shufflelo_epi16(left_u16, 0xaa);
+  const __m128i row3 = _mm_shufflelo_epi16(left_u16, 0xff);
+  (void)above;
+  (void)bd;
+  _mm_storel_epi64((__m128i *)dst, row0);
+  dst += stride;
+  _mm_storel_epi64((__m128i *)dst, row1);
+  dst += stride;
+  _mm_storel_epi64((__m128i *)dst, row2);
+  dst += stride;
+  _mm_storel_epi64((__m128i *)dst, row3);
+}
+
+void aom_highbd_h_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t stride,
+                                     const uint16_t *above,
+                                     const uint16_t *left, int bd) {
+  aom_highbd_h_predictor_4x4_sse2(dst, stride, above, left, bd);
+  dst += stride << 2;
+  left += 4;
+  aom_highbd_h_predictor_4x4_sse2(dst, stride, above, left, bd);
+}
+
+void aom_highbd_h_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t stride,
+                                     const uint16_t *above,
+                                     const uint16_t *left, int bd) {
+  const __m128i left_u16 = _mm_load_si128((const __m128i *)left);
+  const __m128i row0 = _mm_shufflelo_epi16(left_u16, 0x0);
+  const __m128i row1 = _mm_shufflelo_epi16(left_u16, 0x55);
+  const __m128i row2 = _mm_shufflelo_epi16(left_u16, 0xaa);
+  const __m128i row3 = _mm_shufflelo_epi16(left_u16, 0xff);
+  (void)above;
+  (void)bd;
+  _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row0, row0));
+  dst += stride;
+  _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row1, row1));
+  dst += stride;
+  _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row2, row2));
+  dst += stride;
+  _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row3, row3));
+}
+
+void aom_highbd_h_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t stride,
+                                     const uint16_t *above,
+                                     const uint16_t *left, int bd) {
+  const __m128i left_u16 = _mm_load_si128((const __m128i *)left);
+  const __m128i row0 = _mm_shufflelo_epi16(left_u16, 0x0);
+  const __m128i row1 = _mm_shufflelo_epi16(left_u16, 0x55);
+  const __m128i row2 = _mm_shufflelo_epi16(left_u16, 0xaa);
+  const __m128i row3 = _mm_shufflelo_epi16(left_u16, 0xff);
+  const __m128i row4 = _mm_shufflehi_epi16(left_u16, 0x0);
+  const __m128i row5 = _mm_shufflehi_epi16(left_u16, 0x55);
+  const __m128i row6 = _mm_shufflehi_epi16(left_u16, 0xaa);
+  const __m128i row7 = _mm_shufflehi_epi16(left_u16, 0xff);
+  (void)above;
+  (void)bd;
+  _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row0, row0));
+  dst += stride;
+  _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row1, row1));
+  dst += stride;
+  _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row2, row2));
+  dst += stride;
+  _mm_store_si128((__m128i *)dst, _mm_unpacklo_epi64(row3, row3));
+  dst += stride;
+  _mm_store_si128((__m128i *)dst, _mm_unpackhi_epi64(row4, row4));
+  dst += stride;
+  _mm_store_si128((__m128i *)dst, _mm_unpackhi_epi64(row5, row5));
+  dst += stride;
+  _mm_store_si128((__m128i *)dst, _mm_unpackhi_epi64(row6, row6));
+  dst += stride;
+  _mm_store_si128((__m128i *)dst, _mm_unpackhi_epi64(row7, row7));
+}
+
+void aom_highbd_h_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t stride,
+                                      const uint16_t *above,
+                                      const uint16_t *left, int bd) {
+  aom_highbd_h_predictor_8x8_sse2(dst, stride, above, left, bd);
+  dst += stride << 3;
+  left += 8;
+  aom_highbd_h_predictor_8x8_sse2(dst, stride, above, left, bd);
+}
+
+static INLINE void h_store_16_unpacklo(uint16_t **dst, const ptrdiff_t stride,
+                                       const __m128i *row) {
+  const __m128i val = _mm_unpacklo_epi64(*row, *row);
+  _mm_store_si128((__m128i *)*dst, val);
+  _mm_store_si128((__m128i *)(*dst + 8), val);
+  *dst += stride;
+}
+
+static INLINE void h_store_16_unpackhi(uint16_t **dst, const ptrdiff_t stride,
+                                       const __m128i *row) {
+  const __m128i val = _mm_unpackhi_epi64(*row, *row);
+  _mm_store_si128((__m128i *)(*dst), val);
+  _mm_store_si128((__m128i *)(*dst + 8), val);
+  *dst += stride;
+}
+
+static INLINE void h_predictor_16x8(uint16_t *dst, ptrdiff_t stride,
+                                    const uint16_t *left) {
+  const __m128i left_u16 = _mm_load_si128((const __m128i *)left);
+  const __m128i row0 = _mm_shufflelo_epi16(left_u16, 0x0);
+  const __m128i row1 = _mm_shufflelo_epi16(left_u16, 0x55);
+  const __m128i row2 = _mm_shufflelo_epi16(left_u16, 0xaa);
+  const __m128i row3 = _mm_shufflelo_epi16(left_u16, 0xff);
+  const __m128i row4 = _mm_shufflehi_epi16(left_u16, 0x0);
+  const __m128i row5 = _mm_shufflehi_epi16(left_u16, 0x55);
+  const __m128i row6 = _mm_shufflehi_epi16(left_u16, 0xaa);
+  const __m128i row7 = _mm_shufflehi_epi16(left_u16, 0xff);
+  h_store_16_unpacklo(&dst, stride, &row0);
+  h_store_16_unpacklo(&dst, stride, &row1);
+  h_store_16_unpacklo(&dst, stride, &row2);
+  h_store_16_unpacklo(&dst, stride, &row3);
+  h_store_16_unpackhi(&dst, stride, &row4);
+  h_store_16_unpackhi(&dst, stride, &row5);
+  h_store_16_unpackhi(&dst, stride, &row6);
+  h_store_16_unpackhi(&dst, stride, &row7);
+}
+
+void aom_highbd_h_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t stride,
+                                      const uint16_t *above,
+                                      const uint16_t *left, int bd) {
+  (void)above;
+  (void)bd;
+  h_predictor_16x8(dst, stride, left);
+}
+
+void aom_highbd_h_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t stride,
+                                       const uint16_t *above,
+                                       const uint16_t *left, int bd) {
+  int i;
+  (void)above;
+  (void)bd;
+
+  for (i = 0; i < 2; i++, left += 8) {
+    h_predictor_16x8(dst, stride, left);
+    dst += stride << 3;
+  }
+}
+
+void aom_highbd_h_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t stride,
+                                       const uint16_t *above,
+                                       const uint16_t *left, int bd) {
+  int i;
+  (void)above;
+  (void)bd;
+
+  for (i = 0; i < 4; i++, left += 8) {
+    h_predictor_16x8(dst, stride, left);
+    dst += stride << 3;
+  }
+}
+
+static INLINE void h_store_32_unpacklo(uint16_t **dst, const ptrdiff_t stride,
+                                       const __m128i *row) {
+  const __m128i val = _mm_unpacklo_epi64(*row, *row);
+  _mm_store_si128((__m128i *)(*dst), val);
+  _mm_store_si128((__m128i *)(*dst + 8), val);
+  _mm_store_si128((__m128i *)(*dst + 16), val);
+  _mm_store_si128((__m128i *)(*dst + 24), val);
+  *dst += stride;
+}
+
+static INLINE void h_store_32_unpackhi(uint16_t **dst, const ptrdiff_t stride,
+                                       const __m128i *row) {
+  const __m128i val = _mm_unpackhi_epi64(*row, *row);
+  _mm_store_si128((__m128i *)(*dst), val);
+  _mm_store_si128((__m128i *)(*dst + 8), val);
+  _mm_store_si128((__m128i *)(*dst + 16), val);
+  _mm_store_si128((__m128i *)(*dst + 24), val);
+  *dst += stride;
+}
+
+static INLINE void h_predictor_32x8(uint16_t *dst, ptrdiff_t stride,
+                                    const uint16_t *left) {
+  const __m128i left_u16 = _mm_load_si128((const __m128i *)left);
+  const __m128i row0 = _mm_shufflelo_epi16(left_u16, 0x0);
+  const __m128i row1 = _mm_shufflelo_epi16(left_u16, 0x55);
+  const __m128i row2 = _mm_shufflelo_epi16(left_u16, 0xaa);
+  const __m128i row3 = _mm_shufflelo_epi16(left_u16, 0xff);
+  const __m128i row4 = _mm_shufflehi_epi16(left_u16, 0x0);
+  const __m128i row5 = _mm_shufflehi_epi16(left_u16, 0x55);
+  const __m128i row6 = _mm_shufflehi_epi16(left_u16, 0xaa);
+  const __m128i row7 = _mm_shufflehi_epi16(left_u16, 0xff);
+  h_store_32_unpacklo(&dst, stride, &row0);
+  h_store_32_unpacklo(&dst, stride, &row1);
+  h_store_32_unpacklo(&dst, stride, &row2);
+  h_store_32_unpacklo(&dst, stride, &row3);
+  h_store_32_unpackhi(&dst, stride, &row4);
+  h_store_32_unpackhi(&dst, stride, &row5);
+  h_store_32_unpackhi(&dst, stride, &row6);
+  h_store_32_unpackhi(&dst, stride, &row7);
+}
+
+void aom_highbd_h_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t stride,
+                                       const uint16_t *above,
+                                       const uint16_t *left, int bd) {
+  int i;
+  (void)above;
+  (void)bd;
+
+  for (i = 0; i < 2; i++, left += 8) {
+    h_predictor_32x8(dst, stride, left);
+    dst += stride << 3;
+  }
+}
+
+void aom_highbd_h_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t stride,
+                                       const uint16_t *above,
+                                       const uint16_t *left, int bd) {
+  int i;
+  (void)above;
+  (void)bd;
+
+  for (i = 0; i < 4; i++, left += 8) {
+    h_predictor_32x8(dst, stride, left);
+    dst += stride << 3;
+  }
+}
+
+// -----------------------------------------------------------------------------
+// DC_TOP, DC_LEFT, DC_128
+
+// 4x4
+
+static INLINE __m128i dc_sum_4(const uint16_t *ref) {
+  const __m128i _dcba = _mm_loadl_epi64((const __m128i *)ref);
+  const __m128i _xxdc = _mm_shufflelo_epi16(_dcba, 0xe);
+  const __m128i a = _mm_add_epi16(_dcba, _xxdc);
+  return _mm_add_epi16(a, _mm_shufflelo_epi16(a, 0x1));
+}
+
+static INLINE void dc_store_4x4(uint16_t *dst, ptrdiff_t stride,
+                                const __m128i *dc) {
+  const __m128i dc_dup = _mm_shufflelo_epi16(*dc, 0x0);
+  int i;
+  for (i = 0; i < 4; ++i, dst += stride) {
+    _mm_storel_epi64((__m128i *)dst, dc_dup);
+  }
+}
+
+void aom_highbd_dc_left_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t stride,
+                                           const uint16_t *above,
+                                           const uint16_t *left, int bd) {
+  const __m128i two = _mm_cvtsi32_si128(2);
+  const __m128i sum = dc_sum_4(left);
+  const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, two), 2);
+  (void)above;
+  (void)bd;
+  dc_store_4x4(dst, stride, &dc);
+}
+
+void aom_highbd_dc_top_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t stride,
+                                          const uint16_t *above,
+                                          const uint16_t *left, int bd) {
+  const __m128i two = _mm_cvtsi32_si128(2);
+  const __m128i sum = dc_sum_4(above);
+  const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, two), 2);
+  (void)left;
+  (void)bd;
+  dc_store_4x4(dst, stride, &dc);
+}
+
+void aom_highbd_dc_128_predictor_4x4_sse2(uint16_t *dst, ptrdiff_t stride,
+                                          const uint16_t *above,
+                                          const uint16_t *left, int bd) {
+  const __m128i dc = _mm_cvtsi32_si128(1 << (bd - 1));
+  const __m128i dc_dup = _mm_shufflelo_epi16(dc, 0x0);
+  (void)above;
+  (void)left;
+  dc_store_4x4(dst, stride, &dc_dup);
+}
+
+// -----------------------------------------------------------------------------
+// 4x8
+
+static INLINE void dc_store_4x8(uint16_t *dst, ptrdiff_t stride,
+                                const __m128i *dc) {
+  const __m128i dc_dup = _mm_shufflelo_epi16(*dc, 0x0);
+  int i;
+  for (i = 0; i < 8; ++i, dst += stride) {
+    _mm_storel_epi64((__m128i *)dst, dc_dup);
+  }
+}
+
+// Shared with DC 8xh
+static INLINE __m128i dc_sum_8(const uint16_t *ref) {
+  const __m128i ref_u16 = _mm_load_si128((const __m128i *)ref);
+  const __m128i _dcba = _mm_add_epi16(ref_u16, _mm_srli_si128(ref_u16, 8));
+  const __m128i _xxdc = _mm_shufflelo_epi16(_dcba, 0xe);
+  const __m128i a = _mm_add_epi16(_dcba, _xxdc);
+
+  return _mm_add_epi16(a, _mm_shufflelo_epi16(a, 0x1));
+}
+
+void aom_highbd_dc_left_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t stride,
+                                           const uint16_t *above,
+                                           const uint16_t *left, int bd) {
+  const __m128i sum = dc_sum_8(left);
+  const __m128i four = _mm_cvtsi32_si128(4);
+  const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, four), 3);
+  (void)above;
+  (void)bd;
+  dc_store_4x8(dst, stride, &dc);
+}
+
+void aom_highbd_dc_top_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t stride,
+                                          const uint16_t *above,
+                                          const uint16_t *left, int bd) {
+  const __m128i two = _mm_cvtsi32_si128(2);
+  const __m128i sum = dc_sum_4(above);
+  const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, two), 2);
+  (void)left;
+  (void)bd;
+  dc_store_4x8(dst, stride, &dc);
+}
+
+void aom_highbd_dc_128_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t stride,
+                                          const uint16_t *above,
+                                          const uint16_t *left, int bd) {
+  const __m128i dc = _mm_cvtsi32_si128(1 << (bd - 1));
+  const __m128i dc_dup = _mm_shufflelo_epi16(dc, 0x0);
+  (void)above;
+  (void)left;
+  dc_store_4x8(dst, stride, &dc_dup);
+}
+
+// -----------------------------------------------------------------------------
+// 8xh
+
+static INLINE void dc_store_8xh(uint16_t *dst, ptrdiff_t stride, int height,
+                                const __m128i *dc) {
+  const __m128i dc_dup_lo = _mm_shufflelo_epi16(*dc, 0);
+  const __m128i dc_dup = _mm_unpacklo_epi64(dc_dup_lo, dc_dup_lo);
+  int i;
+  for (i = 0; i < height; ++i, dst += stride) {
+    _mm_store_si128((__m128i *)dst, dc_dup);
+  }
+}
+
+// -----------------------------------------------------------------------------
+// DC_TOP
+
+static INLINE void dc_top_predictor_8xh(uint16_t *dst, ptrdiff_t stride,
+                                        int height, const uint16_t *above) {
+  const __m128i four = _mm_cvtsi32_si128(4);
+  const __m128i sum = dc_sum_8(above);
+  const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, four), 3);
+  dc_store_8xh(dst, stride, height, &dc);
+}
+
+void aom_highbd_dc_top_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t stride,
+                                          const uint16_t *above,
+                                          const uint16_t *left, int bd) {
+  (void)left;
+  (void)bd;
+  dc_top_predictor_8xh(dst, stride, 4, above);
+}
+
+void aom_highbd_dc_top_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t stride,
+                                          const uint16_t *above,
+                                          const uint16_t *left, int bd) {
+  (void)left;
+  (void)bd;
+  dc_top_predictor_8xh(dst, stride, 8, above);
+}
+
+void aom_highbd_dc_top_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t stride,
+                                           const uint16_t *above,
+                                           const uint16_t *left, int bd) {
+  (void)left;
+  (void)bd;
+  dc_top_predictor_8xh(dst, stride, 16, above);
+}
+
+// -----------------------------------------------------------------------------
+// DC_LEFT
+
+void aom_highbd_dc_left_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t stride,
+                                           const uint16_t *above,
+                                           const uint16_t *left, int bd) {
+  const __m128i two = _mm_cvtsi32_si128(2);
+  const __m128i sum = dc_sum_4(left);
+  const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, two), 2);
+  (void)above;
+  (void)bd;
+  dc_store_8xh(dst, stride, 4, &dc);
+}
+
+void aom_highbd_dc_left_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t stride,
+                                           const uint16_t *above,
+                                           const uint16_t *left, int bd) {
+  const __m128i four = _mm_cvtsi32_si128(4);
+  const __m128i sum = dc_sum_8(left);
+  const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, four), 3);
+  (void)above;
+  (void)bd;
+  dc_store_8xh(dst, stride, 8, &dc);
+}
+
+// Shared with DC 16xh
+static INLINE __m128i dc_sum_16(const uint16_t *ref) {
+  const __m128i sum_lo = dc_sum_8(ref);
+  const __m128i sum_hi = dc_sum_8(ref + 8);
+  return _mm_add_epi16(sum_lo, sum_hi);
+}
+
+void aom_highbd_dc_left_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t stride,
+                                            const uint16_t *above,
+                                            const uint16_t *left, int bd) {
+  const __m128i eight = _mm_cvtsi32_si128(8);
+  const __m128i sum = dc_sum_16(left);
+  const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, eight), 4);
+  (void)above;
+  (void)bd;
+  dc_store_8xh(dst, stride, 16, &dc);
+}
+
+// -----------------------------------------------------------------------------
+// DC_128
+
+static INLINE void dc_128_predictor_8xh(uint16_t *dst, ptrdiff_t stride,
+                                        int height, int bd) {
+  const __m128i dc = _mm_cvtsi32_si128(1 << (bd - 1));
+  const __m128i dc_dup = _mm_shufflelo_epi16(dc, 0x0);
+  dc_store_8xh(dst, stride, height, &dc_dup);
+}
+
+void aom_highbd_dc_128_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t stride,
+                                          const uint16_t *above,
+                                          const uint16_t *left, int bd) {
+  (void)above;
+  (void)left;
+  dc_128_predictor_8xh(dst, stride, 4, bd);
+}
+
+void aom_highbd_dc_128_predictor_8x8_sse2(uint16_t *dst, ptrdiff_t stride,
+                                          const uint16_t *above,
+                                          const uint16_t *left, int bd) {
+  (void)above;
+  (void)left;
+  dc_128_predictor_8xh(dst, stride, 8, bd);
+}
+
+void aom_highbd_dc_128_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t stride,
+                                           const uint16_t *above,
+                                           const uint16_t *left, int bd) {
+  (void)above;
+  (void)left;
+  dc_128_predictor_8xh(dst, stride, 16, bd);
+}
+
+// -----------------------------------------------------------------------------
+// 16xh
+
+static INLINE void dc_store_16xh(uint16_t *dst, ptrdiff_t stride, int height,
+                                 const __m128i *dc) {
+  const __m128i dc_dup_lo = _mm_shufflelo_epi16(*dc, 0);
+  const __m128i dc_dup = _mm_unpacklo_epi64(dc_dup_lo, dc_dup_lo);
+  int i;
+  for (i = 0; i < height; ++i, dst += stride) {
+    _mm_store_si128((__m128i *)dst, dc_dup);
+    _mm_store_si128((__m128i *)(dst + 8), dc_dup);
+  }
+}
+
+// -----------------------------------------------------------------------------
+// DC_LEFT
+
+void aom_highbd_dc_left_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t stride,
+                                            const uint16_t *above,
+                                            const uint16_t *left, int bd) {
+  const __m128i four = _mm_cvtsi32_si128(4);
+  const __m128i sum = dc_sum_8(left);
+  const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, four), 3);
+  (void)above;
+  (void)bd;
+  dc_store_16xh(dst, stride, 8, &dc);
+}
+
+void aom_highbd_dc_left_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t stride,
+                                             const uint16_t *above,
+                                             const uint16_t *left, int bd) {
+  const __m128i eight = _mm_cvtsi32_si128(8);
+  const __m128i sum = dc_sum_16(left);
+  const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, eight), 4);
+  (void)above;
+  (void)bd;
+  dc_store_16xh(dst, stride, 16, &dc);
+}
+
+// Shared with 32xh
+static INLINE __m128i dc_sum_32(const uint16_t *ref) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i sum_a = dc_sum_16(ref);
+  const __m128i sum_b = dc_sum_16(ref + 16);
+  // 12 bit bd will outrange, so expand to 32 bit before adding final total
+  return _mm_add_epi32(_mm_unpacklo_epi16(sum_a, zero),
+                       _mm_unpacklo_epi16(sum_b, zero));
+}
+
+void aom_highbd_dc_left_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t stride,
+                                             const uint16_t *above,
+                                             const uint16_t *left, int bd) {
+  const __m128i sixteen = _mm_cvtsi32_si128(16);
+  const __m128i sum = dc_sum_32(left);
+  const __m128i dc = _mm_srli_epi32(_mm_add_epi32(sum, sixteen), 5);
+  (void)above;
+  (void)bd;
+  dc_store_16xh(dst, stride, 32, &dc);
+}
+
+// -----------------------------------------------------------------------------
+// DC_TOP
+
+void aom_highbd_dc_top_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t stride,
+                                           const uint16_t *above,
+                                           const uint16_t *left, int bd) {
+  const __m128i eight = _mm_cvtsi32_si128(8);
+  const __m128i sum = dc_sum_16(above);
+  const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, eight), 4);
+  (void)left;
+  (void)bd;
+  dc_store_16xh(dst, stride, 8, &dc);
+}
+
+void aom_highbd_dc_top_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t stride,
+                                            const uint16_t *above,
+                                            const uint16_t *left, int bd) {
+  const __m128i eight = _mm_cvtsi32_si128(8);
+  const __m128i sum = dc_sum_16(above);
+  const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, eight), 4);
+  (void)left;
+  (void)bd;
+  dc_store_16xh(dst, stride, 16, &dc);
+}
+
+void aom_highbd_dc_top_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t stride,
+                                            const uint16_t *above,
+                                            const uint16_t *left, int bd) {
+  const __m128i eight = _mm_cvtsi32_si128(8);
+  const __m128i sum = dc_sum_16(above);
+  const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, eight), 4);
+  (void)left;
+  (void)bd;
+  dc_store_16xh(dst, stride, 32, &dc);
+}
+
+// -----------------------------------------------------------------------------
+// DC_128
+
+void aom_highbd_dc_128_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t stride,
+                                           const uint16_t *above,
+                                           const uint16_t *left, int bd) {
+  const __m128i dc = _mm_cvtsi32_si128(1 << (bd - 1));
+  const __m128i dc_dup = _mm_shufflelo_epi16(dc, 0x0);
+  (void)above;
+  (void)left;
+  dc_store_16xh(dst, stride, 8, &dc_dup);
+}
+
+void aom_highbd_dc_128_predictor_16x16_sse2(uint16_t *dst, ptrdiff_t stride,
+                                            const uint16_t *above,
+                                            const uint16_t *left, int bd) {
+  const __m128i dc = _mm_cvtsi32_si128(1 << (bd - 1));
+  const __m128i dc_dup = _mm_shufflelo_epi16(dc, 0x0);
+  (void)above;
+  (void)left;
+  dc_store_16xh(dst, stride, 16, &dc_dup);
+}
+
+void aom_highbd_dc_128_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t stride,
+                                            const uint16_t *above,
+                                            const uint16_t *left, int bd) {
+  const __m128i dc = _mm_cvtsi32_si128(1 << (bd - 1));
+  const __m128i dc_dup = _mm_shufflelo_epi16(dc, 0x0);
+  (void)above;
+  (void)left;
+  dc_store_16xh(dst, stride, 32, &dc_dup);
+}
+
+// -----------------------------------------------------------------------------
+// 32xh
+
+static INLINE void dc_store_32xh(uint16_t *dst, ptrdiff_t stride, int height,
+                                 const __m128i *dc) {
+  const __m128i dc_dup_lo = _mm_shufflelo_epi16(*dc, 0);
+  const __m128i dc_dup = _mm_unpacklo_epi64(dc_dup_lo, dc_dup_lo);
+  int i;
+  for (i = 0; i < height; ++i, dst += stride) {
+    _mm_store_si128((__m128i *)dst, dc_dup);
+    _mm_store_si128((__m128i *)(dst + 8), dc_dup);
+    _mm_store_si128((__m128i *)(dst + 16), dc_dup);
+    _mm_store_si128((__m128i *)(dst + 24), dc_dup);
+  }
+}
+
+void aom_highbd_dc_left_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t stride,
+                                             const uint16_t *above,
+                                             const uint16_t *left, int bd) {
+  const __m128i eight = _mm_cvtsi32_si128(8);
+  const __m128i sum = dc_sum_16(left);
+  const __m128i dc = _mm_srli_epi16(_mm_add_epi16(sum, eight), 4);
+  (void)above;
+  (void)bd;
+  dc_store_32xh(dst, stride, 16, &dc);
+}
+
+void aom_highbd_dc_left_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t stride,
+                                             const uint16_t *above,
+                                             const uint16_t *left, int bd) {
+  const __m128i sixteen = _mm_cvtsi32_si128(16);
+  const __m128i sum = dc_sum_32(left);
+  const __m128i dc = _mm_srli_epi32(_mm_add_epi32(sum, sixteen), 5);
+  (void)above;
+  (void)bd;
+  dc_store_32xh(dst, stride, 32, &dc);
+}
+
+void aom_highbd_dc_top_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t stride,
+                                            const uint16_t *above,
+                                            const uint16_t *left, int bd) {
+  const __m128i sixteen = _mm_cvtsi32_si128(16);
+  const __m128i sum = dc_sum_32(above);
+  const __m128i dc = _mm_srli_epi32(_mm_add_epi32(sum, sixteen), 5);
+  (void)left;
+  (void)bd;
+  dc_store_32xh(dst, stride, 16, &dc);
+}
+
+void aom_highbd_dc_128_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t stride,
+                                            const uint16_t *above,
+                                            const uint16_t *left, int bd) {
+  const __m128i dc = _mm_cvtsi32_si128(1 << (bd - 1));
+  const __m128i dc_dup = _mm_shufflelo_epi16(dc, 0x0);
+  (void)above;
+  (void)left;
+  dc_store_32xh(dst, stride, 16, &dc_dup);
+}
+
+void aom_highbd_dc_top_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t stride,
+                                            const uint16_t *above,
+                                            const uint16_t *left, int bd) {
+  const __m128i sixteen = _mm_cvtsi32_si128(16);
+  const __m128i sum = dc_sum_32(above);
+  const __m128i dc = _mm_srli_epi32(_mm_add_epi32(sum, sixteen), 5);
+  (void)left;
+  (void)bd;
+  dc_store_32xh(dst, stride, 32, &dc);
+}
+
+void aom_highbd_dc_128_predictor_32x32_sse2(uint16_t *dst, ptrdiff_t stride,
+                                            const uint16_t *above,
+                                            const uint16_t *left, int bd) {
+  const __m128i dc = _mm_cvtsi32_si128(1 << (bd - 1));
+  const __m128i dc_dup = _mm_shufflelo_epi16(dc, 0x0);
+  (void)above;
+  (void)left;
+  dc_store_32xh(dst, stride, 32, &dc_dup);
+}
+
+// -----------------------------------------------------------------------------
+// V_PRED
+
+void aom_highbd_v_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t stride,
+                                     const uint16_t *above,
+                                     const uint16_t *left, int bd) {
+  (void)left;
+  (void)bd;
+  const __m128i above_u16 = _mm_loadl_epi64((const __m128i *)above);
+  int i;
+  for (i = 0; i < 2; ++i) {
+    _mm_storel_epi64((__m128i *)dst, above_u16);
+    _mm_storel_epi64((__m128i *)(dst + stride), above_u16);
+    _mm_storel_epi64((__m128i *)(dst + 2 * stride), above_u16);
+    _mm_storel_epi64((__m128i *)(dst + 3 * stride), above_u16);
+    dst += stride << 2;
+  }
+}
+
+void aom_highbd_v_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t stride,
+                                     const uint16_t *above,
+                                     const uint16_t *left, int bd) {
+  (void)left;
+  (void)bd;
+  const __m128i above_u16 = _mm_load_si128((const __m128i *)above);
+  _mm_store_si128((__m128i *)dst, above_u16);
+  _mm_store_si128((__m128i *)(dst + stride), above_u16);
+  _mm_store_si128((__m128i *)(dst + 2 * stride), above_u16);
+  _mm_store_si128((__m128i *)(dst + 3 * stride), above_u16);
+}
+
+void aom_highbd_v_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t stride,
+                                      const uint16_t *above,
+                                      const uint16_t *left, int bd) {
+  (void)left;
+  (void)bd;
+  const __m128i above_u16 = _mm_load_si128((const __m128i *)above);
+  int i;
+  for (i = 0; i < 4; ++i) {
+    _mm_store_si128((__m128i *)dst, above_u16);
+    _mm_store_si128((__m128i *)(dst + stride), above_u16);
+    _mm_store_si128((__m128i *)(dst + 2 * stride), above_u16);
+    _mm_store_si128((__m128i *)(dst + 3 * stride), above_u16);
+    dst += stride << 2;
+  }
+}
+
+void aom_highbd_v_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t stride,
+                                      const uint16_t *above,
+                                      const uint16_t *left, int bd) {
+  (void)left;
+  (void)bd;
+  const __m128i above0_u16 = _mm_load_si128((const __m128i *)above);
+  const __m128i above1_u16 = _mm_load_si128((const __m128i *)(above + 8));
+  int i;
+  for (i = 0; i < 2; ++i) {
+    _mm_store_si128((__m128i *)dst, above0_u16);
+    _mm_store_si128((__m128i *)(dst + 8), above1_u16);
+    dst += stride;
+    _mm_store_si128((__m128i *)dst, above0_u16);
+    _mm_store_si128((__m128i *)(dst + 8), above1_u16);
+    dst += stride;
+    _mm_store_si128((__m128i *)dst, above0_u16);
+    _mm_store_si128((__m128i *)(dst + 8), above1_u16);
+    dst += stride;
+    _mm_store_si128((__m128i *)dst, above0_u16);
+    _mm_store_si128((__m128i *)(dst + 8), above1_u16);
+    dst += stride;
+  }
+}
+
+void aom_highbd_v_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t stride,
+                                       const uint16_t *above,
+                                       const uint16_t *left, int bd) {
+  (void)left;
+  (void)bd;
+  const __m128i above0_u16 = _mm_load_si128((const __m128i *)above);
+  const __m128i above1_u16 = _mm_load_si128((const __m128i *)(above + 8));
+  int i;
+  for (i = 0; i < 8; ++i) {
+    _mm_store_si128((__m128i *)dst, above0_u16);
+    _mm_store_si128((__m128i *)(dst + 8), above1_u16);
+    dst += stride;
+    _mm_store_si128((__m128i *)dst, above0_u16);
+    _mm_store_si128((__m128i *)(dst + 8), above1_u16);
+    dst += stride;
+    _mm_store_si128((__m128i *)dst, above0_u16);
+    _mm_store_si128((__m128i *)(dst + 8), above1_u16);
+    dst += stride;
+    _mm_store_si128((__m128i *)dst, above0_u16);
+    _mm_store_si128((__m128i *)(dst + 8), above1_u16);
+    dst += stride;
+  }
+}
+
+void aom_highbd_v_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t stride,
+                                       const uint16_t *above,
+                                       const uint16_t *left, int bd) {
+  (void)left;
+  (void)bd;
+  const __m128i above0_u16 = _mm_load_si128((const __m128i *)above);
+  const __m128i above1_u16 = _mm_load_si128((const __m128i *)(above + 8));
+  const __m128i above2_u16 = _mm_load_si128((const __m128i *)(above + 16));
+  const __m128i above3_u16 = _mm_load_si128((const __m128i *)(above + 24));
+  int i;
+  for (i = 0; i < 4; ++i) {
+    _mm_store_si128((__m128i *)dst, above0_u16);
+    _mm_store_si128((__m128i *)(dst + 8), above1_u16);
+    _mm_store_si128((__m128i *)(dst + 16), above2_u16);
+    _mm_store_si128((__m128i *)(dst + 24), above3_u16);
+    dst += stride;
+    _mm_store_si128((__m128i *)dst, above0_u16);
+    _mm_store_si128((__m128i *)(dst + 8), above1_u16);
+    _mm_store_si128((__m128i *)(dst + 16), above2_u16);
+    _mm_store_si128((__m128i *)(dst + 24), above3_u16);
+    dst += stride;
+    _mm_store_si128((__m128i *)dst, above0_u16);
+    _mm_store_si128((__m128i *)(dst + 8), above1_u16);
+    _mm_store_si128((__m128i *)(dst + 16), above2_u16);
+    _mm_store_si128((__m128i *)(dst + 24), above3_u16);
+    dst += stride;
+    _mm_store_si128((__m128i *)dst, above0_u16);
+    _mm_store_si128((__m128i *)(dst + 8), above1_u16);
+    _mm_store_si128((__m128i *)(dst + 16), above2_u16);
+    _mm_store_si128((__m128i *)(dst + 24), above3_u16);
+    dst += stride;
+  }
+}
+
+// -----------------------------------------------------------------------------
+// DC_PRED
+
+void aom_highbd_dc_predictor_4x8_sse2(uint16_t *dst, ptrdiff_t stride,
+                                      const uint16_t *above,
+                                      const uint16_t *left, int bd) {
+  (void)bd;
+  const __m128i sum_above = dc_sum_4(above);
+  const __m128i sum_left = dc_sum_8(left);
+  const __m128i sum = _mm_add_epi16(sum_above, sum_left);
+  uint32_t sum32 = (uint32_t)_mm_cvtsi128_si32(sum);
+  sum32 >>= 16;
+  sum32 += 6;
+  sum32 /= 12;
+  const __m128i row = _mm_set1_epi16((int16_t)sum32);
+  int i;
+  for (i = 0; i < 4; ++i) {
+    _mm_storel_epi64((__m128i *)dst, row);
+    dst += stride;
+    _mm_storel_epi64((__m128i *)dst, row);
+    dst += stride;
+  }
+}
+
+void aom_highbd_dc_predictor_8x4_sse2(uint16_t *dst, ptrdiff_t stride,
+                                      const uint16_t *above,
+                                      const uint16_t *left, int bd) {
+  (void)bd;
+  const __m128i sum_left = dc_sum_4(left);
+  const __m128i sum_above = dc_sum_8(above);
+  const __m128i sum = _mm_add_epi16(sum_above, sum_left);
+  uint32_t sum32 = (uint32_t)_mm_cvtsi128_si32(sum);
+  sum32 >>= 16;
+  sum32 += 6;
+  sum32 /= 12;
+  const __m128i row = _mm_set1_epi16((int16_t)sum32);
+
+  _mm_store_si128((__m128i *)dst, row);
+  dst += stride;
+  _mm_store_si128((__m128i *)dst, row);
+  dst += stride;
+  _mm_store_si128((__m128i *)dst, row);
+  dst += stride;
+  _mm_store_si128((__m128i *)dst, row);
+}
+
+void aom_highbd_dc_predictor_8x16_sse2(uint16_t *dst, ptrdiff_t stride,
+                                       const uint16_t *above,
+                                       const uint16_t *left, int bd) {
+  (void)bd;
+  __m128i sum_left = dc_sum_16(left);
+  __m128i sum_above = dc_sum_8(above);
+  const __m128i zero = _mm_setzero_si128();
+  sum_left = _mm_unpacklo_epi16(sum_left, zero);
+  sum_above = _mm_unpacklo_epi16(sum_above, zero);
+  const __m128i sum = _mm_add_epi32(sum_left, sum_above);
+  uint32_t sum32 = (uint32_t)_mm_cvtsi128_si32(sum);
+  sum32 += 12;
+  sum32 /= 24;
+  const __m128i row = _mm_set1_epi16((int16_t)sum32);
+  int i;
+  for (i = 0; i < 4; ++i) {
+    _mm_store_si128((__m128i *)dst, row);
+    dst += stride;
+    _mm_store_si128((__m128i *)dst, row);
+    dst += stride;
+    _mm_store_si128((__m128i *)dst, row);
+    dst += stride;
+    _mm_store_si128((__m128i *)dst, row);
+    dst += stride;
+  }
+}
+
+void aom_highbd_dc_predictor_16x8_sse2(uint16_t *dst, ptrdiff_t stride,
+                                       const uint16_t *above,
+                                       const uint16_t *left, int bd) {
+  (void)bd;
+  __m128i sum_left = dc_sum_8(left);
+  __m128i sum_above = dc_sum_16(above);
+  const __m128i zero = _mm_setzero_si128();
+  sum_left = _mm_unpacklo_epi16(sum_left, zero);
+  sum_above = _mm_unpacklo_epi16(sum_above, zero);
+  const __m128i sum = _mm_add_epi32(sum_left, sum_above);
+  uint32_t sum32 = (uint32_t)_mm_cvtsi128_si32(sum);
+  sum32 += 12;
+  sum32 /= 24;
+  const __m128i row = _mm_set1_epi16((int16_t)sum32);
+  int i;
+  for (i = 0; i < 2; ++i) {
+    _mm_store_si128((__m128i *)dst, row);
+    _mm_store_si128((__m128i *)(dst + 8), row);
+    dst += stride;
+    _mm_store_si128((__m128i *)dst, row);
+    _mm_store_si128((__m128i *)(dst + 8), row);
+    dst += stride;
+    _mm_store_si128((__m128i *)dst, row);
+    _mm_store_si128((__m128i *)(dst + 8), row);
+    dst += stride;
+    _mm_store_si128((__m128i *)dst, row);
+    _mm_store_si128((__m128i *)(dst + 8), row);
+    dst += stride;
+  }
+}
+
+void aom_highbd_dc_predictor_16x32_sse2(uint16_t *dst, ptrdiff_t stride,
+                                        const uint16_t *above,
+                                        const uint16_t *left, int bd) {
+  (void)bd;
+  __m128i sum_left = dc_sum_32(left);
+  __m128i sum_above = dc_sum_16(above);
+  const __m128i zero = _mm_setzero_si128();
+  sum_above = _mm_unpacklo_epi16(sum_above, zero);
+  const __m128i sum = _mm_add_epi32(sum_left, sum_above);
+  uint32_t sum32 = (uint32_t)_mm_cvtsi128_si32(sum);
+  sum32 += 24;
+  sum32 /= 48;
+  const __m128i row = _mm_set1_epi16((int16_t)sum32);
+  int i;
+  for (i = 0; i < 8; ++i) {
+    _mm_store_si128((__m128i *)dst, row);
+    _mm_store_si128((__m128i *)(dst + 8), row);
+    dst += stride;
+    _mm_store_si128((__m128i *)dst, row);
+    _mm_store_si128((__m128i *)(dst + 8), row);
+    dst += stride;
+    _mm_store_si128((__m128i *)dst, row);
+    _mm_store_si128((__m128i *)(dst + 8), row);
+    dst += stride;
+    _mm_store_si128((__m128i *)dst, row);
+    _mm_store_si128((__m128i *)(dst + 8), row);
+    dst += stride;
+  }
+}
+
+void aom_highbd_dc_predictor_32x16_sse2(uint16_t *dst, ptrdiff_t stride,
+                                        const uint16_t *above,
+                                        const uint16_t *left, int bd) {
+  (void)bd;
+  __m128i sum_left = dc_sum_16(left);
+  __m128i sum_above = dc_sum_32(above);
+  const __m128i zero = _mm_setzero_si128();
+  sum_left = _mm_unpacklo_epi16(sum_left, zero);
+  const __m128i sum = _mm_add_epi32(sum_left, sum_above);
+  uint32_t sum32 = (uint32_t)_mm_cvtsi128_si32(sum);
+  sum32 += 24;
+  sum32 /= 48;
+  const __m128i row = _mm_set1_epi16((int16_t)sum32);
+  int i;
+  for (i = 0; i < 4; ++i) {
+    _mm_store_si128((__m128i *)dst, row);
+    _mm_store_si128((__m128i *)(dst + 8), row);
+    _mm_store_si128((__m128i *)(dst + 16), row);
+    _mm_store_si128((__m128i *)(dst + 24), row);
+    dst += stride;
+    _mm_store_si128((__m128i *)dst, row);
+    _mm_store_si128((__m128i *)(dst + 8), row);
+    _mm_store_si128((__m128i *)(dst + 16), row);
+    _mm_store_si128((__m128i *)(dst + 24), row);
+    dst += stride;
+    _mm_store_si128((__m128i *)dst, row);
+    _mm_store_si128((__m128i *)(dst + 8), row);
+    _mm_store_si128((__m128i *)(dst + 16), row);
+    _mm_store_si128((__m128i *)(dst + 24), row);
+    dst += stride;
+    _mm_store_si128((__m128i *)dst, row);
+    _mm_store_si128((__m128i *)(dst + 8), row);
+    _mm_store_si128((__m128i *)(dst + 16), row);
+    _mm_store_si128((__m128i *)(dst + 24), row);
+    dst += stride;
+  }
+}
diff --git a/third_party/aom/aom_dsp/x86/highbd_loopfilter_avx2.c b/third_party/aom/aom_dsp/x86/highbd_loopfilter_avx2.c
new file mode 100644
index 0000000000..c954da94e5
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/highbd_loopfilter_avx2.c
@@ -0,0 +1,66 @@
+/*
+ * Copyright (c) 2017, 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"
+
+#include "aom_dsp/x86/common_avx2.h"
+#include "aom_dsp/x86/lpf_common_sse2.h"
+#include "aom/aom_integer.h"
+
+void aom_highbd_lpf_horizontal_14_dual_avx2(
+    uint16_t *s, int p, const uint8_t *blimit0, const uint8_t *limit0,
+    const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1,
+    const uint8_t *thresh1, int bd) {
+  aom_highbd_lpf_horizontal_14_dual_sse2(s, p, blimit0, limit0, thresh0,
+                                         blimit1, limit1, thresh1, bd);
+}
+
+void aom_highbd_lpf_vertical_14_dual_avx2(
+    uint16_t *s, int p, const uint8_t *blimit0, const uint8_t *limit0,
+    const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1,
+    const uint8_t *thresh1, int bd) {
+  aom_highbd_lpf_vertical_14_dual_sse2(s, p, blimit0, limit0, thresh0, blimit1,
+                                       limit1, thresh1, bd);
+}
+
+void aom_highbd_lpf_horizontal_4_dual_avx2(
+    uint16_t *s, int p, const uint8_t *blimit0, const uint8_t *limit0,
+    const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1,
+    const uint8_t *thresh1, int bd) {
+  aom_highbd_lpf_horizontal_4_dual_sse2(s, p, blimit0, limit0, thresh0, blimit1,
+                                        limit1, thresh1, bd);
+}
+
+void aom_highbd_lpf_horizontal_8_dual_avx2(
+    uint16_t *s, int p, const uint8_t *blimit0, const uint8_t *limit0,
+    const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1,
+    const uint8_t *thresh1, int bd) {
+  aom_highbd_lpf_horizontal_8_dual_sse2(s, p, blimit0, limit0, thresh0, blimit1,
+                                        limit1, thresh1, bd);
+}
+
+void aom_highbd_lpf_vertical_4_dual_avx2(
+    uint16_t *s, int p, const uint8_t *blimit0, const uint8_t *limit0,
+    const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1,
+    const uint8_t *thresh1, int bd) {
+  aom_highbd_lpf_vertical_4_dual_sse2(s, p, blimit0, limit0, thresh0, blimit1,
+                                      limit1, thresh1, bd);
+}
+
+void aom_highbd_lpf_vertical_8_dual_avx2(
+    uint16_t *s, int p, const uint8_t *blimit0, const uint8_t *limit0,
+    const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1,
+    const uint8_t *thresh1, int bd) {
+  aom_highbd_lpf_vertical_8_dual_sse2(s, p, blimit0, limit0, thresh0, blimit1,
+                                      limit1, thresh1, bd);
+}
diff --git a/third_party/aom/aom_dsp/x86/highbd_loopfilter_sse2.c b/third_party/aom/aom_dsp/x86/highbd_loopfilter_sse2.c
new file mode 100644
index 0000000000..ea7dc6a9e5
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/highbd_loopfilter_sse2.c
@@ -0,0 +1,1698 @@
+/*
+ * Copyright (c) 2016, 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 <emmintrin.h>  // SSE2
+
+#include "config/aom_dsp_rtcd.h"
+
+#include "aom_dsp/x86/lpf_common_sse2.h"
+
+static AOM_FORCE_INLINE void pixel_clamp(const __m128i *min, const __m128i *max,
+                                         __m128i *pixel) {
+  *pixel = _mm_min_epi16(*pixel, *max);
+  *pixel = _mm_max_epi16(*pixel, *min);
+}
+
+static AOM_FORCE_INLINE __m128i abs_diff16(__m128i a, __m128i b) {
+  return _mm_or_si128(_mm_subs_epu16(a, b), _mm_subs_epu16(b, a));
+}
+
+static INLINE void get_limit(const uint8_t *bl, const uint8_t *l,
+                             const uint8_t *t, int bd, __m128i *blt,
+                             __m128i *lt, __m128i *thr, __m128i *t80_out) {
+  const int shift = bd - 8;
+  const __m128i zero = _mm_setzero_si128();
+
+  __m128i x = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)bl), zero);
+  *blt = _mm_slli_epi16(x, shift);
+
+  x = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)l), zero);
+  *lt = _mm_slli_epi16(x, shift);
+
+  x = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)t), zero);
+  *thr = _mm_slli_epi16(x, shift);
+
+  *t80_out = _mm_set1_epi16(1 << (bd - 1));
+}
+
+static INLINE void get_limit_dual(
+    const uint8_t *_blimit0, const uint8_t *_limit0, const uint8_t *_thresh0,
+    const uint8_t *_blimit1, const uint8_t *_limit1, const uint8_t *_thresh1,
+    int bd, __m128i *blt_out, __m128i *lt_out, __m128i *thr_out,
+    __m128i *t80_out) {
+  const int shift = bd - 8;
+  const __m128i zero = _mm_setzero_si128();
+
+  __m128i x0 =
+      _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit0), zero);
+  __m128i x1 =
+      _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_blimit1), zero);
+  x0 = _mm_unpacklo_epi64(x0, x1);
+  *blt_out = _mm_slli_epi16(x0, shift);
+
+  x0 = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit0), zero);
+  x1 = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_limit1), zero);
+  x0 = _mm_unpacklo_epi64(x0, x1);
+  *lt_out = _mm_slli_epi16(x0, shift);
+
+  x0 = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh0), zero);
+  x1 = _mm_unpacklo_epi8(_mm_load_si128((const __m128i *)_thresh1), zero);
+  x0 = _mm_unpacklo_epi64(x0, x1);
+  *thr_out = _mm_slli_epi16(x0, shift);
+
+  *t80_out = _mm_set1_epi16(1 << (bd - 1));
+}
+
+static INLINE void load_highbd_pixel(const uint16_t *s, int size, int pitch,
+                                     __m128i *p, __m128i *q) {
+  int i;
+  for (i = 0; i < size; i++) {
+    p[i] = _mm_loadu_si128((__m128i *)(s - (i + 1) * pitch));
+    q[i] = _mm_loadu_si128((__m128i *)(s + i * pitch));
+  }
+}
+
+static INLINE void highbd_filter_mask_dual(const __m128i *p, const __m128i *q,
+                                           const __m128i *l, const __m128i *bl,
+                                           __m128i *mask) {
+  __m128i abs_p0q0 = abs_diff16(p[0], q[0]);
+  __m128i abs_p1q1 = abs_diff16(p[1], q[1]);
+  abs_p0q0 = _mm_adds_epu16(abs_p0q0, abs_p0q0);
+  abs_p1q1 = _mm_srli_epi16(abs_p1q1, 1);
+
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i one = _mm_set1_epi16(1);
+  const __m128i ffff = _mm_set1_epi16((short)0xFFFF);
+
+  __m128i max = _mm_subs_epu16(_mm_adds_epu16(abs_p0q0, abs_p1q1), *bl);
+  max = _mm_xor_si128(_mm_cmpeq_epi16(max, zero), ffff);
+  max = _mm_and_si128(max, _mm_adds_epu16(*l, one));
+
+  int i;
+  for (i = 1; i < 4; ++i) {
+    max = _mm_max_epi16(max, abs_diff16(p[i], p[i - 1]));
+    max = _mm_max_epi16(max, abs_diff16(q[i], q[i - 1]));
+  }
+  max = _mm_subs_epu16(max, *l);
+  *mask = _mm_cmpeq_epi16(max, zero);  // return ~mask
+}
+
+static INLINE void highbd_hev_filter_mask_x_sse2(__m128i *pq, int x,
+                                                 __m128i *p1p0, __m128i *q1q0,
+                                                 __m128i *abs_p1p0, __m128i *l,
+                                                 __m128i *bl, __m128i *t,
+                                                 __m128i *hev, __m128i *mask) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i one = _mm_set1_epi16(1);
+  const __m128i ffff = _mm_set1_epi16((short)0xFFFF);
+  __m128i abs_p0q0_p1q1, abs_p0q0, abs_p1q1, abs_q1q0;
+  __m128i max, max01, h;
+
+  *p1p0 = _mm_unpacklo_epi64(pq[0], pq[1]);
+  *q1q0 = _mm_unpackhi_epi64(pq[0], pq[1]);
+
+  abs_p0q0_p1q1 = abs_diff16(*p1p0, *q1q0);
+  abs_p0q0 = _mm_adds_epu16(abs_p0q0_p1q1, abs_p0q0_p1q1);
+  abs_p0q0 = _mm_unpacklo_epi64(abs_p0q0, zero);
+
+  abs_p1q1 = _mm_srli_si128(abs_p0q0_p1q1, 8);
+  abs_p1q1 = _mm_srli_epi16(abs_p1q1, 1);  // divide by 2
+
+  max = _mm_subs_epu16(_mm_adds_epu16(abs_p0q0, abs_p1q1), *bl);
+  max = _mm_xor_si128(_mm_cmpeq_epi16(max, zero), ffff);
+  // mask |= (abs(*p0 - *q0) * 2 + abs(*p1 - *q1) / 2  > blimit) * -1;
+  // So taking maximums continues to work:
+  max = _mm_and_si128(max, _mm_adds_epu16(*l, one));
+
+  *abs_p1p0 = abs_diff16(pq[0], pq[1]);
+  abs_q1q0 = _mm_srli_si128(*abs_p1p0, 8);
+  max01 = _mm_max_epi16(*abs_p1p0, abs_q1q0);
+  // mask |= (abs(*p1 - *p0) > limit) * -1;
+  // mask |= (abs(*q1 - *q0) > limit) * -1;
+  h = _mm_subs_epu16(max01, *t);
+
+  *hev = _mm_xor_si128(_mm_cmpeq_epi16(h, zero), ffff);
+  // replicate for the further "merged variables" usage
+  *hev = _mm_unpacklo_epi64(*hev, *hev);
+
+  max = _mm_max_epi16(max, max01);
+  int i;
+  for (i = 2; i < x; ++i) {
+    max = _mm_max_epi16(max, abs_diff16(pq[i], pq[i - 1]));
+  }
+  max = _mm_max_epi16(max, _mm_srli_si128(max, 8));
+
+  max = _mm_subs_epu16(max, *l);
+  *mask = _mm_cmpeq_epi16(max, zero);  //  ~mask
+}
+
+static INLINE void flat_mask_internal(const __m128i *th, const __m128i *pq,
+                                      int start, int end, __m128i *flat) {
+  int i;
+  __m128i max = _mm_max_epi16(abs_diff16(pq[start], pq[0]),
+                              abs_diff16(pq[start + 1], pq[0]));
+
+  for (i = start + 2; i < end; ++i) {
+    max = _mm_max_epi16(max, abs_diff16(pq[i], pq[0]));
+  }
+  max = _mm_max_epi16(max, _mm_srli_si128(max, 8));
+
+  __m128i ft;
+  ft = _mm_subs_epu16(max, *th);
+
+  const __m128i zero = _mm_setzero_si128();
+  *flat = _mm_cmpeq_epi16(ft, zero);
+}
+
+static INLINE void flat_mask_internal_dual(const __m128i *th, const __m128i *p,
+                                           const __m128i *q, int start, int end,
+                                           __m128i *flat) {
+  int i;
+  __m128i max =
+      _mm_max_epi16(abs_diff16(q[start], q[0]), abs_diff16(p[start], p[0]));
+
+  for (i = start + 1; i < end; ++i) {
+    max = _mm_max_epi16(max, abs_diff16(p[i], p[0]));
+    max = _mm_max_epi16(max, abs_diff16(q[i], q[0]));
+  }
+
+  __m128i ft;
+  ft = _mm_subs_epu16(max, *th);
+
+  const __m128i zero = _mm_setzero_si128();
+  *flat = _mm_cmpeq_epi16(ft, zero);
+}
+
+static INLINE void highbd_flat_mask4_sse2(__m128i *pq, __m128i *flat,
+                                          __m128i *flat2, int bd) {
+  // check the distance 1,2,3 against 0
+  __m128i th = _mm_set1_epi16(1);
+  th = _mm_slli_epi16(th, bd - 8);
+  flat_mask_internal(&th, pq, 1, 4, flat);
+  flat_mask_internal(&th, pq, 4, 7, flat2);
+}
+
+static INLINE void highbd_flat_mask4_dual_sse2(const __m128i *p,
+                                               const __m128i *q, __m128i *flat,
+                                               __m128i *flat2, int bd) {
+  // check the distance 1,2,3 against 0
+  __m128i th = _mm_set1_epi16(1);
+  th = _mm_slli_epi16(th, bd - 8);
+  flat_mask_internal_dual(&th, p, q, 1, 4, flat);
+  flat_mask_internal_dual(&th, p, q, 4, 7, flat2);
+}
+
+static AOM_FORCE_INLINE void highbd_filter4_sse2(__m128i *p1p0, __m128i *q1q0,
+                                                 __m128i *hev, __m128i *mask,
+                                                 __m128i *qs1qs0,
+                                                 __m128i *ps1ps0, __m128i *t80,
+                                                 int bd) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i one = _mm_set1_epi16(1);
+  const __m128i pmax =
+      _mm_subs_epi16(_mm_subs_epi16(_mm_slli_epi16(one, bd), one), *t80);
+  const __m128i pmin = _mm_subs_epi16(zero, *t80);
+
+  const __m128i t3t4 = _mm_set_epi16(3, 3, 3, 3, 4, 4, 4, 4);
+  __m128i ps1ps0_work, qs1qs0_work, work;
+  __m128i filt, filter2filter1, filter2filt, filter1filt;
+
+  ps1ps0_work = _mm_subs_epi16(*p1p0, *t80);
+  qs1qs0_work = _mm_subs_epi16(*q1q0, *t80);
+
+  work = _mm_subs_epi16(ps1ps0_work, qs1qs0_work);
+  pixel_clamp(&pmin, &pmax, &work);
+  filt = _mm_and_si128(_mm_srli_si128(work, 8), *hev);
+
+  filt = _mm_subs_epi16(filt, work);
+  filt = _mm_subs_epi16(filt, work);
+  filt = _mm_subs_epi16(filt, work);
+  // (aom_filter + 3 * (qs0 - ps0)) & mask
+  pixel_clamp(&pmin, &pmax, &filt);
+  filt = _mm_and_si128(filt, *mask);
+  filt = _mm_unpacklo_epi64(filt, filt);
+
+  filter2filter1 = _mm_adds_epi16(filt, t3t4); /* signed_short_clamp */
+  pixel_clamp(&pmin, &pmax, &filter2filter1);
+  filter2filter1 = _mm_srai_epi16(filter2filter1, 3); /* >> 3 */
+
+  filt = _mm_unpacklo_epi64(filter2filter1, filter2filter1);
+
+  // filt >> 1
+  filt = _mm_adds_epi16(filt, one);
+  filt = _mm_srai_epi16(filt, 1);
+  filt = _mm_andnot_si128(*hev, filt);
+
+  filter2filt = _mm_unpackhi_epi64(filter2filter1, filt);
+  filter1filt = _mm_unpacklo_epi64(filter2filter1, filt);
+
+  qs1qs0_work = _mm_subs_epi16(qs1qs0_work, filter1filt);
+  ps1ps0_work = _mm_adds_epi16(ps1ps0_work, filter2filt);
+
+  pixel_clamp(&pmin, &pmax, &qs1qs0_work);
+  pixel_clamp(&pmin, &pmax, &ps1ps0_work);
+
+  *qs1qs0 = _mm_adds_epi16(qs1qs0_work, *t80);
+  *ps1ps0 = _mm_adds_epi16(ps1ps0_work, *t80);
+}
+
+static INLINE void highbd_filter4_dual_sse2(__m128i *p, __m128i *q, __m128i *ps,
+                                            __m128i *qs, const __m128i *mask,
+                                            const __m128i *th, int bd,
+                                            __m128i *t80) {
+  __m128i ps0 = _mm_subs_epi16(p[0], *t80);
+  __m128i ps1 = _mm_subs_epi16(p[1], *t80);
+  __m128i qs0 = _mm_subs_epi16(q[0], *t80);
+  __m128i qs1 = _mm_subs_epi16(q[1], *t80);
+  const __m128i one = _mm_set1_epi16(1);
+  const __m128i pmax =
+      _mm_subs_epi16(_mm_subs_epi16(_mm_slli_epi16(one, bd), one), *t80);
+
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i pmin = _mm_subs_epi16(zero, *t80);
+  __m128i filter = _mm_subs_epi16(ps1, qs1);
+  pixel_clamp(&pmin, &pmax, &filter);
+
+  // hev_filter
+  __m128i hev;
+  const __m128i abs_p1p0 = abs_diff16(p[1], p[0]);
+  const __m128i abs_q1q0 = abs_diff16(q[1], q[0]);
+  __m128i h = _mm_max_epi16(abs_p1p0, abs_q1q0);
+  h = _mm_subs_epu16(h, *th);
+  const __m128i ffff = _mm_cmpeq_epi16(h, h);
+  hev = _mm_xor_si128(_mm_cmpeq_epi16(h, zero), ffff);
+
+  filter = _mm_and_si128(filter, hev);
+
+  const __m128i x = _mm_subs_epi16(qs0, ps0);
+  filter = _mm_adds_epi16(filter, x);
+  filter = _mm_adds_epi16(filter, x);
+  filter = _mm_adds_epi16(filter, x);
+  pixel_clamp(&pmin, &pmax, &filter);
+  filter = _mm_and_si128(filter, *mask);
+  const __m128i t3 = _mm_set1_epi16(3);
+  const __m128i t4 = _mm_set1_epi16(4);
+  __m128i filter1 = _mm_adds_epi16(filter, t4);
+  __m128i filter2 = _mm_adds_epi16(filter, t3);
+  pixel_clamp(&pmin, &pmax, &filter1);
+  pixel_clamp(&pmin, &pmax, &filter2);
+  filter1 = _mm_srai_epi16(filter1, 3);
+  filter2 = _mm_srai_epi16(filter2, 3);
+  qs0 = _mm_subs_epi16(qs0, filter1);
+  pixel_clamp(&pmin, &pmax, &qs0);
+  ps0 = _mm_adds_epi16(ps0, filter2);
+  pixel_clamp(&pmin, &pmax, &ps0);
+  qs[0] = _mm_adds_epi16(qs0, *t80);
+  ps[0] = _mm_adds_epi16(ps0, *t80);
+  filter = _mm_adds_epi16(filter1, one);
+  filter = _mm_srai_epi16(filter, 1);
+  filter = _mm_andnot_si128(hev, filter);
+  qs1 = _mm_subs_epi16(qs1, filter);
+  pixel_clamp(&pmin, &pmax, &qs1);
+  ps1 = _mm_adds_epi16(ps1, filter);
+  pixel_clamp(&pmin, &pmax, &ps1);
+  qs[1] = _mm_adds_epi16(qs1, *t80);
+  ps[1] = _mm_adds_epi16(ps1, *t80);
+}
+
+static AOM_FORCE_INLINE void highbd_lpf_internal_14_sse2(
+    __m128i *p, __m128i *q, __m128i *pq, const unsigned char *blt,
+    const unsigned char *lt, const unsigned char *thr, int bd) {
+  int i;
+  const __m128i zero = _mm_setzero_si128();
+  __m128i blimit, limit, thresh;
+  __m128i t80;
+  get_limit(blt, lt, thr, bd, &blimit, &limit, &thresh, &t80);
+
+  for (i = 0; i < 7; i++) {
+    pq[i] = _mm_unpacklo_epi64(p[i], q[i]);
+  }
+  __m128i mask, hevhev;
+  __m128i p1p0, q1q0, abs_p1p0;
+
+  highbd_hev_filter_mask_x_sse2(pq, 4, &p1p0, &q1q0, &abs_p1p0, &limit, &blimit,
+                                &thresh, &hevhev, &mask);
+
+  __m128i ps0ps1, qs0qs1;
+  // filter4
+  highbd_filter4_sse2(&p1p0, &q1q0, &hevhev, &mask, &qs0qs1, &ps0ps1, &t80, bd);
+
+  __m128i flat, flat2;
+  highbd_flat_mask4_sse2(pq, &flat, &flat2, bd);
+
+  flat = _mm_and_si128(flat, mask);
+  flat2 = _mm_and_si128(flat2, flat);
+
+  // replicate for the further "merged variables" usage
+  flat = _mm_unpacklo_epi64(flat, flat);
+  flat2 = _mm_unpacklo_epi64(flat2, flat2);
+
+  // flat and wide flat calculations
+
+  // if flat ==0 then flat2 is zero as well and we don't need any calc below
+  // sse4.1 if (0==_mm_test_all_zeros(flat,ff))
+  if (0xffff != _mm_movemask_epi8(_mm_cmpeq_epi16(flat, zero))) {
+    __m128i flat_p[3], flat_q[3], flat_pq[3];
+    __m128i flat2_p[6], flat2_q[6];
+    __m128i flat2_pq[6];
+    __m128i sum_p6, sum_p3;
+    const __m128i eight = _mm_set1_epi16(8);
+    const __m128i four = _mm_set1_epi16(4);
+
+    __m128i work0, work0_0, work0_1, sum_p_0;
+    __m128i sum_p = _mm_add_epi16(pq[5], _mm_add_epi16(pq[4], pq[3]));
+    __m128i sum_lp = _mm_add_epi16(pq[0], _mm_add_epi16(pq[2], pq[1]));
+    sum_p = _mm_add_epi16(sum_p, sum_lp);
+
+    __m128i sum_lq = _mm_srli_si128(sum_lp, 8);
+    __m128i sum_q = _mm_srli_si128(sum_p, 8);
+
+    sum_p_0 = _mm_add_epi16(eight, _mm_add_epi16(sum_p, sum_q));
+    sum_lp = _mm_add_epi16(four, _mm_add_epi16(sum_lp, sum_lq));
+
+    flat_p[0] = _mm_add_epi16(sum_lp, _mm_add_epi16(pq[3], pq[0]));
+    flat_q[0] = _mm_add_epi16(sum_lp, _mm_add_epi16(q[3], q[0]));
+
+    sum_p6 = _mm_add_epi16(pq[6], pq[6]);
+    sum_p3 = _mm_add_epi16(pq[3], pq[3]);
+
+    sum_q = _mm_sub_epi16(sum_p_0, pq[5]);
+    sum_p = _mm_sub_epi16(sum_p_0, q[5]);
+
+    work0_0 = _mm_add_epi16(_mm_add_epi16(pq[6], pq[0]), pq[1]);
+    work0_1 = _mm_add_epi16(sum_p6,
+                            _mm_add_epi16(pq[1], _mm_add_epi16(pq[2], pq[0])));
+
+    sum_lq = _mm_sub_epi16(sum_lp, pq[2]);
+    sum_lp = _mm_sub_epi16(sum_lp, q[2]);
+
+    work0 = _mm_add_epi16(sum_p3, pq[1]);
+    flat_p[1] = _mm_add_epi16(sum_lp, work0);
+    flat_q[1] = _mm_add_epi16(sum_lq, _mm_srli_si128(work0, 8));
+
+    flat_pq[0] = _mm_srli_epi16(_mm_unpacklo_epi64(flat_p[0], flat_q[0]), 3);
+    flat_pq[1] = _mm_srli_epi16(_mm_unpacklo_epi64(flat_p[1], flat_q[1]), 3);
+
+    sum_lp = _mm_sub_epi16(sum_lp, q[1]);
+    sum_lq = _mm_sub_epi16(sum_lq, pq[1]);
+
+    sum_p3 = _mm_add_epi16(sum_p3, pq[3]);
+    work0 = _mm_add_epi16(sum_p3, pq[2]);
+
+    flat_p[2] = _mm_add_epi16(sum_lp, work0);
+    flat_q[2] = _mm_add_epi16(sum_lq, _mm_srli_si128(work0, 8));
+    flat_pq[2] = _mm_srli_epi16(_mm_unpacklo_epi64(flat_p[2], flat_q[2]), 3);
+
+    int flat2_mask =
+        (0xffff != _mm_movemask_epi8(_mm_cmpeq_epi16(flat2, zero)));
+    if (flat2_mask) {
+      flat2_p[0] = _mm_add_epi16(sum_p_0, _mm_add_epi16(work0_0, q[0]));
+      flat2_q[0] = _mm_add_epi16(
+          sum_p_0, _mm_add_epi16(_mm_srli_si128(work0_0, 8), pq[0]));
+
+      flat2_p[1] = _mm_add_epi16(sum_p, work0_1);
+      flat2_q[1] = _mm_add_epi16(sum_q, _mm_srli_si128(work0_1, 8));
+
+      flat2_pq[0] =
+          _mm_srli_epi16(_mm_unpacklo_epi64(flat2_p[0], flat2_q[0]), 4);
+      flat2_pq[1] =
+          _mm_srli_epi16(_mm_unpacklo_epi64(flat2_p[1], flat2_q[1]), 4);
+
+      sum_p = _mm_sub_epi16(sum_p, q[4]);
+      sum_q = _mm_sub_epi16(sum_q, pq[4]);
+
+      sum_p6 = _mm_add_epi16(sum_p6, pq[6]);
+      work0 = _mm_add_epi16(sum_p6,
+                            _mm_add_epi16(pq[2], _mm_add_epi16(pq[3], pq[1])));
+      flat2_p[2] = _mm_add_epi16(sum_p, work0);
+      flat2_q[2] = _mm_add_epi16(sum_q, _mm_srli_si128(work0, 8));
+      flat2_pq[2] =
+          _mm_srli_epi16(_mm_unpacklo_epi64(flat2_p[2], flat2_q[2]), 4);
+
+      sum_p6 = _mm_add_epi16(sum_p6, pq[6]);
+      sum_p = _mm_sub_epi16(sum_p, q[3]);
+      sum_q = _mm_sub_epi16(sum_q, pq[3]);
+
+      work0 = _mm_add_epi16(sum_p6,
+                            _mm_add_epi16(pq[3], _mm_add_epi16(pq[4], pq[2])));
+      flat2_p[3] = _mm_add_epi16(sum_p, work0);
+      flat2_q[3] = _mm_add_epi16(sum_q, _mm_srli_si128(work0, 8));
+      flat2_pq[3] =
+          _mm_srli_epi16(_mm_unpacklo_epi64(flat2_p[3], flat2_q[3]), 4);
+
+      sum_p6 = _mm_add_epi16(sum_p6, pq[6]);
+      sum_p = _mm_sub_epi16(sum_p, q[2]);
+      sum_q = _mm_sub_epi16(sum_q, pq[2]);
+
+      work0 = _mm_add_epi16(sum_p6,
+                            _mm_add_epi16(pq[4], _mm_add_epi16(pq[5], pq[3])));
+      flat2_p[4] = _mm_add_epi16(sum_p, work0);
+      flat2_q[4] = _mm_add_epi16(sum_q, _mm_srli_si128(work0, 8));
+      flat2_pq[4] =
+          _mm_srli_epi16(_mm_unpacklo_epi64(flat2_p[4], flat2_q[4]), 4);
+
+      sum_p6 = _mm_add_epi16(sum_p6, pq[6]);
+      sum_p = _mm_sub_epi16(sum_p, q[1]);
+      sum_q = _mm_sub_epi16(sum_q, pq[1]);
+
+      work0 = _mm_add_epi16(sum_p6,
+                            _mm_add_epi16(pq[5], _mm_add_epi16(pq[6], pq[4])));
+      flat2_p[5] = _mm_add_epi16(sum_p, work0);
+      flat2_q[5] = _mm_add_epi16(sum_q, _mm_srli_si128(work0, 8));
+      flat2_pq[5] =
+          _mm_srli_epi16(_mm_unpacklo_epi64(flat2_p[5], flat2_q[5]), 4);
+    }  // flat2
+       // ~~~~~~~~~~ apply flat ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+    // highbd_filter8
+    pq[0] = _mm_unpacklo_epi64(ps0ps1, qs0qs1);
+    pq[1] = _mm_unpackhi_epi64(ps0ps1, qs0qs1);
+
+    for (i = 0; i < 3; i++) {
+      pq[i] = _mm_andnot_si128(flat, pq[i]);
+      flat_pq[i] = _mm_and_si128(flat, flat_pq[i]);
+      pq[i] = _mm_or_si128(pq[i], flat_pq[i]);
+    }
+
+    // wide flat
+    // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+    if (flat2_mask) {
+      for (i = 0; i < 6; i++) {
+        pq[i] = _mm_andnot_si128(flat2, pq[i]);
+        flat2_pq[i] = _mm_and_si128(flat2, flat2_pq[i]);
+        pq[i] = _mm_or_si128(pq[i], flat2_pq[i]);  // full list of pq values
+      }
+    }
+  } else {
+    pq[0] = _mm_unpacklo_epi64(ps0ps1, qs0qs1);
+    pq[1] = _mm_unpackhi_epi64(ps0ps1, qs0qs1);
+  }
+}
+
+void aom_highbd_lpf_horizontal_14_sse2(uint16_t *s, int pitch,
+                                       const uint8_t *blimit,
+                                       const uint8_t *limit,
+                                       const uint8_t *thresh, int bd) {
+  __m128i p[7], q[7], pq[7];
+  int i;
+
+  for (i = 0; i < 7; i++) {
+    p[i] = _mm_loadl_epi64((__m128i *)(s - (i + 1) * pitch));
+    q[i] = _mm_loadl_epi64((__m128i *)(s + i * pitch));
+  }
+
+  highbd_lpf_internal_14_sse2(p, q, pq, blimit, limit, thresh, bd);
+
+  for (i = 0; i < 6; i++) {
+    _mm_storel_epi64((__m128i *)(s - (i + 1) * pitch), pq[i]);
+    _mm_storel_epi64((__m128i *)(s + i * pitch), _mm_srli_si128(pq[i], 8));
+  }
+}
+
+static AOM_FORCE_INLINE void highbd_lpf_internal_14_dual_sse2(
+    __m128i *p, __m128i *q, const uint8_t *blt0, const uint8_t *lt0,
+    const uint8_t *thr0, const uint8_t *blt1, const uint8_t *lt1,
+    const uint8_t *thr1, int bd) {
+  __m128i blimit, limit, thresh, t80;
+  const __m128i zero = _mm_setzero_si128();
+
+  get_limit_dual(blt0, lt0, thr0, blt1, lt1, thr1, bd, &blimit, &limit, &thresh,
+                 &t80);
+  __m128i mask;
+  highbd_filter_mask_dual(p, q, &limit, &blimit, &mask);
+  __m128i flat, flat2;
+  highbd_flat_mask4_dual_sse2(p, q, &flat, &flat2, bd);
+
+  flat = _mm_and_si128(flat, mask);
+  flat2 = _mm_and_si128(flat2, flat);
+  __m128i ps[2], qs[2];
+  highbd_filter4_dual_sse2(p, q, ps, qs, &mask, &thresh, bd, &t80);
+  // flat and wide flat calculations
+
+  // if flat ==0 then flat2 is zero as well and we don't need any calc below
+  // sse4.1 if (0==_mm_test_all_zeros(flat,ff))
+  if (0xffff != _mm_movemask_epi8(_mm_cmpeq_epi16(flat, zero))) {
+    __m128i flat_p[3], flat_q[3];
+    __m128i flat2_p[6], flat2_q[6];
+    const __m128i eight = _mm_set1_epi16(8);
+    const __m128i four = _mm_set1_epi16(4);
+    __m128i sum_p_0 = _mm_add_epi16(p[5], _mm_add_epi16(p[4], p[3]));
+    __m128i sum_q = _mm_add_epi16(q[5], _mm_add_epi16(q[4], q[3]));
+    __m128i sum_lp = _mm_add_epi16(p[0], _mm_add_epi16(p[2], p[1]));
+    sum_p_0 = _mm_add_epi16(sum_p_0, sum_lp);
+    __m128i sum_lq = _mm_add_epi16(q[0], _mm_add_epi16(q[2], q[1]));
+    sum_q = _mm_add_epi16(sum_q, sum_lq);
+    sum_p_0 = _mm_add_epi16(eight, _mm_add_epi16(sum_p_0, sum_q));
+    sum_lp = _mm_add_epi16(four, _mm_add_epi16(sum_lp, sum_lq));
+    flat_p[0] =
+        _mm_srli_epi16(_mm_add_epi16(sum_lp, _mm_add_epi16(p[3], p[0])), 3);
+    flat_q[0] =
+        _mm_srli_epi16(_mm_add_epi16(sum_lp, _mm_add_epi16(q[3], q[0])), 3);
+    __m128i sum_p6 = _mm_add_epi16(p[6], p[6]);
+    __m128i sum_q6 = _mm_add_epi16(q[6], q[6]);
+    __m128i sum_p3 = _mm_add_epi16(p[3], p[3]);
+    __m128i sum_q3 = _mm_add_epi16(q[3], q[3]);
+
+    sum_q = _mm_sub_epi16(sum_p_0, p[5]);
+    __m128i sum_p = _mm_sub_epi16(sum_p_0, q[5]);
+
+    sum_lq = _mm_sub_epi16(sum_lp, p[2]);
+    sum_lp = _mm_sub_epi16(sum_lp, q[2]);
+    flat_p[1] =
+        _mm_srli_epi16(_mm_add_epi16(sum_lp, _mm_add_epi16(sum_p3, p[1])), 3);
+    flat_q[1] =
+        _mm_srli_epi16(_mm_add_epi16(sum_lq, _mm_add_epi16(sum_q3, q[1])), 3);
+
+    sum_lp = _mm_sub_epi16(sum_lp, q[1]);
+    sum_lq = _mm_sub_epi16(sum_lq, p[1]);
+    sum_p3 = _mm_add_epi16(sum_p3, p[3]);
+    sum_q3 = _mm_add_epi16(sum_q3, q[3]);
+    flat_p[2] =
+        _mm_srli_epi16(_mm_add_epi16(sum_lp, _mm_add_epi16(sum_p3, p[2])), 3);
+    flat_q[2] =
+        _mm_srli_epi16(_mm_add_epi16(sum_lq, _mm_add_epi16(sum_q3, q[2])), 3);
+
+    int flat2_mask =
+        (0xffff != _mm_movemask_epi8(_mm_cmpeq_epi16(flat2, zero)));
+    if (flat2_mask) {
+      flat2_p[0] = _mm_srli_epi16(
+          _mm_add_epi16(sum_p_0, _mm_add_epi16(_mm_add_epi16(p[6], p[0]),
+                                               _mm_add_epi16(p[1], q[0]))),
+          4);
+      flat2_q[0] = _mm_srli_epi16(
+          _mm_add_epi16(sum_p_0, _mm_add_epi16(_mm_add_epi16(q[6], q[0]),
+                                               _mm_add_epi16(p[0], q[1]))),
+          4);
+
+      flat2_p[1] = _mm_srli_epi16(
+          _mm_add_epi16(
+              sum_p,
+              _mm_add_epi16(sum_p6,
+                            _mm_add_epi16(p[1], _mm_add_epi16(p[2], p[0])))),
+          4);
+      flat2_q[1] = _mm_srli_epi16(
+          _mm_add_epi16(
+              sum_q,
+              _mm_add_epi16(sum_q6,
+                            _mm_add_epi16(q[1], _mm_add_epi16(q[0], q[2])))),
+          4);
+      sum_p6 = _mm_add_epi16(sum_p6, p[6]);
+      sum_q6 = _mm_add_epi16(sum_q6, q[6]);
+      sum_p = _mm_sub_epi16(sum_p, q[4]);
+      sum_q = _mm_sub_epi16(sum_q, p[4]);
+      flat2_p[2] = _mm_srli_epi16(
+          _mm_add_epi16(
+              sum_p,
+              _mm_add_epi16(sum_p6,
+                            _mm_add_epi16(p[2], _mm_add_epi16(p[3], p[1])))),
+          4);
+      flat2_q[2] = _mm_srli_epi16(
+          _mm_add_epi16(
+              sum_q,
+              _mm_add_epi16(sum_q6,
+                            _mm_add_epi16(q[2], _mm_add_epi16(q[1], q[3])))),
+          4);
+      sum_p6 = _mm_add_epi16(sum_p6, p[6]);
+      sum_q6 = _mm_add_epi16(sum_q6, q[6]);
+      sum_p = _mm_sub_epi16(sum_p, q[3]);
+      sum_q = _mm_sub_epi16(sum_q, p[3]);
+      flat2_p[3] = _mm_srli_epi16(
+          _mm_add_epi16(
+              sum_p,
+              _mm_add_epi16(sum_p6,
+                            _mm_add_epi16(p[3], _mm_add_epi16(p[4], p[2])))),
+          4);
+      flat2_q[3] = _mm_srli_epi16(
+          _mm_add_epi16(
+              sum_q,
+              _mm_add_epi16(sum_q6,
+                            _mm_add_epi16(q[3], _mm_add_epi16(q[2], q[4])))),
+          4);
+      sum_p6 = _mm_add_epi16(sum_p6, p[6]);
+      sum_q6 = _mm_add_epi16(sum_q6, q[6]);
+      sum_p = _mm_sub_epi16(sum_p, q[2]);
+      sum_q = _mm_sub_epi16(sum_q, p[2]);
+      flat2_p[4] = _mm_srli_epi16(
+          _mm_add_epi16(
+              sum_p,
+              _mm_add_epi16(sum_p6,
+                            _mm_add_epi16(p[4], _mm_add_epi16(p[5], p[3])))),
+          4);
+      flat2_q[4] = _mm_srli_epi16(
+          _mm_add_epi16(
+              sum_q,
+              _mm_add_epi16(sum_q6,
+                            _mm_add_epi16(q[4], _mm_add_epi16(q[3], q[5])))),
+          4);
+      sum_p6 = _mm_add_epi16(sum_p6, p[6]);
+      sum_q6 = _mm_add_epi16(sum_q6, q[6]);
+      sum_p = _mm_sub_epi16(sum_p, q[1]);
+      sum_q = _mm_sub_epi16(sum_q, p[1]);
+      flat2_p[5] = _mm_srli_epi16(
+          _mm_add_epi16(
+              sum_p,
+              _mm_add_epi16(sum_p6,
+                            _mm_add_epi16(p[5], _mm_add_epi16(p[6], p[4])))),
+          4);
+      flat2_q[5] = _mm_srli_epi16(
+          _mm_add_epi16(
+              sum_q,
+              _mm_add_epi16(sum_q6,
+                            _mm_add_epi16(q[5], _mm_add_epi16(q[4], q[6])))),
+          4);
+    }
+    // highbd_filter8
+    int i;
+    for (i = 0; i < 2; i++) {
+      ps[i] = _mm_andnot_si128(flat, ps[i]);
+      flat_p[i] = _mm_and_si128(flat, flat_p[i]);
+      p[i] = _mm_or_si128(ps[i], flat_p[i]);
+      qs[i] = _mm_andnot_si128(flat, qs[i]);
+      flat_q[i] = _mm_and_si128(flat, flat_q[i]);
+      q[i] = _mm_or_si128(qs[i], flat_q[i]);
+    }
+    p[2] = _mm_andnot_si128(flat, p[2]);
+    //  p2 remains unchanged if !(flat && mask)
+    flat_p[2] = _mm_and_si128(flat, flat_p[2]);
+    //  when (flat && mask)
+    p[2] = _mm_or_si128(p[2], flat_p[2]);  // full list of p2 values
+    q[2] = _mm_andnot_si128(flat, q[2]);
+    flat_q[2] = _mm_and_si128(flat, flat_q[2]);
+    q[2] = _mm_or_si128(q[2], flat_q[2]);  // full list of q2 values
+
+    for (i = 0; i < 2; i++) {
+      ps[i] = _mm_andnot_si128(flat, ps[i]);
+      flat_p[i] = _mm_and_si128(flat, flat_p[i]);
+      p[i] = _mm_or_si128(ps[i], flat_p[i]);
+      qs[i] = _mm_andnot_si128(flat, qs[i]);
+      flat_q[i] = _mm_and_si128(flat, flat_q[i]);
+      q[i] = _mm_or_si128(qs[i], flat_q[i]);
+    }
+    // highbd_filter16
+    if (flat2_mask) {
+      for (i = 0; i < 6; i++) {
+        //  p[i] remains unchanged if !(flat2 && flat && mask)
+        p[i] = _mm_andnot_si128(flat2, p[i]);
+        flat2_p[i] = _mm_and_si128(flat2, flat2_p[i]);
+        //  get values for when (flat2 && flat && mask)
+        p[i] = _mm_or_si128(p[i], flat2_p[i]);  // full list of p values
+        q[i] = _mm_andnot_si128(flat2, q[i]);
+        flat2_q[i] = _mm_and_si128(flat2, flat2_q[i]);
+        q[i] = _mm_or_si128(q[i], flat2_q[i]);
+      }
+    }
+  } else {
+    p[0] = ps[0];
+    q[0] = qs[0];
+    p[1] = ps[1];
+    q[1] = qs[1];
+  }
+}
+
+void aom_highbd_lpf_horizontal_14_dual_sse2(
+    uint16_t *s, int pitch, const uint8_t *_blimit0, const uint8_t *_limit0,
+    const uint8_t *_thresh0, const uint8_t *_blimit1, const uint8_t *_limit1,
+    const uint8_t *_thresh1, int bd) {
+  __m128i p[7], q[7];
+  int i;
+  load_highbd_pixel(s, 7, pitch, p, q);
+
+  highbd_lpf_internal_14_dual_sse2(p, q, _blimit0, _limit0, _thresh0, _blimit1,
+                                   _limit1, _thresh1, bd);
+
+  for (i = 0; i < 6; i++) {
+    _mm_storeu_si128((__m128i *)(s - (i + 1) * pitch), p[i]);
+    _mm_storeu_si128((__m128i *)(s + i * pitch), q[i]);
+  }
+}
+
+static AOM_FORCE_INLINE void highbd_lpf_internal_6_sse2(
+    __m128i *p2, __m128i *p1, __m128i *p0, __m128i *q0, __m128i *q1,
+    __m128i *q2, __m128i *p1p0_out, __m128i *q1q0_out, const uint8_t *_blimit,
+    const uint8_t *_limit, const uint8_t *_thresh, int bd) {
+  __m128i blimit, limit, thresh;
+  __m128i mask, hev, flat;
+  __m128i pq[3];
+  __m128i p1p0, q1q0, abs_p1p0, ps1ps0, qs1qs0;
+  __m128i flat_p1p0, flat_q0q1;
+
+  pq[0] = _mm_unpacklo_epi64(*p0, *q0);
+  pq[1] = _mm_unpacklo_epi64(*p1, *q1);
+  pq[2] = _mm_unpacklo_epi64(*p2, *q2);
+
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i four = _mm_set1_epi16(4);
+  __m128i t80;
+  const __m128i one = _mm_set1_epi16(0x1);
+
+  get_limit(_blimit, _limit, _thresh, bd, &blimit, &limit, &thresh, &t80);
+
+  highbd_hev_filter_mask_x_sse2(pq, 3, &p1p0, &q1q0, &abs_p1p0, &limit, &blimit,
+                                &thresh, &hev, &mask);
+
+  // lp filter
+  highbd_filter4_sse2(&p1p0, &q1q0, &hev, &mask, q1q0_out, p1p0_out, &t80, bd);
+
+  // flat_mask
+  flat = _mm_max_epi16(abs_diff16(pq[2], pq[0]), abs_p1p0);
+  flat = _mm_max_epi16(flat, _mm_srli_si128(flat, 8));
+
+  flat = _mm_subs_epu16(flat, _mm_slli_epi16(one, bd - 8));
+
+  flat = _mm_cmpeq_epi16(flat, zero);
+  flat = _mm_and_si128(flat, mask);
+  // replicate for the further "merged variables" usage
+  flat = _mm_unpacklo_epi64(flat, flat);
+
+  // 5 tap filter
+  // need it only if flat !=0
+  if (0xffff != _mm_movemask_epi8(_mm_cmpeq_epi16(flat, zero))) {
+    __m128i workp_a, workp_b, workp_c;
+    __m128i pq0x2_pq1, pq1_pq2;
+
+    // op1
+    pq0x2_pq1 =
+        _mm_add_epi16(_mm_add_epi16(pq[0], pq[0]), pq[1]);  // p0 *2 + p1
+    pq1_pq2 = _mm_add_epi16(pq[1], pq[2]);                  // p1 + p2
+    workp_a = _mm_add_epi16(_mm_add_epi16(pq0x2_pq1, four),
+                            pq1_pq2);  // p2 + p0 * 2 + p1 * 2 + 4
+
+    workp_b = _mm_add_epi16(_mm_add_epi16(pq[2], pq[2]), *q0);
+    workp_b =
+        _mm_add_epi16(workp_a, workp_b);  // p2 * 3 + p1 * 2 + p0 * 2 + q0 + 4
+
+    // op0
+    workp_c = _mm_srli_si128(pq0x2_pq1, 8);  // q0 * 2 + q1
+    workp_a = _mm_add_epi16(workp_a,
+                            workp_c);  // p2 + p0 * 2 + p1 * 2 + q0 * 2 + q1 + 4
+    workp_b = _mm_unpacklo_epi64(workp_a, workp_b);
+    flat_p1p0 = _mm_srli_epi16(workp_b, 3);
+
+    // oq0
+    workp_a = _mm_sub_epi16(_mm_sub_epi16(workp_a, pq[2]),
+                            pq[1]);  // p0 * 2 + p1  + q0 * 2 + q1 + 4
+    workp_b = _mm_srli_si128(pq1_pq2, 8);
+    workp_a = _mm_add_epi16(
+        workp_a, workp_b);  // p0 * 2 + p1  + q0 * 2 + q1 * 2 + q2 + 4
+    // workp_shft0 = _mm_srli_epi16(workp_a, 3);
+
+    // oq1
+    workp_c = _mm_sub_epi16(_mm_sub_epi16(workp_a, pq[1]),
+                            pq[0]);  // p0   + q0 * 2 + q1 * 2 + q2 + 4
+    workp_b = _mm_add_epi16(*q2, *q2);
+    workp_b =
+        _mm_add_epi16(workp_c, workp_b);  // p0  + q0 * 2 + q1 * 2 + q2 * 3 + 4
+
+    workp_a = _mm_unpacklo_epi64(workp_a, workp_b);
+    flat_q0q1 = _mm_srli_epi16(workp_a, 3);
+
+    qs1qs0 = _mm_andnot_si128(flat, *q1q0_out);
+    q1q0 = _mm_and_si128(flat, flat_q0q1);
+    *q1q0_out = _mm_or_si128(qs1qs0, q1q0);
+
+    ps1ps0 = _mm_andnot_si128(flat, *p1p0_out);
+    p1p0 = _mm_and_si128(flat, flat_p1p0);
+    *p1p0_out = _mm_or_si128(ps1ps0, p1p0);
+  }
+}
+
+static AOM_FORCE_INLINE void highbd_lpf_internal_6_dual_sse2(
+    __m128i *p2, __m128i *p1, __m128i *p0, __m128i *q0, __m128i *q1,
+    __m128i *q2, const unsigned char *_blimit0, const unsigned char *_limit0,
+    const unsigned char *_thresh0, const unsigned char *_blimit1,
+    const unsigned char *_limit1, const unsigned char *_thresh1, int bd) {
+  const __m128i zero = _mm_setzero_si128();
+  __m128i blimit0, limit0, thresh0;
+  __m128i t80;
+  __m128i mask, flat, work;
+  __m128i abs_p1q1, abs_p0q0, abs_p1p0, abs_p2p1, abs_q1q0, abs_q2q1;
+  __m128i op1, op0, oq0, oq1;
+  const __m128i four = _mm_set1_epi16(4);
+  const __m128i one = _mm_set1_epi16(0x1);
+  const __m128i ffff = _mm_cmpeq_epi16(one, one);
+
+  get_limit_dual(_blimit0, _limit0, _thresh0, _blimit1, _limit1, _thresh1, bd,
+                 &blimit0, &limit0, &thresh0, &t80);
+
+  abs_p2p1 = abs_diff16(*p2, *p1);
+  abs_p1p0 = abs_diff16(*p1, *p0);
+  abs_q1q0 = abs_diff16(*q1, *q0);
+  abs_q2q1 = abs_diff16(*q2, *q1);
+
+  abs_p0q0 = abs_diff16(*p0, *q0);
+  abs_p1q1 = abs_diff16(*p1, *q1);
+
+  abs_p0q0 = _mm_adds_epu16(abs_p0q0, abs_p0q0);
+  abs_p1q1 = _mm_srli_epi16(abs_p1q1, 1);
+  mask = _mm_subs_epu16(_mm_adds_epu16(abs_p0q0, abs_p1q1), blimit0);
+  mask = _mm_xor_si128(_mm_cmpeq_epi16(mask, zero), ffff);
+  // mask |= (abs(*p0 - *q0) * 2 + abs(*p1 - *q1) / 2  > blimit) * -1;
+  // So taking maximums continues to work:
+  mask = _mm_and_si128(mask, _mm_adds_epu16(limit0, one));
+
+  mask = _mm_max_epi16(abs_q2q1, mask);
+  work = _mm_max_epi16(abs_p1p0, abs_q1q0);
+  mask = _mm_max_epi16(work, mask);
+  mask = _mm_max_epi16(mask, abs_p2p1);
+  mask = _mm_subs_epu16(mask, limit0);
+  mask = _mm_cmpeq_epi16(mask, zero);
+
+  // lp filter
+  __m128i ps[2], qs[2], p[2], q[2];
+  {
+    p[0] = *p0;
+    p[1] = *p1;
+    q[0] = *q0;
+    q[1] = *q1;
+    // filter_mask and hev_mask
+    highbd_filter4_dual_sse2(p, q, ps, qs, &mask, &thresh0, bd, &t80);
+  }
+
+  // flat_mask
+  flat = _mm_max_epi16(abs_diff16(*q2, *q0), abs_diff16(*p2, *p0));
+  flat = _mm_max_epi16(flat, work);
+
+  flat = _mm_subs_epu16(flat, _mm_slli_epi16(one, bd - 8));
+
+  flat = _mm_cmpeq_epi16(flat, zero);
+  flat = _mm_and_si128(flat, mask);  // flat & mask
+
+  // 5 tap filter
+  // need it only if flat !=0
+  if (0xffff != _mm_movemask_epi8(_mm_cmpeq_epi16(flat, zero))) {
+    __m128i workp_a, workp_b, workp_shft0, workp_shft1;
+
+    // op1
+    workp_a = _mm_add_epi16(_mm_add_epi16(*p0, *p0),
+                            _mm_add_epi16(*p1, *p1));  // *p0 *2 + *p1 * 2
+    workp_a = _mm_add_epi16(_mm_add_epi16(workp_a, four),
+                            *p2);  // *p2 + *p0 * 2 + *p1 * 2 + 4
+
+    workp_b = _mm_add_epi16(_mm_add_epi16(*p2, *p2), *q0);
+    workp_shft0 = _mm_add_epi16(
+        workp_a, workp_b);  // *p2 * 3 + *p1 * 2 + *p0 * 2 + *q0 + 4
+    op1 = _mm_srli_epi16(workp_shft0, 3);
+
+    // op0
+    workp_b = _mm_add_epi16(_mm_add_epi16(*q0, *q0), *q1);  // *q0 * 2 + *q1
+    workp_a =
+        _mm_add_epi16(workp_a,
+                      workp_b);  // *p2 + *p0 * 2 + *p1 * 2 + *q0 * 2 + *q1 + 4
+    op0 = _mm_srli_epi16(workp_a, 3);
+
+    // oq0
+    workp_a = _mm_sub_epi16(_mm_sub_epi16(workp_a, *p2),
+                            *p1);  // *p0 * 2 + *p1  + *q0 * 2 + *q1 + 4
+    workp_b = _mm_add_epi16(*q1, *q2);
+    workp_shft0 = _mm_add_epi16(
+        workp_a, workp_b);  // *p0 * 2 + *p1  + *q0 * 2 + *q1 * 2 + *q2 + 4
+    oq0 = _mm_srli_epi16(workp_shft0, 3);
+
+    // oq1
+    workp_a = _mm_sub_epi16(_mm_sub_epi16(workp_shft0, *p1),
+                            *p0);  // *p0   + *q0 * 2 + *q1 * 2 + *q2 + 4
+    workp_b = _mm_add_epi16(*q2, *q2);
+    workp_shft1 = _mm_add_epi16(
+        workp_a, workp_b);  // *p0  + *q0 * 2 + *q1 * 2 + *q2 * 3 + 4
+    oq1 = _mm_srli_epi16(workp_shft1, 3);
+
+    qs[0] = _mm_andnot_si128(flat, qs[0]);
+    oq0 = _mm_and_si128(flat, oq0);
+    *q0 = _mm_or_si128(qs[0], oq0);
+
+    qs[1] = _mm_andnot_si128(flat, qs[1]);
+    oq1 = _mm_and_si128(flat, oq1);
+    *q1 = _mm_or_si128(qs[1], oq1);
+
+    ps[0] = _mm_andnot_si128(flat, ps[0]);
+    op0 = _mm_and_si128(flat, op0);
+    *p0 = _mm_or_si128(ps[0], op0);
+
+    ps[1] = _mm_andnot_si128(flat, ps[1]);
+    op1 = _mm_and_si128(flat, op1);
+    *p1 = _mm_or_si128(ps[1], op1);
+  } else {
+    *q0 = qs[0];
+    *q1 = qs[1];
+    *p0 = ps[0];
+    *p1 = ps[1];
+  }
+}
+
+void aom_highbd_lpf_horizontal_6_sse2(uint16_t *s, int p,
+                                      const uint8_t *_blimit,
+                                      const uint8_t *_limit,
+                                      const uint8_t *_thresh, int bd) {
+  __m128i p2, p1, p0, q0, q1, q2, p1p0_out, q1q0_out;
+
+  p2 = _mm_loadl_epi64((__m128i *)(s - 3 * p));
+  p1 = _mm_loadl_epi64((__m128i *)(s - 2 * p));
+  p0 = _mm_loadl_epi64((__m128i *)(s - 1 * p));
+  q0 = _mm_loadl_epi64((__m128i *)(s + 0 * p));
+  q1 = _mm_loadl_epi64((__m128i *)(s + 1 * p));
+  q2 = _mm_loadl_epi64((__m128i *)(s + 2 * p));
+
+  highbd_lpf_internal_6_sse2(&p2, &p1, &p0, &q0, &q1, &q2, &p1p0_out, &q1q0_out,
+                             _blimit, _limit, _thresh, bd);
+
+  _mm_storel_epi64((__m128i *)(s - 2 * p), _mm_srli_si128(p1p0_out, 8));
+  _mm_storel_epi64((__m128i *)(s - 1 * p), p1p0_out);
+  _mm_storel_epi64((__m128i *)(s + 0 * p), q1q0_out);
+  _mm_storel_epi64((__m128i *)(s + 1 * p), _mm_srli_si128(q1q0_out, 8));
+}
+
+void aom_highbd_lpf_horizontal_6_dual_sse2(
+    uint16_t *s, int p, const uint8_t *_blimit0, const uint8_t *_limit0,
+    const uint8_t *_thresh0, const uint8_t *_blimit1, const uint8_t *_limit1,
+    const uint8_t *_thresh1, int bd) {
+  __m128i p2, p1, p0, q0, q1, q2;
+
+  p2 = _mm_loadu_si128((__m128i *)(s - 3 * p));
+  p1 = _mm_loadu_si128((__m128i *)(s - 2 * p));
+  p0 = _mm_loadu_si128((__m128i *)(s - 1 * p));
+  q0 = _mm_loadu_si128((__m128i *)(s + 0 * p));
+  q1 = _mm_loadu_si128((__m128i *)(s + 1 * p));
+  q2 = _mm_loadu_si128((__m128i *)(s + 2 * p));
+
+  highbd_lpf_internal_6_dual_sse2(&p2, &p1, &p0, &q0, &q1, &q2, _blimit0,
+                                  _limit0, _thresh0, _blimit1, _limit1,
+                                  _thresh1, bd);
+
+  _mm_storeu_si128((__m128i *)(s - 2 * p), p1);
+  _mm_storeu_si128((__m128i *)(s - 1 * p), p0);
+  _mm_storeu_si128((__m128i *)(s + 0 * p), q0);
+  _mm_storeu_si128((__m128i *)(s + 1 * p), q1);
+}
+
+static AOM_FORCE_INLINE void highbd_lpf_internal_8_sse2(
+    __m128i *p3, __m128i *q3, __m128i *p2, __m128i *q2, __m128i *p1,
+    __m128i *q1, __m128i *p0, __m128i *q0, __m128i *q1q0_out, __m128i *p1p0_out,
+    const unsigned char *_blimit, const unsigned char *_limit,
+    const unsigned char *_thresh, int bd) {
+  const __m128i zero = _mm_setzero_si128();
+  __m128i blimit, limit, thresh;
+  __m128i mask, hev, flat;
+  __m128i pq[4];
+  __m128i p1p0, q1q0, ps1ps0, qs1qs0;
+  __m128i work_a, opq2, flat_p1p0, flat_q0q1;
+
+  pq[0] = _mm_unpacklo_epi64(*p0, *q0);
+  pq[1] = _mm_unpacklo_epi64(*p1, *q1);
+  pq[2] = _mm_unpacklo_epi64(*p2, *q2);
+  pq[3] = _mm_unpacklo_epi64(*p3, *q3);
+
+  __m128i abs_p1p0;
+
+  const __m128i four = _mm_set1_epi16(4);
+  __m128i t80;
+  const __m128i one = _mm_set1_epi16(0x1);
+
+  get_limit(_blimit, _limit, _thresh, bd, &blimit, &limit, &thresh, &t80);
+
+  highbd_hev_filter_mask_x_sse2(pq, 4, &p1p0, &q1q0, &abs_p1p0, &limit, &blimit,
+                                &thresh, &hev, &mask);
+
+  // lp filter
+  highbd_filter4_sse2(&p1p0, &q1q0, &hev, &mask, q1q0_out, p1p0_out, &t80, bd);
+
+  // flat_mask4
+  flat = _mm_max_epi16(abs_diff16(pq[2], pq[0]), abs_diff16(pq[3], pq[0]));
+  flat = _mm_max_epi16(abs_p1p0, flat);
+  flat = _mm_max_epi16(flat, _mm_srli_si128(flat, 8));
+
+  flat = _mm_subs_epu16(flat, _mm_slli_epi16(one, bd - 8));
+
+  flat = _mm_cmpeq_epi16(flat, zero);
+  flat = _mm_and_si128(flat, mask);
+  // replicate for the further "merged variables" usage
+  flat = _mm_unpacklo_epi64(flat, flat);
+
+  if (0xffff != _mm_movemask_epi8(_mm_cmpeq_epi16(flat, zero))) {
+    __m128i workp_a, workp_b, workp_c, workp_shft0, workp_shft1;
+    // Added before shift for rounding part of ROUND_POWER_OF_TWO
+
+    // o*p2
+    workp_a = _mm_add_epi16(_mm_add_epi16(*p3, *p3), _mm_add_epi16(*p2, *p1));
+    workp_a = _mm_add_epi16(_mm_add_epi16(workp_a, four), *p0);
+    workp_c = _mm_add_epi16(_mm_add_epi16(*q0, *p2), *p3);
+    workp_c = _mm_add_epi16(workp_a, workp_c);
+
+    // o*p1
+    workp_b = _mm_add_epi16(_mm_add_epi16(*q0, *q1), *p1);
+    workp_shft0 = _mm_add_epi16(workp_a, workp_b);
+
+    // o*p0
+    workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, *p3), *q2);
+    workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, *p1), *p0);
+    workp_shft1 = _mm_add_epi16(workp_a, workp_b);
+
+    flat_p1p0 = _mm_srli_epi16(_mm_unpacklo_epi64(workp_shft1, workp_shft0), 3);
+
+    // oq0
+    workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, *p3), *q3);
+    workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, *p0), *q0);
+    workp_shft0 = _mm_add_epi16(workp_a, workp_b);
+
+    // oq1
+    workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, *p2), *q3);
+    workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, *q0), *q1);
+    workp_shft1 = _mm_add_epi16(workp_a, workp_b);
+
+    flat_q0q1 = _mm_srli_epi16(_mm_unpacklo_epi64(workp_shft0, workp_shft1), 3);
+
+    // oq2
+    workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, *p1), *q3);
+    workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, *q1), *q2);
+    workp_a = _mm_add_epi16(workp_a, workp_b);
+    opq2 = _mm_srli_epi16(_mm_unpacklo_epi64(workp_c, workp_a), 3);
+
+    qs1qs0 = _mm_andnot_si128(flat, *q1q0_out);
+    q1q0 = _mm_and_si128(flat, flat_q0q1);
+    *q1q0_out = _mm_or_si128(qs1qs0, q1q0);
+
+    ps1ps0 = _mm_andnot_si128(flat, *p1p0_out);
+    p1p0 = _mm_and_si128(flat, flat_p1p0);
+    *p1p0_out = _mm_or_si128(ps1ps0, p1p0);
+
+    work_a = _mm_andnot_si128(flat, pq[2]);
+    *p2 = _mm_and_si128(flat, opq2);
+    *p2 = _mm_or_si128(work_a, *p2);
+    *q2 = _mm_srli_si128(*p2, 8);
+  }
+}
+
+static AOM_FORCE_INLINE void highbd_lpf_internal_8_dual_sse2(
+    __m128i *p3, __m128i *q3, __m128i *p2, __m128i *q2, __m128i *p1,
+    __m128i *q1, __m128i *p0, __m128i *q0, const unsigned char *_blimit0,
+    const unsigned char *_limit0, const unsigned char *_thresh0,
+    const unsigned char *_blimit1, const unsigned char *_limit1,
+    const unsigned char *_thresh1, int bd) {
+  __m128i blimit0, limit0, thresh0;
+  __m128i t80;
+  __m128i mask, flat;
+  __m128i work_a, op2, oq2, op1, op0, oq0, oq1;
+  __m128i abs_p1q1, abs_p0q0, work0, work1, work2;
+
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i four = _mm_set1_epi16(4);
+  const __m128i one = _mm_set1_epi16(0x1);
+  const __m128i ffff = _mm_cmpeq_epi16(one, one);
+
+  get_limit_dual(_blimit0, _limit0, _thresh0, _blimit1, _limit1, _thresh1, bd,
+                 &blimit0, &limit0, &thresh0, &t80);
+
+  abs_p0q0 = abs_diff16(*p0, *q0);
+  abs_p1q1 = abs_diff16(*p1, *q1);
+
+  abs_p0q0 = _mm_adds_epu16(abs_p0q0, abs_p0q0);
+  abs_p1q1 = _mm_srli_epi16(abs_p1q1, 1);
+  mask = _mm_subs_epu16(_mm_adds_epu16(abs_p0q0, abs_p1q1), blimit0);
+  mask = _mm_xor_si128(_mm_cmpeq_epi16(mask, zero), ffff);
+  // mask |= (abs(*p0 - q0) * 2 + abs(*p1 - q1) / 2  > blimit) * -1;
+
+  // So taking maximums continues to work:
+  mask = _mm_and_si128(mask, _mm_adds_epu16(limit0, one));
+
+  work0 = _mm_max_epi16(abs_diff16(*p3, *p2), abs_diff16(*p2, *p1));
+  work1 =
+      _mm_max_epi16(abs_diff16(*p1, *p0), abs_diff16(*q1, *q0));  // tbu 4 flat
+  work0 = _mm_max_epi16(work0, work1);
+  work2 = _mm_max_epi16(abs_diff16(*q2, *q1), abs_diff16(*q2, *q3));
+  work2 = _mm_max_epi16(work2, work0);
+  mask = _mm_max_epi16(work2, mask);
+
+  mask = _mm_subs_epu16(mask, limit0);
+  mask = _mm_cmpeq_epi16(mask, zero);
+
+  // lp filter
+  __m128i ps[2], qs[2], p[2], q[2];
+  {
+    p[0] = *p0;
+    p[1] = *p1;
+    q[0] = *q0;
+    q[1] = *q1;
+    // filter_mask and hev_mask
+    highbd_filter4_dual_sse2(p, q, ps, qs, &mask, &thresh0, bd, &t80);
+  }
+
+  flat = _mm_max_epi16(abs_diff16(*p2, *p0), abs_diff16(*q2, *q0));
+  flat = _mm_max_epi16(work1, flat);
+  work0 = _mm_max_epi16(abs_diff16(*p3, *p0), abs_diff16(*q3, *q0));
+  flat = _mm_max_epi16(work0, flat);
+
+  flat = _mm_subs_epu16(flat, _mm_slli_epi16(one, bd - 8));
+  flat = _mm_cmpeq_epi16(flat, zero);
+  flat = _mm_and_si128(flat, mask);  // flat & mask
+
+  // filter8 need it only if flat !=0
+  if (0xffff != _mm_movemask_epi8(_mm_cmpeq_epi16(flat, zero))) {
+    __m128i workp_a, workp_b;
+    // Added before shift for rounding part of ROUND_POWER_OF_TWO
+
+    // o*p2
+    workp_a = _mm_add_epi16(_mm_add_epi16(*p3, *p3), _mm_add_epi16(*p2, *p1));
+    workp_a = _mm_add_epi16(_mm_add_epi16(workp_a, four), *p0);
+    workp_b = _mm_add_epi16(_mm_add_epi16(*q0, *p2), *p3);
+    op2 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+
+    // o*p1
+    workp_b = _mm_add_epi16(_mm_add_epi16(*q0, *q1), *p1);
+    op1 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+
+    // o*p0
+    workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, *p3), *q2);
+    workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, *p1), *p0);
+    op0 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+
+    // oq0
+    workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, *p3), *q3);
+    workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, *p0), *q0);
+    oq0 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+
+    // oq1
+    workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, *p2), *q3);
+    workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, *q0), *q1);
+    oq1 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+
+    // oq2
+    workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, *p1), *q3);
+    workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, *q1), *q2);
+    oq2 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+
+    qs[0] = _mm_andnot_si128(flat, qs[0]);
+    oq0 = _mm_and_si128(flat, oq0);
+    *q0 = _mm_or_si128(qs[0], oq0);
+
+    qs[1] = _mm_andnot_si128(flat, qs[1]);
+    oq1 = _mm_and_si128(flat, oq1);
+    *q1 = _mm_or_si128(qs[1], oq1);
+
+    ps[0] = _mm_andnot_si128(flat, ps[0]);
+    op0 = _mm_and_si128(flat, op0);
+    *p0 = _mm_or_si128(ps[0], op0);
+
+    ps[1] = _mm_andnot_si128(flat, ps[1]);
+    op1 = _mm_and_si128(flat, op1);
+    *p1 = _mm_or_si128(ps[1], op1);
+
+    work_a = _mm_andnot_si128(flat, *q2);
+    *q2 = _mm_and_si128(flat, oq2);
+    *q2 = _mm_or_si128(work_a, *q2);
+
+    work_a = _mm_andnot_si128(flat, *p2);
+    *p2 = _mm_and_si128(flat, op2);
+    *p2 = _mm_or_si128(work_a, *p2);
+  } else {
+    *q0 = qs[0];
+    *q1 = qs[1];
+    *p0 = ps[0];
+    *p1 = ps[1];
+  }
+}
+
+void aom_highbd_lpf_horizontal_8_sse2(uint16_t *s, int p,
+                                      const uint8_t *_blimit,
+                                      const uint8_t *_limit,
+                                      const uint8_t *_thresh, int bd) {
+  __m128i p2, p1, p0, q0, q1, q2, p3, q3;
+  __m128i q1q0, p1p0;
+
+  p3 = _mm_loadl_epi64((__m128i *)(s - 4 * p));
+  q3 = _mm_loadl_epi64((__m128i *)(s + 3 * p));
+  p2 = _mm_loadl_epi64((__m128i *)(s - 3 * p));
+  q2 = _mm_loadl_epi64((__m128i *)(s + 2 * p));
+  p1 = _mm_loadl_epi64((__m128i *)(s - 2 * p));
+  q1 = _mm_loadl_epi64((__m128i *)(s + 1 * p));
+  p0 = _mm_loadl_epi64((__m128i *)(s - 1 * p));
+  q0 = _mm_loadl_epi64((__m128i *)(s + 0 * p));
+
+  highbd_lpf_internal_8_sse2(&p3, &q3, &p2, &q2, &p1, &q1, &p0, &q0, &q1q0,
+                             &p1p0, _blimit, _limit, _thresh, bd);
+
+  _mm_storel_epi64((__m128i *)(s - 3 * p), p2);
+  _mm_storel_epi64((__m128i *)(s - 2 * p), _mm_srli_si128(p1p0, 8));
+  _mm_storel_epi64((__m128i *)(s - 1 * p), p1p0);
+  _mm_storel_epi64((__m128i *)(s + 0 * p), q1q0);
+  _mm_storel_epi64((__m128i *)(s + 1 * p), _mm_srli_si128(q1q0, 8));
+  _mm_storel_epi64((__m128i *)(s + 2 * p), q2);
+}
+
+void aom_highbd_lpf_horizontal_8_dual_sse2(
+    uint16_t *s, int p, const uint8_t *_blimit0, const uint8_t *_limit0,
+    const uint8_t *_thresh0, const uint8_t *_blimit1, const uint8_t *_limit1,
+    const uint8_t *_thresh1, int bd) {
+  __m128i p2, p1, p0, q0, q1, q2, p3, q3;
+
+  p3 = _mm_loadu_si128((__m128i *)(s - 4 * p));
+  q3 = _mm_loadu_si128((__m128i *)(s + 3 * p));
+  p2 = _mm_loadu_si128((__m128i *)(s - 3 * p));
+  q2 = _mm_loadu_si128((__m128i *)(s + 2 * p));
+  p1 = _mm_loadu_si128((__m128i *)(s - 2 * p));
+  q1 = _mm_loadu_si128((__m128i *)(s + 1 * p));
+  p0 = _mm_loadu_si128((__m128i *)(s - 1 * p));
+  q0 = _mm_loadu_si128((__m128i *)(s + 0 * p));
+
+  highbd_lpf_internal_8_dual_sse2(&p3, &q3, &p2, &q2, &p1, &q1, &p0, &q0,
+                                  _blimit0, _limit0, _thresh0, _blimit1,
+                                  _limit1, _thresh1, bd);
+
+  _mm_storeu_si128((__m128i *)(s - 3 * p), p2);
+  _mm_storeu_si128((__m128i *)(s - 2 * p), p1);
+  _mm_storeu_si128((__m128i *)(s - 1 * p), p0);
+  _mm_storeu_si128((__m128i *)(s + 0 * p), q0);
+  _mm_storeu_si128((__m128i *)(s + 1 * p), q1);
+  _mm_storeu_si128((__m128i *)(s + 2 * p), q2);
+}
+
+static AOM_FORCE_INLINE void highbd_lpf_internal_4_sse2(
+    __m128i *p1, __m128i *p0, __m128i *q0, __m128i *q1, __m128i *q1q0_out,
+    __m128i *p1p0_out, const uint8_t *_blimit, const uint8_t *_limit,
+    const uint8_t *_thresh, int bd) {
+  __m128i blimit, limit, thresh;
+  __m128i mask, hev;
+  __m128i p1p0, q1q0;
+  __m128i pq[2];
+
+  __m128i abs_p1p0;
+
+  __m128i t80;
+  get_limit(_blimit, _limit, _thresh, bd, &blimit, &limit, &thresh, &t80);
+
+  pq[0] = _mm_unpacklo_epi64(*p0, *q0);
+  pq[1] = _mm_unpacklo_epi64(*p1, *q1);
+
+  highbd_hev_filter_mask_x_sse2(pq, 2, &p1p0, &q1q0, &abs_p1p0, &limit, &blimit,
+                                &thresh, &hev, &mask);
+
+  highbd_filter4_sse2(&p1p0, &q1q0, &hev, &mask, q1q0_out, p1p0_out, &t80, bd);
+}
+
+static AOM_FORCE_INLINE void highbd_lpf_internal_4_dual_sse2(
+    __m128i *p1, __m128i *p0, __m128i *q0, __m128i *q1, __m128i *ps,
+    __m128i *qs, const uint8_t *_blimit0, const uint8_t *_limit0,
+    const uint8_t *_thresh0, const uint8_t *_blimit1, const uint8_t *_limit1,
+    const uint8_t *_thresh1, int bd) {
+  __m128i blimit0, limit0, thresh0;
+  __m128i mask, flat;
+  __m128i p[2], q[2];
+
+  const __m128i zero = _mm_setzero_si128();
+  __m128i abs_p0q0 = abs_diff16(*q0, *p0);
+  __m128i abs_p1q1 = abs_diff16(*q1, *p1);
+
+  __m128i abs_p1p0 = abs_diff16(*p1, *p0);
+  __m128i abs_q1q0 = abs_diff16(*q1, *q0);
+
+  const __m128i ffff = _mm_cmpeq_epi16(abs_p1p0, abs_p1p0);
+  const __m128i one = _mm_set1_epi16(1);
+
+  __m128i t80;
+
+  get_limit_dual(_blimit0, _limit0, _thresh0, _blimit1, _limit1, _thresh1, bd,
+                 &blimit0, &limit0, &thresh0, &t80);
+
+  // filter_mask and hev_mask
+  flat = _mm_max_epi16(abs_p1p0, abs_q1q0);
+
+  abs_p0q0 = _mm_adds_epu16(abs_p0q0, abs_p0q0);
+  abs_p1q1 = _mm_srli_epi16(abs_p1q1, 1);
+
+  mask = _mm_subs_epu16(_mm_adds_epu16(abs_p0q0, abs_p1q1), blimit0);
+  mask = _mm_xor_si128(_mm_cmpeq_epi16(mask, zero), ffff);
+  // mask |= (abs(*p0 - *q0) * 2 + abs(*p1 - *q1) / 2  > blimit) * -1;
+  // So taking maximums continues to work:
+  mask = _mm_and_si128(mask, _mm_adds_epu16(limit0, one));
+  mask = _mm_max_epi16(flat, mask);
+
+  mask = _mm_subs_epu16(mask, limit0);
+  mask = _mm_cmpeq_epi16(mask, zero);
+
+  p[0] = *p0;
+  p[1] = *p1;
+  q[0] = *q0;
+  q[1] = *q1;
+
+  highbd_filter4_dual_sse2(p, q, ps, qs, &mask, &thresh0, bd, &t80);
+}
+
+void aom_highbd_lpf_horizontal_4_sse2(uint16_t *s, int p,
+                                      const uint8_t *_blimit,
+                                      const uint8_t *_limit,
+                                      const uint8_t *_thresh, int bd) {
+  __m128i p1p0, q1q0;
+  __m128i p1 = _mm_loadl_epi64((__m128i *)(s - 2 * p));
+  __m128i p0 = _mm_loadl_epi64((__m128i *)(s - 1 * p));
+  __m128i q0 = _mm_loadl_epi64((__m128i *)(s - 0 * p));
+  __m128i q1 = _mm_loadl_epi64((__m128i *)(s + 1 * p));
+
+  highbd_lpf_internal_4_sse2(&p1, &p0, &q0, &q1, &q1q0, &p1p0, _blimit, _limit,
+                             _thresh, bd);
+
+  _mm_storel_epi64((__m128i *)(s - 2 * p), _mm_srli_si128(p1p0, 8));
+  _mm_storel_epi64((__m128i *)(s - 1 * p), p1p0);
+  _mm_storel_epi64((__m128i *)(s + 0 * p), q1q0);
+  _mm_storel_epi64((__m128i *)(s + 1 * p), _mm_srli_si128(q1q0, 8));
+}
+
+void aom_highbd_lpf_horizontal_4_dual_sse2(
+    uint16_t *s, int p, const uint8_t *_blimit0, const uint8_t *_limit0,
+    const uint8_t *_thresh0, const uint8_t *_blimit1, const uint8_t *_limit1,
+    const uint8_t *_thresh1, int bd) {
+  __m128i p1 = _mm_loadu_si128((__m128i *)(s - 2 * p));
+  __m128i p0 = _mm_loadu_si128((__m128i *)(s - 1 * p));
+  __m128i q0 = _mm_loadu_si128((__m128i *)(s - 0 * p));
+  __m128i q1 = _mm_loadu_si128((__m128i *)(s + 1 * p));
+  __m128i ps[2], qs[2];
+
+  highbd_lpf_internal_4_dual_sse2(&p1, &p0, &q0, &q1, ps, qs, _blimit0, _limit0,
+                                  _thresh0, _blimit1, _limit1, _thresh1, bd);
+
+  _mm_storeu_si128((__m128i *)(s - 2 * p), ps[1]);
+  _mm_storeu_si128((__m128i *)(s - 1 * p), ps[0]);
+  _mm_storeu_si128((__m128i *)(s + 0 * p), qs[0]);
+  _mm_storeu_si128((__m128i *)(s + 1 * p), qs[1]);
+}
+
+void aom_highbd_lpf_vertical_4_sse2(uint16_t *s, int p, const uint8_t *blimit,
+                                    const uint8_t *limit, const uint8_t *thresh,
+                                    int bd) {
+  __m128i x0, x1, x2, x3, d0, d1, d2, d3;
+  __m128i p1p0, q1q0;
+  __m128i p1, q1;
+
+  x0 = _mm_loadl_epi64((__m128i *)(s - 2 + 0 * p));
+  x1 = _mm_loadl_epi64((__m128i *)(s - 2 + 1 * p));
+  x2 = _mm_loadl_epi64((__m128i *)(s - 2 + 2 * p));
+  x3 = _mm_loadl_epi64((__m128i *)(s - 2 + 3 * p));
+
+  highbd_transpose4x8_8x4_low_sse2(&x0, &x1, &x2, &x3, &d0, &d1, &d2, &d3);
+
+  highbd_lpf_internal_4_sse2(&d0, &d1, &d2, &d3, &q1q0, &p1p0, blimit, limit,
+                             thresh, bd);
+
+  p1 = _mm_srli_si128(p1p0, 8);
+  q1 = _mm_srli_si128(q1q0, 8);
+
+  // transpose from 8x4 to 4x8
+  highbd_transpose4x8_8x4_low_sse2(&p1, &p1p0, &q1q0, &q1, &d0, &d1, &d2, &d3);
+
+  _mm_storel_epi64((__m128i *)(s - 2 + 0 * p), d0);
+  _mm_storel_epi64((__m128i *)(s - 2 + 1 * p), d1);
+  _mm_storel_epi64((__m128i *)(s - 2 + 2 * p), d2);
+  _mm_storel_epi64((__m128i *)(s - 2 + 3 * p), d3);
+}
+
+void aom_highbd_lpf_vertical_4_dual_sse2(
+    uint16_t *s, int p, const uint8_t *blimit0, const uint8_t *limit0,
+    const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1,
+    const uint8_t *thresh1, int bd) {
+  __m128i x0, x1, x2, x3, x4, x5, x6, x7;
+  __m128i d0, d1, d2, d3, d4, d5, d6, d7;
+  __m128i ps[2], qs[2];
+
+  x0 = _mm_loadl_epi64((__m128i *)(s - 2 + 0 * p));
+  x1 = _mm_loadl_epi64((__m128i *)(s - 2 + 1 * p));
+  x2 = _mm_loadl_epi64((__m128i *)(s - 2 + 2 * p));
+  x3 = _mm_loadl_epi64((__m128i *)(s - 2 + 3 * p));
+  x4 = _mm_loadl_epi64((__m128i *)(s - 2 + 4 * p));
+  x5 = _mm_loadl_epi64((__m128i *)(s - 2 + 5 * p));
+  x6 = _mm_loadl_epi64((__m128i *)(s - 2 + 6 * p));
+  x7 = _mm_loadl_epi64((__m128i *)(s - 2 + 7 * p));
+
+  highbd_transpose8x8_low_sse2(&x0, &x1, &x2, &x3, &x4, &x5, &x6, &x7, &d0, &d1,
+                               &d2, &d3);
+
+  highbd_lpf_internal_4_dual_sse2(&d0, &d1, &d2, &d3, ps, qs, blimit0, limit0,
+                                  thresh0, blimit1, limit1, thresh1, bd);
+
+  highbd_transpose4x8_8x4_sse2(&ps[1], &ps[0], &qs[0], &qs[1], &d0, &d1, &d2,
+                               &d3, &d4, &d5, &d6, &d7);
+
+  _mm_storel_epi64((__m128i *)(s - 2 + 0 * p), d0);
+  _mm_storel_epi64((__m128i *)(s - 2 + 1 * p), d1);
+  _mm_storel_epi64((__m128i *)(s - 2 + 2 * p), d2);
+  _mm_storel_epi64((__m128i *)(s - 2 + 3 * p), d3);
+  _mm_storel_epi64((__m128i *)(s - 2 + 4 * p), d4);
+  _mm_storel_epi64((__m128i *)(s - 2 + 5 * p), d5);
+  _mm_storel_epi64((__m128i *)(s - 2 + 6 * p), d6);
+  _mm_storel_epi64((__m128i *)(s - 2 + 7 * p), d7);
+}
+
+void aom_highbd_lpf_vertical_6_sse2(uint16_t *s, int p, const uint8_t *blimit,
+                                    const uint8_t *limit, const uint8_t *thresh,
+                                    int bd) {
+  __m128i d0, d1, d2, d3, d4, d5, d6, d7;
+  __m128i x3, x2, x1, x0, p0, q0;
+  __m128i p1p0, q1q0;
+
+  x3 = _mm_loadu_si128((__m128i *)((s - 3) + 0 * p));
+  x2 = _mm_loadu_si128((__m128i *)((s - 3) + 1 * p));
+  x1 = _mm_loadu_si128((__m128i *)((s - 3) + 2 * p));
+  x0 = _mm_loadu_si128((__m128i *)((s - 3) + 3 * p));
+
+  highbd_transpose4x8_8x4_sse2(&x3, &x2, &x1, &x0, &d0, &d1, &d2, &d3, &d4, &d5,
+                               &d6, &d7);
+
+  highbd_lpf_internal_6_sse2(&d0, &d1, &d2, &d3, &d4, &d5, &p1p0, &q1q0, blimit,
+                             limit, thresh, bd);
+
+  p0 = _mm_srli_si128(p1p0, 8);
+  q0 = _mm_srli_si128(q1q0, 8);
+
+  highbd_transpose4x8_8x4_low_sse2(&p0, &p1p0, &q1q0, &q0, &d0, &d1, &d2, &d3);
+
+  _mm_storel_epi64((__m128i *)(s - 2 + 0 * p), d0);
+  _mm_storel_epi64((__m128i *)(s - 2 + 1 * p), d1);
+  _mm_storel_epi64((__m128i *)(s - 2 + 2 * p), d2);
+  _mm_storel_epi64((__m128i *)(s - 2 + 3 * p), d3);
+}
+
+void aom_highbd_lpf_vertical_6_dual_sse2(
+    uint16_t *s, int p, const uint8_t *_blimit0, const uint8_t *_limit0,
+    const uint8_t *_thresh0, const uint8_t *_blimit1, const uint8_t *_limit1,
+    const uint8_t *_thresh1, int bd) {
+  __m128i d0, d1, d2, d3, d4, d5, d6, d7;
+  __m128i x0, x1, x2, x3, x4, x5, x6, x7;
+  __m128i p0, q0, p1, q1, p2, q2;
+
+  x0 = _mm_loadu_si128((__m128i *)((s - 3) + 0 * p));
+  x1 = _mm_loadu_si128((__m128i *)((s - 3) + 1 * p));
+  x2 = _mm_loadu_si128((__m128i *)((s - 3) + 2 * p));
+  x3 = _mm_loadu_si128((__m128i *)((s - 3) + 3 * p));
+  x4 = _mm_loadu_si128((__m128i *)((s - 3) + 4 * p));
+  x5 = _mm_loadu_si128((__m128i *)((s - 3) + 5 * p));
+  x6 = _mm_loadu_si128((__m128i *)((s - 3) + 6 * p));
+  x7 = _mm_loadu_si128((__m128i *)((s - 3) + 7 * p));
+
+  highbd_transpose8x8_sse2(&x0, &x1, &x2, &x3, &x4, &x5, &x6, &x7, &p2, &p1,
+                           &p0, &q0, &q1, &q2, &d6, &d7);
+
+  highbd_lpf_internal_6_dual_sse2(&p2, &p1, &p0, &q0, &q1, &q2, _blimit0,
+                                  _limit0, _thresh0, _blimit1, _limit1,
+                                  _thresh1, bd);
+
+  highbd_transpose4x8_8x4_sse2(&p1, &p0, &q0, &q1, &d0, &d1, &d2, &d3, &d4, &d5,
+                               &d6, &d7);
+
+  _mm_storel_epi64((__m128i *)(s - 2 + 0 * p), d0);
+  _mm_storel_epi64((__m128i *)(s - 2 + 1 * p), d1);
+  _mm_storel_epi64((__m128i *)(s - 2 + 2 * p), d2);
+  _mm_storel_epi64((__m128i *)(s - 2 + 3 * p), d3);
+  _mm_storel_epi64((__m128i *)(s - 2 + 4 * p), d4);
+  _mm_storel_epi64((__m128i *)(s - 2 + 5 * p), d5);
+  _mm_storel_epi64((__m128i *)(s - 2 + 6 * p), d6);
+  _mm_storel_epi64((__m128i *)(s - 2 + 7 * p), d7);
+}
+
+void aom_highbd_lpf_vertical_8_sse2(uint16_t *s, int p, const uint8_t *blimit,
+                                    const uint8_t *limit, const uint8_t *thresh,
+                                    int bd) {
+  __m128i d0, d1, d2, d3, d4, d5, d6, d7;
+  __m128i p2, p1, p0, p3, q0;
+  __m128i q1q0, p1p0;
+
+  p3 = _mm_loadu_si128((__m128i *)((s - 4) + 0 * p));
+  p2 = _mm_loadu_si128((__m128i *)((s - 4) + 1 * p));
+  p1 = _mm_loadu_si128((__m128i *)((s - 4) + 2 * p));
+  p0 = _mm_loadu_si128((__m128i *)((s - 4) + 3 * p));
+
+  highbd_transpose4x8_8x4_sse2(&p3, &p2, &p1, &p0, &d0, &d1, &d2, &d3, &d4, &d5,
+                               &d6, &d7);
+
+  // Loop filtering
+  highbd_lpf_internal_8_sse2(&d0, &d7, &d1, &d6, &d2, &d5, &d3, &d4, &q1q0,
+                             &p1p0, blimit, limit, thresh, bd);
+
+  p0 = _mm_srli_si128(p1p0, 8);
+  q0 = _mm_srli_si128(q1q0, 8);
+
+  highbd_transpose8x8_low_sse2(&d0, &d1, &p0, &p1p0, &q1q0, &q0, &d6, &d7, &d0,
+                               &d1, &d2, &d3);
+
+  _mm_storeu_si128((__m128i *)(s - 4 + 0 * p), d0);
+  _mm_storeu_si128((__m128i *)(s - 4 + 1 * p), d1);
+  _mm_storeu_si128((__m128i *)(s - 4 + 2 * p), d2);
+  _mm_storeu_si128((__m128i *)(s - 4 + 3 * p), d3);
+}
+
+void aom_highbd_lpf_vertical_8_dual_sse2(
+    uint16_t *s, int p, const uint8_t *blimit0, const uint8_t *limit0,
+    const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1,
+    const uint8_t *thresh1, int bd) {
+  __m128i x0, x1, x2, x3, x4, x5, x6, x7;
+  __m128i d0, d1, d2, d3, d4, d5, d6, d7;
+
+  x0 = _mm_loadu_si128((__m128i *)(s - 4 + 0 * p));
+  x1 = _mm_loadu_si128((__m128i *)(s - 4 + 1 * p));
+  x2 = _mm_loadu_si128((__m128i *)(s - 4 + 2 * p));
+  x3 = _mm_loadu_si128((__m128i *)(s - 4 + 3 * p));
+  x4 = _mm_loadu_si128((__m128i *)(s - 4 + 4 * p));
+  x5 = _mm_loadu_si128((__m128i *)(s - 4 + 5 * p));
+  x6 = _mm_loadu_si128((__m128i *)(s - 4 + 6 * p));
+  x7 = _mm_loadu_si128((__m128i *)(s - 4 + 7 * p));
+
+  highbd_transpose8x8_sse2(&x0, &x1, &x2, &x3, &x4, &x5, &x6, &x7, &d0, &d1,
+                           &d2, &d3, &d4, &d5, &d6, &d7);
+
+  highbd_lpf_internal_8_dual_sse2(&d0, &d7, &d1, &d6, &d2, &d5, &d3, &d4,
+                                  blimit0, limit0, thresh0, blimit1, limit1,
+                                  thresh1, bd);
+
+  highbd_transpose8x8_sse2(&d0, &d1, &d2, &d3, &d4, &d5, &d6, &d7, &x0, &x1,
+                           &x2, &x3, &x4, &x5, &x6, &x7);
+
+  _mm_storeu_si128((__m128i *)(s - 4 + 0 * p), x0);
+  _mm_storeu_si128((__m128i *)(s - 4 + 1 * p), x1);
+  _mm_storeu_si128((__m128i *)(s - 4 + 2 * p), x2);
+  _mm_storeu_si128((__m128i *)(s - 4 + 3 * p), x3);
+  _mm_storeu_si128((__m128i *)(s - 4 + 4 * p), x4);
+  _mm_storeu_si128((__m128i *)(s - 4 + 5 * p), x5);
+  _mm_storeu_si128((__m128i *)(s - 4 + 6 * p), x6);
+  _mm_storeu_si128((__m128i *)(s - 4 + 7 * p), x7);
+}
+
+void aom_highbd_lpf_vertical_14_sse2(uint16_t *s, int pitch,
+                                     const uint8_t *blimit,
+                                     const uint8_t *limit,
+                                     const uint8_t *thresh, int bd) {
+  __m128i q[7], p[7], pq[7];
+  __m128i p6, p5, p4, p3;
+  __m128i p6_2, p5_2, p4_2, p3_2;
+  __m128i d0, d1, d2, d3;
+  __m128i d0_2, d1_2, d2_2, d3_2, d7_2;
+
+  p6 = _mm_loadu_si128((__m128i *)((s - 8) + 0 * pitch));
+  p5 = _mm_loadu_si128((__m128i *)((s - 8) + 1 * pitch));
+  p4 = _mm_loadu_si128((__m128i *)((s - 8) + 2 * pitch));
+  p3 = _mm_loadu_si128((__m128i *)((s - 8) + 3 * pitch));
+
+  highbd_transpose4x8_8x4_sse2(&p6, &p5, &p4, &p3, &d0, &p[6], &p[5], &p[4],
+                               &p[3], &p[2], &p[1], &p[0]);
+
+  p6_2 = _mm_loadu_si128((__m128i *)(s + 0 * pitch));
+  p5_2 = _mm_loadu_si128((__m128i *)(s + 1 * pitch));
+  p4_2 = _mm_loadu_si128((__m128i *)(s + 2 * pitch));
+  p3_2 = _mm_loadu_si128((__m128i *)(s + 3 * pitch));
+
+  highbd_transpose4x8_8x4_sse2(&p6_2, &p5_2, &p4_2, &p3_2, &q[0], &q[1], &q[2],
+                               &q[3], &q[4], &q[5], &q[6], &d7_2);
+
+  highbd_lpf_internal_14_sse2(p, q, pq, blimit, limit, thresh, bd);
+
+  highbd_transpose8x8_low_sse2(&d0, &p[6], &pq[5], &pq[4], &pq[3], &pq[2],
+                               &pq[1], &pq[0], &d0, &d1, &d2, &d3);
+
+  q[0] = _mm_srli_si128(pq[0], 8);
+  q[1] = _mm_srli_si128(pq[1], 8);
+  q[2] = _mm_srli_si128(pq[2], 8);
+  q[3] = _mm_srli_si128(pq[3], 8);
+  q[4] = _mm_srli_si128(pq[4], 8);
+  q[5] = _mm_srli_si128(pq[5], 8);
+
+  highbd_transpose8x8_low_sse2(&q[0], &q[1], &q[2], &q[3], &q[4], &q[5], &q[6],
+                               &d7_2, &d0_2, &d1_2, &d2_2, &d3_2);
+
+  _mm_storeu_si128((__m128i *)(s - 8 + 0 * pitch), d0);
+  _mm_storeu_si128((__m128i *)(s + 0 * pitch), d0_2);
+
+  _mm_storeu_si128((__m128i *)(s - 8 + 1 * pitch), d1);
+  _mm_storeu_si128((__m128i *)(s + 1 * pitch), d1_2);
+
+  _mm_storeu_si128((__m128i *)(s - 8 + 2 * pitch), d2);
+  _mm_storeu_si128((__m128i *)(s + 2 * pitch), d2_2);
+
+  _mm_storeu_si128((__m128i *)(s - 8 + 3 * pitch), d3);
+  _mm_storeu_si128((__m128i *)(s + 3 * pitch), d3_2);
+}
+
+void aom_highbd_lpf_vertical_14_dual_sse2(
+    uint16_t *s, int pitch, const uint8_t *blimit0, const uint8_t *limit0,
+    const uint8_t *thresh0, const uint8_t *blimit1, const uint8_t *limit1,
+    const uint8_t *thresh1, int bd) {
+  __m128i q[7], p[7];
+  __m128i p6, p5, p4, p3, p2, p1, p0, q0;
+  __m128i p6_2, p5_2, p4_2, p3_2, p2_2, p1_2, q0_2, p0_2;
+  __m128i d0, d7;
+  __m128i d0_out, d1_out, d2_out, d3_out, d4_out, d5_out, d6_out, d7_out;
+
+  p6 = _mm_loadu_si128((__m128i *)((s - 8) + 0 * pitch));
+  p5 = _mm_loadu_si128((__m128i *)((s - 8) + 1 * pitch));
+  p4 = _mm_loadu_si128((__m128i *)((s - 8) + 2 * pitch));
+  p3 = _mm_loadu_si128((__m128i *)((s - 8) + 3 * pitch));
+  p2 = _mm_loadu_si128((__m128i *)((s - 8) + 4 * pitch));
+  p1 = _mm_loadu_si128((__m128i *)((s - 8) + 5 * pitch));
+  p0 = _mm_loadu_si128((__m128i *)((s - 8) + 6 * pitch));
+  q0 = _mm_loadu_si128((__m128i *)((s - 8) + 7 * pitch));
+
+  highbd_transpose8x8_sse2(&p6, &p5, &p4, &p3, &p2, &p1, &p0, &q0, &d0, &p[6],
+                           &p[5], &p[4], &p[3], &p[2], &p[1], &p[0]);
+
+  p6_2 = _mm_loadu_si128((__m128i *)(s + 0 * pitch));
+  p5_2 = _mm_loadu_si128((__m128i *)(s + 1 * pitch));
+  p4_2 = _mm_loadu_si128((__m128i *)(s + 2 * pitch));
+  p3_2 = _mm_loadu_si128((__m128i *)(s + 3 * pitch));
+  p2_2 = _mm_loadu_si128((__m128i *)(s + 4 * pitch));
+  p1_2 = _mm_loadu_si128((__m128i *)(s + 5 * pitch));
+  p0_2 = _mm_loadu_si128((__m128i *)(s + 6 * pitch));
+  q0_2 = _mm_loadu_si128((__m128i *)(s + 7 * pitch));
+
+  highbd_transpose8x8_sse2(&p6_2, &p5_2, &p4_2, &p3_2, &p2_2, &p1_2, &p0_2,
+                           &q0_2, &q[0], &q[1], &q[2], &q[3], &q[4], &q[5],
+                           &q[6], &d7);
+
+  highbd_lpf_internal_14_dual_sse2(p, q, blimit0, limit0, thresh0, blimit1,
+                                   limit1, thresh1, bd);
+
+  highbd_transpose8x8_sse2(&d0, &p[6], &p[5], &p[4], &p[3], &p[2], &p[1], &p[0],
+                           &d0_out, &d1_out, &d2_out, &d3_out, &d4_out, &d5_out,
+                           &d6_out, &d7_out);
+
+  _mm_storeu_si128((__m128i *)(s - 8 + 0 * pitch), d0_out);
+  _mm_storeu_si128((__m128i *)(s - 8 + 1 * pitch), d1_out);
+  _mm_storeu_si128((__m128i *)(s - 8 + 2 * pitch), d2_out);
+  _mm_storeu_si128((__m128i *)(s - 8 + 3 * pitch), d3_out);
+  _mm_storeu_si128((__m128i *)(s - 8 + 4 * pitch), d4_out);
+  _mm_storeu_si128((__m128i *)(s - 8 + 5 * pitch), d5_out);
+  _mm_storeu_si128((__m128i *)(s - 8 + 6 * pitch), d6_out);
+  _mm_storeu_si128((__m128i *)(s - 8 + 7 * pitch), d7_out);
+
+  highbd_transpose8x8_sse2(&q[0], &q[1], &q[2], &q[3], &q[4], &q[5], &q[6], &d7,
+                           &d0_out, &d1_out, &d2_out, &d3_out, &d4_out, &d5_out,
+                           &d6_out, &d7_out);
+
+  _mm_storeu_si128((__m128i *)(s + 0 * pitch), d0_out);
+  _mm_storeu_si128((__m128i *)(s + 1 * pitch), d1_out);
+  _mm_storeu_si128((__m128i *)(s + 2 * pitch), d2_out);
+  _mm_storeu_si128((__m128i *)(s + 3 * pitch), d3_out);
+  _mm_storeu_si128((__m128i *)(s + 4 * pitch), d4_out);
+  _mm_storeu_si128((__m128i *)(s + 5 * pitch), d5_out);
+  _mm_storeu_si128((__m128i *)(s + 6 * pitch), d6_out);
+  _mm_storeu_si128((__m128i *)(s + 7 * pitch), d7_out);
+}
diff --git a/third_party/aom/aom_dsp/x86/highbd_quantize_intrin_avx2.c b/third_party/aom/aom_dsp/x86/highbd_quantize_intrin_avx2.c
new file mode 100644
index 0000000000..950465cf46
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/highbd_quantize_intrin_avx2.c
@@ -0,0 +1,294 @@
+/*
+ * Copyright (c) 2017, 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"
+
+#include "aom/aom_integer.h"
+
+static INLINE void init_one_qp(const __m128i *p, __m256i *qp) {
+  const __m128i sign = _mm_srai_epi16(*p, 15);
+  const __m128i dc = _mm_unpacklo_epi16(*p, sign);
+  const __m128i ac = _mm_unpackhi_epi16(*p, sign);
+  *qp = _mm256_insertf128_si256(_mm256_castsi128_si256(dc), ac, 1);
+}
+
+static INLINE void update_qp(__m256i *qp) {
+  int i;
+  for (i = 0; i < 5; ++i) {
+    qp[i] = _mm256_permute2x128_si256(qp[i], qp[i], 0x11);
+  }
+}
+
+static INLINE void init_qp(const int16_t *zbin_ptr, const int16_t *round_ptr,
+                           const int16_t *quant_ptr, const int16_t *dequant_ptr,
+                           const int16_t *quant_shift_ptr, __m256i *qp,
+                           int log_scale) {
+  const __m128i zbin = _mm_loadu_si128((const __m128i *)zbin_ptr);
+  const __m128i round = _mm_loadu_si128((const __m128i *)round_ptr);
+  const __m128i quant = _mm_loadu_si128((const __m128i *)quant_ptr);
+  const __m128i dequant = _mm_loadu_si128((const __m128i *)dequant_ptr);
+  const __m128i quant_shift = _mm_loadu_si128((const __m128i *)quant_shift_ptr);
+  init_one_qp(&zbin, &qp[0]);
+  init_one_qp(&round, &qp[1]);
+  init_one_qp(&quant, &qp[2]);
+  init_one_qp(&dequant, &qp[3]);
+  init_one_qp(&quant_shift, &qp[4]);
+  if (log_scale > 0) {
+    const __m256i rnd = _mm256_set1_epi32((int16_t)(1 << (log_scale - 1)));
+    qp[0] = _mm256_add_epi32(qp[0], rnd);
+    qp[0] = _mm256_srai_epi32(qp[0], log_scale);
+
+    qp[1] = _mm256_add_epi32(qp[1], rnd);
+    qp[1] = _mm256_srai_epi32(qp[1], log_scale);
+  }
+  // Subtracting 1 here eliminates a _mm256_cmpeq_epi32() instruction when
+  // calculating the zbin mask.
+  qp[0] = _mm256_sub_epi32(qp[0], _mm256_set1_epi32(1));
+}
+
+// Note:
+// *x is vector multiplied by *y which is 16 int32_t parallel multiplication
+// and right shift 16.  The output, 16 int32_t is save in *p.
+static INLINE __m256i mm256_mul_shift_epi32(const __m256i *x,
+                                            const __m256i *y) {
+  __m256i prod_lo = _mm256_mul_epi32(*x, *y);
+  __m256i prod_hi = _mm256_srli_epi64(*x, 32);
+  const __m256i mult_hi = _mm256_srli_epi64(*y, 32);
+  prod_hi = _mm256_mul_epi32(prod_hi, mult_hi);
+
+  prod_lo = _mm256_srli_epi64(prod_lo, 16);
+  const __m256i mask = _mm256_set_epi32(0, -1, 0, -1, 0, -1, 0, -1);
+  prod_lo = _mm256_and_si256(prod_lo, mask);
+  prod_hi = _mm256_srli_epi64(prod_hi, 16);
+
+  prod_hi = _mm256_slli_epi64(prod_hi, 32);
+  return _mm256_or_si256(prod_lo, prod_hi);
+}
+
+static AOM_FORCE_INLINE __m256i get_max_lane_eob(const int16_t *iscan_ptr,
+                                                 __m256i eobmax,
+                                                 __m256i nz_mask) {
+  const __m256i packed_nz_mask = _mm256_packs_epi32(nz_mask, nz_mask);
+  const __m256i packed_nz_mask_perm =
+      _mm256_permute4x64_epi64(packed_nz_mask, 0xD8);
+  const __m256i iscan =
+      _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)iscan_ptr));
+  const __m256i iscan_plus1 = _mm256_sub_epi16(iscan, packed_nz_mask_perm);
+  const __m256i nz_iscan = _mm256_and_si256(iscan_plus1, packed_nz_mask_perm);
+  return _mm256_max_epi16(eobmax, nz_iscan);
+}
+
+// Get the max eob from the lower 128 bits.
+static AOM_FORCE_INLINE uint16_t get_max_eob(__m256i eob) {
+  __m256i eob_s;
+  eob_s = _mm256_shuffle_epi32(eob, 0xe);
+  eob = _mm256_max_epi16(eob, eob_s);
+  eob_s = _mm256_shufflelo_epi16(eob, 0xe);
+  eob = _mm256_max_epi16(eob, eob_s);
+  eob_s = _mm256_shufflelo_epi16(eob, 1);
+  eob = _mm256_max_epi16(eob, eob_s);
+  return (uint16_t)_mm256_extract_epi16(eob, 0);
+}
+
+static AOM_FORCE_INLINE __m256i mm256_mul_shift_epi32_logscale(const __m256i *x,
+                                                               const __m256i *y,
+                                                               int log_scale) {
+  __m256i prod_lo = _mm256_mul_epi32(*x, *y);
+  __m256i prod_hi = _mm256_srli_epi64(*x, 32);
+  const __m256i mult_hi = _mm256_srli_epi64(*y, 32);
+  prod_hi = _mm256_mul_epi32(prod_hi, mult_hi);
+  prod_lo = _mm256_srli_epi64(prod_lo, 16 - log_scale);
+  const __m256i mask = _mm256_set_epi32(0, -1, 0, -1, 0, -1, 0, -1);
+  prod_lo = _mm256_and_si256(prod_lo, mask);
+  prod_hi = _mm256_srli_epi64(prod_hi, 16 - log_scale);
+  prod_hi = _mm256_slli_epi64(prod_hi, 32);
+  return _mm256_or_si256(prod_lo, prod_hi);
+}
+
+static AOM_FORCE_INLINE void quantize_logscale(
+    const __m256i *qp, const tran_low_t *coeff_ptr, const int16_t *iscan_ptr,
+    tran_low_t *qcoeff, tran_low_t *dqcoeff, __m256i *eob, int log_scale) {
+  const __m256i coeff = _mm256_loadu_si256((const __m256i *)coeff_ptr);
+  const __m256i abs_coeff = _mm256_abs_epi32(coeff);
+  const __m256i zbin_mask = _mm256_cmpgt_epi32(abs_coeff, qp[0]);
+
+  if (UNLIKELY(_mm256_movemask_epi8(zbin_mask) == 0)) {
+    const __m256i zero = _mm256_setzero_si256();
+    _mm256_storeu_si256((__m256i *)qcoeff, zero);
+    _mm256_storeu_si256((__m256i *)dqcoeff, zero);
+    return;
+  }
+
+  const __m256i tmp_rnd =
+      _mm256_and_si256(_mm256_add_epi32(abs_coeff, qp[1]), zbin_mask);
+  // const int64_t tmp2 = ((tmpw * quant_ptr[rc != 0]) >> 16) + tmpw;
+  const __m256i tmp = mm256_mul_shift_epi32_logscale(&tmp_rnd, &qp[2], 0);
+  const __m256i tmp2 = _mm256_add_epi32(tmp, tmp_rnd);
+  // const int abs_qcoeff = (int)((tmp2 * quant_shift_ptr[rc != 0]) >>
+  //                              (16 - log_scale + AOM_QM_BITS));
+  const __m256i abs_q =
+      mm256_mul_shift_epi32_logscale(&tmp2, &qp[4], log_scale);
+  const __m256i abs_dq =
+      _mm256_srli_epi32(_mm256_mullo_epi32(abs_q, qp[3]), log_scale);
+  const __m256i nz_mask = _mm256_cmpgt_epi32(abs_q, _mm256_setzero_si256());
+  const __m256i q = _mm256_sign_epi32(abs_q, coeff);
+  const __m256i dq = _mm256_sign_epi32(abs_dq, coeff);
+
+  _mm256_storeu_si256((__m256i *)qcoeff, q);
+  _mm256_storeu_si256((__m256i *)dqcoeff, dq);
+
+  *eob = get_max_lane_eob(iscan_ptr, *eob, nz_mask);
+}
+
+static AOM_FORCE_INLINE void quantize(const __m256i *qp,
+                                      const tran_low_t *coeff_ptr,
+                                      const int16_t *iscan_ptr,
+                                      tran_low_t *qcoeff, tran_low_t *dqcoeff,
+                                      __m256i *eob) {
+  const __m256i coeff = _mm256_loadu_si256((const __m256i *)coeff_ptr);
+  const __m256i abs_coeff = _mm256_abs_epi32(coeff);
+  const __m256i zbin_mask = _mm256_cmpgt_epi32(abs_coeff, qp[0]);
+
+  if (UNLIKELY(_mm256_movemask_epi8(zbin_mask) == 0)) {
+    const __m256i zero = _mm256_setzero_si256();
+    _mm256_storeu_si256((__m256i *)qcoeff, zero);
+    _mm256_storeu_si256((__m256i *)dqcoeff, zero);
+    return;
+  }
+
+  const __m256i tmp_rnd =
+      _mm256_and_si256(_mm256_add_epi32(abs_coeff, qp[1]), zbin_mask);
+  const __m256i tmp = mm256_mul_shift_epi32(&tmp_rnd, &qp[2]);
+  const __m256i tmp2 = _mm256_add_epi32(tmp, tmp_rnd);
+  const __m256i abs_q = mm256_mul_shift_epi32(&tmp2, &qp[4]);
+  const __m256i abs_dq = _mm256_mullo_epi32(abs_q, qp[3]);
+  const __m256i nz_mask = _mm256_cmpgt_epi32(abs_q, _mm256_setzero_si256());
+  const __m256i q = _mm256_sign_epi32(abs_q, coeff);
+  const __m256i dq = _mm256_sign_epi32(abs_dq, coeff);
+
+  _mm256_storeu_si256((__m256i *)qcoeff, q);
+  _mm256_storeu_si256((__m256i *)dqcoeff, dq);
+
+  *eob = get_max_lane_eob(iscan_ptr, *eob, nz_mask);
+}
+
+void aom_highbd_quantize_b_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                                const int16_t *zbin_ptr,
+                                const int16_t *round_ptr,
+                                const int16_t *quant_ptr,
+                                const int16_t *quant_shift_ptr,
+                                tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                                const int16_t *dequant_ptr, uint16_t *eob_ptr,
+                                const int16_t *scan, const int16_t *iscan) {
+  (void)scan;
+  const int step = 8;
+
+  __m256i eob = _mm256_setzero_si256();
+  __m256i qp[5];
+
+  init_qp(zbin_ptr, round_ptr, quant_ptr, dequant_ptr, quant_shift_ptr, qp, 0);
+
+  quantize(qp, coeff_ptr, iscan, qcoeff_ptr, dqcoeff_ptr, &eob);
+
+  coeff_ptr += step;
+  qcoeff_ptr += step;
+  dqcoeff_ptr += step;
+  iscan += step;
+  n_coeffs -= step;
+
+  update_qp(qp);
+
+  while (n_coeffs > 0) {
+    quantize(qp, coeff_ptr, iscan, qcoeff_ptr, dqcoeff_ptr, &eob);
+
+    coeff_ptr += step;
+    qcoeff_ptr += step;
+    dqcoeff_ptr += step;
+    iscan += step;
+    n_coeffs -= step;
+  }
+
+  *eob_ptr = get_max_eob(eob);
+}
+
+void aom_highbd_quantize_b_32x32_avx2(
+    const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr,
+    const int16_t *round_ptr, const int16_t *quant_ptr,
+    const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
+    tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
+    const int16_t *scan, const int16_t *iscan) {
+  (void)scan;
+  const unsigned int step = 8;
+
+  __m256i eob = _mm256_setzero_si256();
+  __m256i qp[5];
+  init_qp(zbin_ptr, round_ptr, quant_ptr, dequant_ptr, quant_shift_ptr, qp, 1);
+
+  quantize_logscale(qp, coeff_ptr, iscan, qcoeff_ptr, dqcoeff_ptr, &eob, 1);
+
+  coeff_ptr += step;
+  qcoeff_ptr += step;
+  dqcoeff_ptr += step;
+  iscan += step;
+  n_coeffs -= step;
+
+  update_qp(qp);
+
+  while (n_coeffs > 0) {
+    quantize_logscale(qp, coeff_ptr, iscan, qcoeff_ptr, dqcoeff_ptr, &eob, 1);
+
+    coeff_ptr += step;
+    qcoeff_ptr += step;
+    dqcoeff_ptr += step;
+    iscan += step;
+    n_coeffs -= step;
+  }
+
+  *eob_ptr = get_max_eob(eob);
+}
+
+void aom_highbd_quantize_b_64x64_avx2(
+    const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr,
+    const int16_t *round_ptr, const int16_t *quant_ptr,
+    const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
+    tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
+    const int16_t *scan, const int16_t *iscan) {
+  (void)scan;
+  const int step = 8;
+
+  __m256i eob = _mm256_setzero_si256();
+  __m256i qp[5];
+  init_qp(zbin_ptr, round_ptr, quant_ptr, dequant_ptr, quant_shift_ptr, qp, 2);
+
+  quantize_logscale(qp, coeff_ptr, iscan, qcoeff_ptr, dqcoeff_ptr, &eob, 2);
+
+  coeff_ptr += step;
+  qcoeff_ptr += step;
+  dqcoeff_ptr += step;
+  iscan += step;
+  n_coeffs -= step;
+
+  update_qp(qp);
+
+  while (n_coeffs > 0) {
+    quantize_logscale(qp, coeff_ptr, iscan, qcoeff_ptr, dqcoeff_ptr, &eob, 2);
+
+    coeff_ptr += step;
+    qcoeff_ptr += step;
+    dqcoeff_ptr += step;
+    iscan += step;
+    n_coeffs -= step;
+  }
+
+  *eob_ptr = get_max_eob(eob);
+}
diff --git a/third_party/aom/aom_dsp/x86/highbd_quantize_intrin_sse2.c b/third_party/aom/aom_dsp/x86/highbd_quantize_intrin_sse2.c
new file mode 100644
index 0000000000..3b0c42c4f5
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/highbd_quantize_intrin_sse2.c
@@ -0,0 +1,208 @@
+/*
+ * Copyright (c) 2016, 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 <emmintrin.h>
+
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_mem/aom_mem.h"
+#include "aom_ports/mem.h"
+#include "config/aom_dsp_rtcd.h"
+
+void aom_highbd_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t count,
+                                const int16_t *zbin_ptr,
+                                const int16_t *round_ptr,
+                                const int16_t *quant_ptr,
+                                const int16_t *quant_shift_ptr,
+                                tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                                const int16_t *dequant_ptr, uint16_t *eob_ptr,
+                                const int16_t *scan, const int16_t *iscan) {
+  int i, j, non_zero_regs = (int)count / 4, eob_i = -1;
+  __m128i zbins[2];
+  __m128i nzbins[2];
+
+  zbins[0] = _mm_set_epi32((int)zbin_ptr[1], (int)zbin_ptr[1], (int)zbin_ptr[1],
+                           (int)zbin_ptr[0]);
+  zbins[1] = _mm_set1_epi32((int)zbin_ptr[1]);
+
+  nzbins[0] = _mm_setzero_si128();
+  nzbins[1] = _mm_setzero_si128();
+  nzbins[0] = _mm_sub_epi32(nzbins[0], zbins[0]);
+  nzbins[1] = _mm_sub_epi32(nzbins[1], zbins[1]);
+
+  (void)scan;
+
+  memset(qcoeff_ptr, 0, count * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, count * sizeof(*dqcoeff_ptr));
+
+  // Pre-scan pass
+  for (i = ((int)count / 4) - 1; i >= 0; i--) {
+    __m128i coeffs, cmp1, cmp2;
+    int test;
+    coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4));
+    cmp1 = _mm_cmplt_epi32(coeffs, zbins[i != 0]);
+    cmp2 = _mm_cmpgt_epi32(coeffs, nzbins[i != 0]);
+    cmp1 = _mm_and_si128(cmp1, cmp2);
+    test = _mm_movemask_epi8(cmp1);
+    if (test == 0xffff)
+      non_zero_regs--;
+    else
+      break;
+  }
+
+  // Quantization pass:
+  for (i = 0; i < non_zero_regs; i++) {
+    __m128i coeffs, coeffs_sign, tmp1, tmp2;
+    int test;
+    int abs_coeff[4];
+    int coeff_sign[4];
+
+    coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4));
+    coeffs_sign = _mm_srai_epi32(coeffs, 31);
+    coeffs = _mm_sub_epi32(_mm_xor_si128(coeffs, coeffs_sign), coeffs_sign);
+    tmp1 = _mm_cmpgt_epi32(coeffs, zbins[i != 0]);
+    tmp2 = _mm_cmpeq_epi32(coeffs, zbins[i != 0]);
+    tmp1 = _mm_or_si128(tmp1, tmp2);
+    test = _mm_movemask_epi8(tmp1);
+    _mm_storeu_si128((__m128i *)abs_coeff, coeffs);
+    _mm_storeu_si128((__m128i *)coeff_sign, coeffs_sign);
+
+    for (j = 0; j < 4; j++) {
+      if (test & (1 << (4 * j))) {
+        int k = 4 * i + j;
+        const int64_t tmp3 = abs_coeff[j] + round_ptr[k != 0];
+        const int64_t tmp4 = ((tmp3 * quant_ptr[k != 0]) >> 16) + tmp3;
+        const uint32_t abs_qcoeff =
+            (uint32_t)((tmp4 * quant_shift_ptr[k != 0]) >> 16);
+        qcoeff_ptr[k] =
+            (int)(abs_qcoeff ^ (uint32_t)coeff_sign[j]) - coeff_sign[j];
+        dqcoeff_ptr[k] = qcoeff_ptr[k] * dequant_ptr[k != 0];
+        if (abs_qcoeff) eob_i = iscan[k] > eob_i ? iscan[k] : eob_i;
+      }
+    }
+  }
+  *eob_ptr = eob_i + 1;
+}
+
+void aom_highbd_quantize_b_32x32_sse2(
+    const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr,
+    const int16_t *round_ptr, const int16_t *quant_ptr,
+    const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
+    tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
+    const int16_t *scan, const int16_t *iscan) {
+  __m128i zbins[2];
+  __m128i nzbins[2];
+  int idx = 0;
+  int idx_arr[1024];
+  int i, eob = -1;
+  const int zbin0_tmp = ROUND_POWER_OF_TWO(zbin_ptr[0], 1);
+  const int zbin1_tmp = ROUND_POWER_OF_TWO(zbin_ptr[1], 1);
+  (void)scan;
+  zbins[0] = _mm_set_epi32(zbin1_tmp, zbin1_tmp, zbin1_tmp, zbin0_tmp);
+  zbins[1] = _mm_set1_epi32(zbin1_tmp);
+
+  nzbins[0] = _mm_setzero_si128();
+  nzbins[1] = _mm_setzero_si128();
+  nzbins[0] = _mm_sub_epi32(nzbins[0], zbins[0]);
+  nzbins[1] = _mm_sub_epi32(nzbins[1], zbins[1]);
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  // Pre-scan pass
+  for (i = 0; i < n_coeffs / 4; i++) {
+    __m128i coeffs, cmp1, cmp2;
+    int test;
+    coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4));
+    cmp1 = _mm_cmplt_epi32(coeffs, zbins[i != 0]);
+    cmp2 = _mm_cmpgt_epi32(coeffs, nzbins[i != 0]);
+    cmp1 = _mm_and_si128(cmp1, cmp2);
+    test = _mm_movemask_epi8(cmp1);
+    if (!(test & 0xf)) idx_arr[idx++] = i * 4;
+    if (!(test & 0xf0)) idx_arr[idx++] = i * 4 + 1;
+    if (!(test & 0xf00)) idx_arr[idx++] = i * 4 + 2;
+    if (!(test & 0xf000)) idx_arr[idx++] = i * 4 + 3;
+  }
+
+  // Quantization pass: only process the coefficients selected in
+  // pre-scan pass. Note: idx can be zero.
+  for (i = 0; i < idx; i++) {
+    const int rc = idx_arr[i];
+    const int coeff = coeff_ptr[rc];
+    const int coeff_sign = AOMSIGN(coeff);
+    const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+    const int64_t tmp1 = abs_coeff + ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1);
+    const int64_t tmp2 = ((tmp1 * quant_ptr[rc != 0]) >> 16) + tmp1;
+    const uint32_t abs_qcoeff =
+        (uint32_t)((tmp2 * quant_shift_ptr[rc != 0]) >> 15);
+    qcoeff_ptr[rc] = (int)(abs_qcoeff ^ (uint32_t)coeff_sign) - coeff_sign;
+    dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0] / 2;
+    if (abs_qcoeff) eob = iscan[idx_arr[i]] > eob ? iscan[idx_arr[i]] : eob;
+  }
+  *eob_ptr = eob + 1;
+}
+
+void aom_highbd_quantize_b_64x64_sse2(
+    const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr,
+    const int16_t *round_ptr, const int16_t *quant_ptr,
+    const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
+    tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
+    const int16_t *scan, const int16_t *iscan) {
+  __m128i zbins[2];
+  __m128i nzbins[2];
+  int idx = 0;
+  int idx_arr[1024];
+  int i, eob = -1;
+  const int zbin0_tmp = ROUND_POWER_OF_TWO(zbin_ptr[0], 2);
+  const int zbin1_tmp = ROUND_POWER_OF_TWO(zbin_ptr[1], 2);
+  (void)scan;
+  zbins[0] = _mm_set_epi32(zbin1_tmp, zbin1_tmp, zbin1_tmp, zbin0_tmp);
+  zbins[1] = _mm_set1_epi32(zbin1_tmp);
+
+  nzbins[0] = _mm_setzero_si128();
+  nzbins[1] = _mm_setzero_si128();
+  nzbins[0] = _mm_sub_epi32(nzbins[0], zbins[0]);
+  nzbins[1] = _mm_sub_epi32(nzbins[1], zbins[1]);
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  // Pre-scan pass
+  for (i = 0; i < n_coeffs / 4; i++) {
+    __m128i coeffs, cmp1, cmp2;
+    int test;
+    coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4));
+    cmp1 = _mm_cmplt_epi32(coeffs, zbins[i != 0]);
+    cmp2 = _mm_cmpgt_epi32(coeffs, nzbins[i != 0]);
+    cmp1 = _mm_and_si128(cmp1, cmp2);
+    test = _mm_movemask_epi8(cmp1);
+    if (!(test & 0xf)) idx_arr[idx++] = i * 4;
+    if (!(test & 0xf0)) idx_arr[idx++] = i * 4 + 1;
+    if (!(test & 0xf00)) idx_arr[idx++] = i * 4 + 2;
+    if (!(test & 0xf000)) idx_arr[idx++] = i * 4 + 3;
+  }
+
+  // Quantization pass: only process the coefficients selected in
+  // pre-scan pass. Note: idx can be zero.
+  for (i = 0; i < idx; i++) {
+    const int rc = idx_arr[i];
+    const int coeff = coeff_ptr[rc];
+    const int coeff_sign = AOMSIGN(coeff);
+    const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+    const int64_t tmp1 = abs_coeff + ROUND_POWER_OF_TWO(round_ptr[rc != 0], 2);
+    const int64_t tmp2 = ((tmp1 * quant_ptr[rc != 0]) >> 16) + tmp1;
+    const uint32_t abs_qcoeff =
+        (uint32_t)((tmp2 * quant_shift_ptr[rc != 0]) >> 14);
+    qcoeff_ptr[rc] = (int)(abs_qcoeff ^ (uint32_t)coeff_sign) - coeff_sign;
+    dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0] / 4;
+    if (abs_qcoeff) eob = iscan[idx_arr[i]] > eob ? iscan[idx_arr[i]] : eob;
+  }
+  *eob_ptr = eob + 1;
+}
diff --git a/third_party/aom/aom_dsp/x86/highbd_sad4d_sse2.asm b/third_party/aom/aom_dsp/x86/highbd_sad4d_sse2.asm
new file mode 100644
index 0000000000..03839b493c
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/highbd_sad4d_sse2.asm
@@ -0,0 +1,344 @@
+;
+; Copyright (c) 2016, 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 "third_party/x86inc/x86inc.asm"
+
+SECTION .text
+
+; HIGH_PROCESS_4x2x4 first, off_{first,second}_{src,ref}, advance_at_end
+%macro HIGH_PROCESS_4x2x4 5-6 0
+  movh                  m0, [srcq +%2*2]
+%if %1 == 1
+  movu                  m4, [ref1q+%3*2]
+  movu                  m5, [ref2q+%3*2]
+  movu                  m6, [ref3q+%3*2]
+  movu                  m7, [ref4q+%3*2]
+  movhps                m0, [srcq +%4*2]
+  movhps                m4, [ref1q+%5*2]
+  movhps                m5, [ref2q+%5*2]
+  movhps                m6, [ref3q+%5*2]
+  movhps                m7, [ref4q+%5*2]
+  mova                  m3, m0
+  mova                  m2, m0
+  psubusw               m3, m4
+  psubusw               m2, m5
+  psubusw               m4, m0
+  psubusw               m5, m0
+  por                   m4, m3
+  por                   m5, m2
+  pmaddwd               m4, m1
+  pmaddwd               m5, m1
+  mova                  m3, m0
+  mova                  m2, m0
+  psubusw               m3, m6
+  psubusw               m2, m7
+  psubusw               m6, m0
+  psubusw               m7, m0
+  por                   m6, m3
+  por                   m7, m2
+  pmaddwd               m6, m1
+  pmaddwd               m7, m1
+%else
+  movu                  m2, [ref1q+%3*2]
+  movhps                m0, [srcq +%4*2]
+  movhps                m2, [ref1q+%5*2]
+  mova                  m3, m0
+  psubusw               m3, m2
+  psubusw               m2, m0
+  por                   m2, m3
+  pmaddwd               m2, m1
+  paddd                 m4, m2
+
+  movu                  m2, [ref2q+%3*2]
+  mova                  m3, m0
+  movhps                m2, [ref2q+%5*2]
+  psubusw               m3, m2
+  psubusw               m2, m0
+  por                   m2, m3
+  pmaddwd               m2, m1
+  paddd                 m5, m2
+
+  movu                  m2, [ref3q+%3*2]
+  mova                  m3, m0
+  movhps                m2, [ref3q+%5*2]
+  psubusw               m3, m2
+  psubusw               m2, m0
+  por                   m2, m3
+  pmaddwd               m2, m1
+  paddd                 m6, m2
+
+  movu                  m2, [ref4q+%3*2]
+  mova                  m3, m0
+  movhps                m2, [ref4q+%5*2]
+  psubusw               m3, m2
+  psubusw               m2, m0
+  por                   m2, m3
+  pmaddwd               m2, m1
+  paddd                 m7, m2
+%endif
+%if %6 == 1
+  lea                 srcq, [srcq +src_strideq*4]
+  lea                ref1q, [ref1q+ref_strideq*4]
+  lea                ref2q, [ref2q+ref_strideq*4]
+  lea                ref3q, [ref3q+ref_strideq*4]
+  lea                ref4q, [ref4q+ref_strideq*4]
+%endif
+%endmacro
+
+; PROCESS_8x2x4 first, off_{first,second}_{src,ref}, advance_at_end
+%macro HIGH_PROCESS_8x2x4 5-6 0
+  ; 1st 8 px
+  mova                  m0, [srcq +%2*2]
+%if %1 == 1
+  movu                  m4, [ref1q+%3*2]
+  movu                  m5, [ref2q+%3*2]
+  movu                  m6, [ref3q+%3*2]
+  movu                  m7, [ref4q+%3*2]
+  mova                  m3, m0
+  mova                  m2, m0
+  psubusw               m3, m4
+  psubusw               m2, m5
+  psubusw               m4, m0
+  psubusw               m5, m0
+  por                   m4, m3
+  por                   m5, m2
+  pmaddwd               m4, m1
+  pmaddwd               m5, m1
+  mova                  m3, m0
+  mova                  m2, m0
+  psubusw               m3, m6
+  psubusw               m2, m7
+  psubusw               m6, m0
+  psubusw               m7, m0
+  por                   m6, m3
+  por                   m7, m2
+  pmaddwd               m6, m1
+  pmaddwd               m7, m1
+%else
+  mova                  m3, m0
+  movu                  m2, [ref1q+%3*2]
+  psubusw               m3, m2
+  psubusw               m2, m0
+  por                   m2, m3
+  mova                  m3, m0
+  pmaddwd               m2, m1
+  paddd                 m4, m2
+  movu                  m2, [ref2q+%3*2]
+  psubusw               m3, m2
+  psubusw               m2, m0
+  por                   m2, m3
+  mova                  m3, m0
+  pmaddwd               m2, m1
+  paddd                 m5, m2
+  movu                  m2, [ref3q+%3*2]
+  psubusw               m3, m2
+  psubusw               m2, m0
+  por                   m2, m3
+  mova                  m3, m0
+  pmaddwd               m2, m1
+  paddd                 m6, m2
+  movu                  m2, [ref4q+%3*2]
+  psubusw               m3, m2
+  psubusw               m2, m0
+  por                   m2, m3
+  pmaddwd               m2, m1
+  paddd                 m7, m2
+%endif
+
+  ; 2nd 8 px
+  mova                  m0, [srcq +(%4)*2]
+  mova                  m3, m0
+  movu                  m2, [ref1q+(%5)*2]
+  psubusw               m3, m2
+  psubusw               m2, m0
+  por                   m2, m3
+  mova                  m3, m0
+  pmaddwd               m2, m1
+  paddd                 m4, m2
+  movu                  m2, [ref2q+(%5)*2]
+  psubusw               m3, m2
+  psubusw               m2, m0
+  por                   m2, m3
+  mova                  m3, m0
+  pmaddwd               m2, m1
+  paddd                 m5, m2
+  movu                  m2, [ref3q+(%5)*2]
+  psubusw               m3, m2
+  psubusw               m2, m0
+  por                   m2, m3
+  mova                  m3, m0
+  pmaddwd               m2, m1
+  paddd                 m6, m2
+  movu                  m2, [ref4q+(%5)*2]
+  psubusw               m3, m2
+  psubusw               m2, m0
+%if %6 == 1
+  lea                 srcq, [srcq +src_strideq*4]
+  lea                ref1q, [ref1q+ref_strideq*4]
+  lea                ref2q, [ref2q+ref_strideq*4]
+  lea                ref3q, [ref3q+ref_strideq*4]
+  lea                ref4q, [ref4q+ref_strideq*4]
+%endif
+  por                   m2, m3
+  pmaddwd               m2, m1
+  paddd                 m7, m2
+%endmacro
+
+; HIGH_PROCESS_16x2x4 first, off_{first,second}_{src,ref}, advance_at_end
+%macro HIGH_PROCESS_16x2x4 5-6 0
+  HIGH_PROCESS_8x2x4 %1, %2, %3, (%2 + 8), (%3 + 8)
+  HIGH_PROCESS_8x2x4  0, %4, %5, (%4 + 8), (%5 + 8), %6
+%endmacro
+
+; HIGH_PROCESS_32x2x4 first, off_{first,second}_{src,ref}, advance_at_end
+%macro HIGH_PROCESS_32x2x4 5-6 0
+  HIGH_PROCESS_16x2x4 %1, %2, %3, (%2 + 16), (%3 + 16)
+  HIGH_PROCESS_16x2x4  0, %4, %5, (%4 + 16), (%5 + 16), %6
+%endmacro
+
+; HIGH_PROCESS_64x2x4 first, off_{first,second}_{src,ref}, advance_at_end
+%macro HIGH_PROCESS_64x2x4 5-6 0
+  HIGH_PROCESS_32x2x4 %1, %2, %3, (%2 + 32), (%3 + 32)
+  HIGH_PROCESS_32x2x4  0, %4, %5, (%4 + 32), (%5 + 32), %6
+%endmacro
+
+; void aom_highbd_sadNxNx4d_sse2(uint8_t *src,    int src_stride,
+;                         uint8_t *ref[4], int ref_stride,
+;                         uint32_t res[4]);
+; Macro Arguments:
+;   1: Width
+;   2: Height
+;   3: If 0, then normal sad, if 2, then skip every other row
+%macro HIGH_SADNXN4D 2-3 0
+%if %3 == 0  ; normal sad
+%if AOM_ARCH_X86_64
+cglobal highbd_sad%1x%2x4d, 5, 8, 8, src, src_stride, ref1, ref_stride, \
+                              res, ref2, ref3, ref4
+%else
+cglobal highbd_sad%1x%2x4d, 4, 7, 8, src, src_stride, ref1, ref_stride, \
+                              ref2, ref3, ref4
+%endif  ; AOM_ARCH_X86_64
+%else  ; %3 == 2, downsample
+%if AOM_ARCH_X86_64
+cglobal highbd_sad_skip_%1x%2x4d, 5, 8, 8, src, src_stride, ref1, ref_stride, \
+                              res, ref2, ref3, ref4
+%else
+cglobal highbd_sad_skip_%1x%2x4d, 4, 7, 8, src, src_stride, ref1, ref_stride, \
+                              ref2, ref3, ref4
+%endif  ; AOM_ARCH_X86_64
+%endif  ; sad/avg/skip
+
+; set m1
+  push                srcq
+  mov                 srcd, 0x00010001
+  movd                  m1, srcd
+  pshufd                m1, m1, 0x0
+  pop                 srcq
+
+%if %3 == 2  ; skip rows
+  lea          src_strided, [2*src_strided]
+  lea          ref_strided, [2*ref_strided]
+%endif  ; skip rows
+  movsxdifnidn src_strideq, src_strided
+  movsxdifnidn ref_strideq, ref_strided
+  mov                ref2q, [ref1q+gprsize*1]
+  mov                ref3q, [ref1q+gprsize*2]
+  mov                ref4q, [ref1q+gprsize*3]
+  mov                ref1q, [ref1q+gprsize*0]
+
+; convert byte pointers to short pointers
+  shl                 srcq, 1
+  shl                ref2q, 1
+  shl                ref3q, 1
+  shl                ref4q, 1
+  shl                ref1q, 1
+
+  HIGH_PROCESS_%1x2x4 1, 0, 0, src_strideq, ref_strideq, 1
+%if %3 == 2  ;  Downsampling by two
+%define num_rep (%2-8)/4
+%else
+%define num_rep (%2-4)/2
+%endif
+%rep num_rep
+  HIGH_PROCESS_%1x2x4 0, 0, 0, src_strideq, ref_strideq, 1
+%endrep
+%undef rep
+  HIGH_PROCESS_%1x2x4 0, 0, 0, src_strideq, ref_strideq, 0
+  ; N.B. HIGH_PROCESS outputs dwords (32 bits)
+  ; so in high bit depth even the smallest width (4) needs 128bits i.e. XMM
+  movhlps               m0, m4
+  movhlps               m1, m5
+  movhlps               m2, m6
+  movhlps               m3, m7
+  paddd                 m4, m0
+  paddd                 m5, m1
+  paddd                 m6, m2
+  paddd                 m7, m3
+  punpckldq             m4, m5
+  punpckldq             m6, m7
+  movhlps               m0, m4
+  movhlps               m1, m6
+  paddd                 m4, m0
+  paddd                 m6, m1
+  punpcklqdq            m4, m6
+%if %3 == 2  ; skip rows
+  pslld                 m4, 1
+%endif
+  movifnidn             r4, r4mp
+  movu                [r4], m4
+  RET
+%endmacro
+
+
+INIT_XMM sse2
+HIGH_SADNXN4D 64, 64
+HIGH_SADNXN4D 64, 32
+HIGH_SADNXN4D 32, 64
+HIGH_SADNXN4D 32, 32
+HIGH_SADNXN4D 32, 16
+HIGH_SADNXN4D 16, 32
+HIGH_SADNXN4D 16, 16
+HIGH_SADNXN4D 16,  8
+HIGH_SADNXN4D  8, 16
+HIGH_SADNXN4D  8,  8
+HIGH_SADNXN4D  8,  4
+HIGH_SADNXN4D  4,  8
+HIGH_SADNXN4D  4,  4
+HIGH_SADNXN4D  4, 16
+HIGH_SADNXN4D 16,  4
+HIGH_SADNXN4D  8, 32
+HIGH_SADNXN4D 32,  8
+HIGH_SADNXN4D 16, 64
+HIGH_SADNXN4D 64, 16
+
+HIGH_SADNXN4D 64, 64, 2
+HIGH_SADNXN4D 64, 32, 2
+HIGH_SADNXN4D 32, 64, 2
+HIGH_SADNXN4D 32, 32, 2
+HIGH_SADNXN4D 32, 16, 2
+HIGH_SADNXN4D 16, 32, 2
+HIGH_SADNXN4D 16, 16, 2
+HIGH_SADNXN4D 16,  8, 2
+HIGH_SADNXN4D  8, 16, 2
+HIGH_SADNXN4D  8,  8, 2
+HIGH_SADNXN4D  4,  8, 2
+HIGH_SADNXN4D  4, 16, 2
+HIGH_SADNXN4D  8, 32, 2
+HIGH_SADNXN4D 32,  8, 2
+HIGH_SADNXN4D 16, 64, 2
+HIGH_SADNXN4D 64, 16, 2
+
+; Current code cannot handle the case when the height is downsampled to 2
+; HIGH_SADNXN4D 16,  4, 2
+; HIGH_SADNXN4D  8,  4, 2
+; HIGH_SADNXN4D  4,  4, 2
diff --git a/third_party/aom/aom_dsp/x86/highbd_sad_avx2.c b/third_party/aom/aom_dsp/x86/highbd_sad_avx2.c
new file mode 100644
index 0000000000..6c78eeeefb
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/highbd_sad_avx2.c
@@ -0,0 +1,720 @@
+/*
+ * Copyright (c) 2016, 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_config.h"
+#include "config/aom_dsp_rtcd.h"
+
+#include "aom/aom_integer.h"
+#include "aom_dsp/x86/synonyms_avx2.h"
+#include "aom_ports/mem.h"
+
+// SAD
+static INLINE unsigned int get_sad_from_mm256_epi32(const __m256i *v) {
+  // input 8 32-bit summation
+  __m128i lo128, hi128;
+  __m256i u = _mm256_srli_si256(*v, 8);
+  u = _mm256_add_epi32(u, *v);
+
+  // 4 32-bit summation
+  hi128 = _mm256_extracti128_si256(u, 1);
+  lo128 = _mm256_castsi256_si128(u);
+  lo128 = _mm_add_epi32(hi128, lo128);
+
+  // 2 32-bit summation
+  hi128 = _mm_srli_si128(lo128, 4);
+  lo128 = _mm_add_epi32(lo128, hi128);
+
+  return (unsigned int)_mm_cvtsi128_si32(lo128);
+}
+
+static INLINE void highbd_sad16x4_core_avx2(__m256i *s, __m256i *r,
+                                            __m256i *sad_acc) {
+  const __m256i zero = _mm256_setzero_si256();
+  int i;
+  for (i = 0; i < 4; i++) {
+    s[i] = _mm256_sub_epi16(s[i], r[i]);
+    s[i] = _mm256_abs_epi16(s[i]);
+  }
+
+  s[0] = _mm256_add_epi16(s[0], s[1]);
+  s[0] = _mm256_add_epi16(s[0], s[2]);
+  s[0] = _mm256_add_epi16(s[0], s[3]);
+
+  r[0] = _mm256_unpacklo_epi16(s[0], zero);
+  r[1] = _mm256_unpackhi_epi16(s[0], zero);
+
+  r[0] = _mm256_add_epi32(r[0], r[1]);
+  *sad_acc = _mm256_add_epi32(*sad_acc, r[0]);
+}
+
+// If sec_ptr = 0, calculate regular SAD. Otherwise, calculate average SAD.
+static INLINE void sad16x4(const uint16_t *src_ptr, int src_stride,
+                           const uint16_t *ref_ptr, int ref_stride,
+                           const uint16_t *sec_ptr, __m256i *sad_acc) {
+  __m256i s[4], r[4];
+  s[0] = _mm256_loadu_si256((const __m256i *)src_ptr);
+  s[1] = _mm256_loadu_si256((const __m256i *)(src_ptr + src_stride));
+  s[2] = _mm256_loadu_si256((const __m256i *)(src_ptr + 2 * src_stride));
+  s[3] = _mm256_loadu_si256((const __m256i *)(src_ptr + 3 * src_stride));
+
+  r[0] = _mm256_loadu_si256((const __m256i *)ref_ptr);
+  r[1] = _mm256_loadu_si256((const __m256i *)(ref_ptr + ref_stride));
+  r[2] = _mm256_loadu_si256((const __m256i *)(ref_ptr + 2 * ref_stride));
+  r[3] = _mm256_loadu_si256((const __m256i *)(ref_ptr + 3 * ref_stride));
+
+  if (sec_ptr) {
+    r[0] = _mm256_avg_epu16(r[0], _mm256_loadu_si256((const __m256i *)sec_ptr));
+    r[1] = _mm256_avg_epu16(
+        r[1], _mm256_loadu_si256((const __m256i *)(sec_ptr + 16)));
+    r[2] = _mm256_avg_epu16(
+        r[2], _mm256_loadu_si256((const __m256i *)(sec_ptr + 32)));
+    r[3] = _mm256_avg_epu16(
+        r[3], _mm256_loadu_si256((const __m256i *)(sec_ptr + 48)));
+  }
+  highbd_sad16x4_core_avx2(s, r, sad_acc);
+}
+
+static AOM_FORCE_INLINE unsigned int aom_highbd_sad16xN_avx2(int N,
+                                                             const uint8_t *src,
+                                                             int src_stride,
+                                                             const uint8_t *ref,
+                                                             int ref_stride) {
+  const uint16_t *src_ptr = CONVERT_TO_SHORTPTR(src);
+  const uint16_t *ref_ptr = CONVERT_TO_SHORTPTR(ref);
+  int i;
+  __m256i sad = _mm256_setzero_si256();
+  for (i = 0; i < N; i += 4) {
+    sad16x4(src_ptr, src_stride, ref_ptr, ref_stride, NULL, &sad);
+    src_ptr += src_stride << 2;
+    ref_ptr += ref_stride << 2;
+  }
+  return (unsigned int)get_sad_from_mm256_epi32(&sad);
+}
+
+static void sad32x4(const uint16_t *src_ptr, int src_stride,
+                    const uint16_t *ref_ptr, int ref_stride,
+                    const uint16_t *sec_ptr, __m256i *sad_acc) {
+  __m256i s[4], r[4];
+  int row_sections = 0;
+
+  while (row_sections < 2) {
+    s[0] = _mm256_loadu_si256((const __m256i *)src_ptr);
+    s[1] = _mm256_loadu_si256((const __m256i *)(src_ptr + 16));
+    s[2] = _mm256_loadu_si256((const __m256i *)(src_ptr + src_stride));
+    s[3] = _mm256_loadu_si256((const __m256i *)(src_ptr + src_stride + 16));
+
+    r[0] = _mm256_loadu_si256((const __m256i *)ref_ptr);
+    r[1] = _mm256_loadu_si256((const __m256i *)(ref_ptr + 16));
+    r[2] = _mm256_loadu_si256((const __m256i *)(ref_ptr + ref_stride));
+    r[3] = _mm256_loadu_si256((const __m256i *)(ref_ptr + ref_stride + 16));
+
+    if (sec_ptr) {
+      r[0] =
+          _mm256_avg_epu16(r[0], _mm256_loadu_si256((const __m256i *)sec_ptr));
+      r[1] = _mm256_avg_epu16(
+          r[1], _mm256_loadu_si256((const __m256i *)(sec_ptr + 16)));
+      r[2] = _mm256_avg_epu16(
+          r[2], _mm256_loadu_si256((const __m256i *)(sec_ptr + 32)));
+      r[3] = _mm256_avg_epu16(
+          r[3], _mm256_loadu_si256((const __m256i *)(sec_ptr + 48)));
+      sec_ptr += 32 << 1;
+    }
+    highbd_sad16x4_core_avx2(s, r, sad_acc);
+
+    row_sections += 1;
+    src_ptr += src_stride << 1;
+    ref_ptr += ref_stride << 1;
+  }
+}
+
+static AOM_FORCE_INLINE unsigned int aom_highbd_sad32xN_avx2(int N,
+                                                             const uint8_t *src,
+                                                             int src_stride,
+                                                             const uint8_t *ref,
+                                                             int ref_stride) {
+  __m256i sad = _mm256_setzero_si256();
+  uint16_t *srcp = CONVERT_TO_SHORTPTR(src);
+  uint16_t *refp = CONVERT_TO_SHORTPTR(ref);
+  const int left_shift = 2;
+  int i;
+
+  for (i = 0; i < N; i += 4) {
+    sad32x4(srcp, src_stride, refp, ref_stride, NULL, &sad);
+    srcp += src_stride << left_shift;
+    refp += ref_stride << left_shift;
+  }
+  return get_sad_from_mm256_epi32(&sad);
+}
+
+static void sad64x2(const uint16_t *src_ptr, int src_stride,
+                    const uint16_t *ref_ptr, int ref_stride,
+                    const uint16_t *sec_ptr, __m256i *sad_acc) {
+  __m256i s[4], r[4];
+  int i;
+  for (i = 0; i < 2; i++) {
+    s[0] = _mm256_loadu_si256((const __m256i *)src_ptr);
+    s[1] = _mm256_loadu_si256((const __m256i *)(src_ptr + 16));
+    s[2] = _mm256_loadu_si256((const __m256i *)(src_ptr + 32));
+    s[3] = _mm256_loadu_si256((const __m256i *)(src_ptr + 48));
+
+    r[0] = _mm256_loadu_si256((const __m256i *)ref_ptr);
+    r[1] = _mm256_loadu_si256((const __m256i *)(ref_ptr + 16));
+    r[2] = _mm256_loadu_si256((const __m256i *)(ref_ptr + 32));
+    r[3] = _mm256_loadu_si256((const __m256i *)(ref_ptr + 48));
+    if (sec_ptr) {
+      r[0] =
+          _mm256_avg_epu16(r[0], _mm256_loadu_si256((const __m256i *)sec_ptr));
+      r[1] = _mm256_avg_epu16(
+          r[1], _mm256_loadu_si256((const __m256i *)(sec_ptr + 16)));
+      r[2] = _mm256_avg_epu16(
+          r[2], _mm256_loadu_si256((const __m256i *)(sec_ptr + 32)));
+      r[3] = _mm256_avg_epu16(
+          r[3], _mm256_loadu_si256((const __m256i *)(sec_ptr + 48)));
+      sec_ptr += 64;
+    }
+    highbd_sad16x4_core_avx2(s, r, sad_acc);
+    src_ptr += src_stride;
+    ref_ptr += ref_stride;
+  }
+}
+
+static AOM_FORCE_INLINE unsigned int aom_highbd_sad64xN_avx2(int N,
+                                                             const uint8_t *src,
+                                                             int src_stride,
+                                                             const uint8_t *ref,
+                                                             int ref_stride) {
+  __m256i sad = _mm256_setzero_si256();
+  uint16_t *srcp = CONVERT_TO_SHORTPTR(src);
+  uint16_t *refp = CONVERT_TO_SHORTPTR(ref);
+  const int left_shift = 1;
+  int i;
+  for (i = 0; i < N; i += 2) {
+    sad64x2(srcp, src_stride, refp, ref_stride, NULL, &sad);
+    srcp += src_stride << left_shift;
+    refp += ref_stride << left_shift;
+  }
+  return get_sad_from_mm256_epi32(&sad);
+}
+
+static void sad128x1(const uint16_t *src_ptr, const uint16_t *ref_ptr,
+                     const uint16_t *sec_ptr, __m256i *sad_acc) {
+  __m256i s[4], r[4];
+  int i;
+  for (i = 0; i < 2; i++) {
+    s[0] = _mm256_loadu_si256((const __m256i *)src_ptr);
+    s[1] = _mm256_loadu_si256((const __m256i *)(src_ptr + 16));
+    s[2] = _mm256_loadu_si256((const __m256i *)(src_ptr + 32));
+    s[3] = _mm256_loadu_si256((const __m256i *)(src_ptr + 48));
+    r[0] = _mm256_loadu_si256((const __m256i *)ref_ptr);
+    r[1] = _mm256_loadu_si256((const __m256i *)(ref_ptr + 16));
+    r[2] = _mm256_loadu_si256((const __m256i *)(ref_ptr + 32));
+    r[3] = _mm256_loadu_si256((const __m256i *)(ref_ptr + 48));
+    if (sec_ptr) {
+      r[0] =
+          _mm256_avg_epu16(r[0], _mm256_loadu_si256((const __m256i *)sec_ptr));
+      r[1] = _mm256_avg_epu16(
+          r[1], _mm256_loadu_si256((const __m256i *)(sec_ptr + 16)));
+      r[2] = _mm256_avg_epu16(
+          r[2], _mm256_loadu_si256((const __m256i *)(sec_ptr + 32)));
+      r[3] = _mm256_avg_epu16(
+          r[3], _mm256_loadu_si256((const __m256i *)(sec_ptr + 48)));
+      sec_ptr += 64;
+    }
+    highbd_sad16x4_core_avx2(s, r, sad_acc);
+    src_ptr += 64;
+    ref_ptr += 64;
+  }
+}
+
+static AOM_FORCE_INLINE unsigned int aom_highbd_sad128xN_avx2(
+    int N, const uint8_t *src, int src_stride, const uint8_t *ref,
+    int ref_stride) {
+  __m256i sad = _mm256_setzero_si256();
+  uint16_t *srcp = CONVERT_TO_SHORTPTR(src);
+  uint16_t *refp = CONVERT_TO_SHORTPTR(ref);
+  int row = 0;
+  while (row < N) {
+    sad128x1(srcp, refp, NULL, &sad);
+    srcp += src_stride;
+    refp += ref_stride;
+    row++;
+  }
+  return get_sad_from_mm256_epi32(&sad);
+}
+
+#define HIGHBD_SADMXN_AVX2(m, n)                                            \
+  unsigned int aom_highbd_sad##m##x##n##_avx2(                              \
+      const uint8_t *src, int src_stride, const uint8_t *ref,               \
+      int ref_stride) {                                                     \
+    return aom_highbd_sad##m##xN_avx2(n, src, src_stride, ref, ref_stride); \
+  }
+
+#define HIGHBD_SAD_SKIP_MXN_AVX2(m, n)                                       \
+  unsigned int aom_highbd_sad_skip_##m##x##n##_avx2(                         \
+      const uint8_t *src, int src_stride, const uint8_t *ref,                \
+      int ref_stride) {                                                      \
+    return 2 * aom_highbd_sad##m##xN_avx2((n / 2), src, 2 * src_stride, ref, \
+                                          2 * ref_stride);                   \
+  }
+
+HIGHBD_SADMXN_AVX2(16, 4)
+HIGHBD_SADMXN_AVX2(16, 8)
+HIGHBD_SADMXN_AVX2(16, 16)
+HIGHBD_SADMXN_AVX2(16, 32)
+HIGHBD_SADMXN_AVX2(16, 64)
+
+HIGHBD_SADMXN_AVX2(32, 8)
+HIGHBD_SADMXN_AVX2(32, 16)
+HIGHBD_SADMXN_AVX2(32, 32)
+HIGHBD_SADMXN_AVX2(32, 64)
+
+HIGHBD_SADMXN_AVX2(64, 16)
+HIGHBD_SADMXN_AVX2(64, 32)
+HIGHBD_SADMXN_AVX2(64, 64)
+HIGHBD_SADMXN_AVX2(64, 128)
+
+HIGHBD_SADMXN_AVX2(128, 64)
+HIGHBD_SADMXN_AVX2(128, 128)
+
+HIGHBD_SAD_SKIP_MXN_AVX2(16, 8)
+HIGHBD_SAD_SKIP_MXN_AVX2(16, 16)
+HIGHBD_SAD_SKIP_MXN_AVX2(16, 32)
+HIGHBD_SAD_SKIP_MXN_AVX2(16, 64)
+
+HIGHBD_SAD_SKIP_MXN_AVX2(32, 8)
+HIGHBD_SAD_SKIP_MXN_AVX2(32, 16)
+HIGHBD_SAD_SKIP_MXN_AVX2(32, 32)
+HIGHBD_SAD_SKIP_MXN_AVX2(32, 64)
+
+HIGHBD_SAD_SKIP_MXN_AVX2(64, 16)
+HIGHBD_SAD_SKIP_MXN_AVX2(64, 32)
+HIGHBD_SAD_SKIP_MXN_AVX2(64, 64)
+HIGHBD_SAD_SKIP_MXN_AVX2(64, 128)
+
+HIGHBD_SAD_SKIP_MXN_AVX2(128, 64)
+HIGHBD_SAD_SKIP_MXN_AVX2(128, 128)
+
+unsigned int aom_highbd_sad16x4_avg_avx2(const uint8_t *src, int src_stride,
+                                         const uint8_t *ref, int ref_stride,
+                                         const uint8_t *second_pred) {
+  __m256i sad = _mm256_setzero_si256();
+  uint16_t *srcp = CONVERT_TO_SHORTPTR(src);
+  uint16_t *refp = CONVERT_TO_SHORTPTR(ref);
+  uint16_t *secp = CONVERT_TO_SHORTPTR(second_pred);
+  sad16x4(srcp, src_stride, refp, ref_stride, secp, &sad);
+
+  return get_sad_from_mm256_epi32(&sad);
+}
+
+unsigned int aom_highbd_sad16x8_avg_avx2(const uint8_t *src, int src_stride,
+                                         const uint8_t *ref, int ref_stride,
+                                         const uint8_t *second_pred) {
+  __m256i sad = _mm256_setzero_si256();
+  uint16_t *srcp = CONVERT_TO_SHORTPTR(src);
+  uint16_t *refp = CONVERT_TO_SHORTPTR(ref);
+  uint16_t *secp = CONVERT_TO_SHORTPTR(second_pred);
+
+  sad16x4(srcp, src_stride, refp, ref_stride, secp, &sad);
+
+  // Next 4 rows
+  srcp += src_stride << 2;
+  refp += ref_stride << 2;
+  secp += 64;
+  sad16x4(srcp, src_stride, refp, ref_stride, secp, &sad);
+  return get_sad_from_mm256_epi32(&sad);
+}
+
+unsigned int aom_highbd_sad16x16_avg_avx2(const uint8_t *src, int src_stride,
+                                          const uint8_t *ref, int ref_stride,
+                                          const uint8_t *second_pred) {
+  const int left_shift = 3;
+  uint32_t sum = aom_highbd_sad16x8_avg_avx2(src, src_stride, ref, ref_stride,
+                                             second_pred);
+  src += src_stride << left_shift;
+  ref += ref_stride << left_shift;
+  second_pred += 16 << left_shift;
+  sum += aom_highbd_sad16x8_avg_avx2(src, src_stride, ref, ref_stride,
+                                     second_pred);
+  return sum;
+}
+
+unsigned int aom_highbd_sad16x32_avg_avx2(const uint8_t *src, int src_stride,
+                                          const uint8_t *ref, int ref_stride,
+                                          const uint8_t *second_pred) {
+  const int left_shift = 4;
+  uint32_t sum = aom_highbd_sad16x16_avg_avx2(src, src_stride, ref, ref_stride,
+                                              second_pred);
+  src += src_stride << left_shift;
+  ref += ref_stride << left_shift;
+  second_pred += 16 << left_shift;
+  sum += aom_highbd_sad16x16_avg_avx2(src, src_stride, ref, ref_stride,
+                                      second_pred);
+  return sum;
+}
+
+unsigned int aom_highbd_sad16x64_avg_avx2(const uint8_t *src, int src_stride,
+                                          const uint8_t *ref, int ref_stride,
+                                          const uint8_t *second_pred) {
+  const int left_shift = 5;
+  uint32_t sum = aom_highbd_sad16x32_avg_avx2(src, src_stride, ref, ref_stride,
+                                              second_pred);
+  src += src_stride << left_shift;
+  ref += ref_stride << left_shift;
+  second_pred += 16 << left_shift;
+  sum += aom_highbd_sad16x32_avg_avx2(src, src_stride, ref, ref_stride,
+                                      second_pred);
+  return sum;
+}
+
+unsigned int aom_highbd_sad32x8_avg_avx2(const uint8_t *src, int src_stride,
+                                         const uint8_t *ref, int ref_stride,
+                                         const uint8_t *second_pred) {
+  __m256i sad = _mm256_setzero_si256();
+  uint16_t *srcp = CONVERT_TO_SHORTPTR(src);
+  uint16_t *refp = CONVERT_TO_SHORTPTR(ref);
+  uint16_t *secp = CONVERT_TO_SHORTPTR(second_pred);
+  const int left_shift = 2;
+  int row_section = 0;
+
+  while (row_section < 2) {
+    sad32x4(srcp, src_stride, refp, ref_stride, secp, &sad);
+    srcp += src_stride << left_shift;
+    refp += ref_stride << left_shift;
+    secp += 32 << left_shift;
+    row_section += 1;
+  }
+  return get_sad_from_mm256_epi32(&sad);
+}
+
+unsigned int aom_highbd_sad32x16_avg_avx2(const uint8_t *src, int src_stride,
+                                          const uint8_t *ref, int ref_stride,
+                                          const uint8_t *second_pred) {
+  __m256i sad = _mm256_setzero_si256();
+  uint16_t *srcp = CONVERT_TO_SHORTPTR(src);
+  uint16_t *refp = CONVERT_TO_SHORTPTR(ref);
+  uint16_t *secp = CONVERT_TO_SHORTPTR(second_pred);
+  const int left_shift = 2;
+  int row_section = 0;
+
+  while (row_section < 4) {
+    sad32x4(srcp, src_stride, refp, ref_stride, secp, &sad);
+    srcp += src_stride << left_shift;
+    refp += ref_stride << left_shift;
+    secp += 32 << left_shift;
+    row_section += 1;
+  }
+  return get_sad_from_mm256_epi32(&sad);
+}
+
+unsigned int aom_highbd_sad32x32_avg_avx2(const uint8_t *src, int src_stride,
+                                          const uint8_t *ref, int ref_stride,
+                                          const uint8_t *second_pred) {
+  const int left_shift = 4;
+  uint32_t sum = aom_highbd_sad32x16_avg_avx2(src, src_stride, ref, ref_stride,
+                                              second_pred);
+  src += src_stride << left_shift;
+  ref += ref_stride << left_shift;
+  second_pred += 32 << left_shift;
+  sum += aom_highbd_sad32x16_avg_avx2(src, src_stride, ref, ref_stride,
+                                      second_pred);
+  return sum;
+}
+
+unsigned int aom_highbd_sad32x64_avg_avx2(const uint8_t *src, int src_stride,
+                                          const uint8_t *ref, int ref_stride,
+                                          const uint8_t *second_pred) {
+  const int left_shift = 5;
+  uint32_t sum = aom_highbd_sad32x32_avg_avx2(src, src_stride, ref, ref_stride,
+                                              second_pred);
+  src += src_stride << left_shift;
+  ref += ref_stride << left_shift;
+  second_pred += 32 << left_shift;
+  sum += aom_highbd_sad32x32_avg_avx2(src, src_stride, ref, ref_stride,
+                                      second_pred);
+  return sum;
+}
+
+unsigned int aom_highbd_sad64x16_avg_avx2(const uint8_t *src, int src_stride,
+                                          const uint8_t *ref, int ref_stride,
+                                          const uint8_t *second_pred) {
+  __m256i sad = _mm256_setzero_si256();
+  uint16_t *srcp = CONVERT_TO_SHORTPTR(src);
+  uint16_t *refp = CONVERT_TO_SHORTPTR(ref);
+  uint16_t *secp = CONVERT_TO_SHORTPTR(second_pred);
+  const int left_shift = 1;
+  int row_section = 0;
+
+  while (row_section < 8) {
+    sad64x2(srcp, src_stride, refp, ref_stride, secp, &sad);
+    srcp += src_stride << left_shift;
+    refp += ref_stride << left_shift;
+    secp += 64 << left_shift;
+    row_section += 1;
+  }
+  return get_sad_from_mm256_epi32(&sad);
+}
+
+unsigned int aom_highbd_sad64x32_avg_avx2(const uint8_t *src, int src_stride,
+                                          const uint8_t *ref, int ref_stride,
+                                          const uint8_t *second_pred) {
+  __m256i sad = _mm256_setzero_si256();
+  uint16_t *srcp = CONVERT_TO_SHORTPTR(src);
+  uint16_t *refp = CONVERT_TO_SHORTPTR(ref);
+  uint16_t *secp = CONVERT_TO_SHORTPTR(second_pred);
+  const int left_shift = 1;
+  int row_section = 0;
+
+  while (row_section < 16) {
+    sad64x2(srcp, src_stride, refp, ref_stride, secp, &sad);
+    srcp += src_stride << left_shift;
+    refp += ref_stride << left_shift;
+    secp += 64 << left_shift;
+    row_section += 1;
+  }
+  return get_sad_from_mm256_epi32(&sad);
+}
+
+unsigned int aom_highbd_sad64x64_avg_avx2(const uint8_t *src, int src_stride,
+                                          const uint8_t *ref, int ref_stride,
+                                          const uint8_t *second_pred) {
+  const int left_shift = 5;
+  uint32_t sum = aom_highbd_sad64x32_avg_avx2(src, src_stride, ref, ref_stride,
+                                              second_pred);
+  src += src_stride << left_shift;
+  ref += ref_stride << left_shift;
+  second_pred += 64 << left_shift;
+  sum += aom_highbd_sad64x32_avg_avx2(src, src_stride, ref, ref_stride,
+                                      second_pred);
+  return sum;
+}
+
+unsigned int aom_highbd_sad64x128_avg_avx2(const uint8_t *src, int src_stride,
+                                           const uint8_t *ref, int ref_stride,
+                                           const uint8_t *second_pred) {
+  const int left_shift = 6;
+  uint32_t sum = aom_highbd_sad64x64_avg_avx2(src, src_stride, ref, ref_stride,
+                                              second_pred);
+  src += src_stride << left_shift;
+  ref += ref_stride << left_shift;
+  second_pred += 64 << left_shift;
+  sum += aom_highbd_sad64x64_avg_avx2(src, src_stride, ref, ref_stride,
+                                      second_pred);
+  return sum;
+}
+
+unsigned int aom_highbd_sad128x64_avg_avx2(const uint8_t *src, int src_stride,
+                                           const uint8_t *ref, int ref_stride,
+                                           const uint8_t *second_pred) {
+  __m256i sad = _mm256_setzero_si256();
+  uint16_t *srcp = CONVERT_TO_SHORTPTR(src);
+  uint16_t *refp = CONVERT_TO_SHORTPTR(ref);
+  uint16_t *secp = CONVERT_TO_SHORTPTR(second_pred);
+  int row = 0;
+  while (row < 64) {
+    sad128x1(srcp, refp, secp, &sad);
+    srcp += src_stride;
+    refp += ref_stride;
+    secp += 16 << 3;
+    row += 1;
+  }
+  return get_sad_from_mm256_epi32(&sad);
+}
+
+unsigned int aom_highbd_sad128x128_avg_avx2(const uint8_t *src, int src_stride,
+                                            const uint8_t *ref, int ref_stride,
+                                            const uint8_t *second_pred) {
+  unsigned int sum;
+  const int left_shift = 6;
+
+  sum = aom_highbd_sad128x64_avg_avx2(src, src_stride, ref, ref_stride,
+                                      second_pred);
+  src += src_stride << left_shift;
+  ref += ref_stride << left_shift;
+  second_pred += 128 << left_shift;
+  sum += aom_highbd_sad128x64_avg_avx2(src, src_stride, ref, ref_stride,
+                                       second_pred);
+  return sum;
+}
+
+// SAD 4D
+// Combine 4 __m256i input vectors  v to uint32_t result[4]
+static INLINE void get_4d_sad_from_mm256_epi32(const __m256i *v,
+                                               uint32_t *res) {
+  __m256i u0, u1, u2, u3;
+  const __m256i mask = yy_set1_64_from_32i(~0);
+  __m128i sad;
+
+  // 8 32-bit summation
+  u0 = _mm256_srli_si256(v[0], 4);
+  u1 = _mm256_srli_si256(v[1], 4);
+  u2 = _mm256_srli_si256(v[2], 4);
+  u3 = _mm256_srli_si256(v[3], 4);
+
+  u0 = _mm256_add_epi32(u0, v[0]);
+  u1 = _mm256_add_epi32(u1, v[1]);
+  u2 = _mm256_add_epi32(u2, v[2]);
+  u3 = _mm256_add_epi32(u3, v[3]);
+
+  u0 = _mm256_and_si256(u0, mask);
+  u1 = _mm256_and_si256(u1, mask);
+  u2 = _mm256_and_si256(u2, mask);
+  u3 = _mm256_and_si256(u3, mask);
+  // 4 32-bit summation, evenly positioned
+
+  u1 = _mm256_slli_si256(u1, 4);
+  u3 = _mm256_slli_si256(u3, 4);
+
+  u0 = _mm256_or_si256(u0, u1);
+  u2 = _mm256_or_si256(u2, u3);
+  // 8 32-bit summation, interleaved
+
+  u1 = _mm256_unpacklo_epi64(u0, u2);
+  u3 = _mm256_unpackhi_epi64(u0, u2);
+
+  u0 = _mm256_add_epi32(u1, u3);
+  sad = _mm_add_epi32(_mm256_extractf128_si256(u0, 1),
+                      _mm256_castsi256_si128(u0));
+  _mm_storeu_si128((__m128i *)res, sad);
+}
+
+static void convert_pointers(const uint8_t *const ref8[],
+                             const uint16_t *ref[]) {
+  ref[0] = CONVERT_TO_SHORTPTR(ref8[0]);
+  ref[1] = CONVERT_TO_SHORTPTR(ref8[1]);
+  ref[2] = CONVERT_TO_SHORTPTR(ref8[2]);
+  ref[3] = CONVERT_TO_SHORTPTR(ref8[3]);
+}
+
+static void init_sad(__m256i *s) {
+  s[0] = _mm256_setzero_si256();
+  s[1] = _mm256_setzero_si256();
+  s[2] = _mm256_setzero_si256();
+  s[3] = _mm256_setzero_si256();
+}
+
+static AOM_FORCE_INLINE void aom_highbd_sadMxNxD_avx2(
+    int M, int N, int D, const uint8_t *src, int src_stride,
+    const uint8_t *const ref_array[4], int ref_stride, uint32_t sad_array[4]) {
+  __m256i sad_vec[4];
+  const uint16_t *refp[4];
+  const uint16_t *keep = CONVERT_TO_SHORTPTR(src);
+  const uint16_t *srcp;
+  const int shift_for_rows = (M < 128) + (M < 64);
+  const int row_units = 1 << shift_for_rows;
+  int i, r;
+
+  init_sad(sad_vec);
+  convert_pointers(ref_array, refp);
+
+  for (i = 0; i < D; ++i) {
+    srcp = keep;
+    for (r = 0; r < N; r += row_units) {
+      if (M == 128) {
+        sad128x1(srcp, refp[i], NULL, &sad_vec[i]);
+      } else if (M == 64) {
+        sad64x2(srcp, src_stride, refp[i], ref_stride, NULL, &sad_vec[i]);
+      } else if (M == 32) {
+        sad32x4(srcp, src_stride, refp[i], ref_stride, 0, &sad_vec[i]);
+      } else if (M == 16) {
+        sad16x4(srcp, src_stride, refp[i], ref_stride, 0, &sad_vec[i]);
+      } else {
+        assert(0);
+      }
+      srcp += src_stride << shift_for_rows;
+      refp[i] += ref_stride << shift_for_rows;
+    }
+  }
+  get_4d_sad_from_mm256_epi32(sad_vec, sad_array);
+}
+
+#define HIGHBD_SAD_MXNX4D_AVX2(m, n)                                          \
+  void aom_highbd_sad##m##x##n##x4d_avx2(                                     \
+      const uint8_t *src, int src_stride, const uint8_t *const ref_array[4],  \
+      int ref_stride, uint32_t sad_array[4]) {                                \
+    aom_highbd_sadMxNxD_avx2(m, n, 4, src, src_stride, ref_array, ref_stride, \
+                             sad_array);                                      \
+  }
+#define HIGHBD_SAD_SKIP_MXNX4D_AVX2(m, n)                                    \
+  void aom_highbd_sad_skip_##m##x##n##x4d_avx2(                              \
+      const uint8_t *src, int src_stride, const uint8_t *const ref_array[4], \
+      int ref_stride, uint32_t sad_array[4]) {                               \
+    aom_highbd_sadMxNxD_avx2(m, (n / 2), 4, src, 2 * src_stride, ref_array,  \
+                             2 * ref_stride, sad_array);                     \
+    sad_array[0] <<= 1;                                                      \
+    sad_array[1] <<= 1;                                                      \
+    sad_array[2] <<= 1;                                                      \
+    sad_array[3] <<= 1;                                                      \
+  }
+#define HIGHBD_SAD_MXNX3D_AVX2(m, n)                                          \
+  void aom_highbd_sad##m##x##n##x3d_avx2(                                     \
+      const uint8_t *src, int src_stride, const uint8_t *const ref_array[4],  \
+      int ref_stride, uint32_t sad_array[4]) {                                \
+    aom_highbd_sadMxNxD_avx2(m, n, 3, src, src_stride, ref_array, ref_stride, \
+                             sad_array);                                      \
+  }
+
+HIGHBD_SAD_MXNX4D_AVX2(16, 4)
+HIGHBD_SAD_MXNX4D_AVX2(16, 8)
+HIGHBD_SAD_MXNX4D_AVX2(16, 16)
+HIGHBD_SAD_MXNX4D_AVX2(16, 32)
+HIGHBD_SAD_MXNX4D_AVX2(16, 64)
+
+HIGHBD_SAD_MXNX4D_AVX2(32, 8)
+HIGHBD_SAD_MXNX4D_AVX2(32, 16)
+HIGHBD_SAD_MXNX4D_AVX2(32, 32)
+HIGHBD_SAD_MXNX4D_AVX2(32, 64)
+
+HIGHBD_SAD_MXNX4D_AVX2(64, 16)
+HIGHBD_SAD_MXNX4D_AVX2(64, 32)
+HIGHBD_SAD_MXNX4D_AVX2(64, 64)
+HIGHBD_SAD_MXNX4D_AVX2(64, 128)
+
+HIGHBD_SAD_MXNX4D_AVX2(128, 64)
+HIGHBD_SAD_MXNX4D_AVX2(128, 128)
+
+HIGHBD_SAD_SKIP_MXNX4D_AVX2(16, 8)
+HIGHBD_SAD_SKIP_MXNX4D_AVX2(16, 16)
+HIGHBD_SAD_SKIP_MXNX4D_AVX2(16, 32)
+HIGHBD_SAD_SKIP_MXNX4D_AVX2(16, 64)
+
+HIGHBD_SAD_SKIP_MXNX4D_AVX2(32, 8)
+HIGHBD_SAD_SKIP_MXNX4D_AVX2(32, 16)
+HIGHBD_SAD_SKIP_MXNX4D_AVX2(32, 32)
+HIGHBD_SAD_SKIP_MXNX4D_AVX2(32, 64)
+
+HIGHBD_SAD_SKIP_MXNX4D_AVX2(64, 16)
+HIGHBD_SAD_SKIP_MXNX4D_AVX2(64, 32)
+HIGHBD_SAD_SKIP_MXNX4D_AVX2(64, 64)
+HIGHBD_SAD_SKIP_MXNX4D_AVX2(64, 128)
+
+HIGHBD_SAD_SKIP_MXNX4D_AVX2(128, 64)
+HIGHBD_SAD_SKIP_MXNX4D_AVX2(128, 128)
+
+HIGHBD_SAD_MXNX3D_AVX2(16, 4)
+HIGHBD_SAD_MXNX3D_AVX2(16, 8)
+HIGHBD_SAD_MXNX3D_AVX2(16, 16)
+HIGHBD_SAD_MXNX3D_AVX2(16, 32)
+HIGHBD_SAD_MXNX3D_AVX2(16, 64)
+
+HIGHBD_SAD_MXNX3D_AVX2(32, 8)
+HIGHBD_SAD_MXNX3D_AVX2(32, 16)
+HIGHBD_SAD_MXNX3D_AVX2(32, 32)
+HIGHBD_SAD_MXNX3D_AVX2(32, 64)
+
+HIGHBD_SAD_MXNX3D_AVX2(64, 16)
+HIGHBD_SAD_MXNX3D_AVX2(64, 32)
+HIGHBD_SAD_MXNX3D_AVX2(64, 64)
+HIGHBD_SAD_MXNX3D_AVX2(64, 128)
+
+HIGHBD_SAD_MXNX3D_AVX2(128, 64)
+HIGHBD_SAD_MXNX3D_AVX2(128, 128)
diff --git a/third_party/aom/aom_dsp/x86/highbd_sad_sse2.asm b/third_party/aom/aom_dsp/x86/highbd_sad_sse2.asm
new file mode 100644
index 0000000000..3dc4e4e0a2
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/highbd_sad_sse2.asm
@@ -0,0 +1,524 @@
+;
+; Copyright (c) 2016, 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 "third_party/x86inc/x86inc.asm"
+
+SECTION .text
+
+; Macro Arguments
+; Arg 1: Width
+; Arg 2: Height
+; Arg 3: Number of general purpose registers: 5 for 32-bit build, 6 for 64-bit
+; Arg 4: Type of function: if 0, normal sad; if 1, avg; if 2, skip rows
+; Arg 5: Number of xmm registers. 8xh needs 8, others only need 7
+%macro HIGH_SAD_FN 4-5 7
+%if %4 == 0
+%if %3 == 5
+cglobal highbd_sad%1x%2, 4, %3, %5, src, src_stride, ref, ref_stride, n_rows
+%else ; %3 == 7
+cglobal highbd_sad%1x%2, 4, %3, %5, src, src_stride, ref, ref_stride, \
+                            src_stride3, ref_stride3, n_rows
+%endif ; %3 == 5/7
+%elif %4 == 1 ; avg
+%if %3 == 5
+cglobal highbd_sad%1x%2_avg, 5, 1 + %3, %5, src, src_stride, ref, ref_stride, \
+                                    second_pred, n_rows
+%else ; %3 == 7
+cglobal highbd_sad%1x%2_avg, 5, AOM_ARCH_X86_64 + %3, %5, src, src_stride, \
+                                              ref, ref_stride, \
+                                              second_pred, \
+                                              src_stride3, ref_stride3
+%if AOM_ARCH_X86_64
+%define n_rowsd r7d
+%else ; x86-32
+%define n_rowsd dword r0m
+%endif ; x86-32/64
+%endif ; %3 == 5/7
+%else  ; %4 == 2, skip rows
+%if %3 == 5
+cglobal highbd_sad_skip_%1x%2, 4, %3, %5, src, src_stride, ref, ref_stride, n_rows
+%else ; %3 == 7
+cglobal highbd_sad_skip_%1x%2, 4, %3, %5, src, src_stride, ref, ref_stride, \
+                            src_stride3, ref_stride3, n_rows
+%endif ; %3 == 5/7
+%endif ; sad/avg/skip
+%if %4 == 2  ; double the stride if we are skipping rows
+  lea          src_strided, [src_strided*2]
+  lea          ref_strided, [ref_strided*2]
+%endif
+  movsxdifnidn src_strideq, src_strided
+  movsxdifnidn ref_strideq, ref_strided
+%if %3 == 7
+  lea         src_stride3q, [src_strideq*3]
+  lea         ref_stride3q, [ref_strideq*3]
+%endif ; %3 == 7
+; convert src, ref & second_pred to short ptrs (from byte ptrs)
+  shl                 srcq, 1
+  shl                 refq, 1
+%if %4 == 1
+  shl         second_predq, 1
+%endif
+%endmacro
+
+; unsigned int aom_highbd_sad64x{16,32,64}_sse2(uint8_t *src, int src_stride,
+;                                    uint8_t *ref, int ref_stride);
+%macro HIGH_SAD64XN 1-2 0
+  HIGH_SAD_FN 64, %1, 5, %2
+%if %2 == 2  ; skip rows, so divide number of rows by 2
+  mov              n_rowsd, %1/2
+%else
+  mov              n_rowsd, %1
+%endif
+  pxor                  m0, m0
+  pxor                  m6, m6
+
+.loop:
+  ; first half of each row
+  movu                  m1, [refq]
+  movu                  m2, [refq+16]
+  movu                  m3, [refq+32]
+  movu                  m4, [refq+48]
+%if %2 == 1
+  pavgw                 m1, [second_predq+mmsize*0]
+  pavgw                 m2, [second_predq+mmsize*1]
+  pavgw                 m3, [second_predq+mmsize*2]
+  pavgw                 m4, [second_predq+mmsize*3]
+  lea         second_predq, [second_predq+mmsize*4]
+%endif
+  mova                  m5, [srcq]
+  psubusw               m5, m1
+  psubusw               m1, [srcq]
+  por                   m1, m5
+  mova                  m5, [srcq+16]
+  psubusw               m5, m2
+  psubusw               m2, [srcq+16]
+  por                   m2, m5
+  mova                  m5, [srcq+32]
+  psubusw               m5, m3
+  psubusw               m3, [srcq+32]
+  por                   m3, m5
+  mova                  m5, [srcq+48]
+  psubusw               m5, m4
+  psubusw               m4, [srcq+48]
+  por                   m4, m5
+  paddw                 m1, m2
+  paddw                 m3, m4
+  movhlps               m2, m1
+  movhlps               m4, m3
+  paddw                 m1, m2
+  paddw                 m3, m4
+  punpcklwd             m1, m6
+  punpcklwd             m3, m6
+  paddd                 m0, m1
+  paddd                 m0, m3
+  ; second half of each row
+  movu                  m1, [refq+64]
+  movu                  m2, [refq+80]
+  movu                  m3, [refq+96]
+  movu                  m4, [refq+112]
+%if %2 == 1
+  pavgw                 m1, [second_predq+mmsize*0]
+  pavgw                 m2, [second_predq+mmsize*1]
+  pavgw                 m3, [second_predq+mmsize*2]
+  pavgw                 m4, [second_predq+mmsize*3]
+  lea         second_predq, [second_predq+mmsize*4]
+%endif
+  mova                  m5, [srcq+64]
+  psubusw               m5, m1
+  psubusw               m1, [srcq+64]
+  por                   m1, m5
+  mova                  m5, [srcq+80]
+  psubusw               m5, m2
+  psubusw               m2, [srcq+80]
+  por                   m2, m5
+  mova                  m5, [srcq+96]
+  psubusw               m5, m3
+  psubusw               m3, [srcq+96]
+  por                   m3, m5
+  mova                  m5, [srcq+112]
+  psubusw               m5, m4
+  psubusw               m4, [srcq+112]
+  por                   m4, m5
+  paddw                 m1, m2
+  paddw                 m3, m4
+  movhlps               m2, m1
+  movhlps               m4, m3
+  paddw                 m1, m2
+  paddw                 m3, m4
+  punpcklwd             m1, m6
+  punpcklwd             m3, m6
+  lea                 refq, [refq+ref_strideq*2]
+  paddd                 m0, m1
+  lea                 srcq, [srcq+src_strideq*2]
+  paddd                 m0, m3
+
+  dec              n_rowsd
+  jg .loop
+
+  movhlps               m1, m0
+  paddd                 m0, m1
+  punpckldq             m0, m6
+  movhlps               m1, m0
+  paddd                 m0, m1
+%if %2 == 2  ; we skipped rows, so we need to double the sad
+  pslld                 m0, 1
+%endif
+  movd                 eax, m0
+  RET
+%endmacro
+
+INIT_XMM sse2
+HIGH_SAD64XN 64 ; highbd_sad64x64_sse2
+HIGH_SAD64XN 32 ; highbd_sad64x32_sse2
+HIGH_SAD64XN 16 ; highbd_sad_64x16_sse2
+HIGH_SAD64XN 64, 1 ; highbd_sad64x64_avg_sse2
+HIGH_SAD64XN 32, 1 ; highbd_sad64x32_avg_sse2
+HIGH_SAD64XN 16, 1 ; highbd_sad_64x16_avg_sse2
+HIGH_SAD64XN 64, 2 ; highbd_sad_skip_64x64_sse2
+HIGH_SAD64XN 32, 2 ; highbd_sad_skip_64x32_sse2
+HIGH_SAD64XN 16, 2 ; highbd_sad_skip_64x16_sse2
+
+; unsigned int aom_highbd_sad32x{16,32,64}_sse2(uint8_t *src, int src_stride,
+;                                    uint8_t *ref, int ref_stride);
+%macro HIGH_SAD32XN 1-2 0
+  HIGH_SAD_FN 32, %1, 5, %2
+%if %2 == 2  ; skip rows, so divide number of rows by 2
+  mov              n_rowsd, %1/2
+%else
+  mov              n_rowsd, %1
+%endif
+  pxor                  m0, m0
+  pxor                  m6, m6
+
+.loop:
+  movu                  m1, [refq]
+  movu                  m2, [refq+16]
+  movu                  m3, [refq+32]
+  movu                  m4, [refq+48]
+%if %2 == 1
+  pavgw                 m1, [second_predq+mmsize*0]
+  pavgw                 m2, [second_predq+mmsize*1]
+  pavgw                 m3, [second_predq+mmsize*2]
+  pavgw                 m4, [second_predq+mmsize*3]
+  lea         second_predq, [second_predq+mmsize*4]
+%endif
+  mova                  m5, [srcq]
+  psubusw               m5, m1
+  psubusw               m1, [srcq]
+  por                   m1, m5
+  mova                  m5, [srcq+16]
+  psubusw               m5, m2
+  psubusw               m2, [srcq+16]
+  por                   m2, m5
+  mova                  m5, [srcq+32]
+  psubusw               m5, m3
+  psubusw               m3, [srcq+32]
+  por                   m3, m5
+  mova                  m5, [srcq+48]
+  psubusw               m5, m4
+  psubusw               m4, [srcq+48]
+  por                   m4, m5
+  paddw                 m1, m2
+  paddw                 m3, m4
+  movhlps               m2, m1
+  movhlps               m4, m3
+  paddw                 m1, m2
+  paddw                 m3, m4
+  punpcklwd             m1, m6
+  punpcklwd             m3, m6
+  lea                 refq, [refq+ref_strideq*2]
+  paddd                 m0, m1
+  lea                 srcq, [srcq+src_strideq*2]
+  paddd                 m0, m3
+  dec              n_rowsd
+  jg .loop
+
+  movhlps               m1, m0
+  paddd                 m0, m1
+  punpckldq             m0, m6
+  movhlps               m1, m0
+  paddd                 m0, m1
+%if %2 == 2  ; we skipped rows, so we need to double the sad
+  pslld                 m0, 1
+%endif
+  movd                 eax, m0
+  RET
+%endmacro
+
+INIT_XMM sse2
+HIGH_SAD32XN 64 ; highbd_sad32x64_sse2
+HIGH_SAD32XN 32 ; highbd_sad32x32_sse2
+HIGH_SAD32XN 16 ; highbd_sad32x16_sse2
+HIGH_SAD32XN  8 ; highbd_sad_32x8_sse2
+HIGH_SAD32XN 64, 1 ; highbd_sad32x64_avg_sse2
+HIGH_SAD32XN 32, 1 ; highbd_sad32x32_avg_sse2
+HIGH_SAD32XN 16, 1 ; highbd_sad32x16_avg_sse2
+HIGH_SAD32XN  8, 1 ; highbd_sad_32x8_avg_sse2
+HIGH_SAD32XN 64, 2 ; highbd_sad_skip_32x64_sse2
+HIGH_SAD32XN 32, 2 ; highbd_sad_skip_32x32_sse2
+HIGH_SAD32XN 16, 2 ; highbd_sad_skip_32x16_sse2
+HIGH_SAD32XN  8, 2 ; highbd_sad_skip_32x8_sse2
+
+; unsigned int aom_highbd_sad16x{8,16,32}_sse2(uint8_t *src, int src_stride,
+;                                    uint8_t *ref, int ref_stride);
+%macro HIGH_SAD16XN 1-2 0
+  HIGH_SAD_FN 16, %1, 5, %2
+%if %2 == 2  ; skip rows, so divide number of rows by 2
+  mov              n_rowsd, %1/4
+%else
+  mov              n_rowsd, %1/2
+%endif
+  pxor                  m0, m0
+  pxor                  m6, m6
+
+.loop:
+  movu                  m1, [refq]
+  movu                  m2, [refq+16]
+  movu                  m3, [refq+ref_strideq*2]
+  movu                  m4, [refq+ref_strideq*2+16]
+%if %2 == 1
+  pavgw                 m1, [second_predq+mmsize*0]
+  pavgw                 m2, [second_predq+16]
+  pavgw                 m3, [second_predq+mmsize*2]
+  pavgw                 m4, [second_predq+mmsize*2+16]
+  lea         second_predq, [second_predq+mmsize*4]
+%endif
+  mova                  m5, [srcq]
+  psubusw               m5, m1
+  psubusw               m1, [srcq]
+  por                   m1, m5
+  mova                  m5, [srcq+16]
+  psubusw               m5, m2
+  psubusw               m2, [srcq+16]
+  por                   m2, m5
+  mova                  m5, [srcq+src_strideq*2]
+  psubusw               m5, m3
+  psubusw               m3, [srcq+src_strideq*2]
+  por                   m3, m5
+  mova                  m5, [srcq+src_strideq*2+16]
+  psubusw               m5, m4
+  psubusw               m4, [srcq+src_strideq*2+16]
+  por                   m4, m5
+  paddw                 m1, m2
+  paddw                 m3, m4
+  movhlps               m2, m1
+  movhlps               m4, m3
+  paddw                 m1, m2
+  paddw                 m3, m4
+  punpcklwd             m1, m6
+  punpcklwd             m3, m6
+  lea                 refq, [refq+ref_strideq*4]
+  paddd                 m0, m1
+  lea                 srcq, [srcq+src_strideq*4]
+  paddd                 m0, m3
+  dec              n_rowsd
+  jg .loop
+
+  movhlps               m1, m0
+  paddd                 m0, m1
+  punpckldq             m0, m6
+  movhlps               m1, m0
+  paddd                 m0, m1
+%if %2 == 2  ; we skipped rows, so we need to double the sad
+  pslld                 m0, 1
+%endif
+  movd                 eax, m0
+  RET
+%endmacro
+
+INIT_XMM sse2
+HIGH_SAD16XN 64 ; highbd_sad_16x64_sse2
+HIGH_SAD16XN 32 ; highbd_sad16x32_sse2
+HIGH_SAD16XN 16 ; highbd_sad16x16_sse2
+HIGH_SAD16XN  8 ; highbd_sad16x8_sse2
+HIGH_SAD16XN  4 ; highbd_sad_16x4_sse2
+HIGH_SAD16XN 64, 1 ; highbd_sad_16x64_avg_sse2
+HIGH_SAD16XN 32, 1 ; highbd_sad16x32_avg_sse2
+HIGH_SAD16XN 16, 1 ; highbd_sad16x16_avg_sse2
+HIGH_SAD16XN  8, 1 ; highbd_sad16x8_avg_sse2
+HIGH_SAD16XN  4, 1 ; highbd_sad_16x4_avg_sse2
+HIGH_SAD16XN 64, 2 ; highbd_sad_skip_16x64_sse2
+HIGH_SAD16XN 32, 2 ; highbd_sad_skip_16x32_sse2
+HIGH_SAD16XN 16, 2 ; highbd_sad_skip_16x16_sse2
+HIGH_SAD16XN  8, 2 ; highbd_sad_skip_16x8_sse2
+; Current code fails there are only 2 rows
+; HIGH_SAD16XN  4, 2 ; highbd_sad_skip_16x4_sse2
+
+; unsigned int aom_highbd_sad8x{4,8,16}_sse2(uint8_t *src, int src_stride,
+;                                    uint8_t *ref, int ref_stride);
+%macro HIGH_SAD8XN 1-2 0
+  HIGH_SAD_FN 8, %1, 7, %2, 8
+%if %2 == 2  ; skip rows, so divide number of rows by 2
+  mov              n_rowsd, %1/8
+%else
+  mov              n_rowsd, %1/4
+%endif
+  pxor                  m0, m0
+  pxor                  m6, m6
+
+.loop:
+  movu                  m1, [refq]
+  movu                  m2, [refq+ref_strideq*2]
+  movu                  m3, [refq+ref_strideq*4]
+  movu                  m4, [refq+ref_stride3q*2]
+%if %2 == 1
+  pavgw                 m1, [second_predq+mmsize*0]
+  pavgw                 m2, [second_predq+mmsize*1]
+  pavgw                 m3, [second_predq+mmsize*2]
+  pavgw                 m4, [second_predq+mmsize*3]
+  lea         second_predq, [second_predq+mmsize*4]
+%endif
+  mova                  m7, m1
+  movu                  m5, [srcq]
+  psubusw               m1, m5
+  psubusw               m5, m7
+  por                   m1, m5
+
+  mova                  m7, m2
+  movu                  m5, [srcq+src_strideq*2]
+  psubusw               m2, m5
+  psubusw               m5, m7
+  por                   m2, m5
+
+  mova                  m7, m3
+  movu                  m5, [srcq+src_strideq*4]
+  psubusw               m3, m5
+  psubusw               m5, m7
+  por                   m3, m5
+
+  mova                  m7, m4
+  movu                  m5, [srcq+src_stride3q*2]
+  psubusw               m4, m5
+  psubusw               m5, m7
+  por                   m4, m5
+
+  paddw                 m1, m2
+  paddw                 m3, m4
+  movhlps               m2, m1
+  movhlps               m4, m3
+  paddw                 m1, m2
+  paddw                 m3, m4
+  punpcklwd             m1, m6
+  punpcklwd             m3, m6
+  lea                 refq, [refq+ref_strideq*8]
+  paddd                 m0, m1
+  lea                 srcq, [srcq+src_strideq*8]
+  paddd                 m0, m3
+  dec              n_rowsd
+  jg .loop
+
+  movhlps               m1, m0
+  paddd                 m0, m1
+  punpckldq             m0, m6
+  movhlps               m1, m0
+  paddd                 m0, m1
+%if %2 == 2  ; we skipped rows, so we need to double the sad
+  pslld                 m0, 1
+%endif
+  movd                 eax, m0
+  RET
+%endmacro
+
+INIT_XMM sse2
+HIGH_SAD8XN 32 ; highbd_sad_8x32_sse2
+HIGH_SAD8XN 16 ; highbd_sad8x16_sse2
+HIGH_SAD8XN  8 ; highbd_sad8x8_sse2
+HIGH_SAD8XN  4 ; highbd_sad8x4_sse2
+HIGH_SAD8XN 32, 1 ; highbd_sad_8x32_avg_sse2
+HIGH_SAD8XN 16, 1 ; highbd_sad8x16_avg_sse2
+HIGH_SAD8XN  8, 1 ; highbd_sad8x8_avg_sse2
+HIGH_SAD8XN  4, 1 ; highbd_sad8x4_avg_sse2
+HIGH_SAD8XN 32, 2 ; highbd_sad_skip_8x32_sse2
+HIGH_SAD8XN 16, 2 ; highbd_sad_skip_8x16_sse2
+HIGH_SAD8XN  8, 2 ; highbd_sad_skip_8x8_sse2
+; Current code fails there are only 2 rows
+; HIGH_SAD8XN  4, 2 ; highbd_sad8x4_avg_sse2
+
+; unsigned int aom_highbd_sad4x{4,8,16}_sse2(uint8_t *src, int src_stride,
+;                                    uint8_t *ref, int ref_stride);
+%macro HIGH_SAD4XN 1-2 0
+  HIGH_SAD_FN 4, %1, 7, %2
+%if %2 == 2  ; skip rows, so divide number of rows by 2
+  mov              n_rowsd, %1/8
+%else
+  mov              n_rowsd, %1/4
+%endif
+  pxor                  m0, m0
+  pxor                  m6, m6
+
+.loop:
+  movq                  m1, [refq]
+  movq                  m2, [refq+ref_strideq*2]
+  movq                  m3, [refq+ref_strideq*4]
+  movq                  m4, [refq+ref_stride3q*2]
+  punpcklwd             m1, m3
+  punpcklwd             m2, m4
+%if %2 == 1
+  movq                  m3, [second_predq+8*0]
+  movq                  m5, [second_predq+8*2]
+  punpcklwd             m3, m5
+  movq                  m4, [second_predq+8*1]
+  movq                  m5, [second_predq+8*3]
+  punpcklwd             m4, m5
+  lea         second_predq, [second_predq+8*4]
+  pavgw                 m1, m3
+  pavgw                 m2, m4
+%endif
+  movq                  m5, [srcq]
+  movq                  m3, [srcq+src_strideq*4]
+  punpcklwd             m5, m3
+  movdqa                m3, m1
+  psubusw               m1, m5
+  psubusw               m5, m3
+  por                   m1, m5
+  movq                  m5, [srcq+src_strideq*2]
+  movq                  m4, [srcq+src_stride3q*2]
+  punpcklwd             m5, m4
+  movdqa                m4, m2
+  psubusw               m2, m5
+  psubusw               m5, m4
+  por                   m2, m5
+  paddw                 m1, m2
+  movdqa                m2, m1
+  punpcklwd             m1, m6
+  punpckhwd             m2, m6
+  lea                 refq, [refq+ref_strideq*8]
+  paddd                 m0, m1
+  lea                 srcq, [srcq+src_strideq*8]
+  paddd                 m0, m2
+  dec              n_rowsd
+  jg .loop
+
+  movhlps               m1, m0
+  paddd                 m0, m1
+  punpckldq             m0, m6
+  movhlps               m1, m0
+  paddd                 m0, m1
+%if %2 == 2  ; we skipped rows, so we need to double the sad
+  pslld                 m0, 1
+%endif
+  movd                 eax, m0
+  RET
+%endmacro
+
+INIT_XMM sse2
+HIGH_SAD4XN 16 ; highbd_sad4x16_sse2
+HIGH_SAD4XN  8 ; highbd_sad4x8_sse2
+HIGH_SAD4XN  4 ; highbd_sad4x4_sse2
+HIGH_SAD4XN 16, 1 ; highbd_sad4x16_avg_sse2
+HIGH_SAD4XN  8, 1 ; highbd_sad4x8_avg_sse2
+HIGH_SAD4XN  4, 1 ; highbd_sad4x4_avg_sse2
+HIGH_SAD4XN 16, 2 ; highbd_sad_skip_4x16_sse2
+HIGH_SAD4XN  8, 2 ; highbd_sad_skip_4x8_sse2
+; Current code fails there are only 2 rows
+; HIGH_SAD4XN  4, 2 ; highbd_sad_skip_4x4_sse2
diff --git a/third_party/aom/aom_dsp/x86/highbd_subpel_variance_impl_sse2.asm b/third_party/aom/aom_dsp/x86/highbd_subpel_variance_impl_sse2.asm
new file mode 100644
index 0000000000..c0ccc182b4
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/highbd_subpel_variance_impl_sse2.asm
@@ -0,0 +1,1024 @@
+;
+; Copyright (c) 2016, 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 "third_party/x86inc/x86inc.asm"
+
+SECTION_RODATA
+pw_8: times  8 dw  8
+bilin_filter_m_sse2: times  8 dw 16
+                     times  8 dw  0
+                     times  8 dw 14
+                     times  8 dw  2
+                     times  8 dw 12
+                     times  8 dw  4
+                     times  8 dw 10
+                     times  8 dw  6
+                     times 16 dw  8
+                     times  8 dw  6
+                     times  8 dw 10
+                     times  8 dw  4
+                     times  8 dw 12
+                     times  8 dw  2
+                     times  8 dw 14
+
+SECTION .text
+
+; int aom_sub_pixel_varianceNxh(const uint8_t *src, ptrdiff_t src_stride,
+;                               int x_offset, int y_offset,
+;                               const uint8_t *dst, ptrdiff_t dst_stride,
+;                               int height, unsigned int *sse);
+;
+; This function returns the SE and stores SSE in the given pointer.
+
+%macro SUM_SSE 6 ; src1, dst1, src2, dst2, sum, sse
+  psubw                %3, %4
+  psubw                %1, %2
+  mova                 %4, %3       ; make copies to manipulate to calc sum
+  mova                 %2, %1       ; use originals for calc sse
+  pmaddwd              %3, %3
+  paddw                %4, %2
+  pmaddwd              %1, %1
+  movhlps              %2, %4
+  paddd                %6, %3
+  paddw                %4, %2
+  pxor                 %2, %2
+  pcmpgtw              %2, %4       ; mask for 0 > %4 (sum)
+  punpcklwd            %4, %2       ; sign-extend word to dword
+  paddd                %6, %1
+  paddd                %5, %4
+
+%endmacro
+
+%macro STORE_AND_RET 0
+%if mmsize == 16
+  ; if H=64 and W=16, we have 8 words of each 2(1bit)x64(6bit)x9bit=16bit
+  ; in m6, i.e. it _exactly_ fits in a signed word per word in the xmm reg.
+  ; We have to sign-extend it before adding the words within the register
+  ; and outputing to a dword.
+  movhlps              m3, m7
+  movhlps              m4, m6
+  paddd                m7, m3
+  paddd                m6, m4
+  pshufd               m3, m7, 0x1
+  pshufd               m4, m6, 0x1
+  paddd                m7, m3
+  paddd                m6, m4
+  mov                  r1, ssem         ; r1 = unsigned int *sse
+  movd               [r1], m7           ; store sse
+  movd                eax, m6           ; store sum as return value
+%endif
+  RET
+%endmacro
+
+%macro INC_SRC_BY_SRC_STRIDE  0
+%if AOM_ARCH_X86=1 && CONFIG_PIC=1
+  add                srcq, src_stridemp
+  add                srcq, src_stridemp
+%else
+  lea                srcq, [srcq + src_strideq*2]
+%endif
+%endmacro
+
+%macro SUBPEL_VARIANCE 1-2 0 ; W
+%define bilin_filter_m bilin_filter_m_sse2
+%define filter_idx_shift 5
+
+
+%if AOM_ARCH_X86_64
+  %if %2 == 1 ; avg
+    cglobal highbd_sub_pixel_avg_variance%1xh, 9, 10, 13, src, src_stride, \
+                                      x_offset, y_offset, \
+                                      dst, dst_stride, \
+                                      sec, sec_stride, height, sse
+    %define sec_str sec_strideq
+  %else
+    cglobal highbd_sub_pixel_variance%1xh, 7, 8, 13, src, src_stride, \
+                                  x_offset, y_offset, \
+                                  dst, dst_stride, height, sse
+  %endif
+  %define block_height heightd
+  %define bilin_filter sseq
+%else
+  %if CONFIG_PIC=1
+    %if %2 == 1 ; avg
+      cglobal highbd_sub_pixel_avg_variance%1xh, 7, 7, 13, src, src_stride, \
+                                        x_offset, y_offset, \
+                                        dst, dst_stride, \
+                                        sec, sec_stride, height, sse
+      %define block_height dword heightm
+      %define sec_str sec_stridemp
+    %else
+      cglobal highbd_sub_pixel_variance%1xh, 7, 7, 13, src, src_stride, \
+                                    x_offset, y_offset, \
+                                    dst, dst_stride, height, sse
+      %define block_height heightd
+    %endif
+
+    ; reuse argument stack space
+    %define g_bilin_filterm x_offsetm
+    %define g_pw_8m y_offsetm
+
+    ; Store bilin_filter and pw_8 location in stack
+    %if GET_GOT_DEFINED == 1
+      GET_GOT eax
+      add esp, 4                ; restore esp
+    %endif
+
+    lea ecx, [GLOBAL(bilin_filter_m)]
+    mov g_bilin_filterm, ecx
+
+    lea ecx, [GLOBAL(pw_8)]
+    mov g_pw_8m, ecx
+
+    LOAD_IF_USED 0, 1         ; load eax, ecx back
+  %else
+    %if %2 == 1 ; avg
+      cglobal highbd_sub_pixel_avg_variance%1xh, 7, 7, 13, src, src_stride, \
+                                        x_offset, y_offset, \
+                                        dst, dst_stride, \
+                                        sec, sec_stride, height, sse
+      %define block_height dword heightm
+      %define sec_str sec_stridemp
+    %else
+      cglobal highbd_sub_pixel_variance%1xh, 7, 7, 13, src, src_stride, \
+                                    x_offset, y_offset, \
+                                    dst, dst_stride, height, sse
+      %define block_height heightd
+    %endif
+
+    %define bilin_filter bilin_filter_m
+  %endif
+%endif
+
+  ASSERT               %1 <= 16         ; m6 overflows if w > 16
+  pxor                 m6, m6           ; sum
+  pxor                 m7, m7           ; sse
+
+%if %1 < 16
+  sar                   block_height, 1
+%endif
+%if %2 == 1 ; avg
+  shl             sec_str, 1
+%endif
+
+  ; FIXME(rbultje) replace by jumptable?
+  test          x_offsetd, x_offsetd
+  jnz .x_nonzero
+  ; x_offset == 0
+  test          y_offsetd, y_offsetd
+  jnz .x_zero_y_nonzero
+
+  ; x_offset == 0 && y_offset == 0
+.x_zero_y_zero_loop:
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m2, [srcq + 16]
+  mova                 m1, [dstq]
+  mova                 m3, [dstq + 16]
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  pavgw                m2, [secq+16]
+%endif
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  lea                srcq, [srcq + src_strideq*2]
+  lea                dstq, [dstq + dst_strideq*2]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%else ; %1 < 16
+  movu                 m0, [srcq]
+  movu                 m2, [srcq + src_strideq*2]
+  mova                 m1, [dstq]
+  mova                 m3, [dstq + dst_strideq*2]
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  add                secq, sec_str
+  pavgw                m2, [secq]
+%endif
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  lea                srcq, [srcq + src_strideq*4]
+  lea                dstq, [dstq + dst_strideq*4]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%endif
+  dec                   block_height
+  jg .x_zero_y_zero_loop
+  STORE_AND_RET
+
+.x_zero_y_nonzero:
+  cmp           y_offsetd, 8
+  jne .x_zero_y_nonhalf
+
+  ; x_offset == 0 && y_offset == 0.5
+.x_zero_y_half_loop:
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq+16]
+  movu                 m4, [srcq+src_strideq*2]
+  movu                 m5, [srcq+src_strideq*2+16]
+  mova                 m2, [dstq]
+  mova                 m3, [dstq+16]
+  pavgw                m0, m4
+  pavgw                m1, m5
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  pavgw                m1, [secq+16]
+%endif
+  SUM_SSE              m0, m2, m1, m3, m6, m7
+
+  lea                srcq, [srcq + src_strideq*2]
+  lea                dstq, [dstq + dst_strideq*2]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%else ; %1 < 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq+src_strideq*2]
+  movu                 m5, [srcq+src_strideq*4]
+  mova                 m2, [dstq]
+  mova                 m3, [dstq+dst_strideq*2]
+  pavgw                m0, m1
+  pavgw                m1, m5
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  add                secq, sec_str
+  pavgw                m1, [secq]
+%endif
+  SUM_SSE              m0, m2, m1, m3, m6, m7
+
+  lea                srcq, [srcq + src_strideq*4]
+  lea                dstq, [dstq + dst_strideq*4]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%endif
+  dec                   block_height
+  jg .x_zero_y_half_loop
+  STORE_AND_RET
+
+.x_zero_y_nonhalf:
+  ; x_offset == 0 && y_offset == bilin interpolation
+%if AOM_ARCH_X86_64
+  lea        bilin_filter, [GLOBAL(bilin_filter_m)]
+%endif
+  shl           y_offsetd, filter_idx_shift
+%if AOM_ARCH_X86_64 && mmsize == 16
+  mova                 m8, [bilin_filter+y_offsetq]
+  mova                 m9, [bilin_filter+y_offsetq+16]
+  mova                m10, [GLOBAL(pw_8)]
+%define filter_y_a m8
+%define filter_y_b m9
+%define filter_rnd m10
+%else ; x86-32 or mmx
+%if AOM_ARCH_X86=1 && CONFIG_PIC=1
+; x_offset == 0, reuse x_offset reg
+%define tempq x_offsetq
+  add y_offsetq, g_bilin_filterm
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+  mov tempq, g_pw_8m
+%define filter_rnd [tempq]
+%else
+  add           y_offsetq, bilin_filter
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+%define filter_rnd [GLOBAL(pw_8)]
+%endif
+%endif
+
+.x_zero_y_other_loop:
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq + 16]
+  movu                 m4, [srcq+src_strideq*2]
+  movu                 m5, [srcq+src_strideq*2+16]
+  mova                 m2, [dstq]
+  mova                 m3, [dstq+16]
+  ; FIXME(rbultje) instead of out=((num-x)*in1+x*in2+rnd)>>log2(num), we can
+  ; also do out=in1+(((num-x)*(in2-in1)+rnd)>>log2(num)). Total number of
+  ; instructions is the same (5), but it is 1 mul instead of 2, so might be
+  ; slightly faster because of pmullw latency. It would also cut our rodata
+  ; tables in half for this function, and save 1-2 registers on x86-64.
+  pmullw               m1, filter_y_a
+  pmullw               m5, filter_y_b
+  paddw                m1, filter_rnd
+  pmullw               m0, filter_y_a
+  pmullw               m4, filter_y_b
+  paddw                m0, filter_rnd
+  paddw                m1, m5
+  paddw                m0, m4
+  psrlw                m1, 4
+  psrlw                m0, 4
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  pavgw                m1, [secq+16]
+%endif
+  SUM_SSE              m0, m2, m1, m3, m6, m7
+
+  lea                srcq, [srcq + src_strideq*2]
+  lea                dstq, [dstq + dst_strideq*2]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%else ; %1 < 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq+src_strideq*2]
+  movu                 m5, [srcq+src_strideq*4]
+  mova                 m4, m1
+  mova                 m2, [dstq]
+  mova                 m3, [dstq+dst_strideq*2]
+  pmullw               m1, filter_y_a
+  pmullw               m5, filter_y_b
+  paddw                m1, filter_rnd
+  pmullw               m0, filter_y_a
+  pmullw               m4, filter_y_b
+  paddw                m0, filter_rnd
+  paddw                m1, m5
+  paddw                m0, m4
+  psrlw                m1, 4
+  psrlw                m0, 4
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  add                secq, sec_str
+  pavgw                m1, [secq]
+%endif
+  SUM_SSE              m0, m2, m1, m3, m6, m7
+
+  lea                srcq, [srcq + src_strideq*4]
+  lea                dstq, [dstq + dst_strideq*4]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%endif
+  dec                   block_height
+  jg .x_zero_y_other_loop
+%undef filter_y_a
+%undef filter_y_b
+%undef filter_rnd
+  STORE_AND_RET
+
+.x_nonzero:
+  cmp           x_offsetd, 8
+  jne .x_nonhalf
+  ; x_offset == 0.5
+  test          y_offsetd, y_offsetd
+  jnz .x_half_y_nonzero
+
+  ; x_offset == 0.5 && y_offset == 0
+.x_half_y_zero_loop:
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq + 16]
+  movu                 m4, [srcq + 2]
+  movu                 m5, [srcq + 18]
+  mova                 m2, [dstq]
+  mova                 m3, [dstq + 16]
+  pavgw                m0, m4
+  pavgw                m1, m5
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  pavgw                m1, [secq+16]
+%endif
+  SUM_SSE              m0, m2, m1, m3, m6, m7
+
+  lea                srcq, [srcq + src_strideq*2]
+  lea                dstq, [dstq + dst_strideq*2]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%else ; %1 < 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq + src_strideq*2]
+  movu                 m4, [srcq + 2]
+  movu                 m5, [srcq + src_strideq*2 + 2]
+  mova                 m2, [dstq]
+  mova                 m3, [dstq + dst_strideq*2]
+  pavgw                m0, m4
+  pavgw                m1, m5
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  add                secq, sec_str
+  pavgw                m1, [secq]
+%endif
+  SUM_SSE              m0, m2, m1, m3, m6, m7
+
+  lea                srcq, [srcq + src_strideq*4]
+  lea                dstq, [dstq + dst_strideq*4]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%endif
+  dec                   block_height
+  jg .x_half_y_zero_loop
+  STORE_AND_RET
+
+.x_half_y_nonzero:
+  cmp           y_offsetd, 8
+  jne .x_half_y_nonhalf
+
+  ; x_offset == 0.5 && y_offset == 0.5
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq+16]
+  movu                 m2, [srcq+2]
+  movu                 m3, [srcq+18]
+  lea                srcq, [srcq + src_strideq*2]
+  pavgw                m0, m2
+  pavgw                m1, m3
+.x_half_y_half_loop:
+  movu                 m2, [srcq]
+  movu                 m3, [srcq + 16]
+  movu                 m4, [srcq + 2]
+  movu                 m5, [srcq + 18]
+  pavgw                m2, m4
+  pavgw                m3, m5
+  pavgw                m0, m2
+  pavgw                m1, m3
+  mova                 m4, [dstq]
+  mova                 m5, [dstq + 16]
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  pavgw                m1, [secq+16]
+%endif
+  SUM_SSE              m0, m4, m1, m5, m6, m7
+  mova                 m0, m2
+  mova                 m1, m3
+
+  lea                srcq, [srcq + src_strideq*2]
+  lea                dstq, [dstq + dst_strideq*2]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%else ; %1 < 16
+  movu                 m0, [srcq]
+  movu                 m2, [srcq+2]
+  lea                srcq, [srcq + src_strideq*2]
+  pavgw                m0, m2
+.x_half_y_half_loop:
+  movu                 m2, [srcq]
+  movu                 m3, [srcq + src_strideq*2]
+  movu                 m4, [srcq + 2]
+  movu                 m5, [srcq + src_strideq*2 + 2]
+  pavgw                m2, m4
+  pavgw                m3, m5
+  pavgw                m0, m2
+  pavgw                m2, m3
+  mova                 m4, [dstq]
+  mova                 m5, [dstq + dst_strideq*2]
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  add                secq, sec_str
+  pavgw                m2, [secq]
+%endif
+  SUM_SSE              m0, m4, m2, m5, m6, m7
+  mova                 m0, m3
+
+  lea                srcq, [srcq + src_strideq*4]
+  lea                dstq, [dstq + dst_strideq*4]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%endif
+  dec                   block_height
+  jg .x_half_y_half_loop
+  STORE_AND_RET
+
+.x_half_y_nonhalf:
+  ; x_offset == 0.5 && y_offset == bilin interpolation
+%if AOM_ARCH_X86_64
+  lea        bilin_filter, [GLOBAL(bilin_filter_m)]
+%endif
+  shl           y_offsetd, filter_idx_shift
+%if AOM_ARCH_X86_64 && mmsize == 16
+  mova                 m8, [bilin_filter+y_offsetq]
+  mova                 m9, [bilin_filter+y_offsetq+16]
+  mova                m10, [GLOBAL(pw_8)]
+%define filter_y_a m8
+%define filter_y_b m9
+%define filter_rnd m10
+%else  ; x86_32
+%if AOM_ARCH_X86=1 && CONFIG_PIC=1
+; x_offset == 0.5. We can reuse x_offset reg
+%define tempq x_offsetq
+  add y_offsetq, g_bilin_filterm
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+  mov tempq, g_pw_8m
+%define filter_rnd [tempq]
+%else
+  add           y_offsetq, bilin_filter
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+%define filter_rnd [GLOBAL(pw_8)]
+%endif
+%endif
+
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq+16]
+  movu                 m2, [srcq+2]
+  movu                 m3, [srcq+18]
+  lea                srcq, [srcq + src_strideq*2]
+  pavgw                m0, m2
+  pavgw                m1, m3
+.x_half_y_other_loop:
+  movu                 m2, [srcq]
+  movu                 m3, [srcq+16]
+  movu                 m4, [srcq+2]
+  movu                 m5, [srcq+18]
+  pavgw                m2, m4
+  pavgw                m3, m5
+  mova                 m4, m2
+  mova                 m5, m3
+  pmullw               m1, filter_y_a
+  pmullw               m3, filter_y_b
+  paddw                m1, filter_rnd
+  paddw                m1, m3
+  pmullw               m0, filter_y_a
+  pmullw               m2, filter_y_b
+  paddw                m0, filter_rnd
+  psrlw                m1, 4
+  paddw                m0, m2
+  mova                 m2, [dstq]
+  psrlw                m0, 4
+  mova                 m3, [dstq+16]
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  pavgw                m1, [secq+16]
+%endif
+  SUM_SSE              m0, m2, m1, m3, m6, m7
+  mova                 m0, m4
+  mova                 m1, m5
+
+  lea                srcq, [srcq + src_strideq*2]
+  lea                dstq, [dstq + dst_strideq*2]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%else ; %1 < 16
+  movu                 m0, [srcq]
+  movu                 m2, [srcq+2]
+  lea                srcq, [srcq + src_strideq*2]
+  pavgw                m0, m2
+.x_half_y_other_loop:
+  movu                 m2, [srcq]
+  movu                 m3, [srcq+src_strideq*2]
+  movu                 m4, [srcq+2]
+  movu                 m5, [srcq+src_strideq*2+2]
+  pavgw                m2, m4
+  pavgw                m3, m5
+  mova                 m4, m2
+  mova                 m5, m3
+  pmullw               m4, filter_y_a
+  pmullw               m3, filter_y_b
+  paddw                m4, filter_rnd
+  paddw                m4, m3
+  pmullw               m0, filter_y_a
+  pmullw               m2, filter_y_b
+  paddw                m0, filter_rnd
+  psrlw                m4, 4
+  paddw                m0, m2
+  mova                 m2, [dstq]
+  psrlw                m0, 4
+  mova                 m3, [dstq+dst_strideq*2]
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  add                secq, sec_str
+  pavgw                m4, [secq]
+%endif
+  SUM_SSE              m0, m2, m4, m3, m6, m7
+  mova                 m0, m5
+
+  lea                srcq, [srcq + src_strideq*4]
+  lea                dstq, [dstq + dst_strideq*4]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%endif
+  dec                   block_height
+  jg .x_half_y_other_loop
+%undef filter_y_a
+%undef filter_y_b
+%undef filter_rnd
+  STORE_AND_RET
+
+.x_nonhalf:
+  test          y_offsetd, y_offsetd
+  jnz .x_nonhalf_y_nonzero
+
+  ; x_offset == bilin interpolation && y_offset == 0
+%if AOM_ARCH_X86_64
+  lea        bilin_filter, [GLOBAL(bilin_filter_m)]
+%endif
+  shl           x_offsetd, filter_idx_shift
+%if AOM_ARCH_X86_64 && mmsize == 16
+  mova                 m8, [bilin_filter+x_offsetq]
+  mova                 m9, [bilin_filter+x_offsetq+16]
+  mova                m10, [GLOBAL(pw_8)]
+%define filter_x_a m8
+%define filter_x_b m9
+%define filter_rnd m10
+%else    ; x86-32
+%if AOM_ARCH_X86=1 && CONFIG_PIC=1
+; y_offset == 0. We can reuse y_offset reg.
+%define tempq y_offsetq
+  add x_offsetq, g_bilin_filterm
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+  mov tempq, g_pw_8m
+%define filter_rnd [tempq]
+%else
+  add           x_offsetq, bilin_filter
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+%define filter_rnd [GLOBAL(pw_8)]
+%endif
+%endif
+
+.x_other_y_zero_loop:
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq+16]
+  movu                 m2, [srcq+2]
+  movu                 m3, [srcq+18]
+  mova                 m4, [dstq]
+  mova                 m5, [dstq+16]
+  pmullw               m1, filter_x_a
+  pmullw               m3, filter_x_b
+  paddw                m1, filter_rnd
+  pmullw               m0, filter_x_a
+  pmullw               m2, filter_x_b
+  paddw                m0, filter_rnd
+  paddw                m1, m3
+  paddw                m0, m2
+  psrlw                m1, 4
+  psrlw                m0, 4
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  pavgw                m1, [secq+16]
+%endif
+  SUM_SSE              m0, m4, m1, m5, m6, m7
+
+  lea                srcq, [srcq+src_strideq*2]
+  lea                dstq, [dstq+dst_strideq*2]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%else ; %1 < 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq+src_strideq*2]
+  movu                 m2, [srcq+2]
+  movu                 m3, [srcq+src_strideq*2+2]
+  mova                 m4, [dstq]
+  mova                 m5, [dstq+dst_strideq*2]
+  pmullw               m1, filter_x_a
+  pmullw               m3, filter_x_b
+  paddw                m1, filter_rnd
+  pmullw               m0, filter_x_a
+  pmullw               m2, filter_x_b
+  paddw                m0, filter_rnd
+  paddw                m1, m3
+  paddw                m0, m2
+  psrlw                m1, 4
+  psrlw                m0, 4
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  add                secq, sec_str
+  pavgw                m1, [secq]
+%endif
+  SUM_SSE              m0, m4, m1, m5, m6, m7
+
+  lea                srcq, [srcq+src_strideq*4]
+  lea                dstq, [dstq+dst_strideq*4]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%endif
+  dec                   block_height
+  jg .x_other_y_zero_loop
+%undef filter_x_a
+%undef filter_x_b
+%undef filter_rnd
+  STORE_AND_RET
+
+.x_nonhalf_y_nonzero:
+  cmp           y_offsetd, 8
+  jne .x_nonhalf_y_nonhalf
+
+  ; x_offset == bilin interpolation && y_offset == 0.5
+%if AOM_ARCH_X86_64
+  lea        bilin_filter, [GLOBAL(bilin_filter_m)]
+%endif
+  shl           x_offsetd, filter_idx_shift
+%if AOM_ARCH_X86_64 && mmsize == 16
+  mova                 m8, [bilin_filter+x_offsetq]
+  mova                 m9, [bilin_filter+x_offsetq+16]
+  mova                m10, [GLOBAL(pw_8)]
+%define filter_x_a m8
+%define filter_x_b m9
+%define filter_rnd m10
+%else    ; x86-32
+%if AOM_ARCH_X86=1 && CONFIG_PIC=1
+; y_offset == 0.5. We can reuse y_offset reg.
+%define tempq y_offsetq
+  add x_offsetq, g_bilin_filterm
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+  mov tempq, g_pw_8m
+%define filter_rnd [tempq]
+%else
+  add           x_offsetq, bilin_filter
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+%define filter_rnd [GLOBAL(pw_8)]
+%endif
+%endif
+
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq+16]
+  movu                 m2, [srcq+2]
+  movu                 m3, [srcq+18]
+  pmullw               m0, filter_x_a
+  pmullw               m2, filter_x_b
+  paddw                m0, filter_rnd
+  pmullw               m1, filter_x_a
+  pmullw               m3, filter_x_b
+  paddw                m1, filter_rnd
+  paddw                m0, m2
+  paddw                m1, m3
+  psrlw                m0, 4
+  psrlw                m1, 4
+  lea                srcq, [srcq+src_strideq*2]
+.x_other_y_half_loop:
+  movu                 m2, [srcq]
+  movu                 m3, [srcq+16]
+  movu                 m4, [srcq+2]
+  movu                 m5, [srcq+18]
+  pmullw               m2, filter_x_a
+  pmullw               m4, filter_x_b
+  paddw                m2, filter_rnd
+  pmullw               m3, filter_x_a
+  pmullw               m5, filter_x_b
+  paddw                m3, filter_rnd
+  paddw                m2, m4
+  paddw                m3, m5
+  mova                 m4, [dstq]
+  mova                 m5, [dstq+16]
+  psrlw                m2, 4
+  psrlw                m3, 4
+  pavgw                m0, m2
+  pavgw                m1, m3
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  pavgw                m1, [secq+16]
+%endif
+  SUM_SSE              m0, m4, m1, m5, m6, m7
+  mova                 m0, m2
+  mova                 m1, m3
+
+  lea                srcq, [srcq+src_strideq*2]
+  lea                dstq, [dstq+dst_strideq*2]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%else ; %1 < 16
+  movu                 m0, [srcq]
+  movu                 m2, [srcq+2]
+  pmullw               m0, filter_x_a
+  pmullw               m2, filter_x_b
+  paddw                m0, filter_rnd
+  paddw                m0, m2
+  psrlw                m0, 4
+  lea                srcq, [srcq+src_strideq*2]
+.x_other_y_half_loop:
+  movu                 m2, [srcq]
+  movu                 m3, [srcq+src_strideq*2]
+  movu                 m4, [srcq+2]
+  movu                 m5, [srcq+src_strideq*2+2]
+  pmullw               m2, filter_x_a
+  pmullw               m4, filter_x_b
+  paddw                m2, filter_rnd
+  pmullw               m3, filter_x_a
+  pmullw               m5, filter_x_b
+  paddw                m3, filter_rnd
+  paddw                m2, m4
+  paddw                m3, m5
+  mova                 m4, [dstq]
+  mova                 m5, [dstq+dst_strideq*2]
+  psrlw                m2, 4
+  psrlw                m3, 4
+  pavgw                m0, m2
+  pavgw                m2, m3
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  add                secq, sec_str
+  pavgw                m2, [secq]
+%endif
+  SUM_SSE              m0, m4, m2, m5, m6, m7
+  mova                 m0, m3
+
+  lea                srcq, [srcq+src_strideq*4]
+  lea                dstq, [dstq+dst_strideq*4]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%endif
+  dec                   block_height
+  jg .x_other_y_half_loop
+%undef filter_x_a
+%undef filter_x_b
+%undef filter_rnd
+  STORE_AND_RET
+
+.x_nonhalf_y_nonhalf:
+; loading filter - this is same as in 8-bit depth
+%if AOM_ARCH_X86_64
+  lea        bilin_filter, [GLOBAL(bilin_filter_m)]
+%endif
+  shl           x_offsetd, filter_idx_shift ; filter_idx_shift = 5
+  shl           y_offsetd, filter_idx_shift
+%if AOM_ARCH_X86_64 && mmsize == 16
+  mova                 m8, [bilin_filter+x_offsetq]
+  mova                 m9, [bilin_filter+x_offsetq+16]
+  mova                m10, [bilin_filter+y_offsetq]
+  mova                m11, [bilin_filter+y_offsetq+16]
+  mova                m12, [GLOBAL(pw_8)]
+%define filter_x_a m8
+%define filter_x_b m9
+%define filter_y_a m10
+%define filter_y_b m11
+%define filter_rnd m12
+%else   ; x86-32
+%if AOM_ARCH_X86=1 && CONFIG_PIC=1
+; In this case, there is NO unused register. Used src_stride register. Later,
+; src_stride has to be loaded from stack when it is needed.
+%define tempq src_strideq
+  mov tempq, g_bilin_filterm
+  add           x_offsetq, tempq
+  add           y_offsetq, tempq
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+
+  mov tempq, g_pw_8m
+%define filter_rnd [tempq]
+%else
+  add           x_offsetq, bilin_filter
+  add           y_offsetq, bilin_filter
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+%define filter_rnd [GLOBAL(pw_8)]
+%endif
+%endif
+; end of load filter
+
+  ; x_offset == bilin interpolation && y_offset == bilin interpolation
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m2, [srcq+2]
+  movu                 m1, [srcq+16]
+  movu                 m3, [srcq+18]
+  pmullw               m0, filter_x_a
+  pmullw               m2, filter_x_b
+  paddw                m0, filter_rnd
+  pmullw               m1, filter_x_a
+  pmullw               m3, filter_x_b
+  paddw                m1, filter_rnd
+  paddw                m0, m2
+  paddw                m1, m3
+  psrlw                m0, 4
+  psrlw                m1, 4
+
+  INC_SRC_BY_SRC_STRIDE
+
+.x_other_y_other_loop:
+  movu                 m2, [srcq]
+  movu                 m4, [srcq+2]
+  movu                 m3, [srcq+16]
+  movu                 m5, [srcq+18]
+  pmullw               m2, filter_x_a
+  pmullw               m4, filter_x_b
+  paddw                m2, filter_rnd
+  pmullw               m3, filter_x_a
+  pmullw               m5, filter_x_b
+  paddw                m3, filter_rnd
+  paddw                m2, m4
+  paddw                m3, m5
+  psrlw                m2, 4
+  psrlw                m3, 4
+  mova                 m4, m2
+  mova                 m5, m3
+  pmullw               m0, filter_y_a
+  pmullw               m2, filter_y_b
+  paddw                m0, filter_rnd
+  pmullw               m1, filter_y_a
+  pmullw               m3, filter_y_b
+  paddw                m0, m2
+  paddw                m1, filter_rnd
+  mova                 m2, [dstq]
+  paddw                m1, m3
+  psrlw                m0, 4
+  psrlw                m1, 4
+  mova                 m3, [dstq+16]
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  pavgw                m1, [secq+16]
+%endif
+  SUM_SSE              m0, m2, m1, m3, m6, m7
+  mova                 m0, m4
+  mova                 m1, m5
+
+  INC_SRC_BY_SRC_STRIDE
+  lea                dstq, [dstq + dst_strideq * 2]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%else ; %1 < 16
+  movu                 m0, [srcq]
+  movu                 m2, [srcq+2]
+  pmullw               m0, filter_x_a
+  pmullw               m2, filter_x_b
+  paddw                m0, filter_rnd
+  paddw                m0, m2
+  psrlw                m0, 4
+
+  INC_SRC_BY_SRC_STRIDE
+
+.x_other_y_other_loop:
+  movu                 m2, [srcq]
+  movu                 m4, [srcq+2]
+  INC_SRC_BY_SRC_STRIDE
+  movu                 m3, [srcq]
+  movu                 m5, [srcq+2]
+  pmullw               m2, filter_x_a
+  pmullw               m4, filter_x_b
+  paddw                m2, filter_rnd
+  pmullw               m3, filter_x_a
+  pmullw               m5, filter_x_b
+  paddw                m3, filter_rnd
+  paddw                m2, m4
+  paddw                m3, m5
+  psrlw                m2, 4
+  psrlw                m3, 4
+  mova                 m4, m2
+  mova                 m5, m3
+  pmullw               m0, filter_y_a
+  pmullw               m2, filter_y_b
+  paddw                m0, filter_rnd
+  pmullw               m4, filter_y_a
+  pmullw               m3, filter_y_b
+  paddw                m0, m2
+  paddw                m4, filter_rnd
+  mova                 m2, [dstq]
+  paddw                m4, m3
+  psrlw                m0, 4
+  psrlw                m4, 4
+  mova                 m3, [dstq+dst_strideq*2]
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  add                secq, sec_str
+  pavgw                m4, [secq]
+%endif
+  SUM_SSE              m0, m2, m4, m3, m6, m7
+  mova                 m0, m5
+
+  INC_SRC_BY_SRC_STRIDE
+  lea                dstq, [dstq + dst_strideq * 4]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%endif
+  dec                   block_height
+  jg .x_other_y_other_loop
+%undef filter_x_a
+%undef filter_x_b
+%undef filter_y_a
+%undef filter_y_b
+%undef filter_rnd
+  STORE_AND_RET
+%endmacro
+
+INIT_XMM sse2
+SUBPEL_VARIANCE  8
+SUBPEL_VARIANCE 16
+
+INIT_XMM sse2
+SUBPEL_VARIANCE  8, 1
+SUBPEL_VARIANCE 16, 1
diff --git a/third_party/aom/aom_dsp/x86/highbd_subtract_sse2.c b/third_party/aom/aom_dsp/x86/highbd_subtract_sse2.c
new file mode 100644
index 0000000000..3c3253bdf9
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/highbd_subtract_sse2.c
@@ -0,0 +1,266 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <emmintrin.h>
+#include <stddef.h>
+
+#include "config/aom_config.h"
+#include "config/aom_dsp_rtcd.h"
+
+typedef void (*SubtractWxHFuncType)(int16_t *diff, ptrdiff_t diff_stride,
+                                    const uint16_t *src, ptrdiff_t src_stride,
+                                    const uint16_t *pred,
+                                    ptrdiff_t pred_stride);
+
+static void subtract_4x4(int16_t *diff, ptrdiff_t diff_stride,
+                         const uint16_t *src, ptrdiff_t src_stride,
+                         const uint16_t *pred, ptrdiff_t pred_stride) {
+  __m128i u0, u1, u2, u3;
+  __m128i v0, v1, v2, v3;
+  __m128i x0, x1, x2, x3;
+  int64_t *store_diff = (int64_t *)(diff + 0 * diff_stride);
+
+  u0 = _mm_loadl_epi64((__m128i const *)(src + 0 * src_stride));
+  u1 = _mm_loadl_epi64((__m128i const *)(src + 1 * src_stride));
+  u2 = _mm_loadl_epi64((__m128i const *)(src + 2 * src_stride));
+  u3 = _mm_loadl_epi64((__m128i const *)(src + 3 * src_stride));
+
+  v0 = _mm_loadl_epi64((__m128i const *)(pred + 0 * pred_stride));
+  v1 = _mm_loadl_epi64((__m128i const *)(pred + 1 * pred_stride));
+  v2 = _mm_loadl_epi64((__m128i const *)(pred + 2 * pred_stride));
+  v3 = _mm_loadl_epi64((__m128i const *)(pred + 3 * pred_stride));
+
+  x0 = _mm_sub_epi16(u0, v0);
+  x1 = _mm_sub_epi16(u1, v1);
+  x2 = _mm_sub_epi16(u2, v2);
+  x3 = _mm_sub_epi16(u3, v3);
+
+  _mm_storel_epi64((__m128i *)store_diff, x0);
+  store_diff = (int64_t *)(diff + 1 * diff_stride);
+  _mm_storel_epi64((__m128i *)store_diff, x1);
+  store_diff = (int64_t *)(diff + 2 * diff_stride);
+  _mm_storel_epi64((__m128i *)store_diff, x2);
+  store_diff = (int64_t *)(diff + 3 * diff_stride);
+  _mm_storel_epi64((__m128i *)store_diff, x3);
+}
+
+static void subtract_4x8(int16_t *diff, ptrdiff_t diff_stride,
+                         const uint16_t *src, ptrdiff_t src_stride,
+                         const uint16_t *pred, ptrdiff_t pred_stride) {
+  __m128i u0, u1, u2, u3, u4, u5, u6, u7;
+  __m128i v0, v1, v2, v3, v4, v5, v6, v7;
+  __m128i x0, x1, x2, x3, x4, x5, x6, x7;
+  int64_t *store_diff = (int64_t *)(diff + 0 * diff_stride);
+
+  u0 = _mm_loadl_epi64((__m128i const *)(src + 0 * src_stride));
+  u1 = _mm_loadl_epi64((__m128i const *)(src + 1 * src_stride));
+  u2 = _mm_loadl_epi64((__m128i const *)(src + 2 * src_stride));
+  u3 = _mm_loadl_epi64((__m128i const *)(src + 3 * src_stride));
+  u4 = _mm_loadl_epi64((__m128i const *)(src + 4 * src_stride));
+  u5 = _mm_loadl_epi64((__m128i const *)(src + 5 * src_stride));
+  u6 = _mm_loadl_epi64((__m128i const *)(src + 6 * src_stride));
+  u7 = _mm_loadl_epi64((__m128i const *)(src + 7 * src_stride));
+
+  v0 = _mm_loadl_epi64((__m128i const *)(pred + 0 * pred_stride));
+  v1 = _mm_loadl_epi64((__m128i const *)(pred + 1 * pred_stride));
+  v2 = _mm_loadl_epi64((__m128i const *)(pred + 2 * pred_stride));
+  v3 = _mm_loadl_epi64((__m128i const *)(pred + 3 * pred_stride));
+  v4 = _mm_loadl_epi64((__m128i const *)(pred + 4 * pred_stride));
+  v5 = _mm_loadl_epi64((__m128i const *)(pred + 5 * pred_stride));
+  v6 = _mm_loadl_epi64((__m128i const *)(pred + 6 * pred_stride));
+  v7 = _mm_loadl_epi64((__m128i const *)(pred + 7 * pred_stride));
+
+  x0 = _mm_sub_epi16(u0, v0);
+  x1 = _mm_sub_epi16(u1, v1);
+  x2 = _mm_sub_epi16(u2, v2);
+  x3 = _mm_sub_epi16(u3, v3);
+  x4 = _mm_sub_epi16(u4, v4);
+  x5 = _mm_sub_epi16(u5, v5);
+  x6 = _mm_sub_epi16(u6, v6);
+  x7 = _mm_sub_epi16(u7, v7);
+
+  _mm_storel_epi64((__m128i *)store_diff, x0);
+  store_diff = (int64_t *)(diff + 1 * diff_stride);
+  _mm_storel_epi64((__m128i *)store_diff, x1);
+  store_diff = (int64_t *)(diff + 2 * diff_stride);
+  _mm_storel_epi64((__m128i *)store_diff, x2);
+  store_diff = (int64_t *)(diff + 3 * diff_stride);
+  _mm_storel_epi64((__m128i *)store_diff, x3);
+  store_diff = (int64_t *)(diff + 4 * diff_stride);
+  _mm_storel_epi64((__m128i *)store_diff, x4);
+  store_diff = (int64_t *)(diff + 5 * diff_stride);
+  _mm_storel_epi64((__m128i *)store_diff, x5);
+  store_diff = (int64_t *)(diff + 6 * diff_stride);
+  _mm_storel_epi64((__m128i *)store_diff, x6);
+  store_diff = (int64_t *)(diff + 7 * diff_stride);
+  _mm_storel_epi64((__m128i *)store_diff, x7);
+}
+
+static void subtract_8x4(int16_t *diff, ptrdiff_t diff_stride,
+                         const uint16_t *src, ptrdiff_t src_stride,
+                         const uint16_t *pred, ptrdiff_t pred_stride) {
+  __m128i u0, u1, u2, u3;
+  __m128i v0, v1, v2, v3;
+  __m128i x0, x1, x2, x3;
+
+  u0 = _mm_loadu_si128((__m128i const *)(src + 0 * src_stride));
+  u1 = _mm_loadu_si128((__m128i const *)(src + 1 * src_stride));
+  u2 = _mm_loadu_si128((__m128i const *)(src + 2 * src_stride));
+  u3 = _mm_loadu_si128((__m128i const *)(src + 3 * src_stride));
+
+  v0 = _mm_loadu_si128((__m128i const *)(pred + 0 * pred_stride));
+  v1 = _mm_loadu_si128((__m128i const *)(pred + 1 * pred_stride));
+  v2 = _mm_loadu_si128((__m128i const *)(pred + 2 * pred_stride));
+  v3 = _mm_loadu_si128((__m128i const *)(pred + 3 * pred_stride));
+
+  x0 = _mm_sub_epi16(u0, v0);
+  x1 = _mm_sub_epi16(u1, v1);
+  x2 = _mm_sub_epi16(u2, v2);
+  x3 = _mm_sub_epi16(u3, v3);
+
+  _mm_storeu_si128((__m128i *)(diff + 0 * diff_stride), x0);
+  _mm_storeu_si128((__m128i *)(diff + 1 * diff_stride), x1);
+  _mm_storeu_si128((__m128i *)(diff + 2 * diff_stride), x2);
+  _mm_storeu_si128((__m128i *)(diff + 3 * diff_stride), x3);
+}
+
+static void subtract_8x8(int16_t *diff, ptrdiff_t diff_stride,
+                         const uint16_t *src, ptrdiff_t src_stride,
+                         const uint16_t *pred, ptrdiff_t pred_stride) {
+  __m128i u0, u1, u2, u3, u4, u5, u6, u7;
+  __m128i v0, v1, v2, v3, v4, v5, v6, v7;
+  __m128i x0, x1, x2, x3, x4, x5, x6, x7;
+
+  u0 = _mm_loadu_si128((__m128i const *)(src + 0 * src_stride));
+  u1 = _mm_loadu_si128((__m128i const *)(src + 1 * src_stride));
+  u2 = _mm_loadu_si128((__m128i const *)(src + 2 * src_stride));
+  u3 = _mm_loadu_si128((__m128i const *)(src + 3 * src_stride));
+  u4 = _mm_loadu_si128((__m128i const *)(src + 4 * src_stride));
+  u5 = _mm_loadu_si128((__m128i const *)(src + 5 * src_stride));
+  u6 = _mm_loadu_si128((__m128i const *)(src + 6 * src_stride));
+  u7 = _mm_loadu_si128((__m128i const *)(src + 7 * src_stride));
+
+  v0 = _mm_loadu_si128((__m128i const *)(pred + 0 * pred_stride));
+  v1 = _mm_loadu_si128((__m128i const *)(pred + 1 * pred_stride));
+  v2 = _mm_loadu_si128((__m128i const *)(pred + 2 * pred_stride));
+  v3 = _mm_loadu_si128((__m128i const *)(pred + 3 * pred_stride));
+  v4 = _mm_loadu_si128((__m128i const *)(pred + 4 * pred_stride));
+  v5 = _mm_loadu_si128((__m128i const *)(pred + 5 * pred_stride));
+  v6 = _mm_loadu_si128((__m128i const *)(pred + 6 * pred_stride));
+  v7 = _mm_loadu_si128((__m128i const *)(pred + 7 * pred_stride));
+
+  x0 = _mm_sub_epi16(u0, v0);
+  x1 = _mm_sub_epi16(u1, v1);
+  x2 = _mm_sub_epi16(u2, v2);
+  x3 = _mm_sub_epi16(u3, v3);
+  x4 = _mm_sub_epi16(u4, v4);
+  x5 = _mm_sub_epi16(u5, v5);
+  x6 = _mm_sub_epi16(u6, v6);
+  x7 = _mm_sub_epi16(u7, v7);
+
+  _mm_storeu_si128((__m128i *)(diff + 0 * diff_stride), x0);
+  _mm_storeu_si128((__m128i *)(diff + 1 * diff_stride), x1);
+  _mm_storeu_si128((__m128i *)(diff + 2 * diff_stride), x2);
+  _mm_storeu_si128((__m128i *)(diff + 3 * diff_stride), x3);
+  _mm_storeu_si128((__m128i *)(diff + 4 * diff_stride), x4);
+  _mm_storeu_si128((__m128i *)(diff + 5 * diff_stride), x5);
+  _mm_storeu_si128((__m128i *)(diff + 6 * diff_stride), x6);
+  _mm_storeu_si128((__m128i *)(diff + 7 * diff_stride), x7);
+}
+
+#define STACK_V(h, fun)                                                        \
+  do {                                                                         \
+    fun(diff, diff_stride, src, src_stride, pred, pred_stride);                \
+    fun(diff + diff_stride * h, diff_stride, src + src_stride * h, src_stride, \
+        pred + pred_stride * h, pred_stride);                                  \
+  } while (0)
+
+#define STACK_H(w, fun)                                                     \
+  do {                                                                      \
+    fun(diff, diff_stride, src, src_stride, pred, pred_stride);             \
+    fun(diff + w, diff_stride, src + w, src_stride, pred + w, pred_stride); \
+  } while (0)
+
+#define SUBTRACT_FUN(size)                                               \
+  static void subtract_##size(int16_t *diff, ptrdiff_t diff_stride,      \
+                              const uint16_t *src, ptrdiff_t src_stride, \
+                              const uint16_t *pred, ptrdiff_t pred_stride)
+
+SUBTRACT_FUN(8x16) { STACK_V(8, subtract_8x8); }
+SUBTRACT_FUN(16x8) { STACK_H(8, subtract_8x8); }
+SUBTRACT_FUN(16x16) { STACK_V(8, subtract_16x8); }
+SUBTRACT_FUN(16x32) { STACK_V(16, subtract_16x16); }
+SUBTRACT_FUN(32x16) { STACK_H(16, subtract_16x16); }
+SUBTRACT_FUN(32x32) { STACK_V(16, subtract_32x16); }
+SUBTRACT_FUN(32x64) { STACK_V(32, subtract_32x32); }
+SUBTRACT_FUN(64x32) { STACK_H(32, subtract_32x32); }
+SUBTRACT_FUN(64x64) { STACK_V(32, subtract_64x32); }
+SUBTRACT_FUN(64x128) { STACK_V(64, subtract_64x64); }
+SUBTRACT_FUN(128x64) { STACK_H(64, subtract_64x64); }
+SUBTRACT_FUN(128x128) { STACK_V(64, subtract_128x64); }
+SUBTRACT_FUN(4x16) { STACK_V(8, subtract_4x8); }
+SUBTRACT_FUN(16x4) { STACK_H(8, subtract_8x4); }
+SUBTRACT_FUN(8x32) { STACK_V(16, subtract_8x16); }
+SUBTRACT_FUN(32x8) { STACK_H(16, subtract_16x8); }
+SUBTRACT_FUN(16x64) { STACK_V(32, subtract_16x32); }
+SUBTRACT_FUN(64x16) { STACK_H(32, subtract_32x16); }
+
+static SubtractWxHFuncType getSubtractFunc(int rows, int cols) {
+  if (rows == 4) {
+    if (cols == 4) return subtract_4x4;
+    if (cols == 8) return subtract_8x4;
+    if (cols == 16) return subtract_16x4;
+  }
+  if (rows == 8) {
+    if (cols == 4) return subtract_4x8;
+    if (cols == 8) return subtract_8x8;
+    if (cols == 16) return subtract_16x8;
+    if (cols == 32) return subtract_32x8;
+  }
+  if (rows == 16) {
+    if (cols == 4) return subtract_4x16;
+    if (cols == 8) return subtract_8x16;
+    if (cols == 16) return subtract_16x16;
+    if (cols == 32) return subtract_32x16;
+    if (cols == 64) return subtract_64x16;
+  }
+  if (rows == 32) {
+    if (cols == 8) return subtract_8x32;
+    if (cols == 16) return subtract_16x32;
+    if (cols == 32) return subtract_32x32;
+    if (cols == 64) return subtract_64x32;
+  }
+  if (rows == 64) {
+    if (cols == 16) return subtract_16x64;
+    if (cols == 32) return subtract_32x64;
+    if (cols == 64) return subtract_64x64;
+    if (cols == 128) return subtract_128x64;
+  }
+  if (rows == 128) {
+    if (cols == 64) return subtract_64x128;
+    if (cols == 128) return subtract_128x128;
+  }
+  assert(0);
+  return NULL;
+}
+
+void aom_highbd_subtract_block_sse2(int rows, int cols, int16_t *diff,
+                                    ptrdiff_t diff_stride, const uint8_t *src8,
+                                    ptrdiff_t src_stride, const uint8_t *pred8,
+                                    ptrdiff_t pred_stride) {
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *pred = CONVERT_TO_SHORTPTR(pred8);
+  SubtractWxHFuncType func;
+
+  func = getSubtractFunc(rows, cols);
+  func(diff, diff_stride, src, src_stride, pred, pred_stride);
+}
diff --git a/third_party/aom/aom_dsp/x86/highbd_variance_avx2.c b/third_party/aom/aom_dsp/x86/highbd_variance_avx2.c
new file mode 100644
index 0000000000..b4ff91d856
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/highbd_variance_avx2.c
@@ -0,0 +1,904 @@
+/*
+ * Copyright (c) 2020, 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>  // AVX2
+
+#include "config/aom_dsp_rtcd.h"
+#include "aom_dsp/aom_filter.h"
+#include "aom_dsp/x86/synonyms.h"
+
+typedef void (*high_variance_fn_t)(const uint16_t *src, int src_stride,
+                                   const uint16_t *ref, int ref_stride,
+                                   uint32_t *sse, int *sum);
+
+static uint32_t aom_highbd_var_filter_block2d_bil_avx2(
+    const uint8_t *src_ptr8, unsigned int src_pixels_per_line, int pixel_step,
+    unsigned int output_height, unsigned int output_width,
+    const uint32_t xoffset, const uint32_t yoffset, const uint8_t *dst_ptr8,
+    int dst_stride, uint32_t *sse) {
+  const __m256i filter1 =
+      _mm256_set1_epi32((int)(bilinear_filters_2t[xoffset][1] << 16) |
+                        bilinear_filters_2t[xoffset][0]);
+  const __m256i filter2 =
+      _mm256_set1_epi32((int)(bilinear_filters_2t[yoffset][1] << 16) |
+                        bilinear_filters_2t[yoffset][0]);
+  const __m256i one = _mm256_set1_epi16(1);
+  const int bitshift = 0x40;
+  (void)pixel_step;
+  unsigned int i, j, prev = 0, curr = 2;
+  uint16_t *src_ptr = CONVERT_TO_SHORTPTR(src_ptr8);
+  uint16_t *dst_ptr = CONVERT_TO_SHORTPTR(dst_ptr8);
+  uint16_t *src_ptr_ref = src_ptr;
+  uint16_t *dst_ptr_ref = dst_ptr;
+  int64_t sum_long = 0;
+  uint64_t sse_long = 0;
+  unsigned int rshift = 0, inc = 1;
+  __m256i rbias = _mm256_set1_epi32(bitshift);
+  __m256i opointer[8];
+  unsigned int range;
+  if (xoffset == 0) {
+    if (yoffset == 0) {  // xoffset==0 && yoffset==0
+      range = output_width / 16;
+      if (output_height == 8) inc = 2;
+      if (output_height == 4) inc = 4;
+      for (j = 0; j < range * output_height * inc / 16; j++) {
+        if (j % (output_height * inc / 16) == 0) {
+          src_ptr = src_ptr_ref;
+          src_ptr_ref += 16;
+          dst_ptr = dst_ptr_ref;
+          dst_ptr_ref += 16;
+        }
+        __m256i sum1 = _mm256_setzero_si256();
+        __m256i sse1 = _mm256_setzero_si256();
+        for (i = 0; i < 16 / inc; ++i) {
+          __m256i V_S_SRC = _mm256_loadu_si256((const __m256i *)src_ptr);
+          src_ptr += src_pixels_per_line;
+          __m256i V_D_DST = _mm256_loadu_si256((const __m256i *)dst_ptr);
+          dst_ptr += dst_stride;
+
+          __m256i V_R_SUB = _mm256_sub_epi16(V_S_SRC, V_D_DST);
+          __m256i V_R_MAD = _mm256_madd_epi16(V_R_SUB, V_R_SUB);
+
+          sum1 = _mm256_add_epi16(sum1, V_R_SUB);
+          sse1 = _mm256_add_epi32(sse1, V_R_MAD);
+        }
+
+        __m256i v_sum0 = _mm256_madd_epi16(sum1, one);
+        __m256i v_d_l = _mm256_unpacklo_epi32(v_sum0, sse1);
+        __m256i v_d_h = _mm256_unpackhi_epi32(v_sum0, sse1);
+        __m256i v_d_lh = _mm256_add_epi32(v_d_l, v_d_h);
+        const __m128i v_d0_d = _mm256_castsi256_si128(v_d_lh);
+        const __m128i v_d1_d = _mm256_extracti128_si256(v_d_lh, 1);
+        __m128i v_d = _mm_add_epi32(v_d0_d, v_d1_d);
+        v_d = _mm_add_epi32(v_d, _mm_srli_si128(v_d, 8));
+        sum_long += _mm_extract_epi32(v_d, 0);
+        sse_long += _mm_extract_epi32(v_d, 1);
+      }
+
+      rshift = get_msb(output_height) + get_msb(output_width);
+
+    } else if (yoffset == 4) {  // xoffset==0 && yoffset==4
+      range = output_width / 16;
+      if (output_height == 8) inc = 2;
+      if (output_height == 4) inc = 4;
+      for (j = 0; j < range * output_height * inc / 16; j++) {
+        if (j % (output_height * inc / 16) == 0) {
+          src_ptr = src_ptr_ref;
+          src_ptr_ref += 16;
+          dst_ptr = dst_ptr_ref;
+          dst_ptr_ref += 16;
+
+          opointer[0] = _mm256_loadu_si256((const __m256i *)src_ptr);
+          src_ptr += src_pixels_per_line;
+          curr = 0;
+        }
+
+        __m256i sum1 = _mm256_setzero_si256();
+        __m256i sse1 = _mm256_setzero_si256();
+
+        for (i = 0; i < 16 / inc; ++i) {
+          prev = curr;
+          curr = (curr == 0) ? 1 : 0;
+          opointer[curr] = _mm256_loadu_si256((const __m256i *)src_ptr);
+          src_ptr += src_pixels_per_line;
+
+          __m256i V_S_SRC = _mm256_avg_epu16(opointer[curr], opointer[prev]);
+
+          __m256i V_D_DST = _mm256_loadu_si256((const __m256i *)dst_ptr);
+          dst_ptr += dst_stride;
+          __m256i V_R_SUB = _mm256_sub_epi16(V_S_SRC, V_D_DST);
+          __m256i V_R_MAD = _mm256_madd_epi16(V_R_SUB, V_R_SUB);
+          sum1 = _mm256_add_epi16(sum1, V_R_SUB);
+          sse1 = _mm256_add_epi32(sse1, V_R_MAD);
+        }
+
+        __m256i v_sum0 = _mm256_madd_epi16(sum1, one);
+        __m256i v_d_l = _mm256_unpacklo_epi32(v_sum0, sse1);
+        __m256i v_d_h = _mm256_unpackhi_epi32(v_sum0, sse1);
+        __m256i v_d_lh = _mm256_add_epi32(v_d_l, v_d_h);
+        const __m128i v_d0_d = _mm256_castsi256_si128(v_d_lh);
+        const __m128i v_d1_d = _mm256_extracti128_si256(v_d_lh, 1);
+        __m128i v_d = _mm_add_epi32(v_d0_d, v_d1_d);
+        v_d = _mm_add_epi32(v_d, _mm_srli_si128(v_d, 8));
+        sum_long += _mm_extract_epi32(v_d, 0);
+        sse_long += _mm_extract_epi32(v_d, 1);
+      }
+
+      rshift = get_msb(output_height) + get_msb(output_width);
+
+    } else {  // xoffset==0 && yoffset==1,2,3,5,6,7
+      range = output_width / 16;
+      if (output_height == 8) inc = 2;
+      if (output_height == 4) inc = 4;
+      for (j = 0; j < range * output_height * inc / 16; j++) {
+        if (j % (output_height * inc / 16) == 0) {
+          src_ptr = src_ptr_ref;
+          src_ptr_ref += 16;
+          dst_ptr = dst_ptr_ref;
+          dst_ptr_ref += 16;
+
+          opointer[0] = _mm256_loadu_si256((const __m256i *)src_ptr);
+          src_ptr += src_pixels_per_line;
+          curr = 0;
+        }
+
+        __m256i sum1 = _mm256_setzero_si256();
+        __m256i sse1 = _mm256_setzero_si256();
+
+        for (i = 0; i < 16 / inc; ++i) {
+          prev = curr;
+          curr = (curr == 0) ? 1 : 0;
+          opointer[curr] = _mm256_loadu_si256((const __m256i *)src_ptr);
+          src_ptr += src_pixels_per_line;
+
+          __m256i V_S_M1 =
+              _mm256_unpacklo_epi16(opointer[prev], opointer[curr]);
+          __m256i V_S_M2 =
+              _mm256_unpackhi_epi16(opointer[prev], opointer[curr]);
+
+          __m256i V_S_MAD1 = _mm256_madd_epi16(V_S_M1, filter2);
+          __m256i V_S_MAD2 = _mm256_madd_epi16(V_S_M2, filter2);
+
+          __m256i V_S_S1 =
+              _mm256_srli_epi32(_mm256_add_epi32(V_S_MAD1, rbias), 7);
+          __m256i V_S_S2 =
+              _mm256_srli_epi32(_mm256_add_epi32(V_S_MAD2, rbias), 7);
+
+          __m256i V_S_SRC = _mm256_packus_epi32(V_S_S1, V_S_S2);
+
+          __m256i V_D_DST = _mm256_loadu_si256((const __m256i *)dst_ptr);
+          dst_ptr += dst_stride;
+
+          __m256i V_R_SUB = _mm256_sub_epi16(V_S_SRC, V_D_DST);
+          __m256i V_R_MAD = _mm256_madd_epi16(V_R_SUB, V_R_SUB);
+
+          sum1 = _mm256_add_epi16(sum1, V_R_SUB);
+          sse1 = _mm256_add_epi32(sse1, V_R_MAD);
+        }
+
+        __m256i v_sum0 = _mm256_madd_epi16(sum1, one);
+        __m256i v_d_l = _mm256_unpacklo_epi32(v_sum0, sse1);
+        __m256i v_d_h = _mm256_unpackhi_epi32(v_sum0, sse1);
+        __m256i v_d_lh = _mm256_add_epi32(v_d_l, v_d_h);
+        const __m128i v_d0_d = _mm256_castsi256_si128(v_d_lh);
+        const __m128i v_d1_d = _mm256_extracti128_si256(v_d_lh, 1);
+        __m128i v_d = _mm_add_epi32(v_d0_d, v_d1_d);
+        v_d = _mm_add_epi32(v_d, _mm_srli_si128(v_d, 8));
+        sum_long += _mm_extract_epi32(v_d, 0);
+        sse_long += _mm_extract_epi32(v_d, 1);
+      }
+
+      rshift = get_msb(output_height) + get_msb(output_width);
+    }
+  } else if (xoffset == 4) {
+    if (yoffset == 0) {  // xoffset==4 && yoffset==0
+      range = output_width / 16;
+      if (output_height == 8) inc = 2;
+      if (output_height == 4) inc = 4;
+      for (j = 0; j < range * output_height * inc / 16; j++) {
+        if (j % (output_height * inc / 16) == 0) {
+          src_ptr = src_ptr_ref;
+          src_ptr_ref += 16;
+          dst_ptr = dst_ptr_ref;
+          dst_ptr_ref += 16;
+          __m256i V_H_D1 = _mm256_loadu_si256((const __m256i *)src_ptr);
+          __m256i V_H_D2 = _mm256_loadu_si256((const __m256i *)(src_ptr + 1));
+          src_ptr += src_pixels_per_line;
+
+          opointer[0] = _mm256_avg_epu16(V_H_D1, V_H_D2);
+
+          curr = 0;
+        }
+
+        __m256i sum1 = _mm256_setzero_si256();
+        __m256i sse1 = _mm256_setzero_si256();
+
+        for (i = 0; i < 16 / inc; ++i) {
+          prev = curr;
+          curr = (curr == 0) ? 1 : 0;
+          __m256i V_V_D1 = _mm256_loadu_si256((const __m256i *)src_ptr);
+          __m256i V_V_D2 = _mm256_loadu_si256((const __m256i *)(src_ptr + 1));
+          src_ptr += src_pixels_per_line;
+
+          opointer[curr] = _mm256_avg_epu16(V_V_D1, V_V_D2);
+
+          __m256i V_S_M1 =
+              _mm256_unpacklo_epi16(opointer[prev], opointer[curr]);
+          __m256i V_S_M2 =
+              _mm256_unpackhi_epi16(opointer[prev], opointer[curr]);
+
+          __m256i V_S_MAD1 = _mm256_madd_epi16(V_S_M1, filter2);
+          __m256i V_S_MAD2 = _mm256_madd_epi16(V_S_M2, filter2);
+
+          __m256i V_S_S1 =
+              _mm256_srli_epi32(_mm256_add_epi32(V_S_MAD1, rbias), 7);
+          __m256i V_S_S2 =
+              _mm256_srli_epi32(_mm256_add_epi32(V_S_MAD2, rbias), 7);
+
+          __m256i V_S_SRC = _mm256_packus_epi32(V_S_S1, V_S_S2);
+
+          __m256i V_D_DST = _mm256_loadu_si256((const __m256i *)dst_ptr);
+          dst_ptr += dst_stride;
+
+          __m256i V_R_SUB = _mm256_sub_epi16(V_S_SRC, V_D_DST);
+          __m256i V_R_MAD = _mm256_madd_epi16(V_R_SUB, V_R_SUB);
+
+          sum1 = _mm256_add_epi16(sum1, V_R_SUB);
+          sse1 = _mm256_add_epi32(sse1, V_R_MAD);
+        }
+
+        __m256i v_sum0 = _mm256_madd_epi16(sum1, one);
+        __m256i v_d_l = _mm256_unpacklo_epi32(v_sum0, sse1);
+        __m256i v_d_h = _mm256_unpackhi_epi32(v_sum0, sse1);
+        __m256i v_d_lh = _mm256_add_epi32(v_d_l, v_d_h);
+        const __m128i v_d0_d = _mm256_castsi256_si128(v_d_lh);
+        const __m128i v_d1_d = _mm256_extracti128_si256(v_d_lh, 1);
+        __m128i v_d = _mm_add_epi32(v_d0_d, v_d1_d);
+        v_d = _mm_add_epi32(v_d, _mm_srli_si128(v_d, 8));
+        sum_long += _mm_extract_epi32(v_d, 0);
+        sse_long += _mm_extract_epi32(v_d, 1);
+      }
+
+      rshift = get_msb(output_height) + get_msb(output_width);
+
+    } else if (yoffset == 4) {  // xoffset==4 && yoffset==4
+      range = output_width / 16;
+      if (output_height == 8) inc = 2;
+      if (output_height == 4) inc = 4;
+      for (j = 0; j < range * output_height * inc / 16; j++) {
+        if (j % (output_height * inc / 16) == 0) {
+          src_ptr = src_ptr_ref;
+          src_ptr_ref += 16;
+          dst_ptr = dst_ptr_ref;
+          dst_ptr_ref += 16;
+
+          __m256i V_H_D1 = _mm256_loadu_si256((const __m256i *)src_ptr);
+          __m256i V_H_D2 = _mm256_loadu_si256((const __m256i *)(src_ptr + 1));
+          src_ptr += src_pixels_per_line;
+          opointer[0] = _mm256_avg_epu16(V_H_D1, V_H_D2);
+          curr = 0;
+        }
+
+        __m256i sum1 = _mm256_setzero_si256();
+        __m256i sse1 = _mm256_setzero_si256();
+
+        for (i = 0; i < 16 / inc; ++i) {
+          prev = curr;
+          curr = (curr == 0) ? 1 : 0;
+          __m256i V_V_D1 = _mm256_loadu_si256((const __m256i *)src_ptr);
+          __m256i V_V_D2 = _mm256_loadu_si256((const __m256i *)(src_ptr + 1));
+          src_ptr += src_pixels_per_line;
+          opointer[curr] = _mm256_avg_epu16(V_V_D1, V_V_D2);
+          __m256i V_S_SRC = _mm256_avg_epu16(opointer[curr], opointer[prev]);
+
+          __m256i V_D_DST = _mm256_loadu_si256((const __m256i *)dst_ptr);
+          dst_ptr += dst_stride;
+          __m256i V_R_SUB = _mm256_sub_epi16(V_S_SRC, V_D_DST);
+          __m256i V_R_MAD = _mm256_madd_epi16(V_R_SUB, V_R_SUB);
+          sum1 = _mm256_add_epi16(sum1, V_R_SUB);
+          sse1 = _mm256_add_epi32(sse1, V_R_MAD);
+        }
+
+        __m256i v_sum0 = _mm256_madd_epi16(sum1, one);
+        __m256i v_d_l = _mm256_unpacklo_epi32(v_sum0, sse1);
+        __m256i v_d_h = _mm256_unpackhi_epi32(v_sum0, sse1);
+        __m256i v_d_lh = _mm256_add_epi32(v_d_l, v_d_h);
+        const __m128i v_d0_d = _mm256_castsi256_si128(v_d_lh);
+        const __m128i v_d1_d = _mm256_extracti128_si256(v_d_lh, 1);
+        __m128i v_d = _mm_add_epi32(v_d0_d, v_d1_d);
+        v_d = _mm_add_epi32(v_d, _mm_srli_si128(v_d, 8));
+        sum_long += _mm_extract_epi32(v_d, 0);
+        sse_long += _mm_extract_epi32(v_d, 1);
+      }
+
+      rshift = get_msb(output_height) + get_msb(output_width);
+
+    } else {  // xoffset==4 && yoffset==1,2,3,5,6,7
+      range = output_width / 16;
+      if (output_height == 8) inc = 2;
+      if (output_height == 4) inc = 4;
+      for (j = 0; j < range * output_height * inc / 16; j++) {
+        if (j % (output_height * inc / 16) == 0) {
+          src_ptr = src_ptr_ref;
+          src_ptr_ref += 16;
+          dst_ptr = dst_ptr_ref;
+          dst_ptr_ref += 16;
+
+          __m256i V_H_D1 = _mm256_loadu_si256((const __m256i *)src_ptr);
+          __m256i V_H_D2 = _mm256_loadu_si256((const __m256i *)(src_ptr + 1));
+          src_ptr += src_pixels_per_line;
+          opointer[0] = _mm256_avg_epu16(V_H_D1, V_H_D2);
+          curr = 0;
+        }
+
+        __m256i sum1 = _mm256_setzero_si256();
+        __m256i sse1 = _mm256_setzero_si256();
+
+        for (i = 0; i < 16 / inc; ++i) {
+          prev = curr;
+          curr = (curr == 0) ? 1 : 0;
+          __m256i V_V_D1 = _mm256_loadu_si256((const __m256i *)src_ptr);
+          __m256i V_V_D2 = _mm256_loadu_si256((const __m256i *)(src_ptr + 1));
+          src_ptr += src_pixels_per_line;
+          opointer[curr] = _mm256_avg_epu16(V_V_D1, V_V_D2);
+
+          __m256i V_S_M1 =
+              _mm256_unpacklo_epi16(opointer[prev], opointer[curr]);
+          __m256i V_S_M2 =
+              _mm256_unpackhi_epi16(opointer[prev], opointer[curr]);
+
+          __m256i V_S_MAD1 = _mm256_madd_epi16(V_S_M1, filter2);
+          __m256i V_S_MAD2 = _mm256_madd_epi16(V_S_M2, filter2);
+
+          __m256i V_S_S1 =
+              _mm256_srli_epi32(_mm256_add_epi32(V_S_MAD1, rbias), 7);
+          __m256i V_S_S2 =
+              _mm256_srli_epi32(_mm256_add_epi32(V_S_MAD2, rbias), 7);
+
+          __m256i V_S_SRC = _mm256_packus_epi32(V_S_S1, V_S_S2);
+
+          __m256i V_D_DST = _mm256_loadu_si256((const __m256i *)dst_ptr);
+          dst_ptr += dst_stride;
+
+          __m256i V_R_SUB = _mm256_sub_epi16(V_S_SRC, V_D_DST);
+          __m256i V_R_MAD = _mm256_madd_epi16(V_R_SUB, V_R_SUB);
+
+          sum1 = _mm256_add_epi16(sum1, V_R_SUB);
+          sse1 = _mm256_add_epi32(sse1, V_R_MAD);
+        }
+
+        __m256i v_sum0 = _mm256_madd_epi16(sum1, one);
+        __m256i v_d_l = _mm256_unpacklo_epi32(v_sum0, sse1);
+        __m256i v_d_h = _mm256_unpackhi_epi32(v_sum0, sse1);
+        __m256i v_d_lh = _mm256_add_epi32(v_d_l, v_d_h);
+        const __m128i v_d0_d = _mm256_castsi256_si128(v_d_lh);
+        const __m128i v_d1_d = _mm256_extracti128_si256(v_d_lh, 1);
+        __m128i v_d = _mm_add_epi32(v_d0_d, v_d1_d);
+        v_d = _mm_add_epi32(v_d, _mm_srli_si128(v_d, 8));
+        sum_long += _mm_extract_epi32(v_d, 0);
+        sse_long += _mm_extract_epi32(v_d, 1);
+      }
+
+      rshift = get_msb(output_height) + get_msb(output_width);
+    }
+  } else if (yoffset == 0) {  // xoffset==1,2,3,5,6,7 && yoffset==0
+    range = output_width / 16;
+    if (output_height == 8) inc = 2;
+    if (output_height == 4) inc = 4;
+    for (j = 0; j < range * output_height * inc / 16; j++) {
+      if (j % (output_height * inc / 16) == 0) {
+        src_ptr = src_ptr_ref;
+        src_ptr_ref += 16;
+        dst_ptr = dst_ptr_ref;
+        dst_ptr_ref += 16;
+
+        curr = 0;
+      }
+
+      __m256i sum1 = _mm256_setzero_si256();
+      __m256i sse1 = _mm256_setzero_si256();
+
+      for (i = 0; i < 16 / inc; ++i) {
+        __m256i V_V_D1 = _mm256_loadu_si256((const __m256i *)src_ptr);
+        __m256i V_V_D2 = _mm256_loadu_si256((const __m256i *)(src_ptr + 1));
+        src_ptr += src_pixels_per_line;
+        __m256i V_V_M1 = _mm256_unpacklo_epi16(V_V_D1, V_V_D2);
+        __m256i V_V_M2 = _mm256_unpackhi_epi16(V_V_D1, V_V_D2);
+        __m256i V_V_MAD1 = _mm256_madd_epi16(V_V_M1, filter1);
+        __m256i V_V_MAD2 = _mm256_madd_epi16(V_V_M2, filter1);
+        __m256i V_V_S1 =
+            _mm256_srli_epi32(_mm256_add_epi32(V_V_MAD1, rbias), 7);
+        __m256i V_V_S2 =
+            _mm256_srli_epi32(_mm256_add_epi32(V_V_MAD2, rbias), 7);
+        opointer[curr] = _mm256_packus_epi32(V_V_S1, V_V_S2);
+
+        __m256i V_D_DST = _mm256_loadu_si256((const __m256i *)dst_ptr);
+        dst_ptr += dst_stride;
+        __m256i V_R_SUB = _mm256_sub_epi16(opointer[curr], V_D_DST);
+        __m256i V_R_MAD = _mm256_madd_epi16(V_R_SUB, V_R_SUB);
+
+        sum1 = _mm256_add_epi16(sum1, V_R_SUB);
+        sse1 = _mm256_add_epi32(sse1, V_R_MAD);
+      }
+
+      __m256i v_sum0 = _mm256_madd_epi16(sum1, one);
+      __m256i v_d_l = _mm256_unpacklo_epi32(v_sum0, sse1);
+      __m256i v_d_h = _mm256_unpackhi_epi32(v_sum0, sse1);
+      __m256i v_d_lh = _mm256_add_epi32(v_d_l, v_d_h);
+      const __m128i v_d0_d = _mm256_castsi256_si128(v_d_lh);
+      const __m128i v_d1_d = _mm256_extracti128_si256(v_d_lh, 1);
+      __m128i v_d = _mm_add_epi32(v_d0_d, v_d1_d);
+      v_d = _mm_add_epi32(v_d, _mm_srli_si128(v_d, 8));
+      sum_long += _mm_extract_epi32(v_d, 0);
+      sse_long += _mm_extract_epi32(v_d, 1);
+    }
+
+    rshift = get_msb(output_height) + get_msb(output_width);
+
+  } else if (yoffset == 4) {  // xoffset==1,2,3,5,6,7 && yoffset==4
+
+    range = output_width / 16;
+    if (output_height == 8) inc = 2;
+    if (output_height == 4) inc = 4;
+    for (j = 0; j < range * output_height * inc / 16; j++) {
+      if (j % (output_height * inc / 16) == 0) {
+        src_ptr = src_ptr_ref;
+        src_ptr_ref += 16;
+        dst_ptr = dst_ptr_ref;
+        dst_ptr_ref += 16;
+
+        __m256i V_H_D1 = _mm256_loadu_si256((const __m256i *)src_ptr);
+        __m256i V_H_D2 = _mm256_loadu_si256((const __m256i *)(src_ptr + 1));
+        src_ptr += src_pixels_per_line;
+
+        __m256i V_H_M1 = _mm256_unpacklo_epi16(V_H_D1, V_H_D2);
+        __m256i V_H_M2 = _mm256_unpackhi_epi16(V_H_D1, V_H_D2);
+
+        __m256i V_H_MAD1 = _mm256_madd_epi16(V_H_M1, filter1);
+        __m256i V_H_MAD2 = _mm256_madd_epi16(V_H_M2, filter1);
+
+        __m256i V_H_S1 =
+            _mm256_srli_epi32(_mm256_add_epi32(V_H_MAD1, rbias), 7);
+        __m256i V_H_S2 =
+            _mm256_srli_epi32(_mm256_add_epi32(V_H_MAD2, rbias), 7);
+
+        opointer[0] = _mm256_packus_epi32(V_H_S1, V_H_S2);
+
+        curr = 0;
+      }
+
+      __m256i sum1 = _mm256_setzero_si256();
+      __m256i sse1 = _mm256_setzero_si256();
+
+      for (i = 0; i < 16 / inc; ++i) {
+        prev = curr;
+        curr = (curr == 0) ? 1 : 0;
+        __m256i V_V_D1 = _mm256_loadu_si256((const __m256i *)src_ptr);
+        __m256i V_V_D2 = _mm256_loadu_si256((const __m256i *)(src_ptr + 1));
+        src_ptr += src_pixels_per_line;
+        __m256i V_V_M1 = _mm256_unpacklo_epi16(V_V_D1, V_V_D2);
+        __m256i V_V_M2 = _mm256_unpackhi_epi16(V_V_D1, V_V_D2);
+        __m256i V_V_MAD1 = _mm256_madd_epi16(V_V_M1, filter1);
+        __m256i V_V_MAD2 = _mm256_madd_epi16(V_V_M2, filter1);
+        __m256i V_V_S1 =
+            _mm256_srli_epi32(_mm256_add_epi32(V_V_MAD1, rbias), 7);
+        __m256i V_V_S2 =
+            _mm256_srli_epi32(_mm256_add_epi32(V_V_MAD2, rbias), 7);
+        opointer[curr] = _mm256_packus_epi32(V_V_S1, V_V_S2);
+
+        __m256i V_S_SRC = _mm256_avg_epu16(opointer[prev], opointer[curr]);
+
+        __m256i V_D_DST = _mm256_loadu_si256((const __m256i *)dst_ptr);
+        dst_ptr += dst_stride;
+
+        __m256i V_R_SUB = _mm256_sub_epi16(V_S_SRC, V_D_DST);
+        __m256i V_R_MAD = _mm256_madd_epi16(V_R_SUB, V_R_SUB);
+
+        sum1 = _mm256_add_epi16(sum1, V_R_SUB);
+        sse1 = _mm256_add_epi32(sse1, V_R_MAD);
+      }
+
+      __m256i v_sum0 = _mm256_madd_epi16(sum1, one);
+      __m256i v_d_l = _mm256_unpacklo_epi32(v_sum0, sse1);
+      __m256i v_d_h = _mm256_unpackhi_epi32(v_sum0, sse1);
+      __m256i v_d_lh = _mm256_add_epi32(v_d_l, v_d_h);
+      const __m128i v_d0_d = _mm256_castsi256_si128(v_d_lh);
+      const __m128i v_d1_d = _mm256_extracti128_si256(v_d_lh, 1);
+      __m128i v_d = _mm_add_epi32(v_d0_d, v_d1_d);
+      v_d = _mm_add_epi32(v_d, _mm_srli_si128(v_d, 8));
+      sum_long += _mm_extract_epi32(v_d, 0);
+      sse_long += _mm_extract_epi32(v_d, 1);
+    }
+
+    rshift = get_msb(output_height) + get_msb(output_width);
+
+  } else {  // xoffset==1,2,3,5,6,7 && yoffset==1,2,3,5,6,7
+    range = output_width / 16;
+    if (output_height == 8) inc = 2;
+    if (output_height == 4) inc = 4;
+    unsigned int nloop = 16 / inc;
+    for (j = 0; j < range * output_height * inc / 16; j++) {
+      if (j % (output_height * inc / 16) == 0) {
+        src_ptr = src_ptr_ref;
+        src_ptr_ref += 16;
+        dst_ptr = dst_ptr_ref;
+        dst_ptr_ref += 16;
+
+        __m256i V_H_D1 = _mm256_loadu_si256((const __m256i *)src_ptr);
+        __m256i V_H_D2 = _mm256_loadu_si256((const __m256i *)(src_ptr + 1));
+        src_ptr += src_pixels_per_line;
+
+        __m256i V_H_M1 = _mm256_unpacklo_epi16(V_H_D1, V_H_D2);
+        __m256i V_H_M2 = _mm256_unpackhi_epi16(V_H_D1, V_H_D2);
+
+        __m256i V_H_MAD1 = _mm256_madd_epi16(V_H_M1, filter1);
+        __m256i V_H_MAD2 = _mm256_madd_epi16(V_H_M2, filter1);
+
+        __m256i V_H_S1 =
+            _mm256_srli_epi32(_mm256_add_epi32(V_H_MAD1, rbias), 7);
+        __m256i V_H_S2 =
+            _mm256_srli_epi32(_mm256_add_epi32(V_H_MAD2, rbias), 7);
+
+        opointer[0] = _mm256_packus_epi32(V_H_S1, V_H_S2);
+
+        curr = 0;
+      }
+
+      __m256i sum1 = _mm256_setzero_si256();
+      __m256i sse1 = _mm256_setzero_si256();
+
+      for (i = 0; i < nloop; ++i) {
+        prev = curr;
+        curr = !curr;
+        __m256i V_V_D1 = _mm256_loadu_si256((const __m256i *)src_ptr);
+        __m256i V_V_D2 = _mm256_loadu_si256((const __m256i *)(src_ptr + 1));
+        src_ptr += src_pixels_per_line;
+        __m256i V_V_M1 = _mm256_unpacklo_epi16(V_V_D1, V_V_D2);
+        __m256i V_V_M2 = _mm256_unpackhi_epi16(V_V_D1, V_V_D2);
+        __m256i V_V_MAD1 = _mm256_madd_epi16(V_V_M1, filter1);
+        __m256i V_V_MAD2 = _mm256_madd_epi16(V_V_M2, filter1);
+        __m256i V_V_S1 =
+            _mm256_srli_epi32(_mm256_add_epi32(V_V_MAD1, rbias), 7);
+        __m256i V_V_S2 =
+            _mm256_srli_epi32(_mm256_add_epi32(V_V_MAD2, rbias), 7);
+        opointer[curr] = _mm256_packus_epi32(V_V_S1, V_V_S2);
+
+        __m256i V_S_M1 = _mm256_unpacklo_epi16(opointer[prev], opointer[curr]);
+        __m256i V_S_M2 = _mm256_unpackhi_epi16(opointer[prev], opointer[curr]);
+
+        __m256i V_S_MAD1 = _mm256_madd_epi16(V_S_M1, filter2);
+        __m256i V_S_MAD2 = _mm256_madd_epi16(V_S_M2, filter2);
+
+        __m256i V_S_S1 =
+            _mm256_srli_epi32(_mm256_add_epi32(V_S_MAD1, rbias), 7);
+        __m256i V_S_S2 =
+            _mm256_srli_epi32(_mm256_add_epi32(V_S_MAD2, rbias), 7);
+
+        __m256i V_S_SRC = _mm256_packus_epi32(V_S_S1, V_S_S2);
+
+        __m256i V_D_DST = _mm256_loadu_si256((const __m256i *)dst_ptr);
+        dst_ptr += dst_stride;
+
+        __m256i V_R_SUB = _mm256_sub_epi16(V_S_SRC, V_D_DST);
+        __m256i V_R_MAD = _mm256_madd_epi16(V_R_SUB, V_R_SUB);
+
+        sum1 = _mm256_add_epi16(sum1, V_R_SUB);
+        sse1 = _mm256_add_epi32(sse1, V_R_MAD);
+      }
+
+      __m256i v_sum0 = _mm256_madd_epi16(sum1, one);
+      __m256i v_d_l = _mm256_unpacklo_epi32(v_sum0, sse1);
+      __m256i v_d_h = _mm256_unpackhi_epi32(v_sum0, sse1);
+      __m256i v_d_lh = _mm256_add_epi32(v_d_l, v_d_h);
+      const __m128i v_d0_d = _mm256_castsi256_si128(v_d_lh);
+      const __m128i v_d1_d = _mm256_extracti128_si256(v_d_lh, 1);
+      __m128i v_d = _mm_add_epi32(v_d0_d, v_d1_d);
+      v_d = _mm_add_epi32(v_d, _mm_srli_si128(v_d, 8));
+      sum_long += _mm_extract_epi32(v_d, 0);
+      sse_long += _mm_extract_epi32(v_d, 1);
+    }
+
+    rshift = get_msb(output_height) + get_msb(output_width);
+  }
+
+  *sse = (uint32_t)ROUND_POWER_OF_TWO(sse_long, 4);
+  int sum = (int)ROUND_POWER_OF_TWO(sum_long, 2);
+
+  int32_t var = *sse - (uint32_t)(((int64_t)sum * sum) >> rshift);
+
+  return (var > 0) ? var : 0;
+}
+
+void aom_highbd_calc8x8var_avx2(const uint16_t *src, int src_stride,
+                                const uint16_t *ref, int ref_stride,
+                                uint32_t *sse, int *sum) {
+  __m256i v_sum_d = _mm256_setzero_si256();
+  __m256i v_sse_d = _mm256_setzero_si256();
+  for (int i = 0; i < 8; i += 2) {
+    const __m128i v_p_a0 = _mm_loadu_si128((const __m128i *)src);
+    const __m128i v_p_a1 = _mm_loadu_si128((const __m128i *)(src + src_stride));
+    const __m128i v_p_b0 = _mm_loadu_si128((const __m128i *)ref);
+    const __m128i v_p_b1 = _mm_loadu_si128((const __m128i *)(ref + ref_stride));
+    __m256i v_p_a = _mm256_castsi128_si256(v_p_a0);
+    __m256i v_p_b = _mm256_castsi128_si256(v_p_b0);
+    v_p_a = _mm256_inserti128_si256(v_p_a, v_p_a1, 1);
+    v_p_b = _mm256_inserti128_si256(v_p_b, v_p_b1, 1);
+    const __m256i v_diff = _mm256_sub_epi16(v_p_a, v_p_b);
+    const __m256i v_sqrdiff = _mm256_madd_epi16(v_diff, v_diff);
+    v_sum_d = _mm256_add_epi16(v_sum_d, v_diff);
+    v_sse_d = _mm256_add_epi32(v_sse_d, v_sqrdiff);
+    src += src_stride * 2;
+    ref += ref_stride * 2;
+  }
+  __m256i v_sum00 = _mm256_cvtepi16_epi32(_mm256_castsi256_si128(v_sum_d));
+  __m256i v_sum01 = _mm256_cvtepi16_epi32(_mm256_extracti128_si256(v_sum_d, 1));
+  __m256i v_sum0 = _mm256_add_epi32(v_sum00, v_sum01);
+  __m256i v_d_l = _mm256_unpacklo_epi32(v_sum0, v_sse_d);
+  __m256i v_d_h = _mm256_unpackhi_epi32(v_sum0, v_sse_d);
+  __m256i v_d_lh = _mm256_add_epi32(v_d_l, v_d_h);
+  const __m128i v_d0_d = _mm256_castsi256_si128(v_d_lh);
+  const __m128i v_d1_d = _mm256_extracti128_si256(v_d_lh, 1);
+  __m128i v_d = _mm_add_epi32(v_d0_d, v_d1_d);
+  v_d = _mm_add_epi32(v_d, _mm_srli_si128(v_d, 8));
+  *sum = _mm_extract_epi32(v_d, 0);
+  *sse = _mm_extract_epi32(v_d, 1);
+}
+
+void aom_highbd_calc16x16var_avx2(const uint16_t *src, int src_stride,
+                                  const uint16_t *ref, int ref_stride,
+                                  uint32_t *sse, int *sum) {
+  __m256i v_sum_d = _mm256_setzero_si256();
+  __m256i v_sse_d = _mm256_setzero_si256();
+  const __m256i one = _mm256_set1_epi16(1);
+  for (int i = 0; i < 16; ++i) {
+    const __m256i v_p_a = _mm256_loadu_si256((const __m256i *)src);
+    const __m256i v_p_b = _mm256_loadu_si256((const __m256i *)ref);
+    const __m256i v_diff = _mm256_sub_epi16(v_p_a, v_p_b);
+    const __m256i v_sqrdiff = _mm256_madd_epi16(v_diff, v_diff);
+    v_sum_d = _mm256_add_epi16(v_sum_d, v_diff);
+    v_sse_d = _mm256_add_epi32(v_sse_d, v_sqrdiff);
+    src += src_stride;
+    ref += ref_stride;
+  }
+  __m256i v_sum0 = _mm256_madd_epi16(v_sum_d, one);
+  __m256i v_d_l = _mm256_unpacklo_epi32(v_sum0, v_sse_d);
+  __m256i v_d_h = _mm256_unpackhi_epi32(v_sum0, v_sse_d);
+  __m256i v_d_lh = _mm256_add_epi32(v_d_l, v_d_h);
+  const __m128i v_d0_d = _mm256_castsi256_si128(v_d_lh);
+  const __m128i v_d1_d = _mm256_extracti128_si256(v_d_lh, 1);
+  __m128i v_d = _mm_add_epi32(v_d0_d, v_d1_d);
+  v_d = _mm_add_epi32(v_d, _mm_srli_si128(v_d, 8));
+  *sum = _mm_extract_epi32(v_d, 0);
+  *sse = _mm_extract_epi32(v_d, 1);
+}
+
+static void highbd_10_variance_avx2(const uint16_t *src, int src_stride,
+                                    const uint16_t *ref, int ref_stride, int w,
+                                    int h, uint32_t *sse, int *sum,
+                                    high_variance_fn_t var_fn, int block_size) {
+  int i, j;
+  uint64_t sse_long = 0;
+  int32_t sum_long = 0;
+
+  for (i = 0; i < h; i += block_size) {
+    for (j = 0; j < w; j += block_size) {
+      unsigned int sse0;
+      int sum0;
+      var_fn(src + src_stride * i + j, src_stride, ref + ref_stride * i + j,
+             ref_stride, &sse0, &sum0);
+      sse_long += sse0;
+      sum_long += sum0;
+    }
+  }
+  *sum = ROUND_POWER_OF_TWO(sum_long, 2);
+  *sse = (uint32_t)ROUND_POWER_OF_TWO(sse_long, 4);
+}
+
+#define VAR_FN(w, h, block_size, shift)                                    \
+  uint32_t aom_highbd_10_variance##w##x##h##_avx2(                         \
+      const uint8_t *src8, int src_stride, const uint8_t *ref8,            \
+      int ref_stride, uint32_t *sse) {                                     \
+    int sum;                                                               \
+    int64_t var;                                                           \
+    uint16_t *src = CONVERT_TO_SHORTPTR(src8);                             \
+    uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);                             \
+    highbd_10_variance_avx2(                                               \
+        src, src_stride, ref, ref_stride, w, h, sse, &sum,                 \
+        aom_highbd_calc##block_size##x##block_size##var_avx2, block_size); \
+    var = (int64_t)(*sse) - (((int64_t)sum * sum) >> shift);               \
+    return (var >= 0) ? (uint32_t)var : 0;                                 \
+  }
+
+VAR_FN(128, 128, 16, 14)
+VAR_FN(128, 64, 16, 13)
+VAR_FN(64, 128, 16, 13)
+VAR_FN(64, 64, 16, 12)
+VAR_FN(64, 32, 16, 11)
+VAR_FN(32, 64, 16, 11)
+VAR_FN(32, 32, 16, 10)
+VAR_FN(32, 16, 16, 9)
+VAR_FN(16, 32, 16, 9)
+VAR_FN(16, 16, 16, 8)
+VAR_FN(16, 8, 8, 7)
+VAR_FN(8, 16, 8, 7)
+VAR_FN(8, 8, 8, 6)
+
+#if !CONFIG_REALTIME_ONLY
+VAR_FN(16, 64, 16, 10)
+VAR_FN(32, 8, 8, 8)
+VAR_FN(64, 16, 16, 10)
+VAR_FN(8, 32, 8, 8)
+#endif  // !CONFIG_REALTIME_ONLY
+
+#undef VAR_FN
+
+#define SSE2_HEIGHT(H)                                                 \
+  uint32_t aom_highbd_10_sub_pixel_variance8x##H##_sse2(               \
+      const uint8_t *src8, int src_stride, int x_offset, int y_offset, \
+      const uint8_t *dst8, int dst_stride, uint32_t *sse_ptr);
+
+SSE2_HEIGHT(8)
+SSE2_HEIGHT(16)
+
+#undef SSE2_Height
+
+#define HIGHBD_SUBPIX_VAR(W, H)                                              \
+  uint32_t aom_highbd_10_sub_pixel_variance##W##x##H##_avx2(                 \
+      const uint8_t *src, int src_stride, int xoffset, int yoffset,          \
+      const uint8_t *dst, int dst_stride, uint32_t *sse) {                   \
+    if (W == 8 && H == 16)                                                   \
+      return aom_highbd_10_sub_pixel_variance8x16_sse2(                      \
+          src, src_stride, xoffset, yoffset, dst, dst_stride, sse);          \
+    else if (W == 8 && H == 8)                                               \
+      return aom_highbd_10_sub_pixel_variance8x8_sse2(                       \
+          src, src_stride, xoffset, yoffset, dst, dst_stride, sse);          \
+    else                                                                     \
+      return aom_highbd_var_filter_block2d_bil_avx2(                         \
+          src, src_stride, 1, H, W, xoffset, yoffset, dst, dst_stride, sse); \
+  }
+
+HIGHBD_SUBPIX_VAR(128, 128)
+HIGHBD_SUBPIX_VAR(128, 64)
+HIGHBD_SUBPIX_VAR(64, 128)
+HIGHBD_SUBPIX_VAR(64, 64)
+HIGHBD_SUBPIX_VAR(64, 32)
+HIGHBD_SUBPIX_VAR(32, 64)
+HIGHBD_SUBPIX_VAR(32, 32)
+HIGHBD_SUBPIX_VAR(32, 16)
+HIGHBD_SUBPIX_VAR(16, 32)
+HIGHBD_SUBPIX_VAR(16, 16)
+HIGHBD_SUBPIX_VAR(16, 8)
+HIGHBD_SUBPIX_VAR(8, 16)
+HIGHBD_SUBPIX_VAR(8, 8)
+
+#undef HIGHBD_SUBPIX_VAR
+
+uint64_t aom_mse_4xh_16bit_highbd_avx2(uint16_t *dst, int dstride,
+                                       uint16_t *src, int sstride, int h) {
+  uint64_t sum = 0;
+  __m128i reg0_4x16, reg1_4x16, reg2_4x16, reg3_4x16;
+  __m256i src0_8x16, src1_8x16, src_16x16;
+  __m256i dst0_8x16, dst1_8x16, dst_16x16;
+  __m256i res0_4x64, res1_4x64, res2_4x64, res3_4x64;
+  __m256i sub_result;
+  const __m256i zeros = _mm256_broadcastsi128_si256(_mm_setzero_si128());
+  __m256i square_result = _mm256_broadcastsi128_si256(_mm_setzero_si128());
+  for (int i = 0; i < h; i += 4) {
+    reg0_4x16 = _mm_loadl_epi64((__m128i const *)(&dst[(i + 0) * dstride]));
+    reg1_4x16 = _mm_loadl_epi64((__m128i const *)(&dst[(i + 1) * dstride]));
+    reg2_4x16 = _mm_loadl_epi64((__m128i const *)(&dst[(i + 2) * dstride]));
+    reg3_4x16 = _mm_loadl_epi64((__m128i const *)(&dst[(i + 3) * dstride]));
+    dst0_8x16 =
+        _mm256_castsi128_si256(_mm_unpacklo_epi64(reg0_4x16, reg1_4x16));
+    dst1_8x16 =
+        _mm256_castsi128_si256(_mm_unpacklo_epi64(reg2_4x16, reg3_4x16));
+    dst_16x16 = _mm256_permute2x128_si256(dst0_8x16, dst1_8x16, 0x20);
+
+    reg0_4x16 = _mm_loadl_epi64((__m128i const *)(&src[(i + 0) * sstride]));
+    reg1_4x16 = _mm_loadl_epi64((__m128i const *)(&src[(i + 1) * sstride]));
+    reg2_4x16 = _mm_loadl_epi64((__m128i const *)(&src[(i + 2) * sstride]));
+    reg3_4x16 = _mm_loadl_epi64((__m128i const *)(&src[(i + 3) * sstride]));
+    src0_8x16 =
+        _mm256_castsi128_si256(_mm_unpacklo_epi64(reg0_4x16, reg1_4x16));
+    src1_8x16 =
+        _mm256_castsi128_si256(_mm_unpacklo_epi64(reg2_4x16, reg3_4x16));
+    src_16x16 = _mm256_permute2x128_si256(src0_8x16, src1_8x16, 0x20);
+
+    sub_result = _mm256_abs_epi16(_mm256_sub_epi16(src_16x16, dst_16x16));
+
+    src_16x16 = _mm256_unpacklo_epi16(sub_result, zeros);
+    dst_16x16 = _mm256_unpackhi_epi16(sub_result, zeros);
+
+    src_16x16 = _mm256_madd_epi16(src_16x16, src_16x16);
+    dst_16x16 = _mm256_madd_epi16(dst_16x16, dst_16x16);
+
+    res0_4x64 = _mm256_unpacklo_epi32(src_16x16, zeros);
+    res1_4x64 = _mm256_unpackhi_epi32(src_16x16, zeros);
+    res2_4x64 = _mm256_unpacklo_epi32(dst_16x16, zeros);
+    res3_4x64 = _mm256_unpackhi_epi32(dst_16x16, zeros);
+
+    square_result = _mm256_add_epi64(
+        square_result,
+        _mm256_add_epi64(
+            _mm256_add_epi64(_mm256_add_epi64(res0_4x64, res1_4x64), res2_4x64),
+            res3_4x64));
+  }
+  const __m128i sum_2x64 =
+      _mm_add_epi64(_mm256_castsi256_si128(square_result),
+                    _mm256_extracti128_si256(square_result, 1));
+  const __m128i sum_1x64 = _mm_add_epi64(sum_2x64, _mm_srli_si128(sum_2x64, 8));
+  xx_storel_64(&sum, sum_1x64);
+  return sum;
+}
+
+uint64_t aom_mse_8xh_16bit_highbd_avx2(uint16_t *dst, int dstride,
+                                       uint16_t *src, int sstride, int h) {
+  uint64_t sum = 0;
+  __m256i src0_8x16, src1_8x16, src_16x16;
+  __m256i dst0_8x16, dst1_8x16, dst_16x16;
+  __m256i res0_4x64, res1_4x64, res2_4x64, res3_4x64;
+  __m256i sub_result;
+  const __m256i zeros = _mm256_broadcastsi128_si256(_mm_setzero_si128());
+  __m256i square_result = _mm256_broadcastsi128_si256(_mm_setzero_si128());
+
+  for (int i = 0; i < h; i += 2) {
+    dst0_8x16 =
+        _mm256_castsi128_si256(_mm_loadu_si128((__m128i *)&dst[i * dstride]));
+    dst1_8x16 = _mm256_castsi128_si256(
+        _mm_loadu_si128((__m128i *)&dst[(i + 1) * dstride]));
+    dst_16x16 = _mm256_permute2x128_si256(dst0_8x16, dst1_8x16, 0x20);
+
+    src0_8x16 =
+        _mm256_castsi128_si256(_mm_loadu_si128((__m128i *)&src[i * sstride]));
+    src1_8x16 = _mm256_castsi128_si256(
+        _mm_loadu_si128((__m128i *)&src[(i + 1) * sstride]));
+    src_16x16 = _mm256_permute2x128_si256(src0_8x16, src1_8x16, 0x20);
+
+    sub_result = _mm256_abs_epi16(_mm256_sub_epi16(src_16x16, dst_16x16));
+
+    src_16x16 = _mm256_unpacklo_epi16(sub_result, zeros);
+    dst_16x16 = _mm256_unpackhi_epi16(sub_result, zeros);
+
+    src_16x16 = _mm256_madd_epi16(src_16x16, src_16x16);
+    dst_16x16 = _mm256_madd_epi16(dst_16x16, dst_16x16);
+
+    res0_4x64 = _mm256_unpacklo_epi32(src_16x16, zeros);
+    res1_4x64 = _mm256_unpackhi_epi32(src_16x16, zeros);
+    res2_4x64 = _mm256_unpacklo_epi32(dst_16x16, zeros);
+    res3_4x64 = _mm256_unpackhi_epi32(dst_16x16, zeros);
+
+    square_result = _mm256_add_epi64(
+        square_result,
+        _mm256_add_epi64(
+            _mm256_add_epi64(_mm256_add_epi64(res0_4x64, res1_4x64), res2_4x64),
+            res3_4x64));
+  }
+
+  const __m128i sum_2x64 =
+      _mm_add_epi64(_mm256_castsi256_si128(square_result),
+                    _mm256_extracti128_si256(square_result, 1));
+  const __m128i sum_1x64 = _mm_add_epi64(sum_2x64, _mm_srli_si128(sum_2x64, 8));
+  xx_storel_64(&sum, sum_1x64);
+  return sum;
+}
+
+uint64_t aom_mse_wxh_16bit_highbd_avx2(uint16_t *dst, int dstride,
+                                       uint16_t *src, int sstride, int w,
+                                       int h) {
+  assert((w == 8 || w == 4) && (h == 8 || h == 4) &&
+         "w=8/4 and h=8/4 must satisfy");
+  switch (w) {
+    case 4: return aom_mse_4xh_16bit_highbd_avx2(dst, dstride, src, sstride, h);
+    case 8: return aom_mse_8xh_16bit_highbd_avx2(dst, dstride, src, sstride, h);
+    default: assert(0 && "unsupported width"); return -1;
+  }
+}
diff --git a/third_party/aom/aom_dsp/x86/highbd_variance_impl_sse2.asm b/third_party/aom/aom_dsp/x86/highbd_variance_impl_sse2.asm
new file mode 100644
index 0000000000..ec6c7e9fa7
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/highbd_variance_impl_sse2.asm
@@ -0,0 +1,318 @@
+;
+; Copyright (c) 2016, 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 "aom_ports/x86_abi_support.asm"
+
+SECTION .text
+
+;unsigned int aom_highbd_calc16x16var_sse2
+;(
+;    unsigned char   *  src_ptr,
+;    int             source_stride,
+;    unsigned char   *  ref_ptr,
+;    int             recon_stride,
+;    unsigned int    *  SSE,
+;    int             *  Sum
+;)
+globalsym(aom_highbd_calc16x16var_sse2)
+sym(aom_highbd_calc16x16var_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push rbx
+    push rsi
+    push rdi
+    ; end prolog
+
+        mov         rsi,            arg(0) ;[src_ptr]
+        mov         rdi,            arg(2) ;[ref_ptr]
+
+        movsxd      rax,            DWORD PTR arg(1) ;[source_stride]
+        movsxd      rdx,            DWORD PTR arg(3) ;[recon_stride]
+        add         rax,            rax ; source stride in bytes
+        add         rdx,            rdx ; recon stride in bytes
+
+        ; Prefetch data
+        prefetcht0      [rsi]
+        prefetcht0      [rsi+16]
+        prefetcht0      [rsi+rax]
+        prefetcht0      [rsi+rax+16]
+        lea             rbx,    [rsi+rax*2]
+        prefetcht0      [rbx]
+        prefetcht0      [rbx+16]
+        prefetcht0      [rbx+rax]
+        prefetcht0      [rbx+rax+16]
+
+        prefetcht0      [rdi]
+        prefetcht0      [rdi+16]
+        prefetcht0      [rdi+rdx]
+        prefetcht0      [rdi+rdx+16]
+        lea             rbx,    [rdi+rdx*2]
+        prefetcht0      [rbx]
+        prefetcht0      [rbx+16]
+        prefetcht0      [rbx+rdx]
+        prefetcht0      [rbx+rdx+16]
+
+        pxor        xmm0,           xmm0     ; clear xmm0 for unpack
+        pxor        xmm7,           xmm7     ; clear xmm7 for accumulating diffs
+
+        pxor        xmm6,           xmm6     ; clear xmm6 for accumulating sse
+        mov         rcx,            16
+
+.var16loop:
+        movdqu      xmm1,           XMMWORD PTR [rsi]
+        movdqu      xmm2,           XMMWORD PTR [rdi]
+
+        lea             rbx,    [rsi+rax*2]
+        prefetcht0      [rbx]
+        prefetcht0      [rbx+16]
+        prefetcht0      [rbx+rax]
+        prefetcht0      [rbx+rax+16]
+        lea             rbx,    [rdi+rdx*2]
+        prefetcht0      [rbx]
+        prefetcht0      [rbx+16]
+        prefetcht0      [rbx+rdx]
+        prefetcht0      [rbx+rdx+16]
+
+        pxor        xmm5,           xmm5
+
+        psubw       xmm1,           xmm2
+        movdqu      xmm3,           XMMWORD PTR [rsi+16]
+        paddw       xmm5,           xmm1
+        pmaddwd     xmm1,           xmm1
+        movdqu      xmm2,           XMMWORD PTR [rdi+16]
+        paddd       xmm6,           xmm1
+
+        psubw       xmm3,           xmm2
+        movdqu      xmm1,           XMMWORD PTR [rsi+rax]
+        paddw       xmm5,           xmm3
+        pmaddwd     xmm3,           xmm3
+        movdqu      xmm2,           XMMWORD PTR [rdi+rdx]
+        paddd       xmm6,           xmm3
+
+        psubw       xmm1,           xmm2
+        movdqu      xmm3,           XMMWORD PTR [rsi+rax+16]
+        paddw       xmm5,           xmm1
+        pmaddwd     xmm1,           xmm1
+        movdqu      xmm2,           XMMWORD PTR [rdi+rdx+16]
+        paddd       xmm6,           xmm1
+
+        psubw       xmm3,           xmm2
+        paddw       xmm5,           xmm3
+        pmaddwd     xmm3,           xmm3
+        paddd       xmm6,           xmm3
+
+        movdqa      xmm1,           xmm5
+        movdqa      xmm2,           xmm5
+        pcmpgtw     xmm1,           xmm0
+        pcmpeqw     xmm2,           xmm0
+        por         xmm1,           xmm2
+        pcmpeqw     xmm1,           xmm0
+        movdqa      xmm2,           xmm5
+        punpcklwd   xmm5,           xmm1
+        punpckhwd   xmm2,           xmm1
+        paddd       xmm7,           xmm5
+        paddd       xmm7,           xmm2
+
+        lea         rsi,            [rsi + 2*rax]
+        lea         rdi,            [rdi + 2*rdx]
+        sub         rcx,            2
+        jnz         .var16loop
+
+        movdqa      xmm4,           xmm6
+        punpckldq   xmm6,           xmm0
+
+        punpckhdq   xmm4,           xmm0
+        movdqa      xmm5,           xmm7
+
+        paddd       xmm6,           xmm4
+        punpckldq   xmm7,           xmm0
+
+        punpckhdq   xmm5,           xmm0
+        paddd       xmm7,           xmm5
+
+        movdqa      xmm4,           xmm6
+        movdqa      xmm5,           xmm7
+
+        psrldq      xmm4,           8
+        psrldq      xmm5,           8
+
+        paddd       xmm6,           xmm4
+        paddd       xmm7,           xmm5
+
+        mov         rdi,            arg(4)   ; [SSE]
+        mov         rax,            arg(5)   ; [Sum]
+
+        movd DWORD PTR [rdi],       xmm6
+        movd DWORD PTR [rax],       xmm7
+
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    pop rbx
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+
+;unsigned int aom_highbd_calc8x8var_sse2
+;(
+;    unsigned char   *  src_ptr,
+;    int             source_stride,
+;    unsigned char   *  ref_ptr,
+;    int             recon_stride,
+;    unsigned int    *  SSE,
+;    int             *  Sum
+;)
+globalsym(aom_highbd_calc8x8var_sse2)
+sym(aom_highbd_calc8x8var_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 6
+    SAVE_XMM 7
+    push rbx
+    push rsi
+    push rdi
+    ; end prolog
+
+        mov         rsi,            arg(0) ;[src_ptr]
+        mov         rdi,            arg(2) ;[ref_ptr]
+
+        movsxd      rax,            DWORD PTR arg(1) ;[source_stride]
+        movsxd      rdx,            DWORD PTR arg(3) ;[recon_stride]
+        add         rax,            rax ; source stride in bytes
+        add         rdx,            rdx ; recon stride in bytes
+
+        ; Prefetch data
+        prefetcht0      [rsi]
+        prefetcht0      [rsi+rax]
+        lea             rbx,    [rsi+rax*2]
+        prefetcht0      [rbx]
+        prefetcht0      [rbx+rax]
+
+        prefetcht0      [rdi]
+        prefetcht0      [rdi+rdx]
+        lea             rbx,    [rdi+rdx*2]
+        prefetcht0      [rbx]
+        prefetcht0      [rbx+rdx]
+
+        pxor        xmm0,           xmm0     ; clear xmm0 for unpack
+        pxor        xmm7,           xmm7     ; clear xmm7 for accumulating diffs
+
+        pxor        xmm6,           xmm6     ; clear xmm6 for accumulating sse
+        mov         rcx,            8
+
+.var8loop:
+        movdqu      xmm1,           XMMWORD PTR [rsi]
+        movdqu      xmm2,           XMMWORD PTR [rdi]
+
+        lea             rbx,    [rsi+rax*4]
+        prefetcht0      [rbx]
+        prefetcht0      [rbx+rax]
+        lea             rbx,    [rbx+rax*2]
+        prefetcht0      [rbx]
+        prefetcht0      [rbx+rax]
+        lea             rbx,    [rdi+rdx*4]
+        prefetcht0      [rbx]
+        prefetcht0      [rbx+rdx]
+        lea             rbx,    [rbx+rdx*2]
+        prefetcht0      [rbx]
+        prefetcht0      [rbx+rdx]
+
+        pxor        xmm5,           xmm5
+
+        psubw       xmm1,           xmm2
+        movdqu      xmm3,           XMMWORD PTR [rsi+rax]
+        paddw       xmm5,           xmm1
+        pmaddwd     xmm1,           xmm1
+        movdqu      xmm2,           XMMWORD PTR [rdi+rdx]
+        paddd       xmm6,           xmm1
+
+        lea         rsi,            [rsi + 2*rax]
+        lea         rdi,            [rdi + 2*rdx]
+
+        psubw       xmm3,           xmm2
+        movdqu      xmm1,           XMMWORD PTR [rsi]
+        paddw       xmm5,           xmm3
+        pmaddwd     xmm3,           xmm3
+        movdqu      xmm2,           XMMWORD PTR [rdi]
+        paddd       xmm6,           xmm3
+
+        psubw       xmm1,           xmm2
+        movdqu      xmm3,           XMMWORD PTR [rsi+rax]
+        paddw       xmm5,           xmm1
+        pmaddwd     xmm1,           xmm1
+        movdqu      xmm2,           XMMWORD PTR [rdi+rdx]
+        paddd       xmm6,           xmm1
+
+        psubw       xmm3,           xmm2
+        paddw       xmm5,           xmm3
+        pmaddwd     xmm3,           xmm3
+        paddd       xmm6,           xmm3
+
+        movdqa      xmm1,           xmm5
+        movdqa      xmm2,           xmm5
+        pcmpgtw     xmm1,           xmm0
+        pcmpeqw     xmm2,           xmm0
+        por         xmm1,           xmm2
+        pcmpeqw     xmm1,           xmm0
+        movdqa      xmm2,           xmm5
+        punpcklwd   xmm5,           xmm1
+        punpckhwd   xmm2,           xmm1
+        paddd       xmm7,           xmm5
+        paddd       xmm7,           xmm2
+
+        lea         rsi,            [rsi + 2*rax]
+        lea         rdi,            [rdi + 2*rdx]
+        sub         rcx,            4
+        jnz         .var8loop
+
+        movdqa      xmm4,           xmm6
+        punpckldq   xmm6,           xmm0
+
+        punpckhdq   xmm4,           xmm0
+        movdqa      xmm5,           xmm7
+
+        paddd       xmm6,           xmm4
+        punpckldq   xmm7,           xmm0
+
+        punpckhdq   xmm5,           xmm0
+        paddd       xmm7,           xmm5
+
+        movdqa      xmm4,           xmm6
+        movdqa      xmm5,           xmm7
+
+        psrldq      xmm4,           8
+        psrldq      xmm5,           8
+
+        paddd       xmm6,           xmm4
+        paddd       xmm7,           xmm5
+
+        mov         rdi,            arg(4)   ; [SSE]
+        mov         rax,            arg(5)   ; [Sum]
+
+        movd DWORD PTR [rdi],       xmm6
+        movd DWORD PTR [rax],       xmm7
+
+    ; begin epilog
+    pop rdi
+    pop rsi
+    pop rbx
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
diff --git a/third_party/aom/aom_dsp/x86/highbd_variance_sse2.c b/third_party/aom/aom_dsp/x86/highbd_variance_sse2.c
new file mode 100644
index 0000000000..e897aab645
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/highbd_variance_sse2.c
@@ -0,0 +1,735 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <emmintrin.h>  // SSE2
+
+#include "config/aom_config.h"
+#include "config/aom_dsp_rtcd.h"
+
+#include "aom_dsp/x86/synonyms.h"
+#include "aom_ports/mem.h"
+
+#include "av1/common/filter.h"
+#include "av1/common/reconinter.h"
+
+typedef uint32_t (*high_variance_fn_t)(const uint16_t *src, int src_stride,
+                                       const uint16_t *ref, int ref_stride,
+                                       uint32_t *sse, int *sum);
+
+uint32_t aom_highbd_calc8x8var_sse2(const uint16_t *src, int src_stride,
+                                    const uint16_t *ref, int ref_stride,
+                                    uint32_t *sse, int *sum);
+
+uint32_t aom_highbd_calc16x16var_sse2(const uint16_t *src, int src_stride,
+                                      const uint16_t *ref, int ref_stride,
+                                      uint32_t *sse, int *sum);
+
+static void highbd_8_variance_sse2(const uint16_t *src, int src_stride,
+                                   const uint16_t *ref, int ref_stride, int w,
+                                   int h, uint32_t *sse, int *sum,
+                                   high_variance_fn_t var_fn, int block_size) {
+  int i, j;
+
+  *sse = 0;
+  *sum = 0;
+
+  for (i = 0; i < h; i += block_size) {
+    for (j = 0; j < w; j += block_size) {
+      unsigned int sse0;
+      int sum0;
+      var_fn(src + src_stride * i + j, src_stride, ref + ref_stride * i + j,
+             ref_stride, &sse0, &sum0);
+      *sse += sse0;
+      *sum += sum0;
+    }
+  }
+}
+
+static void highbd_10_variance_sse2(const uint16_t *src, int src_stride,
+                                    const uint16_t *ref, int ref_stride, int w,
+                                    int h, uint32_t *sse, int *sum,
+                                    high_variance_fn_t var_fn, int block_size) {
+  int i, j;
+  uint64_t sse_long = 0;
+  int32_t sum_long = 0;
+
+  for (i = 0; i < h; i += block_size) {
+    for (j = 0; j < w; j += block_size) {
+      unsigned int sse0;
+      int sum0;
+      var_fn(src + src_stride * i + j, src_stride, ref + ref_stride * i + j,
+             ref_stride, &sse0, &sum0);
+      sse_long += sse0;
+      sum_long += sum0;
+    }
+  }
+  *sum = ROUND_POWER_OF_TWO(sum_long, 2);
+  *sse = (uint32_t)ROUND_POWER_OF_TWO(sse_long, 4);
+}
+
+static void highbd_12_variance_sse2(const uint16_t *src, int src_stride,
+                                    const uint16_t *ref, int ref_stride, int w,
+                                    int h, uint32_t *sse, int *sum,
+                                    high_variance_fn_t var_fn, int block_size) {
+  int i, j;
+  uint64_t sse_long = 0;
+  int32_t sum_long = 0;
+
+  for (i = 0; i < h; i += block_size) {
+    for (j = 0; j < w; j += block_size) {
+      unsigned int sse0;
+      int sum0;
+      var_fn(src + src_stride * i + j, src_stride, ref + ref_stride * i + j,
+             ref_stride, &sse0, &sum0);
+      sse_long += sse0;
+      sum_long += sum0;
+    }
+  }
+  *sum = ROUND_POWER_OF_TWO(sum_long, 4);
+  *sse = (uint32_t)ROUND_POWER_OF_TWO(sse_long, 8);
+}
+
+#define VAR_FN(w, h, block_size, shift)                                    \
+  uint32_t aom_highbd_8_variance##w##x##h##_sse2(                          \
+      const uint8_t *src8, int src_stride, const uint8_t *ref8,            \
+      int ref_stride, uint32_t *sse) {                                     \
+    int sum;                                                               \
+    uint16_t *src = CONVERT_TO_SHORTPTR(src8);                             \
+    uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);                             \
+    highbd_8_variance_sse2(                                                \
+        src, src_stride, ref, ref_stride, w, h, sse, &sum,                 \
+        aom_highbd_calc##block_size##x##block_size##var_sse2, block_size); \
+    return *sse - (uint32_t)(((int64_t)sum * sum) >> shift);               \
+  }                                                                        \
+                                                                           \
+  uint32_t aom_highbd_10_variance##w##x##h##_sse2(                         \
+      const uint8_t *src8, int src_stride, const uint8_t *ref8,            \
+      int ref_stride, uint32_t *sse) {                                     \
+    int sum;                                                               \
+    int64_t var;                                                           \
+    uint16_t *src = CONVERT_TO_SHORTPTR(src8);                             \
+    uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);                             \
+    highbd_10_variance_sse2(                                               \
+        src, src_stride, ref, ref_stride, w, h, sse, &sum,                 \
+        aom_highbd_calc##block_size##x##block_size##var_sse2, block_size); \
+    var = (int64_t)(*sse) - (((int64_t)sum * sum) >> shift);               \
+    return (var >= 0) ? (uint32_t)var : 0;                                 \
+  }                                                                        \
+                                                                           \
+  uint32_t aom_highbd_12_variance##w##x##h##_sse2(                         \
+      const uint8_t *src8, int src_stride, const uint8_t *ref8,            \
+      int ref_stride, uint32_t *sse) {                                     \
+    int sum;                                                               \
+    int64_t var;                                                           \
+    uint16_t *src = CONVERT_TO_SHORTPTR(src8);                             \
+    uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);                             \
+    highbd_12_variance_sse2(                                               \
+        src, src_stride, ref, ref_stride, w, h, sse, &sum,                 \
+        aom_highbd_calc##block_size##x##block_size##var_sse2, block_size); \
+    var = (int64_t)(*sse) - (((int64_t)sum * sum) >> shift);               \
+    return (var >= 0) ? (uint32_t)var : 0;                                 \
+  }
+
+VAR_FN(128, 128, 16, 14)
+VAR_FN(128, 64, 16, 13)
+VAR_FN(64, 128, 16, 13)
+VAR_FN(64, 64, 16, 12)
+VAR_FN(64, 32, 16, 11)
+VAR_FN(32, 64, 16, 11)
+VAR_FN(32, 32, 16, 10)
+VAR_FN(32, 16, 16, 9)
+VAR_FN(16, 32, 16, 9)
+VAR_FN(16, 16, 16, 8)
+VAR_FN(16, 8, 8, 7)
+VAR_FN(8, 16, 8, 7)
+VAR_FN(8, 8, 8, 6)
+VAR_FN(8, 32, 8, 8)
+VAR_FN(32, 8, 8, 8)
+VAR_FN(16, 64, 16, 10)
+VAR_FN(64, 16, 16, 10)
+
+#undef VAR_FN
+
+unsigned int aom_highbd_8_mse16x16_sse2(const uint8_t *src8, int src_stride,
+                                        const uint8_t *ref8, int ref_stride,
+                                        unsigned int *sse) {
+  int sum;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);
+  highbd_8_variance_sse2(src, src_stride, ref, ref_stride, 16, 16, sse, &sum,
+                         aom_highbd_calc16x16var_sse2, 16);
+  return *sse;
+}
+
+unsigned int aom_highbd_10_mse16x16_sse2(const uint8_t *src8, int src_stride,
+                                         const uint8_t *ref8, int ref_stride,
+                                         unsigned int *sse) {
+  int sum;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);
+  highbd_10_variance_sse2(src, src_stride, ref, ref_stride, 16, 16, sse, &sum,
+                          aom_highbd_calc16x16var_sse2, 16);
+  return *sse;
+}
+
+unsigned int aom_highbd_12_mse16x16_sse2(const uint8_t *src8, int src_stride,
+                                         const uint8_t *ref8, int ref_stride,
+                                         unsigned int *sse) {
+  int sum;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);
+  highbd_12_variance_sse2(src, src_stride, ref, ref_stride, 16, 16, sse, &sum,
+                          aom_highbd_calc16x16var_sse2, 16);
+  return *sse;
+}
+
+unsigned int aom_highbd_8_mse8x8_sse2(const uint8_t *src8, int src_stride,
+                                      const uint8_t *ref8, int ref_stride,
+                                      unsigned int *sse) {
+  int sum;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);
+  highbd_8_variance_sse2(src, src_stride, ref, ref_stride, 8, 8, sse, &sum,
+                         aom_highbd_calc8x8var_sse2, 8);
+  return *sse;
+}
+
+unsigned int aom_highbd_10_mse8x8_sse2(const uint8_t *src8, int src_stride,
+                                       const uint8_t *ref8, int ref_stride,
+                                       unsigned int *sse) {
+  int sum;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);
+  highbd_10_variance_sse2(src, src_stride, ref, ref_stride, 8, 8, sse, &sum,
+                          aom_highbd_calc8x8var_sse2, 8);
+  return *sse;
+}
+
+unsigned int aom_highbd_12_mse8x8_sse2(const uint8_t *src8, int src_stride,
+                                       const uint8_t *ref8, int ref_stride,
+                                       unsigned int *sse) {
+  int sum;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);
+  highbd_12_variance_sse2(src, src_stride, ref, ref_stride, 8, 8, sse, &sum,
+                          aom_highbd_calc8x8var_sse2, 8);
+  return *sse;
+}
+
+// The 2 unused parameters are place holders for PIC enabled build.
+// These definitions are for functions defined in
+// highbd_subpel_variance_impl_sse2.asm
+#define DECL(w, opt)                                                         \
+  int aom_highbd_sub_pixel_variance##w##xh_##opt(                            \
+      const uint16_t *src, ptrdiff_t src_stride, int x_offset, int y_offset, \
+      const uint16_t *dst, ptrdiff_t dst_stride, int height,                 \
+      unsigned int *sse, void *unused0, void *unused);
+#define DECLS(opt) \
+  DECL(8, opt)     \
+  DECL(16, opt)
+
+DECLS(sse2)
+
+#undef DECLS
+#undef DECL
+
+#define FN(w, h, wf, wlog2, hlog2, opt, cast)                                  \
+  uint32_t aom_highbd_8_sub_pixel_variance##w##x##h##_##opt(                   \
+      const uint8_t *src8, int src_stride, int x_offset, int y_offset,         \
+      const uint8_t *dst8, int dst_stride, uint32_t *sse_ptr) {                \
+    uint16_t *src = CONVERT_TO_SHORTPTR(src8);                                 \
+    uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);                                 \
+    int se = 0;                                                                \
+    unsigned int sse = 0;                                                      \
+    unsigned int sse2;                                                         \
+    int row_rep = (w > 64) ? 2 : 1;                                            \
+    for (int wd_64 = 0; wd_64 < row_rep; wd_64++) {                            \
+      src += wd_64 * 64;                                                       \
+      dst += wd_64 * 64;                                                       \
+      int se2 = aom_highbd_sub_pixel_variance##wf##xh_##opt(                   \
+          src, src_stride, x_offset, y_offset, dst, dst_stride, h, &sse2,      \
+          NULL, NULL);                                                         \
+      se += se2;                                                               \
+      sse += sse2;                                                             \
+      if (w > wf) {                                                            \
+        se2 = aom_highbd_sub_pixel_variance##wf##xh_##opt(                     \
+            src + wf, src_stride, x_offset, y_offset, dst + wf, dst_stride, h, \
+            &sse2, NULL, NULL);                                                \
+        se += se2;                                                             \
+        sse += sse2;                                                           \
+        if (w > wf * 2) {                                                      \
+          se2 = aom_highbd_sub_pixel_variance##wf##xh_##opt(                   \
+              src + 2 * wf, src_stride, x_offset, y_offset, dst + 2 * wf,      \
+              dst_stride, h, &sse2, NULL, NULL);                               \
+          se += se2;                                                           \
+          sse += sse2;                                                         \
+          se2 = aom_highbd_sub_pixel_variance##wf##xh_##opt(                   \
+              src + 3 * wf, src_stride, x_offset, y_offset, dst + 3 * wf,      \
+              dst_stride, h, &sse2, NULL, NULL);                               \
+          se += se2;                                                           \
+          sse += sse2;                                                         \
+        }                                                                      \
+      }                                                                        \
+    }                                                                          \
+    *sse_ptr = sse;                                                            \
+    return sse - (uint32_t)((cast se * se) >> (wlog2 + hlog2));                \
+  }                                                                            \
+                                                                               \
+  uint32_t aom_highbd_10_sub_pixel_variance##w##x##h##_##opt(                  \
+      const uint8_t *src8, int src_stride, int x_offset, int y_offset,         \
+      const uint8_t *dst8, int dst_stride, uint32_t *sse_ptr) {                \
+    int64_t var;                                                               \
+    uint32_t sse;                                                              \
+    uint64_t long_sse = 0;                                                     \
+    uint16_t *src = CONVERT_TO_SHORTPTR(src8);                                 \
+    uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);                                 \
+    int se = 0;                                                                \
+    int row_rep = (w > 64) ? 2 : 1;                                            \
+    for (int wd_64 = 0; wd_64 < row_rep; wd_64++) {                            \
+      src += wd_64 * 64;                                                       \
+      dst += wd_64 * 64;                                                       \
+      int se2 = aom_highbd_sub_pixel_variance##wf##xh_##opt(                   \
+          src, src_stride, x_offset, y_offset, dst, dst_stride, h, &sse, NULL, \
+          NULL);                                                               \
+      se += se2;                                                               \
+      long_sse += sse;                                                         \
+      if (w > wf) {                                                            \
+        uint32_t sse2;                                                         \
+        se2 = aom_highbd_sub_pixel_variance##wf##xh_##opt(                     \
+            src + wf, src_stride, x_offset, y_offset, dst + wf, dst_stride, h, \
+            &sse2, NULL, NULL);                                                \
+        se += se2;                                                             \
+        long_sse += sse2;                                                      \
+        if (w > wf * 2) {                                                      \
+          se2 = aom_highbd_sub_pixel_variance##wf##xh_##opt(                   \
+              src + 2 * wf, src_stride, x_offset, y_offset, dst + 2 * wf,      \
+              dst_stride, h, &sse2, NULL, NULL);                               \
+          se += se2;                                                           \
+          long_sse += sse2;                                                    \
+          se2 = aom_highbd_sub_pixel_variance##wf##xh_##opt(                   \
+              src + 3 * wf, src_stride, x_offset, y_offset, dst + 3 * wf,      \
+              dst_stride, h, &sse2, NULL, NULL);                               \
+          se += se2;                                                           \
+          long_sse += sse2;                                                    \
+        }                                                                      \
+      }                                                                        \
+    }                                                                          \
+    se = ROUND_POWER_OF_TWO(se, 2);                                            \
+    sse = (uint32_t)ROUND_POWER_OF_TWO(long_sse, 4);                           \
+    *sse_ptr = sse;                                                            \
+    var = (int64_t)(sse) - ((cast se * se) >> (wlog2 + hlog2));                \
+    return (var >= 0) ? (uint32_t)var : 0;                                     \
+  }                                                                            \
+                                                                               \
+  uint32_t aom_highbd_12_sub_pixel_variance##w##x##h##_##opt(                  \
+      const uint8_t *src8, int src_stride, int x_offset, int y_offset,         \
+      const uint8_t *dst8, int dst_stride, uint32_t *sse_ptr) {                \
+    int start_row;                                                             \
+    uint32_t sse;                                                              \
+    int se = 0;                                                                \
+    int64_t var;                                                               \
+    uint64_t long_sse = 0;                                                     \
+    uint16_t *src = CONVERT_TO_SHORTPTR(src8);                                 \
+    uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);                                 \
+    int row_rep = (w > 64) ? 2 : 1;                                            \
+    for (start_row = 0; start_row < h; start_row += 16) {                      \
+      uint32_t sse2;                                                           \
+      int height = h - start_row < 16 ? h - start_row : 16;                    \
+      uint16_t *src_tmp = src + (start_row * src_stride);                      \
+      uint16_t *dst_tmp = dst + (start_row * dst_stride);                      \
+      for (int wd_64 = 0; wd_64 < row_rep; wd_64++) {                          \
+        src_tmp += wd_64 * 64;                                                 \
+        dst_tmp += wd_64 * 64;                                                 \
+        int se2 = aom_highbd_sub_pixel_variance##wf##xh_##opt(                 \
+            src_tmp, src_stride, x_offset, y_offset, dst_tmp, dst_stride,      \
+            height, &sse2, NULL, NULL);                                        \
+        se += se2;                                                             \
+        long_sse += sse2;                                                      \
+        if (w > wf) {                                                          \
+          se2 = aom_highbd_sub_pixel_variance##wf##xh_##opt(                   \
+              src_tmp + wf, src_stride, x_offset, y_offset, dst_tmp + wf,      \
+              dst_stride, height, &sse2, NULL, NULL);                          \
+          se += se2;                                                           \
+          long_sse += sse2;                                                    \
+          if (w > wf * 2) {                                                    \
+            se2 = aom_highbd_sub_pixel_variance##wf##xh_##opt(                 \
+                src_tmp + 2 * wf, src_stride, x_offset, y_offset,              \
+                dst_tmp + 2 * wf, dst_stride, height, &sse2, NULL, NULL);      \
+            se += se2;                                                         \
+            long_sse += sse2;                                                  \
+            se2 = aom_highbd_sub_pixel_variance##wf##xh_##opt(                 \
+                src_tmp + 3 * wf, src_stride, x_offset, y_offset,              \
+                dst_tmp + 3 * wf, dst_stride, height, &sse2, NULL, NULL);      \
+            se += se2;                                                         \
+            long_sse += sse2;                                                  \
+          }                                                                    \
+        }                                                                      \
+      }                                                                        \
+    }                                                                          \
+    se = ROUND_POWER_OF_TWO(se, 4);                                            \
+    sse = (uint32_t)ROUND_POWER_OF_TWO(long_sse, 8);                           \
+    *sse_ptr = sse;                                                            \
+    var = (int64_t)(sse) - ((cast se * se) >> (wlog2 + hlog2));                \
+    return (var >= 0) ? (uint32_t)var : 0;                                     \
+  }
+
+#define FNS(opt)                         \
+  FN(128, 128, 16, 7, 7, opt, (int64_t)) \
+  FN(128, 64, 16, 7, 6, opt, (int64_t))  \
+  FN(64, 128, 16, 6, 7, opt, (int64_t))  \
+  FN(64, 64, 16, 6, 6, opt, (int64_t))   \
+  FN(64, 32, 16, 6, 5, opt, (int64_t))   \
+  FN(32, 64, 16, 5, 6, opt, (int64_t))   \
+  FN(32, 32, 16, 5, 5, opt, (int64_t))   \
+  FN(32, 16, 16, 5, 4, opt, (int64_t))   \
+  FN(16, 32, 16, 4, 5, opt, (int64_t))   \
+  FN(16, 16, 16, 4, 4, opt, (int64_t))   \
+  FN(16, 8, 16, 4, 3, opt, (int64_t))    \
+  FN(8, 16, 8, 3, 4, opt, (int64_t))     \
+  FN(8, 8, 8, 3, 3, opt, (int64_t))      \
+  FN(8, 4, 8, 3, 2, opt, (int64_t))      \
+  FN(16, 4, 16, 4, 2, opt, (int64_t))    \
+  FN(8, 32, 8, 3, 5, opt, (int64_t))     \
+  FN(32, 8, 16, 5, 3, opt, (int64_t))    \
+  FN(16, 64, 16, 4, 6, opt, (int64_t))   \
+  FN(64, 16, 16, 6, 4, opt, (int64_t))
+
+FNS(sse2)
+
+#undef FNS
+#undef FN
+
+// The 2 unused parameters are place holders for PIC enabled build.
+#define DECL(w, opt)                                                         \
+  int aom_highbd_sub_pixel_avg_variance##w##xh_##opt(                        \
+      const uint16_t *src, ptrdiff_t src_stride, int x_offset, int y_offset, \
+      const uint16_t *dst, ptrdiff_t dst_stride, const uint16_t *sec,        \
+      ptrdiff_t sec_stride, int height, unsigned int *sse, void *unused0,    \
+      void *unused);
+#define DECLS(opt) \
+  DECL(16, opt)    \
+  DECL(8, opt)
+
+DECLS(sse2)
+#undef DECL
+#undef DECLS
+
+#define FN(w, h, wf, wlog2, hlog2, opt, cast)                                  \
+  uint32_t aom_highbd_8_sub_pixel_avg_variance##w##x##h##_##opt(               \
+      const uint8_t *src8, int src_stride, int x_offset, int y_offset,         \
+      const uint8_t *dst8, int dst_stride, uint32_t *sse_ptr,                  \
+      const uint8_t *sec8) {                                                   \
+    uint32_t sse;                                                              \
+    uint16_t *src = CONVERT_TO_SHORTPTR(src8);                                 \
+    uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);                                 \
+    uint16_t *sec = CONVERT_TO_SHORTPTR(sec8);                                 \
+    int se = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt(                  \
+        src, src_stride, x_offset, y_offset, dst, dst_stride, sec, w, h, &sse, \
+        NULL, NULL);                                                           \
+    if (w > wf) {                                                              \
+      uint32_t sse2;                                                           \
+      int se2 = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt(               \
+          src + wf, src_stride, x_offset, y_offset, dst + wf, dst_stride,      \
+          sec + wf, w, h, &sse2, NULL, NULL);                                  \
+      se += se2;                                                               \
+      sse += sse2;                                                             \
+      if (w > wf * 2) {                                                        \
+        se2 = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt(                 \
+            src + 2 * wf, src_stride, x_offset, y_offset, dst + 2 * wf,        \
+            dst_stride, sec + 2 * wf, w, h, &sse2, NULL, NULL);                \
+        se += se2;                                                             \
+        sse += sse2;                                                           \
+        se2 = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt(                 \
+            src + 3 * wf, src_stride, x_offset, y_offset, dst + 3 * wf,        \
+            dst_stride, sec + 3 * wf, w, h, &sse2, NULL, NULL);                \
+        se += se2;                                                             \
+        sse += sse2;                                                           \
+      }                                                                        \
+    }                                                                          \
+    *sse_ptr = sse;                                                            \
+    return sse - (uint32_t)((cast se * se) >> (wlog2 + hlog2));                \
+  }                                                                            \
+                                                                               \
+  uint32_t aom_highbd_10_sub_pixel_avg_variance##w##x##h##_##opt(              \
+      const uint8_t *src8, int src_stride, int x_offset, int y_offset,         \
+      const uint8_t *dst8, int dst_stride, uint32_t *sse_ptr,                  \
+      const uint8_t *sec8) {                                                   \
+    int64_t var;                                                               \
+    uint32_t sse;                                                              \
+    uint16_t *src = CONVERT_TO_SHORTPTR(src8);                                 \
+    uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);                                 \
+    uint16_t *sec = CONVERT_TO_SHORTPTR(sec8);                                 \
+    int se = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt(                  \
+        src, src_stride, x_offset, y_offset, dst, dst_stride, sec, w, h, &sse, \
+        NULL, NULL);                                                           \
+    if (w > wf) {                                                              \
+      uint32_t sse2;                                                           \
+      int se2 = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt(               \
+          src + wf, src_stride, x_offset, y_offset, dst + wf, dst_stride,      \
+          sec + wf, w, h, &sse2, NULL, NULL);                                  \
+      se += se2;                                                               \
+      sse += sse2;                                                             \
+      if (w > wf * 2) {                                                        \
+        se2 = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt(                 \
+            src + 2 * wf, src_stride, x_offset, y_offset, dst + 2 * wf,        \
+            dst_stride, sec + 2 * wf, w, h, &sse2, NULL, NULL);                \
+        se += se2;                                                             \
+        sse += sse2;                                                           \
+        se2 = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt(                 \
+            src + 3 * wf, src_stride, x_offset, y_offset, dst + 3 * wf,        \
+            dst_stride, sec + 3 * wf, w, h, &sse2, NULL, NULL);                \
+        se += se2;                                                             \
+        sse += sse2;                                                           \
+      }                                                                        \
+    }                                                                          \
+    se = ROUND_POWER_OF_TWO(se, 2);                                            \
+    sse = ROUND_POWER_OF_TWO(sse, 4);                                          \
+    *sse_ptr = sse;                                                            \
+    var = (int64_t)(sse) - ((cast se * se) >> (wlog2 + hlog2));                \
+    return (var >= 0) ? (uint32_t)var : 0;                                     \
+  }                                                                            \
+                                                                               \
+  uint32_t aom_highbd_12_sub_pixel_avg_variance##w##x##h##_##opt(              \
+      const uint8_t *src8, int src_stride, int x_offset, int y_offset,         \
+      const uint8_t *dst8, int dst_stride, uint32_t *sse_ptr,                  \
+      const uint8_t *sec8) {                                                   \
+    int start_row;                                                             \
+    int64_t var;                                                               \
+    uint32_t sse;                                                              \
+    int se = 0;                                                                \
+    uint64_t long_sse = 0;                                                     \
+    uint16_t *src = CONVERT_TO_SHORTPTR(src8);                                 \
+    uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);                                 \
+    uint16_t *sec = CONVERT_TO_SHORTPTR(sec8);                                 \
+    for (start_row = 0; start_row < h; start_row += 16) {                      \
+      uint32_t sse2;                                                           \
+      int height = h - start_row < 16 ? h - start_row : 16;                    \
+      int se2 = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt(               \
+          src + (start_row * src_stride), src_stride, x_offset, y_offset,      \
+          dst + (start_row * dst_stride), dst_stride, sec + (start_row * w),   \
+          w, height, &sse2, NULL, NULL);                                       \
+      se += se2;                                                               \
+      long_sse += sse2;                                                        \
+      if (w > wf) {                                                            \
+        se2 = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt(                 \
+            src + wf + (start_row * src_stride), src_stride, x_offset,         \
+            y_offset, dst + wf + (start_row * dst_stride), dst_stride,         \
+            sec + wf + (start_row * w), w, height, &sse2, NULL, NULL);         \
+        se += se2;                                                             \
+        long_sse += sse2;                                                      \
+        if (w > wf * 2) {                                                      \
+          se2 = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt(               \
+              src + 2 * wf + (start_row * src_stride), src_stride, x_offset,   \
+              y_offset, dst + 2 * wf + (start_row * dst_stride), dst_stride,   \
+              sec + 2 * wf + (start_row * w), w, height, &sse2, NULL, NULL);   \
+          se += se2;                                                           \
+          long_sse += sse2;                                                    \
+          se2 = aom_highbd_sub_pixel_avg_variance##wf##xh_##opt(               \
+              src + 3 * wf + (start_row * src_stride), src_stride, x_offset,   \
+              y_offset, dst + 3 * wf + (start_row * dst_stride), dst_stride,   \
+              sec + 3 * wf + (start_row * w), w, height, &sse2, NULL, NULL);   \
+          se += se2;                                                           \
+          long_sse += sse2;                                                    \
+        }                                                                      \
+      }                                                                        \
+    }                                                                          \
+    se = ROUND_POWER_OF_TWO(se, 4);                                            \
+    sse = (uint32_t)ROUND_POWER_OF_TWO(long_sse, 8);                           \
+    *sse_ptr = sse;                                                            \
+    var = (int64_t)(sse) - ((cast se * se) >> (wlog2 + hlog2));                \
+    return (var >= 0) ? (uint32_t)var : 0;                                     \
+  }
+
+#define FNS(opt)                       \
+  FN(64, 64, 16, 6, 6, opt, (int64_t)) \
+  FN(64, 32, 16, 6, 5, opt, (int64_t)) \
+  FN(32, 64, 16, 5, 6, opt, (int64_t)) \
+  FN(32, 32, 16, 5, 5, opt, (int64_t)) \
+  FN(32, 16, 16, 5, 4, opt, (int64_t)) \
+  FN(16, 32, 16, 4, 5, opt, (int64_t)) \
+  FN(16, 16, 16, 4, 4, opt, (int64_t)) \
+  FN(16, 8, 16, 4, 3, opt, (int64_t))  \
+  FN(8, 16, 8, 3, 4, opt, (int64_t))   \
+  FN(8, 8, 8, 3, 3, opt, (int64_t))    \
+  FN(8, 4, 8, 3, 2, opt, (int64_t))    \
+  FN(16, 4, 16, 4, 2, opt, (int64_t))  \
+  FN(8, 32, 8, 3, 5, opt, (int64_t))   \
+  FN(32, 8, 16, 5, 3, opt, (int64_t))  \
+  FN(16, 64, 16, 4, 6, opt, (int64_t)) \
+  FN(64, 16, 16, 6, 4, opt, (int64_t))
+
+FNS(sse2)
+
+#undef FNS
+#undef FN
+
+static INLINE void highbd_compute_dist_wtd_comp_avg(__m128i *p0, __m128i *p1,
+                                                    const __m128i *w0,
+                                                    const __m128i *w1,
+                                                    const __m128i *r,
+                                                    void *const result) {
+  assert(DIST_PRECISION_BITS <= 4);
+  __m128i mult0 = _mm_mullo_epi16(*p0, *w0);
+  __m128i mult1 = _mm_mullo_epi16(*p1, *w1);
+  __m128i sum = _mm_adds_epu16(mult0, mult1);
+  __m128i round = _mm_adds_epu16(sum, *r);
+  __m128i shift = _mm_srli_epi16(round, DIST_PRECISION_BITS);
+
+  xx_storeu_128(result, shift);
+}
+
+void aom_highbd_dist_wtd_comp_avg_pred_sse2(
+    uint8_t *comp_pred8, const uint8_t *pred8, int width, int height,
+    const uint8_t *ref8, int ref_stride,
+    const DIST_WTD_COMP_PARAMS *jcp_param) {
+  int i;
+  const int16_t wt0 = (int16_t)jcp_param->fwd_offset;
+  const int16_t wt1 = (int16_t)jcp_param->bck_offset;
+  const __m128i w0 = _mm_set1_epi16(wt0);
+  const __m128i w1 = _mm_set1_epi16(wt1);
+  const int16_t round = (int16_t)((1 << DIST_PRECISION_BITS) >> 1);
+  const __m128i r = _mm_set1_epi16(round);
+  uint16_t *pred = CONVERT_TO_SHORTPTR(pred8);
+  uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);
+  uint16_t *comp_pred = CONVERT_TO_SHORTPTR(comp_pred8);
+
+  if (width >= 8) {
+    // Read 8 pixels one row at a time
+    assert(!(width & 7));
+    for (i = 0; i < height; ++i) {
+      int j;
+      for (j = 0; j < width; j += 8) {
+        __m128i p0 = xx_loadu_128(ref);
+        __m128i p1 = xx_loadu_128(pred);
+
+        highbd_compute_dist_wtd_comp_avg(&p0, &p1, &w0, &w1, &r, comp_pred);
+
+        comp_pred += 8;
+        pred += 8;
+        ref += 8;
+      }
+      ref += ref_stride - width;
+    }
+  } else {
+    // Read 4 pixels two rows at a time
+    assert(!(width & 3));
+    for (i = 0; i < height; i += 2) {
+      __m128i p0_0 = xx_loadl_64(ref + 0 * ref_stride);
+      __m128i p0_1 = xx_loadl_64(ref + 1 * ref_stride);
+      __m128i p0 = _mm_unpacklo_epi64(p0_0, p0_1);
+      __m128i p1 = xx_loadu_128(pred);
+
+      highbd_compute_dist_wtd_comp_avg(&p0, &p1, &w0, &w1, &r, comp_pred);
+
+      comp_pred += 8;
+      pred += 8;
+      ref += 2 * ref_stride;
+    }
+  }
+}
+
+uint64_t aom_mse_4xh_16bit_highbd_sse2(uint16_t *dst, int dstride,
+                                       uint16_t *src, int sstride, int h) {
+  uint64_t sum = 0;
+  __m128i reg0_4x16, reg1_4x16;
+  __m128i src_8x16;
+  __m128i dst_8x16;
+  __m128i res0_4x32, res1_4x32, res0_4x64, res1_4x64, res2_4x64, res3_4x64;
+  __m128i sub_result_8x16;
+  const __m128i zeros = _mm_setzero_si128();
+  __m128i square_result = _mm_setzero_si128();
+  for (int i = 0; i < h; i += 2) {
+    reg0_4x16 = _mm_loadl_epi64((__m128i const *)(&dst[(i + 0) * dstride]));
+    reg1_4x16 = _mm_loadl_epi64((__m128i const *)(&dst[(i + 1) * dstride]));
+    dst_8x16 = _mm_unpacklo_epi64(reg0_4x16, reg1_4x16);
+
+    reg0_4x16 = _mm_loadl_epi64((__m128i const *)(&src[(i + 0) * sstride]));
+    reg1_4x16 = _mm_loadl_epi64((__m128i const *)(&src[(i + 1) * sstride]));
+    src_8x16 = _mm_unpacklo_epi64(reg0_4x16, reg1_4x16);
+
+    sub_result_8x16 = _mm_sub_epi16(src_8x16, dst_8x16);
+
+    res0_4x32 = _mm_unpacklo_epi16(sub_result_8x16, zeros);
+    res1_4x32 = _mm_unpackhi_epi16(sub_result_8x16, zeros);
+
+    res0_4x32 = _mm_madd_epi16(res0_4x32, res0_4x32);
+    res1_4x32 = _mm_madd_epi16(res1_4x32, res1_4x32);
+
+    res0_4x64 = _mm_unpacklo_epi32(res0_4x32, zeros);
+    res1_4x64 = _mm_unpackhi_epi32(res0_4x32, zeros);
+    res2_4x64 = _mm_unpacklo_epi32(res1_4x32, zeros);
+    res3_4x64 = _mm_unpackhi_epi32(res1_4x32, zeros);
+
+    square_result = _mm_add_epi64(
+        square_result,
+        _mm_add_epi64(
+            _mm_add_epi64(_mm_add_epi64(res0_4x64, res1_4x64), res2_4x64),
+            res3_4x64));
+  }
+
+  const __m128i sum_1x64 =
+      _mm_add_epi64(square_result, _mm_srli_si128(square_result, 8));
+  xx_storel_64(&sum, sum_1x64);
+  return sum;
+}
+
+uint64_t aom_mse_8xh_16bit_highbd_sse2(uint16_t *dst, int dstride,
+                                       uint16_t *src, int sstride, int h) {
+  uint64_t sum = 0;
+  __m128i src_8x16;
+  __m128i dst_8x16;
+  __m128i res0_4x32, res1_4x32, res0_4x64, res1_4x64, res2_4x64, res3_4x64;
+  __m128i sub_result_8x16;
+  const __m128i zeros = _mm_setzero_si128();
+  __m128i square_result = _mm_setzero_si128();
+
+  for (int i = 0; i < h; i++) {
+    dst_8x16 = _mm_loadu_si128((__m128i *)&dst[i * dstride]);
+    src_8x16 = _mm_loadu_si128((__m128i *)&src[i * sstride]);
+
+    sub_result_8x16 = _mm_sub_epi16(src_8x16, dst_8x16);
+
+    res0_4x32 = _mm_unpacklo_epi16(sub_result_8x16, zeros);
+    res1_4x32 = _mm_unpackhi_epi16(sub_result_8x16, zeros);
+
+    res0_4x32 = _mm_madd_epi16(res0_4x32, res0_4x32);
+    res1_4x32 = _mm_madd_epi16(res1_4x32, res1_4x32);
+
+    res0_4x64 = _mm_unpacklo_epi32(res0_4x32, zeros);
+    res1_4x64 = _mm_unpackhi_epi32(res0_4x32, zeros);
+    res2_4x64 = _mm_unpacklo_epi32(res1_4x32, zeros);
+    res3_4x64 = _mm_unpackhi_epi32(res1_4x32, zeros);
+
+    square_result = _mm_add_epi64(
+        square_result,
+        _mm_add_epi64(
+            _mm_add_epi64(_mm_add_epi64(res0_4x64, res1_4x64), res2_4x64),
+            res3_4x64));
+  }
+
+  const __m128i sum_1x64 =
+      _mm_add_epi64(square_result, _mm_srli_si128(square_result, 8));
+  xx_storel_64(&sum, sum_1x64);
+  return sum;
+}
+
+uint64_t aom_mse_wxh_16bit_highbd_sse2(uint16_t *dst, int dstride,
+                                       uint16_t *src, int sstride, int w,
+                                       int h) {
+  assert((w == 8 || w == 4) && (h == 8 || h == 4) &&
+         "w=8/4 and h=8/4 must satisfy");
+  switch (w) {
+    case 4: return aom_mse_4xh_16bit_highbd_sse2(dst, dstride, src, sstride, h);
+    case 8: return aom_mse_8xh_16bit_highbd_sse2(dst, dstride, src, sstride, h);
+    default: assert(0 && "unsupported width"); return -1;
+  }
+}
diff --git a/third_party/aom/aom_dsp/x86/highbd_variance_sse4.c b/third_party/aom/aom_dsp/x86/highbd_variance_sse4.c
new file mode 100644
index 0000000000..df5449a9df
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/highbd_variance_sse4.c
@@ -0,0 +1,216 @@
+/*
+ * Copyright (c) 2016, 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 <smmintrin.h> /* SSE4.1 */
+
+#include "config/aom_config.h"
+#include "config/aom_dsp_rtcd.h"
+
+#include "aom_dsp/variance.h"
+#include "aom_dsp/aom_filter.h"
+
+static INLINE void variance4x4_64_sse4_1(const uint8_t *a8, int a_stride,
+                                         const uint8_t *b8, int b_stride,
+                                         uint64_t *sse, int64_t *sum) {
+  __m128i u0, u1, u2, u3;
+  __m128i s0, s1, s2, s3;
+  __m128i t0, t1, x0, y0;
+  __m128i a0, a1, a2, a3;
+  __m128i b0, b1, b2, b3;
+  __m128i k_one_epi16 = _mm_set1_epi16((int16_t)1);
+
+  uint16_t *a = CONVERT_TO_SHORTPTR(a8);
+  uint16_t *b = CONVERT_TO_SHORTPTR(b8);
+
+  a0 = _mm_loadl_epi64((__m128i const *)(a + 0 * a_stride));
+  a1 = _mm_loadl_epi64((__m128i const *)(a + 1 * a_stride));
+  a2 = _mm_loadl_epi64((__m128i const *)(a + 2 * a_stride));
+  a3 = _mm_loadl_epi64((__m128i const *)(a + 3 * a_stride));
+
+  b0 = _mm_loadl_epi64((__m128i const *)(b + 0 * b_stride));
+  b1 = _mm_loadl_epi64((__m128i const *)(b + 1 * b_stride));
+  b2 = _mm_loadl_epi64((__m128i const *)(b + 2 * b_stride));
+  b3 = _mm_loadl_epi64((__m128i const *)(b + 3 * b_stride));
+
+  u0 = _mm_unpacklo_epi16(a0, a1);
+  u1 = _mm_unpacklo_epi16(a2, a3);
+  u2 = _mm_unpacklo_epi16(b0, b1);
+  u3 = _mm_unpacklo_epi16(b2, b3);
+
+  s0 = _mm_sub_epi16(u0, u2);
+  s1 = _mm_sub_epi16(u1, u3);
+
+  t0 = _mm_madd_epi16(s0, k_one_epi16);
+  t1 = _mm_madd_epi16(s1, k_one_epi16);
+
+  s2 = _mm_hadd_epi32(t0, t1);
+  s3 = _mm_hadd_epi32(s2, s2);
+  y0 = _mm_hadd_epi32(s3, s3);
+
+  t0 = _mm_madd_epi16(s0, s0);
+  t1 = _mm_madd_epi16(s1, s1);
+
+  s2 = _mm_hadd_epi32(t0, t1);
+  s3 = _mm_hadd_epi32(s2, s2);
+  x0 = _mm_hadd_epi32(s3, s3);
+
+  *sse = (uint64_t)_mm_extract_epi32(x0, 0);
+  *sum = (int64_t)_mm_extract_epi32(y0, 0);
+}
+
+uint32_t aom_highbd_8_variance4x4_sse4_1(const uint8_t *a, int a_stride,
+                                         const uint8_t *b, int b_stride,
+                                         uint32_t *sse) {
+  int64_t sum, diff;
+  uint64_t local_sse;
+
+  variance4x4_64_sse4_1(a, a_stride, b, b_stride, &local_sse, &sum);
+  *sse = (uint32_t)local_sse;
+
+  diff = (int64_t)*sse - ((sum * sum) >> 4);
+  return (diff >= 0) ? (uint32_t)diff : 0;
+}
+
+uint32_t aom_highbd_10_variance4x4_sse4_1(const uint8_t *a, int a_stride,
+                                          const uint8_t *b, int b_stride,
+                                          uint32_t *sse) {
+  int64_t sum, diff;
+  uint64_t local_sse;
+
+  variance4x4_64_sse4_1(a, a_stride, b, b_stride, &local_sse, &sum);
+  *sse = (uint32_t)ROUND_POWER_OF_TWO(local_sse, 4);
+  sum = ROUND_POWER_OF_TWO(sum, 2);
+
+  diff = (int64_t)*sse - ((sum * sum) >> 4);
+  return (diff >= 0) ? (uint32_t)diff : 0;
+}
+
+uint32_t aom_highbd_12_variance4x4_sse4_1(const uint8_t *a, int a_stride,
+                                          const uint8_t *b, int b_stride,
+                                          uint32_t *sse) {
+  int64_t sum, diff;
+  uint64_t local_sse;
+
+  variance4x4_64_sse4_1(a, a_stride, b, b_stride, &local_sse, &sum);
+  *sse = (uint32_t)ROUND_POWER_OF_TWO(local_sse, 8);
+  sum = ROUND_POWER_OF_TWO(sum, 4);
+
+  diff = (int64_t)*sse - ((sum * sum) >> 4);
+  return diff >= 0 ? (uint32_t)diff : 0;
+}
+
+// Sub-pixel
+uint32_t aom_highbd_8_sub_pixel_variance4x4_sse4_1(
+    const uint8_t *src, int src_stride, int xoffset, int yoffset,
+    const uint8_t *dst, int dst_stride, uint32_t *sse) {
+  uint16_t fdata3[(4 + 1) * 4];
+  uint16_t temp2[4 * 4];
+
+  aom_highbd_var_filter_block2d_bil_first_pass(
+      src, fdata3, src_stride, 1, 4 + 1, 4, bilinear_filters_2t[xoffset]);
+  aom_highbd_var_filter_block2d_bil_second_pass(fdata3, temp2, 4, 4, 4, 4,
+                                                bilinear_filters_2t[yoffset]);
+
+  return aom_highbd_8_variance4x4(CONVERT_TO_BYTEPTR(temp2), 4, dst, dst_stride,
+                                  sse);
+}
+
+uint32_t aom_highbd_10_sub_pixel_variance4x4_sse4_1(
+    const uint8_t *src, int src_stride, int xoffset, int yoffset,
+    const uint8_t *dst, int dst_stride, uint32_t *sse) {
+  uint16_t fdata3[(4 + 1) * 4];
+  uint16_t temp2[4 * 4];
+
+  aom_highbd_var_filter_block2d_bil_first_pass(
+      src, fdata3, src_stride, 1, 4 + 1, 4, bilinear_filters_2t[xoffset]);
+  aom_highbd_var_filter_block2d_bil_second_pass(fdata3, temp2, 4, 4, 4, 4,
+                                                bilinear_filters_2t[yoffset]);
+
+  return aom_highbd_10_variance4x4(CONVERT_TO_BYTEPTR(temp2), 4, dst,
+                                   dst_stride, sse);
+}
+
+uint32_t aom_highbd_12_sub_pixel_variance4x4_sse4_1(
+    const uint8_t *src, int src_stride, int xoffset, int yoffset,
+    const uint8_t *dst, int dst_stride, uint32_t *sse) {
+  uint16_t fdata3[(4 + 1) * 4];
+  uint16_t temp2[4 * 4];
+
+  aom_highbd_var_filter_block2d_bil_first_pass(
+      src, fdata3, src_stride, 1, 4 + 1, 4, bilinear_filters_2t[xoffset]);
+  aom_highbd_var_filter_block2d_bil_second_pass(fdata3, temp2, 4, 4, 4, 4,
+                                                bilinear_filters_2t[yoffset]);
+
+  return aom_highbd_12_variance4x4(CONVERT_TO_BYTEPTR(temp2), 4, dst,
+                                   dst_stride, sse);
+}
+
+// Sub-pixel average
+
+uint32_t aom_highbd_8_sub_pixel_avg_variance4x4_sse4_1(
+    const uint8_t *src, int src_stride, int xoffset, int yoffset,
+    const uint8_t *dst, int dst_stride, uint32_t *sse,
+    const uint8_t *second_pred) {
+  uint16_t fdata3[(4 + 1) * 4];
+  uint16_t temp2[4 * 4];
+  DECLARE_ALIGNED(16, uint16_t, temp3[4 * 4]);
+
+  aom_highbd_var_filter_block2d_bil_first_pass(
+      src, fdata3, src_stride, 1, 4 + 1, 4, bilinear_filters_2t[xoffset]);
+  aom_highbd_var_filter_block2d_bil_second_pass(fdata3, temp2, 4, 4, 4, 4,
+                                                bilinear_filters_2t[yoffset]);
+
+  aom_highbd_comp_avg_pred(CONVERT_TO_BYTEPTR(temp3), second_pred, 4, 4,
+                           CONVERT_TO_BYTEPTR(temp2), 4);
+
+  return aom_highbd_8_variance4x4(CONVERT_TO_BYTEPTR(temp3), 4, dst, dst_stride,
+                                  sse);
+}
+
+uint32_t aom_highbd_10_sub_pixel_avg_variance4x4_sse4_1(
+    const uint8_t *src, int src_stride, int xoffset, int yoffset,
+    const uint8_t *dst, int dst_stride, uint32_t *sse,
+    const uint8_t *second_pred) {
+  uint16_t fdata3[(4 + 1) * 4];
+  uint16_t temp2[4 * 4];
+  DECLARE_ALIGNED(16, uint16_t, temp3[4 * 4]);
+
+  aom_highbd_var_filter_block2d_bil_first_pass(
+      src, fdata3, src_stride, 1, 4 + 1, 4, bilinear_filters_2t[xoffset]);
+  aom_highbd_var_filter_block2d_bil_second_pass(fdata3, temp2, 4, 4, 4, 4,
+                                                bilinear_filters_2t[yoffset]);
+
+  aom_highbd_comp_avg_pred(CONVERT_TO_BYTEPTR(temp3), second_pred, 4, 4,
+                           CONVERT_TO_BYTEPTR(temp2), 4);
+
+  return aom_highbd_10_variance4x4(CONVERT_TO_BYTEPTR(temp3), 4, dst,
+                                   dst_stride, sse);
+}
+
+uint32_t aom_highbd_12_sub_pixel_avg_variance4x4_sse4_1(
+    const uint8_t *src, int src_stride, int xoffset, int yoffset,
+    const uint8_t *dst, int dst_stride, uint32_t *sse,
+    const uint8_t *second_pred) {
+  uint16_t fdata3[(4 + 1) * 4];
+  uint16_t temp2[4 * 4];
+  DECLARE_ALIGNED(16, uint16_t, temp3[4 * 4]);
+
+  aom_highbd_var_filter_block2d_bil_first_pass(
+      src, fdata3, src_stride, 1, 4 + 1, 4, bilinear_filters_2t[xoffset]);
+  aom_highbd_var_filter_block2d_bil_second_pass(fdata3, temp2, 4, 4, 4, 4,
+                                                bilinear_filters_2t[yoffset]);
+
+  aom_highbd_comp_avg_pred(CONVERT_TO_BYTEPTR(temp3), second_pred, 4, 4,
+                           CONVERT_TO_BYTEPTR(temp2), 4);
+
+  return aom_highbd_12_variance4x4(CONVERT_TO_BYTEPTR(temp3), 4, dst,
+                                   dst_stride, sse);
+}
diff --git a/third_party/aom/aom_dsp/x86/intrapred_asm_sse2.asm b/third_party/aom/aom_dsp/x86/intrapred_asm_sse2.asm
new file mode 100644
index 0000000000..0eb632326b
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/intrapred_asm_sse2.asm
@@ -0,0 +1,608 @@
+;
+; Copyright (c) 2016, 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 "third_party/x86inc/x86inc.asm"
+
+SECTION_RODATA
+pb_1: times 16 db 1
+pw_4:  times 8 dw 4
+pw_8:  times 8 dw 8
+pw_16: times 8 dw 16
+pw_32: times 8 dw 32
+dc_128: times 16 db 128
+pw2_4:  times 8 dw 2
+pw2_8:  times 8 dw 4
+pw2_16:  times 8 dw 8
+pw2_32:  times 8 dw 16
+
+SECTION .text
+
+INIT_XMM sse2
+cglobal dc_predictor_4x4, 4, 5, 3, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  movd                  m2, [leftq]
+  movd                  m0, [aboveq]
+  pxor                  m1, m1
+  punpckldq             m0, m2
+  psadbw                m0, m1
+  paddw                 m0, [GLOBAL(pw_4)]
+  psraw                 m0, 3
+  pshuflw               m0, m0, 0x0
+  packuswb              m0, m0
+  movd      [dstq        ], m0
+  movd      [dstq+strideq], m0
+  lea                 dstq, [dstq+strideq*2]
+  movd      [dstq        ], m0
+  movd      [dstq+strideq], m0
+
+  RESTORE_GOT
+  RET
+
+INIT_XMM sse2
+cglobal dc_left_predictor_4x4, 2, 5, 2, dst, stride, above, left, goffset
+  movifnidn          leftq, leftmp
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  movd                  m0, [leftq]
+  psadbw                m0, m1
+  paddw                 m0, [GLOBAL(pw2_4)]
+  psraw                 m0, 2
+  pshuflw               m0, m0, 0x0
+  packuswb              m0, m0
+  movd      [dstq        ], m0
+  movd      [dstq+strideq], m0
+  lea                 dstq, [dstq+strideq*2]
+  movd      [dstq        ], m0
+  movd      [dstq+strideq], m0
+
+  RESTORE_GOT
+  RET
+
+INIT_XMM sse2
+cglobal dc_top_predictor_4x4, 3, 5, 2, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  movd                  m0, [aboveq]
+  psadbw                m0, m1
+  paddw                 m0, [GLOBAL(pw2_4)]
+  psraw                 m0, 2
+  pshuflw               m0, m0, 0x0
+  packuswb              m0, m0
+  movd      [dstq        ], m0
+  movd      [dstq+strideq], m0
+  lea                 dstq, [dstq+strideq*2]
+  movd      [dstq        ], m0
+  movd      [dstq+strideq], m0
+
+  RESTORE_GOT
+  RET
+
+INIT_XMM sse2
+cglobal dc_predictor_8x8, 4, 5, 3, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  movq                  m0, [aboveq]
+  movq                  m2, [leftq]
+  DEFINE_ARGS dst, stride, stride3
+  lea             stride3q, [strideq*3]
+  psadbw                m0, m1
+  psadbw                m2, m1
+  paddw                 m0, m2
+  paddw                 m0, [GLOBAL(pw_8)]
+  psraw                 m0, 4
+  punpcklbw             m0, m0
+  pshuflw               m0, m0, 0x0
+  movq    [dstq          ], m0
+  movq    [dstq+strideq  ], m0
+  movq    [dstq+strideq*2], m0
+  movq    [dstq+stride3q ], m0
+  lea                 dstq, [dstq+strideq*4]
+  movq    [dstq          ], m0
+  movq    [dstq+strideq  ], m0
+  movq    [dstq+strideq*2], m0
+  movq    [dstq+stride3q ], m0
+
+  RESTORE_GOT
+  RET
+
+INIT_XMM sse2
+cglobal dc_top_predictor_8x8, 3, 5, 2, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  movq                  m0, [aboveq]
+  DEFINE_ARGS dst, stride, stride3
+  lea             stride3q, [strideq*3]
+  psadbw                m0, m1
+  paddw                 m0, [GLOBAL(pw2_8)]
+  psraw                 m0, 3
+  punpcklbw             m0, m0
+  pshuflw               m0, m0, 0x0
+  movq    [dstq          ], m0
+  movq    [dstq+strideq  ], m0
+  movq    [dstq+strideq*2], m0
+  movq    [dstq+stride3q ], m0
+  lea                 dstq, [dstq+strideq*4]
+  movq    [dstq          ], m0
+  movq    [dstq+strideq  ], m0
+  movq    [dstq+strideq*2], m0
+  movq    [dstq+stride3q ], m0
+
+  RESTORE_GOT
+  RET
+
+INIT_XMM sse2
+cglobal dc_left_predictor_8x8, 2, 5, 2, dst, stride, above, left, goffset
+  movifnidn          leftq, leftmp
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  movq                  m0, [leftq]
+  DEFINE_ARGS dst, stride, stride3
+  lea             stride3q, [strideq*3]
+  psadbw                m0, m1
+  paddw                 m0, [GLOBAL(pw2_8)]
+  psraw                 m0, 3
+  punpcklbw             m0, m0
+  pshuflw               m0, m0, 0x0
+  movq    [dstq          ], m0
+  movq    [dstq+strideq  ], m0
+  movq    [dstq+strideq*2], m0
+  movq    [dstq+stride3q ], m0
+  lea                 dstq, [dstq+strideq*4]
+  movq    [dstq          ], m0
+  movq    [dstq+strideq  ], m0
+  movq    [dstq+strideq*2], m0
+  movq    [dstq+stride3q ], m0
+
+  RESTORE_GOT
+  RET
+
+INIT_XMM sse2
+cglobal dc_128_predictor_4x4, 2, 5, 1, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  DEFINE_ARGS dst, stride, stride3
+  lea             stride3q, [strideq*3]
+  movd     m0,        [GLOBAL(dc_128)]
+  movd    [dstq          ], m0
+  movd    [dstq+strideq  ], m0
+  movd    [dstq+strideq*2], m0
+  movd    [dstq+stride3q ], m0
+  RESTORE_GOT
+  RET
+
+INIT_XMM sse2
+cglobal dc_128_predictor_8x8, 2, 5, 1, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  DEFINE_ARGS dst, stride, stride3
+  lea             stride3q, [strideq*3]
+  movq    m0,        [GLOBAL(dc_128)]
+  movq    [dstq          ], m0
+  movq    [dstq+strideq  ], m0
+  movq    [dstq+strideq*2], m0
+  movq    [dstq+stride3q ], m0
+  lea                 dstq, [dstq+strideq*4]
+  movq    [dstq          ], m0
+  movq    [dstq+strideq  ], m0
+  movq    [dstq+strideq*2], m0
+  movq    [dstq+stride3q ], m0
+  RESTORE_GOT
+  RET
+
+INIT_XMM sse2
+cglobal dc_predictor_16x16, 4, 5, 3, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  mova                  m0, [aboveq]
+  mova                  m2, [leftq]
+  DEFINE_ARGS dst, stride, stride3, lines4
+  lea             stride3q, [strideq*3]
+  mov              lines4d, 4
+  psadbw                m0, m1
+  psadbw                m2, m1
+  paddw                 m0, m2
+  movhlps               m2, m0
+  paddw                 m0, m2
+  paddw                 m0, [GLOBAL(pw_16)]
+  psraw                 m0, 5
+  pshuflw               m0, m0, 0x0
+  punpcklqdq            m0, m0
+  packuswb              m0, m0
+.loop:
+  mova    [dstq          ], m0
+  mova    [dstq+strideq  ], m0
+  mova    [dstq+strideq*2], m0
+  mova    [dstq+stride3q ], m0
+  lea                 dstq, [dstq+strideq*4]
+  dec              lines4d
+  jnz .loop
+
+  RESTORE_GOT
+  REP_RET
+
+
+INIT_XMM sse2
+cglobal dc_top_predictor_16x16, 4, 5, 3, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  mova                  m0, [aboveq]
+  DEFINE_ARGS dst, stride, stride3, lines4
+  lea             stride3q, [strideq*3]
+  mov              lines4d, 4
+  psadbw                m0, m1
+  movhlps               m2, m0
+  paddw                 m0, m2
+  paddw                 m0, [GLOBAL(pw2_16)]
+  psraw                 m0, 4
+  pshuflw               m0, m0, 0x0
+  punpcklqdq            m0, m0
+  packuswb              m0, m0
+.loop:
+  mova    [dstq          ], m0
+  mova    [dstq+strideq  ], m0
+  mova    [dstq+strideq*2], m0
+  mova    [dstq+stride3q ], m0
+  lea                 dstq, [dstq+strideq*4]
+  dec              lines4d
+  jnz .loop
+
+  RESTORE_GOT
+  REP_RET
+
+INIT_XMM sse2
+cglobal dc_left_predictor_16x16, 4, 5, 3, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  mova                  m0, [leftq]
+  DEFINE_ARGS dst, stride, stride3, lines4
+  lea             stride3q, [strideq*3]
+  mov              lines4d, 4
+  psadbw                m0, m1
+  movhlps               m2, m0
+  paddw                 m0, m2
+  paddw                 m0, [GLOBAL(pw2_16)]
+  psraw                 m0, 4
+  pshuflw               m0, m0, 0x0
+  punpcklqdq            m0, m0
+  packuswb              m0, m0
+.loop:
+  mova    [dstq          ], m0
+  mova    [dstq+strideq  ], m0
+  mova    [dstq+strideq*2], m0
+  mova    [dstq+stride3q ], m0
+  lea                 dstq, [dstq+strideq*4]
+  dec              lines4d
+  jnz .loop
+
+  RESTORE_GOT
+  REP_RET
+
+INIT_XMM sse2
+cglobal dc_128_predictor_16x16, 4, 5, 3, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  DEFINE_ARGS dst, stride, stride3, lines4
+  lea             stride3q, [strideq*3]
+  mov              lines4d, 4
+  mova    m0,        [GLOBAL(dc_128)]
+.loop:
+  mova    [dstq          ], m0
+  mova    [dstq+strideq  ], m0
+  mova    [dstq+strideq*2], m0
+  mova    [dstq+stride3q ], m0
+  lea                 dstq, [dstq+strideq*4]
+  dec              lines4d
+  jnz .loop
+  RESTORE_GOT
+  RET
+
+
+INIT_XMM sse2
+cglobal dc_predictor_32x32, 4, 5, 5, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  mova                  m0, [aboveq]
+  mova                  m2, [aboveq+16]
+  mova                  m3, [leftq]
+  mova                  m4, [leftq+16]
+  DEFINE_ARGS dst, stride, stride3, lines4
+  lea             stride3q, [strideq*3]
+  mov              lines4d, 8
+  psadbw                m0, m1
+  psadbw                m2, m1
+  psadbw                m3, m1
+  psadbw                m4, m1
+  paddw                 m0, m2
+  paddw                 m0, m3
+  paddw                 m0, m4
+  movhlps               m2, m0
+  paddw                 m0, m2
+  paddw                 m0, [GLOBAL(pw_32)]
+  psraw                 m0, 6
+  pshuflw               m0, m0, 0x0
+  punpcklqdq            m0, m0
+  packuswb              m0, m0
+.loop:
+  mova [dstq             ], m0
+  mova [dstq          +16], m0
+  mova [dstq+strideq     ], m0
+  mova [dstq+strideq  +16], m0
+  mova [dstq+strideq*2   ], m0
+  mova [dstq+strideq*2+16], m0
+  mova [dstq+stride3q    ], m0
+  mova [dstq+stride3q +16], m0
+  lea                 dstq, [dstq+strideq*4]
+  dec              lines4d
+  jnz .loop
+
+  RESTORE_GOT
+  REP_RET
+
+INIT_XMM sse2
+cglobal dc_top_predictor_32x32, 4, 5, 5, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  mova                  m0, [aboveq]
+  mova                  m2, [aboveq+16]
+  DEFINE_ARGS dst, stride, stride3, lines4
+  lea             stride3q, [strideq*3]
+  mov              lines4d, 8
+  psadbw                m0, m1
+  psadbw                m2, m1
+  paddw                 m0, m2
+  movhlps               m2, m0
+  paddw                 m0, m2
+  paddw                 m0, [GLOBAL(pw2_32)]
+  psraw                 m0, 5
+  pshuflw               m0, m0, 0x0
+  punpcklqdq            m0, m0
+  packuswb              m0, m0
+.loop:
+  mova [dstq             ], m0
+  mova [dstq          +16], m0
+  mova [dstq+strideq     ], m0
+  mova [dstq+strideq  +16], m0
+  mova [dstq+strideq*2   ], m0
+  mova [dstq+strideq*2+16], m0
+  mova [dstq+stride3q    ], m0
+  mova [dstq+stride3q +16], m0
+  lea                 dstq, [dstq+strideq*4]
+  dec              lines4d
+  jnz .loop
+
+  RESTORE_GOT
+  REP_RET
+
+INIT_XMM sse2
+cglobal dc_left_predictor_32x32, 4, 5, 5, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  pxor                  m1, m1
+  mova                  m0, [leftq]
+  mova                  m2, [leftq+16]
+  DEFINE_ARGS dst, stride, stride3, lines4
+  lea             stride3q, [strideq*3]
+  mov              lines4d, 8
+  psadbw                m0, m1
+  psadbw                m2, m1
+  paddw                 m0, m2
+  movhlps               m2, m0
+  paddw                 m0, m2
+  paddw                 m0, [GLOBAL(pw2_32)]
+  psraw                 m0, 5
+  pshuflw               m0, m0, 0x0
+  punpcklqdq            m0, m0
+  packuswb              m0, m0
+.loop:
+  mova [dstq             ], m0
+  mova [dstq          +16], m0
+  mova [dstq+strideq     ], m0
+  mova [dstq+strideq  +16], m0
+  mova [dstq+strideq*2   ], m0
+  mova [dstq+strideq*2+16], m0
+  mova [dstq+stride3q    ], m0
+  mova [dstq+stride3q +16], m0
+  lea                 dstq, [dstq+strideq*4]
+  dec              lines4d
+  jnz .loop
+
+  RESTORE_GOT
+  REP_RET
+
+INIT_XMM sse2
+cglobal dc_128_predictor_32x32, 4, 5, 3, dst, stride, above, left, goffset
+  GET_GOT     goffsetq
+
+  DEFINE_ARGS dst, stride, stride3, lines4
+  lea             stride3q, [strideq*3]
+  mov              lines4d, 8
+  mova    m0,        [GLOBAL(dc_128)]
+.loop:
+  mova [dstq             ], m0
+  mova [dstq          +16], m0
+  mova [dstq+strideq     ], m0
+  mova [dstq+strideq  +16], m0
+  mova [dstq+strideq*2   ], m0
+  mova [dstq+strideq*2+16], m0
+  mova [dstq+stride3q    ], m0
+  mova [dstq+stride3q +16], m0
+  lea                 dstq, [dstq+strideq*4]
+  dec              lines4d
+  jnz .loop
+  RESTORE_GOT
+  RET
+
+INIT_XMM sse2
+cglobal v_predictor_4x4, 3, 3, 1, dst, stride, above
+  movd                  m0, [aboveq]
+  movd      [dstq        ], m0
+  movd      [dstq+strideq], m0
+  lea                 dstq, [dstq+strideq*2]
+  movd      [dstq        ], m0
+  movd      [dstq+strideq], m0
+  RET
+
+INIT_XMM sse2
+cglobal v_predictor_8x8, 3, 3, 1, dst, stride, above
+  movq                  m0, [aboveq]
+  DEFINE_ARGS dst, stride, stride3
+  lea             stride3q, [strideq*3]
+  movq    [dstq          ], m0
+  movq    [dstq+strideq  ], m0
+  movq    [dstq+strideq*2], m0
+  movq    [dstq+stride3q ], m0
+  lea                 dstq, [dstq+strideq*4]
+  movq    [dstq          ], m0
+  movq    [dstq+strideq  ], m0
+  movq    [dstq+strideq*2], m0
+  movq    [dstq+stride3q ], m0
+  RET
+
+INIT_XMM sse2
+cglobal v_predictor_16x16, 3, 4, 1, dst, stride, above
+  mova                  m0, [aboveq]
+  DEFINE_ARGS dst, stride, stride3, nlines4
+  lea             stride3q, [strideq*3]
+  mov              nlines4d, 4
+.loop:
+  mova    [dstq          ], m0
+  mova    [dstq+strideq  ], m0
+  mova    [dstq+strideq*2], m0
+  mova    [dstq+stride3q ], m0
+  lea                 dstq, [dstq+strideq*4]
+  dec             nlines4d
+  jnz .loop
+  REP_RET
+
+INIT_XMM sse2
+cglobal v_predictor_32x32, 3, 4, 2, dst, stride, above
+  mova                  m0, [aboveq]
+  mova                  m1, [aboveq+16]
+  DEFINE_ARGS dst, stride, stride3, nlines4
+  lea             stride3q, [strideq*3]
+  mov              nlines4d, 8
+.loop:
+  mova [dstq             ], m0
+  mova [dstq          +16], m1
+  mova [dstq+strideq     ], m0
+  mova [dstq+strideq  +16], m1
+  mova [dstq+strideq*2   ], m0
+  mova [dstq+strideq*2+16], m1
+  mova [dstq+stride3q    ], m0
+  mova [dstq+stride3q +16], m1
+  lea                 dstq, [dstq+strideq*4]
+  dec             nlines4d
+  jnz .loop
+  REP_RET
+
+INIT_XMM sse2
+cglobal h_predictor_4x4, 2, 4, 4, dst, stride, line, left
+  movifnidn          leftq, leftmp
+  movd                  m0, [leftq]
+  punpcklbw             m0, m0
+  punpcklbw             m0, m0
+  pshufd                m1, m0, 0x1
+  movd      [dstq        ], m0
+  movd      [dstq+strideq], m1
+  pshufd                m2, m0, 0x2
+  lea                 dstq, [dstq+strideq*2]
+  pshufd                m3, m0, 0x3
+  movd      [dstq        ], m2
+  movd      [dstq+strideq], m3
+  RET
+
+INIT_XMM sse2
+cglobal h_predictor_8x8, 2, 5, 3, dst, stride, line, left
+  movifnidn          leftq, leftmp
+  mov                lineq, -2
+  DEFINE_ARGS  dst, stride, line, left, stride3
+  lea             stride3q, [strideq*3]
+  movq                  m0, [leftq    ]
+  punpcklbw             m0, m0              ; l1 l1 l2 l2 ... l8 l8
+.loop:
+  pshuflw               m1, m0, 0x0         ; l1 l1 l1 l1 l1 l1 l1 l1
+  pshuflw               m2, m0, 0x55        ; l2 l2 l2 l2 l2 l2 l2 l2
+  movq      [dstq        ], m1
+  movq      [dstq+strideq], m2
+  pshuflw               m1, m0, 0xaa
+  pshuflw               m2, m0, 0xff
+  movq    [dstq+strideq*2], m1
+  movq    [dstq+stride3q ], m2
+  pshufd                m0, m0, 0xe         ; [63:0] l5 l5 l6 l6 l7 l7 l8 l8
+  inc                lineq
+  lea                 dstq, [dstq+strideq*4]
+  jnz .loop
+  REP_RET
+
+INIT_XMM sse2
+cglobal h_predictor_16x16, 2, 5, 3, dst, stride, line, left
+  movifnidn          leftq, leftmp
+  mov                lineq, -4
+  DEFINE_ARGS dst, stride, line, left, stride3
+  lea             stride3q, [strideq*3]
+.loop:
+  movd                  m0, [leftq]
+  punpcklbw             m0, m0
+  punpcklbw             m0, m0              ; l1 to l4 each repeated 4 times
+  pshufd            m1, m0, 0x0             ; l1 repeated 16 times
+  pshufd            m2, m0, 0x55            ; l2 repeated 16 times
+  mova    [dstq          ], m1
+  mova    [dstq+strideq  ], m2
+  pshufd            m1, m0, 0xaa
+  pshufd            m2, m0, 0xff
+  mova    [dstq+strideq*2], m1
+  mova    [dstq+stride3q ], m2
+  inc                lineq
+  lea                leftq, [leftq+4       ]
+  lea                 dstq, [dstq+strideq*4]
+  jnz .loop
+  REP_RET
+
+INIT_XMM sse2
+cglobal h_predictor_32x32, 2, 5, 3, dst, stride, line, left
+  movifnidn              leftq, leftmp
+  mov                    lineq, -8
+  DEFINE_ARGS dst, stride, line, left, stride3
+  lea                 stride3q, [strideq*3]
+.loop:
+  movd                      m0, [leftq]
+  punpcklbw                 m0, m0
+  punpcklbw                 m0, m0              ; l1 to l4 each repeated 4 times
+  pshufd                m1, m0, 0x0             ; l1 repeated 16 times
+  pshufd                m2, m0, 0x55            ; l2 repeated 16 times
+  mova     [dstq             ], m1
+  mova     [dstq+16          ], m1
+  mova     [dstq+strideq     ], m2
+  mova     [dstq+strideq+16  ], m2
+  pshufd                m1, m0, 0xaa
+  pshufd                m2, m0, 0xff
+  mova     [dstq+strideq*2   ], m1
+  mova     [dstq+strideq*2+16], m1
+  mova     [dstq+stride3q    ], m2
+  mova     [dstq+stride3q+16 ], m2
+  inc                    lineq
+  lea                    leftq, [leftq+4       ]
+  lea                     dstq, [dstq+strideq*4]
+  jnz .loop
+  REP_RET
diff --git a/third_party/aom/aom_dsp/x86/intrapred_avx2.c b/third_party/aom/aom_dsp/x86/intrapred_avx2.c
new file mode 100644
index 0000000000..242a548df9
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/intrapred_avx2.c
@@ -0,0 +1,4707 @@
+/*
+ * Copyright (c) 2017, 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/av1_rtcd.h"
+#include "aom_dsp/x86/intrapred_x86.h"
+#include "aom_dsp/x86/intrapred_utils.h"
+#include "aom_dsp/x86/lpf_common_sse2.h"
+
+static INLINE __m256i dc_sum_64(const uint8_t *ref) {
+  const __m256i x0 = _mm256_loadu_si256((const __m256i *)ref);
+  const __m256i x1 = _mm256_loadu_si256((const __m256i *)(ref + 32));
+  const __m256i zero = _mm256_setzero_si256();
+  __m256i y0 = _mm256_sad_epu8(x0, zero);
+  __m256i y1 = _mm256_sad_epu8(x1, zero);
+  y0 = _mm256_add_epi64(y0, y1);
+  __m256i u0 = _mm256_permute2x128_si256(y0, y0, 1);
+  y0 = _mm256_add_epi64(u0, y0);
+  u0 = _mm256_unpackhi_epi64(y0, y0);
+  return _mm256_add_epi16(y0, u0);
+}
+
+static INLINE __m256i dc_sum_32(const uint8_t *ref) {
+  const __m256i x = _mm256_loadu_si256((const __m256i *)ref);
+  const __m256i zero = _mm256_setzero_si256();
+  __m256i y = _mm256_sad_epu8(x, zero);
+  __m256i u = _mm256_permute2x128_si256(y, y, 1);
+  y = _mm256_add_epi64(u, y);
+  u = _mm256_unpackhi_epi64(y, y);
+  return _mm256_add_epi16(y, u);
+}
+
+static INLINE void row_store_32xh(const __m256i *r, int height, uint8_t *dst,
+                                  ptrdiff_t stride) {
+  for (int i = 0; i < height; ++i) {
+    _mm256_storeu_si256((__m256i *)dst, *r);
+    dst += stride;
+  }
+}
+
+static INLINE void row_store_32x2xh(const __m256i *r0, const __m256i *r1,
+                                    int height, uint8_t *dst,
+                                    ptrdiff_t stride) {
+  for (int i = 0; i < height; ++i) {
+    _mm256_storeu_si256((__m256i *)dst, *r0);
+    _mm256_storeu_si256((__m256i *)(dst + 32), *r1);
+    dst += stride;
+  }
+}
+
+static INLINE void row_store_64xh(const __m256i *r, int height, uint8_t *dst,
+                                  ptrdiff_t stride) {
+  for (int i = 0; i < height; ++i) {
+    _mm256_storeu_si256((__m256i *)dst, *r);
+    _mm256_storeu_si256((__m256i *)(dst + 32), *r);
+    dst += stride;
+  }
+}
+
+static DECLARE_ALIGNED(16, uint8_t, HighbdLoadMaskx[8][16]) = {
+  { 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 },
+  { 0, 1, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13 },
+  { 0, 1, 0, 1, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11 },
+  { 0, 1, 0, 1, 0, 1, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9 },
+  { 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 2, 3, 4, 5, 6, 7 },
+  { 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 2, 3, 4, 5 },
+  { 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 2, 3 },
+  { 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1 },
+};
+
+static DECLARE_ALIGNED(16, uint8_t, HighbdEvenOddMaskx4[4][16]) = {
+  { 0, 1, 4, 5, 8, 9, 12, 13, 2, 3, 6, 7, 10, 11, 14, 15 },
+  { 0, 1, 2, 3, 6, 7, 10, 11, 14, 15, 4, 5, 8, 9, 12, 13 },
+  { 0, 1, 0, 1, 4, 5, 8, 9, 12, 13, 0, 1, 6, 7, 10, 11 },
+  { 0, 1, 0, 1, 0, 1, 6, 7, 10, 11, 14, 15, 0, 1, 8, 9 }
+};
+
+static DECLARE_ALIGNED(16, uint8_t, HighbdEvenOddMaskx[8][32]) = {
+  { 0, 1, 4, 5, 8,  9,  12, 13, 16, 17, 20, 21, 24, 25, 28, 29,
+    2, 3, 6, 7, 10, 11, 14, 15, 18, 19, 22, 23, 26, 27, 30, 31 },
+  { 0, 1, 2, 3, 6, 7, 10, 11, 14, 15, 18, 19, 22, 23, 26, 27,
+    0, 1, 4, 5, 8, 9, 12, 13, 16, 17, 20, 21, 24, 25, 28, 29 },
+  { 0, 1, 0, 1, 4, 5, 8,  9,  12, 13, 16, 17, 20, 21, 24, 25,
+    0, 1, 0, 1, 6, 7, 10, 11, 14, 15, 18, 19, 22, 23, 26, 27 },
+  { 0, 1, 0, 1, 0, 1, 6, 7, 10, 11, 14, 15, 18, 19, 22, 23,
+    0, 1, 0, 1, 0, 1, 8, 9, 12, 13, 16, 17, 20, 21, 24, 25 },
+  { 0, 1, 0, 1, 0, 1, 0, 1, 8,  9,  12, 13, 16, 17, 20, 21,
+    0, 1, 0, 1, 0, 1, 0, 1, 10, 11, 14, 15, 18, 19, 22, 23 },
+  { 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 10, 11, 14, 15, 18, 19,
+    0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 12, 13, 16, 17, 20, 21 },
+  { 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 12, 13, 16, 17,
+    0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 14, 15, 18, 19 },
+  { 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 14, 15,
+    0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 16, 17 }
+};
+
+static DECLARE_ALIGNED(32, uint16_t, HighbdBaseMask[17][16]) = {
+  { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+  { 0xffff, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+  { 0xffff, 0xffff, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+  { 0xffff, 0xffff, 0xffff, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+  { 0xffff, 0xffff, 0xffff, 0xffff, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+  { 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+  { 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+    0 },
+  { 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0, 0, 0, 0, 0, 0, 0,
+    0, 0 },
+  { 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0, 0, 0, 0,
+    0, 0, 0, 0 },
+  { 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0,
+    0, 0, 0, 0, 0, 0 },
+  { 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff,
+    0xffff, 0, 0, 0, 0, 0, 0 },
+  { 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff,
+    0xffff, 0xffff, 0, 0, 0, 0, 0 },
+  { 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff,
+    0xffff, 0xffff, 0xffff, 0, 0, 0, 0 },
+  { 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff,
+    0xffff, 0xffff, 0xffff, 0xffff, 0, 0, 0 },
+  { 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff,
+    0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0, 0 },
+  { 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff,
+    0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0 },
+  { 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff,
+    0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff, 0xffff }
+};
+
+static INLINE void highbd_transpose16x4_8x8_sse2(__m128i *x, __m128i *d) {
+  __m128i r0, r1, r2, r3, r4, r5, r6, r7, r8, r9, r10, r11, r12, r13, r14, r15;
+
+  r0 = _mm_unpacklo_epi16(x[0], x[1]);
+  r1 = _mm_unpacklo_epi16(x[2], x[3]);
+  r2 = _mm_unpacklo_epi16(x[4], x[5]);
+  r3 = _mm_unpacklo_epi16(x[6], x[7]);
+
+  r4 = _mm_unpacklo_epi16(x[8], x[9]);
+  r5 = _mm_unpacklo_epi16(x[10], x[11]);
+  r6 = _mm_unpacklo_epi16(x[12], x[13]);
+  r7 = _mm_unpacklo_epi16(x[14], x[15]);
+
+  r8 = _mm_unpacklo_epi32(r0, r1);
+  r9 = _mm_unpackhi_epi32(r0, r1);
+  r10 = _mm_unpacklo_epi32(r2, r3);
+  r11 = _mm_unpackhi_epi32(r2, r3);
+
+  r12 = _mm_unpacklo_epi32(r4, r5);
+  r13 = _mm_unpackhi_epi32(r4, r5);
+  r14 = _mm_unpacklo_epi32(r6, r7);
+  r15 = _mm_unpackhi_epi32(r6, r7);
+
+  r0 = _mm_unpacklo_epi64(r8, r9);
+  r1 = _mm_unpackhi_epi64(r8, r9);
+  r2 = _mm_unpacklo_epi64(r10, r11);
+  r3 = _mm_unpackhi_epi64(r10, r11);
+
+  r4 = _mm_unpacklo_epi64(r12, r13);
+  r5 = _mm_unpackhi_epi64(r12, r13);
+  r6 = _mm_unpacklo_epi64(r14, r15);
+  r7 = _mm_unpackhi_epi64(r14, r15);
+
+  d[0] = _mm_unpacklo_epi64(r0, r2);
+  d[1] = _mm_unpacklo_epi64(r4, r6);
+  d[2] = _mm_unpacklo_epi64(r1, r3);
+  d[3] = _mm_unpacklo_epi64(r5, r7);
+
+  d[4] = _mm_unpackhi_epi64(r0, r2);
+  d[5] = _mm_unpackhi_epi64(r4, r6);
+  d[6] = _mm_unpackhi_epi64(r1, r3);
+  d[7] = _mm_unpackhi_epi64(r5, r7);
+}
+
+static INLINE void highbd_transpose4x16_avx2(__m256i *x, __m256i *d) {
+  __m256i w0, w1, w2, w3, ww0, ww1;
+
+  w0 = _mm256_unpacklo_epi16(x[0], x[1]);  // 00 10 01 11 02 12 03 13
+  w1 = _mm256_unpacklo_epi16(x[2], x[3]);  // 20 30 21 31 22 32 23 33
+  w2 = _mm256_unpackhi_epi16(x[0], x[1]);  // 40 50 41 51 42 52 43 53
+  w3 = _mm256_unpackhi_epi16(x[2], x[3]);  // 60 70 61 71 62 72 63 73
+
+  ww0 = _mm256_unpacklo_epi32(w0, w1);  // 00 10 20 30 01 11 21 31
+  ww1 = _mm256_unpacklo_epi32(w2, w3);  // 40 50 60 70 41 51 61 71
+
+  d[0] = _mm256_unpacklo_epi64(ww0, ww1);  // 00 10 20 30 40 50 60 70
+  d[1] = _mm256_unpackhi_epi64(ww0, ww1);  // 01 11 21 31 41 51 61 71
+
+  ww0 = _mm256_unpackhi_epi32(w0, w1);  // 02 12 22 32 03 13 23 33
+  ww1 = _mm256_unpackhi_epi32(w2, w3);  // 42 52 62 72 43 53 63 73
+
+  d[2] = _mm256_unpacklo_epi64(ww0, ww1);  // 02 12 22 32 42 52 62 72
+  d[3] = _mm256_unpackhi_epi64(ww0, ww1);  // 03 13 23 33 43 53 63 73
+}
+
+static INLINE void highbd_transpose8x16_16x8_avx2(__m256i *x, __m256i *d) {
+  __m256i w0, w1, w2, w3, ww0, ww1;
+
+  w0 = _mm256_unpacklo_epi16(x[0], x[1]);  // 00 10 01 11 02 12 03 13
+  w1 = _mm256_unpacklo_epi16(x[2], x[3]);  // 20 30 21 31 22 32 23 33
+  w2 = _mm256_unpacklo_epi16(x[4], x[5]);  // 40 50 41 51 42 52 43 53
+  w3 = _mm256_unpacklo_epi16(x[6], x[7]);  // 60 70 61 71 62 72 63 73
+
+  ww0 = _mm256_unpacklo_epi32(w0, w1);  // 00 10 20 30 01 11 21 31
+  ww1 = _mm256_unpacklo_epi32(w2, w3);  // 40 50 60 70 41 51 61 71
+
+  d[0] = _mm256_unpacklo_epi64(ww0, ww1);  // 00 10 20 30 40 50 60 70
+  d[1] = _mm256_unpackhi_epi64(ww0, ww1);  // 01 11 21 31 41 51 61 71
+
+  ww0 = _mm256_unpackhi_epi32(w0, w1);  // 02 12 22 32 03 13 23 33
+  ww1 = _mm256_unpackhi_epi32(w2, w3);  // 42 52 62 72 43 53 63 73
+
+  d[2] = _mm256_unpacklo_epi64(ww0, ww1);  // 02 12 22 32 42 52 62 72
+  d[3] = _mm256_unpackhi_epi64(ww0, ww1);  // 03 13 23 33 43 53 63 73
+
+  w0 = _mm256_unpackhi_epi16(x[0], x[1]);  // 04 14 05 15 06 16 07 17
+  w1 = _mm256_unpackhi_epi16(x[2], x[3]);  // 24 34 25 35 26 36 27 37
+  w2 = _mm256_unpackhi_epi16(x[4], x[5]);  // 44 54 45 55 46 56 47 57
+  w3 = _mm256_unpackhi_epi16(x[6], x[7]);  // 64 74 65 75 66 76 67 77
+
+  ww0 = _mm256_unpacklo_epi32(w0, w1);  // 04 14 24 34 05 15 25 35
+  ww1 = _mm256_unpacklo_epi32(w2, w3);  // 44 54 64 74 45 55 65 75
+
+  d[4] = _mm256_unpacklo_epi64(ww0, ww1);  // 04 14 24 34 44 54 64 74
+  d[5] = _mm256_unpackhi_epi64(ww0, ww1);  // 05 15 25 35 45 55 65 75
+
+  ww0 = _mm256_unpackhi_epi32(w0, w1);  // 06 16 26 36 07 17 27 37
+  ww1 = _mm256_unpackhi_epi32(w2, w3);  // 46 56 66 76 47 57 67 77
+
+  d[6] = _mm256_unpacklo_epi64(ww0, ww1);  // 06 16 26 36 46 56 66 76
+  d[7] = _mm256_unpackhi_epi64(ww0, ww1);  // 07 17 27 37 47 57 67 77
+}
+
+static INLINE void highbd_transpose16x16_avx2(__m256i *x, __m256i *d) {
+  __m256i w0, w1, w2, w3, ww0, ww1;
+  __m256i dd[16];
+  w0 = _mm256_unpacklo_epi16(x[0], x[1]);
+  w1 = _mm256_unpacklo_epi16(x[2], x[3]);
+  w2 = _mm256_unpacklo_epi16(x[4], x[5]);
+  w3 = _mm256_unpacklo_epi16(x[6], x[7]);
+
+  ww0 = _mm256_unpacklo_epi32(w0, w1);  //
+  ww1 = _mm256_unpacklo_epi32(w2, w3);  //
+
+  dd[0] = _mm256_unpacklo_epi64(ww0, ww1);
+  dd[1] = _mm256_unpackhi_epi64(ww0, ww1);
+
+  ww0 = _mm256_unpackhi_epi32(w0, w1);  //
+  ww1 = _mm256_unpackhi_epi32(w2, w3);  //
+
+  dd[2] = _mm256_unpacklo_epi64(ww0, ww1);
+  dd[3] = _mm256_unpackhi_epi64(ww0, ww1);
+
+  w0 = _mm256_unpackhi_epi16(x[0], x[1]);
+  w1 = _mm256_unpackhi_epi16(x[2], x[3]);
+  w2 = _mm256_unpackhi_epi16(x[4], x[5]);
+  w3 = _mm256_unpackhi_epi16(x[6], x[7]);
+
+  ww0 = _mm256_unpacklo_epi32(w0, w1);  //
+  ww1 = _mm256_unpacklo_epi32(w2, w3);  //
+
+  dd[4] = _mm256_unpacklo_epi64(ww0, ww1);
+  dd[5] = _mm256_unpackhi_epi64(ww0, ww1);
+
+  ww0 = _mm256_unpackhi_epi32(w0, w1);  //
+  ww1 = _mm256_unpackhi_epi32(w2, w3);  //
+
+  dd[6] = _mm256_unpacklo_epi64(ww0, ww1);
+  dd[7] = _mm256_unpackhi_epi64(ww0, ww1);
+
+  w0 = _mm256_unpacklo_epi16(x[8], x[9]);
+  w1 = _mm256_unpacklo_epi16(x[10], x[11]);
+  w2 = _mm256_unpacklo_epi16(x[12], x[13]);
+  w3 = _mm256_unpacklo_epi16(x[14], x[15]);
+
+  ww0 = _mm256_unpacklo_epi32(w0, w1);
+  ww1 = _mm256_unpacklo_epi32(w2, w3);
+
+  dd[8] = _mm256_unpacklo_epi64(ww0, ww1);
+  dd[9] = _mm256_unpackhi_epi64(ww0, ww1);
+
+  ww0 = _mm256_unpackhi_epi32(w0, w1);
+  ww1 = _mm256_unpackhi_epi32(w2, w3);
+
+  dd[10] = _mm256_unpacklo_epi64(ww0, ww1);
+  dd[11] = _mm256_unpackhi_epi64(ww0, ww1);
+
+  w0 = _mm256_unpackhi_epi16(x[8], x[9]);
+  w1 = _mm256_unpackhi_epi16(x[10], x[11]);
+  w2 = _mm256_unpackhi_epi16(x[12], x[13]);
+  w3 = _mm256_unpackhi_epi16(x[14], x[15]);
+
+  ww0 = _mm256_unpacklo_epi32(w0, w1);
+  ww1 = _mm256_unpacklo_epi32(w2, w3);
+
+  dd[12] = _mm256_unpacklo_epi64(ww0, ww1);
+  dd[13] = _mm256_unpackhi_epi64(ww0, ww1);
+
+  ww0 = _mm256_unpackhi_epi32(w0, w1);
+  ww1 = _mm256_unpackhi_epi32(w2, w3);
+
+  dd[14] = _mm256_unpacklo_epi64(ww0, ww1);
+  dd[15] = _mm256_unpackhi_epi64(ww0, ww1);
+
+  for (int i = 0; i < 8; i++) {
+    d[i] = _mm256_insertf128_si256(dd[i], _mm256_castsi256_si128(dd[i + 8]), 1);
+    d[i + 8] = _mm256_insertf128_si256(dd[i + 8],
+                                       _mm256_extracti128_si256(dd[i], 1), 0);
+  }
+}
+
+void aom_dc_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t stride,
+                                 const uint8_t *above, const uint8_t *left) {
+  const __m256i sum_above = dc_sum_32(above);
+  __m256i sum_left = dc_sum_32(left);
+  sum_left = _mm256_add_epi16(sum_left, sum_above);
+  const __m256i thirtytwo = _mm256_set1_epi16(32);
+  sum_left = _mm256_add_epi16(sum_left, thirtytwo);
+  sum_left = _mm256_srai_epi16(sum_left, 6);
+  const __m256i zero = _mm256_setzero_si256();
+  __m256i row = _mm256_shuffle_epi8(sum_left, zero);
+  row_store_32xh(&row, 32, dst, stride);
+}
+
+void aom_dc_top_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  __m256i sum = dc_sum_32(above);
+  (void)left;
+
+  const __m256i sixteen = _mm256_set1_epi16(16);
+  sum = _mm256_add_epi16(sum, sixteen);
+  sum = _mm256_srai_epi16(sum, 5);
+  const __m256i zero = _mm256_setzero_si256();
+  __m256i row = _mm256_shuffle_epi8(sum, zero);
+  row_store_32xh(&row, 32, dst, stride);
+}
+
+void aom_dc_left_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t stride,
+                                      const uint8_t *above,
+                                      const uint8_t *left) {
+  __m256i sum = dc_sum_32(left);
+  (void)above;
+
+  const __m256i sixteen = _mm256_set1_epi16(16);
+  sum = _mm256_add_epi16(sum, sixteen);
+  sum = _mm256_srai_epi16(sum, 5);
+  const __m256i zero = _mm256_setzero_si256();
+  __m256i row = _mm256_shuffle_epi8(sum, zero);
+  row_store_32xh(&row, 32, dst, stride);
+}
+
+void aom_dc_128_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)above;
+  (void)left;
+  const __m256i row = _mm256_set1_epi8((int8_t)0x80);
+  row_store_32xh(&row, 32, dst, stride);
+}
+
+void aom_v_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  const __m256i row = _mm256_loadu_si256((const __m256i *)above);
+  (void)left;
+  row_store_32xh(&row, 32, dst, stride);
+}
+
+// There are 32 rows togeter. This function does line:
+// 0,1,2,3, and 16,17,18,19. The next call would do
+// 4,5,6,7, and 20,21,22,23. So 4 times of calling
+// would finish 32 rows.
+static INLINE void h_predictor_32x8line(const __m256i *row, uint8_t *dst,
+                                        ptrdiff_t stride) {
+  __m256i t[4];
+  __m256i m = _mm256_setzero_si256();
+  const __m256i inc = _mm256_set1_epi8(4);
+  int i;
+
+  for (i = 0; i < 4; i++) {
+    t[i] = _mm256_shuffle_epi8(*row, m);
+    __m256i r0 = _mm256_permute2x128_si256(t[i], t[i], 0);
+    __m256i r1 = _mm256_permute2x128_si256(t[i], t[i], 0x11);
+    _mm256_storeu_si256((__m256i *)dst, r0);
+    _mm256_storeu_si256((__m256i *)(dst + (stride << 4)), r1);
+    dst += stride;
+    m = _mm256_add_epi8(m, inc);
+  }
+}
+
+void aom_h_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  const __m256i left_col = _mm256_loadu_si256((__m256i const *)left);
+
+  __m256i u = _mm256_unpacklo_epi8(left_col, left_col);
+
+  __m256i v = _mm256_unpacklo_epi8(u, u);
+  h_predictor_32x8line(&v, dst, stride);
+  dst += stride << 2;
+
+  v = _mm256_unpackhi_epi8(u, u);
+  h_predictor_32x8line(&v, dst, stride);
+  dst += stride << 2;
+
+  u = _mm256_unpackhi_epi8(left_col, left_col);
+
+  v = _mm256_unpacklo_epi8(u, u);
+  h_predictor_32x8line(&v, dst, stride);
+  dst += stride << 2;
+
+  v = _mm256_unpackhi_epi8(u, u);
+  h_predictor_32x8line(&v, dst, stride);
+}
+
+// -----------------------------------------------------------------------------
+// Rectangle
+void aom_dc_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t stride,
+                                 const uint8_t *above, const uint8_t *left) {
+  const __m128i top_sum = dc_sum_32_sse2(above);
+  __m128i left_sum = dc_sum_16_sse2(left);
+  left_sum = _mm_add_epi16(top_sum, left_sum);
+  uint16_t sum = (uint16_t)_mm_cvtsi128_si32(left_sum);
+  sum += 24;
+  sum /= 48;
+  const __m256i row = _mm256_set1_epi8((int8_t)sum);
+  row_store_32xh(&row, 16, dst, stride);
+}
+
+void aom_dc_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t stride,
+                                 const uint8_t *above, const uint8_t *left) {
+  const __m256i sum_above = dc_sum_32(above);
+  __m256i sum_left = dc_sum_64(left);
+  sum_left = _mm256_add_epi16(sum_left, sum_above);
+  uint16_t sum = (uint16_t)_mm_cvtsi128_si32(_mm256_castsi256_si128(sum_left));
+  sum += 48;
+  sum /= 96;
+  const __m256i row = _mm256_set1_epi8((int8_t)sum);
+  row_store_32xh(&row, 64, dst, stride);
+}
+
+void aom_dc_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t stride,
+                                 const uint8_t *above, const uint8_t *left) {
+  const __m256i sum_above = dc_sum_64(above);
+  __m256i sum_left = dc_sum_64(left);
+  sum_left = _mm256_add_epi16(sum_left, sum_above);
+  uint16_t sum = (uint16_t)_mm_cvtsi128_si32(_mm256_castsi256_si128(sum_left));
+  sum += 64;
+  sum /= 128;
+  const __m256i row = _mm256_set1_epi8((int8_t)sum);
+  row_store_64xh(&row, 64, dst, stride);
+}
+
+void aom_dc_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t stride,
+                                 const uint8_t *above, const uint8_t *left) {
+  const __m256i sum_above = dc_sum_64(above);
+  __m256i sum_left = dc_sum_32(left);
+  sum_left = _mm256_add_epi16(sum_left, sum_above);
+  uint16_t sum = (uint16_t)_mm_cvtsi128_si32(_mm256_castsi256_si128(sum_left));
+  sum += 48;
+  sum /= 96;
+  const __m256i row = _mm256_set1_epi8((int8_t)sum);
+  row_store_64xh(&row, 32, dst, stride);
+}
+
+void aom_dc_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t stride,
+                                 const uint8_t *above, const uint8_t *left) {
+  const __m256i sum_above = dc_sum_64(above);
+  __m256i sum_left = _mm256_castsi128_si256(dc_sum_16_sse2(left));
+  sum_left = _mm256_add_epi16(sum_left, sum_above);
+  uint16_t sum = (uint16_t)_mm_cvtsi128_si32(_mm256_castsi256_si128(sum_left));
+  sum += 40;
+  sum /= 80;
+  const __m256i row = _mm256_set1_epi8((int8_t)sum);
+  row_store_64xh(&row, 16, dst, stride);
+}
+
+void aom_dc_top_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  __m256i sum = dc_sum_32(above);
+  (void)left;
+
+  const __m256i sixteen = _mm256_set1_epi16(16);
+  sum = _mm256_add_epi16(sum, sixteen);
+  sum = _mm256_srai_epi16(sum, 5);
+  const __m256i zero = _mm256_setzero_si256();
+  __m256i row = _mm256_shuffle_epi8(sum, zero);
+  row_store_32xh(&row, 16, dst, stride);
+}
+
+void aom_dc_top_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  __m256i sum = dc_sum_32(above);
+  (void)left;
+
+  const __m256i sixteen = _mm256_set1_epi16(16);
+  sum = _mm256_add_epi16(sum, sixteen);
+  sum = _mm256_srai_epi16(sum, 5);
+  const __m256i zero = _mm256_setzero_si256();
+  __m256i row = _mm256_shuffle_epi8(sum, zero);
+  row_store_32xh(&row, 64, dst, stride);
+}
+
+void aom_dc_top_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  __m256i sum = dc_sum_64(above);
+  (void)left;
+
+  const __m256i thirtytwo = _mm256_set1_epi16(32);
+  sum = _mm256_add_epi16(sum, thirtytwo);
+  sum = _mm256_srai_epi16(sum, 6);
+  const __m256i zero = _mm256_setzero_si256();
+  __m256i row = _mm256_shuffle_epi8(sum, zero);
+  row_store_64xh(&row, 64, dst, stride);
+}
+
+void aom_dc_top_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  __m256i sum = dc_sum_64(above);
+  (void)left;
+
+  const __m256i thirtytwo = _mm256_set1_epi16(32);
+  sum = _mm256_add_epi16(sum, thirtytwo);
+  sum = _mm256_srai_epi16(sum, 6);
+  const __m256i zero = _mm256_setzero_si256();
+  __m256i row = _mm256_shuffle_epi8(sum, zero);
+  row_store_64xh(&row, 32, dst, stride);
+}
+
+void aom_dc_top_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  __m256i sum = dc_sum_64(above);
+  (void)left;
+
+  const __m256i thirtytwo = _mm256_set1_epi16(32);
+  sum = _mm256_add_epi16(sum, thirtytwo);
+  sum = _mm256_srai_epi16(sum, 6);
+  const __m256i zero = _mm256_setzero_si256();
+  __m256i row = _mm256_shuffle_epi8(sum, zero);
+  row_store_64xh(&row, 16, dst, stride);
+}
+
+void aom_dc_left_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t stride,
+                                      const uint8_t *above,
+                                      const uint8_t *left) {
+  __m128i sum = dc_sum_16_sse2(left);
+  (void)above;
+
+  const __m128i eight = _mm_set1_epi16(8);
+  sum = _mm_add_epi16(sum, eight);
+  sum = _mm_srai_epi16(sum, 4);
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i r = _mm_shuffle_epi8(sum, zero);
+  const __m256i row = _mm256_inserti128_si256(_mm256_castsi128_si256(r), r, 1);
+  row_store_32xh(&row, 16, dst, stride);
+}
+
+void aom_dc_left_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t stride,
+                                      const uint8_t *above,
+                                      const uint8_t *left) {
+  __m256i sum = dc_sum_64(left);
+  (void)above;
+
+  const __m256i thirtytwo = _mm256_set1_epi16(32);
+  sum = _mm256_add_epi16(sum, thirtytwo);
+  sum = _mm256_srai_epi16(sum, 6);
+  const __m256i zero = _mm256_setzero_si256();
+  __m256i row = _mm256_shuffle_epi8(sum, zero);
+  row_store_32xh(&row, 64, dst, stride);
+}
+
+void aom_dc_left_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t stride,
+                                      const uint8_t *above,
+                                      const uint8_t *left) {
+  __m256i sum = dc_sum_64(left);
+  (void)above;
+
+  const __m256i thirtytwo = _mm256_set1_epi16(32);
+  sum = _mm256_add_epi16(sum, thirtytwo);
+  sum = _mm256_srai_epi16(sum, 6);
+  const __m256i zero = _mm256_setzero_si256();
+  __m256i row = _mm256_shuffle_epi8(sum, zero);
+  row_store_64xh(&row, 64, dst, stride);
+}
+
+void aom_dc_left_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t stride,
+                                      const uint8_t *above,
+                                      const uint8_t *left) {
+  __m256i sum = dc_sum_32(left);
+  (void)above;
+
+  const __m256i sixteen = _mm256_set1_epi16(16);
+  sum = _mm256_add_epi16(sum, sixteen);
+  sum = _mm256_srai_epi16(sum, 5);
+  const __m256i zero = _mm256_setzero_si256();
+  __m256i row = _mm256_shuffle_epi8(sum, zero);
+  row_store_64xh(&row, 32, dst, stride);
+}
+
+void aom_dc_left_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t stride,
+                                      const uint8_t *above,
+                                      const uint8_t *left) {
+  __m128i sum = dc_sum_16_sse2(left);
+  (void)above;
+
+  const __m128i eight = _mm_set1_epi16(8);
+  sum = _mm_add_epi16(sum, eight);
+  sum = _mm_srai_epi16(sum, 4);
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i r = _mm_shuffle_epi8(sum, zero);
+  const __m256i row = _mm256_inserti128_si256(_mm256_castsi128_si256(r), r, 1);
+  row_store_64xh(&row, 16, dst, stride);
+}
+
+void aom_dc_128_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)above;
+  (void)left;
+  const __m256i row = _mm256_set1_epi8((int8_t)0x80);
+  row_store_32xh(&row, 16, dst, stride);
+}
+
+void aom_dc_128_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)above;
+  (void)left;
+  const __m256i row = _mm256_set1_epi8((int8_t)0x80);
+  row_store_32xh(&row, 64, dst, stride);
+}
+
+void aom_dc_128_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)above;
+  (void)left;
+  const __m256i row = _mm256_set1_epi8((int8_t)0x80);
+  row_store_64xh(&row, 64, dst, stride);
+}
+
+void aom_dc_128_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)above;
+  (void)left;
+  const __m256i row = _mm256_set1_epi8((int8_t)0x80);
+  row_store_64xh(&row, 32, dst, stride);
+}
+
+void aom_dc_128_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)above;
+  (void)left;
+  const __m256i row = _mm256_set1_epi8((int8_t)0x80);
+  row_store_64xh(&row, 16, dst, stride);
+}
+
+void aom_v_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  const __m256i row = _mm256_loadu_si256((const __m256i *)above);
+  (void)left;
+  row_store_32xh(&row, 16, dst, stride);
+}
+
+void aom_v_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  const __m256i row = _mm256_loadu_si256((const __m256i *)above);
+  (void)left;
+  row_store_32xh(&row, 64, dst, stride);
+}
+
+void aom_v_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  const __m256i row0 = _mm256_loadu_si256((const __m256i *)above);
+  const __m256i row1 = _mm256_loadu_si256((const __m256i *)(above + 32));
+  (void)left;
+  row_store_32x2xh(&row0, &row1, 64, dst, stride);
+}
+
+void aom_v_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  const __m256i row0 = _mm256_loadu_si256((const __m256i *)above);
+  const __m256i row1 = _mm256_loadu_si256((const __m256i *)(above + 32));
+  (void)left;
+  row_store_32x2xh(&row0, &row1, 32, dst, stride);
+}
+
+void aom_v_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  const __m256i row0 = _mm256_loadu_si256((const __m256i *)above);
+  const __m256i row1 = _mm256_loadu_si256((const __m256i *)(above + 32));
+  (void)left;
+  row_store_32x2xh(&row0, &row1, 16, dst, stride);
+}
+
+// -----------------------------------------------------------------------------
+// PAETH_PRED
+
+// Return 16 16-bit pixels in one row (__m256i)
+static INLINE __m256i paeth_pred(const __m256i *left, const __m256i *top,
+                                 const __m256i *topleft) {
+  const __m256i base =
+      _mm256_sub_epi16(_mm256_add_epi16(*top, *left), *topleft);
+
+  __m256i pl = _mm256_abs_epi16(_mm256_sub_epi16(base, *left));
+  __m256i pt = _mm256_abs_epi16(_mm256_sub_epi16(base, *top));
+  __m256i ptl = _mm256_abs_epi16(_mm256_sub_epi16(base, *topleft));
+
+  __m256i mask1 = _mm256_cmpgt_epi16(pl, pt);
+  mask1 = _mm256_or_si256(mask1, _mm256_cmpgt_epi16(pl, ptl));
+  __m256i mask2 = _mm256_cmpgt_epi16(pt, ptl);
+
+  pl = _mm256_andnot_si256(mask1, *left);
+
+  ptl = _mm256_and_si256(mask2, *topleft);
+  pt = _mm256_andnot_si256(mask2, *top);
+  pt = _mm256_or_si256(pt, ptl);
+  pt = _mm256_and_si256(mask1, pt);
+
+  return _mm256_or_si256(pt, pl);
+}
+
+// Return 16 8-bit pixels in one row (__m128i)
+static INLINE __m128i paeth_16x1_pred(const __m256i *left, const __m256i *top,
+                                      const __m256i *topleft) {
+  const __m256i p0 = paeth_pred(left, top, topleft);
+  const __m256i p1 = _mm256_permute4x64_epi64(p0, 0xe);
+  const __m256i p = _mm256_packus_epi16(p0, p1);
+  return _mm256_castsi256_si128(p);
+}
+
+static INLINE __m256i get_top_vector(const uint8_t *above) {
+  const __m128i x = _mm_load_si128((const __m128i *)above);
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i t0 = _mm_unpacklo_epi8(x, zero);
+  const __m128i t1 = _mm_unpackhi_epi8(x, zero);
+  return _mm256_inserti128_si256(_mm256_castsi128_si256(t0), t1, 1);
+}
+
+void aom_paeth_predictor_16x8_avx2(uint8_t *dst, ptrdiff_t stride,
+                                   const uint8_t *above, const uint8_t *left) {
+  __m128i x = _mm_loadl_epi64((const __m128i *)left);
+  const __m256i l = _mm256_inserti128_si256(_mm256_castsi128_si256(x), x, 1);
+  const __m256i tl16 = _mm256_set1_epi16((int16_t)above[-1]);
+  __m256i rep = _mm256_set1_epi16((short)0x8000);
+  const __m256i one = _mm256_set1_epi16(1);
+  const __m256i top = get_top_vector(above);
+
+  int i;
+  for (i = 0; i < 8; ++i) {
+    const __m256i l16 = _mm256_shuffle_epi8(l, rep);
+    const __m128i row = paeth_16x1_pred(&l16, &top, &tl16);
+
+    _mm_store_si128((__m128i *)dst, row);
+    dst += stride;
+    rep = _mm256_add_epi16(rep, one);
+  }
+}
+
+static INLINE __m256i get_left_vector(const uint8_t *left) {
+  const __m128i x = _mm_load_si128((const __m128i *)left);
+  return _mm256_inserti128_si256(_mm256_castsi128_si256(x), x, 1);
+}
+
+void aom_paeth_predictor_16x16_avx2(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  const __m256i l = get_left_vector(left);
+  const __m256i tl16 = _mm256_set1_epi16((int16_t)above[-1]);
+  __m256i rep = _mm256_set1_epi16((short)0x8000);
+  const __m256i one = _mm256_set1_epi16(1);
+  const __m256i top = get_top_vector(above);
+
+  int i;
+  for (i = 0; i < 16; ++i) {
+    const __m256i l16 = _mm256_shuffle_epi8(l, rep);
+    const __m128i row = paeth_16x1_pred(&l16, &top, &tl16);
+
+    _mm_store_si128((__m128i *)dst, row);
+    dst += stride;
+    rep = _mm256_add_epi16(rep, one);
+  }
+}
+
+void aom_paeth_predictor_16x32_avx2(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  __m256i l = get_left_vector(left);
+  const __m256i tl16 = _mm256_set1_epi16((int16_t)above[-1]);
+  __m256i rep = _mm256_set1_epi16((short)0x8000);
+  const __m256i one = _mm256_set1_epi16(1);
+  const __m256i top = get_top_vector(above);
+
+  int i;
+  for (i = 0; i < 16; ++i) {
+    const __m256i l16 = _mm256_shuffle_epi8(l, rep);
+    const __m128i row = paeth_16x1_pred(&l16, &top, &tl16);
+
+    _mm_store_si128((__m128i *)dst, row);
+    dst += stride;
+    rep = _mm256_add_epi16(rep, one);
+  }
+
+  l = get_left_vector(left + 16);
+  rep = _mm256_set1_epi16((short)0x8000);
+  for (i = 0; i < 16; ++i) {
+    const __m256i l16 = _mm256_shuffle_epi8(l, rep);
+    const __m128i row = paeth_16x1_pred(&l16, &top, &tl16);
+
+    _mm_store_si128((__m128i *)dst, row);
+    dst += stride;
+    rep = _mm256_add_epi16(rep, one);
+  }
+}
+
+void aom_paeth_predictor_16x64_avx2(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  const __m256i tl16 = _mm256_set1_epi16((int16_t)above[-1]);
+  const __m256i one = _mm256_set1_epi16(1);
+  const __m256i top = get_top_vector(above);
+
+  for (int j = 0; j < 4; ++j) {
+    const __m256i l = get_left_vector(left + j * 16);
+    __m256i rep = _mm256_set1_epi16((short)0x8000);
+    for (int i = 0; i < 16; ++i) {
+      const __m256i l16 = _mm256_shuffle_epi8(l, rep);
+      const __m128i row = paeth_16x1_pred(&l16, &top, &tl16);
+
+      _mm_store_si128((__m128i *)dst, row);
+      dst += stride;
+      rep = _mm256_add_epi16(rep, one);
+    }
+  }
+}
+
+// Return 32 8-bit pixels in one row (__m256i)
+static INLINE __m256i paeth_32x1_pred(const __m256i *left, const __m256i *top0,
+                                      const __m256i *top1,
+                                      const __m256i *topleft) {
+  __m256i p0 = paeth_pred(left, top0, topleft);
+  __m256i p1 = _mm256_permute4x64_epi64(p0, 0xe);
+  const __m256i x0 = _mm256_packus_epi16(p0, p1);
+
+  p0 = paeth_pred(left, top1, topleft);
+  p1 = _mm256_permute4x64_epi64(p0, 0xe);
+  const __m256i x1 = _mm256_packus_epi16(p0, p1);
+
+  return _mm256_permute2x128_si256(x0, x1, 0x20);
+}
+
+void aom_paeth_predictor_32x16_avx2(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  const __m256i l = get_left_vector(left);
+  const __m256i t0 = get_top_vector(above);
+  const __m256i t1 = get_top_vector(above + 16);
+  const __m256i tl = _mm256_set1_epi16((int16_t)above[-1]);
+  __m256i rep = _mm256_set1_epi16((short)0x8000);
+  const __m256i one = _mm256_set1_epi16(1);
+
+  int i;
+  for (i = 0; i < 16; ++i) {
+    const __m256i l16 = _mm256_shuffle_epi8(l, rep);
+
+    const __m256i r = paeth_32x1_pred(&l16, &t0, &t1, &tl);
+
+    _mm256_storeu_si256((__m256i *)dst, r);
+
+    dst += stride;
+    rep = _mm256_add_epi16(rep, one);
+  }
+}
+
+void aom_paeth_predictor_32x32_avx2(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  __m256i l = get_left_vector(left);
+  const __m256i t0 = get_top_vector(above);
+  const __m256i t1 = get_top_vector(above + 16);
+  const __m256i tl = _mm256_set1_epi16((int16_t)above[-1]);
+  __m256i rep = _mm256_set1_epi16((short)0x8000);
+  const __m256i one = _mm256_set1_epi16(1);
+
+  int i;
+  for (i = 0; i < 16; ++i) {
+    const __m256i l16 = _mm256_shuffle_epi8(l, rep);
+
+    const __m128i r0 = paeth_16x1_pred(&l16, &t0, &tl);
+    const __m128i r1 = paeth_16x1_pred(&l16, &t1, &tl);
+
+    _mm_store_si128((__m128i *)dst, r0);
+    _mm_store_si128((__m128i *)(dst + 16), r1);
+
+    dst += stride;
+    rep = _mm256_add_epi16(rep, one);
+  }
+
+  l = get_left_vector(left + 16);
+  rep = _mm256_set1_epi16((short)0x8000);
+  for (i = 0; i < 16; ++i) {
+    const __m256i l16 = _mm256_shuffle_epi8(l, rep);
+
+    const __m128i r0 = paeth_16x1_pred(&l16, &t0, &tl);
+    const __m128i r1 = paeth_16x1_pred(&l16, &t1, &tl);
+
+    _mm_store_si128((__m128i *)dst, r0);
+    _mm_store_si128((__m128i *)(dst + 16), r1);
+
+    dst += stride;
+    rep = _mm256_add_epi16(rep, one);
+  }
+}
+
+void aom_paeth_predictor_32x64_avx2(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  const __m256i t0 = get_top_vector(above);
+  const __m256i t1 = get_top_vector(above + 16);
+  const __m256i tl = _mm256_set1_epi16((int16_t)above[-1]);
+  const __m256i one = _mm256_set1_epi16(1);
+
+  int i, j;
+  for (j = 0; j < 4; ++j) {
+    const __m256i l = get_left_vector(left + j * 16);
+    __m256i rep = _mm256_set1_epi16((short)0x8000);
+    for (i = 0; i < 16; ++i) {
+      const __m256i l16 = _mm256_shuffle_epi8(l, rep);
+
+      const __m128i r0 = paeth_16x1_pred(&l16, &t0, &tl);
+      const __m128i r1 = paeth_16x1_pred(&l16, &t1, &tl);
+
+      _mm_store_si128((__m128i *)dst, r0);
+      _mm_store_si128((__m128i *)(dst + 16), r1);
+
+      dst += stride;
+      rep = _mm256_add_epi16(rep, one);
+    }
+  }
+}
+
+void aom_paeth_predictor_64x32_avx2(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  const __m256i t0 = get_top_vector(above);
+  const __m256i t1 = get_top_vector(above + 16);
+  const __m256i t2 = get_top_vector(above + 32);
+  const __m256i t3 = get_top_vector(above + 48);
+  const __m256i tl = _mm256_set1_epi16((int16_t)above[-1]);
+  const __m256i one = _mm256_set1_epi16(1);
+
+  int i, j;
+  for (j = 0; j < 2; ++j) {
+    const __m256i l = get_left_vector(left + j * 16);
+    __m256i rep = _mm256_set1_epi16((short)0x8000);
+    for (i = 0; i < 16; ++i) {
+      const __m256i l16 = _mm256_shuffle_epi8(l, rep);
+
+      const __m128i r0 = paeth_16x1_pred(&l16, &t0, &tl);
+      const __m128i r1 = paeth_16x1_pred(&l16, &t1, &tl);
+      const __m128i r2 = paeth_16x1_pred(&l16, &t2, &tl);
+      const __m128i r3 = paeth_16x1_pred(&l16, &t3, &tl);
+
+      _mm_store_si128((__m128i *)dst, r0);
+      _mm_store_si128((__m128i *)(dst + 16), r1);
+      _mm_store_si128((__m128i *)(dst + 32), r2);
+      _mm_store_si128((__m128i *)(dst + 48), r3);
+
+      dst += stride;
+      rep = _mm256_add_epi16(rep, one);
+    }
+  }
+}
+
+void aom_paeth_predictor_64x64_avx2(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  const __m256i t0 = get_top_vector(above);
+  const __m256i t1 = get_top_vector(above + 16);
+  const __m256i t2 = get_top_vector(above + 32);
+  const __m256i t3 = get_top_vector(above + 48);
+  const __m256i tl = _mm256_set1_epi16((int16_t)above[-1]);
+  const __m256i one = _mm256_set1_epi16(1);
+
+  int i, j;
+  for (j = 0; j < 4; ++j) {
+    const __m256i l = get_left_vector(left + j * 16);
+    __m256i rep = _mm256_set1_epi16((short)0x8000);
+    for (i = 0; i < 16; ++i) {
+      const __m256i l16 = _mm256_shuffle_epi8(l, rep);
+
+      const __m128i r0 = paeth_16x1_pred(&l16, &t0, &tl);
+      const __m128i r1 = paeth_16x1_pred(&l16, &t1, &tl);
+      const __m128i r2 = paeth_16x1_pred(&l16, &t2, &tl);
+      const __m128i r3 = paeth_16x1_pred(&l16, &t3, &tl);
+
+      _mm_store_si128((__m128i *)dst, r0);
+      _mm_store_si128((__m128i *)(dst + 16), r1);
+      _mm_store_si128((__m128i *)(dst + 32), r2);
+      _mm_store_si128((__m128i *)(dst + 48), r3);
+
+      dst += stride;
+      rep = _mm256_add_epi16(rep, one);
+    }
+  }
+}
+
+void aom_paeth_predictor_64x16_avx2(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  const __m256i t0 = get_top_vector(above);
+  const __m256i t1 = get_top_vector(above + 16);
+  const __m256i t2 = get_top_vector(above + 32);
+  const __m256i t3 = get_top_vector(above + 48);
+  const __m256i tl = _mm256_set1_epi16((int16_t)above[-1]);
+  const __m256i one = _mm256_set1_epi16(1);
+
+  int i;
+  const __m256i l = get_left_vector(left);
+  __m256i rep = _mm256_set1_epi16((short)0x8000);
+  for (i = 0; i < 16; ++i) {
+    const __m256i l16 = _mm256_shuffle_epi8(l, rep);
+
+    const __m128i r0 = paeth_16x1_pred(&l16, &t0, &tl);
+    const __m128i r1 = paeth_16x1_pred(&l16, &t1, &tl);
+    const __m128i r2 = paeth_16x1_pred(&l16, &t2, &tl);
+    const __m128i r3 = paeth_16x1_pred(&l16, &t3, &tl);
+
+    _mm_store_si128((__m128i *)dst, r0);
+    _mm_store_si128((__m128i *)(dst + 16), r1);
+    _mm_store_si128((__m128i *)(dst + 32), r2);
+    _mm_store_si128((__m128i *)(dst + 48), r3);
+
+    dst += stride;
+    rep = _mm256_add_epi16(rep, one);
+  }
+}
+
+#define PERM4x64(c0, c1, c2, c3) c0 + (c1 << 2) + (c2 << 4) + (c3 << 6)
+#define PERM2x128(c0, c1) c0 + (c1 << 4)
+
+static AOM_FORCE_INLINE void highbd_dr_prediction_z1_4xN_internal_avx2(
+    int N, __m128i *dst, const uint16_t *above, int upsample_above, int dx) {
+  const int frac_bits = 6 - upsample_above;
+  const int max_base_x = ((N + 4) - 1) << upsample_above;
+
+  assert(dx > 0);
+  // pre-filter above pixels
+  // store in temp buffers:
+  //   above[x] * 32 + 16
+  //   above[x+1] - above[x]
+  // final pixels will be calculated as:
+  //   (above[x] * 32 + 16 + (above[x+1] - above[x]) * shift) >> 5
+  __m256i a0, a1, a32, a16;
+  __m256i diff, c3f;
+  __m128i a_mbase_x, max_base_x128, base_inc128, mask128;
+  __m128i a0_128, a1_128;
+  a16 = _mm256_set1_epi16(16);
+  a_mbase_x = _mm_set1_epi16(above[max_base_x]);
+  max_base_x128 = _mm_set1_epi16(max_base_x);
+  c3f = _mm256_set1_epi16(0x3f);
+
+  int x = dx;
+  for (int r = 0; r < N; r++) {
+    __m256i b, res, shift;
+    __m128i res1;
+
+    int base = x >> frac_bits;
+    if (base >= max_base_x) {
+      for (int i = r; i < N; ++i) {
+        dst[i] = a_mbase_x;  // save 4 values
+      }
+      return;
+    }
+
+    a0_128 = _mm_loadu_si128((__m128i *)(above + base));
+    a1_128 = _mm_loadu_si128((__m128i *)(above + base + 1));
+
+    if (upsample_above) {
+      a0_128 = _mm_shuffle_epi8(a0_128, *(__m128i *)HighbdEvenOddMaskx4[0]);
+      a1_128 = _mm_srli_si128(a0_128, 8);
+
+      base_inc128 = _mm_setr_epi16(base, base + 2, base + 4, base + 6, base + 8,
+                                   base + 10, base + 12, base + 14);
+      shift = _mm256_srli_epi16(
+          _mm256_and_si256(
+              _mm256_slli_epi16(_mm256_set1_epi16(x), upsample_above),
+              _mm256_set1_epi16(0x3f)),
+          1);
+    } else {
+      base_inc128 = _mm_setr_epi16(base, base + 1, base + 2, base + 3, base + 4,
+                                   base + 5, base + 6, base + 7);
+      shift = _mm256_srli_epi16(_mm256_and_si256(_mm256_set1_epi16(x), c3f), 1);
+    }
+    a0 = _mm256_castsi128_si256(a0_128);
+    a1 = _mm256_castsi128_si256(a1_128);
+    diff = _mm256_sub_epi16(a1, a0);   // a[x+1] - a[x]
+    a32 = _mm256_slli_epi16(a0, 5);    // a[x] * 32
+    a32 = _mm256_add_epi16(a32, a16);  // a[x] * 32 + 16
+
+    b = _mm256_mullo_epi16(diff, shift);
+    res = _mm256_add_epi16(a32, b);
+    res = _mm256_srli_epi16(res, 5);
+    res1 = _mm256_castsi256_si128(res);
+
+    mask128 = _mm_cmpgt_epi16(max_base_x128, base_inc128);
+    dst[r] = _mm_blendv_epi8(a_mbase_x, res1, mask128);
+    x += dx;
+  }
+}
+
+static AOM_FORCE_INLINE void highbd_dr_prediction_32bit_z1_4xN_internal_avx2(
+    int N, __m128i *dst, const uint16_t *above, int upsample_above, int dx) {
+  const int frac_bits = 6 - upsample_above;
+  const int max_base_x = ((N + 4) - 1) << upsample_above;
+
+  assert(dx > 0);
+  // pre-filter above pixels
+  // store in temp buffers:
+  //   above[x] * 32 + 16
+  //   above[x+1] - above[x]
+  // final pixels will be calculated as:
+  //   (above[x] * 32 + 16 + (above[x+1] - above[x]) * shift) >> 5
+  __m256i a0, a1, a32, a16;
+  __m256i diff;
+  __m128i a_mbase_x, max_base_x128, base_inc128, mask128;
+
+  a16 = _mm256_set1_epi32(16);
+  a_mbase_x = _mm_set1_epi16(above[max_base_x]);
+  max_base_x128 = _mm_set1_epi32(max_base_x);
+
+  int x = dx;
+  for (int r = 0; r < N; r++) {
+    __m256i b, res, shift;
+    __m128i res1;
+
+    int base = x >> frac_bits;
+    if (base >= max_base_x) {
+      for (int i = r; i < N; ++i) {
+        dst[i] = a_mbase_x;  // save 4 values
+      }
+      return;
+    }
+
+    a0 = _mm256_cvtepu16_epi32(_mm_loadu_si128((__m128i *)(above + base)));
+    a1 = _mm256_cvtepu16_epi32(_mm_loadu_si128((__m128i *)(above + base + 1)));
+
+    if (upsample_above) {
+      a0 = _mm256_permutevar8x32_epi32(
+          a0, _mm256_set_epi32(7, 5, 3, 1, 6, 4, 2, 0));
+      a1 = _mm256_castsi128_si256(_mm256_extracti128_si256(a0, 1));
+      base_inc128 = _mm_setr_epi32(base, base + 2, base + 4, base + 6);
+      shift = _mm256_srli_epi32(
+          _mm256_and_si256(
+              _mm256_slli_epi32(_mm256_set1_epi32(x), upsample_above),
+              _mm256_set1_epi32(0x3f)),
+          1);
+    } else {
+      base_inc128 = _mm_setr_epi32(base, base + 1, base + 2, base + 3);
+      shift = _mm256_srli_epi32(
+          _mm256_and_si256(_mm256_set1_epi32(x), _mm256_set1_epi32(0x3f)), 1);
+    }
+
+    diff = _mm256_sub_epi32(a1, a0);   // a[x+1] - a[x]
+    a32 = _mm256_slli_epi32(a0, 5);    // a[x] * 32
+    a32 = _mm256_add_epi32(a32, a16);  // a[x] * 32 + 16
+
+    b = _mm256_mullo_epi32(diff, shift);
+    res = _mm256_add_epi32(a32, b);
+    res = _mm256_srli_epi32(res, 5);
+
+    res1 = _mm256_castsi256_si128(res);
+    res1 = _mm_packus_epi32(res1, res1);
+
+    mask128 = _mm_cmpgt_epi32(max_base_x128, base_inc128);
+    mask128 = _mm_packs_epi32(mask128, mask128);  // goto 16 bit
+    dst[r] = _mm_blendv_epi8(a_mbase_x, res1, mask128);
+    x += dx;
+  }
+}
+
+static void highbd_dr_prediction_z1_4xN_avx2(int N, uint16_t *dst,
+                                             ptrdiff_t stride,
+                                             const uint16_t *above,
+                                             int upsample_above, int dx,
+                                             int bd) {
+  __m128i dstvec[16];
+  if (bd < 12) {
+    highbd_dr_prediction_z1_4xN_internal_avx2(N, dstvec, above, upsample_above,
+                                              dx);
+  } else {
+    highbd_dr_prediction_32bit_z1_4xN_internal_avx2(N, dstvec, above,
+                                                    upsample_above, dx);
+  }
+  for (int i = 0; i < N; i++) {
+    _mm_storel_epi64((__m128i *)(dst + stride * i), dstvec[i]);
+  }
+}
+
+static AOM_FORCE_INLINE void highbd_dr_prediction_32bit_z1_8xN_internal_avx2(
+    int N, __m128i *dst, const uint16_t *above, int upsample_above, int dx) {
+  const int frac_bits = 6 - upsample_above;
+  const int max_base_x = ((8 + N) - 1) << upsample_above;
+
+  assert(dx > 0);
+  // pre-filter above pixels
+  // store in temp buffers:
+  //   above[x] * 32 + 16
+  //   above[x+1] - above[x]
+  // final pixels will be calculated as:
+  //   (above[x] * 32 + 16 + (above[x+1] - above[x]) * shift) >> 5
+  __m256i a0, a1, a0_1, a1_1, a32, a16;
+  __m256i a_mbase_x, diff, max_base_x256, base_inc256, mask256;
+
+  a16 = _mm256_set1_epi32(16);
+  a_mbase_x = _mm256_set1_epi16(above[max_base_x]);
+  max_base_x256 = _mm256_set1_epi32(max_base_x);
+
+  int x = dx;
+  for (int r = 0; r < N; r++) {
+    __m256i b, res, res1, shift;
+
+    int base = x >> frac_bits;
+    if (base >= max_base_x) {
+      for (int i = r; i < N; ++i) {
+        dst[i] = _mm256_castsi256_si128(a_mbase_x);  // save 8 values
+      }
+      return;
+    }
+
+    a0 = _mm256_cvtepu16_epi32(_mm_loadu_si128((__m128i *)(above + base)));
+    a1 = _mm256_cvtepu16_epi32(_mm_loadu_si128((__m128i *)(above + base + 1)));
+
+    if (upsample_above) {
+      a0 = _mm256_permutevar8x32_epi32(
+          a0, _mm256_set_epi32(7, 5, 3, 1, 6, 4, 2, 0));
+      a1 = _mm256_castsi128_si256(_mm256_extracti128_si256(a0, 1));
+
+      a0_1 =
+          _mm256_cvtepu16_epi32(_mm_loadu_si128((__m128i *)(above + base + 8)));
+      a0_1 = _mm256_permutevar8x32_epi32(
+          a0_1, _mm256_set_epi32(7, 5, 3, 1, 6, 4, 2, 0));
+      a1_1 = _mm256_castsi128_si256(_mm256_extracti128_si256(a0_1, 1));
+
+      a0 = _mm256_inserti128_si256(a0, _mm256_castsi256_si128(a0_1), 1);
+      a1 = _mm256_inserti128_si256(a1, _mm256_castsi256_si128(a1_1), 1);
+      base_inc256 =
+          _mm256_setr_epi32(base, base + 2, base + 4, base + 6, base + 8,
+                            base + 10, base + 12, base + 14);
+      shift = _mm256_srli_epi32(
+          _mm256_and_si256(
+              _mm256_slli_epi32(_mm256_set1_epi32(x), upsample_above),
+              _mm256_set1_epi32(0x3f)),
+          1);
+    } else {
+      base_inc256 = _mm256_setr_epi32(base, base + 1, base + 2, base + 3,
+                                      base + 4, base + 5, base + 6, base + 7);
+      shift = _mm256_srli_epi32(
+          _mm256_and_si256(_mm256_set1_epi32(x), _mm256_set1_epi32(0x3f)), 1);
+    }
+
+    diff = _mm256_sub_epi32(a1, a0);   // a[x+1] - a[x]
+    a32 = _mm256_slli_epi32(a0, 5);    // a[x] * 32
+    a32 = _mm256_add_epi32(a32, a16);  // a[x] * 32 + 16
+
+    b = _mm256_mullo_epi32(diff, shift);
+    res = _mm256_add_epi32(a32, b);
+    res = _mm256_srli_epi32(res, 5);
+
+    res1 = _mm256_packus_epi32(
+        res, _mm256_castsi128_si256(_mm256_extracti128_si256(res, 1)));
+
+    mask256 = _mm256_cmpgt_epi32(max_base_x256, base_inc256);
+    mask256 = _mm256_packs_epi32(
+        mask256, _mm256_castsi128_si256(
+                     _mm256_extracti128_si256(mask256, 1)));  // goto 16 bit
+    res1 = _mm256_blendv_epi8(a_mbase_x, res1, mask256);
+    dst[r] = _mm256_castsi256_si128(res1);
+    x += dx;
+  }
+}
+
+static AOM_FORCE_INLINE void highbd_dr_prediction_z1_8xN_internal_avx2(
+    int N, __m128i *dst, const uint16_t *above, int upsample_above, int dx) {
+  const int frac_bits = 6 - upsample_above;
+  const int max_base_x = ((8 + N) - 1) << upsample_above;
+
+  assert(dx > 0);
+  // pre-filter above pixels
+  // store in temp buffers:
+  //   above[x] * 32 + 16
+  //   above[x+1] - above[x]
+  // final pixels will be calculated as:
+  //   (above[x] * 32 + 16 + (above[x+1] - above[x]) * shift) >> 5
+  __m256i a0, a1, a32, a16, c3f;
+  __m256i a_mbase_x, diff, max_base_x256, base_inc256, mask256;
+  __m128i a0_x128, a1_x128;
+
+  a16 = _mm256_set1_epi16(16);
+  a_mbase_x = _mm256_set1_epi16(above[max_base_x]);
+  max_base_x256 = _mm256_set1_epi16(max_base_x);
+  c3f = _mm256_set1_epi16(0x3f);
+
+  int x = dx;
+  for (int r = 0; r < N; r++) {
+    __m256i b, res, res1, shift;
+
+    int base = x >> frac_bits;
+    if (base >= max_base_x) {
+      for (int i = r; i < N; ++i) {
+        dst[i] = _mm256_castsi256_si128(a_mbase_x);  // save 8 values
+      }
+      return;
+    }
+
+    a0_x128 = _mm_loadu_si128((__m128i *)(above + base));
+    if (upsample_above) {
+      __m128i mask, atmp0, atmp1, atmp2, atmp3;
+      a1_x128 = _mm_loadu_si128((__m128i *)(above + base + 8));
+      atmp0 = _mm_shuffle_epi8(a0_x128, *(__m128i *)HighbdEvenOddMaskx[0]);
+      atmp1 = _mm_shuffle_epi8(a1_x128, *(__m128i *)HighbdEvenOddMaskx[0]);
+      atmp2 =
+          _mm_shuffle_epi8(a0_x128, *(__m128i *)(HighbdEvenOddMaskx[0] + 16));
+      atmp3 =
+          _mm_shuffle_epi8(a1_x128, *(__m128i *)(HighbdEvenOddMaskx[0] + 16));
+      mask =
+          _mm_cmpgt_epi8(*(__m128i *)HighbdEvenOddMaskx[0], _mm_set1_epi8(15));
+      a0_x128 = _mm_blendv_epi8(atmp0, atmp1, mask);
+      mask = _mm_cmpgt_epi8(*(__m128i *)(HighbdEvenOddMaskx[0] + 16),
+                            _mm_set1_epi8(15));
+      a1_x128 = _mm_blendv_epi8(atmp2, atmp3, mask);
+
+      base_inc256 = _mm256_setr_epi16(base, base + 2, base + 4, base + 6,
+                                      base + 8, base + 10, base + 12, base + 14,
+                                      0, 0, 0, 0, 0, 0, 0, 0);
+      shift = _mm256_srli_epi16(
+          _mm256_and_si256(
+              _mm256_slli_epi16(_mm256_set1_epi16(x), upsample_above), c3f),
+          1);
+    } else {
+      a1_x128 = _mm_loadu_si128((__m128i *)(above + base + 1));
+      base_inc256 = _mm256_setr_epi16(base, base + 1, base + 2, base + 3,
+                                      base + 4, base + 5, base + 6, base + 7, 0,
+                                      0, 0, 0, 0, 0, 0, 0);
+      shift = _mm256_srli_epi16(_mm256_and_si256(_mm256_set1_epi16(x), c3f), 1);
+    }
+    a0 = _mm256_castsi128_si256(a0_x128);
+    a1 = _mm256_castsi128_si256(a1_x128);
+
+    diff = _mm256_sub_epi16(a1, a0);   // a[x+1] - a[x]
+    a32 = _mm256_slli_epi16(a0, 5);    // a[x] * 32
+    a32 = _mm256_add_epi16(a32, a16);  // a[x] * 32 + 16
+
+    b = _mm256_mullo_epi16(diff, shift);
+    res = _mm256_add_epi16(a32, b);
+    res = _mm256_srli_epi16(res, 5);
+
+    mask256 = _mm256_cmpgt_epi16(max_base_x256, base_inc256);
+    res1 = _mm256_blendv_epi8(a_mbase_x, res, mask256);
+    dst[r] = _mm256_castsi256_si128(res1);
+    x += dx;
+  }
+}
+
+static void highbd_dr_prediction_z1_8xN_avx2(int N, uint16_t *dst,
+                                             ptrdiff_t stride,
+                                             const uint16_t *above,
+                                             int upsample_above, int dx,
+                                             int bd) {
+  __m128i dstvec[32];
+  if (bd < 12) {
+    highbd_dr_prediction_z1_8xN_internal_avx2(N, dstvec, above, upsample_above,
+                                              dx);
+  } else {
+    highbd_dr_prediction_32bit_z1_8xN_internal_avx2(N, dstvec, above,
+                                                    upsample_above, dx);
+  }
+  for (int i = 0; i < N; i++) {
+    _mm_storeu_si128((__m128i *)(dst + stride * i), dstvec[i]);
+  }
+}
+
+static AOM_FORCE_INLINE void highbd_dr_prediction_32bit_z1_16xN_internal_avx2(
+    int N, __m256i *dstvec, const uint16_t *above, int upsample_above, int dx) {
+  // here upsample_above is 0 by design of av1_use_intra_edge_upsample
+  (void)upsample_above;
+  const int frac_bits = 6;
+  const int max_base_x = ((16 + N) - 1);
+
+  // pre-filter above pixels
+  // store in temp buffers:
+  //   above[x] * 32 + 16
+  //   above[x+1] - above[x]
+  // final pixels will be calculated as:
+  //   (above[x] * 32 + 16 + (above[x+1] - above[x]) * shift) >> 5
+  __m256i a0, a0_1, a1, a1_1, a32, a16;
+  __m256i a_mbase_x, diff, max_base_x256, base_inc256, mask256;
+
+  a16 = _mm256_set1_epi32(16);
+  a_mbase_x = _mm256_set1_epi16(above[max_base_x]);
+  max_base_x256 = _mm256_set1_epi16(max_base_x);
+
+  int x = dx;
+  for (int r = 0; r < N; r++) {
+    __m256i b, res[2], res1;
+
+    int base = x >> frac_bits;
+    if (base >= max_base_x) {
+      for (int i = r; i < N; ++i) {
+        dstvec[i] = a_mbase_x;  // save 16 values
+      }
+      return;
+    }
+    __m256i shift = _mm256_srli_epi32(
+        _mm256_and_si256(_mm256_set1_epi32(x), _mm256_set1_epi32(0x3f)), 1);
+
+    a0 = _mm256_cvtepu16_epi32(_mm_loadu_si128((__m128i *)(above + base)));
+    a1 = _mm256_cvtepu16_epi32(_mm_loadu_si128((__m128i *)(above + base + 1)));
+
+    diff = _mm256_sub_epi32(a1, a0);   // a[x+1] - a[x]
+    a32 = _mm256_slli_epi32(a0, 5);    // a[x] * 32
+    a32 = _mm256_add_epi32(a32, a16);  // a[x] * 32 + 16
+    b = _mm256_mullo_epi32(diff, shift);
+
+    res[0] = _mm256_add_epi32(a32, b);
+    res[0] = _mm256_srli_epi32(res[0], 5);
+    res[0] = _mm256_packus_epi32(
+        res[0], _mm256_castsi128_si256(_mm256_extracti128_si256(res[0], 1)));
+
+    int mdif = max_base_x - base;
+    if (mdif > 8) {
+      a0_1 =
+          _mm256_cvtepu16_epi32(_mm_loadu_si128((__m128i *)(above + base + 8)));
+      a1_1 =
+          _mm256_cvtepu16_epi32(_mm_loadu_si128((__m128i *)(above + base + 9)));
+
+      diff = _mm256_sub_epi32(a1_1, a0_1);  // a[x+1] - a[x]
+      a32 = _mm256_slli_epi32(a0_1, 5);     // a[x] * 32
+      a32 = _mm256_add_epi32(a32, a16);     // a[x] * 32 + 16
+      b = _mm256_mullo_epi32(diff, shift);
+
+      res[1] = _mm256_add_epi32(a32, b);
+      res[1] = _mm256_srli_epi32(res[1], 5);
+      res[1] = _mm256_packus_epi32(
+          res[1], _mm256_castsi128_si256(_mm256_extracti128_si256(res[1], 1)));
+    } else {
+      res[1] = a_mbase_x;
+    }
+    res1 = _mm256_inserti128_si256(res[0], _mm256_castsi256_si128(res[1]),
+                                   1);  // 16 16bit values
+
+    base_inc256 = _mm256_setr_epi16(base, base + 1, base + 2, base + 3,
+                                    base + 4, base + 5, base + 6, base + 7,
+                                    base + 8, base + 9, base + 10, base + 11,
+                                    base + 12, base + 13, base + 14, base + 15);
+    mask256 = _mm256_cmpgt_epi16(max_base_x256, base_inc256);
+    dstvec[r] = _mm256_blendv_epi8(a_mbase_x, res1, mask256);
+    x += dx;
+  }
+}
+
+static AOM_FORCE_INLINE void highbd_dr_prediction_z1_16xN_internal_avx2(
+    int N, __m256i *dstvec, const uint16_t *above, int upsample_above, int dx) {
+  // here upsample_above is 0 by design of av1_use_intra_edge_upsample
+  (void)upsample_above;
+  const int frac_bits = 6;
+  const int max_base_x = ((16 + N) - 1);
+
+  // pre-filter above pixels
+  // store in temp buffers:
+  //   above[x] * 32 + 16
+  //   above[x+1] - above[x]
+  // final pixels will be calculated as:
+  //   (above[x] * 32 + 16 + (above[x+1] - above[x]) * shift) >> 5
+  __m256i a0, a1, a32, a16, c3f;
+  __m256i a_mbase_x, diff, max_base_x256, base_inc256, mask256;
+
+  a16 = _mm256_set1_epi16(16);
+  a_mbase_x = _mm256_set1_epi16(above[max_base_x]);
+  max_base_x256 = _mm256_set1_epi16(max_base_x);
+  c3f = _mm256_set1_epi16(0x3f);
+
+  int x = dx;
+  for (int r = 0; r < N; r++) {
+    __m256i b, res;
+
+    int base = x >> frac_bits;
+    if (base >= max_base_x) {
+      for (int i = r; i < N; ++i) {
+        dstvec[i] = a_mbase_x;  // save 16 values
+      }
+      return;
+    }
+    __m256i shift =
+        _mm256_srli_epi16(_mm256_and_si256(_mm256_set1_epi16(x), c3f), 1);
+
+    a0 = _mm256_loadu_si256((__m256i *)(above + base));
+    a1 = _mm256_loadu_si256((__m256i *)(above + base + 1));
+
+    diff = _mm256_sub_epi16(a1, a0);   // a[x+1] - a[x]
+    a32 = _mm256_slli_epi16(a0, 5);    // a[x] * 32
+    a32 = _mm256_add_epi16(a32, a16);  // a[x] * 32 + 16
+    b = _mm256_mullo_epi16(diff, shift);
+
+    res = _mm256_add_epi16(a32, b);
+    res = _mm256_srli_epi16(res, 5);  // 16 16bit values
+
+    base_inc256 = _mm256_setr_epi16(base, base + 1, base + 2, base + 3,
+                                    base + 4, base + 5, base + 6, base + 7,
+                                    base + 8, base + 9, base + 10, base + 11,
+                                    base + 12, base + 13, base + 14, base + 15);
+    mask256 = _mm256_cmpgt_epi16(max_base_x256, base_inc256);
+    dstvec[r] = _mm256_blendv_epi8(a_mbase_x, res, mask256);
+    x += dx;
+  }
+}
+
+static void highbd_dr_prediction_z1_16xN_avx2(int N, uint16_t *dst,
+                                              ptrdiff_t stride,
+                                              const uint16_t *above,
+                                              int upsample_above, int dx,
+                                              int bd) {
+  __m256i dstvec[64];
+  if (bd < 12) {
+    highbd_dr_prediction_z1_16xN_internal_avx2(N, dstvec, above, upsample_above,
+                                               dx);
+  } else {
+    highbd_dr_prediction_32bit_z1_16xN_internal_avx2(N, dstvec, above,
+                                                     upsample_above, dx);
+  }
+  for (int i = 0; i < N; i++) {
+    _mm256_storeu_si256((__m256i *)(dst + stride * i), dstvec[i]);
+  }
+}
+
+static AOM_FORCE_INLINE void highbd_dr_prediction_32bit_z1_32xN_internal_avx2(
+    int N, __m256i *dstvec, const uint16_t *above, int upsample_above, int dx) {
+  // here upsample_above is 0 by design of av1_use_intra_edge_upsample
+  (void)upsample_above;
+  const int frac_bits = 6;
+  const int max_base_x = ((32 + N) - 1);
+
+  // pre-filter above pixels
+  // store in temp buffers:
+  //   above[x] * 32 + 16
+  //   above[x+1] - above[x]
+  // final pixels will be calculated as:
+  //   (above[x] * 32 + 16 + (above[x+1] - above[x]) * shift) >> 5
+  __m256i a0, a0_1, a1, a1_1, a32, a16, c3f;
+  __m256i a_mbase_x, diff, max_base_x256, base_inc256, mask256;
+
+  a16 = _mm256_set1_epi32(16);
+  a_mbase_x = _mm256_set1_epi16(above[max_base_x]);
+  max_base_x256 = _mm256_set1_epi16(max_base_x);
+  c3f = _mm256_set1_epi16(0x3f);
+
+  int x = dx;
+  for (int r = 0; r < N; r++) {
+    __m256i b, res[2], res1;
+
+    int base = x >> frac_bits;
+    if (base >= max_base_x) {
+      for (int i = r; i < N; ++i) {
+        dstvec[i] = a_mbase_x;  // save 32 values
+        dstvec[i + N] = a_mbase_x;
+      }
+      return;
+    }
+
+    __m256i shift =
+        _mm256_srli_epi32(_mm256_and_si256(_mm256_set1_epi32(x), c3f), 1);
+
+    for (int j = 0; j < 32; j += 16) {
+      int mdif = max_base_x - (base + j);
+      if (mdif <= 0) {
+        res1 = a_mbase_x;
+      } else {
+        a0 = _mm256_cvtepu16_epi32(
+            _mm_loadu_si128((__m128i *)(above + base + j)));
+        a1 = _mm256_cvtepu16_epi32(
+            _mm_loadu_si128((__m128i *)(above + base + 1 + j)));
+
+        diff = _mm256_sub_epi32(a1, a0);   // a[x+1] - a[x]
+        a32 = _mm256_slli_epi32(a0, 5);    // a[x] * 32
+        a32 = _mm256_add_epi32(a32, a16);  // a[x] * 32 + 16
+        b = _mm256_mullo_epi32(diff, shift);
+
+        res[0] = _mm256_add_epi32(a32, b);
+        res[0] = _mm256_srli_epi32(res[0], 5);
+        res[0] = _mm256_packus_epi32(
+            res[0],
+            _mm256_castsi128_si256(_mm256_extracti128_si256(res[0], 1)));
+        if (mdif > 8) {
+          a0_1 = _mm256_cvtepu16_epi32(
+              _mm_loadu_si128((__m128i *)(above + base + 8 + j)));
+          a1_1 = _mm256_cvtepu16_epi32(
+              _mm_loadu_si128((__m128i *)(above + base + 9 + j)));
+
+          diff = _mm256_sub_epi32(a1_1, a0_1);  // a[x+1] - a[x]
+          a32 = _mm256_slli_epi32(a0_1, 5);     // a[x] * 32
+          a32 = _mm256_add_epi32(a32, a16);     // a[x] * 32 + 16
+          b = _mm256_mullo_epi32(diff, shift);
+
+          res[1] = _mm256_add_epi32(a32, b);
+          res[1] = _mm256_srli_epi32(res[1], 5);
+          res[1] = _mm256_packus_epi32(
+              res[1],
+              _mm256_castsi128_si256(_mm256_extracti128_si256(res[1], 1)));
+        } else {
+          res[1] = a_mbase_x;
+        }
+        res1 = _mm256_inserti128_si256(res[0], _mm256_castsi256_si128(res[1]),
+                                       1);  // 16 16bit values
+        base_inc256 = _mm256_setr_epi16(
+            base + j, base + j + 1, base + j + 2, base + j + 3, base + j + 4,
+            base + j + 5, base + j + 6, base + j + 7, base + j + 8,
+            base + j + 9, base + j + 10, base + j + 11, base + j + 12,
+            base + j + 13, base + j + 14, base + j + 15);
+
+        mask256 = _mm256_cmpgt_epi16(max_base_x256, base_inc256);
+        res1 = _mm256_blendv_epi8(a_mbase_x, res1, mask256);
+      }
+      if (!j) {
+        dstvec[r] = res1;
+      } else {
+        dstvec[r + N] = res1;
+      }
+    }
+    x += dx;
+  }
+}
+
+static AOM_FORCE_INLINE void highbd_dr_prediction_z1_32xN_internal_avx2(
+    int N, __m256i *dstvec, const uint16_t *above, int upsample_above, int dx) {
+  // here upsample_above is 0 by design of av1_use_intra_edge_upsample
+  (void)upsample_above;
+  const int frac_bits = 6;
+  const int max_base_x = ((32 + N) - 1);
+
+  // pre-filter above pixels
+  // store in temp buffers:
+  //   above[x] * 32 + 16
+  //   above[x+1] - above[x]
+  // final pixels will be calculated as:
+  //   (above[x] * 32 + 16 + (above[x+1] - above[x]) * shift) >> 5
+  __m256i a0, a1, a32, a16, c3f;
+  __m256i a_mbase_x, diff, max_base_x256, base_inc256, mask256;
+
+  a16 = _mm256_set1_epi16(16);
+  a_mbase_x = _mm256_set1_epi16(above[max_base_x]);
+  max_base_x256 = _mm256_set1_epi16(max_base_x);
+  c3f = _mm256_set1_epi16(0x3f);
+
+  int x = dx;
+  for (int r = 0; r < N; r++) {
+    __m256i b, res;
+
+    int base = x >> frac_bits;
+    if (base >= max_base_x) {
+      for (int i = r; i < N; ++i) {
+        dstvec[i] = a_mbase_x;  // save 32 values
+        dstvec[i + N] = a_mbase_x;
+      }
+      return;
+    }
+
+    __m256i shift =
+        _mm256_srli_epi16(_mm256_and_si256(_mm256_set1_epi16(x), c3f), 1);
+
+    for (int j = 0; j < 32; j += 16) {
+      int mdif = max_base_x - (base + j);
+      if (mdif <= 0) {
+        res = a_mbase_x;
+      } else {
+        a0 = _mm256_loadu_si256((__m256i *)(above + base + j));
+        a1 = _mm256_loadu_si256((__m256i *)(above + base + 1 + j));
+
+        diff = _mm256_sub_epi16(a1, a0);   // a[x+1] - a[x]
+        a32 = _mm256_slli_epi16(a0, 5);    // a[x] * 32
+        a32 = _mm256_add_epi16(a32, a16);  // a[x] * 32 + 16
+        b = _mm256_mullo_epi16(diff, shift);
+
+        res = _mm256_add_epi16(a32, b);
+        res = _mm256_srli_epi16(res, 5);
+
+        base_inc256 = _mm256_setr_epi16(
+            base + j, base + j + 1, base + j + 2, base + j + 3, base + j + 4,
+            base + j + 5, base + j + 6, base + j + 7, base + j + 8,
+            base + j + 9, base + j + 10, base + j + 11, base + j + 12,
+            base + j + 13, base + j + 14, base + j + 15);
+
+        mask256 = _mm256_cmpgt_epi16(max_base_x256, base_inc256);
+        res = _mm256_blendv_epi8(a_mbase_x, res, mask256);
+      }
+      if (!j) {
+        dstvec[r] = res;
+      } else {
+        dstvec[r + N] = res;
+      }
+    }
+    x += dx;
+  }
+}
+
+static void highbd_dr_prediction_z1_32xN_avx2(int N, uint16_t *dst,
+                                              ptrdiff_t stride,
+                                              const uint16_t *above,
+                                              int upsample_above, int dx,
+                                              int bd) {
+  __m256i dstvec[128];
+  if (bd < 12) {
+    highbd_dr_prediction_z1_32xN_internal_avx2(N, dstvec, above, upsample_above,
+                                               dx);
+  } else {
+    highbd_dr_prediction_32bit_z1_32xN_internal_avx2(N, dstvec, above,
+                                                     upsample_above, dx);
+  }
+  for (int i = 0; i < N; i++) {
+    _mm256_storeu_si256((__m256i *)(dst + stride * i), dstvec[i]);
+    _mm256_storeu_si256((__m256i *)(dst + stride * i + 16), dstvec[i + N]);
+  }
+}
+
+static void highbd_dr_prediction_32bit_z1_64xN_avx2(int N, uint16_t *dst,
+                                                    ptrdiff_t stride,
+                                                    const uint16_t *above,
+                                                    int upsample_above,
+                                                    int dx) {
+  // here upsample_above is 0 by design of av1_use_intra_edge_upsample
+  (void)upsample_above;
+  const int frac_bits = 6;
+  const int max_base_x = ((64 + N) - 1);
+
+  // pre-filter above pixels
+  // store in temp buffers:
+  //   above[x] * 32 + 16
+  //   above[x+1] - above[x]
+  // final pixels will be calculated as:
+  //   (above[x] * 32 + 16 + (above[x+1] - above[x]) * shift) >> 5
+  __m256i a0, a0_1, a1, a1_1, a32, a16;
+  __m256i a_mbase_x, diff, max_base_x256, base_inc256, mask256;
+
+  a16 = _mm256_set1_epi32(16);
+  a_mbase_x = _mm256_set1_epi16(above[max_base_x]);
+  max_base_x256 = _mm256_set1_epi16(max_base_x);
+
+  int x = dx;
+  for (int r = 0; r < N; r++, dst += stride) {
+    __m256i b, res[2], res1;
+
+    int base = x >> frac_bits;
+    if (base >= max_base_x) {
+      for (int i = r; i < N; ++i) {
+        _mm256_storeu_si256((__m256i *)dst, a_mbase_x);  // save 32 values
+        _mm256_storeu_si256((__m256i *)(dst + 16), a_mbase_x);
+        _mm256_storeu_si256((__m256i *)(dst + 32), a_mbase_x);
+        _mm256_storeu_si256((__m256i *)(dst + 48), a_mbase_x);
+        dst += stride;
+      }
+      return;
+    }
+
+    __m256i shift = _mm256_srli_epi32(
+        _mm256_and_si256(_mm256_set1_epi32(x), _mm256_set1_epi32(0x3f)), 1);
+
+    __m128i a0_128, a0_1_128, a1_128, a1_1_128;
+    for (int j = 0; j < 64; j += 16) {
+      int mdif = max_base_x - (base + j);
+      if (mdif <= 0) {
+        _mm256_storeu_si256((__m256i *)(dst + j), a_mbase_x);
+      } else {
+        a0_128 = _mm_loadu_si128((__m128i *)(above + base + j));
+        a1_128 = _mm_loadu_si128((__m128i *)(above + base + 1 + j));
+        a0 = _mm256_cvtepu16_epi32(a0_128);
+        a1 = _mm256_cvtepu16_epi32(a1_128);
+
+        diff = _mm256_sub_epi32(a1, a0);   // a[x+1] - a[x]
+        a32 = _mm256_slli_epi32(a0, 5);    // a[x] * 32
+        a32 = _mm256_add_epi32(a32, a16);  // a[x] * 32 + 16
+        b = _mm256_mullo_epi32(diff, shift);
+
+        res[0] = _mm256_add_epi32(a32, b);
+        res[0] = _mm256_srli_epi32(res[0], 5);
+        res[0] = _mm256_packus_epi32(
+            res[0],
+            _mm256_castsi128_si256(_mm256_extracti128_si256(res[0], 1)));
+        if (mdif > 8) {
+          a0_1_128 = _mm_loadu_si128((__m128i *)(above + base + 8 + j));
+          a1_1_128 = _mm_loadu_si128((__m128i *)(above + base + 9 + j));
+          a0_1 = _mm256_cvtepu16_epi32(a0_1_128);
+          a1_1 = _mm256_cvtepu16_epi32(a1_1_128);
+
+          diff = _mm256_sub_epi32(a1_1, a0_1);  // a[x+1] - a[x]
+          a32 = _mm256_slli_epi32(a0_1, 5);     // a[x] * 32
+          a32 = _mm256_add_epi32(a32, a16);     // a[x] * 32 + 16
+          b = _mm256_mullo_epi32(diff, shift);
+
+          res[1] = _mm256_add_epi32(a32, b);
+          res[1] = _mm256_srli_epi32(res[1], 5);
+          res[1] = _mm256_packus_epi32(
+              res[1],
+              _mm256_castsi128_si256(_mm256_extracti128_si256(res[1], 1)));
+        } else {
+          res[1] = a_mbase_x;
+        }
+        res1 = _mm256_inserti128_si256(res[0], _mm256_castsi256_si128(res[1]),
+                                       1);  // 16 16bit values
+        base_inc256 = _mm256_setr_epi16(
+            base + j, base + j + 1, base + j + 2, base + j + 3, base + j + 4,
+            base + j + 5, base + j + 6, base + j + 7, base + j + 8,
+            base + j + 9, base + j + 10, base + j + 11, base + j + 12,
+            base + j + 13, base + j + 14, base + j + 15);
+
+        mask256 = _mm256_cmpgt_epi16(max_base_x256, base_inc256);
+        res1 = _mm256_blendv_epi8(a_mbase_x, res1, mask256);
+        _mm256_storeu_si256((__m256i *)(dst + j), res1);
+      }
+    }
+    x += dx;
+  }
+}
+
+static void highbd_dr_prediction_z1_64xN_avx2(int N, uint16_t *dst,
+                                              ptrdiff_t stride,
+                                              const uint16_t *above,
+                                              int upsample_above, int dx) {
+  // here upsample_above is 0 by design of av1_use_intra_edge_upsample
+  (void)upsample_above;
+  const int frac_bits = 6;
+  const int max_base_x = ((64 + N) - 1);
+
+  // pre-filter above pixels
+  // store in temp buffers:
+  //   above[x] * 32 + 16
+  //   above[x+1] - above[x]
+  // final pixels will be calculated as:
+  //   (above[x] * 32 + 16 + (above[x+1] - above[x]) * shift) >> 5
+  __m256i a0, a1, a32, a16, c3f;
+  __m256i a_mbase_x, diff, max_base_x256, base_inc256, mask256;
+
+  a16 = _mm256_set1_epi16(16);
+  a_mbase_x = _mm256_set1_epi16(above[max_base_x]);
+  max_base_x256 = _mm256_set1_epi16(max_base_x);
+  c3f = _mm256_set1_epi16(0x3f);
+
+  int x = dx;
+  for (int r = 0; r < N; r++, dst += stride) {
+    __m256i b, res;
+
+    int base = x >> frac_bits;
+    if (base >= max_base_x) {
+      for (int i = r; i < N; ++i) {
+        _mm256_storeu_si256((__m256i *)dst, a_mbase_x);  // save 32 values
+        _mm256_storeu_si256((__m256i *)(dst + 16), a_mbase_x);
+        _mm256_storeu_si256((__m256i *)(dst + 32), a_mbase_x);
+        _mm256_storeu_si256((__m256i *)(dst + 48), a_mbase_x);
+        dst += stride;
+      }
+      return;
+    }
+
+    __m256i shift =
+        _mm256_srli_epi16(_mm256_and_si256(_mm256_set1_epi16(x), c3f), 1);
+
+    for (int j = 0; j < 64; j += 16) {
+      int mdif = max_base_x - (base + j);
+      if (mdif <= 0) {
+        _mm256_storeu_si256((__m256i *)(dst + j), a_mbase_x);
+      } else {
+        a0 = _mm256_loadu_si256((__m256i *)(above + base + j));
+        a1 = _mm256_loadu_si256((__m256i *)(above + base + 1 + j));
+
+        diff = _mm256_sub_epi16(a1, a0);   // a[x+1] - a[x]
+        a32 = _mm256_slli_epi16(a0, 5);    // a[x] * 32
+        a32 = _mm256_add_epi16(a32, a16);  // a[x] * 32 + 16
+        b = _mm256_mullo_epi16(diff, shift);
+
+        res = _mm256_add_epi16(a32, b);
+        res = _mm256_srli_epi16(res, 5);
+
+        base_inc256 = _mm256_setr_epi16(
+            base + j, base + j + 1, base + j + 2, base + j + 3, base + j + 4,
+            base + j + 5, base + j + 6, base + j + 7, base + j + 8,
+            base + j + 9, base + j + 10, base + j + 11, base + j + 12,
+            base + j + 13, base + j + 14, base + j + 15);
+
+        mask256 = _mm256_cmpgt_epi16(max_base_x256, base_inc256);
+        res = _mm256_blendv_epi8(a_mbase_x, res, mask256);
+        _mm256_storeu_si256((__m256i *)(dst + j), res);  // 16 16bit values
+      }
+    }
+    x += dx;
+  }
+}
+
+// Directional prediction, zone 1: 0 < angle < 90
+void av1_highbd_dr_prediction_z1_avx2(uint16_t *dst, ptrdiff_t stride, int bw,
+                                      int bh, const uint16_t *above,
+                                      const uint16_t *left, int upsample_above,
+                                      int dx, int dy, int bd) {
+  (void)left;
+  (void)dy;
+
+  switch (bw) {
+    case 4:
+      highbd_dr_prediction_z1_4xN_avx2(bh, dst, stride, above, upsample_above,
+                                       dx, bd);
+      break;
+    case 8:
+      highbd_dr_prediction_z1_8xN_avx2(bh, dst, stride, above, upsample_above,
+                                       dx, bd);
+      break;
+    case 16:
+      highbd_dr_prediction_z1_16xN_avx2(bh, dst, stride, above, upsample_above,
+                                        dx, bd);
+      break;
+    case 32:
+      highbd_dr_prediction_z1_32xN_avx2(bh, dst, stride, above, upsample_above,
+                                        dx, bd);
+      break;
+    case 64:
+      if (bd < 12) {
+        highbd_dr_prediction_z1_64xN_avx2(bh, dst, stride, above,
+                                          upsample_above, dx);
+      } else {
+        highbd_dr_prediction_32bit_z1_64xN_avx2(bh, dst, stride, above,
+                                                upsample_above, dx);
+      }
+      break;
+    default: break;
+  }
+  return;
+}
+
+static void highbd_transpose_TX_16X16(const uint16_t *src, ptrdiff_t pitchSrc,
+                                      uint16_t *dst, ptrdiff_t pitchDst) {
+  __m256i r[16];
+  __m256i d[16];
+  for (int j = 0; j < 16; j++) {
+    r[j] = _mm256_loadu_si256((__m256i *)(src + j * pitchSrc));
+  }
+  highbd_transpose16x16_avx2(r, d);
+  for (int j = 0; j < 16; j++) {
+    _mm256_storeu_si256((__m256i *)(dst + j * pitchDst), d[j]);
+  }
+}
+
+static void highbd_transpose(const uint16_t *src, ptrdiff_t pitchSrc,
+                             uint16_t *dst, ptrdiff_t pitchDst, int width,
+                             int height) {
+  for (int j = 0; j < height; j += 16)
+    for (int i = 0; i < width; i += 16)
+      highbd_transpose_TX_16X16(src + i * pitchSrc + j, pitchSrc,
+                                dst + j * pitchDst + i, pitchDst);
+}
+
+static void highbd_dr_prediction_32bit_z2_Nx4_avx2(
+    int N, uint16_t *dst, ptrdiff_t stride, const uint16_t *above,
+    const uint16_t *left, int upsample_above, int upsample_left, int dx,
+    int dy) {
+  const int min_base_x = -(1 << upsample_above);
+  const int min_base_y = -(1 << upsample_left);
+  const int frac_bits_x = 6 - upsample_above;
+  const int frac_bits_y = 6 - upsample_left;
+
+  assert(dx > 0);
+  // pre-filter above pixels
+  // store in temp buffers:
+  //   above[x] * 32 + 16
+  //   above[x+1] - above[x]
+  // final pixels will be calculated as:
+  //   (above[x] * 32 + 16 + (above[x+1] - above[x]) * shift) >> 5
+  __m256i a0_x, a1_x, a32, a16;
+  __m256i diff;
+  __m128i c3f, min_base_y128;
+
+  a16 = _mm256_set1_epi32(16);
+  c3f = _mm_set1_epi32(0x3f);
+  min_base_y128 = _mm_set1_epi32(min_base_y);
+
+  for (int r = 0; r < N; r++) {
+    __m256i b, res, shift;
+    __m128i resx, resy, resxy;
+    __m128i a0_x128, a1_x128;
+    int y = r + 1;
+    int base_x = (-y * dx) >> frac_bits_x;
+    int base_shift = 0;
+    if (base_x < (min_base_x - 1)) {
+      base_shift = (min_base_x - base_x - 1) >> upsample_above;
+    }
+    int base_min_diff =
+        (min_base_x - base_x + upsample_above) >> upsample_above;
+    if (base_min_diff > 4) {
+      base_min_diff = 4;
+    } else {
+      if (base_min_diff < 0) base_min_diff = 0;
+    }
+
+    if (base_shift > 3) {
+      a0_x = _mm256_setzero_si256();
+      a1_x = _mm256_setzero_si256();
+      shift = _mm256_setzero_si256();
+    } else {
+      a0_x128 = _mm_loadu_si128((__m128i *)(above + base_x + base_shift));
+      if (upsample_above) {
+        a0_x128 = _mm_shuffle_epi8(a0_x128,
+                                   *(__m128i *)HighbdEvenOddMaskx4[base_shift]);
+        a1_x128 = _mm_srli_si128(a0_x128, 8);
+
+        shift = _mm256_castsi128_si256(_mm_srli_epi32(
+            _mm_and_si128(
+                _mm_slli_epi32(
+                    _mm_setr_epi32(-y * dx, (1 << 6) - y * dx,
+                                   (2 << 6) - y * dx, (3 << 6) - y * dx),
+                    upsample_above),
+                c3f),
+            1));
+      } else {
+        a0_x128 =
+            _mm_shuffle_epi8(a0_x128, *(__m128i *)HighbdLoadMaskx[base_shift]);
+        a1_x128 = _mm_srli_si128(a0_x128, 2);
+
+        shift = _mm256_castsi128_si256(_mm_srli_epi32(
+            _mm_and_si128(_mm_setr_epi32(-y * dx, (1 << 6) - y * dx,
+                                         (2 << 6) - y * dx, (3 << 6) - y * dx),
+                          c3f),
+            1));
+      }
+      a0_x = _mm256_cvtepu16_epi32(a0_x128);
+      a1_x = _mm256_cvtepu16_epi32(a1_x128);
+    }
+    // y calc
+    __m128i a0_y, a1_y, shifty;
+    if (base_x < min_base_x) {
+      __m128i r6, c1234, dy128, y_c128, base_y_c128, mask128;
+      DECLARE_ALIGNED(32, int, base_y_c[4]);
+      r6 = _mm_set1_epi32(r << 6);
+      dy128 = _mm_set1_epi32(dy);
+      c1234 = _mm_setr_epi32(1, 2, 3, 4);
+      y_c128 = _mm_sub_epi32(r6, _mm_mullo_epi32(c1234, dy128));
+      base_y_c128 = _mm_srai_epi32(y_c128, frac_bits_y);
+      mask128 = _mm_cmpgt_epi32(min_base_y128, base_y_c128);
+      base_y_c128 = _mm_andnot_si128(mask128, base_y_c128);
+      _mm_store_si128((__m128i *)base_y_c, base_y_c128);
+
+      a0_y = _mm_setr_epi32(left[base_y_c[0]], left[base_y_c[1]],
+                            left[base_y_c[2]], left[base_y_c[3]]);
+      a1_y = _mm_setr_epi32(left[base_y_c[0] + 1], left[base_y_c[1] + 1],
+                            left[base_y_c[2] + 1], left[base_y_c[3] + 1]);
+
+      if (upsample_left) {
+        shifty = _mm_srli_epi32(
+            _mm_and_si128(_mm_slli_epi32(y_c128, upsample_left), c3f), 1);
+      } else {
+        shifty = _mm_srli_epi32(_mm_and_si128(y_c128, c3f), 1);
+      }
+      a0_x = _mm256_inserti128_si256(a0_x, a0_y, 1);
+      a1_x = _mm256_inserti128_si256(a1_x, a1_y, 1);
+      shift = _mm256_inserti128_si256(shift, shifty, 1);
+    }
+
+    diff = _mm256_sub_epi32(a1_x, a0_x);  // a[x+1] - a[x]
+    a32 = _mm256_slli_epi32(a0_x, 5);     // a[x] * 32
+    a32 = _mm256_add_epi32(a32, a16);     // a[x] * 32 + 16
+
+    b = _mm256_mullo_epi32(diff, shift);
+    res = _mm256_add_epi32(a32, b);
+    res = _mm256_srli_epi32(res, 5);
+
+    resx = _mm256_castsi256_si128(res);
+    resx = _mm_packus_epi32(resx, resx);
+
+    resy = _mm256_extracti128_si256(res, 1);
+    resy = _mm_packus_epi32(resy, resy);
+
+    resxy =
+        _mm_blendv_epi8(resx, resy, *(__m128i *)HighbdBaseMask[base_min_diff]);
+    _mm_storel_epi64((__m128i *)(dst), resxy);
+    dst += stride;
+  }
+}
+
+static void highbd_dr_prediction_z2_Nx4_avx2(
+    int N, uint16_t *dst, ptrdiff_t stride, const uint16_t *above,
+    const uint16_t *left, int upsample_above, int upsample_left, int dx,
+    int dy) {
+  const int min_base_x = -(1 << upsample_above);
+  const int min_base_y = -(1 << upsample_left);
+  const int frac_bits_x = 6 - upsample_above;
+  const int frac_bits_y = 6 - upsample_left;
+
+  assert(dx > 0);
+  // pre-filter above pixels
+  // store in temp buffers:
+  //   above[x] * 32 + 16
+  //   above[x+1] - above[x]
+  // final pixels will be calculated as:
+  //   (above[x] * 32 + 16 + (above[x+1] - above[x]) * shift) >> 5
+  __m256i a0_x, a1_x, a32, a16;
+  __m256i diff;
+  __m128i c3f, min_base_y128;
+
+  a16 = _mm256_set1_epi16(16);
+  c3f = _mm_set1_epi16(0x3f);
+  min_base_y128 = _mm_set1_epi16(min_base_y);
+
+  for (int r = 0; r < N; r++) {
+    __m256i b, res, shift;
+    __m128i resx, resy, resxy;
+    __m128i a0_x128, a1_x128;
+    int y = r + 1;
+    int base_x = (-y * dx) >> frac_bits_x;
+    int base_shift = 0;
+    if (base_x < (min_base_x - 1)) {
+      base_shift = (min_base_x - base_x - 1) >> upsample_above;
+    }
+    int base_min_diff =
+        (min_base_x - base_x + upsample_above) >> upsample_above;
+    if (base_min_diff > 4) {
+      base_min_diff = 4;
+    } else {
+      if (base_min_diff < 0) base_min_diff = 0;
+    }
+
+    if (base_shift > 3) {
+      a0_x = _mm256_setzero_si256();
+      a1_x = _mm256_setzero_si256();
+      shift = _mm256_setzero_si256();
+    } else {
+      a0_x128 = _mm_loadu_si128((__m128i *)(above + base_x + base_shift));
+      if (upsample_above) {
+        a0_x128 = _mm_shuffle_epi8(a0_x128,
+                                   *(__m128i *)HighbdEvenOddMaskx4[base_shift]);
+        a1_x128 = _mm_srli_si128(a0_x128, 8);
+
+        shift = _mm256_castsi128_si256(_mm_srli_epi16(
+            _mm_and_si128(
+                _mm_slli_epi16(_mm_setr_epi16(-y * dx, (1 << 6) - y * dx,
+                                              (2 << 6) - y * dx,
+                                              (3 << 6) - y * dx, 0, 0, 0, 0),
+                               upsample_above),
+                c3f),
+            1));
+      } else {
+        a0_x128 =
+            _mm_shuffle_epi8(a0_x128, *(__m128i *)HighbdLoadMaskx[base_shift]);
+        a1_x128 = _mm_srli_si128(a0_x128, 2);
+
+        shift = _mm256_castsi128_si256(_mm_srli_epi16(
+            _mm_and_si128(
+                _mm_setr_epi16(-y * dx, (1 << 6) - y * dx, (2 << 6) - y * dx,
+                               (3 << 6) - y * dx, 0, 0, 0, 0),
+                c3f),
+            1));
+      }
+      a0_x = _mm256_castsi128_si256(a0_x128);
+      a1_x = _mm256_castsi128_si256(a1_x128);
+    }
+    // y calc
+    __m128i a0_y, a1_y, shifty;
+    if (base_x < min_base_x) {
+      __m128i r6, c1234, dy128, y_c128, base_y_c128, mask128;
+      DECLARE_ALIGNED(32, int16_t, base_y_c[8]);
+      r6 = _mm_set1_epi16(r << 6);
+      dy128 = _mm_set1_epi16(dy);
+      c1234 = _mm_setr_epi16(1, 2, 3, 4, 0, 0, 0, 0);
+      y_c128 = _mm_sub_epi16(r6, _mm_mullo_epi16(c1234, dy128));
+      base_y_c128 = _mm_srai_epi16(y_c128, frac_bits_y);
+      mask128 = _mm_cmpgt_epi16(min_base_y128, base_y_c128);
+      base_y_c128 = _mm_andnot_si128(mask128, base_y_c128);
+      _mm_store_si128((__m128i *)base_y_c, base_y_c128);
+
+      a0_y = _mm_setr_epi16(left[base_y_c[0]], left[base_y_c[1]],
+                            left[base_y_c[2]], left[base_y_c[3]], 0, 0, 0, 0);
+      a1_y = _mm_setr_epi16(left[base_y_c[0] + 1], left[base_y_c[1] + 1],
+                            left[base_y_c[2] + 1], left[base_y_c[3] + 1], 0, 0,
+                            0, 0);
+
+      if (upsample_left) {
+        shifty = _mm_srli_epi16(
+            _mm_and_si128(_mm_slli_epi16(y_c128, upsample_left), c3f), 1);
+      } else {
+        shifty = _mm_srli_epi16(_mm_and_si128(y_c128, c3f), 1);
+      }
+      a0_x = _mm256_inserti128_si256(a0_x, a0_y, 1);
+      a1_x = _mm256_inserti128_si256(a1_x, a1_y, 1);
+      shift = _mm256_inserti128_si256(shift, shifty, 1);
+    }
+
+    diff = _mm256_sub_epi16(a1_x, a0_x);  // a[x+1] - a[x]
+    a32 = _mm256_slli_epi16(a0_x, 5);     // a[x] * 32
+    a32 = _mm256_add_epi16(a32, a16);     // a[x] * 32 + 16
+
+    b = _mm256_mullo_epi16(diff, shift);
+    res = _mm256_add_epi16(a32, b);
+    res = _mm256_srli_epi16(res, 5);
+
+    resx = _mm256_castsi256_si128(res);
+    resy = _mm256_extracti128_si256(res, 1);
+    resxy =
+        _mm_blendv_epi8(resx, resy, *(__m128i *)HighbdBaseMask[base_min_diff]);
+    _mm_storel_epi64((__m128i *)(dst), resxy);
+    dst += stride;
+  }
+}
+
+static void highbd_dr_prediction_32bit_z2_Nx8_avx2(
+    int N, uint16_t *dst, ptrdiff_t stride, const uint16_t *above,
+    const uint16_t *left, int upsample_above, int upsample_left, int dx,
+    int dy) {
+  const int min_base_x = -(1 << upsample_above);
+  const int min_base_y = -(1 << upsample_left);
+  const int frac_bits_x = 6 - upsample_above;
+  const int frac_bits_y = 6 - upsample_left;
+
+  // pre-filter above pixels
+  // store in temp buffers:
+  //   above[x] * 32 + 16
+  //   above[x+1] - above[x]
+  // final pixels will be calculated as:
+  //   (above[x] * 32 + 16 + (above[x+1] - above[x]) * shift) >> 5
+  __m256i a0_x, a1_x, a0_y, a1_y, a32, a16, c3f, min_base_y256;
+  __m256i diff;
+  __m128i a0_x128, a1_x128;
+
+  a16 = _mm256_set1_epi32(16);
+  c3f = _mm256_set1_epi32(0x3f);
+  min_base_y256 = _mm256_set1_epi32(min_base_y);
+
+  for (int r = 0; r < N; r++) {
+    __m256i b, res, shift;
+    __m128i resx, resy, resxy;
+    int y = r + 1;
+    int base_x = (-y * dx) >> frac_bits_x;
+    int base_shift = 0;
+    if (base_x < (min_base_x - 1)) {
+      base_shift = (min_base_x - base_x - 1) >> upsample_above;
+    }
+    int base_min_diff =
+        (min_base_x - base_x + upsample_above) >> upsample_above;
+    if (base_min_diff > 8) {
+      base_min_diff = 8;
+    } else {
+      if (base_min_diff < 0) base_min_diff = 0;
+    }
+
+    if (base_shift > 7) {
+      resx = _mm_setzero_si128();
+    } else {
+      a0_x128 = _mm_loadu_si128((__m128i *)(above + base_x + base_shift));
+      if (upsample_above) {
+        __m128i mask, atmp0, atmp1, atmp2, atmp3;
+        a1_x128 = _mm_loadu_si128((__m128i *)(above + base_x + 8 + base_shift));
+        atmp0 = _mm_shuffle_epi8(a0_x128,
+                                 *(__m128i *)HighbdEvenOddMaskx[base_shift]);
+        atmp1 = _mm_shuffle_epi8(a1_x128,
+                                 *(__m128i *)HighbdEvenOddMaskx[base_shift]);
+        atmp2 = _mm_shuffle_epi8(
+            a0_x128, *(__m128i *)(HighbdEvenOddMaskx[base_shift] + 16));
+        atmp3 = _mm_shuffle_epi8(
+            a1_x128, *(__m128i *)(HighbdEvenOddMaskx[base_shift] + 16));
+        mask = _mm_cmpgt_epi8(*(__m128i *)HighbdEvenOddMaskx[base_shift],
+                              _mm_set1_epi8(15));
+        a0_x128 = _mm_blendv_epi8(atmp0, atmp1, mask);
+        mask = _mm_cmpgt_epi8(*(__m128i *)(HighbdEvenOddMaskx[base_shift] + 16),
+                              _mm_set1_epi8(15));
+        a1_x128 = _mm_blendv_epi8(atmp2, atmp3, mask);
+        shift = _mm256_srli_epi32(
+            _mm256_and_si256(
+                _mm256_slli_epi32(
+                    _mm256_setr_epi32(-y * dx, (1 << 6) - y * dx,
+                                      (2 << 6) - y * dx, (3 << 6) - y * dx,
+                                      (4 << 6) - y * dx, (5 << 6) - y * dx,
+                                      (6 << 6) - y * dx, (7 << 6) - y * dx),
+                    upsample_above),
+                c3f),
+            1);
+      } else {
+        a1_x128 = _mm_loadu_si128((__m128i *)(above + base_x + 1 + base_shift));
+        a0_x128 =
+            _mm_shuffle_epi8(a0_x128, *(__m128i *)HighbdLoadMaskx[base_shift]);
+        a1_x128 =
+            _mm_shuffle_epi8(a1_x128, *(__m128i *)HighbdLoadMaskx[base_shift]);
+
+        shift = _mm256_srli_epi32(
+            _mm256_and_si256(
+                _mm256_setr_epi32(-y * dx, (1 << 6) - y * dx, (2 << 6) - y * dx,
+                                  (3 << 6) - y * dx, (4 << 6) - y * dx,
+                                  (5 << 6) - y * dx, (6 << 6) - y * dx,
+                                  (7 << 6) - y * dx),
+                c3f),
+            1);
+      }
+      a0_x = _mm256_cvtepu16_epi32(a0_x128);
+      a1_x = _mm256_cvtepu16_epi32(a1_x128);
+
+      diff = _mm256_sub_epi32(a1_x, a0_x);  // a[x+1] - a[x]
+      a32 = _mm256_slli_epi32(a0_x, 5);     // a[x] * 32
+      a32 = _mm256_add_epi32(a32, a16);     // a[x] * 32 + 16
+
+      b = _mm256_mullo_epi32(diff, shift);
+      res = _mm256_add_epi32(a32, b);
+      res = _mm256_srli_epi32(res, 5);
+
+      resx = _mm256_castsi256_si128(_mm256_packus_epi32(
+          res, _mm256_castsi128_si256(_mm256_extracti128_si256(res, 1))));
+    }
+    // y calc
+    if (base_x < min_base_x) {
+      DECLARE_ALIGNED(32, int, base_y_c[8]);
+      __m256i r6, c256, dy256, y_c256, base_y_c256, mask256;
+      r6 = _mm256_set1_epi32(r << 6);
+      dy256 = _mm256_set1_epi32(dy);
+      c256 = _mm256_setr_epi32(1, 2, 3, 4, 5, 6, 7, 8);
+      y_c256 = _mm256_sub_epi32(r6, _mm256_mullo_epi32(c256, dy256));
+      base_y_c256 = _mm256_srai_epi32(y_c256, frac_bits_y);
+      mask256 = _mm256_cmpgt_epi32(min_base_y256, base_y_c256);
+      base_y_c256 = _mm256_andnot_si256(mask256, base_y_c256);
+      _mm256_store_si256((__m256i *)base_y_c, base_y_c256);
+
+      a0_y = _mm256_cvtepu16_epi32(_mm_setr_epi16(
+          left[base_y_c[0]], left[base_y_c[1]], left[base_y_c[2]],
+          left[base_y_c[3]], left[base_y_c[4]], left[base_y_c[5]],
+          left[base_y_c[6]], left[base_y_c[7]]));
+      a1_y = _mm256_cvtepu16_epi32(_mm_setr_epi16(
+          left[base_y_c[0] + 1], left[base_y_c[1] + 1], left[base_y_c[2] + 1],
+          left[base_y_c[3] + 1], left[base_y_c[4] + 1], left[base_y_c[5] + 1],
+          left[base_y_c[6] + 1], left[base_y_c[7] + 1]));
+
+      if (upsample_left) {
+        shift = _mm256_srli_epi32(
+            _mm256_and_si256(_mm256_slli_epi32((y_c256), upsample_left), c3f),
+            1);
+      } else {
+        shift = _mm256_srli_epi32(_mm256_and_si256(y_c256, c3f), 1);
+      }
+      diff = _mm256_sub_epi32(a1_y, a0_y);  // a[x+1] - a[x]
+      a32 = _mm256_slli_epi32(a0_y, 5);     // a[x] * 32
+      a32 = _mm256_add_epi32(a32, a16);     // a[x] * 32 + 16
+
+      b = _mm256_mullo_epi32(diff, shift);
+      res = _mm256_add_epi32(a32, b);
+      res = _mm256_srli_epi32(res, 5);
+
+      resy = _mm256_castsi256_si128(_mm256_packus_epi32(
+          res, _mm256_castsi128_si256(_mm256_extracti128_si256(res, 1))));
+    } else {
+      resy = resx;
+    }
+    resxy =
+        _mm_blendv_epi8(resx, resy, *(__m128i *)HighbdBaseMask[base_min_diff]);
+    _mm_storeu_si128((__m128i *)(dst), resxy);
+    dst += stride;
+  }
+}
+
+static void highbd_dr_prediction_z2_Nx8_avx2(
+    int N, uint16_t *dst, ptrdiff_t stride, const uint16_t *above,
+    const uint16_t *left, int upsample_above, int upsample_left, int dx,
+    int dy) {
+  const int min_base_x = -(1 << upsample_above);
+  const int min_base_y = -(1 << upsample_left);
+  const int frac_bits_x = 6 - upsample_above;
+  const int frac_bits_y = 6 - upsample_left;
+
+  // pre-filter above pixels
+  // store in temp buffers:
+  //   above[x] * 32 + 16
+  //   above[x+1] - above[x]
+  // final pixels will be calculated as:
+  //   (above[x] * 32 + 16 + (above[x+1] - above[x]) * shift) >> 5
+  __m128i c3f, min_base_y128;
+  __m256i a0_x, a1_x, diff, a32, a16;
+  __m128i a0_x128, a1_x128;
+
+  a16 = _mm256_set1_epi16(16);
+  c3f = _mm_set1_epi16(0x3f);
+  min_base_y128 = _mm_set1_epi16(min_base_y);
+
+  for (int r = 0; r < N; r++) {
+    __m256i b, res, shift;
+    __m128i resx, resy, resxy;
+    int y = r + 1;
+    int base_x = (-y * dx) >> frac_bits_x;
+    int base_shift = 0;
+    if (base_x < (min_base_x - 1)) {
+      base_shift = (min_base_x - base_x - 1) >> upsample_above;
+    }
+    int base_min_diff =
+        (min_base_x - base_x + upsample_above) >> upsample_above;
+    if (base_min_diff > 8) {
+      base_min_diff = 8;
+    } else {
+      if (base_min_diff < 0) base_min_diff = 0;
+    }
+
+    if (base_shift > 7) {
+      a0_x = _mm256_setzero_si256();
+      a1_x = _mm256_setzero_si256();
+      shift = _mm256_setzero_si256();
+    } else {
+      a0_x128 = _mm_loadu_si128((__m128i *)(above + base_x + base_shift));
+      if (upsample_above) {
+        __m128i mask, atmp0, atmp1, atmp2, atmp3;
+        a1_x128 = _mm_loadu_si128((__m128i *)(above + base_x + 8 + base_shift));
+        atmp0 = _mm_shuffle_epi8(a0_x128,
+                                 *(__m128i *)HighbdEvenOddMaskx[base_shift]);
+        atmp1 = _mm_shuffle_epi8(a1_x128,
+                                 *(__m128i *)HighbdEvenOddMaskx[base_shift]);
+        atmp2 = _mm_shuffle_epi8(
+            a0_x128, *(__m128i *)(HighbdEvenOddMaskx[base_shift] + 16));
+        atmp3 = _mm_shuffle_epi8(
+            a1_x128, *(__m128i *)(HighbdEvenOddMaskx[base_shift] + 16));
+        mask = _mm_cmpgt_epi8(*(__m128i *)HighbdEvenOddMaskx[base_shift],
+                              _mm_set1_epi8(15));
+        a0_x128 = _mm_blendv_epi8(atmp0, atmp1, mask);
+        mask = _mm_cmpgt_epi8(*(__m128i *)(HighbdEvenOddMaskx[base_shift] + 16),
+                              _mm_set1_epi8(15));
+        a1_x128 = _mm_blendv_epi8(atmp2, atmp3, mask);
+
+        shift = _mm256_castsi128_si256(_mm_srli_epi16(
+            _mm_and_si128(
+                _mm_slli_epi16(
+                    _mm_setr_epi16(-y * dx, (1 << 6) - y * dx,
+                                   (2 << 6) - y * dx, (3 << 6) - y * dx,
+                                   (4 << 6) - y * dx, (5 << 6) - y * dx,
+                                   (6 << 6) - y * dx, (7 << 6) - y * dx),
+                    upsample_above),
+                c3f),
+            1));
+      } else {
+        a1_x128 = _mm_loadu_si128((__m128i *)(above + base_x + 1 + base_shift));
+        a0_x128 =
+            _mm_shuffle_epi8(a0_x128, *(__m128i *)HighbdLoadMaskx[base_shift]);
+        a1_x128 =
+            _mm_shuffle_epi8(a1_x128, *(__m128i *)HighbdLoadMaskx[base_shift]);
+
+        shift = _mm256_castsi128_si256(_mm_srli_epi16(
+            _mm_and_si128(_mm_setr_epi16(-y * dx, (1 << 6) - y * dx,
+                                         (2 << 6) - y * dx, (3 << 6) - y * dx,
+                                         (4 << 6) - y * dx, (5 << 6) - y * dx,
+                                         (6 << 6) - y * dx, (7 << 6) - y * dx),
+                          c3f),
+            1));
+      }
+      a0_x = _mm256_castsi128_si256(a0_x128);
+      a1_x = _mm256_castsi128_si256(a1_x128);
+    }
+
+    // y calc
+    __m128i a0_y, a1_y, shifty;
+    if (base_x < min_base_x) {
+      DECLARE_ALIGNED(32, int16_t, base_y_c[8]);
+      __m128i r6, c1234, dy128, y_c128, base_y_c128, mask128;
+      r6 = _mm_set1_epi16(r << 6);
+      dy128 = _mm_set1_epi16(dy);
+      c1234 = _mm_setr_epi16(1, 2, 3, 4, 5, 6, 7, 8);
+      y_c128 = _mm_sub_epi16(r6, _mm_mullo_epi16(c1234, dy128));
+      base_y_c128 = _mm_srai_epi16(y_c128, frac_bits_y);
+      mask128 = _mm_cmpgt_epi16(min_base_y128, base_y_c128);
+      base_y_c128 = _mm_andnot_si128(mask128, base_y_c128);
+      _mm_store_si128((__m128i *)base_y_c, base_y_c128);
+
+      a0_y = _mm_setr_epi16(left[base_y_c[0]], left[base_y_c[1]],
+                            left[base_y_c[2]], left[base_y_c[3]],
+                            left[base_y_c[4]], left[base_y_c[5]],
+                            left[base_y_c[6]], left[base_y_c[7]]);
+      a1_y = _mm_setr_epi16(left[base_y_c[0] + 1], left[base_y_c[1] + 1],
+                            left[base_y_c[2] + 1], left[base_y_c[3] + 1],
+                            left[base_y_c[4] + 1], left[base_y_c[5] + 1],
+                            left[base_y_c[6] + 1], left[base_y_c[7] + 1]);
+
+      if (upsample_left) {
+        shifty = _mm_srli_epi16(
+            _mm_and_si128(_mm_slli_epi16((y_c128), upsample_left), c3f), 1);
+      } else {
+        shifty = _mm_srli_epi16(_mm_and_si128(y_c128, c3f), 1);
+      }
+      a0_x = _mm256_inserti128_si256(a0_x, a0_y, 1);
+      a1_x = _mm256_inserti128_si256(a1_x, a1_y, 1);
+      shift = _mm256_inserti128_si256(shift, shifty, 1);
+    }
+
+    diff = _mm256_sub_epi16(a1_x, a0_x);  // a[x+1] - a[x]
+    a32 = _mm256_slli_epi16(a0_x, 5);     // a[x] * 32
+    a32 = _mm256_add_epi16(a32, a16);     // a[x] * 32 + 16
+
+    b = _mm256_mullo_epi16(diff, shift);
+    res = _mm256_add_epi16(a32, b);
+    res = _mm256_srli_epi16(res, 5);
+
+    resx = _mm256_castsi256_si128(res);
+    resy = _mm256_extracti128_si256(res, 1);
+
+    resxy =
+        _mm_blendv_epi8(resx, resy, *(__m128i *)HighbdBaseMask[base_min_diff]);
+    _mm_storeu_si128((__m128i *)(dst), resxy);
+    dst += stride;
+  }
+}
+
+static void highbd_dr_prediction_32bit_z2_HxW_avx2(
+    int H, int W, uint16_t *dst, ptrdiff_t stride, const uint16_t *above,
+    const uint16_t *left, int upsample_above, int upsample_left, int dx,
+    int dy) {
+  // here upsample_above and upsample_left are 0 by design of
+  // av1_use_intra_edge_upsample
+  const int min_base_x = -1;
+  const int min_base_y = -1;
+  (void)upsample_above;
+  (void)upsample_left;
+  const int frac_bits_x = 6;
+  const int frac_bits_y = 6;
+
+  // pre-filter above pixels
+  // store in temp buffers:
+  //   above[x] * 32 + 16
+  //   above[x+1] - above[x]
+  // final pixels will be calculated as:
+  //   (above[x] * 32 + 16 + (above[x+1] - above[x]) * shift) >> 5
+  __m256i a0_x, a1_x, a0_y, a1_y, a32, a0_1_x, a1_1_x, a16, c1;
+  __m256i diff, min_base_y256, c3f, dy256, c1234, c0123, c8;
+  __m128i a0_x128, a1_x128, a0_1_x128, a1_1_x128;
+  DECLARE_ALIGNED(32, int, base_y_c[16]);
+
+  a16 = _mm256_set1_epi32(16);
+  c1 = _mm256_srli_epi32(a16, 4);
+  c8 = _mm256_srli_epi32(a16, 1);
+  min_base_y256 = _mm256_set1_epi32(min_base_y);
+  c3f = _mm256_set1_epi32(0x3f);
+  dy256 = _mm256_set1_epi32(dy);
+  c0123 = _mm256_setr_epi32(0, 1, 2, 3, 4, 5, 6, 7);
+  c1234 = _mm256_add_epi32(c0123, c1);
+
+  for (int r = 0; r < H; r++) {
+    __m256i b, res, shift, ydx;
+    __m256i resx[2], resy[2];
+    __m256i resxy, j256, r6;
+    for (int j = 0; j < W; j += 16) {
+      j256 = _mm256_set1_epi32(j);
+      int y = r + 1;
+      ydx = _mm256_set1_epi32(y * dx);
+
+      int base_x = ((j << 6) - y * dx) >> frac_bits_x;
+      int base_shift = 0;
+      if ((base_x) < (min_base_x - 1)) {
+        base_shift = (min_base_x - base_x - 1);
+      }
+      int base_min_diff = (min_base_x - base_x);
+      if (base_min_diff > 16) {
+        base_min_diff = 16;
+      } else {
+        if (base_min_diff < 0) base_min_diff = 0;
+      }
+
+      if (base_shift > 7) {
+        resx[0] = _mm256_setzero_si256();
+      } else {
+        a0_x128 = _mm_loadu_si128((__m128i *)(above + base_x + base_shift));
+        a1_x128 = _mm_loadu_si128((__m128i *)(above + base_x + base_shift + 1));
+        a0_x128 =
+            _mm_shuffle_epi8(a0_x128, *(__m128i *)HighbdLoadMaskx[base_shift]);
+        a1_x128 =
+            _mm_shuffle_epi8(a1_x128, *(__m128i *)HighbdLoadMaskx[base_shift]);
+
+        a0_x = _mm256_cvtepu16_epi32(a0_x128);
+        a1_x = _mm256_cvtepu16_epi32(a1_x128);
+
+        r6 = _mm256_slli_epi32(_mm256_add_epi32(c0123, j256), 6);
+        shift = _mm256_srli_epi32(
+            _mm256_and_si256(_mm256_sub_epi32(r6, ydx), c3f), 1);
+
+        diff = _mm256_sub_epi32(a1_x, a0_x);  // a[x+1] - a[x]
+        a32 = _mm256_slli_epi32(a0_x, 5);     // a[x] * 32
+        a32 = _mm256_add_epi32(a32, a16);     // a[x] * 32 + 16
+
+        b = _mm256_mullo_epi32(diff, shift);
+        res = _mm256_add_epi32(a32, b);
+        res = _mm256_srli_epi32(res, 5);
+
+        resx[0] = _mm256_packus_epi32(
+            res, _mm256_castsi128_si256(_mm256_extracti128_si256(res, 1)));
+      }
+      int base_shift8 = 0;
+      if ((base_x + 8) < (min_base_x - 1)) {
+        base_shift8 = (min_base_x - (base_x + 8) - 1);
+      }
+      if (base_shift8 > 7) {
+        resx[1] = _mm256_setzero_si256();
+      } else {
+        a0_1_x128 =
+            _mm_loadu_si128((__m128i *)(above + base_x + base_shift8 + 8));
+        a1_1_x128 =
+            _mm_loadu_si128((__m128i *)(above + base_x + base_shift8 + 9));
+        a0_1_x128 = _mm_shuffle_epi8(a0_1_x128,
+                                     *(__m128i *)HighbdLoadMaskx[base_shift8]);
+        a1_1_x128 = _mm_shuffle_epi8(a1_1_x128,
+                                     *(__m128i *)HighbdLoadMaskx[base_shift8]);
+
+        a0_1_x = _mm256_cvtepu16_epi32(a0_1_x128);
+        a1_1_x = _mm256_cvtepu16_epi32(a1_1_x128);
+
+        r6 = _mm256_slli_epi32(
+            _mm256_add_epi32(c0123, _mm256_add_epi32(j256, c8)), 6);
+        shift = _mm256_srli_epi32(
+            _mm256_and_si256(_mm256_sub_epi32(r6, ydx), c3f), 1);
+
+        diff = _mm256_sub_epi32(a1_1_x, a0_1_x);  // a[x+1] - a[x]
+        a32 = _mm256_slli_epi32(a0_1_x, 5);       // a[x] * 32
+        a32 = _mm256_add_epi32(a32, a16);         // a[x] * 32 + 16
+        b = _mm256_mullo_epi32(diff, shift);
+
+        resx[1] = _mm256_add_epi32(a32, b);
+        resx[1] = _mm256_srli_epi32(resx[1], 5);
+        resx[1] = _mm256_packus_epi32(
+            resx[1],
+            _mm256_castsi128_si256(_mm256_extracti128_si256(resx[1], 1)));
+      }
+      resx[0] =
+          _mm256_inserti128_si256(resx[0], _mm256_castsi256_si128(resx[1]),
+                                  1);  // 16 16bit values
+
+      // y calc
+      resy[0] = _mm256_setzero_si256();
+      if ((base_x < min_base_x)) {
+        __m256i c256, y_c256, y_c_1_256, base_y_c256, mask256;
+        r6 = _mm256_set1_epi32(r << 6);
+        c256 = _mm256_add_epi32(j256, c1234);
+        y_c256 = _mm256_sub_epi32(r6, _mm256_mullo_epi32(c256, dy256));
+        base_y_c256 = _mm256_srai_epi32(y_c256, frac_bits_y);
+        mask256 = _mm256_cmpgt_epi32(min_base_y256, base_y_c256);
+        base_y_c256 = _mm256_andnot_si256(mask256, base_y_c256);
+        _mm256_store_si256((__m256i *)base_y_c, base_y_c256);
+        c256 = _mm256_add_epi32(c256, c8);
+        y_c_1_256 = _mm256_sub_epi32(r6, _mm256_mullo_epi32(c256, dy256));
+        base_y_c256 = _mm256_srai_epi32(y_c_1_256, frac_bits_y);
+        mask256 = _mm256_cmpgt_epi32(min_base_y256, base_y_c256);
+        base_y_c256 = _mm256_andnot_si256(mask256, base_y_c256);
+        _mm256_store_si256((__m256i *)(base_y_c + 8), base_y_c256);
+
+        a0_y = _mm256_cvtepu16_epi32(_mm_setr_epi16(
+            left[base_y_c[0]], left[base_y_c[1]], left[base_y_c[2]],
+            left[base_y_c[3]], left[base_y_c[4]], left[base_y_c[5]],
+            left[base_y_c[6]], left[base_y_c[7]]));
+        a1_y = _mm256_cvtepu16_epi32(_mm_setr_epi16(
+            left[base_y_c[0] + 1], left[base_y_c[1] + 1], left[base_y_c[2] + 1],
+            left[base_y_c[3] + 1], left[base_y_c[4] + 1], left[base_y_c[5] + 1],
+            left[base_y_c[6] + 1], left[base_y_c[7] + 1]));
+
+        shift = _mm256_srli_epi32(_mm256_and_si256(y_c256, c3f), 1);
+
+        diff = _mm256_sub_epi32(a1_y, a0_y);  // a[x+1] - a[x]
+        a32 = _mm256_slli_epi32(a0_y, 5);     // a[x] * 32
+        a32 = _mm256_add_epi32(a32, a16);     // a[x] * 32 + 16
+
+        b = _mm256_mullo_epi32(diff, shift);
+        res = _mm256_add_epi32(a32, b);
+        res = _mm256_srli_epi32(res, 5);
+
+        resy[0] = _mm256_packus_epi32(
+            res, _mm256_castsi128_si256(_mm256_extracti128_si256(res, 1)));
+
+        a0_y = _mm256_cvtepu16_epi32(_mm_setr_epi16(
+            left[base_y_c[8]], left[base_y_c[9]], left[base_y_c[10]],
+            left[base_y_c[11]], left[base_y_c[12]], left[base_y_c[13]],
+            left[base_y_c[14]], left[base_y_c[15]]));
+        a1_y = _mm256_cvtepu16_epi32(
+            _mm_setr_epi16(left[base_y_c[8] + 1], left[base_y_c[9] + 1],
+                           left[base_y_c[10] + 1], left[base_y_c[11] + 1],
+                           left[base_y_c[12] + 1], left[base_y_c[13] + 1],
+                           left[base_y_c[14] + 1], left[base_y_c[15] + 1]));
+        shift = _mm256_srli_epi32(_mm256_and_si256(y_c_1_256, c3f), 1);
+
+        diff = _mm256_sub_epi32(a1_y, a0_y);  // a[x+1] - a[x]
+        a32 = _mm256_slli_epi32(a0_y, 5);     // a[x] * 32
+        a32 = _mm256_add_epi32(a32, a16);     // a[x] * 32 + 16
+
+        b = _mm256_mullo_epi32(diff, shift);
+        res = _mm256_add_epi32(a32, b);
+        res = _mm256_srli_epi32(res, 5);
+
+        resy[1] = _mm256_packus_epi32(
+            res, _mm256_castsi128_si256(_mm256_extracti128_si256(res, 1)));
+
+        resy[0] =
+            _mm256_inserti128_si256(resy[0], _mm256_castsi256_si128(resy[1]),
+                                    1);  // 16 16bit values
+      }
+
+      resxy = _mm256_blendv_epi8(resx[0], resy[0],
+                                 *(__m256i *)HighbdBaseMask[base_min_diff]);
+      _mm256_storeu_si256((__m256i *)(dst + j), resxy);
+    }  // for j
+    dst += stride;
+  }
+}
+
+static void highbd_dr_prediction_z2_HxW_avx2(
+    int H, int W, uint16_t *dst, ptrdiff_t stride, const uint16_t *above,
+    const uint16_t *left, int upsample_above, int upsample_left, int dx,
+    int dy) {
+  // here upsample_above and upsample_left are 0 by design of
+  // av1_use_intra_edge_upsample
+  const int min_base_x = -1;
+  const int min_base_y = -1;
+  (void)upsample_above;
+  (void)upsample_left;
+  const int frac_bits_x = 6;
+  const int frac_bits_y = 6;
+
+  // pre-filter above pixels
+  // store in temp buffers:
+  //   above[x] * 32 + 16
+  //   above[x+1] - above[x]
+  // final pixels will be calculated as:
+  //   (above[x] * 32 + 16 + (above[x+1] - above[x]) * shift) >> 5
+  __m256i a0_x, a1_x, a32, a16, c3f, c1;
+  __m256i diff, min_base_y256, dy256, c1234, c0123;
+  DECLARE_ALIGNED(32, int16_t, base_y_c[16]);
+
+  a16 = _mm256_set1_epi16(16);
+  c1 = _mm256_srli_epi16(a16, 4);
+  min_base_y256 = _mm256_set1_epi16(min_base_y);
+  c3f = _mm256_set1_epi16(0x3f);
+  dy256 = _mm256_set1_epi16(dy);
+  c0123 =
+      _mm256_setr_epi16(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
+  c1234 = _mm256_add_epi16(c0123, c1);
+
+  for (int r = 0; r < H; r++) {
+    __m256i b, res, shift;
+    __m256i resx, resy, ydx;
+    __m256i resxy, j256, r6;
+    __m128i a0_x128, a1_x128, a0_1_x128, a1_1_x128;
+    int y = r + 1;
+    ydx = _mm256_set1_epi16((short)(y * dx));
+
+    for (int j = 0; j < W; j += 16) {
+      j256 = _mm256_set1_epi16(j);
+      int base_x = ((j << 6) - y * dx) >> frac_bits_x;
+      int base_shift = 0;
+      if ((base_x) < (min_base_x - 1)) {
+        base_shift = (min_base_x - (base_x)-1);
+      }
+      int base_min_diff = (min_base_x - base_x);
+      if (base_min_diff > 16) {
+        base_min_diff = 16;
+      } else {
+        if (base_min_diff < 0) base_min_diff = 0;
+      }
+
+      if (base_shift < 8) {
+        a0_x128 = _mm_loadu_si128((__m128i *)(above + base_x + base_shift));
+        a1_x128 = _mm_loadu_si128((__m128i *)(above + base_x + base_shift + 1));
+        a0_x128 =
+            _mm_shuffle_epi8(a0_x128, *(__m128i *)HighbdLoadMaskx[base_shift]);
+        a1_x128 =
+            _mm_shuffle_epi8(a1_x128, *(__m128i *)HighbdLoadMaskx[base_shift]);
+
+        a0_x = _mm256_castsi128_si256(a0_x128);
+        a1_x = _mm256_castsi128_si256(a1_x128);
+      } else {
+        a0_x = _mm256_setzero_si256();
+        a1_x = _mm256_setzero_si256();
+      }
+
+      int base_shift1 = 0;
+      if (base_shift > 8) {
+        base_shift1 = base_shift - 8;
+      }
+      if (base_shift1 < 8) {
+        a0_1_x128 =
+            _mm_loadu_si128((__m128i *)(above + base_x + base_shift1 + 8));
+        a1_1_x128 =
+            _mm_loadu_si128((__m128i *)(above + base_x + base_shift1 + 9));
+        a0_1_x128 = _mm_shuffle_epi8(a0_1_x128,
+                                     *(__m128i *)HighbdLoadMaskx[base_shift1]);
+        a1_1_x128 = _mm_shuffle_epi8(a1_1_x128,
+                                     *(__m128i *)HighbdLoadMaskx[base_shift1]);
+
+        a0_x = _mm256_inserti128_si256(a0_x, a0_1_x128, 1);
+        a1_x = _mm256_inserti128_si256(a1_x, a1_1_x128, 1);
+      }
+      r6 = _mm256_slli_epi16(_mm256_add_epi16(c0123, j256), 6);
+      shift = _mm256_srli_epi16(
+          _mm256_and_si256(_mm256_sub_epi16(r6, ydx), c3f), 1);
+
+      diff = _mm256_sub_epi16(a1_x, a0_x);  // a[x+1] - a[x]
+      a32 = _mm256_slli_epi16(a0_x, 5);     // a[x] * 32
+      a32 = _mm256_add_epi16(a32, a16);     // a[x] * 32 + 16
+
+      b = _mm256_mullo_epi16(diff, shift);
+      res = _mm256_add_epi16(a32, b);
+      resx = _mm256_srli_epi16(res, 5);  // 16 16-bit values
+
+      // y calc
+      resy = _mm256_setzero_si256();
+      __m256i a0_y, a1_y, shifty;
+      if ((base_x < min_base_x)) {
+        __m256i c256, y_c256, base_y_c256, mask256, mul16;
+        r6 = _mm256_set1_epi16(r << 6);
+        c256 = _mm256_add_epi16(j256, c1234);
+        mul16 = _mm256_min_epu16(_mm256_mullo_epi16(c256, dy256),
+                                 _mm256_srli_epi16(min_base_y256, 1));
+        y_c256 = _mm256_sub_epi16(r6, mul16);
+        base_y_c256 = _mm256_srai_epi16(y_c256, frac_bits_y);
+        mask256 = _mm256_cmpgt_epi16(min_base_y256, base_y_c256);
+        base_y_c256 = _mm256_andnot_si256(mask256, base_y_c256);
+        _mm256_store_si256((__m256i *)base_y_c, base_y_c256);
+
+        a0_y = _mm256_setr_epi16(
+            left[base_y_c[0]], left[base_y_c[1]], left[base_y_c[2]],
+            left[base_y_c[3]], left[base_y_c[4]], left[base_y_c[5]],
+            left[base_y_c[6]], left[base_y_c[7]], left[base_y_c[8]],
+            left[base_y_c[9]], left[base_y_c[10]], left[base_y_c[11]],
+            left[base_y_c[12]], left[base_y_c[13]], left[base_y_c[14]],
+            left[base_y_c[15]]);
+        base_y_c256 = _mm256_add_epi16(base_y_c256, c1);
+        _mm256_store_si256((__m256i *)base_y_c, base_y_c256);
+
+        a1_y = _mm256_setr_epi16(
+            left[base_y_c[0]], left[base_y_c[1]], left[base_y_c[2]],
+            left[base_y_c[3]], left[base_y_c[4]], left[base_y_c[5]],
+            left[base_y_c[6]], left[base_y_c[7]], left[base_y_c[8]],
+            left[base_y_c[9]], left[base_y_c[10]], left[base_y_c[11]],
+            left[base_y_c[12]], left[base_y_c[13]], left[base_y_c[14]],
+            left[base_y_c[15]]);
+
+        shifty = _mm256_srli_epi16(_mm256_and_si256(y_c256, c3f), 1);
+
+        diff = _mm256_sub_epi16(a1_y, a0_y);  // a[x+1] - a[x]
+        a32 = _mm256_slli_epi16(a0_y, 5);     // a[x] * 32
+        a32 = _mm256_add_epi16(a32, a16);     // a[x] * 32 + 16
+
+        b = _mm256_mullo_epi16(diff, shifty);
+        res = _mm256_add_epi16(a32, b);
+        resy = _mm256_srli_epi16(res, 5);
+      }
+
+      resxy = _mm256_blendv_epi8(resx, resy,
+                                 *(__m256i *)HighbdBaseMask[base_min_diff]);
+      _mm256_storeu_si256((__m256i *)(dst + j), resxy);
+    }  // for j
+    dst += stride;
+  }
+}
+
+// Directional prediction, zone 2: 90 < angle < 180
+void av1_highbd_dr_prediction_z2_avx2(uint16_t *dst, ptrdiff_t stride, int bw,
+                                      int bh, const uint16_t *above,
+                                      const uint16_t *left, int upsample_above,
+                                      int upsample_left, int dx, int dy,
+                                      int bd) {
+  (void)bd;
+  assert(dx > 0);
+  assert(dy > 0);
+  switch (bw) {
+    case 4:
+      if (bd < 12) {
+        highbd_dr_prediction_z2_Nx4_avx2(bh, dst, stride, above, left,
+                                         upsample_above, upsample_left, dx, dy);
+      } else {
+        highbd_dr_prediction_32bit_z2_Nx4_avx2(bh, dst, stride, above, left,
+                                               upsample_above, upsample_left,
+                                               dx, dy);
+      }
+      break;
+    case 8:
+      if (bd < 12) {
+        highbd_dr_prediction_z2_Nx8_avx2(bh, dst, stride, above, left,
+                                         upsample_above, upsample_left, dx, dy);
+      } else {
+        highbd_dr_prediction_32bit_z2_Nx8_avx2(bh, dst, stride, above, left,
+                                               upsample_above, upsample_left,
+                                               dx, dy);
+      }
+      break;
+    default:
+      if (bd < 12) {
+        highbd_dr_prediction_z2_HxW_avx2(bh, bw, dst, stride, above, left,
+                                         upsample_above, upsample_left, dx, dy);
+      } else {
+        highbd_dr_prediction_32bit_z2_HxW_avx2(bh, bw, dst, stride, above, left,
+                                               upsample_above, upsample_left,
+                                               dx, dy);
+      }
+      break;
+  }
+}
+
+//  Directional prediction, zone 3 functions
+static void highbd_dr_prediction_z3_4x4_avx2(uint16_t *dst, ptrdiff_t stride,
+                                             const uint16_t *left,
+                                             int upsample_left, int dy,
+                                             int bd) {
+  __m128i dstvec[4], d[4];
+  if (bd < 12) {
+    highbd_dr_prediction_z1_4xN_internal_avx2(4, dstvec, left, upsample_left,
+                                              dy);
+  } else {
+    highbd_dr_prediction_32bit_z1_4xN_internal_avx2(4, dstvec, left,
+                                                    upsample_left, dy);
+  }
+  highbd_transpose4x8_8x4_low_sse2(&dstvec[0], &dstvec[1], &dstvec[2],
+                                   &dstvec[3], &d[0], &d[1], &d[2], &d[3]);
+  _mm_storel_epi64((__m128i *)(dst + 0 * stride), d[0]);
+  _mm_storel_epi64((__m128i *)(dst + 1 * stride), d[1]);
+  _mm_storel_epi64((__m128i *)(dst + 2 * stride), d[2]);
+  _mm_storel_epi64((__m128i *)(dst + 3 * stride), d[3]);
+  return;
+}
+
+static void highbd_dr_prediction_z3_8x8_avx2(uint16_t *dst, ptrdiff_t stride,
+                                             const uint16_t *left,
+                                             int upsample_left, int dy,
+                                             int bd) {
+  __m128i dstvec[8], d[8];
+  if (bd < 12) {
+    highbd_dr_prediction_z1_8xN_internal_avx2(8, dstvec, left, upsample_left,
+                                              dy);
+  } else {
+    highbd_dr_prediction_32bit_z1_8xN_internal_avx2(8, dstvec, left,
+                                                    upsample_left, dy);
+  }
+  highbd_transpose8x8_sse2(&dstvec[0], &dstvec[1], &dstvec[2], &dstvec[3],
+                           &dstvec[4], &dstvec[5], &dstvec[6], &dstvec[7],
+                           &d[0], &d[1], &d[2], &d[3], &d[4], &d[5], &d[6],
+                           &d[7]);
+  for (int i = 0; i < 8; i++) {
+    _mm_storeu_si128((__m128i *)(dst + i * stride), d[i]);
+  }
+}
+
+static void highbd_dr_prediction_z3_4x8_avx2(uint16_t *dst, ptrdiff_t stride,
+                                             const uint16_t *left,
+                                             int upsample_left, int dy,
+                                             int bd) {
+  __m128i dstvec[4], d[8];
+  if (bd < 12) {
+    highbd_dr_prediction_z1_8xN_internal_avx2(4, dstvec, left, upsample_left,
+                                              dy);
+  } else {
+    highbd_dr_prediction_32bit_z1_8xN_internal_avx2(4, dstvec, left,
+                                                    upsample_left, dy);
+  }
+
+  highbd_transpose4x8_8x4_sse2(&dstvec[0], &dstvec[1], &dstvec[2], &dstvec[3],
+                               &d[0], &d[1], &d[2], &d[3], &d[4], &d[5], &d[6],
+                               &d[7]);
+  for (int i = 0; i < 8; i++) {
+    _mm_storel_epi64((__m128i *)(dst + i * stride), d[i]);
+  }
+}
+
+static void highbd_dr_prediction_z3_8x4_avx2(uint16_t *dst, ptrdiff_t stride,
+                                             const uint16_t *left,
+                                             int upsample_left, int dy,
+                                             int bd) {
+  __m128i dstvec[8], d[4];
+  if (bd < 12) {
+    highbd_dr_prediction_z1_4xN_internal_avx2(8, dstvec, left, upsample_left,
+                                              dy);
+  } else {
+    highbd_dr_prediction_32bit_z1_4xN_internal_avx2(8, dstvec, left,
+                                                    upsample_left, dy);
+  }
+
+  highbd_transpose8x8_low_sse2(&dstvec[0], &dstvec[1], &dstvec[2], &dstvec[3],
+                               &dstvec[4], &dstvec[5], &dstvec[6], &dstvec[7],
+                               &d[0], &d[1], &d[2], &d[3]);
+  _mm_storeu_si128((__m128i *)(dst + 0 * stride), d[0]);
+  _mm_storeu_si128((__m128i *)(dst + 1 * stride), d[1]);
+  _mm_storeu_si128((__m128i *)(dst + 2 * stride), d[2]);
+  _mm_storeu_si128((__m128i *)(dst + 3 * stride), d[3]);
+}
+
+static void highbd_dr_prediction_z3_8x16_avx2(uint16_t *dst, ptrdiff_t stride,
+                                              const uint16_t *left,
+                                              int upsample_left, int dy,
+                                              int bd) {
+  __m256i dstvec[8], d[8];
+  if (bd < 12) {
+    highbd_dr_prediction_z1_16xN_internal_avx2(8, dstvec, left, upsample_left,
+                                               dy);
+  } else {
+    highbd_dr_prediction_32bit_z1_16xN_internal_avx2(8, dstvec, left,
+                                                     upsample_left, dy);
+  }
+  highbd_transpose8x16_16x8_avx2(dstvec, d);
+  for (int i = 0; i < 8; i++) {
+    _mm_storeu_si128((__m128i *)(dst + i * stride),
+                     _mm256_castsi256_si128(d[i]));
+  }
+  for (int i = 8; i < 16; i++) {
+    _mm_storeu_si128((__m128i *)(dst + i * stride),
+                     _mm256_extracti128_si256(d[i - 8], 1));
+  }
+}
+
+static void highbd_dr_prediction_z3_16x8_avx2(uint16_t *dst, ptrdiff_t stride,
+                                              const uint16_t *left,
+                                              int upsample_left, int dy,
+                                              int bd) {
+  __m128i dstvec[16], d[16];
+  if (bd < 12) {
+    highbd_dr_prediction_z1_8xN_internal_avx2(16, dstvec, left, upsample_left,
+                                              dy);
+  } else {
+    highbd_dr_prediction_32bit_z1_8xN_internal_avx2(16, dstvec, left,
+                                                    upsample_left, dy);
+  }
+  for (int i = 0; i < 16; i += 8) {
+    highbd_transpose8x8_sse2(&dstvec[0 + i], &dstvec[1 + i], &dstvec[2 + i],
+                             &dstvec[3 + i], &dstvec[4 + i], &dstvec[5 + i],
+                             &dstvec[6 + i], &dstvec[7 + i], &d[0 + i],
+                             &d[1 + i], &d[2 + i], &d[3 + i], &d[4 + i],
+                             &d[5 + i], &d[6 + i], &d[7 + i]);
+  }
+  for (int i = 0; i < 8; i++) {
+    _mm_storeu_si128((__m128i *)(dst + i * stride), d[i]);
+    _mm_storeu_si128((__m128i *)(dst + i * stride + 8), d[i + 8]);
+  }
+}
+
+static void highbd_dr_prediction_z3_4x16_avx2(uint16_t *dst, ptrdiff_t stride,
+                                              const uint16_t *left,
+                                              int upsample_left, int dy,
+                                              int bd) {
+  __m256i dstvec[4], d[4], d1;
+  if (bd < 12) {
+    highbd_dr_prediction_z1_16xN_internal_avx2(4, dstvec, left, upsample_left,
+                                               dy);
+  } else {
+    highbd_dr_prediction_32bit_z1_16xN_internal_avx2(4, dstvec, left,
+                                                     upsample_left, dy);
+  }
+  highbd_transpose4x16_avx2(dstvec, d);
+  for (int i = 0; i < 4; i++) {
+    _mm_storel_epi64((__m128i *)(dst + i * stride),
+                     _mm256_castsi256_si128(d[i]));
+    d1 = _mm256_bsrli_epi128(d[i], 8);
+    _mm_storel_epi64((__m128i *)(dst + (i + 4) * stride),
+                     _mm256_castsi256_si128(d1));
+    _mm_storel_epi64((__m128i *)(dst + (i + 8) * stride),
+                     _mm256_extracti128_si256(d[i], 1));
+    _mm_storel_epi64((__m128i *)(dst + (i + 12) * stride),
+                     _mm256_extracti128_si256(d1, 1));
+  }
+}
+
+static void highbd_dr_prediction_z3_16x4_avx2(uint16_t *dst, ptrdiff_t stride,
+                                              const uint16_t *left,
+                                              int upsample_left, int dy,
+                                              int bd) {
+  __m128i dstvec[16], d[8];
+  if (bd < 12) {
+    highbd_dr_prediction_z1_4xN_internal_avx2(16, dstvec, left, upsample_left,
+                                              dy);
+  } else {
+    highbd_dr_prediction_32bit_z1_4xN_internal_avx2(16, dstvec, left,
+                                                    upsample_left, dy);
+  }
+  highbd_transpose16x4_8x8_sse2(dstvec, d);
+
+  _mm_storeu_si128((__m128i *)(dst + 0 * stride), d[0]);
+  _mm_storeu_si128((__m128i *)(dst + 0 * stride + 8), d[1]);
+  _mm_storeu_si128((__m128i *)(dst + 1 * stride), d[2]);
+  _mm_storeu_si128((__m128i *)(dst + 1 * stride + 8), d[3]);
+  _mm_storeu_si128((__m128i *)(dst + 2 * stride), d[4]);
+  _mm_storeu_si128((__m128i *)(dst + 2 * stride + 8), d[5]);
+  _mm_storeu_si128((__m128i *)(dst + 3 * stride), d[6]);
+  _mm_storeu_si128((__m128i *)(dst + 3 * stride + 8), d[7]);
+}
+
+static void highbd_dr_prediction_z3_8x32_avx2(uint16_t *dst, ptrdiff_t stride,
+                                              const uint16_t *left,
+                                              int upsample_left, int dy,
+                                              int bd) {
+  __m256i dstvec[16], d[16];
+  if (bd < 12) {
+    highbd_dr_prediction_z1_32xN_internal_avx2(8, dstvec, left, upsample_left,
+                                               dy);
+  } else {
+    highbd_dr_prediction_32bit_z1_32xN_internal_avx2(8, dstvec, left,
+                                                     upsample_left, dy);
+  }
+
+  for (int i = 0; i < 16; i += 8) {
+    highbd_transpose8x16_16x8_avx2(dstvec + i, d + i);
+  }
+
+  for (int i = 0; i < 8; i++) {
+    _mm_storeu_si128((__m128i *)(dst + i * stride),
+                     _mm256_castsi256_si128(d[i]));
+  }
+  for (int i = 0; i < 8; i++) {
+    _mm_storeu_si128((__m128i *)(dst + (i + 8) * stride),
+                     _mm256_extracti128_si256(d[i], 1));
+  }
+  for (int i = 8; i < 16; i++) {
+    _mm_storeu_si128((__m128i *)(dst + (i + 8) * stride),
+                     _mm256_castsi256_si128(d[i]));
+  }
+  for (int i = 8; i < 16; i++) {
+    _mm_storeu_si128((__m128i *)(dst + (i + 16) * stride),
+                     _mm256_extracti128_si256(d[i], 1));
+  }
+}
+
+static void highbd_dr_prediction_z3_32x8_avx2(uint16_t *dst, ptrdiff_t stride,
+                                              const uint16_t *left,
+                                              int upsample_left, int dy,
+                                              int bd) {
+  __m128i dstvec[32], d[32];
+  if (bd < 12) {
+    highbd_dr_prediction_z1_8xN_internal_avx2(32, dstvec, left, upsample_left,
+                                              dy);
+  } else {
+    highbd_dr_prediction_32bit_z1_8xN_internal_avx2(32, dstvec, left,
+                                                    upsample_left, dy);
+  }
+
+  for (int i = 0; i < 32; i += 8) {
+    highbd_transpose8x8_sse2(&dstvec[0 + i], &dstvec[1 + i], &dstvec[2 + i],
+                             &dstvec[3 + i], &dstvec[4 + i], &dstvec[5 + i],
+                             &dstvec[6 + i], &dstvec[7 + i], &d[0 + i],
+                             &d[1 + i], &d[2 + i], &d[3 + i], &d[4 + i],
+                             &d[5 + i], &d[6 + i], &d[7 + i]);
+  }
+  for (int i = 0; i < 8; i++) {
+    _mm_storeu_si128((__m128i *)(dst + i * stride), d[i]);
+    _mm_storeu_si128((__m128i *)(dst + i * stride + 8), d[i + 8]);
+    _mm_storeu_si128((__m128i *)(dst + i * stride + 16), d[i + 16]);
+    _mm_storeu_si128((__m128i *)(dst + i * stride + 24), d[i + 24]);
+  }
+}
+
+static void highbd_dr_prediction_z3_16x16_avx2(uint16_t *dst, ptrdiff_t stride,
+                                               const uint16_t *left,
+                                               int upsample_left, int dy,
+                                               int bd) {
+  __m256i dstvec[16], d[16];
+  if (bd < 12) {
+    highbd_dr_prediction_z1_16xN_internal_avx2(16, dstvec, left, upsample_left,
+                                               dy);
+  } else {
+    highbd_dr_prediction_32bit_z1_16xN_internal_avx2(16, dstvec, left,
+                                                     upsample_left, dy);
+  }
+
+  highbd_transpose16x16_avx2(dstvec, d);
+
+  for (int i = 0; i < 16; i++) {
+    _mm256_storeu_si256((__m256i *)(dst + i * stride), d[i]);
+  }
+}
+
+static void highbd_dr_prediction_z3_32x32_avx2(uint16_t *dst, ptrdiff_t stride,
+                                               const uint16_t *left,
+                                               int upsample_left, int dy,
+                                               int bd) {
+  __m256i dstvec[64], d[16];
+  if (bd < 12) {
+    highbd_dr_prediction_z1_32xN_internal_avx2(32, dstvec, left, upsample_left,
+                                               dy);
+  } else {
+    highbd_dr_prediction_32bit_z1_32xN_internal_avx2(32, dstvec, left,
+                                                     upsample_left, dy);
+  }
+  highbd_transpose16x16_avx2(dstvec, d);
+  for (int j = 0; j < 16; j++) {
+    _mm256_storeu_si256((__m256i *)(dst + j * stride), d[j]);
+  }
+  highbd_transpose16x16_avx2(dstvec + 16, d);
+  for (int j = 0; j < 16; j++) {
+    _mm256_storeu_si256((__m256i *)(dst + j * stride + 16), d[j]);
+  }
+  highbd_transpose16x16_avx2(dstvec + 32, d);
+  for (int j = 0; j < 16; j++) {
+    _mm256_storeu_si256((__m256i *)(dst + (j + 16) * stride), d[j]);
+  }
+  highbd_transpose16x16_avx2(dstvec + 48, d);
+  for (int j = 0; j < 16; j++) {
+    _mm256_storeu_si256((__m256i *)(dst + (j + 16) * stride + 16), d[j]);
+  }
+}
+
+static void highbd_dr_prediction_z3_64x64_avx2(uint16_t *dst, ptrdiff_t stride,
+                                               const uint16_t *left,
+                                               int upsample_left, int dy,
+                                               int bd) {
+  DECLARE_ALIGNED(16, uint16_t, dstT[64 * 64]);
+  if (bd < 12) {
+    highbd_dr_prediction_z1_64xN_avx2(64, dstT, 64, left, upsample_left, dy);
+  } else {
+    highbd_dr_prediction_32bit_z1_64xN_avx2(64, dstT, 64, left, upsample_left,
+                                            dy);
+  }
+  highbd_transpose(dstT, 64, dst, stride, 64, 64);
+}
+
+static void highbd_dr_prediction_z3_16x32_avx2(uint16_t *dst, ptrdiff_t stride,
+                                               const uint16_t *left,
+                                               int upsample_left, int dy,
+                                               int bd) {
+  __m256i dstvec[32], d[32];
+  if (bd < 12) {
+    highbd_dr_prediction_z1_32xN_internal_avx2(16, dstvec, left, upsample_left,
+                                               dy);
+  } else {
+    highbd_dr_prediction_32bit_z1_32xN_internal_avx2(16, dstvec, left,
+                                                     upsample_left, dy);
+  }
+  for (int i = 0; i < 32; i += 8) {
+    highbd_transpose8x16_16x8_avx2(dstvec + i, d + i);
+  }
+  // store
+  for (int j = 0; j < 32; j += 16) {
+    for (int i = 0; i < 8; i++) {
+      _mm_storeu_si128((__m128i *)(dst + (i + j) * stride),
+                       _mm256_castsi256_si128(d[(i + j)]));
+    }
+    for (int i = 0; i < 8; i++) {
+      _mm_storeu_si128((__m128i *)(dst + (i + j) * stride + 8),
+                       _mm256_castsi256_si128(d[(i + j) + 8]));
+    }
+    for (int i = 8; i < 16; i++) {
+      _mm256_storeu_si256(
+          (__m256i *)(dst + (i + j) * stride),
+          _mm256_inserti128_si256(
+              d[(i + j)], _mm256_extracti128_si256(d[(i + j) - 8], 1), 0));
+    }
+  }
+}
+
+static void highbd_dr_prediction_z3_32x16_avx2(uint16_t *dst, ptrdiff_t stride,
+                                               const uint16_t *left,
+                                               int upsample_left, int dy,
+                                               int bd) {
+  __m256i dstvec[32], d[16];
+  if (bd < 12) {
+    highbd_dr_prediction_z1_16xN_internal_avx2(32, dstvec, left, upsample_left,
+                                               dy);
+  } else {
+    highbd_dr_prediction_32bit_z1_16xN_internal_avx2(32, dstvec, left,
+                                                     upsample_left, dy);
+  }
+  for (int i = 0; i < 32; i += 16) {
+    highbd_transpose16x16_avx2((dstvec + i), d);
+    for (int j = 0; j < 16; j++) {
+      _mm256_storeu_si256((__m256i *)(dst + j * stride + i), d[j]);
+    }
+  }
+}
+
+static void highbd_dr_prediction_z3_32x64_avx2(uint16_t *dst, ptrdiff_t stride,
+                                               const uint16_t *left,
+                                               int upsample_left, int dy,
+                                               int bd) {
+  uint16_t dstT[64 * 32];
+  if (bd < 12) {
+    highbd_dr_prediction_z1_64xN_avx2(32, dstT, 64, left, upsample_left, dy);
+  } else {
+    highbd_dr_prediction_32bit_z1_64xN_avx2(32, dstT, 64, left, upsample_left,
+                                            dy);
+  }
+  highbd_transpose(dstT, 64, dst, stride, 32, 64);
+}
+
+static void highbd_dr_prediction_z3_64x32_avx2(uint16_t *dst, ptrdiff_t stride,
+                                               const uint16_t *left,
+                                               int upsample_left, int dy,
+                                               int bd) {
+  DECLARE_ALIGNED(16, uint16_t, dstT[32 * 64]);
+  highbd_dr_prediction_z1_32xN_avx2(64, dstT, 32, left, upsample_left, dy, bd);
+  highbd_transpose(dstT, 32, dst, stride, 64, 32);
+  return;
+}
+
+static void highbd_dr_prediction_z3_16x64_avx2(uint16_t *dst, ptrdiff_t stride,
+                                               const uint16_t *left,
+                                               int upsample_left, int dy,
+                                               int bd) {
+  DECLARE_ALIGNED(16, uint16_t, dstT[64 * 16]);
+  if (bd < 12) {
+    highbd_dr_prediction_z1_64xN_avx2(16, dstT, 64, left, upsample_left, dy);
+  } else {
+    highbd_dr_prediction_32bit_z1_64xN_avx2(16, dstT, 64, left, upsample_left,
+                                            dy);
+  }
+  highbd_transpose(dstT, 64, dst, stride, 16, 64);
+}
+
+static void highbd_dr_prediction_z3_64x16_avx2(uint16_t *dst, ptrdiff_t stride,
+                                               const uint16_t *left,
+                                               int upsample_left, int dy,
+                                               int bd) {
+  __m256i dstvec[64], d[16];
+  if (bd < 12) {
+    highbd_dr_prediction_z1_16xN_internal_avx2(64, dstvec, left, upsample_left,
+                                               dy);
+  } else {
+    highbd_dr_prediction_32bit_z1_16xN_internal_avx2(64, dstvec, left,
+                                                     upsample_left, dy);
+  }
+  for (int i = 0; i < 64; i += 16) {
+    highbd_transpose16x16_avx2((dstvec + i), d);
+    for (int j = 0; j < 16; j++) {
+      _mm256_storeu_si256((__m256i *)(dst + j * stride + i), d[j]);
+    }
+  }
+}
+
+void av1_highbd_dr_prediction_z3_avx2(uint16_t *dst, ptrdiff_t stride, int bw,
+                                      int bh, const uint16_t *above,
+                                      const uint16_t *left, int upsample_left,
+                                      int dx, int dy, int bd) {
+  (void)above;
+  (void)dx;
+
+  assert(dx == 1);
+  assert(dy > 0);
+  if (bw == bh) {
+    switch (bw) {
+      case 4:
+        highbd_dr_prediction_z3_4x4_avx2(dst, stride, left, upsample_left, dy,
+                                         bd);
+        break;
+      case 8:
+        highbd_dr_prediction_z3_8x8_avx2(dst, stride, left, upsample_left, dy,
+                                         bd);
+        break;
+      case 16:
+        highbd_dr_prediction_z3_16x16_avx2(dst, stride, left, upsample_left, dy,
+                                           bd);
+        break;
+      case 32:
+        highbd_dr_prediction_z3_32x32_avx2(dst, stride, left, upsample_left, dy,
+                                           bd);
+        break;
+      case 64:
+        highbd_dr_prediction_z3_64x64_avx2(dst, stride, left, upsample_left, dy,
+                                           bd);
+        break;
+    }
+  } else {
+    if (bw < bh) {
+      if (bw + bw == bh) {
+        switch (bw) {
+          case 4:
+            highbd_dr_prediction_z3_4x8_avx2(dst, stride, left, upsample_left,
+                                             dy, bd);
+            break;
+          case 8:
+            highbd_dr_prediction_z3_8x16_avx2(dst, stride, left, upsample_left,
+                                              dy, bd);
+            break;
+          case 16:
+            highbd_dr_prediction_z3_16x32_avx2(dst, stride, left, upsample_left,
+                                               dy, bd);
+            break;
+          case 32:
+            highbd_dr_prediction_z3_32x64_avx2(dst, stride, left, upsample_left,
+                                               dy, bd);
+            break;
+        }
+      } else {
+        switch (bw) {
+          case 4:
+            highbd_dr_prediction_z3_4x16_avx2(dst, stride, left, upsample_left,
+                                              dy, bd);
+            break;
+          case 8:
+            highbd_dr_prediction_z3_8x32_avx2(dst, stride, left, upsample_left,
+                                              dy, bd);
+            break;
+          case 16:
+            highbd_dr_prediction_z3_16x64_avx2(dst, stride, left, upsample_left,
+                                               dy, bd);
+            break;
+        }
+      }
+    } else {
+      if (bh + bh == bw) {
+        switch (bh) {
+          case 4:
+            highbd_dr_prediction_z3_8x4_avx2(dst, stride, left, upsample_left,
+                                             dy, bd);
+            break;
+          case 8:
+            highbd_dr_prediction_z3_16x8_avx2(dst, stride, left, upsample_left,
+                                              dy, bd);
+            break;
+          case 16:
+            highbd_dr_prediction_z3_32x16_avx2(dst, stride, left, upsample_left,
+                                               dy, bd);
+            break;
+          case 32:
+            highbd_dr_prediction_z3_64x32_avx2(dst, stride, left, upsample_left,
+                                               dy, bd);
+            break;
+        }
+      } else {
+        switch (bh) {
+          case 4:
+            highbd_dr_prediction_z3_16x4_avx2(dst, stride, left, upsample_left,
+                                              dy, bd);
+            break;
+          case 8:
+            highbd_dr_prediction_z3_32x8_avx2(dst, stride, left, upsample_left,
+                                              dy, bd);
+            break;
+          case 16:
+            highbd_dr_prediction_z3_64x16_avx2(dst, stride, left, upsample_left,
+                                               dy, bd);
+            break;
+        }
+      }
+    }
+  }
+  return;
+}
+
+// Low bit depth functions
+static DECLARE_ALIGNED(32, uint8_t, BaseMask[33][32]) = {
+  { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+  { 0xff, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+    0,    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+  { 0xff, 0xff, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+    0,    0,    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+  { 0xff, 0xff, 0xff, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+    0,    0,    0,    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+  { 0xff, 0xff, 0xff, 0xff, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+    0,    0,    0,    0,    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+  { 0xff, 0xff, 0xff, 0xff, 0xff, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+    0,    0,    0,    0,    0,    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+  { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+    0,    0,    0,    0,    0,    0,    0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+  { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+    0,    0,    0,    0,    0,    0,    0,    0, 0, 0, 0, 0, 0, 0, 0, 0 },
+  { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0, 0, 0, 0, 0, 0, 0, 0,
+    0,    0,    0,    0,    0,    0,    0,    0,    0, 0, 0, 0, 0, 0, 0, 0 },
+  { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0, 0, 0, 0, 0, 0, 0,
+    0,    0,    0,    0,    0,    0,    0,    0,    0,    0, 0, 0, 0, 0, 0, 0 },
+  { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0,
+    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,
+    0,    0,    0,    0,    0,    0,    0,    0,    0,    0 },
+  { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,    0,
+    0,    0,    0,    0,    0,    0,    0,    0,    0,    0 },
+  { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+    0xff, 0,    0,    0,    0,    0,    0,    0,    0,    0,    0,
+    0,    0,    0,    0,    0,    0,    0,    0,    0,    0 },
+  { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+    0xff, 0xff, 0,    0,    0,    0,    0,    0,    0,    0,    0,
+    0,    0,    0,    0,    0,    0,    0,    0,    0,    0 },
+  { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+    0xff, 0xff, 0xff, 0,    0,    0,    0,    0,    0,    0,    0,
+    0,    0,    0,    0,    0,    0,    0,    0,    0,    0 },
+  { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+    0xff, 0xff, 0xff, 0xff, 0,    0,    0,    0,    0,    0,    0,
+    0,    0,    0,    0,    0,    0,    0,    0,    0,    0 },
+  { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+    0xff, 0xff, 0xff, 0xff, 0xff, 0,    0,    0,    0,    0,    0,
+    0,    0,    0,    0,    0,    0,    0,    0,    0,    0 },
+  { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0,    0,    0,    0,    0,
+    0,    0,    0,    0,    0,    0,    0,    0,    0,    0 },
+  { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0,    0,    0,    0,
+    0,    0,    0,    0,    0,    0,    0,    0,    0,    0 },
+  { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0,    0,    0,
+    0,    0,    0,    0,    0,    0,    0,    0,    0,    0 },
+  { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0,    0,
+    0,    0,    0,    0,    0,    0,    0,    0,    0,    0 },
+  { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0,
+    0,    0,    0,    0,    0,    0,    0,    0,    0,    0 },
+  { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+    0,    0,    0,    0,    0,    0,    0,    0,    0,    0 },
+  { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+    0xff, 0,    0,    0,    0,    0,    0,    0,    0,    0 },
+  { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+    0xff, 0xff, 0,    0,    0,    0,    0,    0,    0,    0 },
+  { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+    0xff, 0xff, 0xff, 0,    0,    0,    0,    0,    0,    0 },
+  { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+    0xff, 0xff, 0xff, 0xff, 0,    0,    0,    0,    0,    0 },
+  { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+    0xff, 0xff, 0xff, 0xff, 0xff, 0,    0,    0,    0,    0 },
+  { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0,    0,    0,    0 },
+  { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0,    0,    0 },
+  { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0,    0 },
+  { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0 },
+  { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+    0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff },
+};
+
+/* clang-format on */
+static AOM_FORCE_INLINE void dr_prediction_z1_HxW_internal_avx2(
+    int H, int W, __m128i *dst, const uint8_t *above, int upsample_above,
+    int dx) {
+  const int frac_bits = 6 - upsample_above;
+  const int max_base_x = ((W + H) - 1) << upsample_above;
+
+  assert(dx > 0);
+  // pre-filter above pixels
+  // store in temp buffers:
+  //   above[x] * 32 + 16
+  //   above[x+1] - above[x]
+  // final pixels will be calculated as:
+  //   (above[x] * 32 + 16 + (above[x+1] - above[x]) * shift) >> 5
+  __m256i a0, a1, a32, a16;
+  __m256i diff, c3f;
+  __m128i a_mbase_x;
+
+  a16 = _mm256_set1_epi16(16);
+  a_mbase_x = _mm_set1_epi8((int8_t)above[max_base_x]);
+  c3f = _mm256_set1_epi16(0x3f);
+
+  int x = dx;
+  for (int r = 0; r < W; r++) {
+    __m256i b, res, shift;
+    __m128i res1, a0_128, a1_128;
+
+    int base = x >> frac_bits;
+    int base_max_diff = (max_base_x - base) >> upsample_above;
+    if (base_max_diff <= 0) {
+      for (int i = r; i < W; ++i) {
+        dst[i] = a_mbase_x;  // save 4 values
+      }
+      return;
+    }
+    if (base_max_diff > H) base_max_diff = H;
+    a0_128 = _mm_loadu_si128((__m128i *)(above + base));
+    a1_128 = _mm_loadu_si128((__m128i *)(above + base + 1));
+
+    if (upsample_above) {
+      a0_128 = _mm_shuffle_epi8(a0_128, *(__m128i *)EvenOddMaskx[0]);
+      a1_128 = _mm_srli_si128(a0_128, 8);
+
+      shift = _mm256_srli_epi16(
+          _mm256_and_si256(
+              _mm256_slli_epi16(_mm256_set1_epi16(x), upsample_above), c3f),
+          1);
+    } else {
+      shift = _mm256_srli_epi16(_mm256_and_si256(_mm256_set1_epi16(x), c3f), 1);
+    }
+    a0 = _mm256_cvtepu8_epi16(a0_128);
+    a1 = _mm256_cvtepu8_epi16(a1_128);
+
+    diff = _mm256_sub_epi16(a1, a0);   // a[x+1] - a[x]
+    a32 = _mm256_slli_epi16(a0, 5);    // a[x] * 32
+    a32 = _mm256_add_epi16(a32, a16);  // a[x] * 32 + 16
+
+    b = _mm256_mullo_epi16(diff, shift);
+    res = _mm256_add_epi16(a32, b);
+    res = _mm256_srli_epi16(res, 5);
+
+    res = _mm256_packus_epi16(
+        res, _mm256_castsi128_si256(
+                 _mm256_extracti128_si256(res, 1)));  // goto 8 bit
+    res1 = _mm256_castsi256_si128(res);               // 16 8bit values
+
+    dst[r] =
+        _mm_blendv_epi8(a_mbase_x, res1, *(__m128i *)BaseMask[base_max_diff]);
+    x += dx;
+  }
+}
+
+static void dr_prediction_z1_4xN_avx2(int N, uint8_t *dst, ptrdiff_t stride,
+                                      const uint8_t *above, int upsample_above,
+                                      int dx) {
+  __m128i dstvec[16];
+
+  dr_prediction_z1_HxW_internal_avx2(4, N, dstvec, above, upsample_above, dx);
+  for (int i = 0; i < N; i++) {
+    *(int *)(dst + stride * i) = _mm_cvtsi128_si32(dstvec[i]);
+  }
+}
+
+static void dr_prediction_z1_8xN_avx2(int N, uint8_t *dst, ptrdiff_t stride,
+                                      const uint8_t *above, int upsample_above,
+                                      int dx) {
+  __m128i dstvec[32];
+
+  dr_prediction_z1_HxW_internal_avx2(8, N, dstvec, above, upsample_above, dx);
+  for (int i = 0; i < N; i++) {
+    _mm_storel_epi64((__m128i *)(dst + stride * i), dstvec[i]);
+  }
+}
+
+static void dr_prediction_z1_16xN_avx2(int N, uint8_t *dst, ptrdiff_t stride,
+                                       const uint8_t *above, int upsample_above,
+                                       int dx) {
+  __m128i dstvec[64];
+
+  dr_prediction_z1_HxW_internal_avx2(16, N, dstvec, above, upsample_above, dx);
+  for (int i = 0; i < N; i++) {
+    _mm_storeu_si128((__m128i *)(dst + stride * i), dstvec[i]);
+  }
+}
+
+static AOM_FORCE_INLINE void dr_prediction_z1_32xN_internal_avx2(
+    int N, __m256i *dstvec, const uint8_t *above, int upsample_above, int dx) {
+  // here upsample_above is 0 by design of av1_use_intra_edge_upsample
+  (void)upsample_above;
+  const int frac_bits = 6;
+  const int max_base_x = ((32 + N) - 1);
+
+  // pre-filter above pixels
+  // store in temp buffers:
+  //   above[x] * 32 + 16
+  //   above[x+1] - above[x]
+  // final pixels will be calculated as:
+  //   (above[x] * 32 + 16 + (above[x+1] - above[x]) * shift) >> 5
+  __m256i a0, a1, a32, a16;
+  __m256i a_mbase_x, diff, c3f;
+
+  a16 = _mm256_set1_epi16(16);
+  a_mbase_x = _mm256_set1_epi8((int8_t)above[max_base_x]);
+  c3f = _mm256_set1_epi16(0x3f);
+
+  int x = dx;
+  for (int r = 0; r < N; r++) {
+    __m256i b, res, res16[2];
+    __m128i a0_128, a1_128;
+
+    int base = x >> frac_bits;
+    int base_max_diff = (max_base_x - base);
+    if (base_max_diff <= 0) {
+      for (int i = r; i < N; ++i) {
+        dstvec[i] = a_mbase_x;  // save 32 values
+      }
+      return;
+    }
+    if (base_max_diff > 32) base_max_diff = 32;
+    __m256i shift =
+        _mm256_srli_epi16(_mm256_and_si256(_mm256_set1_epi16(x), c3f), 1);
+
+    for (int j = 0, jj = 0; j < 32; j += 16, jj++) {
+      int mdiff = base_max_diff - j;
+      if (mdiff <= 0) {
+        res16[jj] = a_mbase_x;
+      } else {
+        a0_128 = _mm_loadu_si128((__m128i *)(above + base + j));
+        a1_128 = _mm_loadu_si128((__m128i *)(above + base + j + 1));
+        a0 = _mm256_cvtepu8_epi16(a0_128);
+        a1 = _mm256_cvtepu8_epi16(a1_128);
+
+        diff = _mm256_sub_epi16(a1, a0);   // a[x+1] - a[x]
+        a32 = _mm256_slli_epi16(a0, 5);    // a[x] * 32
+        a32 = _mm256_add_epi16(a32, a16);  // a[x] * 32 + 16
+        b = _mm256_mullo_epi16(diff, shift);
+
+        res = _mm256_add_epi16(a32, b);
+        res = _mm256_srli_epi16(res, 5);
+        res16[jj] = _mm256_packus_epi16(
+            res, _mm256_castsi128_si256(
+                     _mm256_extracti128_si256(res, 1)));  // 16 8bit values
+      }
+    }
+    res16[1] =
+        _mm256_inserti128_si256(res16[0], _mm256_castsi256_si128(res16[1]),
+                                1);  // 32 8bit values
+
+    dstvec[r] = _mm256_blendv_epi8(
+        a_mbase_x, res16[1],
+        *(__m256i *)BaseMask[base_max_diff]);  // 32 8bit values
+    x += dx;
+  }
+}
+
+static void dr_prediction_z1_32xN_avx2(int N, uint8_t *dst, ptrdiff_t stride,
+                                       const uint8_t *above, int upsample_above,
+                                       int dx) {
+  __m256i dstvec[64];
+  dr_prediction_z1_32xN_internal_avx2(N, dstvec, above, upsample_above, dx);
+  for (int i = 0; i < N; i++) {
+    _mm256_storeu_si256((__m256i *)(dst + stride * i), dstvec[i]);
+  }
+}
+
+static void dr_prediction_z1_64xN_avx2(int N, uint8_t *dst, ptrdiff_t stride,
+                                       const uint8_t *above, int upsample_above,
+                                       int dx) {
+  // here upsample_above is 0 by design of av1_use_intra_edge_upsample
+  (void)upsample_above;
+  const int frac_bits = 6;
+  const int max_base_x = ((64 + N) - 1);
+
+  // pre-filter above pixels
+  // store in temp buffers:
+  //   above[x] * 32 + 16
+  //   above[x+1] - above[x]
+  // final pixels will be calculated as:
+  //   (above[x] * 32 + 16 + (above[x+1] - above[x]) * shift) >> 5
+  __m256i a0, a1, a32, a16;
+  __m256i a_mbase_x, diff, c3f;
+  __m128i max_base_x128, base_inc128, mask128;
+
+  a16 = _mm256_set1_epi16(16);
+  a_mbase_x = _mm256_set1_epi8((int8_t)above[max_base_x]);
+  max_base_x128 = _mm_set1_epi8(max_base_x);
+  c3f = _mm256_set1_epi16(0x3f);
+
+  int x = dx;
+  for (int r = 0; r < N; r++, dst += stride) {
+    __m256i b, res;
+    int base = x >> frac_bits;
+    if (base >= max_base_x) {
+      for (int i = r; i < N; ++i) {
+        _mm256_storeu_si256((__m256i *)dst, a_mbase_x);  // save 32 values
+        _mm256_storeu_si256((__m256i *)(dst + 32), a_mbase_x);
+        dst += stride;
+      }
+      return;
+    }
+
+    __m256i shift =
+        _mm256_srli_epi16(_mm256_and_si256(_mm256_set1_epi16(x), c3f), 1);
+
+    __m128i a0_128, a1_128, res128;
+    for (int j = 0; j < 64; j += 16) {
+      int mdif = max_base_x - (base + j);
+      if (mdif <= 0) {
+        _mm_storeu_si128((__m128i *)(dst + j),
+                         _mm256_castsi256_si128(a_mbase_x));
+      } else {
+        a0_128 = _mm_loadu_si128((__m128i *)(above + base + j));
+        a1_128 = _mm_loadu_si128((__m128i *)(above + base + 1 + j));
+        a0 = _mm256_cvtepu8_epi16(a0_128);
+        a1 = _mm256_cvtepu8_epi16(a1_128);
+
+        diff = _mm256_sub_epi16(a1, a0);   // a[x+1] - a[x]
+        a32 = _mm256_slli_epi16(a0, 5);    // a[x] * 32
+        a32 = _mm256_add_epi16(a32, a16);  // a[x] * 32 + 16
+        b = _mm256_mullo_epi16(diff, shift);
+
+        res = _mm256_add_epi16(a32, b);
+        res = _mm256_srli_epi16(res, 5);
+        res = _mm256_packus_epi16(
+            res, _mm256_castsi128_si256(
+                     _mm256_extracti128_si256(res, 1)));  // 16 8bit values
+
+        base_inc128 =
+            _mm_setr_epi8((int8_t)(base + j), (int8_t)(base + j + 1),
+                          (int8_t)(base + j + 2), (int8_t)(base + j + 3),
+                          (int8_t)(base + j + 4), (int8_t)(base + j + 5),
+                          (int8_t)(base + j + 6), (int8_t)(base + j + 7),
+                          (int8_t)(base + j + 8), (int8_t)(base + j + 9),
+                          (int8_t)(base + j + 10), (int8_t)(base + j + 11),
+                          (int8_t)(base + j + 12), (int8_t)(base + j + 13),
+                          (int8_t)(base + j + 14), (int8_t)(base + j + 15));
+
+        mask128 = _mm_cmpgt_epi8(_mm_subs_epu8(max_base_x128, base_inc128),
+                                 _mm_setzero_si128());
+        res128 = _mm_blendv_epi8(_mm256_castsi256_si128(a_mbase_x),
+                                 _mm256_castsi256_si128(res), mask128);
+        _mm_storeu_si128((__m128i *)(dst + j), res128);
+      }
+    }
+    x += dx;
+  }
+}
+
+// Directional prediction, zone 1: 0 < angle < 90
+void av1_dr_prediction_z1_avx2(uint8_t *dst, ptrdiff_t stride, int bw, int bh,
+                               const uint8_t *above, const uint8_t *left,
+                               int upsample_above, int dx, int dy) {
+  (void)left;
+  (void)dy;
+  switch (bw) {
+    case 4:
+      dr_prediction_z1_4xN_avx2(bh, dst, stride, above, upsample_above, dx);
+      break;
+    case 8:
+      dr_prediction_z1_8xN_avx2(bh, dst, stride, above, upsample_above, dx);
+      break;
+    case 16:
+      dr_prediction_z1_16xN_avx2(bh, dst, stride, above, upsample_above, dx);
+      break;
+    case 32:
+      dr_prediction_z1_32xN_avx2(bh, dst, stride, above, upsample_above, dx);
+      break;
+    case 64:
+      dr_prediction_z1_64xN_avx2(bh, dst, stride, above, upsample_above, dx);
+      break;
+    default: break;
+  }
+  return;
+}
+
+static void dr_prediction_z2_Nx4_avx2(int N, uint8_t *dst, ptrdiff_t stride,
+                                      const uint8_t *above, const uint8_t *left,
+                                      int upsample_above, int upsample_left,
+                                      int dx, int dy) {
+  const int min_base_x = -(1 << upsample_above);
+  const int min_base_y = -(1 << upsample_left);
+  const int frac_bits_x = 6 - upsample_above;
+  const int frac_bits_y = 6 - upsample_left;
+
+  assert(dx > 0);
+  // pre-filter above pixels
+  // store in temp buffers:
+  //   above[x] * 32 + 16
+  //   above[x+1] - above[x]
+  // final pixels will be calculated as:
+  //   (above[x] * 32 + 16 + (above[x+1] - above[x]) * shift) >> 5
+  __m128i a0_x, a1_x, a32, a16, diff;
+  __m128i c3f, min_base_y128, c1234, dy128;
+
+  a16 = _mm_set1_epi16(16);
+  c3f = _mm_set1_epi16(0x3f);
+  min_base_y128 = _mm_set1_epi16(min_base_y);
+  c1234 = _mm_setr_epi16(0, 1, 2, 3, 4, 0, 0, 0);
+  dy128 = _mm_set1_epi16(dy);
+
+  for (int r = 0; r < N; r++) {
+    __m128i b, res, shift, r6, ydx;
+    __m128i resx, resy, resxy;
+    __m128i a0_x128, a1_x128;
+    int y = r + 1;
+    int base_x = (-y * dx) >> frac_bits_x;
+    int base_shift = 0;
+    if (base_x < (min_base_x - 1)) {
+      base_shift = (min_base_x - base_x - 1) >> upsample_above;
+    }
+    int base_min_diff =
+        (min_base_x - base_x + upsample_above) >> upsample_above;
+    if (base_min_diff > 4) {
+      base_min_diff = 4;
+    } else {
+      if (base_min_diff < 0) base_min_diff = 0;
+    }
+
+    if (base_shift > 3) {
+      a0_x = _mm_setzero_si128();
+      a1_x = _mm_setzero_si128();
+      shift = _mm_setzero_si128();
+    } else {
+      a0_x128 = _mm_loadu_si128((__m128i *)(above + base_x + base_shift));
+      ydx = _mm_set1_epi16(y * dx);
+      r6 = _mm_slli_epi16(c1234, 6);
+
+      if (upsample_above) {
+        a0_x128 =
+            _mm_shuffle_epi8(a0_x128, *(__m128i *)EvenOddMaskx[base_shift]);
+        a1_x128 = _mm_srli_si128(a0_x128, 8);
+
+        shift = _mm_srli_epi16(
+            _mm_and_si128(
+                _mm_slli_epi16(_mm_sub_epi16(r6, ydx), upsample_above), c3f),
+            1);
+      } else {
+        a0_x128 = _mm_shuffle_epi8(a0_x128, *(__m128i *)LoadMaskx[base_shift]);
+        a1_x128 = _mm_srli_si128(a0_x128, 1);
+
+        shift = _mm_srli_epi16(_mm_and_si128(_mm_sub_epi16(r6, ydx), c3f), 1);
+      }
+      a0_x = _mm_cvtepu8_epi16(a0_x128);
+      a1_x = _mm_cvtepu8_epi16(a1_x128);
+    }
+    // y calc
+    __m128i a0_y, a1_y, shifty;
+    if (base_x < min_base_x) {
+      DECLARE_ALIGNED(32, int16_t, base_y_c[8]);
+      __m128i y_c128, base_y_c128, mask128, c1234_;
+      c1234_ = _mm_srli_si128(c1234, 2);
+      r6 = _mm_set1_epi16(r << 6);
+      y_c128 = _mm_sub_epi16(r6, _mm_mullo_epi16(c1234_, dy128));
+      base_y_c128 = _mm_srai_epi16(y_c128, frac_bits_y);
+      mask128 = _mm_cmpgt_epi16(min_base_y128, base_y_c128);
+      base_y_c128 = _mm_andnot_si128(mask128, base_y_c128);
+      _mm_store_si128((__m128i *)base_y_c, base_y_c128);
+
+      a0_y = _mm_setr_epi16(left[base_y_c[0]], left[base_y_c[1]],
+                            left[base_y_c[2]], left[base_y_c[3]], 0, 0, 0, 0);
+      base_y_c128 = _mm_add_epi16(base_y_c128, _mm_srli_epi16(a16, 4));
+      _mm_store_si128((__m128i *)base_y_c, base_y_c128);
+      a1_y = _mm_setr_epi16(left[base_y_c[0]], left[base_y_c[1]],
+                            left[base_y_c[2]], left[base_y_c[3]], 0, 0, 0, 0);
+
+      if (upsample_left) {
+        shifty = _mm_srli_epi16(
+            _mm_and_si128(_mm_slli_epi16(y_c128, upsample_left), c3f), 1);
+      } else {
+        shifty = _mm_srli_epi16(_mm_and_si128(y_c128, c3f), 1);
+      }
+      a0_x = _mm_unpacklo_epi64(a0_x, a0_y);
+      a1_x = _mm_unpacklo_epi64(a1_x, a1_y);
+      shift = _mm_unpacklo_epi64(shift, shifty);
+    }
+
+    diff = _mm_sub_epi16(a1_x, a0_x);  // a[x+1] - a[x]
+    a32 = _mm_slli_epi16(a0_x, 5);     // a[x] * 32
+    a32 = _mm_add_epi16(a32, a16);     // a[x] * 32 + 16
+
+    b = _mm_mullo_epi16(diff, shift);
+    res = _mm_add_epi16(a32, b);
+    res = _mm_srli_epi16(res, 5);
+
+    resx = _mm_packus_epi16(res, res);
+    resy = _mm_srli_si128(resx, 4);
+
+    resxy = _mm_blendv_epi8(resx, resy, *(__m128i *)BaseMask[base_min_diff]);
+    *(int *)(dst) = _mm_cvtsi128_si32(resxy);
+    dst += stride;
+  }
+}
+
+static void dr_prediction_z2_Nx8_avx2(int N, uint8_t *dst, ptrdiff_t stride,
+                                      const uint8_t *above, const uint8_t *left,
+                                      int upsample_above, int upsample_left,
+                                      int dx, int dy) {
+  const int min_base_x = -(1 << upsample_above);
+  const int min_base_y = -(1 << upsample_left);
+  const int frac_bits_x = 6 - upsample_above;
+  const int frac_bits_y = 6 - upsample_left;
+
+  // pre-filter above pixels
+  // store in temp buffers:
+  //   above[x] * 32 + 16
+  //   above[x+1] - above[x]
+  // final pixels will be calculated as:
+  //   (above[x] * 32 + 16 + (above[x+1] - above[x]) * shift) >> 5
+  __m256i diff, a32, a16;
+  __m256i a0_x, a1_x;
+  __m128i a0_x128, a1_x128, min_base_y128, c3f;
+  __m128i c1234, dy128;
+
+  a16 = _mm256_set1_epi16(16);
+  c3f = _mm_set1_epi16(0x3f);
+  min_base_y128 = _mm_set1_epi16(min_base_y);
+  dy128 = _mm_set1_epi16(dy);
+  c1234 = _mm_setr_epi16(1, 2, 3, 4, 5, 6, 7, 8);
+
+  for (int r = 0; r < N; r++) {
+    __m256i b, res, shift;
+    __m128i resx, resy, resxy, r6, ydx;
+
+    int y = r + 1;
+    int base_x = (-y * dx) >> frac_bits_x;
+    int base_shift = 0;
+    if (base_x < (min_base_x - 1)) {
+      base_shift = (min_base_x - base_x - 1) >> upsample_above;
+    }
+    int base_min_diff =
+        (min_base_x - base_x + upsample_above) >> upsample_above;
+    if (base_min_diff > 8) {
+      base_min_diff = 8;
+    } else {
+      if (base_min_diff < 0) base_min_diff = 0;
+    }
+
+    if (base_shift > 7) {
+      a0_x = _mm256_setzero_si256();
+      a1_x = _mm256_setzero_si256();
+      shift = _mm256_setzero_si256();
+    } else {
+      a0_x128 = _mm_loadu_si128((__m128i *)(above + base_x + base_shift));
+      ydx = _mm_set1_epi16(y * dx);
+      r6 = _mm_slli_epi16(_mm_srli_si128(c1234, 2), 6);
+      if (upsample_above) {
+        a0_x128 =
+            _mm_shuffle_epi8(a0_x128, *(__m128i *)EvenOddMaskx[base_shift]);
+        a1_x128 = _mm_srli_si128(a0_x128, 8);
+
+        shift = _mm256_castsi128_si256(_mm_srli_epi16(
+            _mm_and_si128(
+                _mm_slli_epi16(_mm_sub_epi16(r6, ydx), upsample_above), c3f),
+            1));
+      } else {
+        a1_x128 = _mm_srli_si128(a0_x128, 1);
+        a0_x128 = _mm_shuffle_epi8(a0_x128, *(__m128i *)LoadMaskx[base_shift]);
+        a1_x128 = _mm_shuffle_epi8(a1_x128, *(__m128i *)LoadMaskx[base_shift]);
+
+        shift = _mm256_castsi128_si256(
+            _mm_srli_epi16(_mm_and_si128(_mm_sub_epi16(r6, ydx), c3f), 1));
+      }
+      a0_x = _mm256_castsi128_si256(_mm_cvtepu8_epi16(a0_x128));
+      a1_x = _mm256_castsi128_si256(_mm_cvtepu8_epi16(a1_x128));
+    }
+
+    // y calc
+    __m128i a0_y, a1_y, shifty;
+    if (base_x < min_base_x) {
+      DECLARE_ALIGNED(32, int16_t, base_y_c[16]);
+      __m128i y_c128, base_y_c128, mask128;
+      r6 = _mm_set1_epi16(r << 6);
+      y_c128 = _mm_sub_epi16(r6, _mm_mullo_epi16(c1234, dy128));
+      base_y_c128 = _mm_srai_epi16(y_c128, frac_bits_y);
+      mask128 = _mm_cmpgt_epi16(min_base_y128, base_y_c128);
+      base_y_c128 = _mm_andnot_si128(mask128, base_y_c128);
+      _mm_store_si128((__m128i *)base_y_c, base_y_c128);
+
+      a0_y = _mm_setr_epi16(left[base_y_c[0]], left[base_y_c[1]],
+                            left[base_y_c[2]], left[base_y_c[3]],
+                            left[base_y_c[4]], left[base_y_c[5]],
+                            left[base_y_c[6]], left[base_y_c[7]]);
+      base_y_c128 = _mm_add_epi16(
+          base_y_c128, _mm_srli_epi16(_mm256_castsi256_si128(a16), 4));
+      _mm_store_si128((__m128i *)base_y_c, base_y_c128);
+
+      a1_y = _mm_setr_epi16(left[base_y_c[0]], left[base_y_c[1]],
+                            left[base_y_c[2]], left[base_y_c[3]],
+                            left[base_y_c[4]], left[base_y_c[5]],
+                            left[base_y_c[6]], left[base_y_c[7]]);
+
+      if (upsample_left) {
+        shifty = _mm_srli_epi16(
+            _mm_and_si128(_mm_slli_epi16(y_c128, upsample_left), c3f), 1);
+      } else {
+        shifty = _mm_srli_epi16(_mm_and_si128(y_c128, c3f), 1);
+      }
+
+      a0_x = _mm256_inserti128_si256(a0_x, a0_y, 1);
+      a1_x = _mm256_inserti128_si256(a1_x, a1_y, 1);
+      shift = _mm256_inserti128_si256(shift, shifty, 1);
+    }
+
+    diff = _mm256_sub_epi16(a1_x, a0_x);  // a[x+1] - a[x]
+    a32 = _mm256_slli_epi16(a0_x, 5);     // a[x] * 32
+    a32 = _mm256_add_epi16(a32, a16);     // a[x] * 32 + 16
+
+    b = _mm256_mullo_epi16(diff, shift);
+    res = _mm256_add_epi16(a32, b);
+    res = _mm256_srli_epi16(res, 5);
+
+    resx = _mm_packus_epi16(_mm256_castsi256_si128(res),
+                            _mm256_castsi256_si128(res));
+    resy = _mm256_extracti128_si256(res, 1);
+    resy = _mm_packus_epi16(resy, resy);
+
+    resxy = _mm_blendv_epi8(resx, resy, *(__m128i *)BaseMask[base_min_diff]);
+    _mm_storel_epi64((__m128i *)(dst), resxy);
+    dst += stride;
+  }
+}
+
+static void dr_prediction_z2_HxW_avx2(int H, int W, uint8_t *dst,
+                                      ptrdiff_t stride, const uint8_t *above,
+                                      const uint8_t *left, int upsample_above,
+                                      int upsample_left, int dx, int dy) {
+  // here upsample_above and upsample_left are 0 by design of
+  // av1_use_intra_edge_upsample
+  const int min_base_x = -1;
+  const int min_base_y = -1;
+  (void)upsample_above;
+  (void)upsample_left;
+  const int frac_bits_x = 6;
+  const int frac_bits_y = 6;
+
+  __m256i a0_x, a1_x, a0_y, a1_y, a32, a16, c1234, c0123;
+  __m256i diff, min_base_y256, c3f, shifty, dy256, c1;
+  __m128i a0_x128, a1_x128;
+
+  DECLARE_ALIGNED(32, int16_t, base_y_c[16]);
+  a16 = _mm256_set1_epi16(16);
+  c1 = _mm256_srli_epi16(a16, 4);
+  min_base_y256 = _mm256_set1_epi16(min_base_y);
+  c3f = _mm256_set1_epi16(0x3f);
+  dy256 = _mm256_set1_epi16(dy);
+  c0123 =
+      _mm256_setr_epi16(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15);
+  c1234 = _mm256_add_epi16(c0123, c1);
+
+  for (int r = 0; r < H; r++) {
+    __m256i b, res, shift, j256, r6, ydx;
+    __m128i resx, resy;
+    __m128i resxy;
+    int y = r + 1;
+    ydx = _mm256_set1_epi16((int16_t)(y * dx));
+
+    int base_x = (-y * dx) >> frac_bits_x;
+    for (int j = 0; j < W; j += 16) {
+      j256 = _mm256_set1_epi16(j);
+      int base_shift = 0;
+      if ((base_x + j) < (min_base_x - 1)) {
+        base_shift = (min_base_x - (base_x + j) - 1);
+      }
+      int base_min_diff = (min_base_x - base_x - j);
+      if (base_min_diff > 16) {
+        base_min_diff = 16;
+      } else {
+        if (base_min_diff < 0) base_min_diff = 0;
+      }
+
+      if (base_shift < 16) {
+        a0_x128 = _mm_loadu_si128((__m128i *)(above + base_x + base_shift + j));
+        a1_x128 =
+            _mm_loadu_si128((__m128i *)(above + base_x + base_shift + 1 + j));
+        a0_x128 = _mm_shuffle_epi8(a0_x128, *(__m128i *)LoadMaskx[base_shift]);
+        a1_x128 = _mm_shuffle_epi8(a1_x128, *(__m128i *)LoadMaskx[base_shift]);
+
+        a0_x = _mm256_cvtepu8_epi16(a0_x128);
+        a1_x = _mm256_cvtepu8_epi16(a1_x128);
+
+        r6 = _mm256_slli_epi16(_mm256_add_epi16(c0123, j256), 6);
+        shift = _mm256_srli_epi16(
+            _mm256_and_si256(_mm256_sub_epi16(r6, ydx), c3f), 1);
+
+        diff = _mm256_sub_epi16(a1_x, a0_x);  // a[x+1] - a[x]
+        a32 = _mm256_slli_epi16(a0_x, 5);     // a[x] * 32
+        a32 = _mm256_add_epi16(a32, a16);     // a[x] * 32 + 16
+
+        b = _mm256_mullo_epi16(diff, shift);
+        res = _mm256_add_epi16(a32, b);
+        res = _mm256_srli_epi16(res, 5);  // 16 16-bit values
+        resx = _mm256_castsi256_si128(_mm256_packus_epi16(
+            res, _mm256_castsi128_si256(_mm256_extracti128_si256(res, 1))));
+      } else {
+        resx = _mm_setzero_si128();
+      }
+
+      // y calc
+      if (base_x < min_base_x) {
+        __m256i c256, y_c256, base_y_c256, mask256, mul16;
+        r6 = _mm256_set1_epi16(r << 6);
+        c256 = _mm256_add_epi16(j256, c1234);
+        mul16 = _mm256_min_epu16(_mm256_mullo_epi16(c256, dy256),
+                                 _mm256_srli_epi16(min_base_y256, 1));
+        y_c256 = _mm256_sub_epi16(r6, mul16);
+
+        base_y_c256 = _mm256_srai_epi16(y_c256, frac_bits_y);
+        mask256 = _mm256_cmpgt_epi16(min_base_y256, base_y_c256);
+
+        base_y_c256 = _mm256_blendv_epi8(base_y_c256, min_base_y256, mask256);
+        int16_t min_y = (int16_t)_mm_extract_epi16(
+            _mm256_extracti128_si256(base_y_c256, 1), 7);
+        int16_t max_y =
+            (int16_t)_mm_extract_epi16(_mm256_castsi256_si128(base_y_c256), 0);
+        int16_t offset_diff = max_y - min_y;
+
+        if (offset_diff < 16) {
+          __m256i min_y256 = _mm256_set1_epi16(min_y);
+
+          __m256i base_y_offset = _mm256_sub_epi16(base_y_c256, min_y256);
+          __m128i base_y_offset128 =
+              _mm_packs_epi16(_mm256_extracti128_si256(base_y_offset, 0),
+                              _mm256_extracti128_si256(base_y_offset, 1));
+
+          __m128i a0_y128 = _mm_maskload_epi32(
+              (int *)(left + min_y), *(__m128i *)LoadMaskz2[offset_diff / 4]);
+          __m128i a1_y128 =
+              _mm_maskload_epi32((int *)(left + min_y + 1),
+                                 *(__m128i *)LoadMaskz2[offset_diff / 4]);
+          a0_y128 = _mm_shuffle_epi8(a0_y128, base_y_offset128);
+          a1_y128 = _mm_shuffle_epi8(a1_y128, base_y_offset128);
+          a0_y = _mm256_cvtepu8_epi16(a0_y128);
+          a1_y = _mm256_cvtepu8_epi16(a1_y128);
+        } else {
+          base_y_c256 = _mm256_andnot_si256(mask256, base_y_c256);
+          _mm256_store_si256((__m256i *)base_y_c, base_y_c256);
+
+          a0_y = _mm256_setr_epi16(
+              left[base_y_c[0]], left[base_y_c[1]], left[base_y_c[2]],
+              left[base_y_c[3]], left[base_y_c[4]], left[base_y_c[5]],
+              left[base_y_c[6]], left[base_y_c[7]], left[base_y_c[8]],
+              left[base_y_c[9]], left[base_y_c[10]], left[base_y_c[11]],
+              left[base_y_c[12]], left[base_y_c[13]], left[base_y_c[14]],
+              left[base_y_c[15]]);
+          base_y_c256 = _mm256_add_epi16(base_y_c256, c1);
+          _mm256_store_si256((__m256i *)base_y_c, base_y_c256);
+
+          a1_y = _mm256_setr_epi16(
+              left[base_y_c[0]], left[base_y_c[1]], left[base_y_c[2]],
+              left[base_y_c[3]], left[base_y_c[4]], left[base_y_c[5]],
+              left[base_y_c[6]], left[base_y_c[7]], left[base_y_c[8]],
+              left[base_y_c[9]], left[base_y_c[10]], left[base_y_c[11]],
+              left[base_y_c[12]], left[base_y_c[13]], left[base_y_c[14]],
+              left[base_y_c[15]]);
+        }
+        shifty = _mm256_srli_epi16(_mm256_and_si256(y_c256, c3f), 1);
+
+        diff = _mm256_sub_epi16(a1_y, a0_y);  // a[x+1] - a[x]
+        a32 = _mm256_slli_epi16(a0_y, 5);     // a[x] * 32
+        a32 = _mm256_add_epi16(a32, a16);     // a[x] * 32 + 16
+
+        b = _mm256_mullo_epi16(diff, shifty);
+        res = _mm256_add_epi16(a32, b);
+        res = _mm256_srli_epi16(res, 5);  // 16 16-bit values
+        resy = _mm256_castsi256_si128(_mm256_packus_epi16(
+            res, _mm256_castsi128_si256(_mm256_extracti128_si256(res, 1))));
+      } else {
+        resy = _mm_setzero_si128();
+      }
+      resxy = _mm_blendv_epi8(resx, resy, *(__m128i *)BaseMask[base_min_diff]);
+      _mm_storeu_si128((__m128i *)(dst + j), resxy);
+    }  // for j
+    dst += stride;
+  }
+}
+
+// Directional prediction, zone 2: 90 < angle < 180
+void av1_dr_prediction_z2_avx2(uint8_t *dst, ptrdiff_t stride, int bw, int bh,
+                               const uint8_t *above, const uint8_t *left,
+                               int upsample_above, int upsample_left, int dx,
+                               int dy) {
+  assert(dx > 0);
+  assert(dy > 0);
+  switch (bw) {
+    case 4:
+      dr_prediction_z2_Nx4_avx2(bh, dst, stride, above, left, upsample_above,
+                                upsample_left, dx, dy);
+      break;
+    case 8:
+      dr_prediction_z2_Nx8_avx2(bh, dst, stride, above, left, upsample_above,
+                                upsample_left, dx, dy);
+      break;
+    default:
+      dr_prediction_z2_HxW_avx2(bh, bw, dst, stride, above, left,
+                                upsample_above, upsample_left, dx, dy);
+      break;
+  }
+  return;
+}
+
+// z3 functions
+static INLINE void transpose16x32_avx2(__m256i *x, __m256i *d) {
+  __m256i w0, w1, w2, w3, w4, w5, w6, w7, w8, w9;
+  __m256i w10, w11, w12, w13, w14, w15;
+
+  w0 = _mm256_unpacklo_epi8(x[0], x[1]);
+  w1 = _mm256_unpacklo_epi8(x[2], x[3]);
+  w2 = _mm256_unpacklo_epi8(x[4], x[5]);
+  w3 = _mm256_unpacklo_epi8(x[6], x[7]);
+
+  w8 = _mm256_unpacklo_epi8(x[8], x[9]);
+  w9 = _mm256_unpacklo_epi8(x[10], x[11]);
+  w10 = _mm256_unpacklo_epi8(x[12], x[13]);
+  w11 = _mm256_unpacklo_epi8(x[14], x[15]);
+
+  w4 = _mm256_unpacklo_epi16(w0, w1);
+  w5 = _mm256_unpacklo_epi16(w2, w3);
+  w12 = _mm256_unpacklo_epi16(w8, w9);
+  w13 = _mm256_unpacklo_epi16(w10, w11);
+
+  w6 = _mm256_unpacklo_epi32(w4, w5);
+  w7 = _mm256_unpackhi_epi32(w4, w5);
+  w14 = _mm256_unpacklo_epi32(w12, w13);
+  w15 = _mm256_unpackhi_epi32(w12, w13);
+
+  // Store first 4-line result
+  d[0] = _mm256_unpacklo_epi64(w6, w14);
+  d[1] = _mm256_unpackhi_epi64(w6, w14);
+  d[2] = _mm256_unpacklo_epi64(w7, w15);
+  d[3] = _mm256_unpackhi_epi64(w7, w15);
+
+  w4 = _mm256_unpackhi_epi16(w0, w1);
+  w5 = _mm256_unpackhi_epi16(w2, w3);
+  w12 = _mm256_unpackhi_epi16(w8, w9);
+  w13 = _mm256_unpackhi_epi16(w10, w11);
+
+  w6 = _mm256_unpacklo_epi32(w4, w5);
+  w7 = _mm256_unpackhi_epi32(w4, w5);
+  w14 = _mm256_unpacklo_epi32(w12, w13);
+  w15 = _mm256_unpackhi_epi32(w12, w13);
+
+  // Store second 4-line result
+  d[4] = _mm256_unpacklo_epi64(w6, w14);
+  d[5] = _mm256_unpackhi_epi64(w6, w14);
+  d[6] = _mm256_unpacklo_epi64(w7, w15);
+  d[7] = _mm256_unpackhi_epi64(w7, w15);
+
+  // upper half
+  w0 = _mm256_unpackhi_epi8(x[0], x[1]);
+  w1 = _mm256_unpackhi_epi8(x[2], x[3]);
+  w2 = _mm256_unpackhi_epi8(x[4], x[5]);
+  w3 = _mm256_unpackhi_epi8(x[6], x[7]);
+
+  w8 = _mm256_unpackhi_epi8(x[8], x[9]);
+  w9 = _mm256_unpackhi_epi8(x[10], x[11]);
+  w10 = _mm256_unpackhi_epi8(x[12], x[13]);
+  w11 = _mm256_unpackhi_epi8(x[14], x[15]);
+
+  w4 = _mm256_unpacklo_epi16(w0, w1);
+  w5 = _mm256_unpacklo_epi16(w2, w3);
+  w12 = _mm256_unpacklo_epi16(w8, w9);
+  w13 = _mm256_unpacklo_epi16(w10, w11);
+
+  w6 = _mm256_unpacklo_epi32(w4, w5);
+  w7 = _mm256_unpackhi_epi32(w4, w5);
+  w14 = _mm256_unpacklo_epi32(w12, w13);
+  w15 = _mm256_unpackhi_epi32(w12, w13);
+
+  // Store first 4-line result
+  d[8] = _mm256_unpacklo_epi64(w6, w14);
+  d[9] = _mm256_unpackhi_epi64(w6, w14);
+  d[10] = _mm256_unpacklo_epi64(w7, w15);
+  d[11] = _mm256_unpackhi_epi64(w7, w15);
+
+  w4 = _mm256_unpackhi_epi16(w0, w1);
+  w5 = _mm256_unpackhi_epi16(w2, w3);
+  w12 = _mm256_unpackhi_epi16(w8, w9);
+  w13 = _mm256_unpackhi_epi16(w10, w11);
+
+  w6 = _mm256_unpacklo_epi32(w4, w5);
+  w7 = _mm256_unpackhi_epi32(w4, w5);
+  w14 = _mm256_unpacklo_epi32(w12, w13);
+  w15 = _mm256_unpackhi_epi32(w12, w13);
+
+  // Store second 4-line result
+  d[12] = _mm256_unpacklo_epi64(w6, w14);
+  d[13] = _mm256_unpackhi_epi64(w6, w14);
+  d[14] = _mm256_unpacklo_epi64(w7, w15);
+  d[15] = _mm256_unpackhi_epi64(w7, w15);
+}
+
+static void dr_prediction_z3_4x4_avx2(uint8_t *dst, ptrdiff_t stride,
+                                      const uint8_t *left, int upsample_left,
+                                      int dy) {
+  __m128i dstvec[4], d[4];
+
+  dr_prediction_z1_HxW_internal_avx2(4, 4, dstvec, left, upsample_left, dy);
+  transpose4x8_8x4_low_sse2(&dstvec[0], &dstvec[1], &dstvec[2], &dstvec[3],
+                            &d[0], &d[1], &d[2], &d[3]);
+
+  *(int *)(dst + stride * 0) = _mm_cvtsi128_si32(d[0]);
+  *(int *)(dst + stride * 1) = _mm_cvtsi128_si32(d[1]);
+  *(int *)(dst + stride * 2) = _mm_cvtsi128_si32(d[2]);
+  *(int *)(dst + stride * 3) = _mm_cvtsi128_si32(d[3]);
+  return;
+}
+
+static void dr_prediction_z3_8x8_avx2(uint8_t *dst, ptrdiff_t stride,
+                                      const uint8_t *left, int upsample_left,
+                                      int dy) {
+  __m128i dstvec[8], d[8];
+
+  dr_prediction_z1_HxW_internal_avx2(8, 8, dstvec, left, upsample_left, dy);
+  transpose8x8_sse2(&dstvec[0], &dstvec[1], &dstvec[2], &dstvec[3], &dstvec[4],
+                    &dstvec[5], &dstvec[6], &dstvec[7], &d[0], &d[1], &d[2],
+                    &d[3]);
+
+  _mm_storel_epi64((__m128i *)(dst + 0 * stride), d[0]);
+  _mm_storel_epi64((__m128i *)(dst + 1 * stride), _mm_srli_si128(d[0], 8));
+  _mm_storel_epi64((__m128i *)(dst + 2 * stride), d[1]);
+  _mm_storel_epi64((__m128i *)(dst + 3 * stride), _mm_srli_si128(d[1], 8));
+  _mm_storel_epi64((__m128i *)(dst + 4 * stride), d[2]);
+  _mm_storel_epi64((__m128i *)(dst + 5 * stride), _mm_srli_si128(d[2], 8));
+  _mm_storel_epi64((__m128i *)(dst + 6 * stride), d[3]);
+  _mm_storel_epi64((__m128i *)(dst + 7 * stride), _mm_srli_si128(d[3], 8));
+}
+
+static void dr_prediction_z3_4x8_avx2(uint8_t *dst, ptrdiff_t stride,
+                                      const uint8_t *left, int upsample_left,
+                                      int dy) {
+  __m128i dstvec[4], d[8];
+
+  dr_prediction_z1_HxW_internal_avx2(8, 4, dstvec, left, upsample_left, dy);
+  transpose4x8_8x4_sse2(&dstvec[0], &dstvec[1], &dstvec[2], &dstvec[3], &d[0],
+                        &d[1], &d[2], &d[3], &d[4], &d[5], &d[6], &d[7]);
+  for (int i = 0; i < 8; i++) {
+    *(int *)(dst + stride * i) = _mm_cvtsi128_si32(d[i]);
+  }
+}
+
+static void dr_prediction_z3_8x4_avx2(uint8_t *dst, ptrdiff_t stride,
+                                      const uint8_t *left, int upsample_left,
+                                      int dy) {
+  __m128i dstvec[8], d[4];
+
+  dr_prediction_z1_HxW_internal_avx2(4, 8, dstvec, left, upsample_left, dy);
+  transpose8x8_low_sse2(&dstvec[0], &dstvec[1], &dstvec[2], &dstvec[3],
+                        &dstvec[4], &dstvec[5], &dstvec[6], &dstvec[7], &d[0],
+                        &d[1], &d[2], &d[3]);
+  _mm_storel_epi64((__m128i *)(dst + 0 * stride), d[0]);
+  _mm_storel_epi64((__m128i *)(dst + 1 * stride), d[1]);
+  _mm_storel_epi64((__m128i *)(dst + 2 * stride), d[2]);
+  _mm_storel_epi64((__m128i *)(dst + 3 * stride), d[3]);
+}
+
+static void dr_prediction_z3_8x16_avx2(uint8_t *dst, ptrdiff_t stride,
+                                       const uint8_t *left, int upsample_left,
+                                       int dy) {
+  __m128i dstvec[8], d[8];
+
+  dr_prediction_z1_HxW_internal_avx2(16, 8, dstvec, left, upsample_left, dy);
+  transpose8x16_16x8_sse2(dstvec, dstvec + 1, dstvec + 2, dstvec + 3,
+                          dstvec + 4, dstvec + 5, dstvec + 6, dstvec + 7, d,
+                          d + 1, d + 2, d + 3, d + 4, d + 5, d + 6, d + 7);
+  for (int i = 0; i < 8; i++) {
+    _mm_storel_epi64((__m128i *)(dst + i * stride), d[i]);
+    _mm_storel_epi64((__m128i *)(dst + (i + 8) * stride),
+                     _mm_srli_si128(d[i], 8));
+  }
+}
+
+static void dr_prediction_z3_16x8_avx2(uint8_t *dst, ptrdiff_t stride,
+                                       const uint8_t *left, int upsample_left,
+                                       int dy) {
+  __m128i dstvec[16], d[16];
+
+  dr_prediction_z1_HxW_internal_avx2(8, 16, dstvec, left, upsample_left, dy);
+  transpose16x8_8x16_sse2(
+      &dstvec[0], &dstvec[1], &dstvec[2], &dstvec[3], &dstvec[4], &dstvec[5],
+      &dstvec[6], &dstvec[7], &dstvec[8], &dstvec[9], &dstvec[10], &dstvec[11],
+      &dstvec[12], &dstvec[13], &dstvec[14], &dstvec[15], &d[0], &d[1], &d[2],
+      &d[3], &d[4], &d[5], &d[6], &d[7]);
+
+  for (int i = 0; i < 8; i++) {
+    _mm_storeu_si128((__m128i *)(dst + i * stride), d[i]);
+  }
+}
+
+static void dr_prediction_z3_4x16_avx2(uint8_t *dst, ptrdiff_t stride,
+                                       const uint8_t *left, int upsample_left,
+                                       int dy) {
+  __m128i dstvec[4], d[16];
+
+  dr_prediction_z1_HxW_internal_avx2(16, 4, dstvec, left, upsample_left, dy);
+  transpose4x16_sse2(dstvec, d);
+  for (int i = 0; i < 16; i++) {
+    *(int *)(dst + stride * i) = _mm_cvtsi128_si32(d[i]);
+  }
+}
+
+static void dr_prediction_z3_16x4_avx2(uint8_t *dst, ptrdiff_t stride,
+                                       const uint8_t *left, int upsample_left,
+                                       int dy) {
+  __m128i dstvec[16], d[8];
+
+  dr_prediction_z1_HxW_internal_avx2(4, 16, dstvec, left, upsample_left, dy);
+  for (int i = 4; i < 8; i++) {
+    d[i] = _mm_setzero_si128();
+  }
+  transpose16x8_8x16_sse2(
+      &dstvec[0], &dstvec[1], &dstvec[2], &dstvec[3], &dstvec[4], &dstvec[5],
+      &dstvec[6], &dstvec[7], &dstvec[8], &dstvec[9], &dstvec[10], &dstvec[11],
+      &dstvec[12], &dstvec[13], &dstvec[14], &dstvec[15], &d[0], &d[1], &d[2],
+      &d[3], &d[4], &d[5], &d[6], &d[7]);
+
+  for (int i = 0; i < 4; i++) {
+    _mm_storeu_si128((__m128i *)(dst + i * stride), d[i]);
+  }
+}
+
+static void dr_prediction_z3_8x32_avx2(uint8_t *dst, ptrdiff_t stride,
+                                       const uint8_t *left, int upsample_left,
+                                       int dy) {
+  __m256i dstvec[16], d[16];
+
+  dr_prediction_z1_32xN_internal_avx2(8, dstvec, left, upsample_left, dy);
+  for (int i = 8; i < 16; i++) {
+    dstvec[i] = _mm256_setzero_si256();
+  }
+  transpose16x32_avx2(dstvec, d);
+
+  for (int i = 0; i < 16; i++) {
+    _mm_storel_epi64((__m128i *)(dst + i * stride),
+                     _mm256_castsi256_si128(d[i]));
+  }
+  for (int i = 0; i < 16; i++) {
+    _mm_storel_epi64((__m128i *)(dst + (i + 16) * stride),
+                     _mm256_extracti128_si256(d[i], 1));
+  }
+}
+
+static void dr_prediction_z3_32x8_avx2(uint8_t *dst, ptrdiff_t stride,
+                                       const uint8_t *left, int upsample_left,
+                                       int dy) {
+  __m128i dstvec[32], d[16];
+
+  dr_prediction_z1_HxW_internal_avx2(8, 32, dstvec, left, upsample_left, dy);
+
+  transpose16x8_8x16_sse2(
+      &dstvec[0], &dstvec[1], &dstvec[2], &dstvec[3], &dstvec[4], &dstvec[5],
+      &dstvec[6], &dstvec[7], &dstvec[8], &dstvec[9], &dstvec[10], &dstvec[11],
+      &dstvec[12], &dstvec[13], &dstvec[14], &dstvec[15], &d[0], &d[1], &d[2],
+      &d[3], &d[4], &d[5], &d[6], &d[7]);
+  transpose16x8_8x16_sse2(
+      &dstvec[0 + 16], &dstvec[1 + 16], &dstvec[2 + 16], &dstvec[3 + 16],
+      &dstvec[4 + 16], &dstvec[5 + 16], &dstvec[6 + 16], &dstvec[7 + 16],
+      &dstvec[8 + 16], &dstvec[9 + 16], &dstvec[10 + 16], &dstvec[11 + 16],
+      &dstvec[12 + 16], &dstvec[13 + 16], &dstvec[14 + 16], &dstvec[15 + 16],
+      &d[0 + 8], &d[1 + 8], &d[2 + 8], &d[3 + 8], &d[4 + 8], &d[5 + 8],
+      &d[6 + 8], &d[7 + 8]);
+
+  for (int i = 0; i < 8; i++) {
+    _mm_storeu_si128((__m128i *)(dst + i * stride), d[i]);
+    _mm_storeu_si128((__m128i *)(dst + i * stride + 16), d[i + 8]);
+  }
+}
+
+static void dr_prediction_z3_16x16_avx2(uint8_t *dst, ptrdiff_t stride,
+                                        const uint8_t *left, int upsample_left,
+                                        int dy) {
+  __m128i dstvec[16], d[16];
+
+  dr_prediction_z1_HxW_internal_avx2(16, 16, dstvec, left, upsample_left, dy);
+  transpose16x16_sse2(dstvec, d);
+
+  for (int i = 0; i < 16; i++) {
+    _mm_storeu_si128((__m128i *)(dst + i * stride), d[i]);
+  }
+}
+
+static void dr_prediction_z3_32x32_avx2(uint8_t *dst, ptrdiff_t stride,
+                                        const uint8_t *left, int upsample_left,
+                                        int dy) {
+  __m256i dstvec[32], d[32];
+
+  dr_prediction_z1_32xN_internal_avx2(32, dstvec, left, upsample_left, dy);
+  transpose16x32_avx2(dstvec, d);
+  transpose16x32_avx2(dstvec + 16, d + 16);
+  for (int j = 0; j < 16; j++) {
+    _mm_storeu_si128((__m128i *)(dst + j * stride),
+                     _mm256_castsi256_si128(d[j]));
+    _mm_storeu_si128((__m128i *)(dst + j * stride + 16),
+                     _mm256_castsi256_si128(d[j + 16]));
+  }
+  for (int j = 0; j < 16; j++) {
+    _mm_storeu_si128((__m128i *)(dst + (j + 16) * stride),
+                     _mm256_extracti128_si256(d[j], 1));
+    _mm_storeu_si128((__m128i *)(dst + (j + 16) * stride + 16),
+                     _mm256_extracti128_si256(d[j + 16], 1));
+  }
+}
+
+static void dr_prediction_z3_64x64_avx2(uint8_t *dst, ptrdiff_t stride,
+                                        const uint8_t *left, int upsample_left,
+                                        int dy) {
+  DECLARE_ALIGNED(16, uint8_t, dstT[64 * 64]);
+  dr_prediction_z1_64xN_avx2(64, dstT, 64, left, upsample_left, dy);
+  transpose(dstT, 64, dst, stride, 64, 64);
+}
+
+static void dr_prediction_z3_16x32_avx2(uint8_t *dst, ptrdiff_t stride,
+                                        const uint8_t *left, int upsample_left,
+                                        int dy) {
+  __m256i dstvec[16], d[16];
+
+  dr_prediction_z1_32xN_internal_avx2(16, dstvec, left, upsample_left, dy);
+  transpose16x32_avx2(dstvec, d);
+  // store
+  for (int j = 0; j < 16; j++) {
+    _mm_storeu_si128((__m128i *)(dst + j * stride),
+                     _mm256_castsi256_si128(d[j]));
+    _mm_storeu_si128((__m128i *)(dst + (j + 16) * stride),
+                     _mm256_extracti128_si256(d[j], 1));
+  }
+}
+
+static void dr_prediction_z3_32x16_avx2(uint8_t *dst, ptrdiff_t stride,
+                                        const uint8_t *left, int upsample_left,
+                                        int dy) {
+  __m128i dstvec[32], d[16];
+
+  dr_prediction_z1_HxW_internal_avx2(16, 32, dstvec, left, upsample_left, dy);
+  for (int i = 0; i < 32; i += 16) {
+    transpose16x16_sse2((dstvec + i), d);
+    for (int j = 0; j < 16; j++) {
+      _mm_storeu_si128((__m128i *)(dst + j * stride + i), d[j]);
+    }
+  }
+}
+
+static void dr_prediction_z3_32x64_avx2(uint8_t *dst, ptrdiff_t stride,
+                                        const uint8_t *left, int upsample_left,
+                                        int dy) {
+  uint8_t dstT[64 * 32];
+  dr_prediction_z1_64xN_avx2(32, dstT, 64, left, upsample_left, dy);
+  transpose(dstT, 64, dst, stride, 32, 64);
+}
+
+static void dr_prediction_z3_64x32_avx2(uint8_t *dst, ptrdiff_t stride,
+                                        const uint8_t *left, int upsample_left,
+                                        int dy) {
+  uint8_t dstT[32 * 64];
+  dr_prediction_z1_32xN_avx2(64, dstT, 32, left, upsample_left, dy);
+  transpose(dstT, 32, dst, stride, 64, 32);
+  return;
+}
+
+static void dr_prediction_z3_16x64_avx2(uint8_t *dst, ptrdiff_t stride,
+                                        const uint8_t *left, int upsample_left,
+                                        int dy) {
+  uint8_t dstT[64 * 16];
+  dr_prediction_z1_64xN_avx2(16, dstT, 64, left, upsample_left, dy);
+  transpose(dstT, 64, dst, stride, 16, 64);
+}
+
+static void dr_prediction_z3_64x16_avx2(uint8_t *dst, ptrdiff_t stride,
+                                        const uint8_t *left, int upsample_left,
+                                        int dy) {
+  __m128i dstvec[64], d[16];
+
+  dr_prediction_z1_HxW_internal_avx2(16, 64, dstvec, left, upsample_left, dy);
+  for (int i = 0; i < 64; i += 16) {
+    transpose16x16_sse2((dstvec + i), d);
+    for (int j = 0; j < 16; j++) {
+      _mm_storeu_si128((__m128i *)(dst + j * stride + i), d[j]);
+    }
+  }
+}
+
+void av1_dr_prediction_z3_avx2(uint8_t *dst, ptrdiff_t stride, int bw, int bh,
+                               const uint8_t *above, const uint8_t *left,
+                               int upsample_left, int dx, int dy) {
+  (void)above;
+  (void)dx;
+  assert(dx == 1);
+  assert(dy > 0);
+
+  if (bw == bh) {
+    switch (bw) {
+      case 4:
+        dr_prediction_z3_4x4_avx2(dst, stride, left, upsample_left, dy);
+        break;
+      case 8:
+        dr_prediction_z3_8x8_avx2(dst, stride, left, upsample_left, dy);
+        break;
+      case 16:
+        dr_prediction_z3_16x16_avx2(dst, stride, left, upsample_left, dy);
+        break;
+      case 32:
+        dr_prediction_z3_32x32_avx2(dst, stride, left, upsample_left, dy);
+        break;
+      case 64:
+        dr_prediction_z3_64x64_avx2(dst, stride, left, upsample_left, dy);
+        break;
+    }
+  } else {
+    if (bw < bh) {
+      if (bw + bw == bh) {
+        switch (bw) {
+          case 4:
+            dr_prediction_z3_4x8_avx2(dst, stride, left, upsample_left, dy);
+            break;
+          case 8:
+            dr_prediction_z3_8x16_avx2(dst, stride, left, upsample_left, dy);
+            break;
+          case 16:
+            dr_prediction_z3_16x32_avx2(dst, stride, left, upsample_left, dy);
+            break;
+          case 32:
+            dr_prediction_z3_32x64_avx2(dst, stride, left, upsample_left, dy);
+            break;
+        }
+      } else {
+        switch (bw) {
+          case 4:
+            dr_prediction_z3_4x16_avx2(dst, stride, left, upsample_left, dy);
+            break;
+          case 8:
+            dr_prediction_z3_8x32_avx2(dst, stride, left, upsample_left, dy);
+            break;
+          case 16:
+            dr_prediction_z3_16x64_avx2(dst, stride, left, upsample_left, dy);
+            break;
+        }
+      }
+    } else {
+      if (bh + bh == bw) {
+        switch (bh) {
+          case 4:
+            dr_prediction_z3_8x4_avx2(dst, stride, left, upsample_left, dy);
+            break;
+          case 8:
+            dr_prediction_z3_16x8_avx2(dst, stride, left, upsample_left, dy);
+            break;
+          case 16:
+            dr_prediction_z3_32x16_avx2(dst, stride, left, upsample_left, dy);
+            break;
+          case 32:
+            dr_prediction_z3_64x32_avx2(dst, stride, left, upsample_left, dy);
+            break;
+        }
+      } else {
+        switch (bh) {
+          case 4:
+            dr_prediction_z3_16x4_avx2(dst, stride, left, upsample_left, dy);
+            break;
+          case 8:
+            dr_prediction_z3_32x8_avx2(dst, stride, left, upsample_left, dy);
+            break;
+          case 16:
+            dr_prediction_z3_64x16_avx2(dst, stride, left, upsample_left, dy);
+            break;
+        }
+      }
+    }
+  }
+}
diff --git a/third_party/aom/aom_dsp/x86/intrapred_sse2.c b/third_party/aom/aom_dsp/x86/intrapred_sse2.c
new file mode 100644
index 0000000000..61e29731c4
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/intrapred_sse2.c
@@ -0,0 +1,1411 @@
+/*
+ * Copyright (c) 2017, 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 <emmintrin.h>
+#include "aom_dsp/x86/intrapred_x86.h"
+#include "config/aom_dsp_rtcd.h"
+
+static INLINE void dc_store_4xh(uint32_t dc, int height, uint8_t *dst,
+                                ptrdiff_t stride) {
+  for (int i = 0; i < height; i += 2) {
+    *(uint32_t *)dst = dc;
+    dst += stride;
+    *(uint32_t *)dst = dc;
+    dst += stride;
+  }
+}
+
+static INLINE void dc_store_8xh(const __m128i *row, int height, uint8_t *dst,
+                                ptrdiff_t stride) {
+  int i;
+  for (i = 0; i < height; ++i) {
+    _mm_storel_epi64((__m128i *)dst, *row);
+    dst += stride;
+  }
+}
+
+static INLINE void dc_store_16xh(const __m128i *row, int height, uint8_t *dst,
+                                 ptrdiff_t stride) {
+  int i;
+  for (i = 0; i < height; ++i) {
+    _mm_store_si128((__m128i *)dst, *row);
+    dst += stride;
+  }
+}
+
+static INLINE void dc_store_32xh(const __m128i *row, int height, uint8_t *dst,
+                                 ptrdiff_t stride) {
+  int i;
+  for (i = 0; i < height; ++i) {
+    _mm_store_si128((__m128i *)dst, *row);
+    _mm_store_si128((__m128i *)(dst + 16), *row);
+    dst += stride;
+  }
+}
+
+static INLINE void dc_store_64xh(const __m128i *row, int height, uint8_t *dst,
+                                 ptrdiff_t stride) {
+  for (int i = 0; i < height; ++i) {
+    _mm_store_si128((__m128i *)dst, *row);
+    _mm_store_si128((__m128i *)(dst + 16), *row);
+    _mm_store_si128((__m128i *)(dst + 32), *row);
+    _mm_store_si128((__m128i *)(dst + 48), *row);
+    dst += stride;
+  }
+}
+
+static INLINE __m128i dc_sum_4(const uint8_t *ref) {
+  __m128i x = _mm_loadl_epi64((__m128i const *)ref);
+  const __m128i zero = _mm_setzero_si128();
+  x = _mm_unpacklo_epi8(x, zero);
+  return _mm_sad_epu8(x, zero);
+}
+
+static INLINE __m128i dc_sum_8(const uint8_t *ref) {
+  __m128i x = _mm_loadl_epi64((__m128i const *)ref);
+  const __m128i zero = _mm_setzero_si128();
+  return _mm_sad_epu8(x, zero);
+}
+
+static INLINE __m128i dc_sum_64(const uint8_t *ref) {
+  __m128i x0 = _mm_load_si128((__m128i const *)ref);
+  __m128i x1 = _mm_load_si128((__m128i const *)(ref + 16));
+  __m128i x2 = _mm_load_si128((__m128i const *)(ref + 32));
+  __m128i x3 = _mm_load_si128((__m128i const *)(ref + 48));
+  const __m128i zero = _mm_setzero_si128();
+  x0 = _mm_sad_epu8(x0, zero);
+  x1 = _mm_sad_epu8(x1, zero);
+  x2 = _mm_sad_epu8(x2, zero);
+  x3 = _mm_sad_epu8(x3, zero);
+  x0 = _mm_add_epi16(x0, x1);
+  x2 = _mm_add_epi16(x2, x3);
+  x0 = _mm_add_epi16(x0, x2);
+  const __m128i high = _mm_unpackhi_epi64(x0, x0);
+  return _mm_add_epi16(x0, high);
+}
+
+#define DC_MULTIPLIER_1X2 0x5556
+#define DC_MULTIPLIER_1X4 0x3334
+
+#define DC_SHIFT2 16
+
+static INLINE int divide_using_multiply_shift(int num, int shift1,
+                                              int multiplier) {
+  const int interm = num >> shift1;
+  return interm * multiplier >> DC_SHIFT2;
+}
+
+// -----------------------------------------------------------------------------
+// DC_PRED
+
+void aom_dc_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t stride,
+                               const uint8_t *above, const uint8_t *left) {
+  const __m128i sum_left = dc_sum_8(left);
+  __m128i sum_above = dc_sum_4(above);
+  sum_above = _mm_add_epi16(sum_left, sum_above);
+
+  uint32_t sum = (uint32_t)_mm_cvtsi128_si32(sum_above);
+  sum += 6;
+  sum = divide_using_multiply_shift(sum, 2, DC_MULTIPLIER_1X2);
+
+  const __m128i row = _mm_set1_epi8((int8_t)sum);
+  const uint32_t pred = (uint32_t)_mm_cvtsi128_si32(row);
+  dc_store_4xh(pred, 8, dst, stride);
+}
+
+void aom_dc_predictor_4x16_sse2(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  const __m128i sum_left = dc_sum_16_sse2(left);
+  __m128i sum_above = dc_sum_4(above);
+  sum_above = _mm_add_epi16(sum_left, sum_above);
+
+  uint32_t sum = (uint32_t)_mm_cvtsi128_si32(sum_above);
+  sum += 10;
+  sum = divide_using_multiply_shift(sum, 2, DC_MULTIPLIER_1X4);
+
+  const __m128i row = _mm_set1_epi8((int8_t)sum);
+  const uint32_t pred = (uint32_t)_mm_cvtsi128_si32(row);
+  dc_store_4xh(pred, 16, dst, stride);
+}
+
+void aom_dc_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t stride,
+                               const uint8_t *above, const uint8_t *left) {
+  const __m128i sum_left = dc_sum_4(left);
+  __m128i sum_above = dc_sum_8(above);
+  sum_above = _mm_add_epi16(sum_above, sum_left);
+
+  uint32_t sum = (uint32_t)_mm_cvtsi128_si32(sum_above);
+  sum += 6;
+  sum = divide_using_multiply_shift(sum, 2, DC_MULTIPLIER_1X2);
+
+  const __m128i row = _mm_set1_epi8((int8_t)sum);
+  dc_store_8xh(&row, 4, dst, stride);
+}
+
+void aom_dc_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  const __m128i sum_left = dc_sum_16_sse2(left);
+  __m128i sum_above = dc_sum_8(above);
+  sum_above = _mm_add_epi16(sum_above, sum_left);
+
+  uint32_t sum = (uint32_t)_mm_cvtsi128_si32(sum_above);
+  sum += 12;
+  sum = divide_using_multiply_shift(sum, 3, DC_MULTIPLIER_1X2);
+  const __m128i row = _mm_set1_epi8((int8_t)sum);
+  dc_store_8xh(&row, 16, dst, stride);
+}
+
+void aom_dc_predictor_8x32_sse2(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  const __m128i sum_left = dc_sum_32_sse2(left);
+  __m128i sum_above = dc_sum_8(above);
+  sum_above = _mm_add_epi16(sum_above, sum_left);
+
+  uint32_t sum = (uint32_t)_mm_cvtsi128_si32(sum_above);
+  sum += 20;
+  sum = divide_using_multiply_shift(sum, 3, DC_MULTIPLIER_1X4);
+  const __m128i row = _mm_set1_epi8((int8_t)sum);
+  dc_store_8xh(&row, 32, dst, stride);
+}
+
+void aom_dc_predictor_16x4_sse2(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  const __m128i sum_left = dc_sum_4(left);
+  __m128i sum_above = dc_sum_16_sse2(above);
+  sum_above = _mm_add_epi16(sum_above, sum_left);
+
+  uint32_t sum = (uint32_t)_mm_cvtsi128_si32(sum_above);
+  sum += 10;
+  sum = divide_using_multiply_shift(sum, 2, DC_MULTIPLIER_1X4);
+  const __m128i row = _mm_set1_epi8((int8_t)sum);
+  dc_store_16xh(&row, 4, dst, stride);
+}
+
+void aom_dc_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  const __m128i sum_left = dc_sum_8(left);
+  __m128i sum_above = dc_sum_16_sse2(above);
+  sum_above = _mm_add_epi16(sum_above, sum_left);
+
+  uint32_t sum = (uint32_t)_mm_cvtsi128_si32(sum_above);
+  sum += 12;
+  sum = divide_using_multiply_shift(sum, 3, DC_MULTIPLIER_1X2);
+  const __m128i row = _mm_set1_epi8((int8_t)sum);
+  dc_store_16xh(&row, 8, dst, stride);
+}
+
+void aom_dc_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t stride,
+                                 const uint8_t *above, const uint8_t *left) {
+  const __m128i sum_left = dc_sum_32_sse2(left);
+  __m128i sum_above = dc_sum_16_sse2(above);
+  sum_above = _mm_add_epi16(sum_left, sum_above);
+
+  uint32_t sum = (uint32_t)_mm_cvtsi128_si32(sum_above);
+  sum += 24;
+  sum = divide_using_multiply_shift(sum, 4, DC_MULTIPLIER_1X2);
+  const __m128i row = _mm_set1_epi8((int8_t)sum);
+  dc_store_16xh(&row, 32, dst, stride);
+}
+
+void aom_dc_predictor_16x64_sse2(uint8_t *dst, ptrdiff_t stride,
+                                 const uint8_t *above, const uint8_t *left) {
+  const __m128i sum_left = dc_sum_64(left);
+  __m128i sum_above = dc_sum_16_sse2(above);
+  sum_above = _mm_add_epi16(sum_left, sum_above);
+
+  uint32_t sum = (uint32_t)_mm_cvtsi128_si32(sum_above);
+  sum += 40;
+  sum = divide_using_multiply_shift(sum, 4, DC_MULTIPLIER_1X4);
+  const __m128i row = _mm_set1_epi8((int8_t)sum);
+  dc_store_16xh(&row, 64, dst, stride);
+}
+
+void aom_dc_predictor_32x8_sse2(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  __m128i sum_above = dc_sum_32_sse2(above);
+  const __m128i sum_left = dc_sum_8(left);
+  sum_above = _mm_add_epi16(sum_above, sum_left);
+
+  uint32_t sum = (uint32_t)_mm_cvtsi128_si32(sum_above);
+  sum += 20;
+  sum = divide_using_multiply_shift(sum, 3, DC_MULTIPLIER_1X4);
+  const __m128i row = _mm_set1_epi8((int8_t)sum);
+  dc_store_32xh(&row, 8, dst, stride);
+}
+
+void aom_dc_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t stride,
+                                 const uint8_t *above, const uint8_t *left) {
+  __m128i sum_above = dc_sum_32_sse2(above);
+  const __m128i sum_left = dc_sum_16_sse2(left);
+  sum_above = _mm_add_epi16(sum_above, sum_left);
+
+  uint32_t sum = (uint32_t)_mm_cvtsi128_si32(sum_above);
+  sum += 24;
+  sum = divide_using_multiply_shift(sum, 4, DC_MULTIPLIER_1X2);
+  const __m128i row = _mm_set1_epi8((int8_t)sum);
+  dc_store_32xh(&row, 16, dst, stride);
+}
+
+void aom_dc_predictor_32x64_sse2(uint8_t *dst, ptrdiff_t stride,
+                                 const uint8_t *above, const uint8_t *left) {
+  __m128i sum_above = dc_sum_32_sse2(above);
+  const __m128i sum_left = dc_sum_64(left);
+  sum_above = _mm_add_epi16(sum_above, sum_left);
+
+  uint32_t sum = (uint32_t)_mm_cvtsi128_si32(sum_above);
+  sum += 48;
+  sum = divide_using_multiply_shift(sum, 5, DC_MULTIPLIER_1X2);
+  const __m128i row = _mm_set1_epi8((int8_t)sum);
+  dc_store_32xh(&row, 64, dst, stride);
+}
+
+void aom_dc_predictor_64x64_sse2(uint8_t *dst, ptrdiff_t stride,
+                                 const uint8_t *above, const uint8_t *left) {
+  __m128i sum_above = dc_sum_64(above);
+  const __m128i sum_left = dc_sum_64(left);
+  sum_above = _mm_add_epi16(sum_above, sum_left);
+
+  uint32_t sum = (uint32_t)_mm_cvtsi128_si32(sum_above);
+  sum += 64;
+  sum /= 128;
+  const __m128i row = _mm_set1_epi8((int8_t)sum);
+  dc_store_64xh(&row, 64, dst, stride);
+}
+
+void aom_dc_predictor_64x32_sse2(uint8_t *dst, ptrdiff_t stride,
+                                 const uint8_t *above, const uint8_t *left) {
+  __m128i sum_above = dc_sum_64(above);
+  const __m128i sum_left = dc_sum_32_sse2(left);
+  sum_above = _mm_add_epi16(sum_above, sum_left);
+
+  uint32_t sum = (uint32_t)_mm_cvtsi128_si32(sum_above);
+  sum += 48;
+  sum = divide_using_multiply_shift(sum, 5, DC_MULTIPLIER_1X2);
+  const __m128i row = _mm_set1_epi8((int8_t)sum);
+  dc_store_64xh(&row, 32, dst, stride);
+}
+
+void aom_dc_predictor_64x16_sse2(uint8_t *dst, ptrdiff_t stride,
+                                 const uint8_t *above, const uint8_t *left) {
+  __m128i sum_above = dc_sum_64(above);
+  const __m128i sum_left = dc_sum_16_sse2(left);
+  sum_above = _mm_add_epi16(sum_above, sum_left);
+
+  uint32_t sum = (uint32_t)_mm_cvtsi128_si32(sum_above);
+  sum += 40;
+  sum = divide_using_multiply_shift(sum, 4, DC_MULTIPLIER_1X4);
+  const __m128i row = _mm_set1_epi8((int8_t)sum);
+  dc_store_64xh(&row, 16, dst, stride);
+}
+
+// -----------------------------------------------------------------------------
+// DC_TOP
+
+void aom_dc_top_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t stride,
+                                   const uint8_t *above, const uint8_t *left) {
+  (void)left;
+  __m128i sum_above = dc_sum_4(above);
+  const __m128i two = _mm_set1_epi16(2);
+  sum_above = _mm_add_epi16(sum_above, two);
+  sum_above = _mm_srai_epi16(sum_above, 2);
+  sum_above = _mm_shufflelo_epi16(sum_above, 0);
+  sum_above = _mm_packus_epi16(sum_above, sum_above);
+
+  const uint32_t pred = (uint32_t)_mm_cvtsi128_si32(sum_above);
+  dc_store_4xh(pred, 8, dst, stride);
+}
+
+void aom_dc_top_predictor_4x16_sse2(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  (void)left;
+  __m128i sum_above = dc_sum_4(above);
+  const __m128i two = _mm_set1_epi16(2);
+  sum_above = _mm_add_epi16(sum_above, two);
+  sum_above = _mm_srai_epi16(sum_above, 2);
+  sum_above = _mm_shufflelo_epi16(sum_above, 0);
+  sum_above = _mm_packus_epi16(sum_above, sum_above);
+
+  const uint32_t pred = (uint32_t)_mm_cvtsi128_si32(sum_above);
+  dc_store_4xh(pred, 16, dst, stride);
+}
+
+void aom_dc_top_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t stride,
+                                   const uint8_t *above, const uint8_t *left) {
+  (void)left;
+  __m128i sum_above = dc_sum_8(above);
+  const __m128i four = _mm_set1_epi16(4);
+  sum_above = _mm_add_epi16(sum_above, four);
+  sum_above = _mm_srai_epi16(sum_above, 3);
+  sum_above = _mm_unpacklo_epi8(sum_above, sum_above);
+  const __m128i row = _mm_shufflelo_epi16(sum_above, 0);
+  dc_store_8xh(&row, 4, dst, stride);
+}
+
+void aom_dc_top_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  (void)left;
+  __m128i sum_above = dc_sum_8(above);
+  const __m128i four = _mm_set1_epi16(4);
+  sum_above = _mm_add_epi16(sum_above, four);
+  sum_above = _mm_srai_epi16(sum_above, 3);
+  sum_above = _mm_unpacklo_epi8(sum_above, sum_above);
+  const __m128i row = _mm_shufflelo_epi16(sum_above, 0);
+  dc_store_8xh(&row, 16, dst, stride);
+}
+
+void aom_dc_top_predictor_8x32_sse2(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  (void)left;
+  __m128i sum_above = dc_sum_8(above);
+  const __m128i four = _mm_set1_epi16(4);
+  sum_above = _mm_add_epi16(sum_above, four);
+  sum_above = _mm_srai_epi16(sum_above, 3);
+  sum_above = _mm_unpacklo_epi8(sum_above, sum_above);
+  const __m128i row = _mm_shufflelo_epi16(sum_above, 0);
+  dc_store_8xh(&row, 32, dst, stride);
+}
+
+void aom_dc_top_predictor_16x4_sse2(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  (void)left;
+  __m128i sum_above = dc_sum_16_sse2(above);
+  const __m128i eight = _mm_set1_epi16(8);
+  sum_above = _mm_add_epi16(sum_above, eight);
+  sum_above = _mm_srai_epi16(sum_above, 4);
+  sum_above = _mm_unpacklo_epi8(sum_above, sum_above);
+  sum_above = _mm_shufflelo_epi16(sum_above, 0);
+  const __m128i row = _mm_unpacklo_epi64(sum_above, sum_above);
+  dc_store_16xh(&row, 4, dst, stride);
+}
+
+void aom_dc_top_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  (void)left;
+  __m128i sum_above = dc_sum_16_sse2(above);
+  const __m128i eight = _mm_set1_epi16(8);
+  sum_above = _mm_add_epi16(sum_above, eight);
+  sum_above = _mm_srai_epi16(sum_above, 4);
+  sum_above = _mm_unpacklo_epi8(sum_above, sum_above);
+  sum_above = _mm_shufflelo_epi16(sum_above, 0);
+  const __m128i row = _mm_unpacklo_epi64(sum_above, sum_above);
+  dc_store_16xh(&row, 8, dst, stride);
+}
+
+void aom_dc_top_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)left;
+  __m128i sum_above = dc_sum_16_sse2(above);
+  const __m128i eight = _mm_set1_epi16(8);
+  sum_above = _mm_add_epi16(sum_above, eight);
+  sum_above = _mm_srai_epi16(sum_above, 4);
+  sum_above = _mm_unpacklo_epi8(sum_above, sum_above);
+  sum_above = _mm_shufflelo_epi16(sum_above, 0);
+  const __m128i row = _mm_unpacklo_epi64(sum_above, sum_above);
+  dc_store_16xh(&row, 32, dst, stride);
+}
+
+void aom_dc_top_predictor_16x64_sse2(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)left;
+  __m128i sum_above = dc_sum_16_sse2(above);
+  const __m128i eight = _mm_set1_epi16(8);
+  sum_above = _mm_add_epi16(sum_above, eight);
+  sum_above = _mm_srai_epi16(sum_above, 4);
+  sum_above = _mm_unpacklo_epi8(sum_above, sum_above);
+  sum_above = _mm_shufflelo_epi16(sum_above, 0);
+  const __m128i row = _mm_unpacklo_epi64(sum_above, sum_above);
+  dc_store_16xh(&row, 64, dst, stride);
+}
+
+void aom_dc_top_predictor_32x8_sse2(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  (void)left;
+  __m128i sum_above = dc_sum_32_sse2(above);
+  const __m128i sixteen = _mm_set1_epi16(16);
+  sum_above = _mm_add_epi16(sum_above, sixteen);
+  sum_above = _mm_srai_epi16(sum_above, 5);
+  sum_above = _mm_unpacklo_epi8(sum_above, sum_above);
+  sum_above = _mm_shufflelo_epi16(sum_above, 0);
+  const __m128i row = _mm_unpacklo_epi64(sum_above, sum_above);
+  dc_store_32xh(&row, 8, dst, stride);
+}
+
+void aom_dc_top_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)left;
+  __m128i sum_above = dc_sum_32_sse2(above);
+  const __m128i sixteen = _mm_set1_epi16(16);
+  sum_above = _mm_add_epi16(sum_above, sixteen);
+  sum_above = _mm_srai_epi16(sum_above, 5);
+  sum_above = _mm_unpacklo_epi8(sum_above, sum_above);
+  sum_above = _mm_shufflelo_epi16(sum_above, 0);
+  const __m128i row = _mm_unpacklo_epi64(sum_above, sum_above);
+  dc_store_32xh(&row, 16, dst, stride);
+}
+
+void aom_dc_top_predictor_32x64_sse2(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)left;
+  __m128i sum_above = dc_sum_32_sse2(above);
+  const __m128i sixteen = _mm_set1_epi16(16);
+  sum_above = _mm_add_epi16(sum_above, sixteen);
+  sum_above = _mm_srai_epi16(sum_above, 5);
+  sum_above = _mm_unpacklo_epi8(sum_above, sum_above);
+  sum_above = _mm_shufflelo_epi16(sum_above, 0);
+  const __m128i row = _mm_unpacklo_epi64(sum_above, sum_above);
+  dc_store_32xh(&row, 64, dst, stride);
+}
+
+void aom_dc_top_predictor_64x64_sse2(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)left;
+  __m128i sum_above = dc_sum_64(above);
+  const __m128i thirtytwo = _mm_set1_epi16(32);
+  sum_above = _mm_add_epi16(sum_above, thirtytwo);
+  sum_above = _mm_srai_epi16(sum_above, 6);
+  sum_above = _mm_unpacklo_epi8(sum_above, sum_above);
+  sum_above = _mm_shufflelo_epi16(sum_above, 0);
+  const __m128i row = _mm_unpacklo_epi64(sum_above, sum_above);
+  dc_store_64xh(&row, 64, dst, stride);
+}
+
+void aom_dc_top_predictor_64x32_sse2(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)left;
+  __m128i sum_above = dc_sum_64(above);
+  const __m128i thirtytwo = _mm_set1_epi16(32);
+  sum_above = _mm_add_epi16(sum_above, thirtytwo);
+  sum_above = _mm_srai_epi16(sum_above, 6);
+  sum_above = _mm_unpacklo_epi8(sum_above, sum_above);
+  sum_above = _mm_shufflelo_epi16(sum_above, 0);
+  const __m128i row = _mm_unpacklo_epi64(sum_above, sum_above);
+  dc_store_64xh(&row, 32, dst, stride);
+}
+
+void aom_dc_top_predictor_64x16_sse2(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)left;
+  __m128i sum_above = dc_sum_64(above);
+  const __m128i thirtytwo = _mm_set1_epi16(32);
+  sum_above = _mm_add_epi16(sum_above, thirtytwo);
+  sum_above = _mm_srai_epi16(sum_above, 6);
+  sum_above = _mm_unpacklo_epi8(sum_above, sum_above);
+  sum_above = _mm_shufflelo_epi16(sum_above, 0);
+  const __m128i row = _mm_unpacklo_epi64(sum_above, sum_above);
+  dc_store_64xh(&row, 16, dst, stride);
+}
+
+// -----------------------------------------------------------------------------
+// DC_LEFT
+
+void aom_dc_left_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  __m128i sum_left = dc_sum_8(left);
+  const __m128i four = _mm_set1_epi16(4);
+  sum_left = _mm_add_epi16(sum_left, four);
+  sum_left = _mm_srai_epi16(sum_left, 3);
+  sum_left = _mm_shufflelo_epi16(sum_left, 0);
+  sum_left = _mm_packus_epi16(sum_left, sum_left);
+
+  const uint32_t pred = (uint32_t)_mm_cvtsi128_si32(sum_left);
+  dc_store_4xh(pred, 8, dst, stride);
+}
+
+void aom_dc_left_predictor_4x16_sse2(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)above;
+  __m128i sum_left = dc_sum_16_sse2(left);
+  const __m128i eight = _mm_set1_epi16(8);
+  sum_left = _mm_add_epi16(sum_left, eight);
+  sum_left = _mm_srai_epi16(sum_left, 4);
+  sum_left = _mm_shufflelo_epi16(sum_left, 0);
+  sum_left = _mm_packus_epi16(sum_left, sum_left);
+
+  const uint32_t pred = (uint32_t)_mm_cvtsi128_si32(sum_left);
+  dc_store_4xh(pred, 16, dst, stride);
+}
+
+void aom_dc_left_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  __m128i sum_left = dc_sum_4(left);
+  const __m128i two = _mm_set1_epi16(2);
+  sum_left = _mm_add_epi16(sum_left, two);
+  sum_left = _mm_srai_epi16(sum_left, 2);
+  sum_left = _mm_unpacklo_epi8(sum_left, sum_left);
+  const __m128i row = _mm_shufflelo_epi16(sum_left, 0);
+  dc_store_8xh(&row, 4, dst, stride);
+}
+
+void aom_dc_left_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)above;
+  __m128i sum_left = dc_sum_16_sse2(left);
+  const __m128i eight = _mm_set1_epi16(8);
+  sum_left = _mm_add_epi16(sum_left, eight);
+  sum_left = _mm_srai_epi16(sum_left, 4);
+  sum_left = _mm_unpacklo_epi8(sum_left, sum_left);
+  const __m128i row = _mm_shufflelo_epi16(sum_left, 0);
+  dc_store_8xh(&row, 16, dst, stride);
+}
+
+void aom_dc_left_predictor_8x32_sse2(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)above;
+  __m128i sum_left = dc_sum_32_sse2(left);
+  const __m128i sixteen = _mm_set1_epi16(16);
+  sum_left = _mm_add_epi16(sum_left, sixteen);
+  sum_left = _mm_srai_epi16(sum_left, 5);
+  sum_left = _mm_unpacklo_epi8(sum_left, sum_left);
+  const __m128i row = _mm_shufflelo_epi16(sum_left, 0);
+  dc_store_8xh(&row, 32, dst, stride);
+}
+
+void aom_dc_left_predictor_16x4_sse2(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)above;
+  __m128i sum_left = dc_sum_4(left);
+  const __m128i two = _mm_set1_epi16(2);
+  sum_left = _mm_add_epi16(sum_left, two);
+  sum_left = _mm_srai_epi16(sum_left, 2);
+  sum_left = _mm_unpacklo_epi8(sum_left, sum_left);
+  sum_left = _mm_shufflelo_epi16(sum_left, 0);
+  const __m128i row = _mm_unpacklo_epi64(sum_left, sum_left);
+  dc_store_16xh(&row, 4, dst, stride);
+}
+
+void aom_dc_left_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)above;
+  __m128i sum_left = dc_sum_8(left);
+  const __m128i four = _mm_set1_epi16(4);
+  sum_left = _mm_add_epi16(sum_left, four);
+  sum_left = _mm_srai_epi16(sum_left, 3);
+  sum_left = _mm_unpacklo_epi8(sum_left, sum_left);
+  sum_left = _mm_shufflelo_epi16(sum_left, 0);
+  const __m128i row = _mm_unpacklo_epi64(sum_left, sum_left);
+  dc_store_16xh(&row, 8, dst, stride);
+}
+
+void aom_dc_left_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t stride,
+                                      const uint8_t *above,
+                                      const uint8_t *left) {
+  (void)above;
+  __m128i sum_left = dc_sum_32_sse2(left);
+  const __m128i sixteen = _mm_set1_epi16(16);
+  sum_left = _mm_add_epi16(sum_left, sixteen);
+  sum_left = _mm_srai_epi16(sum_left, 5);
+  sum_left = _mm_unpacklo_epi8(sum_left, sum_left);
+  sum_left = _mm_shufflelo_epi16(sum_left, 0);
+  const __m128i row = _mm_unpacklo_epi64(sum_left, sum_left);
+  dc_store_16xh(&row, 32, dst, stride);
+}
+
+void aom_dc_left_predictor_16x64_sse2(uint8_t *dst, ptrdiff_t stride,
+                                      const uint8_t *above,
+                                      const uint8_t *left) {
+  (void)above;
+  __m128i sum_left = dc_sum_64(left);
+  const __m128i thirtytwo = _mm_set1_epi16(32);
+  sum_left = _mm_add_epi16(sum_left, thirtytwo);
+  sum_left = _mm_srai_epi16(sum_left, 6);
+  sum_left = _mm_unpacklo_epi8(sum_left, sum_left);
+  sum_left = _mm_shufflelo_epi16(sum_left, 0);
+  const __m128i row = _mm_unpacklo_epi64(sum_left, sum_left);
+  dc_store_16xh(&row, 64, dst, stride);
+}
+
+void aom_dc_left_predictor_32x8_sse2(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)above;
+  __m128i sum_left = dc_sum_8(left);
+  const __m128i four = _mm_set1_epi16(4);
+  sum_left = _mm_add_epi16(sum_left, four);
+  sum_left = _mm_srai_epi16(sum_left, 3);
+  sum_left = _mm_unpacklo_epi8(sum_left, sum_left);
+  sum_left = _mm_shufflelo_epi16(sum_left, 0);
+  const __m128i row = _mm_unpacklo_epi64(sum_left, sum_left);
+  dc_store_32xh(&row, 8, dst, stride);
+}
+
+void aom_dc_left_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t stride,
+                                      const uint8_t *above,
+                                      const uint8_t *left) {
+  (void)above;
+  __m128i sum_left = dc_sum_16_sse2(left);
+  const __m128i eight = _mm_set1_epi16(8);
+  sum_left = _mm_add_epi16(sum_left, eight);
+  sum_left = _mm_srai_epi16(sum_left, 4);
+  sum_left = _mm_unpacklo_epi8(sum_left, sum_left);
+  sum_left = _mm_shufflelo_epi16(sum_left, 0);
+  const __m128i row = _mm_unpacklo_epi64(sum_left, sum_left);
+  dc_store_32xh(&row, 16, dst, stride);
+}
+
+void aom_dc_left_predictor_32x64_sse2(uint8_t *dst, ptrdiff_t stride,
+                                      const uint8_t *above,
+                                      const uint8_t *left) {
+  (void)above;
+  __m128i sum_left = dc_sum_64(left);
+  const __m128i thirtytwo = _mm_set1_epi16(32);
+  sum_left = _mm_add_epi16(sum_left, thirtytwo);
+  sum_left = _mm_srai_epi16(sum_left, 6);
+  sum_left = _mm_unpacklo_epi8(sum_left, sum_left);
+  sum_left = _mm_shufflelo_epi16(sum_left, 0);
+  const __m128i row = _mm_unpacklo_epi64(sum_left, sum_left);
+  dc_store_32xh(&row, 64, dst, stride);
+}
+
+void aom_dc_left_predictor_64x64_sse2(uint8_t *dst, ptrdiff_t stride,
+                                      const uint8_t *above,
+                                      const uint8_t *left) {
+  (void)above;
+  __m128i sum_left = dc_sum_64(left);
+  const __m128i thirtytwo = _mm_set1_epi16(32);
+  sum_left = _mm_add_epi16(sum_left, thirtytwo);
+  sum_left = _mm_srai_epi16(sum_left, 6);
+  sum_left = _mm_unpacklo_epi8(sum_left, sum_left);
+  sum_left = _mm_shufflelo_epi16(sum_left, 0);
+  const __m128i row = _mm_unpacklo_epi64(sum_left, sum_left);
+  dc_store_64xh(&row, 64, dst, stride);
+}
+
+void aom_dc_left_predictor_64x32_sse2(uint8_t *dst, ptrdiff_t stride,
+                                      const uint8_t *above,
+                                      const uint8_t *left) {
+  (void)above;
+  __m128i sum_left = dc_sum_32_sse2(left);
+  const __m128i sixteen = _mm_set1_epi16(16);
+  sum_left = _mm_add_epi16(sum_left, sixteen);
+  sum_left = _mm_srai_epi16(sum_left, 5);
+  sum_left = _mm_unpacklo_epi8(sum_left, sum_left);
+  sum_left = _mm_shufflelo_epi16(sum_left, 0);
+  const __m128i row = _mm_unpacklo_epi64(sum_left, sum_left);
+  dc_store_64xh(&row, 32, dst, stride);
+}
+
+void aom_dc_left_predictor_64x16_sse2(uint8_t *dst, ptrdiff_t stride,
+                                      const uint8_t *above,
+                                      const uint8_t *left) {
+  (void)above;
+  __m128i sum_left = dc_sum_16_sse2(left);
+  const __m128i eight = _mm_set1_epi16(8);
+  sum_left = _mm_add_epi16(sum_left, eight);
+  sum_left = _mm_srai_epi16(sum_left, 4);
+  sum_left = _mm_unpacklo_epi8(sum_left, sum_left);
+  sum_left = _mm_shufflelo_epi16(sum_left, 0);
+  const __m128i row = _mm_unpacklo_epi64(sum_left, sum_left);
+  dc_store_64xh(&row, 16, dst, stride);
+}
+
+// -----------------------------------------------------------------------------
+// DC_128
+
+void aom_dc_128_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t stride,
+                                   const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  (void)left;
+  const uint32_t pred = 0x80808080;
+  dc_store_4xh(pred, 8, dst, stride);
+}
+
+void aom_dc_128_predictor_4x16_sse2(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  (void)left;
+  const uint32_t pred = 0x80808080;
+  dc_store_4xh(pred, 16, dst, stride);
+}
+
+void aom_dc_128_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t stride,
+                                   const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  (void)left;
+  const __m128i row = _mm_set1_epi8((int8_t)128);
+  dc_store_8xh(&row, 4, dst, stride);
+}
+
+void aom_dc_128_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  (void)left;
+  const __m128i row = _mm_set1_epi8((int8_t)128);
+  dc_store_8xh(&row, 16, dst, stride);
+}
+
+void aom_dc_128_predictor_8x32_sse2(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  (void)left;
+  const __m128i row = _mm_set1_epi8((int8_t)128);
+  dc_store_8xh(&row, 32, dst, stride);
+}
+
+void aom_dc_128_predictor_16x4_sse2(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  (void)left;
+  const __m128i row = _mm_set1_epi8((int8_t)128);
+  dc_store_16xh(&row, 4, dst, stride);
+}
+
+void aom_dc_128_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  (void)left;
+  const __m128i row = _mm_set1_epi8((int8_t)128);
+  dc_store_16xh(&row, 8, dst, stride);
+}
+
+void aom_dc_128_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)above;
+  (void)left;
+  const __m128i row = _mm_set1_epi8((int8_t)128);
+  dc_store_16xh(&row, 32, dst, stride);
+}
+
+void aom_dc_128_predictor_16x64_sse2(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)above;
+  (void)left;
+  const __m128i row = _mm_set1_epi8((int8_t)128);
+  dc_store_16xh(&row, 64, dst, stride);
+}
+
+void aom_dc_128_predictor_32x8_sse2(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  (void)left;
+  const __m128i row = _mm_set1_epi8((int8_t)128);
+  dc_store_32xh(&row, 8, dst, stride);
+}
+
+void aom_dc_128_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)above;
+  (void)left;
+  const __m128i row = _mm_set1_epi8((int8_t)128);
+  dc_store_32xh(&row, 16, dst, stride);
+}
+
+void aom_dc_128_predictor_32x64_sse2(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)above;
+  (void)left;
+  const __m128i row = _mm_set1_epi8((int8_t)128);
+  dc_store_32xh(&row, 64, dst, stride);
+}
+
+void aom_dc_128_predictor_64x64_sse2(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)above;
+  (void)left;
+  const __m128i row = _mm_set1_epi8((int8_t)128);
+  dc_store_64xh(&row, 64, dst, stride);
+}
+
+void aom_dc_128_predictor_64x32_sse2(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)above;
+  (void)left;
+  const __m128i row = _mm_set1_epi8((int8_t)128);
+  dc_store_64xh(&row, 32, dst, stride);
+}
+
+void aom_dc_128_predictor_64x16_sse2(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)above;
+  (void)left;
+  const __m128i row = _mm_set1_epi8((int8_t)128);
+  dc_store_64xh(&row, 16, dst, stride);
+}
+
+// -----------------------------------------------------------------------------
+// V_PRED
+
+void aom_v_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left) {
+  const uint32_t pred = *(uint32_t *)above;
+  (void)left;
+  dc_store_4xh(pred, 8, dst, stride);
+}
+
+void aom_v_predictor_4x16_sse2(uint8_t *dst, ptrdiff_t stride,
+                               const uint8_t *above, const uint8_t *left) {
+  const uint32_t pred = *(uint32_t *)above;
+  (void)left;
+  dc_store_4xh(pred, 16, dst, stride);
+}
+
+void aom_v_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left) {
+  const __m128i row = _mm_loadl_epi64((__m128i const *)above);
+  (void)left;
+  dc_store_8xh(&row, 4, dst, stride);
+}
+
+void aom_v_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t stride,
+                               const uint8_t *above, const uint8_t *left) {
+  const __m128i row = _mm_loadl_epi64((__m128i const *)above);
+  (void)left;
+  dc_store_8xh(&row, 16, dst, stride);
+}
+
+void aom_v_predictor_8x32_sse2(uint8_t *dst, ptrdiff_t stride,
+                               const uint8_t *above, const uint8_t *left) {
+  const __m128i row = _mm_loadl_epi64((__m128i const *)above);
+  (void)left;
+  dc_store_8xh(&row, 32, dst, stride);
+}
+
+void aom_v_predictor_16x4_sse2(uint8_t *dst, ptrdiff_t stride,
+                               const uint8_t *above, const uint8_t *left) {
+  const __m128i row = _mm_load_si128((__m128i const *)above);
+  (void)left;
+  dc_store_16xh(&row, 4, dst, stride);
+}
+
+void aom_v_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t stride,
+                               const uint8_t *above, const uint8_t *left) {
+  const __m128i row = _mm_load_si128((__m128i const *)above);
+  (void)left;
+  dc_store_16xh(&row, 8, dst, stride);
+}
+
+void aom_v_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  const __m128i row = _mm_load_si128((__m128i const *)above);
+  (void)left;
+  dc_store_16xh(&row, 32, dst, stride);
+}
+
+void aom_v_predictor_16x64_sse2(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  const __m128i row = _mm_load_si128((__m128i const *)above);
+  (void)left;
+  dc_store_16xh(&row, 64, dst, stride);
+}
+
+static INLINE void v_predictor_32xh(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, int height) {
+  const __m128i row0 = _mm_load_si128((__m128i const *)above);
+  const __m128i row1 = _mm_load_si128((__m128i const *)(above + 16));
+  for (int i = 0; i < height; ++i) {
+    _mm_store_si128((__m128i *)dst, row0);
+    _mm_store_si128((__m128i *)(dst + 16), row1);
+    dst += stride;
+  }
+}
+
+void aom_v_predictor_32x8_sse2(uint8_t *dst, ptrdiff_t stride,
+                               const uint8_t *above, const uint8_t *left) {
+  (void)left;
+  v_predictor_32xh(dst, stride, above, 8);
+}
+
+void aom_v_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  (void)left;
+  v_predictor_32xh(dst, stride, above, 16);
+}
+
+void aom_v_predictor_32x64_sse2(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  (void)left;
+  v_predictor_32xh(dst, stride, above, 64);
+}
+
+static INLINE void v_predictor_64xh(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, int height) {
+  const __m128i row0 = _mm_load_si128((__m128i const *)above);
+  const __m128i row1 = _mm_load_si128((__m128i const *)(above + 16));
+  const __m128i row2 = _mm_load_si128((__m128i const *)(above + 32));
+  const __m128i row3 = _mm_load_si128((__m128i const *)(above + 48));
+  for (int i = 0; i < height; ++i) {
+    _mm_store_si128((__m128i *)dst, row0);
+    _mm_store_si128((__m128i *)(dst + 16), row1);
+    _mm_store_si128((__m128i *)(dst + 32), row2);
+    _mm_store_si128((__m128i *)(dst + 48), row3);
+    dst += stride;
+  }
+}
+
+void aom_v_predictor_64x64_sse2(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  (void)left;
+  v_predictor_64xh(dst, stride, above, 64);
+}
+
+void aom_v_predictor_64x32_sse2(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  (void)left;
+  v_predictor_64xh(dst, stride, above, 32);
+}
+
+void aom_v_predictor_64x16_sse2(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  (void)left;
+  v_predictor_64xh(dst, stride, above, 16);
+}
+
+// -----------------------------------------------------------------------------
+// H_PRED
+
+void aom_h_predictor_4x8_sse2(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  __m128i left_col = _mm_loadl_epi64((__m128i const *)left);
+  left_col = _mm_unpacklo_epi8(left_col, left_col);
+  __m128i row0 = _mm_shufflelo_epi16(left_col, 0);
+  __m128i row1 = _mm_shufflelo_epi16(left_col, 0x55);
+  __m128i row2 = _mm_shufflelo_epi16(left_col, 0xaa);
+  __m128i row3 = _mm_shufflelo_epi16(left_col, 0xff);
+  *(int *)dst = _mm_cvtsi128_si32(row0);
+  dst += stride;
+  *(int *)dst = _mm_cvtsi128_si32(row1);
+  dst += stride;
+  *(int *)dst = _mm_cvtsi128_si32(row2);
+  dst += stride;
+  *(int *)dst = _mm_cvtsi128_si32(row3);
+  dst += stride;
+  left_col = _mm_unpackhi_epi64(left_col, left_col);
+  row0 = _mm_shufflelo_epi16(left_col, 0);
+  row1 = _mm_shufflelo_epi16(left_col, 0x55);
+  row2 = _mm_shufflelo_epi16(left_col, 0xaa);
+  row3 = _mm_shufflelo_epi16(left_col, 0xff);
+  *(int *)dst = _mm_cvtsi128_si32(row0);
+  dst += stride;
+  *(int *)dst = _mm_cvtsi128_si32(row1);
+  dst += stride;
+  *(int *)dst = _mm_cvtsi128_si32(row2);
+  dst += stride;
+  *(int *)dst = _mm_cvtsi128_si32(row3);
+}
+
+void aom_h_predictor_4x16_sse2(uint8_t *dst, ptrdiff_t stride,
+                               const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  const __m128i left_col = _mm_load_si128((__m128i const *)left);
+  __m128i left_col_low = _mm_unpacklo_epi8(left_col, left_col);
+  __m128i left_col_high = _mm_unpackhi_epi8(left_col, left_col);
+
+  __m128i row0 = _mm_shufflelo_epi16(left_col_low, 0);
+  __m128i row1 = _mm_shufflelo_epi16(left_col_low, 0x55);
+  __m128i row2 = _mm_shufflelo_epi16(left_col_low, 0xaa);
+  __m128i row3 = _mm_shufflelo_epi16(left_col_low, 0xff);
+  *(int *)dst = _mm_cvtsi128_si32(row0);
+  dst += stride;
+  *(int *)dst = _mm_cvtsi128_si32(row1);
+  dst += stride;
+  *(int *)dst = _mm_cvtsi128_si32(row2);
+  dst += stride;
+  *(int *)dst = _mm_cvtsi128_si32(row3);
+  dst += stride;
+
+  left_col_low = _mm_unpackhi_epi64(left_col_low, left_col_low);
+  row0 = _mm_shufflelo_epi16(left_col_low, 0);
+  row1 = _mm_shufflelo_epi16(left_col_low, 0x55);
+  row2 = _mm_shufflelo_epi16(left_col_low, 0xaa);
+  row3 = _mm_shufflelo_epi16(left_col_low, 0xff);
+  *(int *)dst = _mm_cvtsi128_si32(row0);
+  dst += stride;
+  *(int *)dst = _mm_cvtsi128_si32(row1);
+  dst += stride;
+  *(int *)dst = _mm_cvtsi128_si32(row2);
+  dst += stride;
+  *(int *)dst = _mm_cvtsi128_si32(row3);
+  dst += stride;
+
+  row0 = _mm_shufflelo_epi16(left_col_high, 0);
+  row1 = _mm_shufflelo_epi16(left_col_high, 0x55);
+  row2 = _mm_shufflelo_epi16(left_col_high, 0xaa);
+  row3 = _mm_shufflelo_epi16(left_col_high, 0xff);
+  *(int *)dst = _mm_cvtsi128_si32(row0);
+  dst += stride;
+  *(int *)dst = _mm_cvtsi128_si32(row1);
+  dst += stride;
+  *(int *)dst = _mm_cvtsi128_si32(row2);
+  dst += stride;
+  *(int *)dst = _mm_cvtsi128_si32(row3);
+  dst += stride;
+
+  left_col_high = _mm_unpackhi_epi64(left_col_high, left_col_high);
+  row0 = _mm_shufflelo_epi16(left_col_high, 0);
+  row1 = _mm_shufflelo_epi16(left_col_high, 0x55);
+  row2 = _mm_shufflelo_epi16(left_col_high, 0xaa);
+  row3 = _mm_shufflelo_epi16(left_col_high, 0xff);
+  *(int *)dst = _mm_cvtsi128_si32(row0);
+  dst += stride;
+  *(int *)dst = _mm_cvtsi128_si32(row1);
+  dst += stride;
+  *(int *)dst = _mm_cvtsi128_si32(row2);
+  dst += stride;
+  *(int *)dst = _mm_cvtsi128_si32(row3);
+}
+
+void aom_h_predictor_8x4_sse2(uint8_t *dst, ptrdiff_t stride,
+                              const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  __m128i left_col = _mm_loadl_epi64((__m128i const *)left);
+  left_col = _mm_unpacklo_epi8(left_col, left_col);
+  __m128i row0 = _mm_shufflelo_epi16(left_col, 0);
+  __m128i row1 = _mm_shufflelo_epi16(left_col, 0x55);
+  __m128i row2 = _mm_shufflelo_epi16(left_col, 0xaa);
+  __m128i row3 = _mm_shufflelo_epi16(left_col, 0xff);
+  _mm_storel_epi64((__m128i *)dst, row0);
+  dst += stride;
+  _mm_storel_epi64((__m128i *)dst, row1);
+  dst += stride;
+  _mm_storel_epi64((__m128i *)dst, row2);
+  dst += stride;
+  _mm_storel_epi64((__m128i *)dst, row3);
+}
+
+static INLINE void h_predictor_8x16xc(uint8_t *dst, ptrdiff_t stride,
+                                      const uint8_t *above, const uint8_t *left,
+                                      int count) {
+  (void)above;
+  for (int i = 0; i < count; ++i) {
+    const __m128i left_col = _mm_load_si128((__m128i const *)left);
+    __m128i left_col_low = _mm_unpacklo_epi8(left_col, left_col);
+    __m128i left_col_high = _mm_unpackhi_epi8(left_col, left_col);
+
+    __m128i row0 = _mm_shufflelo_epi16(left_col_low, 0);
+    __m128i row1 = _mm_shufflelo_epi16(left_col_low, 0x55);
+    __m128i row2 = _mm_shufflelo_epi16(left_col_low, 0xaa);
+    __m128i row3 = _mm_shufflelo_epi16(left_col_low, 0xff);
+    _mm_storel_epi64((__m128i *)dst, row0);
+    dst += stride;
+    _mm_storel_epi64((__m128i *)dst, row1);
+    dst += stride;
+    _mm_storel_epi64((__m128i *)dst, row2);
+    dst += stride;
+    _mm_storel_epi64((__m128i *)dst, row3);
+    dst += stride;
+
+    left_col_low = _mm_unpackhi_epi64(left_col_low, left_col_low);
+    row0 = _mm_shufflelo_epi16(left_col_low, 0);
+    row1 = _mm_shufflelo_epi16(left_col_low, 0x55);
+    row2 = _mm_shufflelo_epi16(left_col_low, 0xaa);
+    row3 = _mm_shufflelo_epi16(left_col_low, 0xff);
+    _mm_storel_epi64((__m128i *)dst, row0);
+    dst += stride;
+    _mm_storel_epi64((__m128i *)dst, row1);
+    dst += stride;
+    _mm_storel_epi64((__m128i *)dst, row2);
+    dst += stride;
+    _mm_storel_epi64((__m128i *)dst, row3);
+    dst += stride;
+
+    row0 = _mm_shufflelo_epi16(left_col_high, 0);
+    row1 = _mm_shufflelo_epi16(left_col_high, 0x55);
+    row2 = _mm_shufflelo_epi16(left_col_high, 0xaa);
+    row3 = _mm_shufflelo_epi16(left_col_high, 0xff);
+    _mm_storel_epi64((__m128i *)dst, row0);
+    dst += stride;
+    _mm_storel_epi64((__m128i *)dst, row1);
+    dst += stride;
+    _mm_storel_epi64((__m128i *)dst, row2);
+    dst += stride;
+    _mm_storel_epi64((__m128i *)dst, row3);
+    dst += stride;
+
+    left_col_high = _mm_unpackhi_epi64(left_col_high, left_col_high);
+    row0 = _mm_shufflelo_epi16(left_col_high, 0);
+    row1 = _mm_shufflelo_epi16(left_col_high, 0x55);
+    row2 = _mm_shufflelo_epi16(left_col_high, 0xaa);
+    row3 = _mm_shufflelo_epi16(left_col_high, 0xff);
+    _mm_storel_epi64((__m128i *)dst, row0);
+    dst += stride;
+    _mm_storel_epi64((__m128i *)dst, row1);
+    dst += stride;
+    _mm_storel_epi64((__m128i *)dst, row2);
+    dst += stride;
+    _mm_storel_epi64((__m128i *)dst, row3);
+    dst += stride;
+    left += 16;
+  }
+}
+
+void aom_h_predictor_8x16_sse2(uint8_t *dst, ptrdiff_t stride,
+                               const uint8_t *above, const uint8_t *left) {
+  h_predictor_8x16xc(dst, stride, above, left, 1);
+}
+
+void aom_h_predictor_8x32_sse2(uint8_t *dst, ptrdiff_t stride,
+                               const uint8_t *above, const uint8_t *left) {
+  h_predictor_8x16xc(dst, stride, above, left, 2);
+}
+
+static INLINE void h_pred_store_16xh(const __m128i *row, int h, uint8_t *dst,
+                                     ptrdiff_t stride) {
+  int i;
+  for (i = 0; i < h; ++i) {
+    _mm_store_si128((__m128i *)dst, row[i]);
+    dst += stride;
+  }
+}
+
+static INLINE void repeat_low_4pixels(const __m128i *x, __m128i *row) {
+  const __m128i u0 = _mm_shufflelo_epi16(*x, 0);
+  const __m128i u1 = _mm_shufflelo_epi16(*x, 0x55);
+  const __m128i u2 = _mm_shufflelo_epi16(*x, 0xaa);
+  const __m128i u3 = _mm_shufflelo_epi16(*x, 0xff);
+
+  row[0] = _mm_unpacklo_epi64(u0, u0);
+  row[1] = _mm_unpacklo_epi64(u1, u1);
+  row[2] = _mm_unpacklo_epi64(u2, u2);
+  row[3] = _mm_unpacklo_epi64(u3, u3);
+}
+
+static INLINE void repeat_high_4pixels(const __m128i *x, __m128i *row) {
+  const __m128i u0 = _mm_shufflehi_epi16(*x, 0);
+  const __m128i u1 = _mm_shufflehi_epi16(*x, 0x55);
+  const __m128i u2 = _mm_shufflehi_epi16(*x, 0xaa);
+  const __m128i u3 = _mm_shufflehi_epi16(*x, 0xff);
+
+  row[0] = _mm_unpackhi_epi64(u0, u0);
+  row[1] = _mm_unpackhi_epi64(u1, u1);
+  row[2] = _mm_unpackhi_epi64(u2, u2);
+  row[3] = _mm_unpackhi_epi64(u3, u3);
+}
+
+// Process 16x8, first 4 rows
+// Use first 8 bytes of left register: xxxxxxxx33221100
+static INLINE void h_prediction_16x8_1(const __m128i *left, uint8_t *dst,
+                                       ptrdiff_t stride) {
+  __m128i row[4];
+  repeat_low_4pixels(left, row);
+  h_pred_store_16xh(row, 4, dst, stride);
+}
+
+// Process 16x8, second 4 rows
+// Use second 8 bytes of left register: 77665544xxxxxxxx
+static INLINE void h_prediction_16x8_2(const __m128i *left, uint8_t *dst,
+                                       ptrdiff_t stride) {
+  __m128i row[4];
+  repeat_high_4pixels(left, row);
+  h_pred_store_16xh(row, 4, dst, stride);
+}
+
+void aom_h_predictor_16x4_sse2(uint8_t *dst, ptrdiff_t stride,
+                               const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  const __m128i left_col = _mm_loadl_epi64((const __m128i *)left);
+  const __m128i left_col_8p = _mm_unpacklo_epi8(left_col, left_col);
+  h_prediction_16x8_1(&left_col_8p, dst, stride);
+}
+
+void aom_h_predictor_16x8_sse2(uint8_t *dst, ptrdiff_t stride,
+                               const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  const __m128i left_col = _mm_loadl_epi64((const __m128i *)left);
+  const __m128i left_col_8p = _mm_unpacklo_epi8(left_col, left_col);
+  h_prediction_16x8_1(&left_col_8p, dst, stride);
+  dst += stride << 2;
+  h_prediction_16x8_2(&left_col_8p, dst, stride);
+}
+
+static INLINE void h_predictor_16xh(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *left, int count) {
+  int i = 0;
+  do {
+    const __m128i left_col = _mm_load_si128((const __m128i *)left);
+    const __m128i left_col_8p_lo = _mm_unpacklo_epi8(left_col, left_col);
+    h_prediction_16x8_1(&left_col_8p_lo, dst, stride);
+    dst += stride << 2;
+    h_prediction_16x8_2(&left_col_8p_lo, dst, stride);
+    dst += stride << 2;
+
+    const __m128i left_col_8p_hi = _mm_unpackhi_epi8(left_col, left_col);
+    h_prediction_16x8_1(&left_col_8p_hi, dst, stride);
+    dst += stride << 2;
+    h_prediction_16x8_2(&left_col_8p_hi, dst, stride);
+    dst += stride << 2;
+
+    left += 16;
+    i++;
+  } while (i < count);
+}
+
+void aom_h_predictor_16x32_sse2(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  h_predictor_16xh(dst, stride, left, 2);
+}
+
+void aom_h_predictor_16x64_sse2(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  h_predictor_16xh(dst, stride, left, 4);
+}
+
+static INLINE void h_pred_store_32xh(const __m128i *row, int h, uint8_t *dst,
+                                     ptrdiff_t stride) {
+  int i;
+  for (i = 0; i < h; ++i) {
+    _mm_store_si128((__m128i *)dst, row[i]);
+    _mm_store_si128((__m128i *)(dst + 16), row[i]);
+    dst += stride;
+  }
+}
+
+// Process 32x8, first 4 rows
+// Use first 8 bytes of left register: xxxxxxxx33221100
+static INLINE void h_prediction_32x8_1(const __m128i *left, uint8_t *dst,
+                                       ptrdiff_t stride) {
+  __m128i row[4];
+  repeat_low_4pixels(left, row);
+  h_pred_store_32xh(row, 4, dst, stride);
+}
+
+// Process 32x8, second 4 rows
+// Use second 8 bytes of left register: 77665544xxxxxxxx
+static INLINE void h_prediction_32x8_2(const __m128i *left, uint8_t *dst,
+                                       ptrdiff_t stride) {
+  __m128i row[4];
+  repeat_high_4pixels(left, row);
+  h_pred_store_32xh(row, 4, dst, stride);
+}
+
+void aom_h_predictor_32x8_sse2(uint8_t *dst, ptrdiff_t stride,
+                               const uint8_t *above, const uint8_t *left) {
+  __m128i left_col, left_col_8p;
+  (void)above;
+
+  left_col = _mm_load_si128((const __m128i *)left);
+
+  left_col_8p = _mm_unpacklo_epi8(left_col, left_col);
+  h_prediction_32x8_1(&left_col_8p, dst, stride);
+  dst += stride << 2;
+  h_prediction_32x8_2(&left_col_8p, dst, stride);
+}
+
+void aom_h_predictor_32x16_sse2(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  __m128i left_col, left_col_8p;
+  (void)above;
+
+  left_col = _mm_load_si128((const __m128i *)left);
+
+  left_col_8p = _mm_unpacklo_epi8(left_col, left_col);
+  h_prediction_32x8_1(&left_col_8p, dst, stride);
+  dst += stride << 2;
+  h_prediction_32x8_2(&left_col_8p, dst, stride);
+  dst += stride << 2;
+
+  left_col_8p = _mm_unpackhi_epi8(left_col, left_col);
+  h_prediction_32x8_1(&left_col_8p, dst, stride);
+  dst += stride << 2;
+  h_prediction_32x8_2(&left_col_8p, dst, stride);
+}
+
+static INLINE void h_predictor_32xh(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *left, int height) {
+  int i = height >> 2;
+  do {
+    __m128i left4 = _mm_cvtsi32_si128(((int *)left)[0]);
+    left4 = _mm_unpacklo_epi8(left4, left4);
+    left4 = _mm_unpacklo_epi8(left4, left4);
+    const __m128i r0 = _mm_shuffle_epi32(left4, 0x0);
+    const __m128i r1 = _mm_shuffle_epi32(left4, 0x55);
+    _mm_store_si128((__m128i *)dst, r0);
+    _mm_store_si128((__m128i *)(dst + 16), r0);
+    _mm_store_si128((__m128i *)(dst + stride), r1);
+    _mm_store_si128((__m128i *)(dst + stride + 16), r1);
+    const __m128i r2 = _mm_shuffle_epi32(left4, 0xaa);
+    const __m128i r3 = _mm_shuffle_epi32(left4, 0xff);
+    _mm_store_si128((__m128i *)(dst + stride * 2), r2);
+    _mm_store_si128((__m128i *)(dst + stride * 2 + 16), r2);
+    _mm_store_si128((__m128i *)(dst + stride * 3), r3);
+    _mm_store_si128((__m128i *)(dst + stride * 3 + 16), r3);
+    left += 4;
+    dst += stride * 4;
+  } while (--i);
+}
+
+void aom_h_predictor_32x64_sse2(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  h_predictor_32xh(dst, stride, left, 64);
+}
+
+static INLINE void h_predictor_64xh(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *left, int height) {
+  int i = height >> 2;
+  do {
+    __m128i left4 = _mm_cvtsi32_si128(((int *)left)[0]);
+    left4 = _mm_unpacklo_epi8(left4, left4);
+    left4 = _mm_unpacklo_epi8(left4, left4);
+    const __m128i r0 = _mm_shuffle_epi32(left4, 0x0);
+    const __m128i r1 = _mm_shuffle_epi32(left4, 0x55);
+    _mm_store_si128((__m128i *)dst, r0);
+    _mm_store_si128((__m128i *)(dst + 16), r0);
+    _mm_store_si128((__m128i *)(dst + 32), r0);
+    _mm_store_si128((__m128i *)(dst + 48), r0);
+    _mm_store_si128((__m128i *)(dst + stride), r1);
+    _mm_store_si128((__m128i *)(dst + stride + 16), r1);
+    _mm_store_si128((__m128i *)(dst + stride + 32), r1);
+    _mm_store_si128((__m128i *)(dst + stride + 48), r1);
+    const __m128i r2 = _mm_shuffle_epi32(left4, 0xaa);
+    const __m128i r3 = _mm_shuffle_epi32(left4, 0xff);
+    _mm_store_si128((__m128i *)(dst + stride * 2), r2);
+    _mm_store_si128((__m128i *)(dst + stride * 2 + 16), r2);
+    _mm_store_si128((__m128i *)(dst + stride * 2 + 32), r2);
+    _mm_store_si128((__m128i *)(dst + stride * 2 + 48), r2);
+    _mm_store_si128((__m128i *)(dst + stride * 3), r3);
+    _mm_store_si128((__m128i *)(dst + stride * 3 + 16), r3);
+    _mm_store_si128((__m128i *)(dst + stride * 3 + 32), r3);
+    _mm_store_si128((__m128i *)(dst + stride * 3 + 48), r3);
+    left += 4;
+    dst += stride * 4;
+  } while (--i);
+}
+
+void aom_h_predictor_64x64_sse2(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  h_predictor_64xh(dst, stride, left, 64);
+}
+
+void aom_h_predictor_64x32_sse2(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  h_predictor_64xh(dst, stride, left, 32);
+}
+
+void aom_h_predictor_64x16_sse2(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  (void)above;
+  h_predictor_64xh(dst, stride, left, 16);
+}
diff --git a/third_party/aom/aom_dsp/x86/intrapred_sse4.c b/third_party/aom/aom_dsp/x86/intrapred_sse4.c
new file mode 100644
index 0000000000..9de8bf3c0f
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/intrapred_sse4.c
@@ -0,0 +1,1307 @@
+/*
+ * Copyright (c) 2021, 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 <emmintrin.h>  // SSE2
+#include <smmintrin.h>  /* SSE4.1 */
+
+#include "config/av1_rtcd.h"
+#include "aom_dsp/x86/intrapred_x86.h"
+#include "aom_dsp/x86/intrapred_utils.h"
+#include "aom_dsp/x86/lpf_common_sse2.h"
+
+// Low bit depth functions
+static DECLARE_ALIGNED(16, uint8_t, Mask[2][33][16]) = {
+  { { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+    { 0xff, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+    { 0xff, 0xff, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+    { 0xff, 0xff, 0xff, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+    { 0xff, 0xff, 0xff, 0xff, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+    { 0xff, 0xff, 0xff, 0xff, 0xff, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+    { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+    { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+    { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0, 0, 0, 0, 0, 0, 0, 0 },
+    { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0, 0, 0, 0, 0, 0,
+      0 },
+    { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0, 0, 0, 0, 0,
+      0 },
+    { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0, 0, 0,
+      0, 0 },
+    { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0,
+      0, 0, 0 },
+    { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+      0xff, 0, 0, 0 },
+    { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+      0xff, 0xff, 0, 0 },
+    { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+      0xff, 0xff, 0xff, 0 },
+    { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+      0xff, 0xff, 0xff, 0xff },
+    { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+      0xff, 0xff, 0xff, 0xff },
+    { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+      0xff, 0xff, 0xff, 0xff },
+    { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+      0xff, 0xff, 0xff, 0xff },
+    { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+      0xff, 0xff, 0xff, 0xff },
+    { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+      0xff, 0xff, 0xff, 0xff },
+    { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+      0xff, 0xff, 0xff, 0xff },
+    { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+      0xff, 0xff, 0xff, 0xff },
+    { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+      0xff, 0xff, 0xff, 0xff },
+    { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+      0xff, 0xff, 0xff, 0xff },
+    { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+      0xff, 0xff, 0xff, 0xff },
+    { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+      0xff, 0xff, 0xff, 0xff },
+    { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+      0xff, 0xff, 0xff, 0xff },
+    { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+      0xff, 0xff, 0xff, 0xff },
+    { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+      0xff, 0xff, 0xff, 0xff },
+    { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+      0xff, 0xff, 0xff, 0xff },
+    { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+      0xff, 0xff, 0xff, 0xff } },
+  {
+      { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+      { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+      { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+      { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+      { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+      { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+      { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+      { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+      { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+      { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+      { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+      { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+      { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+      { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+      { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+      { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+      { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+      { 0xff, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+      { 0xff, 0xff, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+      { 0xff, 0xff, 0xff, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+      { 0xff, 0xff, 0xff, 0xff, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+      { 0xff, 0xff, 0xff, 0xff, 0xff, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+      { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+      { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+      { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0, 0, 0, 0, 0, 0, 0,
+        0 },
+      { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0, 0, 0, 0, 0, 0,
+        0 },
+      { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0, 0, 0, 0,
+        0, 0 },
+      { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0, 0,
+        0, 0, 0 },
+      { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+        0, 0, 0, 0 },
+      { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+        0xff, 0, 0, 0 },
+      { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+        0xff, 0xff, 0, 0 },
+      { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+        0xff, 0xff, 0xff, 0 },
+      { 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+        0xff, 0xff, 0xff, 0xff },
+  },
+};
+
+/* clang-format on */
+static AOM_FORCE_INLINE void dr_prediction_z1_HxW_internal_sse4_1(
+    int H, int W, __m128i *dst, const uint8_t *above, int upsample_above,
+    int dx) {
+  const int frac_bits = 6 - upsample_above;
+  const int max_base_x = ((W + H) - 1) << upsample_above;
+
+  assert(dx > 0);
+  // pre-filter above pixels
+  // store in temp buffers:
+  //   above[x] * 32 + 16
+  //   above[x+1] - above[x]
+  // final pixels will be calculated as:
+  //   (above[x] * 32 + 16 + (above[x+1] - above[x]) * shift) >> 5
+  __m128i a0, a1, a32, a16;
+  __m128i diff, c3f;
+  __m128i a_mbase_x;
+
+  a16 = _mm_set1_epi16(16);
+  a_mbase_x = _mm_set1_epi8((char)above[max_base_x]);
+  c3f = _mm_set1_epi16(0x3f);
+
+  int x = dx;
+  for (int r = 0; r < W; r++) {
+    __m128i b, res, res1, shift;
+    __m128i a0_above, a1_above;
+
+    int base = x >> frac_bits;
+    int base_max_diff = (max_base_x - base) >> upsample_above;
+    if (base_max_diff <= 0) {
+      for (int i = r; i < W; ++i) {
+        dst[i] = a_mbase_x;  // save 4 values
+      }
+      return;
+    }
+    if (base_max_diff > H) base_max_diff = H;
+    a0_above = _mm_loadu_si128((__m128i *)(above + base));
+    a1_above = _mm_loadu_si128((__m128i *)(above + base + 1));
+
+    if (upsample_above) {
+      a0_above = _mm_shuffle_epi8(a0_above, *(__m128i *)EvenOddMaskx[0]);
+      a1_above = _mm_srli_si128(a0_above, 8);
+
+      shift = _mm_srli_epi16(
+          _mm_and_si128(_mm_slli_epi16(_mm_set1_epi16(x), upsample_above), c3f),
+          1);
+    } else {
+      shift = _mm_srli_epi16(_mm_and_si128(_mm_set1_epi16(x), c3f), 1);
+    }
+    // lower half
+    a0 = _mm_cvtepu8_epi16(a0_above);
+    a1 = _mm_cvtepu8_epi16(a1_above);
+
+    diff = _mm_sub_epi16(a1, a0);   // a[x+1] - a[x]
+    a32 = _mm_slli_epi16(a0, 5);    // a[x] * 32
+    a32 = _mm_add_epi16(a32, a16);  // a[x] * 32 + 16
+
+    b = _mm_mullo_epi16(diff, shift);
+    res = _mm_add_epi16(a32, b);
+    res = _mm_srli_epi16(res, 5);
+
+    // uppar half
+    a0 = _mm_cvtepu8_epi16(_mm_srli_si128(a0_above, 8));
+    a1 = _mm_cvtepu8_epi16(_mm_srli_si128(a1_above, 8));
+
+    diff = _mm_sub_epi16(a1, a0);   // a[x+1] - a[x]
+    a32 = _mm_slli_epi16(a0, 5);    // a[x] * 32
+    a32 = _mm_add_epi16(a32, a16);  // a[x] * 32 + 16
+
+    b = _mm_mullo_epi16(diff, shift);
+    res1 = _mm_add_epi16(a32, b);
+    res1 = _mm_srli_epi16(res1, 5);
+
+    res = _mm_packus_epi16(res, res1);
+
+    dst[r] =
+        _mm_blendv_epi8(a_mbase_x, res, *(__m128i *)Mask[0][base_max_diff]);
+    x += dx;
+  }
+}
+
+static void dr_prediction_z1_4xN_sse4_1(int N, uint8_t *dst, ptrdiff_t stride,
+                                        const uint8_t *above,
+                                        int upsample_above, int dx) {
+  __m128i dstvec[16];
+
+  dr_prediction_z1_HxW_internal_sse4_1(4, N, dstvec, above, upsample_above, dx);
+  for (int i = 0; i < N; i++) {
+    *(int *)(dst + stride * i) = _mm_cvtsi128_si32(dstvec[i]);
+  }
+}
+
+static void dr_prediction_z1_8xN_sse4_1(int N, uint8_t *dst, ptrdiff_t stride,
+                                        const uint8_t *above,
+                                        int upsample_above, int dx) {
+  __m128i dstvec[32];
+
+  dr_prediction_z1_HxW_internal_sse4_1(8, N, dstvec, above, upsample_above, dx);
+  for (int i = 0; i < N; i++) {
+    _mm_storel_epi64((__m128i *)(dst + stride * i), dstvec[i]);
+  }
+}
+
+static void dr_prediction_z1_16xN_sse4_1(int N, uint8_t *dst, ptrdiff_t stride,
+                                         const uint8_t *above,
+                                         int upsample_above, int dx) {
+  __m128i dstvec[64];
+
+  dr_prediction_z1_HxW_internal_sse4_1(16, N, dstvec, above, upsample_above,
+                                       dx);
+  for (int i = 0; i < N; i++) {
+    _mm_storeu_si128((__m128i *)(dst + stride * i), dstvec[i]);
+  }
+}
+
+static AOM_FORCE_INLINE void dr_prediction_z1_32xN_internal_sse4_1(
+    int N, __m128i *dstvec, __m128i *dstvec_h, const uint8_t *above,
+    int upsample_above, int dx) {
+  // here upsample_above is 0 by design of av1_use_intra_edge_upsample
+  (void)upsample_above;
+  const int frac_bits = 6;
+  const int max_base_x = ((32 + N) - 1);
+
+  // pre-filter above pixels
+  // store in temp buffers:
+  //   above[x] * 32 + 16
+  //   above[x+1] - above[x]
+  // final pixels will be calculated as:
+  //   (above[x] * 32 + 16 + (above[x+1] - above[x]) * shift) >> 5
+  __m128i a0, a1, a32, a16;
+  __m128i a_mbase_x, diff, c3f;
+
+  a16 = _mm_set1_epi16(16);
+  a_mbase_x = _mm_set1_epi8((char)above[max_base_x]);
+  c3f = _mm_set1_epi16(0x3f);
+
+  int x = dx;
+  for (int r = 0; r < N; r++) {
+    __m128i b, res, res1, res16[2];
+    __m128i a0_above, a1_above;
+
+    int base = x >> frac_bits;
+    int base_max_diff = (max_base_x - base);
+    if (base_max_diff <= 0) {
+      for (int i = r; i < N; ++i) {
+        dstvec[i] = a_mbase_x;  // save 32 values
+        dstvec_h[i] = a_mbase_x;
+      }
+      return;
+    }
+    if (base_max_diff > 32) base_max_diff = 32;
+    __m128i shift = _mm_srli_epi16(_mm_and_si128(_mm_set1_epi16(x), c3f), 1);
+
+    for (int j = 0, jj = 0; j < 32; j += 16, jj++) {
+      int mdiff = base_max_diff - j;
+      if (mdiff <= 0) {
+        res16[jj] = a_mbase_x;
+      } else {
+        a0_above = _mm_loadu_si128((__m128i *)(above + base + j));
+        a1_above = _mm_loadu_si128((__m128i *)(above + base + j + 1));
+
+        // lower half
+        a0 = _mm_cvtepu8_epi16(a0_above);
+        a1 = _mm_cvtepu8_epi16(a1_above);
+
+        diff = _mm_sub_epi16(a1, a0);   // a[x+1] - a[x]
+        a32 = _mm_slli_epi16(a0, 5);    // a[x] * 32
+        a32 = _mm_add_epi16(a32, a16);  // a[x] * 32 + 16
+        b = _mm_mullo_epi16(diff, shift);
+
+        res = _mm_add_epi16(a32, b);
+        res = _mm_srli_epi16(res, 5);
+
+        // uppar half
+        a0 = _mm_cvtepu8_epi16(_mm_srli_si128(a0_above, 8));
+        a1 = _mm_cvtepu8_epi16(_mm_srli_si128(a1_above, 8));
+
+        diff = _mm_sub_epi16(a1, a0);   // a[x+1] - a[x]
+        a32 = _mm_slli_epi16(a0, 5);    // a[x] * 32
+        a32 = _mm_add_epi16(a32, a16);  // a[x] * 32 + 16
+
+        b = _mm_mullo_epi16(diff, shift);
+        res1 = _mm_add_epi16(a32, b);
+        res1 = _mm_srli_epi16(res1, 5);
+
+        res16[jj] = _mm_packus_epi16(res, res1);  // 16 8bit values
+      }
+    }
+
+    dstvec[r] =
+        _mm_blendv_epi8(a_mbase_x, res16[0],
+                        *(__m128i *)Mask[0][base_max_diff]);  // 16 8bit values
+
+    dstvec_h[r] =
+        _mm_blendv_epi8(a_mbase_x, res16[1],
+                        *(__m128i *)Mask[1][base_max_diff]);  // 16 8bit values
+    x += dx;
+  }
+}
+
+static void dr_prediction_z1_32xN_sse4_1(int N, uint8_t *dst, ptrdiff_t stride,
+                                         const uint8_t *above,
+                                         int upsample_above, int dx) {
+  __m128i dstvec[64], dstvec_h[64];
+  dr_prediction_z1_32xN_internal_sse4_1(N, dstvec, dstvec_h, above,
+                                        upsample_above, dx);
+  for (int i = 0; i < N; i++) {
+    _mm_storeu_si128((__m128i *)(dst + stride * i), dstvec[i]);
+    _mm_storeu_si128((__m128i *)(dst + stride * i + 16), dstvec_h[i]);
+  }
+}
+
+static void dr_prediction_z1_64xN_sse4_1(int N, uint8_t *dst, ptrdiff_t stride,
+                                         const uint8_t *above,
+                                         int upsample_above, int dx) {
+  // here upsample_above is 0 by design of av1_use_intra_edge_upsample
+  (void)upsample_above;
+  const int frac_bits = 6;
+  const int max_base_x = ((64 + N) - 1);
+
+  // pre-filter above pixels
+  // store in temp buffers:
+  //   above[x] * 32 + 16
+  //   above[x+1] - above[x]
+  // final pixels will be calculated as:
+  //   (above[x] * 32 + 16 + (above[x+1] - above[x]) * shift) >> 5
+  __m128i a0, a1, a32, a16;
+  __m128i a_mbase_x, diff, c3f;
+  __m128i max_base, base_inc, mask;
+
+  a16 = _mm_set1_epi16(16);
+  a_mbase_x = _mm_set1_epi8((char)above[max_base_x]);
+  max_base = _mm_set1_epi8(max_base_x);
+  c3f = _mm_set1_epi16(0x3f);
+
+  int x = dx;
+  for (int r = 0; r < N; r++, dst += stride) {
+    __m128i b, res, res1;
+    int base = x >> frac_bits;
+    if (base >= max_base_x) {
+      for (int i = r; i < N; ++i) {
+        _mm_storeu_si128((__m128i *)dst, a_mbase_x);  // save 32 values
+        _mm_storeu_si128((__m128i *)(dst + 16), a_mbase_x);
+        _mm_storeu_si128((__m128i *)(dst + 32), a_mbase_x);
+        _mm_storeu_si128((__m128i *)(dst + 48), a_mbase_x);
+        dst += stride;
+      }
+      return;
+    }
+
+    __m128i shift =
+        _mm_srli_epi16(_mm_and_si128(_mm_set1_epi16(x), c3f), 1);  // 8 element
+
+    __m128i a0_above, a1_above, res_val;
+    for (int j = 0; j < 64; j += 16) {
+      int mdif = max_base_x - (base + j);
+      if (mdif <= 0) {
+        _mm_storeu_si128((__m128i *)(dst + j), a_mbase_x);
+      } else {
+        a0_above =
+            _mm_loadu_si128((__m128i *)(above + base + j));  // load 16 element
+        a1_above = _mm_loadu_si128((__m128i *)(above + base + 1 + j));
+
+        // lower half
+        a0 = _mm_cvtepu8_epi16(a0_above);
+        a1 = _mm_cvtepu8_epi16(a1_above);
+
+        diff = _mm_sub_epi16(a1, a0);   // a[x+1] - a[x]
+        a32 = _mm_slli_epi16(a0, 5);    // a[x] * 32
+        a32 = _mm_add_epi16(a32, a16);  // a[x] * 32 + 16
+        b = _mm_mullo_epi16(diff, shift);
+
+        res = _mm_add_epi16(a32, b);
+        res = _mm_srli_epi16(res, 5);
+
+        // uppar half
+        a0 = _mm_cvtepu8_epi16(_mm_srli_si128(a0_above, 8));
+        a1 = _mm_cvtepu8_epi16(_mm_srli_si128(a1_above, 8));
+
+        diff = _mm_sub_epi16(a1, a0);   // a[x+1] - a[x]
+        a32 = _mm_slli_epi16(a0, 5);    // a[x] * 32
+        a32 = _mm_add_epi16(a32, a16);  // a[x] * 32 + 16
+
+        b = _mm_mullo_epi16(diff, shift);
+        res1 = _mm_add_epi16(a32, b);
+        res1 = _mm_srli_epi16(res1, 5);
+
+        res = _mm_packus_epi16(res, res1);  // 16 8bit values
+
+        base_inc =
+            _mm_setr_epi8((int8_t)(base + j), (int8_t)(base + j + 1),
+                          (int8_t)(base + j + 2), (int8_t)(base + j + 3),
+                          (int8_t)(base + j + 4), (int8_t)(base + j + 5),
+                          (int8_t)(base + j + 6), (int8_t)(base + j + 7),
+                          (int8_t)(base + j + 8), (int8_t)(base + j + 9),
+                          (int8_t)(base + j + 10), (int8_t)(base + j + 11),
+                          (int8_t)(base + j + 12), (int8_t)(base + j + 13),
+                          (int8_t)(base + j + 14), (int8_t)(base + j + 15));
+
+        mask = _mm_cmpgt_epi8(_mm_subs_epu8(max_base, base_inc),
+                              _mm_setzero_si128());
+        res_val = _mm_blendv_epi8(a_mbase_x, res, mask);
+        _mm_storeu_si128((__m128i *)(dst + j), res_val);
+      }
+    }
+    x += dx;
+  }
+}
+
+// Directional prediction, zone 1: 0 < angle < 90
+void av1_dr_prediction_z1_sse4_1(uint8_t *dst, ptrdiff_t stride, int bw, int bh,
+                                 const uint8_t *above, const uint8_t *left,
+                                 int upsample_above, int dx, int dy) {
+  (void)left;
+  (void)dy;
+  switch (bw) {
+    case 4:
+      dr_prediction_z1_4xN_sse4_1(bh, dst, stride, above, upsample_above, dx);
+      break;
+    case 8:
+      dr_prediction_z1_8xN_sse4_1(bh, dst, stride, above, upsample_above, dx);
+      break;
+    case 16:
+      dr_prediction_z1_16xN_sse4_1(bh, dst, stride, above, upsample_above, dx);
+      break;
+    case 32:
+      dr_prediction_z1_32xN_sse4_1(bh, dst, stride, above, upsample_above, dx);
+      break;
+    case 64:
+      dr_prediction_z1_64xN_sse4_1(bh, dst, stride, above, upsample_above, dx);
+      break;
+    default: assert(0 && "Invalid block size");
+  }
+  return;
+}
+
+static void dr_prediction_z2_Nx4_sse4_1(int N, uint8_t *dst, ptrdiff_t stride,
+                                        const uint8_t *above,
+                                        const uint8_t *left, int upsample_above,
+                                        int upsample_left, int dx, int dy) {
+  const int min_base_x = -(1 << upsample_above);
+  const int min_base_y = -(1 << upsample_left);
+  const int frac_bits_x = 6 - upsample_above;
+  const int frac_bits_y = 6 - upsample_left;
+
+  assert(dx > 0);
+  // pre-filter above pixels
+  // store in temp buffers:
+  //   above[x] * 32 + 16
+  //   above[x+1] - above[x]
+  // final pixels will be calculated as:
+  //   (above[x] * 32 + 16 + (above[x+1] - above[x]) * shift) >> 5
+  __m128i a0_x, a1_x, a32, diff;
+
+  const __m128i c3f = _mm_set1_epi16(0x3f);
+  const __m128i min_y_base = _mm_set1_epi16(min_base_y);
+  const __m128i c1234 = _mm_setr_epi16(0, 1, 2, 3, 4, 0, 0, 0);
+  const __m128i dy_reg = _mm_set1_epi16(dy);
+  const __m128i a16 = _mm_set1_epi16(16);
+
+  for (int r = 0; r < N; r++) {
+    __m128i b, res, shift, r6, ydx;
+    __m128i resx, resy, resxy;
+    __m128i a0_above, a1_above;
+    int y = r + 1;
+    int base_x = (-y * dx) >> frac_bits_x;
+    int base_shift = 0;
+    if (base_x < (min_base_x - 1)) {
+      base_shift = (min_base_x - base_x - 1) >> upsample_above;
+    }
+    int base_min_diff =
+        (min_base_x - base_x + upsample_above) >> upsample_above;
+    if (base_min_diff > 4) {
+      base_min_diff = 4;
+    } else {
+      if (base_min_diff < 0) base_min_diff = 0;
+    }
+
+    if (base_shift > 3) {
+      a0_x = _mm_setzero_si128();
+      a1_x = _mm_setzero_si128();
+      shift = _mm_setzero_si128();
+    } else {
+      a0_above = _mm_loadu_si128((__m128i *)(above + base_x + base_shift));
+      ydx = _mm_set1_epi16(y * dx);
+      r6 = _mm_slli_epi16(c1234, 6);
+
+      if (upsample_above) {
+        a0_above =
+            _mm_shuffle_epi8(a0_above, *(__m128i *)EvenOddMaskx[base_shift]);
+        a1_above = _mm_srli_si128(a0_above, 8);
+
+        shift = _mm_srli_epi16(
+            _mm_and_si128(
+                _mm_slli_epi16(_mm_sub_epi16(r6, ydx), upsample_above), c3f),
+            1);
+      } else {
+        a0_above =
+            _mm_shuffle_epi8(a0_above, *(__m128i *)LoadMaskx[base_shift]);
+        a1_above = _mm_srli_si128(a0_above, 1);
+
+        shift = _mm_srli_epi16(_mm_and_si128(_mm_sub_epi16(r6, ydx), c3f), 1);
+      }
+      a0_x = _mm_cvtepu8_epi16(a0_above);
+      a1_x = _mm_cvtepu8_epi16(a1_above);
+    }
+    // y calc
+    __m128i a0_y, a1_y, shifty;
+    if (base_x < min_base_x) {
+      DECLARE_ALIGNED(32, int16_t, base_y_c[8]);
+      __m128i y_c, base_y_c_reg, mask, c1234_;
+      c1234_ = _mm_srli_si128(c1234, 2);
+      r6 = _mm_set1_epi16(r << 6);
+      y_c = _mm_sub_epi16(r6, _mm_mullo_epi16(c1234_, dy_reg));
+      base_y_c_reg = _mm_srai_epi16(y_c, frac_bits_y);
+      mask = _mm_cmpgt_epi16(min_y_base, base_y_c_reg);
+      base_y_c_reg = _mm_andnot_si128(mask, base_y_c_reg);
+      _mm_store_si128((__m128i *)base_y_c, base_y_c_reg);
+
+      a0_y = _mm_setr_epi16(left[base_y_c[0]], left[base_y_c[1]],
+                            left[base_y_c[2]], left[base_y_c[3]], 0, 0, 0, 0);
+      base_y_c_reg = _mm_add_epi16(base_y_c_reg, _mm_srli_epi16(a16, 4));
+      _mm_store_si128((__m128i *)base_y_c, base_y_c_reg);
+      a1_y = _mm_setr_epi16(left[base_y_c[0]], left[base_y_c[1]],
+                            left[base_y_c[2]], left[base_y_c[3]], 0, 0, 0, 0);
+
+      if (upsample_left) {
+        shifty = _mm_srli_epi16(
+            _mm_and_si128(_mm_slli_epi16(y_c, upsample_left), c3f), 1);
+      } else {
+        shifty = _mm_srli_epi16(_mm_and_si128(y_c, c3f), 1);
+      }
+      a0_x = _mm_unpacklo_epi64(a0_x, a0_y);
+      a1_x = _mm_unpacklo_epi64(a1_x, a1_y);
+      shift = _mm_unpacklo_epi64(shift, shifty);
+    }
+
+    diff = _mm_sub_epi16(a1_x, a0_x);  // a[x+1] - a[x]
+    a32 = _mm_slli_epi16(a0_x, 5);     // a[x] * 32
+    a32 = _mm_add_epi16(a32, a16);     // a[x] * 32 + 16
+
+    b = _mm_mullo_epi16(diff, shift);
+    res = _mm_add_epi16(a32, b);
+    res = _mm_srli_epi16(res, 5);
+
+    resx = _mm_packus_epi16(res, res);
+    resy = _mm_srli_si128(resx, 4);
+
+    resxy = _mm_blendv_epi8(resx, resy, *(__m128i *)Mask[0][base_min_diff]);
+    *(int *)(dst) = _mm_cvtsi128_si32(resxy);
+    dst += stride;
+  }
+}
+
+static void dr_prediction_z2_Nx8_sse4_1(int N, uint8_t *dst, ptrdiff_t stride,
+                                        const uint8_t *above,
+                                        const uint8_t *left, int upsample_above,
+                                        int upsample_left, int dx, int dy) {
+  const int min_base_x = -(1 << upsample_above);
+  const int min_base_y = -(1 << upsample_left);
+  const int frac_bits_x = 6 - upsample_above;
+  const int frac_bits_y = 6 - upsample_left;
+
+  // pre-filter above pixels
+  // store in temp buffers:
+  //   above[x] * 32 + 16
+  //   above[x+1] - above[x]
+  // final pixels will be calculated as:
+  //   (above[x] * 32 + 16 + (above[x+1] - above[x]) * shift) >> 5
+  __m128i diff, a32;
+  __m128i a0_x, a1_x, a0_y, a1_y;
+  __m128i a0_above, a1_above;
+
+  const __m128i a16 = _mm_set1_epi16(16);
+  const __m128i c3f = _mm_set1_epi16(0x3f);
+  const __m128i min_y_base = _mm_set1_epi16(min_base_y);
+  const __m128i dy_reg = _mm_set1_epi16(dy);
+  const __m128i c1234 = _mm_setr_epi16(1, 2, 3, 4, 5, 6, 7, 8);
+
+  for (int r = 0; r < N; r++) {
+    __m128i b, res, res1, shift;
+    __m128i resx, resy, resxy, r6, ydx;
+
+    int y = r + 1;
+    int base_x = (-y * dx) >> frac_bits_x;
+    int base_shift = 0;
+    if (base_x < (min_base_x - 1)) {
+      base_shift = (min_base_x - base_x - 1) >> upsample_above;
+    }
+    int base_min_diff =
+        (min_base_x - base_x + upsample_above) >> upsample_above;
+    if (base_min_diff > 8) {
+      base_min_diff = 8;
+    } else {
+      if (base_min_diff < 0) base_min_diff = 0;
+    }
+
+    if (base_shift > 7) {
+      resx = _mm_setzero_si128();
+    } else {
+      a0_above = _mm_loadu_si128((__m128i *)(above + base_x + base_shift));
+      ydx = _mm_set1_epi16(y * dx);
+      r6 = _mm_slli_epi16(_mm_srli_si128(c1234, 2), 6);
+      if (upsample_above) {
+        a0_above =
+            _mm_shuffle_epi8(a0_above, *(__m128i *)EvenOddMaskx[base_shift]);
+        a1_above = _mm_srli_si128(a0_above, 8);
+
+        shift = _mm_srli_epi16(
+            _mm_and_si128(
+                _mm_slli_epi16(_mm_sub_epi16(r6, ydx), upsample_above), c3f),
+            1);
+      } else {
+        a1_above = _mm_srli_si128(a0_above, 1);
+        a0_above =
+            _mm_shuffle_epi8(a0_above, *(__m128i *)LoadMaskx[base_shift]);
+        a1_above =
+            _mm_shuffle_epi8(a1_above, *(__m128i *)LoadMaskx[base_shift]);
+
+        shift = _mm_srli_epi16(_mm_and_si128(_mm_sub_epi16(r6, ydx), c3f), 1);
+      }
+      a0_x = _mm_cvtepu8_epi16(a0_above);
+      a1_x = _mm_cvtepu8_epi16(a1_above);
+
+      diff = _mm_sub_epi16(a1_x, a0_x);  // a[x+1] - a[x]
+      a32 = _mm_slli_epi16(a0_x, 5);     // a[x] * 32
+      a32 = _mm_add_epi16(a32, a16);     // a[x] * 32 + 16
+
+      b = _mm_mullo_epi16(diff, shift);
+      res = _mm_add_epi16(a32, b);
+      res = _mm_srli_epi16(res, 5);
+      resx = _mm_packus_epi16(res, res);
+    }
+
+    // y calc
+    if (base_x < min_base_x) {
+      DECLARE_ALIGNED(32, int16_t, base_y_c[16]);
+      __m128i y_c, base_y_c_reg, mask;
+      r6 = _mm_set1_epi16(r << 6);
+      y_c = _mm_sub_epi16(r6, _mm_mullo_epi16(c1234, dy_reg));
+      base_y_c_reg = _mm_srai_epi16(y_c, frac_bits_y);
+      mask = _mm_cmpgt_epi16(min_y_base, base_y_c_reg);
+      base_y_c_reg = _mm_andnot_si128(mask, base_y_c_reg);
+      _mm_store_si128((__m128i *)base_y_c, base_y_c_reg);
+
+      a0_y = _mm_setr_epi16(left[base_y_c[0]], left[base_y_c[1]],
+                            left[base_y_c[2]], left[base_y_c[3]],
+                            left[base_y_c[4]], left[base_y_c[5]],
+                            left[base_y_c[6]], left[base_y_c[7]]);
+      base_y_c_reg = _mm_add_epi16(base_y_c_reg, _mm_srli_epi16(a16, 4));
+      _mm_store_si128((__m128i *)base_y_c, base_y_c_reg);
+
+      a1_y = _mm_setr_epi16(left[base_y_c[0]], left[base_y_c[1]],
+                            left[base_y_c[2]], left[base_y_c[3]],
+                            left[base_y_c[4]], left[base_y_c[5]],
+                            left[base_y_c[6]], left[base_y_c[7]]);
+
+      if (upsample_left) {
+        shift = _mm_srli_epi16(
+            _mm_and_si128(_mm_slli_epi16(y_c, upsample_left), c3f), 1);
+      } else {
+        shift = _mm_srli_epi16(_mm_and_si128(y_c, c3f), 1);
+      }
+
+      diff = _mm_sub_epi16(a1_y, a0_y);  // a[x+1] - a[x]
+      a32 = _mm_slli_epi16(a0_y, 5);     // a[x] * 32
+      a32 = _mm_add_epi16(a32, a16);     // a[x] * 32 + 16
+
+      b = _mm_mullo_epi16(diff, shift);
+      res1 = _mm_add_epi16(a32, b);
+      res1 = _mm_srli_epi16(res1, 5);
+
+      resy = _mm_packus_epi16(res1, res1);
+      resxy = _mm_blendv_epi8(resx, resy, *(__m128i *)Mask[0][base_min_diff]);
+      _mm_storel_epi64((__m128i *)dst, resxy);
+    } else {
+      _mm_storel_epi64((__m128i *)dst, resx);
+    }
+
+    dst += stride;
+  }
+}
+
+static void dr_prediction_z2_HxW_sse4_1(int H, int W, uint8_t *dst,
+                                        ptrdiff_t stride, const uint8_t *above,
+                                        const uint8_t *left, int upsample_above,
+                                        int upsample_left, int dx, int dy) {
+  // here upsample_above and upsample_left are 0 by design of
+  // av1_use_intra_edge_upsample
+  const int min_base_x = -1;
+  const int min_base_y = -1;
+  (void)upsample_above;
+  (void)upsample_left;
+  const int frac_bits_x = 6;
+  const int frac_bits_y = 6;
+
+  __m128i a0_x, a1_x, a0_y, a1_y, a0_y_h, a1_y_h, a32;
+  __m128i diff, shifty, shifty_h;
+  __m128i a0_above, a1_above;
+
+  DECLARE_ALIGNED(32, int16_t, base_y_c[16]);
+  const __m128i a16 = _mm_set1_epi16(16);
+  const __m128i c1 = _mm_srli_epi16(a16, 4);
+  const __m128i min_y_base = _mm_set1_epi16(min_base_y);
+  const __m128i c3f = _mm_set1_epi16(0x3f);
+  const __m128i dy256 = _mm_set1_epi16(dy);
+  const __m128i c0123 = _mm_setr_epi16(0, 1, 2, 3, 4, 5, 6, 7);
+  const __m128i c0123_h = _mm_setr_epi16(8, 9, 10, 11, 12, 13, 14, 15);
+  const __m128i c1234 = _mm_add_epi16(c0123, c1);
+  const __m128i c1234_h = _mm_add_epi16(c0123_h, c1);
+
+  for (int r = 0; r < H; r++) {
+    __m128i b, res, res1, shift, reg_j, r6, ydx;
+    __m128i resx, resy;
+    __m128i resxy;
+    int y = r + 1;
+    ydx = _mm_set1_epi16((int16_t)(y * dx));
+
+    int base_x = (-y * dx) >> frac_bits_x;
+    for (int j = 0; j < W; j += 16) {
+      reg_j = _mm_set1_epi16(j);
+      int base_shift = 0;
+      if ((base_x + j) < (min_base_x - 1)) {
+        base_shift = (min_base_x - (base_x + j) - 1);
+      }
+      int base_min_diff = (min_base_x - base_x - j);
+      if (base_min_diff > 16) {
+        base_min_diff = 16;
+      } else {
+        if (base_min_diff < 0) base_min_diff = 0;
+      }
+
+      if (base_shift < 16) {
+        a0_above =
+            _mm_loadu_si128((__m128i *)(above + base_x + base_shift + j));
+        a1_above =
+            _mm_loadu_si128((__m128i *)(above + base_x + base_shift + 1 + j));
+        a0_above =
+            _mm_shuffle_epi8(a0_above, *(__m128i *)LoadMaskx[base_shift]);
+        a1_above =
+            _mm_shuffle_epi8(a1_above, *(__m128i *)LoadMaskx[base_shift]);
+
+        a0_x = _mm_cvtepu8_epi16(a0_above);
+        a1_x = _mm_cvtepu8_epi16(a1_above);
+
+        r6 = _mm_slli_epi16(_mm_add_epi16(c0123, reg_j), 6);
+        shift = _mm_srli_epi16(_mm_and_si128(_mm_sub_epi16(r6, ydx), c3f), 1);
+
+        diff = _mm_sub_epi16(a1_x, a0_x);  // a[x+1] - a[x]
+        a32 = _mm_slli_epi16(a0_x, 5);     // a[x] * 32
+        a32 = _mm_add_epi16(a32, a16);     // a[x] * 32 + 16
+
+        b = _mm_mullo_epi16(diff, shift);
+        res = _mm_add_epi16(a32, b);
+        res = _mm_srli_epi16(res, 5);  // 16 16-bit values
+
+        a0_x = _mm_cvtepu8_epi16(_mm_srli_si128(a0_above, 8));
+        a1_x = _mm_cvtepu8_epi16(_mm_srli_si128(a1_above, 8));
+
+        r6 = _mm_slli_epi16(_mm_add_epi16(c0123_h, reg_j), 6);
+        shift = _mm_srli_epi16(_mm_and_si128(_mm_sub_epi16(r6, ydx), c3f), 1);
+
+        diff = _mm_sub_epi16(a1_x, a0_x);  // a[x+1] - a[x]
+        a32 = _mm_slli_epi16(a0_x, 5);     // a[x] * 32
+        a32 = _mm_add_epi16(a32, a16);     // a[x] * 32 + 16
+
+        b = _mm_mullo_epi16(diff, shift);
+        res1 = _mm_add_epi16(a32, b);
+        res1 = _mm_srli_epi16(res1, 5);  // 16 16-bit values
+
+        resx = _mm_packus_epi16(res, res1);
+      } else {
+        resx = _mm_setzero_si128();
+      }
+
+      // y calc
+      if (base_x < min_base_x) {
+        __m128i c_reg, c_reg_h, y_reg, y_reg_h, base_y, base_y_h;
+        __m128i mask, mask_h, mul16, mul16_h;
+        r6 = _mm_set1_epi16(r << 6);
+        c_reg = _mm_add_epi16(reg_j, c1234);
+        c_reg_h = _mm_add_epi16(reg_j, c1234_h);
+        mul16 = _mm_min_epu16(_mm_mullo_epi16(c_reg, dy256),
+                              _mm_srli_epi16(min_y_base, 1));
+        mul16_h = _mm_min_epu16(_mm_mullo_epi16(c_reg_h, dy256),
+                                _mm_srli_epi16(min_y_base, 1));
+        y_reg = _mm_sub_epi16(r6, mul16);
+        y_reg_h = _mm_sub_epi16(r6, mul16_h);
+
+        base_y = _mm_srai_epi16(y_reg, frac_bits_y);
+        base_y_h = _mm_srai_epi16(y_reg_h, frac_bits_y);
+        mask = _mm_cmpgt_epi16(min_y_base, base_y);
+        mask_h = _mm_cmpgt_epi16(min_y_base, base_y_h);
+
+        base_y = _mm_blendv_epi8(base_y, min_y_base, mask);
+        base_y_h = _mm_blendv_epi8(base_y_h, min_y_base, mask_h);
+        int16_t min_y = (int16_t)_mm_extract_epi16(base_y_h, 7);
+        int16_t max_y = (int16_t)_mm_extract_epi16(base_y, 0);
+        int16_t offset_diff = max_y - min_y;
+
+        if (offset_diff < 16) {
+          __m128i min_y_reg = _mm_set1_epi16(min_y);
+
+          __m128i base_y_offset = _mm_sub_epi16(base_y, min_y_reg);
+          __m128i base_y_offset_h = _mm_sub_epi16(base_y_h, min_y_reg);
+          __m128i y_offset = _mm_packs_epi16(base_y_offset, base_y_offset_h);
+
+          __m128i a0_mask = _mm_loadu_si128((__m128i *)(left + min_y));
+          __m128i a1_mask = _mm_loadu_si128((__m128i *)(left + min_y + 1));
+          __m128i LoadMask =
+              _mm_loadu_si128((__m128i *)(LoadMaskz2[offset_diff / 4]));
+
+          a0_mask = _mm_and_si128(a0_mask, LoadMask);
+          a1_mask = _mm_and_si128(a1_mask, LoadMask);
+
+          a0_mask = _mm_shuffle_epi8(a0_mask, y_offset);
+          a1_mask = _mm_shuffle_epi8(a1_mask, y_offset);
+          a0_y = _mm_cvtepu8_epi16(a0_mask);
+          a1_y = _mm_cvtepu8_epi16(a1_mask);
+          a0_y_h = _mm_cvtepu8_epi16(_mm_srli_si128(a0_mask, 8));
+          a1_y_h = _mm_cvtepu8_epi16(_mm_srli_si128(a1_mask, 8));
+        } else {
+          base_y = _mm_andnot_si128(mask, base_y);
+          base_y_h = _mm_andnot_si128(mask_h, base_y_h);
+          _mm_store_si128((__m128i *)base_y_c, base_y);
+          _mm_store_si128((__m128i *)&base_y_c[8], base_y_h);
+
+          a0_y = _mm_setr_epi16(left[base_y_c[0]], left[base_y_c[1]],
+                                left[base_y_c[2]], left[base_y_c[3]],
+                                left[base_y_c[4]], left[base_y_c[5]],
+                                left[base_y_c[6]], left[base_y_c[7]]);
+          a0_y_h = _mm_setr_epi16(left[base_y_c[8]], left[base_y_c[9]],
+                                  left[base_y_c[10]], left[base_y_c[11]],
+                                  left[base_y_c[12]], left[base_y_c[13]],
+                                  left[base_y_c[14]], left[base_y_c[15]]);
+          base_y = _mm_add_epi16(base_y, c1);
+          base_y_h = _mm_add_epi16(base_y_h, c1);
+          _mm_store_si128((__m128i *)base_y_c, base_y);
+          _mm_store_si128((__m128i *)&base_y_c[8], base_y_h);
+
+          a1_y = _mm_setr_epi16(left[base_y_c[0]], left[base_y_c[1]],
+                                left[base_y_c[2]], left[base_y_c[3]],
+                                left[base_y_c[4]], left[base_y_c[5]],
+                                left[base_y_c[6]], left[base_y_c[7]]);
+          a1_y_h = _mm_setr_epi16(left[base_y_c[8]], left[base_y_c[9]],
+                                  left[base_y_c[10]], left[base_y_c[11]],
+                                  left[base_y_c[12]], left[base_y_c[13]],
+                                  left[base_y_c[14]], left[base_y_c[15]]);
+        }
+        shifty = _mm_srli_epi16(_mm_and_si128(y_reg, c3f), 1);
+        shifty_h = _mm_srli_epi16(_mm_and_si128(y_reg_h, c3f), 1);
+
+        diff = _mm_sub_epi16(a1_y, a0_y);  // a[x+1] - a[x]
+        a32 = _mm_slli_epi16(a0_y, 5);     // a[x] * 32
+        a32 = _mm_add_epi16(a32, a16);     // a[x] * 32 + 16
+
+        b = _mm_mullo_epi16(diff, shifty);
+        res = _mm_add_epi16(a32, b);
+        res = _mm_srli_epi16(res, 5);  // 16 16-bit values
+
+        diff = _mm_sub_epi16(a1_y_h, a0_y_h);  // a[x+1] - a[x]
+        a32 = _mm_slli_epi16(a0_y_h, 5);       // a[x] * 32
+        a32 = _mm_add_epi16(a32, a16);         // a[x] * 32 + 16
+
+        b = _mm_mullo_epi16(diff, shifty_h);
+        res1 = _mm_add_epi16(a32, b);
+        res1 = _mm_srli_epi16(res1, 5);  // 16 16-bit values
+        resy = _mm_packus_epi16(res, res1);
+      } else {
+        resy = _mm_setzero_si128();
+      }
+      resxy = _mm_blendv_epi8(resx, resy, *(__m128i *)Mask[0][base_min_diff]);
+      _mm_storeu_si128((__m128i *)(dst + j), resxy);
+    }  // for j
+    dst += stride;
+  }
+}
+
+// Directional prediction, zone 2: 90 < angle < 180
+void av1_dr_prediction_z2_sse4_1(uint8_t *dst, ptrdiff_t stride, int bw, int bh,
+                                 const uint8_t *above, const uint8_t *left,
+                                 int upsample_above, int upsample_left, int dx,
+                                 int dy) {
+  assert(dx > 0);
+  assert(dy > 0);
+  switch (bw) {
+    case 4:
+      dr_prediction_z2_Nx4_sse4_1(bh, dst, stride, above, left, upsample_above,
+                                  upsample_left, dx, dy);
+      break;
+    case 8:
+      dr_prediction_z2_Nx8_sse4_1(bh, dst, stride, above, left, upsample_above,
+                                  upsample_left, dx, dy);
+      break;
+    default:
+      dr_prediction_z2_HxW_sse4_1(bh, bw, dst, stride, above, left,
+                                  upsample_above, upsample_left, dx, dy);
+  }
+  return;
+}
+
+// z3 functions
+static void dr_prediction_z3_4x4_sse4_1(uint8_t *dst, ptrdiff_t stride,
+                                        const uint8_t *left, int upsample_left,
+                                        int dy) {
+  __m128i dstvec[4], d[4];
+
+  dr_prediction_z1_HxW_internal_sse4_1(4, 4, dstvec, left, upsample_left, dy);
+  transpose4x8_8x4_low_sse2(&dstvec[0], &dstvec[1], &dstvec[2], &dstvec[3],
+                            &d[0], &d[1], &d[2], &d[3]);
+
+  *(int *)(dst + stride * 0) = _mm_cvtsi128_si32(d[0]);
+  *(int *)(dst + stride * 1) = _mm_cvtsi128_si32(d[1]);
+  *(int *)(dst + stride * 2) = _mm_cvtsi128_si32(d[2]);
+  *(int *)(dst + stride * 3) = _mm_cvtsi128_si32(d[3]);
+  return;
+}
+
+static void dr_prediction_z3_8x8_sse4_1(uint8_t *dst, ptrdiff_t stride,
+                                        const uint8_t *left, int upsample_left,
+                                        int dy) {
+  __m128i dstvec[8], d[8];
+
+  dr_prediction_z1_HxW_internal_sse4_1(8, 8, dstvec, left, upsample_left, dy);
+  transpose8x8_sse2(&dstvec[0], &dstvec[1], &dstvec[2], &dstvec[3], &dstvec[4],
+                    &dstvec[5], &dstvec[6], &dstvec[7], &d[0], &d[1], &d[2],
+                    &d[3]);
+
+  _mm_storel_epi64((__m128i *)(dst + 0 * stride), d[0]);
+  _mm_storel_epi64((__m128i *)(dst + 1 * stride), _mm_srli_si128(d[0], 8));
+  _mm_storel_epi64((__m128i *)(dst + 2 * stride), d[1]);
+  _mm_storel_epi64((__m128i *)(dst + 3 * stride), _mm_srli_si128(d[1], 8));
+  _mm_storel_epi64((__m128i *)(dst + 4 * stride), d[2]);
+  _mm_storel_epi64((__m128i *)(dst + 5 * stride), _mm_srli_si128(d[2], 8));
+  _mm_storel_epi64((__m128i *)(dst + 6 * stride), d[3]);
+  _mm_storel_epi64((__m128i *)(dst + 7 * stride), _mm_srli_si128(d[3], 8));
+}
+
+static void dr_prediction_z3_4x8_sse4_1(uint8_t *dst, ptrdiff_t stride,
+                                        const uint8_t *left, int upsample_left,
+                                        int dy) {
+  __m128i dstvec[4], d[8];
+
+  dr_prediction_z1_HxW_internal_sse4_1(8, 4, dstvec, left, upsample_left, dy);
+  transpose4x8_8x4_sse2(&dstvec[0], &dstvec[1], &dstvec[2], &dstvec[3], &d[0],
+                        &d[1], &d[2], &d[3], &d[4], &d[5], &d[6], &d[7]);
+  for (int i = 0; i < 8; i++) {
+    *(int *)(dst + stride * i) = _mm_cvtsi128_si32(d[i]);
+  }
+}
+
+static void dr_prediction_z3_8x4_sse4_1(uint8_t *dst, ptrdiff_t stride,
+                                        const uint8_t *left, int upsample_left,
+                                        int dy) {
+  __m128i dstvec[8], d[4];
+
+  dr_prediction_z1_HxW_internal_sse4_1(4, 8, dstvec, left, upsample_left, dy);
+  transpose8x8_low_sse2(&dstvec[0], &dstvec[1], &dstvec[2], &dstvec[3],
+                        &dstvec[4], &dstvec[5], &dstvec[6], &dstvec[7], &d[0],
+                        &d[1], &d[2], &d[3]);
+  _mm_storel_epi64((__m128i *)(dst + 0 * stride), d[0]);
+  _mm_storel_epi64((__m128i *)(dst + 1 * stride), d[1]);
+  _mm_storel_epi64((__m128i *)(dst + 2 * stride), d[2]);
+  _mm_storel_epi64((__m128i *)(dst + 3 * stride), d[3]);
+}
+
+static void dr_prediction_z3_8x16_sse4_1(uint8_t *dst, ptrdiff_t stride,
+                                         const uint8_t *left, int upsample_left,
+                                         int dy) {
+  __m128i dstvec[8], d[8];
+
+  dr_prediction_z1_HxW_internal_sse4_1(16, 8, dstvec, left, upsample_left, dy);
+  transpose8x16_16x8_sse2(dstvec, dstvec + 1, dstvec + 2, dstvec + 3,
+                          dstvec + 4, dstvec + 5, dstvec + 6, dstvec + 7, d,
+                          d + 1, d + 2, d + 3, d + 4, d + 5, d + 6, d + 7);
+  for (int i = 0; i < 8; i++) {
+    _mm_storel_epi64((__m128i *)(dst + i * stride), d[i]);
+    _mm_storel_epi64((__m128i *)(dst + (i + 8) * stride),
+                     _mm_srli_si128(d[i], 8));
+  }
+}
+
+static void dr_prediction_z3_16x8_sse4_1(uint8_t *dst, ptrdiff_t stride,
+                                         const uint8_t *left, int upsample_left,
+                                         int dy) {
+  __m128i dstvec[16], d[16];
+
+  dr_prediction_z1_HxW_internal_sse4_1(8, 16, dstvec, left, upsample_left, dy);
+  transpose16x8_8x16_sse2(
+      &dstvec[0], &dstvec[1], &dstvec[2], &dstvec[3], &dstvec[4], &dstvec[5],
+      &dstvec[6], &dstvec[7], &dstvec[8], &dstvec[9], &dstvec[10], &dstvec[11],
+      &dstvec[12], &dstvec[13], &dstvec[14], &dstvec[15], &d[0], &d[1], &d[2],
+      &d[3], &d[4], &d[5], &d[6], &d[7]);
+
+  for (int i = 0; i < 8; i++) {
+    _mm_storeu_si128((__m128i *)(dst + i * stride), d[i]);
+  }
+}
+
+static void dr_prediction_z3_4x16_sse4_1(uint8_t *dst, ptrdiff_t stride,
+                                         const uint8_t *left, int upsample_left,
+                                         int dy) {
+  __m128i dstvec[4], d[16];
+
+  dr_prediction_z1_HxW_internal_sse4_1(16, 4, dstvec, left, upsample_left, dy);
+  transpose4x16_sse2(dstvec, d);
+  for (int i = 0; i < 16; i++) {
+    *(int *)(dst + stride * i) = _mm_cvtsi128_si32(d[i]);
+  }
+}
+
+static void dr_prediction_z3_16x4_sse4_1(uint8_t *dst, ptrdiff_t stride,
+                                         const uint8_t *left, int upsample_left,
+                                         int dy) {
+  __m128i dstvec[16], d[8];
+
+  dr_prediction_z1_HxW_internal_sse4_1(4, 16, dstvec, left, upsample_left, dy);
+  for (int i = 4; i < 8; i++) {
+    d[i] = _mm_setzero_si128();
+  }
+  transpose16x8_8x16_sse2(
+      &dstvec[0], &dstvec[1], &dstvec[2], &dstvec[3], &dstvec[4], &dstvec[5],
+      &dstvec[6], &dstvec[7], &dstvec[8], &dstvec[9], &dstvec[10], &dstvec[11],
+      &dstvec[12], &dstvec[13], &dstvec[14], &dstvec[15], &d[0], &d[1], &d[2],
+      &d[3], &d[4], &d[5], &d[6], &d[7]);
+
+  for (int i = 0; i < 4; i++) {
+    _mm_storeu_si128((__m128i *)(dst + i * stride), d[i]);
+  }
+}
+
+static void dr_prediction_z3_8x32_sse4_1(uint8_t *dst, ptrdiff_t stride,
+                                         const uint8_t *left, int upsample_left,
+                                         int dy) {
+  __m128i dstvec[16], d[16], dstvec_h[16], d_h[16];
+
+  dr_prediction_z1_32xN_internal_sse4_1(8, dstvec, dstvec_h, left,
+                                        upsample_left, dy);
+  for (int i = 8; i < 16; i++) {
+    dstvec[i] = _mm_setzero_si128();
+    dstvec_h[i] = _mm_setzero_si128();
+  }
+  transpose16x16_sse2(dstvec, d);
+  transpose16x16_sse2(dstvec_h, d_h);
+
+  for (int i = 0; i < 16; i++) {
+    _mm_storel_epi64((__m128i *)(dst + i * stride), d[i]);
+  }
+  for (int i = 0; i < 16; i++) {
+    _mm_storel_epi64((__m128i *)(dst + (i + 16) * stride), d_h[i]);
+  }
+}
+
+static void dr_prediction_z3_32x8_sse4_1(uint8_t *dst, ptrdiff_t stride,
+                                         const uint8_t *left, int upsample_left,
+                                         int dy) {
+  __m128i dstvec[32], d[16];
+
+  dr_prediction_z1_HxW_internal_sse4_1(8, 32, dstvec, left, upsample_left, dy);
+
+  transpose16x8_8x16_sse2(
+      &dstvec[0], &dstvec[1], &dstvec[2], &dstvec[3], &dstvec[4], &dstvec[5],
+      &dstvec[6], &dstvec[7], &dstvec[8], &dstvec[9], &dstvec[10], &dstvec[11],
+      &dstvec[12], &dstvec[13], &dstvec[14], &dstvec[15], &d[0], &d[1], &d[2],
+      &d[3], &d[4], &d[5], &d[6], &d[7]);
+  transpose16x8_8x16_sse2(
+      &dstvec[0 + 16], &dstvec[1 + 16], &dstvec[2 + 16], &dstvec[3 + 16],
+      &dstvec[4 + 16], &dstvec[5 + 16], &dstvec[6 + 16], &dstvec[7 + 16],
+      &dstvec[8 + 16], &dstvec[9 + 16], &dstvec[10 + 16], &dstvec[11 + 16],
+      &dstvec[12 + 16], &dstvec[13 + 16], &dstvec[14 + 16], &dstvec[15 + 16],
+      &d[0 + 8], &d[1 + 8], &d[2 + 8], &d[3 + 8], &d[4 + 8], &d[5 + 8],
+      &d[6 + 8], &d[7 + 8]);
+
+  for (int i = 0; i < 8; i++) {
+    _mm_storeu_si128((__m128i *)(dst + i * stride), d[i]);
+    _mm_storeu_si128((__m128i *)(dst + i * stride + 16), d[i + 8]);
+  }
+}
+
+static void dr_prediction_z3_16x16_sse4_1(uint8_t *dst, ptrdiff_t stride,
+                                          const uint8_t *left,
+                                          int upsample_left, int dy) {
+  __m128i dstvec[16], d[16];
+
+  dr_prediction_z1_HxW_internal_sse4_1(16, 16, dstvec, left, upsample_left, dy);
+  transpose16x16_sse2(dstvec, d);
+
+  for (int i = 0; i < 16; i++) {
+    _mm_storeu_si128((__m128i *)(dst + i * stride), d[i]);
+  }
+}
+
+static void dr_prediction_z3_32x32_sse4_1(uint8_t *dst, ptrdiff_t stride,
+                                          const uint8_t *left,
+                                          int upsample_left, int dy) {
+  __m128i dstvec[32], d[32], dstvec_h[32], d_h[32];
+
+  dr_prediction_z1_32xN_internal_sse4_1(32, dstvec, dstvec_h, left,
+                                        upsample_left, dy);
+  transpose16x16_sse2(dstvec, d);
+  transpose16x16_sse2(dstvec_h, d_h);
+  transpose16x16_sse2(dstvec + 16, d + 16);
+  transpose16x16_sse2(dstvec_h + 16, d_h + 16);
+  for (int j = 0; j < 16; j++) {
+    _mm_storeu_si128((__m128i *)(dst + j * stride), d[j]);
+    _mm_storeu_si128((__m128i *)(dst + j * stride + 16), d[j + 16]);
+  }
+  for (int j = 0; j < 16; j++) {
+    _mm_storeu_si128((__m128i *)(dst + (j + 16) * stride), d_h[j]);
+    _mm_storeu_si128((__m128i *)(dst + (j + 16) * stride + 16), d_h[j + 16]);
+  }
+}
+
+static void dr_prediction_z3_64x64_sse4_1(uint8_t *dst, ptrdiff_t stride,
+                                          const uint8_t *left,
+                                          int upsample_left, int dy) {
+  uint8_t dstT[64 * 64];
+  dr_prediction_z1_64xN_sse4_1(64, dstT, 64, left, upsample_left, dy);
+  transpose(dstT, 64, dst, stride, 64, 64);
+}
+
+static void dr_prediction_z3_16x32_sse4_1(uint8_t *dst, ptrdiff_t stride,
+                                          const uint8_t *left,
+                                          int upsample_left, int dy) {
+  __m128i dstvec[16], d[16], dstvec_h[16], d_h[16];
+
+  dr_prediction_z1_32xN_internal_sse4_1(16, dstvec, dstvec_h, left,
+                                        upsample_left, dy);
+  transpose16x16_sse2(dstvec, d);
+  transpose16x16_sse2(dstvec_h, d_h);
+  // store
+  for (int j = 0; j < 16; j++) {
+    _mm_storeu_si128((__m128i *)(dst + j * stride), d[j]);
+    _mm_storeu_si128((__m128i *)(dst + (j + 16) * stride), d_h[j]);
+  }
+}
+
+static void dr_prediction_z3_32x16_sse4_1(uint8_t *dst, ptrdiff_t stride,
+                                          const uint8_t *left,
+                                          int upsample_left, int dy) {
+  __m128i dstvec[32], d[16];
+
+  dr_prediction_z1_HxW_internal_sse4_1(16, 32, dstvec, left, upsample_left, dy);
+  for (int i = 0; i < 32; i += 16) {
+    transpose16x16_sse2((dstvec + i), d);
+    for (int j = 0; j < 16; j++) {
+      _mm_storeu_si128((__m128i *)(dst + j * stride + i), d[j]);
+    }
+  }
+}
+
+static void dr_prediction_z3_32x64_sse4_1(uint8_t *dst, ptrdiff_t stride,
+                                          const uint8_t *left,
+                                          int upsample_left, int dy) {
+  uint8_t dstT[64 * 32];
+  dr_prediction_z1_64xN_sse4_1(32, dstT, 64, left, upsample_left, dy);
+  transpose(dstT, 64, dst, stride, 32, 64);
+}
+
+static void dr_prediction_z3_64x32_sse4_1(uint8_t *dst, ptrdiff_t stride,
+                                          const uint8_t *left,
+                                          int upsample_left, int dy) {
+  uint8_t dstT[32 * 64];
+  dr_prediction_z1_32xN_sse4_1(64, dstT, 32, left, upsample_left, dy);
+  transpose(dstT, 32, dst, stride, 64, 32);
+  return;
+}
+
+static void dr_prediction_z3_16x64_sse4_1(uint8_t *dst, ptrdiff_t stride,
+                                          const uint8_t *left,
+                                          int upsample_left, int dy) {
+  uint8_t dstT[64 * 16];
+  dr_prediction_z1_64xN_sse4_1(16, dstT, 64, left, upsample_left, dy);
+  transpose(dstT, 64, dst, stride, 16, 64);
+}
+
+static void dr_prediction_z3_64x16_sse4_1(uint8_t *dst, ptrdiff_t stride,
+                                          const uint8_t *left,
+                                          int upsample_left, int dy) {
+  __m128i dstvec[64], d[16];
+
+  dr_prediction_z1_HxW_internal_sse4_1(16, 64, dstvec, left, upsample_left, dy);
+  for (int i = 0; i < 64; i += 16) {
+    transpose16x16_sse2(dstvec + i, d);
+    for (int j = 0; j < 16; j++) {
+      _mm_storeu_si128((__m128i *)(dst + j * stride + i), d[j]);
+    }
+  }
+}
+
+void av1_dr_prediction_z3_sse4_1(uint8_t *dst, ptrdiff_t stride, int bw, int bh,
+                                 const uint8_t *above, const uint8_t *left,
+                                 int upsample_left, int dx, int dy) {
+  (void)above;
+  (void)dx;
+  assert(dx == 1);
+  assert(dy > 0);
+
+  if (bw == bh) {
+    switch (bw) {
+      case 4:
+        dr_prediction_z3_4x4_sse4_1(dst, stride, left, upsample_left, dy);
+        break;
+      case 8:
+        dr_prediction_z3_8x8_sse4_1(dst, stride, left, upsample_left, dy);
+        break;
+      case 16:
+        dr_prediction_z3_16x16_sse4_1(dst, stride, left, upsample_left, dy);
+        break;
+      case 32:
+        dr_prediction_z3_32x32_sse4_1(dst, stride, left, upsample_left, dy);
+        break;
+      case 64:
+        dr_prediction_z3_64x64_sse4_1(dst, stride, left, upsample_left, dy);
+        break;
+      default: assert(0 && "Invalid block size");
+    }
+  } else {
+    if (bw < bh) {
+      if (bw + bw == bh) {
+        switch (bw) {
+          case 4:
+            dr_prediction_z3_4x8_sse4_1(dst, stride, left, upsample_left, dy);
+            break;
+          case 8:
+            dr_prediction_z3_8x16_sse4_1(dst, stride, left, upsample_left, dy);
+            break;
+          case 16:
+            dr_prediction_z3_16x32_sse4_1(dst, stride, left, upsample_left, dy);
+            break;
+          case 32:
+            dr_prediction_z3_32x64_sse4_1(dst, stride, left, upsample_left, dy);
+            break;
+          default: assert(0 && "Invalid block size");
+        }
+      } else {
+        switch (bw) {
+          case 4:
+            dr_prediction_z3_4x16_sse4_1(dst, stride, left, upsample_left, dy);
+            break;
+          case 8:
+            dr_prediction_z3_8x32_sse4_1(dst, stride, left, upsample_left, dy);
+            break;
+          case 16:
+            dr_prediction_z3_16x64_sse4_1(dst, stride, left, upsample_left, dy);
+            break;
+          default: assert(0 && "Invalid block size");
+        }
+      }
+    } else {
+      if (bh + bh == bw) {
+        switch (bh) {
+          case 4:
+            dr_prediction_z3_8x4_sse4_1(dst, stride, left, upsample_left, dy);
+            break;
+          case 8:
+            dr_prediction_z3_16x8_sse4_1(dst, stride, left, upsample_left, dy);
+            break;
+          case 16:
+            dr_prediction_z3_32x16_sse4_1(dst, stride, left, upsample_left, dy);
+            break;
+          case 32:
+            dr_prediction_z3_64x32_sse4_1(dst, stride, left, upsample_left, dy);
+            break;
+          default: assert(0 && "Invalid block size");
+        }
+      } else {
+        switch (bh) {
+          case 4:
+            dr_prediction_z3_16x4_sse4_1(dst, stride, left, upsample_left, dy);
+            break;
+          case 8:
+            dr_prediction_z3_32x8_sse4_1(dst, stride, left, upsample_left, dy);
+            break;
+          case 16:
+            dr_prediction_z3_64x16_sse4_1(dst, stride, left, upsample_left, dy);
+            break;
+          default: assert(0 && "Invalid block size");
+        }
+      }
+    }
+  }
+}
diff --git a/third_party/aom/aom_dsp/x86/intrapred_ssse3.c b/third_party/aom/aom_dsp/x86/intrapred_ssse3.c
new file mode 100644
index 0000000000..fd48260c6f
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/intrapred_ssse3.c
@@ -0,0 +1,2997 @@
+/*
+ * Copyright (c) 2017, 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_dsp_rtcd.h"
+
+#include "aom_dsp/intrapred_common.h"
+
+// -----------------------------------------------------------------------------
+// PAETH_PRED
+
+// Return 8 16-bit pixels in one row
+static INLINE __m128i paeth_8x1_pred(const __m128i *left, const __m128i *top,
+                                     const __m128i *topleft) {
+  const __m128i base = _mm_sub_epi16(_mm_add_epi16(*top, *left), *topleft);
+
+  __m128i pl = _mm_abs_epi16(_mm_sub_epi16(base, *left));
+  __m128i pt = _mm_abs_epi16(_mm_sub_epi16(base, *top));
+  __m128i ptl = _mm_abs_epi16(_mm_sub_epi16(base, *topleft));
+
+  __m128i mask1 = _mm_cmpgt_epi16(pl, pt);
+  mask1 = _mm_or_si128(mask1, _mm_cmpgt_epi16(pl, ptl));
+  __m128i mask2 = _mm_cmpgt_epi16(pt, ptl);
+
+  pl = _mm_andnot_si128(mask1, *left);
+
+  ptl = _mm_and_si128(mask2, *topleft);
+  pt = _mm_andnot_si128(mask2, *top);
+  pt = _mm_or_si128(pt, ptl);
+  pt = _mm_and_si128(mask1, pt);
+
+  return _mm_or_si128(pl, pt);
+}
+
+void aom_paeth_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t stride,
+                                   const uint8_t *above, const uint8_t *left) {
+  __m128i l = _mm_loadl_epi64((const __m128i *)left);
+  const __m128i t = _mm_loadl_epi64((const __m128i *)above);
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i t16 = _mm_unpacklo_epi8(t, zero);
+  const __m128i tl16 = _mm_set1_epi16((int16_t)above[-1]);
+  __m128i rep = _mm_set1_epi16((short)0x8000);
+  const __m128i one = _mm_set1_epi16(1);
+
+  int i;
+  for (i = 0; i < 4; ++i) {
+    const __m128i l16 = _mm_shuffle_epi8(l, rep);
+    const __m128i row = paeth_8x1_pred(&l16, &t16, &tl16);
+
+    *(int *)dst = _mm_cvtsi128_si32(_mm_packus_epi16(row, row));
+    dst += stride;
+    rep = _mm_add_epi16(rep, one);
+  }
+}
+
+void aom_paeth_predictor_4x8_ssse3(uint8_t *dst, ptrdiff_t stride,
+                                   const uint8_t *above, const uint8_t *left) {
+  __m128i l = _mm_loadl_epi64((const __m128i *)left);
+  const __m128i t = _mm_loadl_epi64((const __m128i *)above);
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i t16 = _mm_unpacklo_epi8(t, zero);
+  const __m128i tl16 = _mm_set1_epi16((int16_t)above[-1]);
+  __m128i rep = _mm_set1_epi16((short)0x8000);
+  const __m128i one = _mm_set1_epi16(1);
+
+  int i;
+  for (i = 0; i < 8; ++i) {
+    const __m128i l16 = _mm_shuffle_epi8(l, rep);
+    const __m128i row = paeth_8x1_pred(&l16, &t16, &tl16);
+
+    *(int *)dst = _mm_cvtsi128_si32(_mm_packus_epi16(row, row));
+    dst += stride;
+    rep = _mm_add_epi16(rep, one);
+  }
+}
+
+void aom_paeth_predictor_4x16_ssse3(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  __m128i l = _mm_load_si128((const __m128i *)left);
+  const __m128i t = _mm_cvtsi32_si128(((const int *)above)[0]);
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i t16 = _mm_unpacklo_epi8(t, zero);
+  const __m128i tl16 = _mm_set1_epi16((int16_t)above[-1]);
+  __m128i rep = _mm_set1_epi16((short)0x8000);
+  const __m128i one = _mm_set1_epi16(1);
+
+  for (int i = 0; i < 16; ++i) {
+    const __m128i l16 = _mm_shuffle_epi8(l, rep);
+    const __m128i row = paeth_8x1_pred(&l16, &t16, &tl16);
+
+    *(int *)dst = _mm_cvtsi128_si32(_mm_packus_epi16(row, row));
+    dst += stride;
+    rep = _mm_add_epi16(rep, one);
+  }
+}
+
+void aom_paeth_predictor_8x4_ssse3(uint8_t *dst, ptrdiff_t stride,
+                                   const uint8_t *above, const uint8_t *left) {
+  __m128i l = _mm_loadl_epi64((const __m128i *)left);
+  const __m128i t = _mm_loadl_epi64((const __m128i *)above);
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i t16 = _mm_unpacklo_epi8(t, zero);
+  const __m128i tl16 = _mm_set1_epi16((int16_t)above[-1]);
+  __m128i rep = _mm_set1_epi16((short)0x8000);
+  const __m128i one = _mm_set1_epi16(1);
+
+  int i;
+  for (i = 0; i < 4; ++i) {
+    const __m128i l16 = _mm_shuffle_epi8(l, rep);
+    const __m128i row = paeth_8x1_pred(&l16, &t16, &tl16);
+
+    _mm_storel_epi64((__m128i *)dst, _mm_packus_epi16(row, row));
+    dst += stride;
+    rep = _mm_add_epi16(rep, one);
+  }
+}
+
+void aom_paeth_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t stride,
+                                   const uint8_t *above, const uint8_t *left) {
+  __m128i l = _mm_loadl_epi64((const __m128i *)left);
+  const __m128i t = _mm_loadl_epi64((const __m128i *)above);
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i t16 = _mm_unpacklo_epi8(t, zero);
+  const __m128i tl16 = _mm_set1_epi16((int16_t)above[-1]);
+  __m128i rep = _mm_set1_epi16((short)0x8000);
+  const __m128i one = _mm_set1_epi16(1);
+
+  int i;
+  for (i = 0; i < 8; ++i) {
+    const __m128i l16 = _mm_shuffle_epi8(l, rep);
+    const __m128i row = paeth_8x1_pred(&l16, &t16, &tl16);
+
+    _mm_storel_epi64((__m128i *)dst, _mm_packus_epi16(row, row));
+    dst += stride;
+    rep = _mm_add_epi16(rep, one);
+  }
+}
+
+void aom_paeth_predictor_8x16_ssse3(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  __m128i l = _mm_load_si128((const __m128i *)left);
+  const __m128i t = _mm_loadl_epi64((const __m128i *)above);
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i t16 = _mm_unpacklo_epi8(t, zero);
+  const __m128i tl16 = _mm_set1_epi16((int16_t)above[-1]);
+  __m128i rep = _mm_set1_epi16((short)0x8000);
+  const __m128i one = _mm_set1_epi16(1);
+
+  int i;
+  for (i = 0; i < 16; ++i) {
+    const __m128i l16 = _mm_shuffle_epi8(l, rep);
+    const __m128i row = paeth_8x1_pred(&l16, &t16, &tl16);
+
+    _mm_storel_epi64((__m128i *)dst, _mm_packus_epi16(row, row));
+    dst += stride;
+    rep = _mm_add_epi16(rep, one);
+  }
+}
+
+void aom_paeth_predictor_8x32_ssse3(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  const __m128i t = _mm_loadl_epi64((const __m128i *)above);
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i t16 = _mm_unpacklo_epi8(t, zero);
+  const __m128i tl16 = _mm_set1_epi16((int16_t)above[-1]);
+  const __m128i one = _mm_set1_epi16(1);
+
+  for (int j = 0; j < 2; ++j) {
+    const __m128i l = _mm_load_si128((const __m128i *)(left + j * 16));
+    __m128i rep = _mm_set1_epi16((short)0x8000);
+    for (int i = 0; i < 16; ++i) {
+      const __m128i l16 = _mm_shuffle_epi8(l, rep);
+      const __m128i row = paeth_8x1_pred(&l16, &t16, &tl16);
+
+      _mm_storel_epi64((__m128i *)dst, _mm_packus_epi16(row, row));
+      dst += stride;
+      rep = _mm_add_epi16(rep, one);
+    }
+  }
+}
+
+// Return 16 8-bit pixels in one row
+static INLINE __m128i paeth_16x1_pred(const __m128i *left, const __m128i *top0,
+                                      const __m128i *top1,
+                                      const __m128i *topleft) {
+  const __m128i p0 = paeth_8x1_pred(left, top0, topleft);
+  const __m128i p1 = paeth_8x1_pred(left, top1, topleft);
+  return _mm_packus_epi16(p0, p1);
+}
+
+void aom_paeth_predictor_16x4_ssse3(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  __m128i l = _mm_cvtsi32_si128(((const int *)left)[0]);
+  const __m128i t = _mm_load_si128((const __m128i *)above);
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i top0 = _mm_unpacklo_epi8(t, zero);
+  const __m128i top1 = _mm_unpackhi_epi8(t, zero);
+  const __m128i tl16 = _mm_set1_epi16((int16_t)above[-1]);
+  __m128i rep = _mm_set1_epi16((short)0x8000);
+  const __m128i one = _mm_set1_epi16(1);
+
+  for (int i = 0; i < 4; ++i) {
+    const __m128i l16 = _mm_shuffle_epi8(l, rep);
+    const __m128i row = paeth_16x1_pred(&l16, &top0, &top1, &tl16);
+
+    _mm_store_si128((__m128i *)dst, row);
+    dst += stride;
+    rep = _mm_add_epi16(rep, one);
+  }
+}
+
+void aom_paeth_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  __m128i l = _mm_loadl_epi64((const __m128i *)left);
+  const __m128i t = _mm_load_si128((const __m128i *)above);
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i top0 = _mm_unpacklo_epi8(t, zero);
+  const __m128i top1 = _mm_unpackhi_epi8(t, zero);
+  const __m128i tl16 = _mm_set1_epi16((int16_t)above[-1]);
+  __m128i rep = _mm_set1_epi16((short)0x8000);
+  const __m128i one = _mm_set1_epi16(1);
+
+  int i;
+  for (i = 0; i < 8; ++i) {
+    const __m128i l16 = _mm_shuffle_epi8(l, rep);
+    const __m128i row = paeth_16x1_pred(&l16, &top0, &top1, &tl16);
+
+    _mm_store_si128((__m128i *)dst, row);
+    dst += stride;
+    rep = _mm_add_epi16(rep, one);
+  }
+}
+
+void aom_paeth_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  __m128i l = _mm_load_si128((const __m128i *)left);
+  const __m128i t = _mm_load_si128((const __m128i *)above);
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i top0 = _mm_unpacklo_epi8(t, zero);
+  const __m128i top1 = _mm_unpackhi_epi8(t, zero);
+  const __m128i tl16 = _mm_set1_epi16((int16_t)above[-1]);
+  __m128i rep = _mm_set1_epi16((short)0x8000);
+  const __m128i one = _mm_set1_epi16(1);
+
+  int i;
+  for (i = 0; i < 16; ++i) {
+    const __m128i l16 = _mm_shuffle_epi8(l, rep);
+    const __m128i row = paeth_16x1_pred(&l16, &top0, &top1, &tl16);
+
+    _mm_store_si128((__m128i *)dst, row);
+    dst += stride;
+    rep = _mm_add_epi16(rep, one);
+  }
+}
+
+void aom_paeth_predictor_16x32_ssse3(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  __m128i l = _mm_load_si128((const __m128i *)left);
+  const __m128i t = _mm_load_si128((const __m128i *)above);
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i top0 = _mm_unpacklo_epi8(t, zero);
+  const __m128i top1 = _mm_unpackhi_epi8(t, zero);
+  const __m128i tl16 = _mm_set1_epi16((int16_t)above[-1]);
+  __m128i rep = _mm_set1_epi16((short)0x8000);
+  const __m128i one = _mm_set1_epi16(1);
+  __m128i l16;
+
+  int i;
+  for (i = 0; i < 16; ++i) {
+    l16 = _mm_shuffle_epi8(l, rep);
+    const __m128i row = paeth_16x1_pred(&l16, &top0, &top1, &tl16);
+
+    _mm_store_si128((__m128i *)dst, row);
+    dst += stride;
+    rep = _mm_add_epi16(rep, one);
+  }
+
+  l = _mm_load_si128((const __m128i *)(left + 16));
+  rep = _mm_set1_epi16((short)0x8000);
+  for (i = 0; i < 16; ++i) {
+    l16 = _mm_shuffle_epi8(l, rep);
+    const __m128i row = paeth_16x1_pred(&l16, &top0, &top1, &tl16);
+
+    _mm_store_si128((__m128i *)dst, row);
+    dst += stride;
+    rep = _mm_add_epi16(rep, one);
+  }
+}
+
+void aom_paeth_predictor_16x64_ssse3(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  const __m128i t = _mm_load_si128((const __m128i *)above);
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i top0 = _mm_unpacklo_epi8(t, zero);
+  const __m128i top1 = _mm_unpackhi_epi8(t, zero);
+  const __m128i tl16 = _mm_set1_epi16((int16_t)above[-1]);
+  const __m128i one = _mm_set1_epi16(1);
+
+  for (int j = 0; j < 4; ++j) {
+    const __m128i l = _mm_load_si128((const __m128i *)(left + j * 16));
+    __m128i rep = _mm_set1_epi16((short)0x8000);
+    for (int i = 0; i < 16; ++i) {
+      const __m128i l16 = _mm_shuffle_epi8(l, rep);
+      const __m128i row = paeth_16x1_pred(&l16, &top0, &top1, &tl16);
+      _mm_store_si128((__m128i *)dst, row);
+      dst += stride;
+      rep = _mm_add_epi16(rep, one);
+    }
+  }
+}
+
+void aom_paeth_predictor_32x8_ssse3(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  const __m128i a = _mm_load_si128((const __m128i *)above);
+  const __m128i b = _mm_load_si128((const __m128i *)(above + 16));
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i al = _mm_unpacklo_epi8(a, zero);
+  const __m128i ah = _mm_unpackhi_epi8(a, zero);
+  const __m128i bl = _mm_unpacklo_epi8(b, zero);
+  const __m128i bh = _mm_unpackhi_epi8(b, zero);
+
+  const __m128i tl16 = _mm_set1_epi16((int16_t)above[-1]);
+  __m128i rep = _mm_set1_epi16((short)0x8000);
+  const __m128i one = _mm_set1_epi16(1);
+  const __m128i l = _mm_loadl_epi64((const __m128i *)left);
+  __m128i l16;
+
+  for (int i = 0; i < 8; ++i) {
+    l16 = _mm_shuffle_epi8(l, rep);
+    const __m128i r32l = paeth_16x1_pred(&l16, &al, &ah, &tl16);
+    const __m128i r32h = paeth_16x1_pred(&l16, &bl, &bh, &tl16);
+
+    _mm_store_si128((__m128i *)dst, r32l);
+    _mm_store_si128((__m128i *)(dst + 16), r32h);
+    dst += stride;
+    rep = _mm_add_epi16(rep, one);
+  }
+}
+
+void aom_paeth_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  const __m128i a = _mm_load_si128((const __m128i *)above);
+  const __m128i b = _mm_load_si128((const __m128i *)(above + 16));
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i al = _mm_unpacklo_epi8(a, zero);
+  const __m128i ah = _mm_unpackhi_epi8(a, zero);
+  const __m128i bl = _mm_unpacklo_epi8(b, zero);
+  const __m128i bh = _mm_unpackhi_epi8(b, zero);
+
+  const __m128i tl16 = _mm_set1_epi16((int16_t)above[-1]);
+  __m128i rep = _mm_set1_epi16((short)0x8000);
+  const __m128i one = _mm_set1_epi16(1);
+  __m128i l = _mm_load_si128((const __m128i *)left);
+  __m128i l16;
+
+  int i;
+  for (i = 0; i < 16; ++i) {
+    l16 = _mm_shuffle_epi8(l, rep);
+    const __m128i r32l = paeth_16x1_pred(&l16, &al, &ah, &tl16);
+    const __m128i r32h = paeth_16x1_pred(&l16, &bl, &bh, &tl16);
+
+    _mm_store_si128((__m128i *)dst, r32l);
+    _mm_store_si128((__m128i *)(dst + 16), r32h);
+    dst += stride;
+    rep = _mm_add_epi16(rep, one);
+  }
+}
+
+void aom_paeth_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  const __m128i a = _mm_load_si128((const __m128i *)above);
+  const __m128i b = _mm_load_si128((const __m128i *)(above + 16));
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i al = _mm_unpacklo_epi8(a, zero);
+  const __m128i ah = _mm_unpackhi_epi8(a, zero);
+  const __m128i bl = _mm_unpacklo_epi8(b, zero);
+  const __m128i bh = _mm_unpackhi_epi8(b, zero);
+
+  const __m128i tl16 = _mm_set1_epi16((int16_t)above[-1]);
+  __m128i rep = _mm_set1_epi16((short)0x8000);
+  const __m128i one = _mm_set1_epi16(1);
+  __m128i l = _mm_load_si128((const __m128i *)left);
+  __m128i l16;
+
+  int i;
+  for (i = 0; i < 16; ++i) {
+    l16 = _mm_shuffle_epi8(l, rep);
+    const __m128i r32l = paeth_16x1_pred(&l16, &al, &ah, &tl16);
+    const __m128i r32h = paeth_16x1_pred(&l16, &bl, &bh, &tl16);
+
+    _mm_store_si128((__m128i *)dst, r32l);
+    _mm_store_si128((__m128i *)(dst + 16), r32h);
+    dst += stride;
+    rep = _mm_add_epi16(rep, one);
+  }
+
+  rep = _mm_set1_epi16((short)0x8000);
+  l = _mm_load_si128((const __m128i *)(left + 16));
+  for (i = 0; i < 16; ++i) {
+    l16 = _mm_shuffle_epi8(l, rep);
+    const __m128i r32l = paeth_16x1_pred(&l16, &al, &ah, &tl16);
+    const __m128i r32h = paeth_16x1_pred(&l16, &bl, &bh, &tl16);
+
+    _mm_store_si128((__m128i *)dst, r32l);
+    _mm_store_si128((__m128i *)(dst + 16), r32h);
+    dst += stride;
+    rep = _mm_add_epi16(rep, one);
+  }
+}
+
+void aom_paeth_predictor_32x64_ssse3(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  const __m128i a = _mm_load_si128((const __m128i *)above);
+  const __m128i b = _mm_load_si128((const __m128i *)(above + 16));
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i al = _mm_unpacklo_epi8(a, zero);
+  const __m128i ah = _mm_unpackhi_epi8(a, zero);
+  const __m128i bl = _mm_unpacklo_epi8(b, zero);
+  const __m128i bh = _mm_unpackhi_epi8(b, zero);
+
+  const __m128i tl16 = _mm_set1_epi16((int16_t)above[-1]);
+  const __m128i one = _mm_set1_epi16(1);
+  __m128i l16;
+
+  int i, j;
+  for (j = 0; j < 4; ++j) {
+    const __m128i l = _mm_load_si128((const __m128i *)(left + j * 16));
+    __m128i rep = _mm_set1_epi16((short)0x8000);
+    for (i = 0; i < 16; ++i) {
+      l16 = _mm_shuffle_epi8(l, rep);
+      const __m128i r32l = paeth_16x1_pred(&l16, &al, &ah, &tl16);
+      const __m128i r32h = paeth_16x1_pred(&l16, &bl, &bh, &tl16);
+
+      _mm_store_si128((__m128i *)dst, r32l);
+      _mm_store_si128((__m128i *)(dst + 16), r32h);
+      dst += stride;
+      rep = _mm_add_epi16(rep, one);
+    }
+  }
+}
+
+void aom_paeth_predictor_64x32_ssse3(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  const __m128i a = _mm_load_si128((const __m128i *)above);
+  const __m128i b = _mm_load_si128((const __m128i *)(above + 16));
+  const __m128i c = _mm_load_si128((const __m128i *)(above + 32));
+  const __m128i d = _mm_load_si128((const __m128i *)(above + 48));
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i al = _mm_unpacklo_epi8(a, zero);
+  const __m128i ah = _mm_unpackhi_epi8(a, zero);
+  const __m128i bl = _mm_unpacklo_epi8(b, zero);
+  const __m128i bh = _mm_unpackhi_epi8(b, zero);
+  const __m128i cl = _mm_unpacklo_epi8(c, zero);
+  const __m128i ch = _mm_unpackhi_epi8(c, zero);
+  const __m128i dl = _mm_unpacklo_epi8(d, zero);
+  const __m128i dh = _mm_unpackhi_epi8(d, zero);
+
+  const __m128i tl16 = _mm_set1_epi16((int16_t)above[-1]);
+  const __m128i one = _mm_set1_epi16(1);
+  __m128i l16;
+
+  int i, j;
+  for (j = 0; j < 2; ++j) {
+    const __m128i l = _mm_load_si128((const __m128i *)(left + j * 16));
+    __m128i rep = _mm_set1_epi16((short)0x8000);
+    for (i = 0; i < 16; ++i) {
+      l16 = _mm_shuffle_epi8(l, rep);
+      const __m128i r0 = paeth_16x1_pred(&l16, &al, &ah, &tl16);
+      const __m128i r1 = paeth_16x1_pred(&l16, &bl, &bh, &tl16);
+      const __m128i r2 = paeth_16x1_pred(&l16, &cl, &ch, &tl16);
+      const __m128i r3 = paeth_16x1_pred(&l16, &dl, &dh, &tl16);
+
+      _mm_store_si128((__m128i *)dst, r0);
+      _mm_store_si128((__m128i *)(dst + 16), r1);
+      _mm_store_si128((__m128i *)(dst + 32), r2);
+      _mm_store_si128((__m128i *)(dst + 48), r3);
+      dst += stride;
+      rep = _mm_add_epi16(rep, one);
+    }
+  }
+}
+
+void aom_paeth_predictor_64x64_ssse3(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  const __m128i a = _mm_load_si128((const __m128i *)above);
+  const __m128i b = _mm_load_si128((const __m128i *)(above + 16));
+  const __m128i c = _mm_load_si128((const __m128i *)(above + 32));
+  const __m128i d = _mm_load_si128((const __m128i *)(above + 48));
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i al = _mm_unpacklo_epi8(a, zero);
+  const __m128i ah = _mm_unpackhi_epi8(a, zero);
+  const __m128i bl = _mm_unpacklo_epi8(b, zero);
+  const __m128i bh = _mm_unpackhi_epi8(b, zero);
+  const __m128i cl = _mm_unpacklo_epi8(c, zero);
+  const __m128i ch = _mm_unpackhi_epi8(c, zero);
+  const __m128i dl = _mm_unpacklo_epi8(d, zero);
+  const __m128i dh = _mm_unpackhi_epi8(d, zero);
+
+  const __m128i tl16 = _mm_set1_epi16((int16_t)above[-1]);
+  const __m128i one = _mm_set1_epi16(1);
+  __m128i l16;
+
+  int i, j;
+  for (j = 0; j < 4; ++j) {
+    const __m128i l = _mm_load_si128((const __m128i *)(left + j * 16));
+    __m128i rep = _mm_set1_epi16((short)0x8000);
+    for (i = 0; i < 16; ++i) {
+      l16 = _mm_shuffle_epi8(l, rep);
+      const __m128i r0 = paeth_16x1_pred(&l16, &al, &ah, &tl16);
+      const __m128i r1 = paeth_16x1_pred(&l16, &bl, &bh, &tl16);
+      const __m128i r2 = paeth_16x1_pred(&l16, &cl, &ch, &tl16);
+      const __m128i r3 = paeth_16x1_pred(&l16, &dl, &dh, &tl16);
+
+      _mm_store_si128((__m128i *)dst, r0);
+      _mm_store_si128((__m128i *)(dst + 16), r1);
+      _mm_store_si128((__m128i *)(dst + 32), r2);
+      _mm_store_si128((__m128i *)(dst + 48), r3);
+      dst += stride;
+      rep = _mm_add_epi16(rep, one);
+    }
+  }
+}
+
+void aom_paeth_predictor_64x16_ssse3(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  const __m128i a = _mm_load_si128((const __m128i *)above);
+  const __m128i b = _mm_load_si128((const __m128i *)(above + 16));
+  const __m128i c = _mm_load_si128((const __m128i *)(above + 32));
+  const __m128i d = _mm_load_si128((const __m128i *)(above + 48));
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i al = _mm_unpacklo_epi8(a, zero);
+  const __m128i ah = _mm_unpackhi_epi8(a, zero);
+  const __m128i bl = _mm_unpacklo_epi8(b, zero);
+  const __m128i bh = _mm_unpackhi_epi8(b, zero);
+  const __m128i cl = _mm_unpacklo_epi8(c, zero);
+  const __m128i ch = _mm_unpackhi_epi8(c, zero);
+  const __m128i dl = _mm_unpacklo_epi8(d, zero);
+  const __m128i dh = _mm_unpackhi_epi8(d, zero);
+
+  const __m128i tl16 = _mm_set1_epi16((int16_t)above[-1]);
+  const __m128i one = _mm_set1_epi16(1);
+  __m128i l16;
+
+  int i;
+  const __m128i l = _mm_load_si128((const __m128i *)left);
+  __m128i rep = _mm_set1_epi16((short)0x8000);
+  for (i = 0; i < 16; ++i) {
+    l16 = _mm_shuffle_epi8(l, rep);
+    const __m128i r0 = paeth_16x1_pred(&l16, &al, &ah, &tl16);
+    const __m128i r1 = paeth_16x1_pred(&l16, &bl, &bh, &tl16);
+    const __m128i r2 = paeth_16x1_pred(&l16, &cl, &ch, &tl16);
+    const __m128i r3 = paeth_16x1_pred(&l16, &dl, &dh, &tl16);
+
+    _mm_store_si128((__m128i *)dst, r0);
+    _mm_store_si128((__m128i *)(dst + 16), r1);
+    _mm_store_si128((__m128i *)(dst + 32), r2);
+    _mm_store_si128((__m128i *)(dst + 48), r3);
+    dst += stride;
+    rep = _mm_add_epi16(rep, one);
+  }
+}
+
+// -----------------------------------------------------------------------------
+// SMOOTH_PRED
+
+// pixels[0]: above and below_pred interleave vector
+// pixels[1]: left vector
+// pixels[2]: right_pred vector
+static INLINE void load_pixel_w4(const uint8_t *above, const uint8_t *left,
+                                 int height, __m128i *pixels) {
+  __m128i d = _mm_cvtsi32_si128(((const int *)above)[0]);
+  if (height == 4)
+    pixels[1] = _mm_cvtsi32_si128(((const int *)left)[0]);
+  else if (height == 8)
+    pixels[1] = _mm_loadl_epi64(((const __m128i *)left));
+  else
+    pixels[1] = _mm_loadu_si128(((const __m128i *)left));
+
+  pixels[2] = _mm_set1_epi16((int16_t)above[3]);
+
+  const __m128i bp = _mm_set1_epi16((int16_t)left[height - 1]);
+  const __m128i zero = _mm_setzero_si128();
+  d = _mm_unpacklo_epi8(d, zero);
+  pixels[0] = _mm_unpacklo_epi16(d, bp);
+}
+
+// weight_h[0]: weight_h vector
+// weight_h[1]: scale - weight_h vector
+// weight_h[2]: same as [0], second half for height = 16 only
+// weight_h[3]: same as [1], second half for height = 16 only
+// weight_w[0]: weights_w and scale - weights_w interleave vector
+static INLINE void load_weight_w4(int height, __m128i *weight_h,
+                                  __m128i *weight_w) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i d = _mm_set1_epi16((int16_t)(1 << SMOOTH_WEIGHT_LOG2_SCALE));
+  const __m128i t = _mm_cvtsi32_si128(((const int *)smooth_weights)[0]);
+  weight_h[0] = _mm_unpacklo_epi8(t, zero);
+  weight_h[1] = _mm_sub_epi16(d, weight_h[0]);
+  weight_w[0] = _mm_unpacklo_epi16(weight_h[0], weight_h[1]);
+
+  if (height == 8) {
+    const __m128i weight = _mm_loadl_epi64((const __m128i *)&smooth_weights[4]);
+    weight_h[0] = _mm_unpacklo_epi8(weight, zero);
+    weight_h[1] = _mm_sub_epi16(d, weight_h[0]);
+  } else if (height == 16) {
+    const __m128i weight =
+        _mm_loadu_si128((const __m128i *)&smooth_weights[12]);
+    weight_h[0] = _mm_unpacklo_epi8(weight, zero);
+    weight_h[1] = _mm_sub_epi16(d, weight_h[0]);
+    weight_h[2] = _mm_unpackhi_epi8(weight, zero);
+    weight_h[3] = _mm_sub_epi16(d, weight_h[2]);
+  }
+}
+
+static INLINE void smooth_pred_4xh(const __m128i *pixel, const __m128i *wh,
+                                   const __m128i *ww, int h, uint8_t *dst,
+                                   ptrdiff_t stride, int second_half) {
+  const __m128i round = _mm_set1_epi32((1 << SMOOTH_WEIGHT_LOG2_SCALE));
+  const __m128i one = _mm_set1_epi16(1);
+  const __m128i inc = _mm_set1_epi16(0x202);
+  const __m128i gat = _mm_set1_epi32(0xc080400);
+  __m128i rep = second_half ? _mm_set1_epi16((short)0x8008)
+                            : _mm_set1_epi16((short)0x8000);
+  __m128i d = _mm_set1_epi16(0x100);
+
+  for (int i = 0; i < h; ++i) {
+    const __m128i wg_wg = _mm_shuffle_epi8(wh[0], d);
+    const __m128i sc_sc = _mm_shuffle_epi8(wh[1], d);
+    const __m128i wh_sc = _mm_unpacklo_epi16(wg_wg, sc_sc);
+    __m128i s = _mm_madd_epi16(pixel[0], wh_sc);
+
+    __m128i b = _mm_shuffle_epi8(pixel[1], rep);
+    b = _mm_unpacklo_epi16(b, pixel[2]);
+    __m128i sum = _mm_madd_epi16(b, ww[0]);
+
+    sum = _mm_add_epi32(s, sum);
+    sum = _mm_add_epi32(sum, round);
+    sum = _mm_srai_epi32(sum, 1 + SMOOTH_WEIGHT_LOG2_SCALE);
+
+    sum = _mm_shuffle_epi8(sum, gat);
+    *(int *)dst = _mm_cvtsi128_si32(sum);
+    dst += stride;
+
+    rep = _mm_add_epi16(rep, one);
+    d = _mm_add_epi16(d, inc);
+  }
+}
+
+void aom_smooth_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  __m128i pixels[3];
+  load_pixel_w4(above, left, 4, pixels);
+
+  __m128i wh[4], ww[2];
+  load_weight_w4(4, wh, ww);
+
+  smooth_pred_4xh(pixels, wh, ww, 4, dst, stride, 0);
+}
+
+void aom_smooth_predictor_4x8_ssse3(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  __m128i pixels[3];
+  load_pixel_w4(above, left, 8, pixels);
+
+  __m128i wh[4], ww[2];
+  load_weight_w4(8, wh, ww);
+
+  smooth_pred_4xh(pixels, wh, ww, 8, dst, stride, 0);
+}
+
+void aom_smooth_predictor_4x16_ssse3(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  __m128i pixels[3];
+  load_pixel_w4(above, left, 16, pixels);
+
+  __m128i wh[4], ww[2];
+  load_weight_w4(16, wh, ww);
+
+  smooth_pred_4xh(pixels, wh, ww, 8, dst, stride, 0);
+  dst += stride << 3;
+  smooth_pred_4xh(pixels, &wh[2], ww, 8, dst, stride, 1);
+}
+
+// pixels[0]: above and below_pred interleave vector, first half
+// pixels[1]: above and below_pred interleave vector, second half
+// pixels[2]: left vector
+// pixels[3]: right_pred vector
+// pixels[4]: above and below_pred interleave vector, first half
+// pixels[5]: above and below_pred interleave vector, second half
+// pixels[6]: left vector + 16
+// pixels[7]: right_pred vector
+static INLINE void load_pixel_w8(const uint8_t *above, const uint8_t *left,
+                                 int height, __m128i *pixels) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i bp = _mm_set1_epi16((int16_t)left[height - 1]);
+  __m128i d = _mm_loadl_epi64((const __m128i *)above);
+  d = _mm_unpacklo_epi8(d, zero);
+  pixels[0] = _mm_unpacklo_epi16(d, bp);
+  pixels[1] = _mm_unpackhi_epi16(d, bp);
+
+  pixels[3] = _mm_set1_epi16((int16_t)above[7]);
+
+  if (height == 4) {
+    pixels[2] = _mm_cvtsi32_si128(((const int *)left)[0]);
+  } else if (height == 8) {
+    pixels[2] = _mm_loadl_epi64((const __m128i *)left);
+  } else if (height == 16) {
+    pixels[2] = _mm_load_si128((const __m128i *)left);
+  } else {
+    pixels[2] = _mm_load_si128((const __m128i *)left);
+    pixels[4] = pixels[0];
+    pixels[5] = pixels[1];
+    pixels[6] = _mm_load_si128((const __m128i *)(left + 16));
+    pixels[7] = pixels[3];
+  }
+}
+
+// weight_h[0]: weight_h vector
+// weight_h[1]: scale - weight_h vector
+// weight_h[2]: same as [0], offset 8
+// weight_h[3]: same as [1], offset 8
+// weight_h[4]: same as [0], offset 16
+// weight_h[5]: same as [1], offset 16
+// weight_h[6]: same as [0], offset 24
+// weight_h[7]: same as [1], offset 24
+// weight_w[0]: weights_w and scale - weights_w interleave vector, first half
+// weight_w[1]: weights_w and scale - weights_w interleave vector, second half
+static INLINE void load_weight_w8(int height, __m128i *weight_h,
+                                  __m128i *weight_w) {
+  const __m128i zero = _mm_setzero_si128();
+  const int we_offset = height < 8 ? 0 : 4;
+  __m128i we = _mm_loadu_si128((const __m128i *)&smooth_weights[we_offset]);
+  weight_h[0] = _mm_unpacklo_epi8(we, zero);
+  const __m128i d = _mm_set1_epi16((int16_t)(1 << SMOOTH_WEIGHT_LOG2_SCALE));
+  weight_h[1] = _mm_sub_epi16(d, weight_h[0]);
+
+  if (height == 4) {
+    we = _mm_srli_si128(we, 4);
+    __m128i tmp1 = _mm_unpacklo_epi8(we, zero);
+    __m128i tmp2 = _mm_sub_epi16(d, tmp1);
+    weight_w[0] = _mm_unpacklo_epi16(tmp1, tmp2);
+    weight_w[1] = _mm_unpackhi_epi16(tmp1, tmp2);
+  } else {
+    weight_w[0] = _mm_unpacklo_epi16(weight_h[0], weight_h[1]);
+    weight_w[1] = _mm_unpackhi_epi16(weight_h[0], weight_h[1]);
+  }
+
+  if (height == 16) {
+    we = _mm_loadu_si128((const __m128i *)&smooth_weights[12]);
+    weight_h[0] = _mm_unpacklo_epi8(we, zero);
+    weight_h[1] = _mm_sub_epi16(d, weight_h[0]);
+    weight_h[2] = _mm_unpackhi_epi8(we, zero);
+    weight_h[3] = _mm_sub_epi16(d, weight_h[2]);
+  } else if (height == 32) {
+    const __m128i weight_lo =
+        _mm_loadu_si128((const __m128i *)&smooth_weights[28]);
+    weight_h[0] = _mm_unpacklo_epi8(weight_lo, zero);
+    weight_h[1] = _mm_sub_epi16(d, weight_h[0]);
+    weight_h[2] = _mm_unpackhi_epi8(weight_lo, zero);
+    weight_h[3] = _mm_sub_epi16(d, weight_h[2]);
+    const __m128i weight_hi =
+        _mm_loadu_si128((const __m128i *)&smooth_weights[28 + 16]);
+    weight_h[4] = _mm_unpacklo_epi8(weight_hi, zero);
+    weight_h[5] = _mm_sub_epi16(d, weight_h[4]);
+    weight_h[6] = _mm_unpackhi_epi8(weight_hi, zero);
+    weight_h[7] = _mm_sub_epi16(d, weight_h[6]);
+  }
+}
+
+static INLINE void smooth_pred_8xh(const __m128i *pixels, const __m128i *wh,
+                                   const __m128i *ww, int h, uint8_t *dst,
+                                   ptrdiff_t stride, int second_half) {
+  const __m128i round = _mm_set1_epi32((1 << SMOOTH_WEIGHT_LOG2_SCALE));
+  const __m128i one = _mm_set1_epi16(1);
+  const __m128i inc = _mm_set1_epi16(0x202);
+  const __m128i gat = _mm_set_epi32(0, 0, 0xe0c0a08, 0x6040200);
+
+  __m128i rep = second_half ? _mm_set1_epi16((short)0x8008)
+                            : _mm_set1_epi16((short)0x8000);
+  __m128i d = _mm_set1_epi16(0x100);
+
+  int i;
+  for (i = 0; i < h; ++i) {
+    const __m128i wg_wg = _mm_shuffle_epi8(wh[0], d);
+    const __m128i sc_sc = _mm_shuffle_epi8(wh[1], d);
+    const __m128i wh_sc = _mm_unpacklo_epi16(wg_wg, sc_sc);
+    __m128i s0 = _mm_madd_epi16(pixels[0], wh_sc);
+    __m128i s1 = _mm_madd_epi16(pixels[1], wh_sc);
+
+    __m128i b = _mm_shuffle_epi8(pixels[2], rep);
+    b = _mm_unpacklo_epi16(b, pixels[3]);
+    __m128i sum0 = _mm_madd_epi16(b, ww[0]);
+    __m128i sum1 = _mm_madd_epi16(b, ww[1]);
+
+    s0 = _mm_add_epi32(s0, sum0);
+    s0 = _mm_add_epi32(s0, round);
+    s0 = _mm_srai_epi32(s0, 1 + SMOOTH_WEIGHT_LOG2_SCALE);
+
+    s1 = _mm_add_epi32(s1, sum1);
+    s1 = _mm_add_epi32(s1, round);
+    s1 = _mm_srai_epi32(s1, 1 + SMOOTH_WEIGHT_LOG2_SCALE);
+
+    sum0 = _mm_packus_epi16(s0, s1);
+    sum0 = _mm_shuffle_epi8(sum0, gat);
+    _mm_storel_epi64((__m128i *)dst, sum0);
+    dst += stride;
+
+    rep = _mm_add_epi16(rep, one);
+    d = _mm_add_epi16(d, inc);
+  }
+}
+
+void aom_smooth_predictor_8x4_ssse3(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  __m128i pixels[4];
+  load_pixel_w8(above, left, 4, pixels);
+
+  __m128i wh[4], ww[2];
+  load_weight_w8(4, wh, ww);
+
+  smooth_pred_8xh(pixels, wh, ww, 4, dst, stride, 0);
+}
+
+void aom_smooth_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t stride,
+                                    const uint8_t *above, const uint8_t *left) {
+  __m128i pixels[4];
+  load_pixel_w8(above, left, 8, pixels);
+
+  __m128i wh[4], ww[2];
+  load_weight_w8(8, wh, ww);
+
+  smooth_pred_8xh(pixels, wh, ww, 8, dst, stride, 0);
+}
+
+void aom_smooth_predictor_8x16_ssse3(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  __m128i pixels[4];
+  load_pixel_w8(above, left, 16, pixels);
+
+  __m128i wh[4], ww[2];
+  load_weight_w8(16, wh, ww);
+
+  smooth_pred_8xh(pixels, wh, ww, 8, dst, stride, 0);
+  dst += stride << 3;
+  smooth_pred_8xh(pixels, &wh[2], ww, 8, dst, stride, 1);
+}
+
+void aom_smooth_predictor_8x32_ssse3(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  __m128i pixels[8];
+  load_pixel_w8(above, left, 32, pixels);
+
+  __m128i wh[8], ww[2];
+  load_weight_w8(32, wh, ww);
+
+  smooth_pred_8xh(&pixels[0], wh, ww, 8, dst, stride, 0);
+  dst += stride << 3;
+  smooth_pred_8xh(&pixels[0], &wh[2], ww, 8, dst, stride, 1);
+  dst += stride << 3;
+  smooth_pred_8xh(&pixels[4], &wh[4], ww, 8, dst, stride, 0);
+  dst += stride << 3;
+  smooth_pred_8xh(&pixels[4], &wh[6], ww, 8, dst, stride, 1);
+}
+
+// TODO(slavarnway): Visual Studio only supports restrict when /std:c11
+// (available in 2019+) or greater is specified; __restrict can be used in that
+// case. This should be moved to rtcd and used consistently between the
+// function declarations and definitions to avoid warnings in Visual Studio
+// when defining LIBAOM_RESTRICT to restrict or __restrict.
+#if defined(_MSC_VER)
+#define LIBAOM_RESTRICT
+#else
+#define LIBAOM_RESTRICT restrict
+#endif
+
+static AOM_FORCE_INLINE __m128i Load4(const void *src) {
+  // With new compilers such as clang 8.0.0 we can use the new _mm_loadu_si32
+  // intrinsic. Both _mm_loadu_si32(src) and the code here are compiled into a
+  // movss instruction.
+  //
+  // Until compiler support of _mm_loadu_si32 is widespread, use of
+  // _mm_loadu_si32 is banned.
+  int val;
+  memcpy(&val, src, sizeof(val));
+  return _mm_cvtsi32_si128(val);
+}
+
+static AOM_FORCE_INLINE __m128i LoadLo8(const void *a) {
+  return _mm_loadl_epi64((const __m128i *)(a));
+}
+
+static AOM_FORCE_INLINE __m128i LoadUnaligned16(const void *a) {
+  return _mm_loadu_si128((const __m128i *)(a));
+}
+
+static AOM_FORCE_INLINE void Store4(void *dst, const __m128i x) {
+  const int val = _mm_cvtsi128_si32(x);
+  memcpy(dst, &val, sizeof(val));
+}
+
+static AOM_FORCE_INLINE void StoreLo8(void *a, const __m128i v) {
+  _mm_storel_epi64((__m128i *)(a), v);
+}
+
+static AOM_FORCE_INLINE void StoreUnaligned16(void *a, const __m128i v) {
+  _mm_storeu_si128((__m128i *)(a), v);
+}
+
+static AOM_FORCE_INLINE __m128i cvtepu8_epi16(__m128i x) {
+  return _mm_unpacklo_epi8((x), _mm_setzero_si128());
+}
+
+static AOM_FORCE_INLINE __m128i cvtepu8_epi32(__m128i x) {
+  const __m128i tmp = _mm_unpacklo_epi8((x), _mm_setzero_si128());
+  return _mm_unpacklo_epi16(tmp, _mm_setzero_si128());
+}
+
+static AOM_FORCE_INLINE __m128i cvtepu16_epi32(__m128i x) {
+  return _mm_unpacklo_epi16((x), _mm_setzero_si128());
+}
+
+void smooth_predictor_wxh(uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
+                          const uint8_t *LIBAOM_RESTRICT top_row,
+                          const uint8_t *LIBAOM_RESTRICT left_column, int width,
+                          int height) {
+  const uint8_t *const sm_weights_h = smooth_weights + height - 4;
+  const uint8_t *const sm_weights_w = smooth_weights + width - 4;
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i scale_value = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
+  const __m128i bottom_left = _mm_cvtsi32_si128(left_column[height - 1]);
+  const __m128i top_right = _mm_set1_epi16(top_row[width - 1]);
+  const __m128i round = _mm_set1_epi32(1 << SMOOTH_WEIGHT_LOG2_SCALE);
+  for (int y = 0; y < height; ++y) {
+    const __m128i weights_y = _mm_cvtsi32_si128(sm_weights_h[y]);
+    const __m128i left_y = _mm_cvtsi32_si128(left_column[y]);
+    const __m128i scale_m_weights_y = _mm_sub_epi16(scale_value, weights_y);
+    __m128i scaled_bottom_left =
+        _mm_mullo_epi16(scale_m_weights_y, bottom_left);
+    const __m128i weight_left_y =
+        _mm_shuffle_epi32(_mm_unpacklo_epi16(weights_y, left_y), 0);
+    scaled_bottom_left = _mm_add_epi32(scaled_bottom_left, round);
+    scaled_bottom_left = _mm_shuffle_epi32(scaled_bottom_left, 0);
+    for (int x = 0; x < width; x += 8) {
+      const __m128i top_x = LoadLo8(top_row + x);
+      const __m128i weights_x = LoadLo8(sm_weights_w + x);
+      const __m128i top_weights_x = _mm_unpacklo_epi8(top_x, weights_x);
+      const __m128i top_weights_x_lo = cvtepu8_epi16(top_weights_x);
+      const __m128i top_weights_x_hi = _mm_unpackhi_epi8(top_weights_x, zero);
+
+      // Here opposite weights and pixels are multiplied, where the order of
+      // interleaving is indicated in the names.
+      __m128i pred_lo = _mm_madd_epi16(top_weights_x_lo, weight_left_y);
+      __m128i pred_hi = _mm_madd_epi16(top_weights_x_hi, weight_left_y);
+
+      // |scaled_bottom_left| is always scaled by the same weight each row, so
+      // we only derive |scaled_top_right| values here.
+      const __m128i inverted_weights_x =
+          _mm_sub_epi16(scale_value, cvtepu8_epi16(weights_x));
+      const __m128i scaled_top_right =
+          _mm_mullo_epi16(inverted_weights_x, top_right);
+      const __m128i scaled_top_right_lo = cvtepu16_epi32(scaled_top_right);
+      const __m128i scaled_top_right_hi =
+          _mm_unpackhi_epi16(scaled_top_right, zero);
+      pred_lo = _mm_add_epi32(pred_lo, scaled_bottom_left);
+      pred_hi = _mm_add_epi32(pred_hi, scaled_bottom_left);
+      pred_lo = _mm_add_epi32(pred_lo, scaled_top_right_lo);
+      pred_hi = _mm_add_epi32(pred_hi, scaled_top_right_hi);
+
+      // The round value for RightShiftWithRounding was added with
+      // |scaled_bottom_left|.
+      pred_lo = _mm_srli_epi32(pred_lo, (1 + SMOOTH_WEIGHT_LOG2_SCALE));
+      pred_hi = _mm_srli_epi32(pred_hi, (1 + SMOOTH_WEIGHT_LOG2_SCALE));
+      const __m128i pred = _mm_packus_epi16(pred_lo, pred_hi);
+      StoreLo8(dst + x, _mm_packus_epi16(pred, pred));
+    }
+    dst += stride;
+  }
+}
+
+void aom_smooth_predictor_16x4_ssse3(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  smooth_predictor_wxh(dst, stride, above, left, 16, 4);
+}
+
+void aom_smooth_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  smooth_predictor_wxh(dst, stride, above, left, 16, 8);
+}
+
+void aom_smooth_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t stride,
+                                      const uint8_t *above,
+                                      const uint8_t *left) {
+  smooth_predictor_wxh(dst, stride, above, left, 16, 16);
+}
+
+void aom_smooth_predictor_16x32_ssse3(uint8_t *dst, ptrdiff_t stride,
+                                      const uint8_t *above,
+                                      const uint8_t *left) {
+  smooth_predictor_wxh(dst, stride, above, left, 16, 32);
+}
+
+void aom_smooth_predictor_16x64_ssse3(uint8_t *dst, ptrdiff_t stride,
+                                      const uint8_t *above,
+                                      const uint8_t *left) {
+  smooth_predictor_wxh(dst, stride, above, left, 16, 64);
+}
+
+void aom_smooth_predictor_32x8_ssse3(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  smooth_predictor_wxh(dst, stride, above, left, 32, 8);
+}
+
+void aom_smooth_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t stride,
+                                      const uint8_t *above,
+                                      const uint8_t *left) {
+  smooth_predictor_wxh(dst, stride, above, left, 32, 16);
+}
+
+void aom_smooth_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t stride,
+                                      const uint8_t *above,
+                                      const uint8_t *left) {
+  smooth_predictor_wxh(dst, stride, above, left, 32, 32);
+}
+
+void aom_smooth_predictor_32x64_ssse3(uint8_t *dst, ptrdiff_t stride,
+                                      const uint8_t *above,
+                                      const uint8_t *left) {
+  smooth_predictor_wxh(dst, stride, above, left, 32, 64);
+}
+
+void aom_smooth_predictor_64x16_ssse3(uint8_t *dst, ptrdiff_t stride,
+                                      const uint8_t *above,
+                                      const uint8_t *left) {
+  smooth_predictor_wxh(dst, stride, above, left, 64, 16);
+}
+
+void aom_smooth_predictor_64x32_ssse3(uint8_t *dst, ptrdiff_t stride,
+                                      const uint8_t *above,
+                                      const uint8_t *left) {
+  smooth_predictor_wxh(dst, stride, above, left, 64, 32);
+}
+
+void aom_smooth_predictor_64x64_ssse3(uint8_t *dst, ptrdiff_t stride,
+                                      const uint8_t *above,
+                                      const uint8_t *left) {
+  smooth_predictor_wxh(dst, stride, above, left, 64, 64);
+}
+
+// -----------------------------------------------------------------------------
+// Smooth horizontal/vertical helper functions.
+
+// For Horizontal, pixels1 and pixels2 are the same repeated value. For
+// Vertical, weights1 and weights2 are the same, and scaled_corner1 and
+// scaled_corner2 are the same.
+static AOM_FORCE_INLINE void write_smooth_directional_sum16(
+    uint8_t *LIBAOM_RESTRICT dst, const __m128i pixels1, const __m128i pixels2,
+    const __m128i weights1, const __m128i weights2,
+    const __m128i scaled_corner1, const __m128i scaled_corner2,
+    const __m128i round) {
+  const __m128i weighted_px1 = _mm_mullo_epi16(pixels1, weights1);
+  const __m128i weighted_px2 = _mm_mullo_epi16(pixels2, weights2);
+  const __m128i pred_sum1 = _mm_add_epi16(scaled_corner1, weighted_px1);
+  const __m128i pred_sum2 = _mm_add_epi16(scaled_corner2, weighted_px2);
+  // Equivalent to RightShiftWithRounding(pred[x][y], 8).
+  const __m128i pred1 = _mm_srli_epi16(_mm_add_epi16(pred_sum1, round), 8);
+  const __m128i pred2 = _mm_srli_epi16(_mm_add_epi16(pred_sum2, round), 8);
+  StoreUnaligned16(dst, _mm_packus_epi16(pred1, pred2));
+}
+
+static AOM_FORCE_INLINE __m128i smooth_directional_sum8(
+    const __m128i pixels, const __m128i weights, const __m128i scaled_corner) {
+  const __m128i weighted_px = _mm_mullo_epi16(pixels, weights);
+  return _mm_add_epi16(scaled_corner, weighted_px);
+}
+
+static AOM_FORCE_INLINE void write_smooth_directional_sum8(
+    uint8_t *LIBAOM_RESTRICT dst, const __m128i *pixels, const __m128i *weights,
+    const __m128i *scaled_corner, const __m128i *round) {
+  const __m128i pred_sum =
+      smooth_directional_sum8(*pixels, *weights, *scaled_corner);
+  // Equivalent to RightShiftWithRounding(pred[x][y], 8).
+  const __m128i pred = _mm_srli_epi16(_mm_add_epi16(pred_sum, *round), 8);
+  StoreLo8(dst, _mm_packus_epi16(pred, pred));
+}
+
+// -----------------------------------------------------------------------------
+// SMOOTH_V_PRED
+
+static AOM_FORCE_INLINE void load_smooth_vertical_pixels4(
+    const uint8_t *LIBAOM_RESTRICT above, const uint8_t *LIBAOM_RESTRICT left,
+    const int height, __m128i *pixels) {
+  __m128i top = Load4(above);
+  const __m128i bottom_left = _mm_set1_epi16(left[height - 1]);
+  top = cvtepu8_epi16(top);
+  pixels[0] = _mm_unpacklo_epi16(top, bottom_left);
+}
+
+// |weight_array| alternates weight vectors from the table with their inverted
+// (256-w) counterparts. This is precomputed by the compiler when the weights
+// table is visible to this module. Removing this visibility can cut speed by up
+// to half in both 4xH and 8xH transforms.
+static AOM_FORCE_INLINE void load_smooth_vertical_weights4(
+    const uint8_t *LIBAOM_RESTRICT weight_array, const int height,
+    __m128i *weights) {
+  const __m128i inverter = _mm_set1_epi16(256);
+
+  if (height == 4) {
+    const __m128i weight = Load4(weight_array);
+    weights[0] = cvtepu8_epi16(weight);
+    weights[1] = _mm_sub_epi16(inverter, weights[0]);
+  } else if (height == 8) {
+    const __m128i weight = LoadLo8(weight_array + 4);
+    weights[0] = cvtepu8_epi16(weight);
+    weights[1] = _mm_sub_epi16(inverter, weights[0]);
+  } else {
+    const __m128i weight = LoadUnaligned16(weight_array + 12);
+    const __m128i zero = _mm_setzero_si128();
+    weights[0] = cvtepu8_epi16(weight);
+    weights[1] = _mm_sub_epi16(inverter, weights[0]);
+    weights[2] = _mm_unpackhi_epi8(weight, zero);
+    weights[3] = _mm_sub_epi16(inverter, weights[2]);
+  }
+}
+
+static AOM_FORCE_INLINE void write_smooth_vertical4xh(
+    const __m128i *pixel, const __m128i *weight, const int height,
+    uint8_t *LIBAOM_RESTRICT dst, const ptrdiff_t stride) {
+  const __m128i pred_round = _mm_set1_epi32(128);
+  const __m128i mask_increment = _mm_set1_epi16(0x0202);
+  const __m128i cvtepu8_epi32 = _mm_set1_epi32(0xC080400);
+  __m128i y_select = _mm_set1_epi16(0x0100);
+
+  for (int y = 0; y < height; ++y) {
+    const __m128i weight_y = _mm_shuffle_epi8(weight[0], y_select);
+    const __m128i inverted_weight_y = _mm_shuffle_epi8(weight[1], y_select);
+    const __m128i alternate_weights =
+        _mm_unpacklo_epi16(weight_y, inverted_weight_y);
+    // Here the pixel vector is top_row[0], corner, top_row[1], corner, ...
+    // The madd instruction yields four results of the form:
+    // (top_row[x] * weight[y] + corner * inverted_weight[y])
+    __m128i sum = _mm_madd_epi16(pixel[0], alternate_weights);
+    sum = _mm_add_epi32(sum, pred_round);
+    sum = _mm_srai_epi32(sum, 8);
+    sum = _mm_shuffle_epi8(sum, cvtepu8_epi32);
+    Store4(dst, sum);
+    dst += stride;
+    y_select = _mm_add_epi16(y_select, mask_increment);
+  }
+}
+
+void aom_smooth_v_predictor_4x4_ssse3(
+    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
+    const uint8_t *LIBAOM_RESTRICT top_row,
+    const uint8_t *LIBAOM_RESTRICT left_column) {
+  __m128i pixels;
+  load_smooth_vertical_pixels4(top_row, left_column, 4, &pixels);
+
+  __m128i weights[2];
+  load_smooth_vertical_weights4(smooth_weights, 4, weights);
+
+  write_smooth_vertical4xh(&pixels, weights, 4, dst, stride);
+}
+
+void aom_smooth_v_predictor_4x8_ssse3(
+    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
+    const uint8_t *LIBAOM_RESTRICT top_row,
+    const uint8_t *LIBAOM_RESTRICT left_column) {
+  __m128i pixels;
+  load_smooth_vertical_pixels4(top_row, left_column, 8, &pixels);
+
+  __m128i weights[2];
+  load_smooth_vertical_weights4(smooth_weights, 8, weights);
+
+  write_smooth_vertical4xh(&pixels, weights, 8, dst, stride);
+}
+
+void aom_smooth_v_predictor_4x16_ssse3(
+    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
+    const uint8_t *LIBAOM_RESTRICT top_row,
+    const uint8_t *LIBAOM_RESTRICT left_column) {
+  __m128i pixels;
+  load_smooth_vertical_pixels4(top_row, left_column, 16, &pixels);
+
+  __m128i weights[4];
+  load_smooth_vertical_weights4(smooth_weights, 16, weights);
+
+  write_smooth_vertical4xh(&pixels, weights, 8, dst, stride);
+  dst += stride << 3;
+  write_smooth_vertical4xh(&pixels, &weights[2], 8, dst, stride);
+}
+
+void aom_smooth_v_predictor_8x4_ssse3(
+    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
+    const uint8_t *LIBAOM_RESTRICT top_row,
+    const uint8_t *LIBAOM_RESTRICT left_column) {
+  const __m128i bottom_left = _mm_set1_epi16(left_column[3]);
+  const __m128i weights = cvtepu8_epi16(Load4(smooth_weights));
+  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
+  const __m128i inverted_weights = _mm_sub_epi16(scale, weights);
+  const __m128i scaled_bottom_left =
+      _mm_mullo_epi16(inverted_weights, bottom_left);
+  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
+  __m128i y_select = _mm_set1_epi32(0x01000100);
+  const __m128i top = cvtepu8_epi16(LoadLo8(top_row));
+  __m128i weights_y = _mm_shuffle_epi8(weights, y_select);
+  __m128i scaled_bottom_left_y = _mm_shuffle_epi8(scaled_bottom_left, y_select);
+  write_smooth_directional_sum8(dst, &top, &weights_y, &scaled_bottom_left_y,
+                                &round);
+  dst += stride;
+  y_select = _mm_set1_epi32(0x03020302);
+  weights_y = _mm_shuffle_epi8(weights, y_select);
+  scaled_bottom_left_y = _mm_shuffle_epi8(scaled_bottom_left, y_select);
+  write_smooth_directional_sum8(dst, &top, &weights_y, &scaled_bottom_left_y,
+                                &round);
+  dst += stride;
+  y_select = _mm_set1_epi32(0x05040504);
+  weights_y = _mm_shuffle_epi8(weights, y_select);
+  scaled_bottom_left_y = _mm_shuffle_epi8(scaled_bottom_left, y_select);
+  write_smooth_directional_sum8(dst, &top, &weights_y, &scaled_bottom_left_y,
+                                &round);
+  dst += stride;
+  y_select = _mm_set1_epi32(0x07060706);
+  weights_y = _mm_shuffle_epi8(weights, y_select);
+  scaled_bottom_left_y = _mm_shuffle_epi8(scaled_bottom_left, y_select);
+  write_smooth_directional_sum8(dst, &top, &weights_y, &scaled_bottom_left_y,
+                                &round);
+}
+
+void aom_smooth_v_predictor_8x8_ssse3(
+    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
+    const uint8_t *LIBAOM_RESTRICT top_row,
+    const uint8_t *LIBAOM_RESTRICT left_column) {
+  const __m128i bottom_left = _mm_set1_epi16(left_column[7]);
+  const __m128i weights = cvtepu8_epi16(LoadLo8(smooth_weights + 4));
+  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
+  const __m128i inverted_weights = _mm_sub_epi16(scale, weights);
+  const __m128i scaled_bottom_left =
+      _mm_mullo_epi16(inverted_weights, bottom_left);
+  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
+  const __m128i top = cvtepu8_epi16(LoadLo8(top_row));
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    const __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i weights_y = _mm_shuffle_epi8(weights, y_select);
+    const __m128i scaled_bottom_left_y =
+        _mm_shuffle_epi8(scaled_bottom_left, y_select);
+    write_smooth_directional_sum8(dst, &top, &weights_y, &scaled_bottom_left_y,
+                                  &round);
+    dst += stride;
+  }
+}
+
+void aom_smooth_v_predictor_8x16_ssse3(
+    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
+    const uint8_t *LIBAOM_RESTRICT top_row,
+    const uint8_t *LIBAOM_RESTRICT left_column) {
+  const __m128i bottom_left = _mm_set1_epi16(left_column[15]);
+  const __m128i weights = LoadUnaligned16(smooth_weights + 12);
+
+  const __m128i weights1 = cvtepu8_epi16(weights);
+  const __m128i weights2 = cvtepu8_epi16(_mm_srli_si128(weights, 8));
+  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
+  const __m128i inverted_weights1 = _mm_sub_epi16(scale, weights1);
+  const __m128i inverted_weights2 = _mm_sub_epi16(scale, weights2);
+  const __m128i scaled_bottom_left1 =
+      _mm_mullo_epi16(inverted_weights1, bottom_left);
+  const __m128i scaled_bottom_left2 =
+      _mm_mullo_epi16(inverted_weights2, bottom_left);
+  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
+  const __m128i top = cvtepu8_epi16(LoadLo8(top_row));
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    const __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i weights_y = _mm_shuffle_epi8(weights1, y_select);
+    const __m128i scaled_bottom_left_y =
+        _mm_shuffle_epi8(scaled_bottom_left1, y_select);
+    write_smooth_directional_sum8(dst, &top, &weights_y, &scaled_bottom_left_y,
+                                  &round);
+    dst += stride;
+  }
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    const __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i weights_y = _mm_shuffle_epi8(weights2, y_select);
+    const __m128i scaled_bottom_left_y =
+        _mm_shuffle_epi8(scaled_bottom_left2, y_select);
+    write_smooth_directional_sum8(dst, &top, &weights_y, &scaled_bottom_left_y,
+                                  &round);
+    dst += stride;
+  }
+}
+
+void aom_smooth_v_predictor_8x32_ssse3(
+    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
+    const uint8_t *LIBAOM_RESTRICT top_row,
+    const uint8_t *LIBAOM_RESTRICT left_column) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i bottom_left = _mm_set1_epi16(left_column[31]);
+  const __m128i weights_lo = LoadUnaligned16(smooth_weights + 28);
+  const __m128i weights_hi = LoadUnaligned16(smooth_weights + 44);
+  const __m128i weights1 = cvtepu8_epi16(weights_lo);
+  const __m128i weights2 = _mm_unpackhi_epi8(weights_lo, zero);
+  const __m128i weights3 = cvtepu8_epi16(weights_hi);
+  const __m128i weights4 = _mm_unpackhi_epi8(weights_hi, zero);
+  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
+  const __m128i inverted_weights1 = _mm_sub_epi16(scale, weights1);
+  const __m128i inverted_weights2 = _mm_sub_epi16(scale, weights2);
+  const __m128i inverted_weights3 = _mm_sub_epi16(scale, weights3);
+  const __m128i inverted_weights4 = _mm_sub_epi16(scale, weights4);
+  const __m128i scaled_bottom_left1 =
+      _mm_mullo_epi16(inverted_weights1, bottom_left);
+  const __m128i scaled_bottom_left2 =
+      _mm_mullo_epi16(inverted_weights2, bottom_left);
+  const __m128i scaled_bottom_left3 =
+      _mm_mullo_epi16(inverted_weights3, bottom_left);
+  const __m128i scaled_bottom_left4 =
+      _mm_mullo_epi16(inverted_weights4, bottom_left);
+  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
+  const __m128i top = cvtepu8_epi16(LoadLo8(top_row));
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    const __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i weights_y = _mm_shuffle_epi8(weights1, y_select);
+    const __m128i scaled_bottom_left_y =
+        _mm_shuffle_epi8(scaled_bottom_left1, y_select);
+    write_smooth_directional_sum8(dst, &top, &weights_y, &scaled_bottom_left_y,
+                                  &round);
+    dst += stride;
+  }
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    const __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i weights_y = _mm_shuffle_epi8(weights2, y_select);
+    const __m128i scaled_bottom_left_y =
+        _mm_shuffle_epi8(scaled_bottom_left2, y_select);
+    write_smooth_directional_sum8(dst, &top, &weights_y, &scaled_bottom_left_y,
+                                  &round);
+    dst += stride;
+  }
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    const __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i weights_y = _mm_shuffle_epi8(weights3, y_select);
+    const __m128i scaled_bottom_left_y =
+        _mm_shuffle_epi8(scaled_bottom_left3, y_select);
+    write_smooth_directional_sum8(dst, &top, &weights_y, &scaled_bottom_left_y,
+                                  &round);
+    dst += stride;
+  }
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    const __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i weights_y = _mm_shuffle_epi8(weights4, y_select);
+    const __m128i scaled_bottom_left_y =
+        _mm_shuffle_epi8(scaled_bottom_left4, y_select);
+    write_smooth_directional_sum8(dst, &top, &weights_y, &scaled_bottom_left_y,
+                                  &round);
+    dst += stride;
+  }
+}
+
+void aom_smooth_v_predictor_16x4_ssse3(
+    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
+    const uint8_t *LIBAOM_RESTRICT top_row,
+    const uint8_t *LIBAOM_RESTRICT left_column) {
+  const __m128i bottom_left = _mm_set1_epi16(left_column[3]);
+  const __m128i weights = cvtepu8_epi16(Load4(smooth_weights));
+  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
+  const __m128i inverted_weights = _mm_sub_epi16(scale, weights);
+  const __m128i scaled_bottom_left =
+      _mm_mullo_epi16(inverted_weights, bottom_left);
+  const __m128i round = _mm_set1_epi16(128);
+  const __m128i top = LoadUnaligned16(top_row);
+  const __m128i top_lo = cvtepu8_epi16(top);
+  const __m128i top_hi = cvtepu8_epi16(_mm_srli_si128(top, 8));
+
+  __m128i y_select = _mm_set1_epi32(0x01000100);
+  __m128i weights_y = _mm_shuffle_epi8(weights, y_select);
+  __m128i scaled_bottom_left_y = _mm_shuffle_epi8(scaled_bottom_left, y_select);
+  write_smooth_directional_sum16(dst, top_lo, top_hi, weights_y, weights_y,
+                                 scaled_bottom_left_y, scaled_bottom_left_y,
+                                 round);
+  dst += stride;
+  y_select = _mm_set1_epi32(0x03020302);
+  weights_y = _mm_shuffle_epi8(weights, y_select);
+  scaled_bottom_left_y = _mm_shuffle_epi8(scaled_bottom_left, y_select);
+  write_smooth_directional_sum16(dst, top_lo, top_hi, weights_y, weights_y,
+                                 scaled_bottom_left_y, scaled_bottom_left_y,
+                                 round);
+  dst += stride;
+  y_select = _mm_set1_epi32(0x05040504);
+  weights_y = _mm_shuffle_epi8(weights, y_select);
+  scaled_bottom_left_y = _mm_shuffle_epi8(scaled_bottom_left, y_select);
+  write_smooth_directional_sum16(dst, top_lo, top_hi, weights_y, weights_y,
+                                 scaled_bottom_left_y, scaled_bottom_left_y,
+                                 round);
+  dst += stride;
+  y_select = _mm_set1_epi32(0x07060706);
+  weights_y = _mm_shuffle_epi8(weights, y_select);
+  scaled_bottom_left_y = _mm_shuffle_epi8(scaled_bottom_left, y_select);
+  write_smooth_directional_sum16(dst, top_lo, top_hi, weights_y, weights_y,
+                                 scaled_bottom_left_y, scaled_bottom_left_y,
+                                 round);
+}
+
+void aom_smooth_v_predictor_16x8_ssse3(
+    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
+    const uint8_t *LIBAOM_RESTRICT top_row,
+    const uint8_t *LIBAOM_RESTRICT left_column) {
+  const __m128i bottom_left = _mm_set1_epi16(left_column[7]);
+  const __m128i weights = cvtepu8_epi16(LoadLo8(smooth_weights + 4));
+  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
+  const __m128i inverted_weights = _mm_sub_epi16(scale, weights);
+  const __m128i scaled_bottom_left =
+      _mm_mullo_epi16(inverted_weights, bottom_left);
+  const __m128i round = _mm_set1_epi16(128);
+  const __m128i top = LoadUnaligned16(top_row);
+  const __m128i top_lo = cvtepu8_epi16(top);
+  const __m128i top_hi = cvtepu8_epi16(_mm_srli_si128(top, 8));
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    const __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i weights_y = _mm_shuffle_epi8(weights, y_select);
+    const __m128i scaled_bottom_left_y =
+        _mm_shuffle_epi8(scaled_bottom_left, y_select);
+    write_smooth_directional_sum16(dst, top_lo, top_hi, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    dst += stride;
+  }
+}
+
+void aom_smooth_v_predictor_16x16_ssse3(
+    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
+    const uint8_t *LIBAOM_RESTRICT top_row,
+    const uint8_t *LIBAOM_RESTRICT left_column) {
+  const __m128i bottom_left = _mm_set1_epi16(left_column[15]);
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
+  const __m128i weights = LoadUnaligned16(smooth_weights + 12);
+  const __m128i weights_lo = cvtepu8_epi16(weights);
+  const __m128i weights_hi = _mm_unpackhi_epi8(weights, zero);
+  const __m128i inverted_weights_lo = _mm_sub_epi16(scale, weights_lo);
+  const __m128i inverted_weights_hi = _mm_sub_epi16(scale, weights_hi);
+  const __m128i scaled_bottom_left_lo =
+      _mm_mullo_epi16(inverted_weights_lo, bottom_left);
+  const __m128i scaled_bottom_left_hi =
+      _mm_mullo_epi16(inverted_weights_hi, bottom_left);
+  const __m128i round = _mm_set1_epi16(128);
+
+  const __m128i top = LoadUnaligned16(top_row);
+  const __m128i top_lo = cvtepu8_epi16(top);
+  const __m128i top_hi = _mm_unpackhi_epi8(top, zero);
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    const __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i weights_y = _mm_shuffle_epi8(weights_lo, y_select);
+    const __m128i scaled_bottom_left_y =
+        _mm_shuffle_epi8(scaled_bottom_left_lo, y_select);
+    write_smooth_directional_sum16(dst, top_lo, top_hi, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    dst += stride;
+  }
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    const __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i weights_y = _mm_shuffle_epi8(weights_hi, y_select);
+    const __m128i scaled_bottom_left_y =
+        _mm_shuffle_epi8(scaled_bottom_left_hi, y_select);
+    write_smooth_directional_sum16(dst, top_lo, top_hi, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    dst += stride;
+  }
+}
+
+void aom_smooth_v_predictor_16x32_ssse3(
+    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
+    const uint8_t *LIBAOM_RESTRICT top_row,
+    const uint8_t *LIBAOM_RESTRICT left_column) {
+  const __m128i bottom_left = _mm_set1_epi16(left_column[31]);
+  const __m128i weights_lo = LoadUnaligned16(smooth_weights + 28);
+  const __m128i weights_hi = LoadUnaligned16(smooth_weights + 44);
+  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i weights1 = cvtepu8_epi16(weights_lo);
+  const __m128i weights2 = _mm_unpackhi_epi8(weights_lo, zero);
+  const __m128i weights3 = cvtepu8_epi16(weights_hi);
+  const __m128i weights4 = _mm_unpackhi_epi8(weights_hi, zero);
+  const __m128i inverted_weights1 = _mm_sub_epi16(scale, weights1);
+  const __m128i inverted_weights2 = _mm_sub_epi16(scale, weights2);
+  const __m128i inverted_weights3 = _mm_sub_epi16(scale, weights3);
+  const __m128i inverted_weights4 = _mm_sub_epi16(scale, weights4);
+  const __m128i scaled_bottom_left1 =
+      _mm_mullo_epi16(inverted_weights1, bottom_left);
+  const __m128i scaled_bottom_left2 =
+      _mm_mullo_epi16(inverted_weights2, bottom_left);
+  const __m128i scaled_bottom_left3 =
+      _mm_mullo_epi16(inverted_weights3, bottom_left);
+  const __m128i scaled_bottom_left4 =
+      _mm_mullo_epi16(inverted_weights4, bottom_left);
+  const __m128i round = _mm_set1_epi16(128);
+
+  const __m128i top = LoadUnaligned16(top_row);
+  const __m128i top_lo = cvtepu8_epi16(top);
+  const __m128i top_hi = _mm_unpackhi_epi8(top, zero);
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    const __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i weights_y = _mm_shuffle_epi8(weights1, y_select);
+    const __m128i scaled_bottom_left_y =
+        _mm_shuffle_epi8(scaled_bottom_left1, y_select);
+    write_smooth_directional_sum16(dst, top_lo, top_hi, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    dst += stride;
+  }
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    const __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i weights_y = _mm_shuffle_epi8(weights2, y_select);
+    const __m128i scaled_bottom_left_y =
+        _mm_shuffle_epi8(scaled_bottom_left2, y_select);
+    write_smooth_directional_sum16(dst, top_lo, top_hi, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    dst += stride;
+  }
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    const __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i weights_y = _mm_shuffle_epi8(weights3, y_select);
+    const __m128i scaled_bottom_left_y =
+        _mm_shuffle_epi8(scaled_bottom_left3, y_select);
+    write_smooth_directional_sum16(dst, top_lo, top_hi, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    dst += stride;
+  }
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    const __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i weights_y = _mm_shuffle_epi8(weights4, y_select);
+    const __m128i scaled_bottom_left_y =
+        _mm_shuffle_epi8(scaled_bottom_left4, y_select);
+    write_smooth_directional_sum16(dst, top_lo, top_hi, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    dst += stride;
+  }
+}
+
+void aom_smooth_v_predictor_16x64_ssse3(
+    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
+    const uint8_t *LIBAOM_RESTRICT top_row,
+    const uint8_t *LIBAOM_RESTRICT left_column) {
+  const __m128i bottom_left = _mm_set1_epi16(left_column[63]);
+  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
+  const __m128i round = _mm_set1_epi16(128);
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i top = LoadUnaligned16(top_row);
+  const __m128i top_lo = cvtepu8_epi16(top);
+  const __m128i top_hi = _mm_unpackhi_epi8(top, zero);
+  const uint8_t *weights_base_ptr = smooth_weights + 60;
+  for (int left_offset = 0; left_offset < 64; left_offset += 16) {
+    const __m128i weights = LoadUnaligned16(weights_base_ptr + left_offset);
+    const __m128i weights_lo = cvtepu8_epi16(weights);
+    const __m128i weights_hi = _mm_unpackhi_epi8(weights, zero);
+    const __m128i inverted_weights_lo = _mm_sub_epi16(scale, weights_lo);
+    const __m128i inverted_weights_hi = _mm_sub_epi16(scale, weights_hi);
+    const __m128i scaled_bottom_left_lo =
+        _mm_mullo_epi16(inverted_weights_lo, bottom_left);
+    const __m128i scaled_bottom_left_hi =
+        _mm_mullo_epi16(inverted_weights_hi, bottom_left);
+
+    for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+      const __m128i y_select = _mm_set1_epi32(y_mask);
+      const __m128i weights_y = _mm_shuffle_epi8(weights_lo, y_select);
+      const __m128i scaled_bottom_left_y =
+          _mm_shuffle_epi8(scaled_bottom_left_lo, y_select);
+      write_smooth_directional_sum16(dst, top_lo, top_hi, weights_y, weights_y,
+                                     scaled_bottom_left_y, scaled_bottom_left_y,
+                                     round);
+      dst += stride;
+    }
+    for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+      const __m128i y_select = _mm_set1_epi32(y_mask);
+      const __m128i weights_y = _mm_shuffle_epi8(weights_hi, y_select);
+      const __m128i scaled_bottom_left_y =
+          _mm_shuffle_epi8(scaled_bottom_left_hi, y_select);
+      write_smooth_directional_sum16(dst, top_lo, top_hi, weights_y, weights_y,
+                                     scaled_bottom_left_y, scaled_bottom_left_y,
+                                     round);
+      dst += stride;
+    }
+  }
+}
+
+void aom_smooth_v_predictor_32x8_ssse3(
+    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
+    const uint8_t *LIBAOM_RESTRICT top_row,
+    const uint8_t *LIBAOM_RESTRICT left_column) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i bottom_left = _mm_set1_epi16(left_column[7]);
+  const __m128i top_lo = LoadUnaligned16(top_row);
+  const __m128i top_hi = LoadUnaligned16(top_row + 16);
+  const __m128i top1 = cvtepu8_epi16(top_lo);
+  const __m128i top2 = _mm_unpackhi_epi8(top_lo, zero);
+  const __m128i top3 = cvtepu8_epi16(top_hi);
+  const __m128i top4 = _mm_unpackhi_epi8(top_hi, zero);
+  __m128i scale = _mm_set1_epi16(256);
+  const __m128i weights = cvtepu8_epi16(LoadLo8(smooth_weights + 4));
+  const __m128i inverted_weights = _mm_sub_epi16(scale, weights);
+  const __m128i scaled_bottom_left =
+      _mm_mullo_epi16(inverted_weights, bottom_left);
+  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i weights_y = _mm_shuffle_epi8(weights, y_select);
+    const __m128i scaled_bottom_left_y =
+        _mm_shuffle_epi8(scaled_bottom_left, y_select);
+    write_smooth_directional_sum16(dst, top1, top2, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    write_smooth_directional_sum16(dst + 16, top3, top4, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    dst += stride;
+  }
+}
+
+void aom_smooth_v_predictor_32x16_ssse3(
+    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
+    const uint8_t *LIBAOM_RESTRICT top_row,
+    const uint8_t *LIBAOM_RESTRICT left_column) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i bottom_left = _mm_set1_epi16(left_column[15]);
+  const __m128i top_lo = LoadUnaligned16(top_row);
+  const __m128i top_hi = LoadUnaligned16(top_row + 16);
+  const __m128i top1 = cvtepu8_epi16(top_lo);
+  const __m128i top2 = _mm_unpackhi_epi8(top_lo, zero);
+  const __m128i top3 = cvtepu8_epi16(top_hi);
+  const __m128i top4 = _mm_unpackhi_epi8(top_hi, zero);
+  const __m128i weights = LoadUnaligned16(smooth_weights + 12);
+  const __m128i weights1 = cvtepu8_epi16(weights);
+  const __m128i weights2 = _mm_unpackhi_epi8(weights, zero);
+  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
+  const __m128i inverted_weights1 = _mm_sub_epi16(scale, weights1);
+  const __m128i inverted_weights2 = _mm_sub_epi16(scale, weights2);
+  const __m128i scaled_bottom_left1 =
+      _mm_mullo_epi16(inverted_weights1, bottom_left);
+  const __m128i scaled_bottom_left2 =
+      _mm_mullo_epi16(inverted_weights2, bottom_left);
+  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i weights_y = _mm_shuffle_epi8(weights1, y_select);
+    const __m128i scaled_bottom_left_y =
+        _mm_shuffle_epi8(scaled_bottom_left1, y_select);
+    write_smooth_directional_sum16(dst, top1, top2, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    write_smooth_directional_sum16(dst + 16, top3, top4, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    dst += stride;
+  }
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i weights_y = _mm_shuffle_epi8(weights2, y_select);
+    const __m128i scaled_bottom_left_y =
+        _mm_shuffle_epi8(scaled_bottom_left2, y_select);
+    write_smooth_directional_sum16(dst, top1, top2, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    write_smooth_directional_sum16(dst + 16, top3, top4, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    dst += stride;
+  }
+}
+
+void aom_smooth_v_predictor_32x32_ssse3(
+    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
+    const uint8_t *LIBAOM_RESTRICT top_row,
+    const uint8_t *LIBAOM_RESTRICT left_column) {
+  const __m128i bottom_left = _mm_set1_epi16(left_column[31]);
+  const __m128i weights_lo = LoadUnaligned16(smooth_weights + 28);
+  const __m128i weights_hi = LoadUnaligned16(smooth_weights + 44);
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
+  const __m128i top_lo = LoadUnaligned16(top_row);
+  const __m128i top_hi = LoadUnaligned16(top_row + 16);
+  const __m128i top1 = cvtepu8_epi16(top_lo);
+  const __m128i top2 = _mm_unpackhi_epi8(top_lo, zero);
+  const __m128i top3 = cvtepu8_epi16(top_hi);
+  const __m128i top4 = _mm_unpackhi_epi8(top_hi, zero);
+  const __m128i weights1 = cvtepu8_epi16(weights_lo);
+  const __m128i weights2 = _mm_unpackhi_epi8(weights_lo, zero);
+  const __m128i weights3 = cvtepu8_epi16(weights_hi);
+  const __m128i weights4 = _mm_unpackhi_epi8(weights_hi, zero);
+  const __m128i inverted_weights1 = _mm_sub_epi16(scale, weights1);
+  const __m128i inverted_weights2 = _mm_sub_epi16(scale, weights2);
+  const __m128i inverted_weights3 = _mm_sub_epi16(scale, weights3);
+  const __m128i inverted_weights4 = _mm_sub_epi16(scale, weights4);
+  const __m128i scaled_bottom_left1 =
+      _mm_mullo_epi16(inverted_weights1, bottom_left);
+  const __m128i scaled_bottom_left2 =
+      _mm_mullo_epi16(inverted_weights2, bottom_left);
+  const __m128i scaled_bottom_left3 =
+      _mm_mullo_epi16(inverted_weights3, bottom_left);
+  const __m128i scaled_bottom_left4 =
+      _mm_mullo_epi16(inverted_weights4, bottom_left);
+  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    const __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i weights_y = _mm_shuffle_epi8(weights1, y_select);
+    const __m128i scaled_bottom_left_y =
+        _mm_shuffle_epi8(scaled_bottom_left1, y_select);
+    write_smooth_directional_sum16(dst, top1, top2, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    write_smooth_directional_sum16(dst + 16, top3, top4, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    dst += stride;
+  }
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    const __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i weights_y = _mm_shuffle_epi8(weights2, y_select);
+    const __m128i scaled_bottom_left_y =
+        _mm_shuffle_epi8(scaled_bottom_left2, y_select);
+    write_smooth_directional_sum16(dst, top1, top2, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    write_smooth_directional_sum16(dst + 16, top3, top4, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    dst += stride;
+  }
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    const __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i weights_y = _mm_shuffle_epi8(weights3, y_select);
+    const __m128i scaled_bottom_left_y =
+        _mm_shuffle_epi8(scaled_bottom_left3, y_select);
+    write_smooth_directional_sum16(dst, top1, top2, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    write_smooth_directional_sum16(dst + 16, top3, top4, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    dst += stride;
+  }
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    const __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i weights_y = _mm_shuffle_epi8(weights4, y_select);
+    const __m128i scaled_bottom_left_y =
+        _mm_shuffle_epi8(scaled_bottom_left4, y_select);
+    write_smooth_directional_sum16(dst, top1, top2, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    write_smooth_directional_sum16(dst + 16, top3, top4, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    dst += stride;
+  }
+}
+
+void aom_smooth_v_predictor_32x64_ssse3(
+    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
+    const uint8_t *LIBAOM_RESTRICT top_row,
+    const uint8_t *LIBAOM_RESTRICT left_column) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i bottom_left = _mm_set1_epi16(left_column[63]);
+  const __m128i top_lo = LoadUnaligned16(top_row);
+  const __m128i top_hi = LoadUnaligned16(top_row + 16);
+  const __m128i top1 = cvtepu8_epi16(top_lo);
+  const __m128i top2 = _mm_unpackhi_epi8(top_lo, zero);
+  const __m128i top3 = cvtepu8_epi16(top_hi);
+  const __m128i top4 = _mm_unpackhi_epi8(top_hi, zero);
+  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
+  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
+  const uint8_t *weights_base_ptr = smooth_weights + 60;
+  for (int left_offset = 0; left_offset < 64; left_offset += 16) {
+    const __m128i weights = LoadUnaligned16(weights_base_ptr + left_offset);
+    const __m128i weights_lo = cvtepu8_epi16(weights);
+    const __m128i weights_hi = _mm_unpackhi_epi8(weights, zero);
+    const __m128i inverted_weights_lo = _mm_sub_epi16(scale, weights_lo);
+    const __m128i inverted_weights_hi = _mm_sub_epi16(scale, weights_hi);
+    const __m128i scaled_bottom_left_lo =
+        _mm_mullo_epi16(inverted_weights_lo, bottom_left);
+    const __m128i scaled_bottom_left_hi =
+        _mm_mullo_epi16(inverted_weights_hi, bottom_left);
+
+    for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+      const __m128i y_select = _mm_set1_epi32(y_mask);
+      const __m128i weights_y = _mm_shuffle_epi8(weights_lo, y_select);
+      const __m128i scaled_bottom_left_y =
+          _mm_shuffle_epi8(scaled_bottom_left_lo, y_select);
+      write_smooth_directional_sum16(dst, top1, top2, weights_y, weights_y,
+                                     scaled_bottom_left_y, scaled_bottom_left_y,
+                                     round);
+      write_smooth_directional_sum16(dst + 16, top3, top4, weights_y, weights_y,
+                                     scaled_bottom_left_y, scaled_bottom_left_y,
+                                     round);
+      dst += stride;
+    }
+    for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+      const __m128i y_select = _mm_set1_epi32(y_mask);
+      const __m128i weights_y = _mm_shuffle_epi8(weights_hi, y_select);
+      const __m128i scaled_bottom_left_y =
+          _mm_shuffle_epi8(scaled_bottom_left_hi, y_select);
+      write_smooth_directional_sum16(dst, top1, top2, weights_y, weights_y,
+                                     scaled_bottom_left_y, scaled_bottom_left_y,
+                                     round);
+      write_smooth_directional_sum16(dst + 16, top3, top4, weights_y, weights_y,
+                                     scaled_bottom_left_y, scaled_bottom_left_y,
+                                     round);
+      dst += stride;
+    }
+  }
+}
+
+void aom_smooth_v_predictor_64x16_ssse3(
+    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
+    const uint8_t *LIBAOM_RESTRICT top_row,
+    const uint8_t *LIBAOM_RESTRICT left_column) {
+  const __m128i bottom_left = _mm_set1_epi16(left_column[15]);
+  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i top_lolo = LoadUnaligned16(top_row);
+  const __m128i top_lohi = LoadUnaligned16(top_row + 16);
+  const __m128i top1 = cvtepu8_epi16(top_lolo);
+  const __m128i top2 = _mm_unpackhi_epi8(top_lolo, zero);
+  const __m128i top3 = cvtepu8_epi16(top_lohi);
+  const __m128i top4 = _mm_unpackhi_epi8(top_lohi, zero);
+
+  const __m128i weights = LoadUnaligned16(smooth_weights + 12);
+  const __m128i weights1 = cvtepu8_epi16(weights);
+  const __m128i weights2 = _mm_unpackhi_epi8(weights, zero);
+  const __m128i inverted_weights1 = _mm_sub_epi16(scale, weights1);
+  const __m128i inverted_weights2 = _mm_sub_epi16(scale, weights2);
+  const __m128i top_hilo = LoadUnaligned16(top_row + 32);
+  const __m128i top_hihi = LoadUnaligned16(top_row + 48);
+  const __m128i top5 = cvtepu8_epi16(top_hilo);
+  const __m128i top6 = _mm_unpackhi_epi8(top_hilo, zero);
+  const __m128i top7 = cvtepu8_epi16(top_hihi);
+  const __m128i top8 = _mm_unpackhi_epi8(top_hihi, zero);
+  const __m128i scaled_bottom_left1 =
+      _mm_mullo_epi16(inverted_weights1, bottom_left);
+  const __m128i scaled_bottom_left2 =
+      _mm_mullo_epi16(inverted_weights2, bottom_left);
+  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    const __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i weights_y = _mm_shuffle_epi8(weights1, y_select);
+    const __m128i scaled_bottom_left_y =
+        _mm_shuffle_epi8(scaled_bottom_left1, y_select);
+    write_smooth_directional_sum16(dst, top1, top2, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    write_smooth_directional_sum16(dst + 16, top3, top4, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    write_smooth_directional_sum16(dst + 32, top5, top6, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    write_smooth_directional_sum16(dst + 48, top7, top8, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    dst += stride;
+  }
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    const __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i weights_y = _mm_shuffle_epi8(weights2, y_select);
+    const __m128i scaled_bottom_left_y =
+        _mm_shuffle_epi8(scaled_bottom_left2, y_select);
+    write_smooth_directional_sum16(dst, top1, top2, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    write_smooth_directional_sum16(dst + 16, top3, top4, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    write_smooth_directional_sum16(dst + 32, top5, top6, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    write_smooth_directional_sum16(dst + 48, top7, top8, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    dst += stride;
+  }
+}
+
+void aom_smooth_v_predictor_64x32_ssse3(
+    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
+    const uint8_t *LIBAOM_RESTRICT top_row,
+    const uint8_t *LIBAOM_RESTRICT left_column) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i bottom_left = _mm_set1_epi16(left_column[31]);
+  const __m128i top_lolo = LoadUnaligned16(top_row);
+  const __m128i top_lohi = LoadUnaligned16(top_row + 16);
+  const __m128i top1 = cvtepu8_epi16(top_lolo);
+  const __m128i top2 = _mm_unpackhi_epi8(top_lolo, zero);
+  const __m128i top3 = cvtepu8_epi16(top_lohi);
+  const __m128i top4 = _mm_unpackhi_epi8(top_lohi, zero);
+  const __m128i top_hilo = LoadUnaligned16(top_row + 32);
+  const __m128i top_hihi = LoadUnaligned16(top_row + 48);
+  const __m128i top5 = cvtepu8_epi16(top_hilo);
+  const __m128i top6 = _mm_unpackhi_epi8(top_hilo, zero);
+  const __m128i top7 = cvtepu8_epi16(top_hihi);
+  const __m128i top8 = _mm_unpackhi_epi8(top_hihi, zero);
+  const __m128i weights_lo = LoadUnaligned16(smooth_weights + 28);
+  const __m128i weights_hi = LoadUnaligned16(smooth_weights + 44);
+  const __m128i weights1 = cvtepu8_epi16(weights_lo);
+  const __m128i weights2 = _mm_unpackhi_epi8(weights_lo, zero);
+  const __m128i weights3 = cvtepu8_epi16(weights_hi);
+  const __m128i weights4 = _mm_unpackhi_epi8(weights_hi, zero);
+  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
+  const __m128i inverted_weights1 = _mm_sub_epi16(scale, weights1);
+  const __m128i inverted_weights2 = _mm_sub_epi16(scale, weights2);
+  const __m128i inverted_weights3 = _mm_sub_epi16(scale, weights3);
+  const __m128i inverted_weights4 = _mm_sub_epi16(scale, weights4);
+  const __m128i scaled_bottom_left1 =
+      _mm_mullo_epi16(inverted_weights1, bottom_left);
+  const __m128i scaled_bottom_left2 =
+      _mm_mullo_epi16(inverted_weights2, bottom_left);
+  const __m128i scaled_bottom_left3 =
+      _mm_mullo_epi16(inverted_weights3, bottom_left);
+  const __m128i scaled_bottom_left4 =
+      _mm_mullo_epi16(inverted_weights4, bottom_left);
+  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
+
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    const __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i weights_y = _mm_shuffle_epi8(weights1, y_select);
+    const __m128i scaled_bottom_left_y =
+        _mm_shuffle_epi8(scaled_bottom_left1, y_select);
+    write_smooth_directional_sum16(dst, top1, top2, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    write_smooth_directional_sum16(dst + 16, top3, top4, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    write_smooth_directional_sum16(dst + 32, top5, top6, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    write_smooth_directional_sum16(dst + 48, top7, top8, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    dst += stride;
+  }
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    const __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i weights_y = _mm_shuffle_epi8(weights2, y_select);
+    const __m128i scaled_bottom_left_y =
+        _mm_shuffle_epi8(scaled_bottom_left2, y_select);
+    write_smooth_directional_sum16(dst, top1, top2, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    write_smooth_directional_sum16(dst + 16, top3, top4, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    write_smooth_directional_sum16(dst + 32, top5, top6, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    write_smooth_directional_sum16(dst + 48, top7, top8, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    dst += stride;
+  }
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    const __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i weights_y = _mm_shuffle_epi8(weights3, y_select);
+    const __m128i scaled_bottom_left_y =
+        _mm_shuffle_epi8(scaled_bottom_left3, y_select);
+    write_smooth_directional_sum16(dst, top1, top2, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    write_smooth_directional_sum16(dst + 16, top3, top4, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    write_smooth_directional_sum16(dst + 32, top5, top6, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    write_smooth_directional_sum16(dst + 48, top7, top8, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    dst += stride;
+  }
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    const __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i weights_y = _mm_shuffle_epi8(weights4, y_select);
+    const __m128i scaled_bottom_left_y =
+        _mm_shuffle_epi8(scaled_bottom_left4, y_select);
+    write_smooth_directional_sum16(dst, top1, top2, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    write_smooth_directional_sum16(dst + 16, top3, top4, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    write_smooth_directional_sum16(dst + 32, top5, top6, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    write_smooth_directional_sum16(dst + 48, top7, top8, weights_y, weights_y,
+                                   scaled_bottom_left_y, scaled_bottom_left_y,
+                                   round);
+    dst += stride;
+  }
+}
+
+void aom_smooth_v_predictor_64x64_ssse3(
+    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
+    const uint8_t *LIBAOM_RESTRICT top_row,
+    const uint8_t *LIBAOM_RESTRICT left_column) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i bottom_left = _mm_set1_epi16(left_column[63]);
+  const __m128i top_lolo = LoadUnaligned16(top_row);
+  const __m128i top_lohi = LoadUnaligned16(top_row + 16);
+  const __m128i top1 = cvtepu8_epi16(top_lolo);
+  const __m128i top2 = _mm_unpackhi_epi8(top_lolo, zero);
+  const __m128i top3 = cvtepu8_epi16(top_lohi);
+  const __m128i top4 = _mm_unpackhi_epi8(top_lohi, zero);
+  const __m128i top_hilo = LoadUnaligned16(top_row + 32);
+  const __m128i top_hihi = LoadUnaligned16(top_row + 48);
+  const __m128i top5 = cvtepu8_epi16(top_hilo);
+  const __m128i top6 = _mm_unpackhi_epi8(top_hilo, zero);
+  const __m128i top7 = cvtepu8_epi16(top_hihi);
+  const __m128i top8 = _mm_unpackhi_epi8(top_hihi, zero);
+  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
+  const __m128i round = _mm_set1_epi16(128);
+  const uint8_t *weights_base_ptr = smooth_weights + 60;
+  for (int left_offset = 0; left_offset < 64; left_offset += 16) {
+    const __m128i weights = LoadUnaligned16(weights_base_ptr + left_offset);
+    const __m128i weights_lo = cvtepu8_epi16(weights);
+    const __m128i weights_hi = _mm_unpackhi_epi8(weights, zero);
+    const __m128i inverted_weights_lo = _mm_sub_epi16(scale, weights_lo);
+    const __m128i inverted_weights_hi = _mm_sub_epi16(scale, weights_hi);
+    const __m128i scaled_bottom_left_lo =
+        _mm_mullo_epi16(inverted_weights_lo, bottom_left);
+    const __m128i scaled_bottom_left_hi =
+        _mm_mullo_epi16(inverted_weights_hi, bottom_left);
+    for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+      const __m128i y_select = _mm_set1_epi32(y_mask);
+      const __m128i weights_y = _mm_shuffle_epi8(weights_lo, y_select);
+      const __m128i scaled_bottom_left_y =
+          _mm_shuffle_epi8(scaled_bottom_left_lo, y_select);
+      write_smooth_directional_sum16(dst, top1, top2, weights_y, weights_y,
+                                     scaled_bottom_left_y, scaled_bottom_left_y,
+                                     round);
+      write_smooth_directional_sum16(dst + 16, top3, top4, weights_y, weights_y,
+                                     scaled_bottom_left_y, scaled_bottom_left_y,
+                                     round);
+      write_smooth_directional_sum16(dst + 32, top5, top6, weights_y, weights_y,
+                                     scaled_bottom_left_y, scaled_bottom_left_y,
+                                     round);
+      write_smooth_directional_sum16(dst + 48, top7, top8, weights_y, weights_y,
+                                     scaled_bottom_left_y, scaled_bottom_left_y,
+                                     round);
+      dst += stride;
+    }
+    for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+      const __m128i y_select = _mm_set1_epi32(y_mask);
+      const __m128i weights_y = _mm_shuffle_epi8(weights_hi, y_select);
+      const __m128i scaled_bottom_left_y =
+          _mm_shuffle_epi8(scaled_bottom_left_hi, y_select);
+      write_smooth_directional_sum16(dst, top1, top2, weights_y, weights_y,
+                                     scaled_bottom_left_y, scaled_bottom_left_y,
+                                     round);
+      write_smooth_directional_sum16(dst + 16, top3, top4, weights_y, weights_y,
+                                     scaled_bottom_left_y, scaled_bottom_left_y,
+                                     round);
+      write_smooth_directional_sum16(dst + 32, top5, top6, weights_y, weights_y,
+                                     scaled_bottom_left_y, scaled_bottom_left_y,
+                                     round);
+      write_smooth_directional_sum16(dst + 48, top7, top8, weights_y, weights_y,
+                                     scaled_bottom_left_y, scaled_bottom_left_y,
+                                     round);
+      dst += stride;
+    }
+  }
+}
+
+// -----------------------------------------------------------------------------
+// SMOOTH_H_PRED
+static AOM_FORCE_INLINE void write_smooth_horizontal_sum4(
+    uint8_t *LIBAOM_RESTRICT dst, const __m128i *left_y, const __m128i *weights,
+    const __m128i *scaled_top_right, const __m128i *round) {
+  const __m128i weighted_left_y = _mm_mullo_epi16(*left_y, *weights);
+  const __m128i pred_sum = _mm_add_epi32(*scaled_top_right, weighted_left_y);
+  // Equivalent to RightShiftWithRounding(pred[x][y], 8).
+  const __m128i pred = _mm_srli_epi32(_mm_add_epi32(pred_sum, *round), 8);
+  const __m128i cvtepi32_epi8 = _mm_set1_epi32(0x0C080400);
+  Store4(dst, _mm_shuffle_epi8(pred, cvtepi32_epi8));
+}
+
+void aom_smooth_h_predictor_4x4_ssse3(
+    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
+    const uint8_t *LIBAOM_RESTRICT top_row,
+    const uint8_t *LIBAOM_RESTRICT left_column) {
+  const __m128i top_right = _mm_set1_epi32(top_row[3]);
+  const __m128i left = cvtepu8_epi32(Load4(left_column));
+  const __m128i weights = cvtepu8_epi32(Load4(smooth_weights));
+  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
+  const __m128i inverted_weights = _mm_sub_epi32(scale, weights);
+  const __m128i scaled_top_right = _mm_mullo_epi16(inverted_weights, top_right);
+  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
+  __m128i left_y = _mm_shuffle_epi32(left, 0);
+  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
+                               &round);
+  dst += stride;
+  left_y = _mm_shuffle_epi32(left, 0x55);
+  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
+                               &round);
+  dst += stride;
+  left_y = _mm_shuffle_epi32(left, 0xaa);
+  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
+                               &round);
+  dst += stride;
+  left_y = _mm_shuffle_epi32(left, 0xff);
+  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
+                               &round);
+}
+
+void aom_smooth_h_predictor_4x8_ssse3(
+    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
+    const uint8_t *LIBAOM_RESTRICT top_row,
+    const uint8_t *LIBAOM_RESTRICT left_column) {
+  const __m128i top_right = _mm_set1_epi32(top_row[3]);
+  const __m128i weights = cvtepu8_epi32(Load4(smooth_weights));
+  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
+  const __m128i inverted_weights = _mm_sub_epi32(scale, weights);
+  const __m128i scaled_top_right = _mm_mullo_epi16(inverted_weights, top_right);
+  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
+  __m128i left = cvtepu8_epi32(Load4(left_column));
+  __m128i left_y = _mm_shuffle_epi32(left, 0);
+  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
+                               &round);
+  dst += stride;
+  left_y = _mm_shuffle_epi32(left, 0x55);
+  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
+                               &round);
+  dst += stride;
+  left_y = _mm_shuffle_epi32(left, 0xaa);
+  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
+                               &round);
+  dst += stride;
+  left_y = _mm_shuffle_epi32(left, 0xff);
+  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
+                               &round);
+  dst += stride;
+
+  left = cvtepu8_epi32(Load4(left_column + 4));
+  left_y = _mm_shuffle_epi32(left, 0);
+  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
+                               &round);
+  dst += stride;
+  left_y = _mm_shuffle_epi32(left, 0x55);
+  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
+                               &round);
+  dst += stride;
+  left_y = _mm_shuffle_epi32(left, 0xaa);
+  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
+                               &round);
+  dst += stride;
+  left_y = _mm_shuffle_epi32(left, 0xff);
+  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
+                               &round);
+}
+
+void aom_smooth_h_predictor_4x16_ssse3(
+    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
+    const uint8_t *LIBAOM_RESTRICT top_row,
+    const uint8_t *LIBAOM_RESTRICT left_column) {
+  const __m128i top_right = _mm_set1_epi32(top_row[3]);
+  const __m128i weights = cvtepu8_epi32(Load4(smooth_weights));
+  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
+  const __m128i inverted_weights = _mm_sub_epi32(scale, weights);
+  const __m128i scaled_top_right = _mm_mullo_epi16(inverted_weights, top_right);
+  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
+  __m128i left = cvtepu8_epi32(Load4(left_column));
+  __m128i left_y = _mm_shuffle_epi32(left, 0);
+  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
+                               &round);
+  dst += stride;
+  left_y = _mm_shuffle_epi32(left, 0x55);
+  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
+                               &round);
+  dst += stride;
+  left_y = _mm_shuffle_epi32(left, 0xaa);
+  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
+                               &round);
+  dst += stride;
+  left_y = _mm_shuffle_epi32(left, 0xff);
+  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
+                               &round);
+  dst += stride;
+
+  left = cvtepu8_epi32(Load4(left_column + 4));
+  left_y = _mm_shuffle_epi32(left, 0);
+  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
+                               &round);
+  dst += stride;
+  left_y = _mm_shuffle_epi32(left, 0x55);
+  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
+                               &round);
+  dst += stride;
+  left_y = _mm_shuffle_epi32(left, 0xaa);
+  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
+                               &round);
+  dst += stride;
+  left_y = _mm_shuffle_epi32(left, 0xff);
+  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
+                               &round);
+  dst += stride;
+
+  left = cvtepu8_epi32(Load4(left_column + 8));
+  left_y = _mm_shuffle_epi32(left, 0);
+  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
+                               &round);
+  dst += stride;
+  left_y = _mm_shuffle_epi32(left, 0x55);
+  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
+                               &round);
+  dst += stride;
+  left_y = _mm_shuffle_epi32(left, 0xaa);
+  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
+                               &round);
+  dst += stride;
+  left_y = _mm_shuffle_epi32(left, 0xff);
+  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
+                               &round);
+  dst += stride;
+
+  left = cvtepu8_epi32(Load4(left_column + 12));
+  left_y = _mm_shuffle_epi32(left, 0);
+  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
+                               &round);
+  dst += stride;
+  left_y = _mm_shuffle_epi32(left, 0x55);
+  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
+                               &round);
+  dst += stride;
+  left_y = _mm_shuffle_epi32(left, 0xaa);
+  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
+                               &round);
+  dst += stride;
+  left_y = _mm_shuffle_epi32(left, 0xff);
+  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
+                               &round);
+}
+
+// For SMOOTH_H, |pixels| is the repeated left value for the row. For SMOOTH_V,
+// |pixels| is a segment of the top row or the whole top row, and |weights| is
+// repeated.
+void aom_smooth_h_predictor_8x4_ssse3(
+    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
+    const uint8_t *LIBAOM_RESTRICT top_row,
+    const uint8_t *LIBAOM_RESTRICT left_column) {
+  const __m128i top_right = _mm_set1_epi16(top_row[7]);
+  const __m128i left = cvtepu8_epi16(Load4(left_column));
+  const __m128i weights = cvtepu8_epi16(LoadLo8(smooth_weights + 4));
+  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
+  const __m128i inverted_weights = _mm_sub_epi16(scale, weights);
+  const __m128i scaled_top_right = _mm_mullo_epi16(inverted_weights, top_right);
+  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
+  __m128i y_select = _mm_set1_epi32(0x01000100);
+  __m128i left_y = _mm_shuffle_epi8(left, y_select);
+  write_smooth_directional_sum8(dst, &left_y, &weights, &scaled_top_right,
+                                &round);
+  dst += stride;
+  y_select = _mm_set1_epi32(0x03020302);
+  left_y = _mm_shuffle_epi8(left, y_select);
+  write_smooth_directional_sum8(dst, &left_y, &weights, &scaled_top_right,
+                                &round);
+  dst += stride;
+  y_select = _mm_set1_epi32(0x05040504);
+  left_y = _mm_shuffle_epi8(left, y_select);
+  write_smooth_directional_sum8(dst, &left_y, &weights, &scaled_top_right,
+                                &round);
+  dst += stride;
+  y_select = _mm_set1_epi32(0x07060706);
+  left_y = _mm_shuffle_epi8(left, y_select);
+  write_smooth_directional_sum8(dst, &left_y, &weights, &scaled_top_right,
+                                &round);
+}
+
+void aom_smooth_h_predictor_8x8_ssse3(
+    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
+    const uint8_t *LIBAOM_RESTRICT top_row,
+    const uint8_t *LIBAOM_RESTRICT left_column) {
+  const __m128i top_right = _mm_set1_epi16(top_row[7]);
+  const __m128i left = cvtepu8_epi16(LoadLo8(left_column));
+  const __m128i weights = cvtepu8_epi16(LoadLo8(smooth_weights + 4));
+  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
+  const __m128i inverted_weights = _mm_sub_epi16(scale, weights);
+  const __m128i scaled_top_right = _mm_mullo_epi16(inverted_weights, top_right);
+  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    const __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i left_y = _mm_shuffle_epi8(left, y_select);
+    write_smooth_directional_sum8(dst, &left_y, &weights, &scaled_top_right,
+                                  &round);
+    dst += stride;
+  }
+}
+
+void aom_smooth_h_predictor_8x16_ssse3(
+    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
+    const uint8_t *LIBAOM_RESTRICT top_row,
+    const uint8_t *LIBAOM_RESTRICT left_column) {
+  const __m128i top_right = _mm_set1_epi16(top_row[7]);
+  const __m128i weights = cvtepu8_epi16(LoadLo8(smooth_weights + 4));
+  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
+  const __m128i inverted_weights = _mm_sub_epi16(scale, weights);
+  const __m128i scaled_top_right = _mm_mullo_epi16(inverted_weights, top_right);
+  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
+  __m128i left = cvtepu8_epi16(LoadLo8(left_column));
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    const __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i left_y = _mm_shuffle_epi8(left, y_select);
+    write_smooth_directional_sum8(dst, &left_y, &weights, &scaled_top_right,
+                                  &round);
+    dst += stride;
+  }
+  left = cvtepu8_epi16(LoadLo8(left_column + 8));
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    const __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i left_y = _mm_shuffle_epi8(left, y_select);
+    write_smooth_directional_sum8(dst, &left_y, &weights, &scaled_top_right,
+                                  &round);
+    dst += stride;
+  }
+}
+
+void aom_smooth_h_predictor_8x32_ssse3(
+    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
+    const uint8_t *LIBAOM_RESTRICT top_row,
+    const uint8_t *LIBAOM_RESTRICT left_column) {
+  const __m128i top_right = _mm_set1_epi16(top_row[7]);
+  const __m128i weights = cvtepu8_epi16(LoadLo8(smooth_weights + 4));
+  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
+  const __m128i inverted_weights = _mm_sub_epi16(scale, weights);
+  const __m128i scaled_top_right = _mm_mullo_epi16(inverted_weights, top_right);
+  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
+  __m128i left = cvtepu8_epi16(LoadLo8(left_column));
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    const __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i left_y = _mm_shuffle_epi8(left, y_select);
+    write_smooth_directional_sum8(dst, &left_y, &weights, &scaled_top_right,
+                                  &round);
+    dst += stride;
+  }
+  left = cvtepu8_epi16(LoadLo8(left_column + 8));
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    const __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i left_y = _mm_shuffle_epi8(left, y_select);
+    write_smooth_directional_sum8(dst, &left_y, &weights, &scaled_top_right,
+                                  &round);
+    dst += stride;
+  }
+  left = cvtepu8_epi16(LoadLo8(left_column + 16));
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    const __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i left_y = _mm_shuffle_epi8(left, y_select);
+    write_smooth_directional_sum8(dst, &left_y, &weights, &scaled_top_right,
+                                  &round);
+    dst += stride;
+  }
+  left = cvtepu8_epi16(LoadLo8(left_column + 24));
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    const __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i left_y = _mm_shuffle_epi8(left, y_select);
+    write_smooth_directional_sum8(dst, &left_y, &weights, &scaled_top_right,
+                                  &round);
+    dst += stride;
+  }
+}
+
+void aom_smooth_h_predictor_16x4_ssse3(
+    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
+    const uint8_t *LIBAOM_RESTRICT top_row,
+    const uint8_t *LIBAOM_RESTRICT left_column) {
+  const __m128i top_right = _mm_set1_epi16(top_row[15]);
+  const __m128i left = cvtepu8_epi16(Load4(left_column));
+  const __m128i weights = LoadUnaligned16(smooth_weights + 12);
+  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
+  const __m128i weights1 = cvtepu8_epi16(weights);
+  const __m128i weights2 = cvtepu8_epi16(_mm_srli_si128(weights, 8));
+  const __m128i inverted_weights1 = _mm_sub_epi16(scale, weights1);
+  const __m128i inverted_weights2 = _mm_sub_epi16(scale, weights2);
+  const __m128i scaled_top_right1 =
+      _mm_mullo_epi16(inverted_weights1, top_right);
+  const __m128i scaled_top_right2 =
+      _mm_mullo_epi16(inverted_weights2, top_right);
+  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
+  __m128i y_mask = _mm_set1_epi32(0x01000100);
+  __m128i left_y = _mm_shuffle_epi8(left, y_mask);
+  write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
+                                 scaled_top_right1, scaled_top_right2, round);
+  dst += stride;
+  y_mask = _mm_set1_epi32(0x03020302);
+  left_y = _mm_shuffle_epi8(left, y_mask);
+  write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
+                                 scaled_top_right1, scaled_top_right2, round);
+  dst += stride;
+  y_mask = _mm_set1_epi32(0x05040504);
+  left_y = _mm_shuffle_epi8(left, y_mask);
+  write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
+                                 scaled_top_right1, scaled_top_right2, round);
+  dst += stride;
+  y_mask = _mm_set1_epi32(0x07060706);
+  left_y = _mm_shuffle_epi8(left, y_mask);
+  write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
+                                 scaled_top_right1, scaled_top_right2, round);
+}
+
+void aom_smooth_h_predictor_16x8_ssse3(
+    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
+    const uint8_t *LIBAOM_RESTRICT top_row,
+    const uint8_t *LIBAOM_RESTRICT left_column) {
+  const __m128i top_right = _mm_set1_epi16(top_row[15]);
+  const __m128i left = cvtepu8_epi16(LoadLo8(left_column));
+  const __m128i weights = LoadUnaligned16(smooth_weights + 12);
+  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
+  const __m128i weights1 = cvtepu8_epi16(weights);
+  const __m128i weights2 = cvtepu8_epi16(_mm_srli_si128(weights, 8));
+  const __m128i inverted_weights1 = _mm_sub_epi16(scale, weights1);
+  const __m128i inverted_weights2 = _mm_sub_epi16(scale, weights2);
+  const __m128i scaled_top_right1 =
+      _mm_mullo_epi16(inverted_weights1, top_right);
+  const __m128i scaled_top_right2 =
+      _mm_mullo_epi16(inverted_weights2, top_right);
+  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    const __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i left_y = _mm_shuffle_epi8(left, y_select);
+    write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
+                                   scaled_top_right1, scaled_top_right2, round);
+    dst += stride;
+  }
+}
+
+void aom_smooth_h_predictor_16x16_ssse3(
+    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
+    const uint8_t *LIBAOM_RESTRICT top_row,
+    const uint8_t *LIBAOM_RESTRICT left_column) {
+  const __m128i top_right = _mm_set1_epi16(top_row[15]);
+  const __m128i weights = LoadUnaligned16(smooth_weights + 12);
+  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
+  const __m128i weights1 = cvtepu8_epi16(weights);
+  const __m128i weights2 = cvtepu8_epi16(_mm_srli_si128(weights, 8));
+  const __m128i inverted_weights1 = _mm_sub_epi16(scale, weights1);
+  const __m128i inverted_weights2 = _mm_sub_epi16(scale, weights2);
+  const __m128i scaled_top_right1 =
+      _mm_mullo_epi16(inverted_weights1, top_right);
+  const __m128i scaled_top_right2 =
+      _mm_mullo_epi16(inverted_weights2, top_right);
+  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
+  __m128i left = cvtepu8_epi16(LoadLo8(left_column));
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    const __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i left_y = _mm_shuffle_epi8(left, y_select);
+    write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
+                                   scaled_top_right1, scaled_top_right2, round);
+    dst += stride;
+  }
+  left = cvtepu8_epi16(LoadLo8(left_column + 8));
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    const __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i left_y = _mm_shuffle_epi8(left, y_select);
+    write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
+                                   scaled_top_right1, scaled_top_right2, round);
+    dst += stride;
+  }
+}
+
+void aom_smooth_h_predictor_16x32_ssse3(
+    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
+    const uint8_t *LIBAOM_RESTRICT top_row,
+    const uint8_t *LIBAOM_RESTRICT left_column) {
+  const __m128i top_right = _mm_set1_epi16(top_row[15]);
+  const __m128i weights = LoadUnaligned16(smooth_weights + 12);
+  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
+  const __m128i weights1 = cvtepu8_epi16(weights);
+  const __m128i weights2 = cvtepu8_epi16(_mm_srli_si128(weights, 8));
+  const __m128i inverted_weights1 = _mm_sub_epi16(scale, weights1);
+  const __m128i inverted_weights2 = _mm_sub_epi16(scale, weights2);
+  const __m128i scaled_top_right1 =
+      _mm_mullo_epi16(inverted_weights1, top_right);
+  const __m128i scaled_top_right2 =
+      _mm_mullo_epi16(inverted_weights2, top_right);
+  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
+  __m128i left = cvtepu8_epi16(LoadLo8(left_column));
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    const __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i left_y = _mm_shuffle_epi8(left, y_select);
+    write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
+                                   scaled_top_right1, scaled_top_right2, round);
+    dst += stride;
+  }
+  left = cvtepu8_epi16(LoadLo8(left_column + 8));
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    const __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i left_y = _mm_shuffle_epi8(left, y_select);
+    write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
+                                   scaled_top_right1, scaled_top_right2, round);
+    dst += stride;
+  }
+  left = cvtepu8_epi16(LoadLo8(left_column + 16));
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    const __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i left_y = _mm_shuffle_epi8(left, y_select);
+    write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
+                                   scaled_top_right1, scaled_top_right2, round);
+    dst += stride;
+  }
+  left = cvtepu8_epi16(LoadLo8(left_column + 24));
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    const __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i left_y = _mm_shuffle_epi8(left, y_select);
+    write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
+                                   scaled_top_right1, scaled_top_right2, round);
+    dst += stride;
+  }
+}
+
+void aom_smooth_h_predictor_16x64_ssse3(
+    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
+    const uint8_t *LIBAOM_RESTRICT top_row,
+    const uint8_t *LIBAOM_RESTRICT left_column) {
+  const __m128i top_right = _mm_set1_epi16(top_row[15]);
+  const __m128i weights = LoadUnaligned16(smooth_weights + 12);
+  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
+  const __m128i weights1 = cvtepu8_epi16(weights);
+  const __m128i weights2 = cvtepu8_epi16(_mm_srli_si128(weights, 8));
+  const __m128i inverted_weights1 = _mm_sub_epi16(scale, weights1);
+  const __m128i inverted_weights2 = _mm_sub_epi16(scale, weights2);
+  const __m128i scaled_top_right1 =
+      _mm_mullo_epi16(inverted_weights1, top_right);
+  const __m128i scaled_top_right2 =
+      _mm_mullo_epi16(inverted_weights2, top_right);
+  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
+  for (int left_offset = 0; left_offset < 64; left_offset += 8) {
+    const __m128i left = cvtepu8_epi16(LoadLo8(left_column + left_offset));
+    for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+      const __m128i y_select = _mm_set1_epi32(y_mask);
+      const __m128i left_y = _mm_shuffle_epi8(left, y_select);
+      write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
+                                     scaled_top_right1, scaled_top_right2,
+                                     round);
+      dst += stride;
+    }
+  }
+}
+
+void aom_smooth_h_predictor_32x8_ssse3(
+    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
+    const uint8_t *LIBAOM_RESTRICT top_row,
+    const uint8_t *LIBAOM_RESTRICT left_column) {
+  const __m128i top_right = _mm_set1_epi16(top_row[31]);
+  const __m128i left = cvtepu8_epi16(LoadLo8(left_column));
+  const __m128i weights_lo = LoadUnaligned16(smooth_weights + 28);
+  const __m128i weights_hi = LoadUnaligned16(smooth_weights + 44);
+  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
+  const __m128i weights1 = cvtepu8_epi16(weights_lo);
+  const __m128i weights2 = cvtepu8_epi16(_mm_srli_si128(weights_lo, 8));
+  const __m128i weights3 = cvtepu8_epi16(weights_hi);
+  const __m128i weights4 = cvtepu8_epi16(_mm_srli_si128(weights_hi, 8));
+  const __m128i inverted_weights1 = _mm_sub_epi16(scale, weights1);
+  const __m128i inverted_weights2 = _mm_sub_epi16(scale, weights2);
+  const __m128i inverted_weights3 = _mm_sub_epi16(scale, weights3);
+  const __m128i inverted_weights4 = _mm_sub_epi16(scale, weights4);
+  const __m128i scaled_top_right1 =
+      _mm_mullo_epi16(inverted_weights1, top_right);
+  const __m128i scaled_top_right2 =
+      _mm_mullo_epi16(inverted_weights2, top_right);
+  const __m128i scaled_top_right3 =
+      _mm_mullo_epi16(inverted_weights3, top_right);
+  const __m128i scaled_top_right4 =
+      _mm_mullo_epi16(inverted_weights4, top_right);
+  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    __m128i y_select = _mm_set1_epi32(y_mask);
+    __m128i left_y = _mm_shuffle_epi8(left, y_select);
+    write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
+                                   scaled_top_right1, scaled_top_right2, round);
+    write_smooth_directional_sum16(dst + 16, left_y, left_y, weights3, weights4,
+                                   scaled_top_right3, scaled_top_right4, round);
+    dst += stride;
+  }
+}
+
+void aom_smooth_h_predictor_32x16_ssse3(
+    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
+    const uint8_t *LIBAOM_RESTRICT top_row,
+    const uint8_t *LIBAOM_RESTRICT left_column) {
+  const __m128i top_right = _mm_set1_epi16(top_row[31]);
+  const __m128i left1 = cvtepu8_epi16(LoadLo8(left_column));
+  const __m128i weights_lo = LoadUnaligned16(smooth_weights + 28);
+  const __m128i weights_hi = LoadUnaligned16(smooth_weights + 44);
+  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
+  const __m128i weights1 = cvtepu8_epi16(weights_lo);
+  const __m128i weights2 = cvtepu8_epi16(_mm_srli_si128(weights_lo, 8));
+  const __m128i weights3 = cvtepu8_epi16(weights_hi);
+  const __m128i weights4 = cvtepu8_epi16(_mm_srli_si128(weights_hi, 8));
+  const __m128i inverted_weights1 = _mm_sub_epi16(scale, weights1);
+  const __m128i inverted_weights2 = _mm_sub_epi16(scale, weights2);
+  const __m128i inverted_weights3 = _mm_sub_epi16(scale, weights3);
+  const __m128i inverted_weights4 = _mm_sub_epi16(scale, weights4);
+  const __m128i scaled_top_right1 =
+      _mm_mullo_epi16(inverted_weights1, top_right);
+  const __m128i scaled_top_right2 =
+      _mm_mullo_epi16(inverted_weights2, top_right);
+  const __m128i scaled_top_right3 =
+      _mm_mullo_epi16(inverted_weights3, top_right);
+  const __m128i scaled_top_right4 =
+      _mm_mullo_epi16(inverted_weights4, top_right);
+  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    __m128i y_select = _mm_set1_epi32(y_mask);
+    __m128i left_y = _mm_shuffle_epi8(left1, y_select);
+    write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
+                                   scaled_top_right1, scaled_top_right2, round);
+    write_smooth_directional_sum16(dst + 16, left_y, left_y, weights3, weights4,
+                                   scaled_top_right3, scaled_top_right4, round);
+    dst += stride;
+  }
+  const __m128i left2 =
+      cvtepu8_epi16(LoadLo8((const uint8_t *)left_column + 8));
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    __m128i y_select = _mm_set1_epi32(y_mask);
+    __m128i left_y = _mm_shuffle_epi8(left2, y_select);
+    write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
+                                   scaled_top_right1, scaled_top_right2, round);
+    write_smooth_directional_sum16(dst + 16, left_y, left_y, weights3, weights4,
+                                   scaled_top_right3, scaled_top_right4, round);
+    dst += stride;
+  }
+}
+
+void aom_smooth_h_predictor_32x32_ssse3(
+    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
+    const uint8_t *LIBAOM_RESTRICT top_row,
+    const uint8_t *LIBAOM_RESTRICT left_column) {
+  const __m128i top_right = _mm_set1_epi16(top_row[31]);
+  const __m128i weights_lo = LoadUnaligned16(smooth_weights + 28);
+  const __m128i weights_hi = LoadUnaligned16(smooth_weights + 44);
+  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
+  const __m128i weights1 = cvtepu8_epi16(weights_lo);
+  const __m128i weights2 = cvtepu8_epi16(_mm_srli_si128(weights_lo, 8));
+  const __m128i weights3 = cvtepu8_epi16(weights_hi);
+  const __m128i weights4 = cvtepu8_epi16(_mm_srli_si128(weights_hi, 8));
+  const __m128i inverted_weights1 = _mm_sub_epi16(scale, weights1);
+  const __m128i inverted_weights2 = _mm_sub_epi16(scale, weights2);
+  const __m128i inverted_weights3 = _mm_sub_epi16(scale, weights3);
+  const __m128i inverted_weights4 = _mm_sub_epi16(scale, weights4);
+  const __m128i scaled_top_right1 =
+      _mm_mullo_epi16(inverted_weights1, top_right);
+  const __m128i scaled_top_right2 =
+      _mm_mullo_epi16(inverted_weights2, top_right);
+  const __m128i scaled_top_right3 =
+      _mm_mullo_epi16(inverted_weights3, top_right);
+  const __m128i scaled_top_right4 =
+      _mm_mullo_epi16(inverted_weights4, top_right);
+  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
+  __m128i left = cvtepu8_epi16(LoadLo8(left_column));
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    __m128i y_select = _mm_set1_epi32(y_mask);
+    __m128i left_y = _mm_shuffle_epi8(left, y_select);
+    write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
+                                   scaled_top_right1, scaled_top_right2, round);
+    write_smooth_directional_sum16(dst + 16, left_y, left_y, weights3, weights4,
+                                   scaled_top_right3, scaled_top_right4, round);
+    dst += stride;
+  }
+  left = cvtepu8_epi16(LoadLo8(left_column + 8));
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    __m128i y_select = _mm_set1_epi32(y_mask);
+    __m128i left_y = _mm_shuffle_epi8(left, y_select);
+    write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
+                                   scaled_top_right1, scaled_top_right2, round);
+    write_smooth_directional_sum16(dst + 16, left_y, left_y, weights3, weights4,
+                                   scaled_top_right3, scaled_top_right4, round);
+    dst += stride;
+  }
+  left = cvtepu8_epi16(LoadLo8(left_column + 16));
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    __m128i y_select = _mm_set1_epi32(y_mask);
+    __m128i left_y = _mm_shuffle_epi8(left, y_select);
+    write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
+                                   scaled_top_right1, scaled_top_right2, round);
+    write_smooth_directional_sum16(dst + 16, left_y, left_y, weights3, weights4,
+                                   scaled_top_right3, scaled_top_right4, round);
+    dst += stride;
+  }
+  left = cvtepu8_epi16(LoadLo8(left_column + 24));
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    __m128i y_select = _mm_set1_epi32(y_mask);
+    __m128i left_y = _mm_shuffle_epi8(left, y_select);
+    write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
+                                   scaled_top_right1, scaled_top_right2, round);
+    write_smooth_directional_sum16(dst + 16, left_y, left_y, weights3, weights4,
+                                   scaled_top_right3, scaled_top_right4, round);
+    dst += stride;
+  }
+}
+
+void aom_smooth_h_predictor_32x64_ssse3(
+    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
+    const uint8_t *LIBAOM_RESTRICT top_row,
+    const uint8_t *LIBAOM_RESTRICT left_column) {
+  const __m128i top_right = _mm_set1_epi16(top_row[31]);
+  const __m128i weights_lo = LoadUnaligned16(smooth_weights + 28);
+  const __m128i weights_hi = LoadUnaligned16(smooth_weights + 44);
+  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
+  const __m128i weights1 = cvtepu8_epi16(weights_lo);
+  const __m128i weights2 = cvtepu8_epi16(_mm_srli_si128(weights_lo, 8));
+  const __m128i weights3 = cvtepu8_epi16(weights_hi);
+  const __m128i weights4 = cvtepu8_epi16(_mm_srli_si128(weights_hi, 8));
+  const __m128i inverted_weights1 = _mm_sub_epi16(scale, weights1);
+  const __m128i inverted_weights2 = _mm_sub_epi16(scale, weights2);
+  const __m128i inverted_weights3 = _mm_sub_epi16(scale, weights3);
+  const __m128i inverted_weights4 = _mm_sub_epi16(scale, weights4);
+  const __m128i scaled_top_right1 =
+      _mm_mullo_epi16(inverted_weights1, top_right);
+  const __m128i scaled_top_right2 =
+      _mm_mullo_epi16(inverted_weights2, top_right);
+  const __m128i scaled_top_right3 =
+      _mm_mullo_epi16(inverted_weights3, top_right);
+  const __m128i scaled_top_right4 =
+      _mm_mullo_epi16(inverted_weights4, top_right);
+  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
+  for (int left_offset = 0; left_offset < 64; left_offset += 8) {
+    const __m128i left = cvtepu8_epi16(LoadLo8(left_column + left_offset));
+    for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+      const __m128i y_select = _mm_set1_epi32(y_mask);
+      const __m128i left_y = _mm_shuffle_epi8(left, y_select);
+      write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
+                                     scaled_top_right1, scaled_top_right2,
+                                     round);
+      write_smooth_directional_sum16(dst + 16, left_y, left_y, weights3,
+                                     weights4, scaled_top_right3,
+                                     scaled_top_right4, round);
+      dst += stride;
+    }
+  }
+}
+
+void aom_smooth_h_predictor_64x16_ssse3(
+    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
+    const uint8_t *LIBAOM_RESTRICT top_row,
+    const uint8_t *LIBAOM_RESTRICT left_column) {
+  const __m128i top_right = _mm_set1_epi16(top_row[63]);
+  const __m128i left1 = cvtepu8_epi16(LoadLo8(left_column));
+  const __m128i weights_lolo = LoadUnaligned16(smooth_weights + 60);
+  const __m128i weights_lohi = LoadUnaligned16(smooth_weights + 76);
+  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
+  const __m128i weights1 = cvtepu8_epi16(weights_lolo);
+  const __m128i weights2 = cvtepu8_epi16(_mm_srli_si128(weights_lolo, 8));
+  const __m128i weights3 = cvtepu8_epi16(weights_lohi);
+  const __m128i weights4 = cvtepu8_epi16(_mm_srli_si128(weights_lohi, 8));
+  const __m128i inverted_weights1 = _mm_sub_epi16(scale, weights1);
+  const __m128i inverted_weights2 = _mm_sub_epi16(scale, weights2);
+  const __m128i inverted_weights3 = _mm_sub_epi16(scale, weights3);
+  const __m128i inverted_weights4 = _mm_sub_epi16(scale, weights4);
+  const __m128i scaled_top_right1 =
+      _mm_mullo_epi16(inverted_weights1, top_right);
+  const __m128i scaled_top_right2 =
+      _mm_mullo_epi16(inverted_weights2, top_right);
+  const __m128i scaled_top_right3 =
+      _mm_mullo_epi16(inverted_weights3, top_right);
+  const __m128i scaled_top_right4 =
+      _mm_mullo_epi16(inverted_weights4, top_right);
+  const __m128i weights_hilo = LoadUnaligned16(smooth_weights + 92);
+  const __m128i weights_hihi = LoadUnaligned16(smooth_weights + 108);
+  const __m128i weights5 = cvtepu8_epi16(weights_hilo);
+  const __m128i weights6 = cvtepu8_epi16(_mm_srli_si128(weights_hilo, 8));
+  const __m128i weights7 = cvtepu8_epi16(weights_hihi);
+  const __m128i weights8 = cvtepu8_epi16(_mm_srli_si128(weights_hihi, 8));
+  const __m128i inverted_weights5 = _mm_sub_epi16(scale, weights5);
+  const __m128i inverted_weights6 = _mm_sub_epi16(scale, weights6);
+  const __m128i inverted_weights7 = _mm_sub_epi16(scale, weights7);
+  const __m128i inverted_weights8 = _mm_sub_epi16(scale, weights8);
+  const __m128i scaled_top_right5 =
+      _mm_mullo_epi16(inverted_weights5, top_right);
+  const __m128i scaled_top_right6 =
+      _mm_mullo_epi16(inverted_weights6, top_right);
+  const __m128i scaled_top_right7 =
+      _mm_mullo_epi16(inverted_weights7, top_right);
+  const __m128i scaled_top_right8 =
+      _mm_mullo_epi16(inverted_weights8, top_right);
+  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    __m128i y_select = _mm_set1_epi32(y_mask);
+    __m128i left_y = _mm_shuffle_epi8(left1, y_select);
+    write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
+                                   scaled_top_right1, scaled_top_right2, round);
+    write_smooth_directional_sum16(dst + 16, left_y, left_y, weights3, weights4,
+                                   scaled_top_right3, scaled_top_right4, round);
+    write_smooth_directional_sum16(dst + 32, left_y, left_y, weights5, weights6,
+                                   scaled_top_right5, scaled_top_right6, round);
+    write_smooth_directional_sum16(dst + 48, left_y, left_y, weights7, weights8,
+                                   scaled_top_right7, scaled_top_right8, round);
+    dst += stride;
+  }
+  const __m128i left2 = cvtepu8_epi16(LoadLo8(left_column + 8));
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    __m128i y_select = _mm_set1_epi32(y_mask);
+    __m128i left_y = _mm_shuffle_epi8(left2, y_select);
+    write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
+                                   scaled_top_right1, scaled_top_right2, round);
+    write_smooth_directional_sum16(dst + 16, left_y, left_y, weights3, weights4,
+                                   scaled_top_right3, scaled_top_right4, round);
+    write_smooth_directional_sum16(dst + 32, left_y, left_y, weights5, weights6,
+                                   scaled_top_right5, scaled_top_right6, round);
+    write_smooth_directional_sum16(dst + 48, left_y, left_y, weights7, weights8,
+                                   scaled_top_right7, scaled_top_right8, round);
+    dst += stride;
+  }
+}
+
+void aom_smooth_h_predictor_64x32_ssse3(
+    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
+    const uint8_t *LIBAOM_RESTRICT top_row,
+    const uint8_t *LIBAOM_RESTRICT left_column) {
+  const __m128i top_right = _mm_set1_epi16(top_row[63]);
+  const __m128i left1 = cvtepu8_epi16(LoadLo8(left_column));
+  const __m128i weights_lolo = LoadUnaligned16(smooth_weights + 60);
+  const __m128i weights_lohi = LoadUnaligned16(smooth_weights + 76);
+  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
+  const __m128i weights1 = cvtepu8_epi16(weights_lolo);
+  const __m128i weights2 = cvtepu8_epi16(_mm_srli_si128(weights_lolo, 8));
+  const __m128i weights3 = cvtepu8_epi16(weights_lohi);
+  const __m128i weights4 = cvtepu8_epi16(_mm_srli_si128(weights_lohi, 8));
+  const __m128i inverted_weights1 = _mm_sub_epi16(scale, weights1);
+  const __m128i inverted_weights2 = _mm_sub_epi16(scale, weights2);
+  const __m128i inverted_weights3 = _mm_sub_epi16(scale, weights3);
+  const __m128i inverted_weights4 = _mm_sub_epi16(scale, weights4);
+  const __m128i scaled_top_right1 =
+      _mm_mullo_epi16(inverted_weights1, top_right);
+  const __m128i scaled_top_right2 =
+      _mm_mullo_epi16(inverted_weights2, top_right);
+  const __m128i scaled_top_right3 =
+      _mm_mullo_epi16(inverted_weights3, top_right);
+  const __m128i scaled_top_right4 =
+      _mm_mullo_epi16(inverted_weights4, top_right);
+  const __m128i weights_hilo = LoadUnaligned16(smooth_weights + 92);
+  const __m128i weights_hihi = LoadUnaligned16(smooth_weights + 108);
+  const __m128i weights5 = cvtepu8_epi16(weights_hilo);
+  const __m128i weights6 = cvtepu8_epi16(_mm_srli_si128(weights_hilo, 8));
+  const __m128i weights7 = cvtepu8_epi16(weights_hihi);
+  const __m128i weights8 = cvtepu8_epi16(_mm_srli_si128(weights_hihi, 8));
+  const __m128i inverted_weights5 = _mm_sub_epi16(scale, weights5);
+  const __m128i inverted_weights6 = _mm_sub_epi16(scale, weights6);
+  const __m128i inverted_weights7 = _mm_sub_epi16(scale, weights7);
+  const __m128i inverted_weights8 = _mm_sub_epi16(scale, weights8);
+  const __m128i scaled_top_right5 =
+      _mm_mullo_epi16(inverted_weights5, top_right);
+  const __m128i scaled_top_right6 =
+      _mm_mullo_epi16(inverted_weights6, top_right);
+  const __m128i scaled_top_right7 =
+      _mm_mullo_epi16(inverted_weights7, top_right);
+  const __m128i scaled_top_right8 =
+      _mm_mullo_epi16(inverted_weights8, top_right);
+  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    const __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i left_y = _mm_shuffle_epi8(left1, y_select);
+    write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
+                                   scaled_top_right1, scaled_top_right2, round);
+    write_smooth_directional_sum16(dst + 16, left_y, left_y, weights3, weights4,
+                                   scaled_top_right3, scaled_top_right4, round);
+    write_smooth_directional_sum16(dst + 32, left_y, left_y, weights5, weights6,
+                                   scaled_top_right5, scaled_top_right6, round);
+    write_smooth_directional_sum16(dst + 48, left_y, left_y, weights7, weights8,
+                                   scaled_top_right7, scaled_top_right8, round);
+    dst += stride;
+  }
+  const __m128i left2 = cvtepu8_epi16(LoadLo8(left_column + 8));
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    const __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i left_y = _mm_shuffle_epi8(left2, y_select);
+    write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
+                                   scaled_top_right1, scaled_top_right2, round);
+    write_smooth_directional_sum16(dst + 16, left_y, left_y, weights3, weights4,
+                                   scaled_top_right3, scaled_top_right4, round);
+    write_smooth_directional_sum16(dst + 32, left_y, left_y, weights5, weights6,
+                                   scaled_top_right5, scaled_top_right6, round);
+    write_smooth_directional_sum16(dst + 48, left_y, left_y, weights7, weights8,
+                                   scaled_top_right7, scaled_top_right8, round);
+    dst += stride;
+  }
+  const __m128i left3 = cvtepu8_epi16(LoadLo8(left_column + 16));
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    const __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i left_y = _mm_shuffle_epi8(left3, y_select);
+    write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
+                                   scaled_top_right1, scaled_top_right2, round);
+    write_smooth_directional_sum16(dst + 16, left_y, left_y, weights3, weights4,
+                                   scaled_top_right3, scaled_top_right4, round);
+    write_smooth_directional_sum16(dst + 32, left_y, left_y, weights5, weights6,
+                                   scaled_top_right5, scaled_top_right6, round);
+    write_smooth_directional_sum16(dst + 48, left_y, left_y, weights7, weights8,
+                                   scaled_top_right7, scaled_top_right8, round);
+    dst += stride;
+  }
+  const __m128i left4 = cvtepu8_epi16(LoadLo8(left_column + 24));
+  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+    const __m128i y_select = _mm_set1_epi32(y_mask);
+    const __m128i left_y = _mm_shuffle_epi8(left4, y_select);
+    write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
+                                   scaled_top_right1, scaled_top_right2, round);
+    write_smooth_directional_sum16(dst + 16, left_y, left_y, weights3, weights4,
+                                   scaled_top_right3, scaled_top_right4, round);
+    write_smooth_directional_sum16(dst + 32, left_y, left_y, weights5, weights6,
+                                   scaled_top_right5, scaled_top_right6, round);
+    write_smooth_directional_sum16(dst + 48, left_y, left_y, weights7, weights8,
+                                   scaled_top_right7, scaled_top_right8, round);
+    dst += stride;
+  }
+}
+
+void aom_smooth_h_predictor_64x64_ssse3(
+    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
+    const uint8_t *LIBAOM_RESTRICT top_row,
+    const uint8_t *LIBAOM_RESTRICT left_column) {
+  const __m128i top_right = _mm_set1_epi16(top_row[63]);
+  const __m128i weights_lolo = LoadUnaligned16(smooth_weights + 60);
+  const __m128i weights_lohi = LoadUnaligned16(smooth_weights + 76);
+  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
+  const __m128i weights1 = cvtepu8_epi16(weights_lolo);
+  const __m128i weights2 = cvtepu8_epi16(_mm_srli_si128(weights_lolo, 8));
+  const __m128i weights3 = cvtepu8_epi16(weights_lohi);
+  const __m128i weights4 = cvtepu8_epi16(_mm_srli_si128(weights_lohi, 8));
+  const __m128i inverted_weights1 = _mm_sub_epi16(scale, weights1);
+  const __m128i inverted_weights2 = _mm_sub_epi16(scale, weights2);
+  const __m128i inverted_weights3 = _mm_sub_epi16(scale, weights3);
+  const __m128i inverted_weights4 = _mm_sub_epi16(scale, weights4);
+  const __m128i scaled_top_right1 =
+      _mm_mullo_epi16(inverted_weights1, top_right);
+  const __m128i scaled_top_right2 =
+      _mm_mullo_epi16(inverted_weights2, top_right);
+  const __m128i scaled_top_right3 =
+      _mm_mullo_epi16(inverted_weights3, top_right);
+  const __m128i scaled_top_right4 =
+      _mm_mullo_epi16(inverted_weights4, top_right);
+  const __m128i weights_hilo = LoadUnaligned16(smooth_weights + 92);
+  const __m128i weights_hihi = LoadUnaligned16(smooth_weights + 108);
+  const __m128i weights5 = cvtepu8_epi16(weights_hilo);
+  const __m128i weights6 = cvtepu8_epi16(_mm_srli_si128(weights_hilo, 8));
+  const __m128i weights7 = cvtepu8_epi16(weights_hihi);
+  const __m128i weights8 = cvtepu8_epi16(_mm_srli_si128(weights_hihi, 8));
+  const __m128i inverted_weights5 = _mm_sub_epi16(scale, weights5);
+  const __m128i inverted_weights6 = _mm_sub_epi16(scale, weights6);
+  const __m128i inverted_weights7 = _mm_sub_epi16(scale, weights7);
+  const __m128i inverted_weights8 = _mm_sub_epi16(scale, weights8);
+  const __m128i scaled_top_right5 =
+      _mm_mullo_epi16(inverted_weights5, top_right);
+  const __m128i scaled_top_right6 =
+      _mm_mullo_epi16(inverted_weights6, top_right);
+  const __m128i scaled_top_right7 =
+      _mm_mullo_epi16(inverted_weights7, top_right);
+  const __m128i scaled_top_right8 =
+      _mm_mullo_epi16(inverted_weights8, top_right);
+  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
+  for (int left_offset = 0; left_offset < 64; left_offset += 8) {
+    const __m128i left = cvtepu8_epi16(LoadLo8(left_column + left_offset));
+    for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
+      const __m128i y_select = _mm_set1_epi32(y_mask);
+      const __m128i left_y = _mm_shuffle_epi8(left, y_select);
+      write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
+                                     scaled_top_right1, scaled_top_right2,
+                                     round);
+      write_smooth_directional_sum16(dst + 16, left_y, left_y, weights3,
+                                     weights4, scaled_top_right3,
+                                     scaled_top_right4, round);
+      write_smooth_directional_sum16(dst + 32, left_y, left_y, weights5,
+                                     weights6, scaled_top_right5,
+                                     scaled_top_right6, round);
+      write_smooth_directional_sum16(dst + 48, left_y, left_y, weights7,
+                                     weights8, scaled_top_right7,
+                                     scaled_top_right8, round);
+      dst += stride;
+    }
+  }
+}
diff --git a/third_party/aom/aom_dsp/x86/intrapred_utils.h b/third_party/aom/aom_dsp/x86/intrapred_utils.h
new file mode 100644
index 0000000000..502574673e
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/intrapred_utils.h
@@ -0,0 +1,205 @@
+/*
+ * Copyright (c) 2021, 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_AOM_DSP_X86_INTRAPRED_UTILS_H_
+#define AOM_AOM_DSP_X86_INTRAPRED_UTILS_H_
+
+#include <emmintrin.h>  // SSE2
+#include "aom/aom_integer.h"
+#include "config/aom_config.h"
+#include "config/aom_dsp_rtcd.h"
+
+static DECLARE_ALIGNED(16, uint8_t, EvenOddMaskx[8][16]) = {
+  { 0, 2, 4, 6, 8, 10, 12, 14, 1, 3, 5, 7, 9, 11, 13, 15 },
+  { 0, 1, 3, 5, 7, 9, 11, 13, 0, 2, 4, 6, 8, 10, 12, 14 },
+  { 0, 0, 2, 4, 6, 8, 10, 12, 0, 0, 3, 5, 7, 9, 11, 13 },
+  { 0, 0, 0, 3, 5, 7, 9, 11, 0, 0, 0, 4, 6, 8, 10, 12 },
+  { 0, 0, 0, 0, 4, 6, 8, 10, 0, 0, 0, 0, 5, 7, 9, 11 },
+  { 0, 0, 0, 0, 0, 5, 7, 9, 0, 0, 0, 0, 0, 6, 8, 10 },
+  { 0, 0, 0, 0, 0, 0, 6, 8, 0, 0, 0, 0, 0, 0, 7, 9 },
+  { 0, 0, 0, 0, 0, 0, 0, 7, 0, 0, 0, 0, 0, 0, 0, 8 }
+};
+
+static DECLARE_ALIGNED(16, uint8_t, LoadMaskx[16][16]) = {
+  { 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 },
+  { 0, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14 },
+  { 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13 },
+  { 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12 },
+  { 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11 },
+  { 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10 },
+  { 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9 },
+  { 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7, 8 },
+  { 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6, 7 },
+  { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5, 6 },
+  { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4, 5 },
+  { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 4 },
+  { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3 },
+  { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2 },
+  { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1 },
+  { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 },
+};
+
+static DECLARE_ALIGNED(32, int, LoadMaskz2[8][8]) = {
+  { -1, 0, 0, 0, 0, 0, 0, 0 },       { -1, -1, 0, 0, 0, 0, 0, 0 },
+  { -1, -1, -1, 0, 0, 0, 0, 0 },     { -1, -1, -1, -1, 0, 0, 0, 0 },
+  { -1, -1, -1, -1, -1, 0, 0, 0 },   { -1, -1, -1, -1, -1, -1, 0, 0 },
+  { -1, -1, -1, -1, -1, -1, -1, 0 }, { -1, -1, -1, -1, -1, -1, -1, -1 },
+};
+
+static INLINE void transpose4x16_sse2(__m128i *x, __m128i *d) {
+  __m128i w0, w1, w2, w3, ww0, ww1, ww2, ww3;
+  w0 = _mm_unpacklo_epi8(x[0], x[1]);
+  w1 = _mm_unpacklo_epi8(x[2], x[3]);
+  w2 = _mm_unpackhi_epi8(x[0], x[1]);
+  w3 = _mm_unpackhi_epi8(x[2], x[3]);
+
+  ww0 = _mm_unpacklo_epi16(w0, w1);
+  ww1 = _mm_unpacklo_epi16(w2, w3);
+  ww2 = _mm_unpackhi_epi16(w0, w1);
+  ww3 = _mm_unpackhi_epi16(w2, w3);
+
+  w0 = _mm_unpacklo_epi32(ww0, ww1);
+  w2 = _mm_unpacklo_epi32(ww2, ww3);
+  w1 = _mm_unpackhi_epi32(ww0, ww1);
+  w3 = _mm_unpackhi_epi32(ww2, ww3);
+
+  d[0] = _mm_unpacklo_epi64(w0, w2);
+  d[1] = _mm_unpackhi_epi64(w0, w2);
+  d[2] = _mm_unpacklo_epi64(w1, w3);
+  d[3] = _mm_unpackhi_epi64(w1, w3);
+
+  d[4] = _mm_srli_si128(d[0], 8);
+  d[5] = _mm_srli_si128(d[1], 8);
+  d[6] = _mm_srli_si128(d[2], 8);
+  d[7] = _mm_srli_si128(d[3], 8);
+
+  d[8] = _mm_srli_si128(d[0], 4);
+  d[9] = _mm_srli_si128(d[1], 4);
+  d[10] = _mm_srli_si128(d[2], 4);
+  d[11] = _mm_srli_si128(d[3], 4);
+
+  d[12] = _mm_srli_si128(d[0], 12);
+  d[13] = _mm_srli_si128(d[1], 12);
+  d[14] = _mm_srli_si128(d[2], 12);
+  d[15] = _mm_srli_si128(d[3], 12);
+}
+
+static INLINE void transpose16x16_sse2(__m128i *x, __m128i *d) {
+  __m128i w0, w1, w2, w3, w4, w5, w6, w7, w8, w9;
+  __m128i w10, w11, w12, w13, w14, w15;
+
+  w0 = _mm_unpacklo_epi8(x[0], x[1]);
+  w1 = _mm_unpacklo_epi8(x[2], x[3]);
+  w2 = _mm_unpacklo_epi8(x[4], x[5]);
+  w3 = _mm_unpacklo_epi8(x[6], x[7]);
+
+  w8 = _mm_unpacklo_epi8(x[8], x[9]);
+  w9 = _mm_unpacklo_epi8(x[10], x[11]);
+  w10 = _mm_unpacklo_epi8(x[12], x[13]);
+  w11 = _mm_unpacklo_epi8(x[14], x[15]);
+
+  w4 = _mm_unpacklo_epi16(w0, w1);
+  w5 = _mm_unpacklo_epi16(w2, w3);
+  w12 = _mm_unpacklo_epi16(w8, w9);
+  w13 = _mm_unpacklo_epi16(w10, w11);
+
+  w6 = _mm_unpacklo_epi32(w4, w5);
+  w7 = _mm_unpackhi_epi32(w4, w5);
+  w14 = _mm_unpacklo_epi32(w12, w13);
+  w15 = _mm_unpackhi_epi32(w12, w13);
+
+  // Store first 4-line result
+  d[0] = _mm_unpacklo_epi64(w6, w14);
+  d[1] = _mm_unpackhi_epi64(w6, w14);
+  d[2] = _mm_unpacklo_epi64(w7, w15);
+  d[3] = _mm_unpackhi_epi64(w7, w15);
+
+  w4 = _mm_unpackhi_epi16(w0, w1);
+  w5 = _mm_unpackhi_epi16(w2, w3);
+  w12 = _mm_unpackhi_epi16(w8, w9);
+  w13 = _mm_unpackhi_epi16(w10, w11);
+
+  w6 = _mm_unpacklo_epi32(w4, w5);
+  w7 = _mm_unpackhi_epi32(w4, w5);
+  w14 = _mm_unpacklo_epi32(w12, w13);
+  w15 = _mm_unpackhi_epi32(w12, w13);
+
+  // Store second 4-line result
+  d[4] = _mm_unpacklo_epi64(w6, w14);
+  d[5] = _mm_unpackhi_epi64(w6, w14);
+  d[6] = _mm_unpacklo_epi64(w7, w15);
+  d[7] = _mm_unpackhi_epi64(w7, w15);
+
+  // upper half
+  w0 = _mm_unpackhi_epi8(x[0], x[1]);
+  w1 = _mm_unpackhi_epi8(x[2], x[3]);
+  w2 = _mm_unpackhi_epi8(x[4], x[5]);
+  w3 = _mm_unpackhi_epi8(x[6], x[7]);
+
+  w8 = _mm_unpackhi_epi8(x[8], x[9]);
+  w9 = _mm_unpackhi_epi8(x[10], x[11]);
+  w10 = _mm_unpackhi_epi8(x[12], x[13]);
+  w11 = _mm_unpackhi_epi8(x[14], x[15]);
+
+  w4 = _mm_unpacklo_epi16(w0, w1);
+  w5 = _mm_unpacklo_epi16(w2, w3);
+  w12 = _mm_unpacklo_epi16(w8, w9);
+  w13 = _mm_unpacklo_epi16(w10, w11);
+
+  w6 = _mm_unpacklo_epi32(w4, w5);
+  w7 = _mm_unpackhi_epi32(w4, w5);
+  w14 = _mm_unpacklo_epi32(w12, w13);
+  w15 = _mm_unpackhi_epi32(w12, w13);
+
+  // Store first 4-line result
+  d[8] = _mm_unpacklo_epi64(w6, w14);
+  d[9] = _mm_unpackhi_epi64(w6, w14);
+  d[10] = _mm_unpacklo_epi64(w7, w15);
+  d[11] = _mm_unpackhi_epi64(w7, w15);
+
+  w4 = _mm_unpackhi_epi16(w0, w1);
+  w5 = _mm_unpackhi_epi16(w2, w3);
+  w12 = _mm_unpackhi_epi16(w8, w9);
+  w13 = _mm_unpackhi_epi16(w10, w11);
+
+  w6 = _mm_unpacklo_epi32(w4, w5);
+  w7 = _mm_unpackhi_epi32(w4, w5);
+  w14 = _mm_unpacklo_epi32(w12, w13);
+  w15 = _mm_unpackhi_epi32(w12, w13);
+
+  // Store second 4-line result
+  d[12] = _mm_unpacklo_epi64(w6, w14);
+  d[13] = _mm_unpackhi_epi64(w6, w14);
+  d[14] = _mm_unpacklo_epi64(w7, w15);
+  d[15] = _mm_unpackhi_epi64(w7, w15);
+}
+
+static void transpose_TX_16X16(const uint8_t *src, ptrdiff_t pitchSrc,
+                               uint8_t *dst, ptrdiff_t pitchDst) {
+  __m128i r[16];
+  __m128i d[16];
+  for (int j = 0; j < 16; j++) {
+    r[j] = _mm_loadu_si128((__m128i *)(src + j * pitchSrc));
+  }
+  transpose16x16_sse2(r, d);
+  for (int j = 0; j < 16; j++) {
+    _mm_storeu_si128((__m128i *)(dst + j * pitchDst), d[j]);
+  }
+}
+
+static void transpose(const uint8_t *src, ptrdiff_t pitchSrc, uint8_t *dst,
+                      ptrdiff_t pitchDst, int width, int height) {
+  for (int j = 0; j < height; j += 16)
+    for (int i = 0; i < width; i += 16)
+      transpose_TX_16X16(src + i * pitchSrc + j, pitchSrc,
+                         dst + j * pitchDst + i, pitchDst);
+}
+
+#endif  // AOM_AOM_DSP_X86_INTRAPRED_UTILS_H_
diff --git a/third_party/aom/aom_dsp/x86/intrapred_x86.h b/third_party/aom/aom_dsp/x86/intrapred_x86.h
new file mode 100644
index 0000000000..b13f575a76
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/intrapred_x86.h
@@ -0,0 +1,38 @@
+/*
+ * Copyright (c) 2020, 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_AOM_DSP_X86_INTRAPRED_X86_H_
+#define AOM_AOM_DSP_X86_INTRAPRED_X86_H_
+
+#include <emmintrin.h>  // SSE2
+#include "aom/aom_integer.h"
+#include "config/aom_config.h"
+
+static INLINE __m128i dc_sum_16_sse2(const uint8_t *ref) {
+  __m128i x = _mm_load_si128((__m128i const *)ref);
+  const __m128i zero = _mm_setzero_si128();
+  x = _mm_sad_epu8(x, zero);
+  const __m128i high = _mm_unpackhi_epi64(x, x);
+  return _mm_add_epi16(x, high);
+}
+
+static INLINE __m128i dc_sum_32_sse2(const uint8_t *ref) {
+  __m128i x0 = _mm_load_si128((__m128i const *)ref);
+  __m128i x1 = _mm_load_si128((__m128i const *)(ref + 16));
+  const __m128i zero = _mm_setzero_si128();
+  x0 = _mm_sad_epu8(x0, zero);
+  x1 = _mm_sad_epu8(x1, zero);
+  x0 = _mm_add_epi16(x0, x1);
+  const __m128i high = _mm_unpackhi_epi64(x0, x0);
+  return _mm_add_epi16(x0, high);
+}
+
+#endif  // AOM_AOM_DSP_X86_INTRAPRED_X86_H_
diff --git a/third_party/aom/aom_dsp/x86/inv_wht_sse2.asm b/third_party/aom/aom_dsp/x86/inv_wht_sse2.asm
new file mode 100644
index 0000000000..0bc841a7a4
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/inv_wht_sse2.asm
@@ -0,0 +1,107 @@
+;
+; Copyright (c) 2016, 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 "third_party/x86inc/x86inc.asm"
+
+SECTION .text
+
+%macro REORDER_INPUTS 0
+  ; a c d b  to  a b c d
+  SWAP 1, 3, 2
+%endmacro
+
+%macro TRANSFORM_COLS 0
+  ; input:
+  ; m0 a
+  ; m1 b
+  ; m2 c
+  ; m3 d
+  paddw           m0,        m2
+  psubw           m3,        m1
+
+  ; wide subtract
+  punpcklwd       m4,        m0
+  punpcklwd       m5,        m3
+  psrad           m4,        16
+  psrad           m5,        16
+  psubd           m4,        m5
+  psrad           m4,        1
+  packssdw        m4,        m4             ; e
+
+  psubw           m5,        m4,        m1  ; b
+  psubw           m4,        m2             ; c
+  psubw           m0,        m5
+  paddw           m3,        m4
+                                ; m0 a
+  SWAP            1,         5  ; m1 b
+  SWAP            2,         4  ; m2 c
+                                ; m3 d
+%endmacro
+
+%macro TRANSPOSE_4X4 0
+  punpcklwd       m0,        m2
+  punpcklwd       m1,        m3
+  mova            m2,        m0
+  punpcklwd       m0,        m1
+  punpckhwd       m2,        m1
+  pshufd          m1,        m0, 0x0e
+  pshufd          m3,        m2, 0x0e
+%endmacro
+
+; transpose a 4x4 int16 matrix in xmm0 and xmm1 to the bottom half of xmm0-xmm3
+%macro TRANSPOSE_4X4_WIDE 0
+  mova            m3, m0
+  punpcklwd       m0, m1
+  punpckhwd       m3, m1
+  mova            m2, m0
+  punpcklwd       m0, m3
+  punpckhwd       m2, m3
+  pshufd          m1, m0, 0x0e
+  pshufd          m3, m2, 0x0e
+%endmacro
+
+%macro ADD_STORE_4P_2X 5  ; src1, src2, tmp1, tmp2, zero
+  movd            m%3,       [outputq]
+  movd            m%4,       [outputq + strideq]
+  punpcklbw       m%3,       m%5
+  punpcklbw       m%4,       m%5
+  paddw           m%1,       m%3
+  paddw           m%2,       m%4
+  packuswb        m%1,       m%5
+  packuswb        m%2,       m%5
+  movd            [outputq], m%1
+  movd            [outputq + strideq], m%2
+%endmacro
+
+INIT_XMM sse2
+cglobal iwht4x4_16_add, 3, 3, 7, input, output, stride
+  mova            m0,        [inputq +  0]
+  packssdw        m0,        [inputq + 16]
+  mova            m1,        [inputq + 32]
+  packssdw        m1,        [inputq + 48]
+  psraw           m0,        2
+  psraw           m1,        2
+
+  TRANSPOSE_4X4_WIDE
+  REORDER_INPUTS
+  TRANSFORM_COLS
+  TRANSPOSE_4X4
+  REORDER_INPUTS
+  TRANSFORM_COLS
+
+  pxor            m4, m4
+  ADD_STORE_4P_2X  0, 1, 5, 6, 4
+  lea             outputq, [outputq + 2 * strideq]
+  ADD_STORE_4P_2X  2, 3, 5, 6, 4
+
+  RET
diff --git a/third_party/aom/aom_dsp/x86/jnt_sad_sse2.c b/third_party/aom/aom_dsp/x86/jnt_sad_sse2.c
new file mode 100644
index 0000000000..16d2f4be7f
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/jnt_sad_sse2.c
@@ -0,0 +1,238 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <emmintrin.h>
+
+#include "config/aom_config.h"
+#include "config/aom_dsp_rtcd.h"
+
+#include "aom_dsp/x86/synonyms.h"
+
+static unsigned int sad4xh_sse2(const uint8_t *a, int a_stride,
+                                const uint8_t *b, int b_stride, int width,
+                                int height) {
+  int i;
+  assert(width == 4);
+  (void)width;
+
+  __m128i sad = _mm_setzero_si128();
+  for (i = 0; i < height; i += 4) {
+    __m128i x0 = xx_loadl_32(a + 0 * a_stride);
+    __m128i x1 = xx_loadl_32(a + 1 * a_stride);
+    __m128i x2 = xx_loadl_32(a + 2 * a_stride);
+    __m128i x3 = xx_loadl_32(a + 3 * a_stride);
+    __m128i x_lo = _mm_unpacklo_epi32(x0, x1);
+    __m128i x_hi = _mm_unpacklo_epi32(x2, x3);
+
+    __m128i x = _mm_unpacklo_epi64(x_lo, x_hi);
+
+    x0 = xx_loadl_32(b + 0 * b_stride);
+    x1 = xx_loadl_32(b + 1 * b_stride);
+    x2 = xx_loadl_32(b + 2 * b_stride);
+    x3 = xx_loadl_32(b + 3 * b_stride);
+    x_lo = _mm_unpacklo_epi32(x0, x1);
+    x_hi = _mm_unpacklo_epi32(x2, x3);
+
+    __m128i y = _mm_unpacklo_epi64(x_lo, x_hi);
+
+    __m128i sad4x4 = _mm_sad_epu8(x, y);
+    sad = _mm_add_epi32(sad, sad4x4);
+
+    a += 4 * a_stride;
+    b += 4 * b_stride;
+  }
+
+  // At this point, we have two 32-bit partial SADs at bit[0:31] and [64:95].
+  const unsigned int res =
+      (unsigned int)(_mm_cvtsi128_si32(sad) +
+                     _mm_cvtsi128_si32(_mm_srli_si128(sad, 8)));
+
+  return res;
+}
+
+static unsigned int sad8xh_sse2(const uint8_t *a, int a_stride,
+                                const uint8_t *b, int b_stride, int width,
+                                int height) {
+  int i;
+  assert(width == 8);
+  (void)width;
+
+  __m128i sad = _mm_setzero_si128();
+  for (i = 0; i < height; i += 2) {
+    __m128i x0 = xx_loadl_64(a + 0 * a_stride);
+    __m128i x1 = xx_loadl_64(a + 1 * a_stride);
+
+    __m128i x = _mm_unpacklo_epi64(x0, x1);
+
+    x0 = xx_loadl_64(b + 0 * b_stride);
+    x1 = xx_loadl_64(b + 1 * b_stride);
+
+    __m128i y = _mm_unpacklo_epi64(x0, x1);
+
+    __m128i sad8x2 = _mm_sad_epu8(x, y);
+    sad = _mm_add_epi32(sad, sad8x2);
+
+    a += 2 * a_stride;
+    b += 2 * b_stride;
+  }
+
+  const unsigned int res =
+      (unsigned int)(_mm_cvtsi128_si32(sad) +
+                     _mm_cvtsi128_si32(_mm_srli_si128(sad, 8)));
+
+  return res;
+}
+
+static unsigned int sad16xh_sse2(const uint8_t *a, int a_stride,
+                                 const uint8_t *b, int b_stride, int width,
+                                 int height) {
+  int i;
+  assert(width == 16);
+  (void)width;
+
+  __m128i sad = _mm_setzero_si128();
+  for (i = 0; i < height; ++i) {
+    __m128i x = xx_loadu_128(a);
+    __m128i y = xx_loadu_128(b);
+
+    __m128i sad16x1 = _mm_sad_epu8(x, y);
+    sad = _mm_add_epi32(sad, sad16x1);
+
+    a += a_stride;
+    b += b_stride;
+  }
+
+  const unsigned int res =
+      (unsigned int)(_mm_cvtsi128_si32(sad) +
+                     _mm_cvtsi128_si32(_mm_srli_si128(sad, 8)));
+
+  return res;
+}
+
+static unsigned int sad32xh_sse2(const uint8_t *a, int a_stride,
+                                 const uint8_t *b, int b_stride, int width,
+                                 int height) {
+  int i, j;
+  assert(width == 32);
+  (void)width;
+
+  __m128i sad = _mm_setzero_si128();
+  for (i = 0; i < height; ++i) {
+    for (j = 0; j < 2; ++j) {
+      __m128i x = xx_loadu_128(a + j * 16);
+      __m128i y = xx_loadu_128(b + j * 16);
+
+      __m128i sad32_half = _mm_sad_epu8(x, y);
+      sad = _mm_add_epi32(sad, sad32_half);
+    }
+
+    a += a_stride;
+    b += b_stride;
+  }
+
+  const unsigned int res =
+      (unsigned int)(_mm_cvtsi128_si32(sad) +
+                     _mm_cvtsi128_si32(_mm_srli_si128(sad, 8)));
+
+  return res;
+}
+
+static unsigned int sad64xh_sse2(const uint8_t *a, int a_stride,
+                                 const uint8_t *b, int b_stride, int width,
+                                 int height) {
+  int i, j;
+  assert(width == 64);
+  (void)width;
+
+  __m128i sad = _mm_setzero_si128();
+  for (i = 0; i < height; ++i) {
+    for (j = 0; j < 4; ++j) {
+      __m128i x = xx_loadu_128(a + j * 16);
+      __m128i y = xx_loadu_128(b + j * 16);
+
+      __m128i sad64_quarter = _mm_sad_epu8(x, y);
+      sad = _mm_add_epi32(sad, sad64_quarter);
+    }
+
+    a += a_stride;
+    b += b_stride;
+  }
+
+  const unsigned int res =
+      (unsigned int)(_mm_cvtsi128_si32(sad) +
+                     _mm_cvtsi128_si32(_mm_srli_si128(sad, 8)));
+
+  return res;
+}
+
+static unsigned int sad128xh_sse2(const uint8_t *a, int a_stride,
+                                  const uint8_t *b, int b_stride, int width,
+                                  int height) {
+  int i, j;
+  assert(width == 128);
+  (void)width;
+
+  __m128i sad = _mm_setzero_si128();
+  for (i = 0; i < height; ++i) {
+    for (j = 0; j < 8; ++j) {
+      __m128i x = xx_loadu_128(a + j * 16);
+      __m128i y = xx_loadu_128(b + j * 16);
+
+      __m128i sad64_quarter = _mm_sad_epu8(x, y);
+      sad = _mm_add_epi32(sad, sad64_quarter);
+    }
+
+    a += a_stride;
+    b += b_stride;
+  }
+
+  const unsigned int res =
+      (unsigned int)(_mm_cvtsi128_si32(sad) +
+                     _mm_cvtsi128_si32(_mm_srli_si128(sad, 8)));
+
+  return res;
+}
+
+#define DIST_WTD_SADMXN_SSE2(m, n)                                            \
+  unsigned int aom_dist_wtd_sad##m##x##n##_avg_sse2(                          \
+      const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, \
+      const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param) {    \
+    uint8_t comp_pred[m * n];                                                 \
+    aom_dist_wtd_comp_avg_pred(comp_pred, second_pred, m, n, ref, ref_stride, \
+                               jcp_param);                                    \
+    return sad##m##xh_sse2(src, src_stride, comp_pred, m, m, n);              \
+  }
+
+DIST_WTD_SADMXN_SSE2(128, 128)
+DIST_WTD_SADMXN_SSE2(128, 64)
+DIST_WTD_SADMXN_SSE2(64, 128)
+DIST_WTD_SADMXN_SSE2(64, 64)
+DIST_WTD_SADMXN_SSE2(64, 32)
+DIST_WTD_SADMXN_SSE2(32, 64)
+DIST_WTD_SADMXN_SSE2(32, 32)
+DIST_WTD_SADMXN_SSE2(32, 16)
+DIST_WTD_SADMXN_SSE2(16, 32)
+DIST_WTD_SADMXN_SSE2(16, 16)
+DIST_WTD_SADMXN_SSE2(16, 8)
+DIST_WTD_SADMXN_SSE2(8, 16)
+DIST_WTD_SADMXN_SSE2(8, 8)
+DIST_WTD_SADMXN_SSE2(8, 4)
+DIST_WTD_SADMXN_SSE2(4, 8)
+DIST_WTD_SADMXN_SSE2(4, 4)
+#if !CONFIG_REALTIME_ONLY
+DIST_WTD_SADMXN_SSE2(4, 16)
+DIST_WTD_SADMXN_SSE2(16, 4)
+DIST_WTD_SADMXN_SSE2(8, 32)
+DIST_WTD_SADMXN_SSE2(32, 8)
+DIST_WTD_SADMXN_SSE2(16, 64)
+DIST_WTD_SADMXN_SSE2(64, 16)
+#endif
diff --git a/third_party/aom/aom_dsp/x86/jnt_variance_ssse3.c b/third_party/aom/aom_dsp/x86/jnt_variance_ssse3.c
new file mode 100644
index 0000000000..dd798ca54a
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/jnt_variance_ssse3.c
@@ -0,0 +1,161 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <emmintrin.h>  // SSE2
+#include <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);
+
+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 compute_dist_wtd_avg(__m128i *p0, __m128i *p1,
+                                        const __m128i *w, const __m128i *r,
+                                        void *const result) {
+  __m128i p_lo = _mm_unpacklo_epi8(*p0, *p1);
+  __m128i mult_lo = _mm_maddubs_epi16(p_lo, *w);
+  __m128i round_lo = _mm_add_epi16(mult_lo, *r);
+  __m128i shift_lo = _mm_srai_epi16(round_lo, DIST_PRECISION_BITS);
+
+  __m128i p_hi = _mm_unpackhi_epi8(*p0, *p1);
+  __m128i mult_hi = _mm_maddubs_epi16(p_hi, *w);
+  __m128i round_hi = _mm_add_epi16(mult_hi, *r);
+  __m128i shift_hi = _mm_srai_epi16(round_hi, DIST_PRECISION_BITS);
+
+  xx_storeu_128(result, _mm_packus_epi16(shift_lo, shift_hi));
+}
+
+void aom_dist_wtd_comp_avg_pred_ssse3(uint8_t *comp_pred, const uint8_t *pred,
+                                      int width, int height, const uint8_t *ref,
+                                      int ref_stride,
+                                      const DIST_WTD_COMP_PARAMS *jcp_param) {
+  int i;
+  const int8_t w0 = (int8_t)jcp_param->fwd_offset;
+  const int8_t w1 = (int8_t)jcp_param->bck_offset;
+  const __m128i w = _mm_set_epi8(w1, w0, w1, w0, w1, w0, w1, w0, w1, w0, w1, w0,
+                                 w1, w0, w1, w0);
+  const int16_t round = (int16_t)((1 << DIST_PRECISION_BITS) >> 1);
+  const __m128i r = _mm_set1_epi16(round);
+
+  if (width >= 16) {
+    // Read 16 pixels one row at a time
+    assert(!(width & 15));
+    for (i = 0; i < height; ++i) {
+      int j;
+      for (j = 0; j < width; j += 16) {
+        __m128i p0 = xx_loadu_128(ref);
+        __m128i p1 = xx_loadu_128(pred);
+
+        compute_dist_wtd_avg(&p0, &p1, &w, &r, comp_pred);
+
+        comp_pred += 16;
+        pred += 16;
+        ref += 16;
+      }
+      ref += ref_stride - width;
+    }
+  } else if (width >= 8) {
+    // Read 8 pixels two row at a time
+    assert(!(width & 7));
+    assert(!(width & 1));
+    for (i = 0; i < height; i += 2) {
+      __m128i p0_0 = xx_loadl_64(ref + 0 * ref_stride);
+      __m128i p0_1 = xx_loadl_64(ref + 1 * ref_stride);
+      __m128i p0 = _mm_unpacklo_epi64(p0_0, p0_1);
+      __m128i p1 = xx_loadu_128(pred);
+
+      compute_dist_wtd_avg(&p0, &p1, &w, &r, comp_pred);
+
+      comp_pred += 16;
+      pred += 16;
+      ref += 2 * ref_stride;
+    }
+  } else {
+    // Read 4 pixels four row at a time
+    assert(!(width & 3));
+    assert(!(height & 3));
+    for (i = 0; i < height; i += 4) {
+      const int8_t *row0 = (const int8_t *)ref + 0 * ref_stride;
+      const int8_t *row1 = (const int8_t *)ref + 1 * ref_stride;
+      const int8_t *row2 = (const int8_t *)ref + 2 * ref_stride;
+      const int8_t *row3 = (const int8_t *)ref + 3 * ref_stride;
+
+      __m128i p0 =
+          _mm_setr_epi8(row0[0], row0[1], row0[2], row0[3], row1[0], row1[1],
+                        row1[2], row1[3], row2[0], row2[1], row2[2], row2[3],
+                        row3[0], row3[1], row3[2], row3[3]);
+      __m128i p1 = xx_loadu_128(pred);
+
+      compute_dist_wtd_avg(&p0, &p1, &w, &r, comp_pred);
+
+      comp_pred += 16;
+      pred += 16;
+      ref += 4 * ref_stride;
+    }
+  }
+}
+
+#define DIST_WTD_SUBPIX_AVG_VAR(W, H)                                      \
+  uint32_t aom_dist_wtd_sub_pixel_avg_variance##W##x##H##_ssse3(           \
+      const uint8_t *a, int a_stride, int xoffset, int yoffset,            \
+      const uint8_t *b, int b_stride, uint32_t *sse,                       \
+      const uint8_t *second_pred, const DIST_WTD_COMP_PARAMS *jcp_param) { \
+    uint16_t fdata3[(H + 1) * W];                                          \
+    uint8_t temp2[H * W];                                                  \
+    DECLARE_ALIGNED(16, uint8_t, temp3[H * W]);                            \
+                                                                           \
+    aom_var_filter_block2d_bil_first_pass_ssse3(                           \
+        a, fdata3, a_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]);          \
+                                                                           \
+    aom_dist_wtd_comp_avg_pred_ssse3(temp3, second_pred, W, H, temp2, W,   \
+                                     jcp_param);                           \
+                                                                           \
+    return aom_variance##W##x##H(temp3, W, b, b_stride, sse);              \
+  }
+
+DIST_WTD_SUBPIX_AVG_VAR(128, 128)
+DIST_WTD_SUBPIX_AVG_VAR(128, 64)
+DIST_WTD_SUBPIX_AVG_VAR(64, 128)
+DIST_WTD_SUBPIX_AVG_VAR(64, 64)
+DIST_WTD_SUBPIX_AVG_VAR(64, 32)
+DIST_WTD_SUBPIX_AVG_VAR(32, 64)
+DIST_WTD_SUBPIX_AVG_VAR(32, 32)
+DIST_WTD_SUBPIX_AVG_VAR(32, 16)
+DIST_WTD_SUBPIX_AVG_VAR(16, 32)
+DIST_WTD_SUBPIX_AVG_VAR(16, 16)
+DIST_WTD_SUBPIX_AVG_VAR(16, 8)
+DIST_WTD_SUBPIX_AVG_VAR(8, 16)
+DIST_WTD_SUBPIX_AVG_VAR(8, 8)
+DIST_WTD_SUBPIX_AVG_VAR(8, 4)
+DIST_WTD_SUBPIX_AVG_VAR(4, 8)
+DIST_WTD_SUBPIX_AVG_VAR(4, 4)
+
+#if !CONFIG_REALTIME_ONLY
+DIST_WTD_SUBPIX_AVG_VAR(4, 16)
+DIST_WTD_SUBPIX_AVG_VAR(16, 4)
+DIST_WTD_SUBPIX_AVG_VAR(8, 32)
+DIST_WTD_SUBPIX_AVG_VAR(32, 8)
+DIST_WTD_SUBPIX_AVG_VAR(16, 64)
+DIST_WTD_SUBPIX_AVG_VAR(64, 16)
+#endif
diff --git a/third_party/aom/aom_dsp/x86/loopfilter_avx2.c b/third_party/aom/aom_dsp/x86/loopfilter_avx2.c
new file mode 100644
index 0000000000..6e77742e3c
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/loopfilter_avx2.c
@@ -0,0 +1,1016 @@
+/*
+ * Copyright (c) 2022, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <immintrin.h> /* AVX2 */
+
+#include "config/aom_dsp_rtcd.h"
+
+DECLARE_ALIGNED(32, static const uint8_t, filt_loopfilter_avx2[32]) = {
+  0, 128, 1, 128, 2,  128, 3,  128, 4,  128, 5,  128, 6,  128, 7,  128,
+  8, 128, 9, 128, 10, 128, 11, 128, 12, 128, 13, 128, 14, 128, 15, 128
+};
+
+void aom_lpf_horizontal_6_quad_avx2(unsigned char *s, int p,
+                                    const unsigned char *_blimit0,
+                                    const unsigned char *_limit0,
+                                    const unsigned char *_thresh0) {
+  __m256i p256_2, q256_2, p256_1, q256_1, p256_0, q256_0;
+  __m128i p2, p1, p0, q0, q1, q2;
+  __m128i mask, flat;
+
+  const __m128i thresh_v =
+      _mm_broadcastb_epi8(_mm_cvtsi32_si128((int)_thresh0[0]));
+  const __m128i limit_v =
+      _mm_broadcastb_epi8(_mm_cvtsi32_si128((int)_limit0[0]));
+  const __m128i blimit_v =
+      _mm_broadcastb_epi8(_mm_cvtsi32_si128((int)_blimit0[0]));
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i ff = _mm_cmpeq_epi8(zero, zero);
+
+  p256_2 =
+      _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 3 * p)));
+  p256_1 =
+      _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 2 * p)));
+  p256_0 =
+      _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 1 * p)));
+  q256_0 =
+      _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 0 * p)));
+  q256_1 =
+      _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s + 1 * p)));
+  q256_2 =
+      _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s + 2 * p)));
+
+  p2 = _mm256_castsi256_si128(p256_2);
+  p1 = _mm256_castsi256_si128(p256_1);
+  p0 = _mm256_castsi256_si128(p256_0);
+  q0 = _mm256_castsi256_si128(q256_0);
+  q1 = _mm256_castsi256_si128(q256_1);
+  q2 = _mm256_castsi256_si128(q256_2);
+
+  {
+    __m128i work;
+    const __m128i fe = _mm_set1_epi8((int8_t)0xfe);
+    const __m128i abs_p1p0 =
+        _mm_or_si128(_mm_subs_epu8(p1, p0), _mm_subs_epu8(p0, p1));
+    const __m128i abs_q1q0 =
+        _mm_or_si128(_mm_subs_epu8(q1, q0), _mm_subs_epu8(q0, q1));
+    __m128i abs_p0q0 =
+        _mm_or_si128(_mm_subs_epu8(p0, q0), _mm_subs_epu8(q0, p0));
+    __m128i abs_p1q1 =
+        _mm_or_si128(_mm_subs_epu8(p1, q1), _mm_subs_epu8(q1, p1));
+
+    flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
+
+    abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0);
+    abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
+    mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit_v);
+    mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
+    // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit) * -1;
+    mask = _mm_max_epu8(flat, mask);
+    // mask |= (abs(p1 - p0) > limit) * -1;
+    // mask |= (abs(q1 - q0) > limit) * -1;
+    work = _mm_max_epu8(
+        _mm_or_si128(_mm_subs_epu8(p2, p1), _mm_subs_epu8(p1, p2)),
+        _mm_or_si128(_mm_subs_epu8(q2, q1), _mm_subs_epu8(q1, q2)));
+    mask = _mm_max_epu8(work, mask);
+    mask = _mm_subs_epu8(mask, limit_v);
+    mask = _mm_cmpeq_epi8(mask, zero);
+  }
+
+  if (0xffff == _mm_movemask_epi8(_mm_cmpeq_epi8(mask, zero))) return;
+
+  // loop filter
+  {
+    const __m128i t4 = _mm_set1_epi8(4);
+    const __m128i t3 = _mm_set1_epi8(3);
+    const __m128i t80 = _mm_set1_epi8((int8_t)0x80);
+    const __m128i te0 = _mm_set1_epi8((int8_t)0xe0);
+    const __m128i t1f = _mm_set1_epi8(0x1f);
+    const __m128i t1 = _mm_set1_epi8(0x1);
+    const __m128i t7f = _mm_set1_epi8(0x7f);
+    const __m128i one = _mm_set1_epi8(1);
+    __m128i hev;
+
+    hev = _mm_subs_epu8(flat, thresh_v);
+    hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+
+    __m128i ps1 = _mm_xor_si128(p1, t80);
+    __m128i ps0 = _mm_xor_si128(p0, t80);
+    __m128i qs0 = _mm_xor_si128(q0, t80);
+    __m128i qs1 = _mm_xor_si128(q1, t80);
+    __m128i filt;
+    __m128i work_a;
+    __m128i filter1, filter2;
+    __m128i flat_p1, flat_p0, flat_q0, flat_q1;
+
+    filt = _mm_and_si128(_mm_subs_epi8(ps1, qs1), hev);
+    work_a = _mm_subs_epi8(qs0, ps0);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_and_si128(filt, mask);
+
+    filter1 = _mm_adds_epi8(filt, t4);
+    filter2 = _mm_adds_epi8(filt, t3);
+
+    work_a = _mm_cmpgt_epi8(zero, filter1);
+    filter1 = _mm_srli_epi16(filter1, 3);
+    work_a = _mm_and_si128(work_a, te0);
+    filter1 = _mm_and_si128(filter1, t1f);
+    filter1 = _mm_or_si128(filter1, work_a);
+    qs0 = _mm_xor_si128(_mm_subs_epi8(qs0, filter1), t80);
+
+    work_a = _mm_cmpgt_epi8(zero, filter2);
+    filter2 = _mm_srli_epi16(filter2, 3);
+    work_a = _mm_and_si128(work_a, te0);
+    filter2 = _mm_and_si128(filter2, t1f);
+    filter2 = _mm_or_si128(filter2, work_a);
+    ps0 = _mm_xor_si128(_mm_adds_epi8(ps0, filter2), t80);
+
+    filt = _mm_adds_epi8(filter1, t1);
+    work_a = _mm_cmpgt_epi8(zero, filt);
+    filt = _mm_srli_epi16(filt, 1);
+    work_a = _mm_and_si128(work_a, t80);
+    filt = _mm_and_si128(filt, t7f);
+    filt = _mm_or_si128(filt, work_a);
+    filt = _mm_andnot_si128(hev, filt);
+    ps1 = _mm_xor_si128(_mm_adds_epi8(ps1, filt), t80);
+    qs1 = _mm_xor_si128(_mm_subs_epi8(qs1, filt), t80);
+
+    __m128i work;
+    work = _mm_max_epu8(
+        _mm_or_si128(_mm_subs_epu8(p2, p0), _mm_subs_epu8(p0, p2)),
+        _mm_or_si128(_mm_subs_epu8(q2, q0), _mm_subs_epu8(q0, q2)));
+    flat = _mm_max_epu8(work, flat);
+    flat = _mm_subs_epu8(flat, one);
+    flat = _mm_cmpeq_epi8(flat, zero);
+    flat = _mm_and_si128(flat, mask);
+
+    if (0xffff != _mm_movemask_epi8(_mm_cmpeq_epi8(flat, zero))) {
+      const __m256i four = _mm256_set1_epi16(4);
+      __m256i pixetFilter, add, res;
+
+      const __m256i filter =
+          _mm256_load_si256((__m256i const *)filt_loopfilter_avx2);
+
+      p256_2 = _mm256_shuffle_epi8(p256_2, filter);
+      p256_1 = _mm256_shuffle_epi8(p256_1, filter);
+      p256_0 = _mm256_shuffle_epi8(p256_0, filter);
+      q256_0 = _mm256_shuffle_epi8(q256_0, filter);
+      q256_1 = _mm256_shuffle_epi8(q256_1, filter);
+      q256_2 = _mm256_shuffle_epi8(q256_2, filter);
+
+      pixetFilter = _mm256_slli_epi16(
+          _mm256_add_epi16(p256_2, _mm256_add_epi16(p256_1, p256_0)), 1);
+      pixetFilter =
+          _mm256_add_epi16(pixetFilter, _mm256_add_epi16(p256_2, q256_0));
+      pixetFilter = _mm256_add_epi16(four, pixetFilter);
+      res = _mm256_srli_epi16(pixetFilter, 3);
+      flat_p1 = _mm256_castsi256_si128(
+          _mm256_permute4x64_epi64(_mm256_packus_epi16(res, res), 168));
+      p1 = _mm_andnot_si128(flat, ps1);
+      flat_p1 = _mm_and_si128(flat, flat_p1);
+      p1 = _mm_or_si128(flat_p1, p1);
+
+      add = _mm256_add_epi16(_mm256_sub_epi16(q256_1, p256_2),
+                             _mm256_sub_epi16(q256_0, p256_2));
+      pixetFilter = _mm256_add_epi16(pixetFilter, add);
+      res = _mm256_srli_epi16(pixetFilter, 3);
+      flat_p0 = _mm256_castsi256_si128(
+          _mm256_permute4x64_epi64(_mm256_packus_epi16(res, res), 168));
+      p0 = _mm_andnot_si128(flat, ps0);
+      flat_p0 = _mm_and_si128(flat, flat_p0);
+      p0 = _mm_or_si128(flat_p0, p0);
+
+      add = _mm256_add_epi16(_mm256_sub_epi16(q256_2, p256_2),
+                             _mm256_sub_epi16(q256_1, p256_1));
+      pixetFilter = _mm256_add_epi16(pixetFilter, add);
+      res = _mm256_srli_epi16(pixetFilter, 3);
+      flat_q0 = _mm256_castsi256_si128(
+          _mm256_permute4x64_epi64(_mm256_packus_epi16(res, res), 168));
+      q0 = _mm_andnot_si128(flat, qs0);
+      flat_q0 = _mm_and_si128(flat, flat_q0);
+      q0 = _mm_or_si128(flat_q0, q0);
+
+      add = _mm256_add_epi16(_mm256_sub_epi16(q256_2, p256_1),
+                             _mm256_sub_epi16(q256_2, p256_0));
+      pixetFilter = _mm256_add_epi16(pixetFilter, add);
+      res = _mm256_srli_epi16(pixetFilter, 3);
+      flat_q1 = _mm256_castsi256_si128(
+          _mm256_permute4x64_epi64(_mm256_packus_epi16(res, res), 168));
+      q1 = _mm_andnot_si128(flat, qs1);
+      flat_q1 = _mm_and_si128(flat, flat_q1);
+      q1 = _mm_or_si128(flat_q1, q1);
+
+      _mm_storeu_si128((__m128i *)(s - 3 * p), p2);
+      _mm_storeu_si128((__m128i *)(s - 2 * p), p1);
+      _mm_storeu_si128((__m128i *)(s - 1 * p), p0);
+      _mm_storeu_si128((__m128i *)(s - 0 * p), q0);
+      _mm_storeu_si128((__m128i *)(s + 1 * p), q1);
+      _mm_storeu_si128((__m128i *)(s + 2 * p), q2);
+    } else {
+      _mm_storeu_si128((__m128i *)(s - 2 * p), ps1);
+      _mm_storeu_si128((__m128i *)(s - 1 * p), ps0);
+      _mm_storeu_si128((__m128i *)(s - 0 * p), qs0);
+      _mm_storeu_si128((__m128i *)(s + 1 * p), qs1);
+    }
+  }
+}
+
+void aom_lpf_horizontal_8_quad_avx2(unsigned char *s, int p,
+                                    const unsigned char *_blimit0,
+                                    const unsigned char *_limit0,
+                                    const unsigned char *_thresh0) {
+  __m256i p256_3, q256_3, p256_2, q256_2, p256_1, q256_1, p256_0, q256_0;
+  __m128i p3, p2, p1, p0, q0, q1, q2, q3;
+  __m128i mask, flat;
+
+  const __m128i thresh_v =
+      _mm_broadcastb_epi8(_mm_cvtsi32_si128((int)_thresh0[0]));
+  const __m128i limit_v =
+      _mm_broadcastb_epi8(_mm_cvtsi32_si128((int)_limit0[0]));
+  const __m128i blimit_v =
+      _mm_broadcastb_epi8(_mm_cvtsi32_si128((int)_blimit0[0]));
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i ff = _mm_cmpeq_epi8(zero, zero);
+
+  p256_3 =
+      _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 4 * p)));
+  p256_2 =
+      _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 3 * p)));
+  p256_1 =
+      _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 2 * p)));
+  p256_0 =
+      _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 1 * p)));
+  q256_0 =
+      _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 0 * p)));
+  q256_1 =
+      _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s + 1 * p)));
+  q256_2 =
+      _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s + 2 * p)));
+  q256_3 =
+      _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s + 3 * p)));
+
+  p3 = _mm256_castsi256_si128(p256_3);
+  p2 = _mm256_castsi256_si128(p256_2);
+  p1 = _mm256_castsi256_si128(p256_1);
+  p0 = _mm256_castsi256_si128(p256_0);
+  q0 = _mm256_castsi256_si128(q256_0);
+  q1 = _mm256_castsi256_si128(q256_1);
+  q2 = _mm256_castsi256_si128(q256_2);
+  q3 = _mm256_castsi256_si128(q256_3);
+
+  {
+    __m128i work;
+    const __m128i fe = _mm_set1_epi8((int8_t)0xfe);
+    const __m128i abs_p1p0 =
+        _mm_or_si128(_mm_subs_epu8(p1, p0), _mm_subs_epu8(p0, p1));
+    const __m128i abs_q1q0 =
+        _mm_or_si128(_mm_subs_epu8(q1, q0), _mm_subs_epu8(q0, q1));
+    __m128i abs_p0q0 =
+        _mm_or_si128(_mm_subs_epu8(p0, q0), _mm_subs_epu8(q0, p0));
+    __m128i abs_p1q1 =
+        _mm_or_si128(_mm_subs_epu8(p1, q1), _mm_subs_epu8(q1, p1));
+
+    flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
+
+    abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0);
+    abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
+    mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit_v);
+    mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
+    // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit) * -1;
+    mask = _mm_max_epu8(flat, mask);
+    // mask |= (abs(p1 - p0) > limit) * -1;
+    // mask |= (abs(q1 - q0) > limit) * -1;
+    work = _mm_max_epu8(
+        _mm_or_si128(_mm_subs_epu8(p2, p1), _mm_subs_epu8(p1, p2)),
+        _mm_or_si128(_mm_subs_epu8(p3, p2), _mm_subs_epu8(p2, p3)));
+    mask = _mm_max_epu8(work, mask);
+    work = _mm_max_epu8(
+        _mm_or_si128(_mm_subs_epu8(q2, q1), _mm_subs_epu8(q1, q2)),
+        _mm_or_si128(_mm_subs_epu8(q3, q2), _mm_subs_epu8(q2, q3)));
+    mask = _mm_max_epu8(work, mask);
+    mask = _mm_subs_epu8(mask, limit_v);
+    mask = _mm_cmpeq_epi8(mask, zero);
+  }
+
+  if (0xffff == _mm_movemask_epi8(_mm_cmpeq_epi8(mask, zero))) return;
+
+  // loop filter
+  {
+    const __m128i t4 = _mm_set1_epi8(4);
+    const __m128i t3 = _mm_set1_epi8(3);
+    const __m128i t80 = _mm_set1_epi8((int8_t)0x80);
+    const __m128i te0 = _mm_set1_epi8((int8_t)0xe0);
+    const __m128i t1f = _mm_set1_epi8(0x1f);
+    const __m128i t1 = _mm_set1_epi8(0x1);
+    const __m128i t7f = _mm_set1_epi8(0x7f);
+    const __m128i one = _mm_set1_epi8(1);
+    __m128i hev;
+
+    hev = _mm_subs_epu8(flat, thresh_v);
+    hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+
+    __m128i ps1 = _mm_xor_si128(p1, t80);
+    __m128i ps0 = _mm_xor_si128(p0, t80);
+    __m128i qs0 = _mm_xor_si128(q0, t80);
+    __m128i qs1 = _mm_xor_si128(q1, t80);
+    __m128i filt;
+    __m128i work_a;
+    __m128i filter1, filter2;
+    __m128i flat_p2, flat_p1, flat_p0, flat_q0, flat_q1, flat_q2;
+
+    filt = _mm_and_si128(_mm_subs_epi8(ps1, qs1), hev);
+    work_a = _mm_subs_epi8(qs0, ps0);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_and_si128(filt, mask);
+
+    filter1 = _mm_adds_epi8(filt, t4);
+    filter2 = _mm_adds_epi8(filt, t3);
+
+    work_a = _mm_cmpgt_epi8(zero, filter1);
+    filter1 = _mm_srli_epi16(filter1, 3);
+    work_a = _mm_and_si128(work_a, te0);
+    filter1 = _mm_and_si128(filter1, t1f);
+    filter1 = _mm_or_si128(filter1, work_a);
+    qs0 = _mm_xor_si128(_mm_subs_epi8(qs0, filter1), t80);
+
+    work_a = _mm_cmpgt_epi8(zero, filter2);
+    filter2 = _mm_srli_epi16(filter2, 3);
+    work_a = _mm_and_si128(work_a, te0);
+    filter2 = _mm_and_si128(filter2, t1f);
+    filter2 = _mm_or_si128(filter2, work_a);
+    ps0 = _mm_xor_si128(_mm_adds_epi8(ps0, filter2), t80);
+
+    filt = _mm_adds_epi8(filter1, t1);
+    work_a = _mm_cmpgt_epi8(zero, filt);
+    filt = _mm_srli_epi16(filt, 1);
+    work_a = _mm_and_si128(work_a, t80);
+    filt = _mm_and_si128(filt, t7f);
+    filt = _mm_or_si128(filt, work_a);
+    filt = _mm_andnot_si128(hev, filt);
+    ps1 = _mm_xor_si128(_mm_adds_epi8(ps1, filt), t80);
+    qs1 = _mm_xor_si128(_mm_subs_epi8(qs1, filt), t80);
+
+    __m128i work;
+    work = _mm_max_epu8(
+        _mm_or_si128(_mm_subs_epu8(p2, p0), _mm_subs_epu8(p0, p2)),
+        _mm_or_si128(_mm_subs_epu8(q2, q0), _mm_subs_epu8(q0, q2)));
+    flat = _mm_max_epu8(work, flat);
+    work = _mm_max_epu8(
+        _mm_or_si128(_mm_subs_epu8(p3, p0), _mm_subs_epu8(p0, p3)),
+        _mm_or_si128(_mm_subs_epu8(q3, q0), _mm_subs_epu8(q0, q3)));
+    flat = _mm_max_epu8(work, flat);
+    flat = _mm_subs_epu8(flat, one);
+    flat = _mm_cmpeq_epi8(flat, zero);
+    flat = _mm_and_si128(flat, mask);
+
+    if (0xffff != _mm_movemask_epi8(_mm_cmpeq_epi8(flat, zero))) {
+      const __m256i four = _mm256_set1_epi16(4);
+      __m256i pixetFilter_p2p1p0, p2p1p0, q2q1q0, pixetFilter_q2q1q0, sum_p,
+          sum_q, res_p, res_q;
+
+      const __m256i filter =
+          _mm256_load_si256((__m256i const *)filt_loopfilter_avx2);
+
+      p256_3 = _mm256_shuffle_epi8(p256_3, filter);
+      p256_2 = _mm256_shuffle_epi8(p256_2, filter);
+      p256_1 = _mm256_shuffle_epi8(p256_1, filter);
+      p256_0 = _mm256_shuffle_epi8(p256_0, filter);
+      q256_0 = _mm256_shuffle_epi8(q256_0, filter);
+      q256_1 = _mm256_shuffle_epi8(q256_1, filter);
+      q256_2 = _mm256_shuffle_epi8(q256_2, filter);
+      q256_3 = _mm256_shuffle_epi8(q256_3, filter);
+
+      p2p1p0 = _mm256_add_epi16(p256_0, _mm256_add_epi16(p256_2, p256_1));
+      q2q1q0 = _mm256_add_epi16(q256_0, _mm256_add_epi16(q256_2, q256_1));
+
+      pixetFilter_p2p1p0 =
+          _mm256_add_epi16(four, _mm256_add_epi16(p2p1p0, q2q1q0));
+      pixetFilter_q2q1q0 = pixetFilter_p2p1p0;
+
+      pixetFilter_p2p1p0 = _mm256_add_epi16(pixetFilter_p2p1p0, p256_3);
+      res_p =
+          _mm256_srli_epi16(_mm256_add_epi16(pixetFilter_p2p1p0, p256_0), 3);
+      flat_p0 = _mm256_castsi256_si128(
+          _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
+      p0 = _mm_andnot_si128(flat, ps0);
+      flat_p0 = _mm_and_si128(flat, flat_p0);
+      p0 = _mm_or_si128(flat_p0, p0);
+
+      pixetFilter_q2q1q0 = _mm256_add_epi16(pixetFilter_q2q1q0, q256_3);
+      res_q =
+          _mm256_srli_epi16(_mm256_add_epi16(pixetFilter_q2q1q0, q256_0), 3);
+      flat_q0 = _mm256_castsi256_si128(
+          _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
+      q0 = _mm_andnot_si128(flat, qs0);
+      flat_q0 = _mm_and_si128(flat, flat_q0);
+      q0 = _mm_or_si128(flat_q0, q0);
+
+      sum_p = _mm256_sub_epi16(p256_3, q256_2);
+      pixetFilter_p2p1p0 = _mm256_add_epi16(pixetFilter_p2p1p0, sum_p);
+      res_p =
+          _mm256_srli_epi16(_mm256_add_epi16(pixetFilter_p2p1p0, p256_1), 3);
+      flat_p1 = _mm256_castsi256_si128(
+          _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
+      p1 = _mm_andnot_si128(flat, ps1);
+      flat_p1 = _mm_and_si128(flat, flat_p1);
+      p1 = _mm_or_si128(flat_p1, p1);
+
+      sum_q = _mm256_sub_epi16(q256_3, p256_2);
+      pixetFilter_q2q1q0 = _mm256_add_epi16(pixetFilter_q2q1q0, sum_q);
+      res_q =
+          _mm256_srli_epi16(_mm256_add_epi16(pixetFilter_q2q1q0, q256_1), 3);
+      flat_q1 = _mm256_castsi256_si128(
+          _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
+      q1 = _mm_andnot_si128(flat, qs1);
+      flat_q1 = _mm_and_si128(flat, flat_q1);
+      q1 = _mm_or_si128(flat_q1, q1);
+
+      sum_p = _mm256_sub_epi16(p256_3, q256_1);
+      pixetFilter_p2p1p0 = _mm256_add_epi16(pixetFilter_p2p1p0, sum_p);
+      res_p =
+          _mm256_srli_epi16(_mm256_add_epi16(pixetFilter_p2p1p0, p256_2), 3);
+      flat_p2 = _mm256_castsi256_si128(
+          _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_p), 168));
+      p2 = _mm_andnot_si128(flat, p2);
+      flat_p2 = _mm_and_si128(flat, flat_p2);
+      p2 = _mm_or_si128(flat_p2, p2);
+
+      sum_q = _mm256_sub_epi16(q256_3, p256_1);
+      pixetFilter_q2q1q0 = _mm256_add_epi16(pixetFilter_q2q1q0, sum_q);
+      res_q =
+          _mm256_srli_epi16(_mm256_add_epi16(pixetFilter_q2q1q0, q256_2), 3);
+      flat_q2 = _mm256_castsi256_si128(
+          _mm256_permute4x64_epi64(_mm256_packus_epi16(res_q, res_q), 168));
+      q2 = _mm_andnot_si128(flat, q2);
+      flat_q2 = _mm_and_si128(flat, flat_q2);
+      q2 = _mm_or_si128(flat_q2, q2);
+
+      _mm_storeu_si128((__m128i *)(s - 3 * p), p2);
+      _mm_storeu_si128((__m128i *)(s - 2 * p), p1);
+      _mm_storeu_si128((__m128i *)(s - 1 * p), p0);
+      _mm_storeu_si128((__m128i *)(s - 0 * p), q0);
+      _mm_storeu_si128((__m128i *)(s + 1 * p), q1);
+      _mm_storeu_si128((__m128i *)(s + 2 * p), q2);
+    } else {
+      _mm_storeu_si128((__m128i *)(s - 2 * p), ps1);
+      _mm_storeu_si128((__m128i *)(s - 1 * p), ps0);
+      _mm_storeu_si128((__m128i *)(s - 0 * p), qs0);
+      _mm_storeu_si128((__m128i *)(s + 1 * p), qs1);
+    }
+  }
+}
+
+static INLINE void trans_store_16x16_lpf_vert14(unsigned char *in0, int in_p,
+                                                unsigned char *out, int out_p,
+                                                int is_store_avx2) {
+  const __m128i x0 = _mm_loadu_si128((__m128i *)in0);
+  const __m128i x1 = _mm_loadu_si128((__m128i *)(in0 + in_p * 1));
+  const __m128i x2 = _mm_loadu_si128((__m128i *)(in0 + in_p * 2));
+  const __m128i x3 = _mm_loadu_si128((__m128i *)(in0 + in_p * 3));
+  const __m128i x4 = _mm_loadu_si128((__m128i *)(in0 + in_p * 4));
+  const __m128i x5 = _mm_loadu_si128((__m128i *)(in0 + in_p * 5));
+  const __m128i x6 = _mm_loadu_si128((__m128i *)(in0 + in_p * 6));
+  const __m128i x7 = _mm_loadu_si128((__m128i *)(in0 + in_p * 7));
+
+  const __m256i y0 = _mm256_insertf128_si256(
+      _mm256_castsi128_si256(x0), _mm_loadu_si128((__m128i *)(in0 + in_p * 8)),
+      0x1);
+  const __m256i y1 = _mm256_insertf128_si256(
+      _mm256_castsi128_si256(x1), _mm_loadu_si128((__m128i *)(in0 + in_p * 9)),
+      0x1);
+  const __m256i y2 = _mm256_insertf128_si256(
+      _mm256_castsi128_si256(x2), _mm_loadu_si128((__m128i *)(in0 + in_p * 10)),
+      0x1);
+  const __m256i y3 = _mm256_insertf128_si256(
+      _mm256_castsi128_si256(x3), _mm_loadu_si128((__m128i *)(in0 + in_p * 11)),
+      0x1);
+  const __m256i y4 = _mm256_insertf128_si256(
+      _mm256_castsi128_si256(x4), _mm_loadu_si128((__m128i *)(in0 + in_p * 12)),
+      0x1);
+  const __m256i y5 = _mm256_insertf128_si256(
+      _mm256_castsi128_si256(x5), _mm_loadu_si128((__m128i *)(in0 + in_p * 13)),
+      0x1);
+  const __m256i y6 = _mm256_insertf128_si256(
+      _mm256_castsi128_si256(x6), _mm_loadu_si128((__m128i *)(in0 + in_p * 14)),
+      0x1);
+  const __m256i y7 = _mm256_insertf128_si256(
+      _mm256_castsi128_si256(x7), _mm_loadu_si128((__m128i *)(in0 + in_p * 15)),
+      0x1);
+
+  const __m256i y_s00 = _mm256_unpacklo_epi8(y0, y1);
+  const __m256i y_s01 = _mm256_unpackhi_epi8(y0, y1);
+  const __m256i y_s02 = _mm256_unpacklo_epi8(y2, y3);
+  const __m256i y_s03 = _mm256_unpackhi_epi8(y2, y3);
+  const __m256i y_s04 = _mm256_unpacklo_epi8(y4, y5);
+  const __m256i y_s05 = _mm256_unpackhi_epi8(y4, y5);
+  const __m256i y_s06 = _mm256_unpacklo_epi8(y6, y7);
+  const __m256i y_s07 = _mm256_unpackhi_epi8(y6, y7);
+
+  const __m256i y_s10 = _mm256_unpacklo_epi16(y_s00, y_s02);
+  const __m256i y_s11 = _mm256_unpackhi_epi16(y_s00, y_s02);
+  const __m256i y_s12 = _mm256_unpacklo_epi16(y_s01, y_s03);
+  const __m256i y_s13 = _mm256_unpackhi_epi16(y_s01, y_s03);
+  const __m256i y_s14 = _mm256_unpacklo_epi16(y_s04, y_s06);
+  const __m256i y_s15 = _mm256_unpackhi_epi16(y_s04, y_s06);
+  const __m256i y_s16 = _mm256_unpacklo_epi16(y_s05, y_s07);
+  const __m256i y_s17 = _mm256_unpackhi_epi16(y_s05, y_s07);
+
+  const __m256i y_s20 = _mm256_unpacklo_epi32(y_s10, y_s14);
+  const __m256i y_s21 = _mm256_unpackhi_epi32(y_s10, y_s14);
+  const __m256i y_s22 = _mm256_unpacklo_epi32(y_s11, y_s15);
+  const __m256i y_s23 = _mm256_unpackhi_epi32(y_s11, y_s15);
+  const __m256i y_s24 = _mm256_unpacklo_epi32(y_s12, y_s16);
+  const __m256i y_s25 = _mm256_unpackhi_epi32(y_s12, y_s16);
+  const __m256i y_s26 = _mm256_unpacklo_epi32(y_s13, y_s17);
+  const __m256i y_s27 = _mm256_unpackhi_epi32(y_s13, y_s17);
+
+  const __m256i row_s01 = _mm256_permute4x64_epi64(y_s20, 0xd8);
+  const __m256i row_s23 = _mm256_permute4x64_epi64(y_s21, 0xd8);
+  const __m256i row_s45 = _mm256_permute4x64_epi64(y_s22, 0xd8);
+  const __m256i row_s67 = _mm256_permute4x64_epi64(y_s23, 0xd8);
+  const __m256i row_s89 = _mm256_permute4x64_epi64(y_s24, 0xd8);
+  const __m256i row_s1011 = _mm256_permute4x64_epi64(y_s25, 0xd8);
+  const __m256i row_s1213 = _mm256_permute4x64_epi64(y_s26, 0xd8);
+  const __m256i row_s1415 = _mm256_permute4x64_epi64(y_s27, 0xd8);
+
+  if (is_store_avx2) {
+    _mm256_storeu_si256((__m256i *)(out), row_s01);
+    _mm256_storeu_si256((__m256i *)(out + (2 * out_p)), row_s23);
+    _mm256_storeu_si256((__m256i *)(out + (4 * out_p)), row_s45);
+    _mm256_storeu_si256((__m256i *)(out + (6 * out_p)), row_s67);
+    _mm256_storeu_si256((__m256i *)(out + (8 * out_p)), row_s89);
+    _mm256_storeu_si256((__m256i *)(out + (10 * out_p)), row_s1011);
+    _mm256_storeu_si256((__m256i *)(out + (12 * out_p)), row_s1213);
+    _mm256_storeu_si256((__m256i *)(out + (14 * out_p)), row_s1415);
+  } else {
+    _mm_storeu_si128((__m128i *)(out), _mm256_castsi256_si128(row_s01));
+    _mm_storeu_si128((__m128i *)(out + (2 * out_p)),
+                     _mm256_castsi256_si128(row_s23));
+    _mm_storeu_si128((__m128i *)(out + (4 * out_p)),
+                     _mm256_castsi256_si128(row_s45));
+    _mm_storeu_si128((__m128i *)(out + (6 * out_p)),
+                     _mm256_castsi256_si128(row_s67));
+    _mm_storeu_si128((__m128i *)(out + (8 * out_p)),
+                     _mm256_castsi256_si128(row_s89));
+    _mm_storeu_si128((__m128i *)(out + (10 * out_p)),
+                     _mm256_castsi256_si128(row_s1011));
+    _mm_storeu_si128((__m128i *)(out + (12 * out_p)),
+                     _mm256_castsi256_si128(row_s1213));
+    _mm_storeu_si128((__m128i *)(out + (14 * out_p)),
+                     _mm256_castsi256_si128(row_s1415));
+    _mm_storeu_si128((__m128i *)(out + (1 * out_p)),
+                     _mm256_extracti128_si256(row_s01, 1));
+    _mm_storeu_si128((__m128i *)(out + (3 * out_p)),
+                     _mm256_extracti128_si256(row_s23, 1));
+    _mm_storeu_si128((__m128i *)(out + (5 * out_p)),
+                     _mm256_extracti128_si256(row_s45, 1));
+    _mm_storeu_si128((__m128i *)(out + (7 * out_p)),
+                     _mm256_extracti128_si256(row_s67, 1));
+    _mm_storeu_si128((__m128i *)(out + (9 * out_p)),
+                     _mm256_extracti128_si256(row_s89, 1));
+    _mm_storeu_si128((__m128i *)(out + (11 * out_p)),
+                     _mm256_extracti128_si256(row_s1011, 1));
+    _mm_storeu_si128((__m128i *)(out + (13 * out_p)),
+                     _mm256_extracti128_si256(row_s1213, 1));
+    _mm_storeu_si128((__m128i *)(out + (15 * out_p)),
+                     _mm256_extracti128_si256(row_s1415, 1));
+  }
+}
+
+void aom_lpf_horizontal_14_quad_avx2(unsigned char *s, int p,
+                                     const unsigned char *_blimit0,
+                                     const unsigned char *_limit0,
+                                     const unsigned char *_thresh0) {
+  __m128i mask, flat;
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i ff = _mm_cmpeq_epi8(zero, zero);
+
+  __m256i p256_3 =
+      _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 4 * p)));
+  __m256i p256_2 =
+      _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 3 * p)));
+  __m256i p256_1 =
+      _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 2 * p)));
+  __m256i p256_0 =
+      _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 1 * p)));
+  __m256i q256_0 =
+      _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s - 0 * p)));
+  __m256i q256_1 =
+      _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s + 1 * p)));
+  __m256i q256_2 =
+      _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s + 2 * p)));
+  __m256i q256_3 =
+      _mm256_castpd_si256(_mm256_broadcast_pd((__m128d const *)(s + 3 * p)));
+
+  __m128i p3 = _mm256_castsi256_si128(p256_3);
+  __m128i p2 = _mm256_castsi256_si128(p256_2);
+  __m128i p1 = _mm256_castsi256_si128(p256_1);
+  __m128i p0 = _mm256_castsi256_si128(p256_0);
+  __m128i q0 = _mm256_castsi256_si128(q256_0);
+  __m128i q1 = _mm256_castsi256_si128(q256_1);
+  __m128i q2 = _mm256_castsi256_si128(q256_2);
+  __m128i q3 = _mm256_castsi256_si128(q256_3);
+
+  {
+    const __m128i limit_v =
+        _mm_broadcastb_epi8(_mm_cvtsi32_si128((int)_limit0[0]));
+    const __m128i blimit_v =
+        _mm_broadcastb_epi8(_mm_cvtsi32_si128((int)_blimit0[0]));
+    const __m128i fe = _mm_set1_epi8((int8_t)0xfe);
+    const __m128i abs_p1p0 =
+        _mm_or_si128(_mm_subs_epu8(p1, p0), _mm_subs_epu8(p0, p1));
+    const __m128i abs_q1q0 =
+        _mm_or_si128(_mm_subs_epu8(q1, q0), _mm_subs_epu8(q0, q1));
+    __m128i abs_p0q0 =
+        _mm_or_si128(_mm_subs_epu8(p0, q0), _mm_subs_epu8(q0, p0));
+    __m128i abs_p1q1 =
+        _mm_or_si128(_mm_subs_epu8(p1, q1), _mm_subs_epu8(q1, p1));
+
+    flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
+
+    abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0);
+    abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
+    mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit_v);
+    mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
+    // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit) * -1;
+    mask = _mm_max_epu8(flat, mask);
+    // mask |= (abs(p1 - p0) > limit) * -1;
+    // mask |= (abs(q1 - q0) > limit) * -1;
+    __m128i work = _mm_max_epu8(
+        _mm_or_si128(_mm_subs_epu8(p2, p1), _mm_subs_epu8(p1, p2)),
+        _mm_or_si128(_mm_subs_epu8(p3, p2), _mm_subs_epu8(p2, p3)));
+    mask = _mm_max_epu8(work, mask);
+    work = _mm_max_epu8(
+        _mm_or_si128(_mm_subs_epu8(q2, q1), _mm_subs_epu8(q1, q2)),
+        _mm_or_si128(_mm_subs_epu8(q3, q2), _mm_subs_epu8(q2, q3)));
+    mask = _mm_max_epu8(work, mask);
+    mask = _mm_subs_epu8(mask, limit_v);
+    mask = _mm_cmpeq_epi8(mask, zero);
+  }
+
+  if (0xffff == _mm_movemask_epi8(_mm_cmpeq_epi8(mask, zero))) return;
+
+  // loop filter
+  {
+    const __m128i thresh_v =
+        _mm_broadcastb_epi8(_mm_cvtsi32_si128((int)_thresh0[0]));
+    const __m128i one = _mm_set1_epi8(1);
+    const __m128i t3 = _mm_set1_epi8(3);
+    const __m128i t4 = _mm_add_epi8(one, t3);
+    const __m128i t80 = _mm_set1_epi8((int8_t)0x80);
+    const __m128i te0 = _mm_set1_epi8((int8_t)0xe0);
+    const __m128i t1f = _mm_set1_epi8(0x1f);
+    const __m128i t7f = _mm_sub_epi8(t80, one);
+
+    __m128i hev = _mm_subs_epu8(flat, thresh_v);
+    hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+
+    __m128i ps1 = _mm_xor_si128(p1, t80);
+    __m128i ps0 = _mm_xor_si128(p0, t80);
+    __m128i qs0 = _mm_xor_si128(q0, t80);
+    __m128i qs1 = _mm_xor_si128(q1, t80);
+
+    __m128i filt = _mm_and_si128(_mm_subs_epi8(ps1, qs1), hev);
+    __m128i work_a = _mm_subs_epi8(qs0, ps0);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_and_si128(filt, mask);
+
+    __m128i filter1 = _mm_adds_epi8(filt, t4);
+    __m128i filter2 = _mm_adds_epi8(filt, t3);
+
+    work_a = _mm_cmpgt_epi8(zero, filter1);
+    filter1 = _mm_srli_epi16(filter1, 3);
+    work_a = _mm_and_si128(work_a, te0);
+    filter1 = _mm_and_si128(filter1, t1f);
+    filter1 = _mm_or_si128(filter1, work_a);
+    qs0 = _mm_xor_si128(_mm_subs_epi8(qs0, filter1), t80);
+
+    work_a = _mm_cmpgt_epi8(zero, filter2);
+    filter2 = _mm_srli_epi16(filter2, 3);
+    work_a = _mm_and_si128(work_a, te0);
+    filter2 = _mm_and_si128(filter2, t1f);
+    filter2 = _mm_or_si128(filter2, work_a);
+    ps0 = _mm_xor_si128(_mm_adds_epi8(ps0, filter2), t80);
+
+    filt = _mm_adds_epi8(filter1, one);
+    work_a = _mm_cmpgt_epi8(zero, filt);
+    filt = _mm_srli_epi16(filt, 1);
+    work_a = _mm_and_si128(work_a, t80);
+    filt = _mm_and_si128(filt, t7f);
+    filt = _mm_or_si128(filt, work_a);
+    filt = _mm_andnot_si128(hev, filt);
+    ps1 = _mm_xor_si128(_mm_adds_epi8(ps1, filt), t80);
+    qs1 = _mm_xor_si128(_mm_subs_epi8(qs1, filt), t80);
+
+    // Derive flat
+    __m256i p0q0256 = _mm256_blend_epi32(p256_0, q256_0, 0xf0);
+    __m256i p2q2256 = _mm256_blend_epi32(p256_2, q256_2, 0xf0);
+    __m256i p3q3256 = _mm256_blend_epi32(p256_3, q256_3, 0xf0);
+    const __m256i ps0qs0256 =
+        _mm256_insertf128_si256(_mm256_castsi128_si256(ps0), qs0, 0x1);
+    const __m256i ps1qs1256 =
+        _mm256_insertf128_si256(_mm256_castsi128_si256(ps1), qs1, 0x1);
+    const __m256i work01 = _mm256_or_si256(_mm256_subs_epu8(p2q2256, p0q0256),
+                                           _mm256_subs_epu8(p0q0256, p2q2256));
+    const __m256i work02 = _mm256_or_si256(_mm256_subs_epu8(p3q3256, p0q0256),
+                                           _mm256_subs_epu8(p0q0256, p3q3256));
+    const __m256i max0_256 = _mm256_max_epu8(work01, work02);
+    const __m128i max1_256 =
+        _mm_max_epu8(_mm256_castsi256_si128(max0_256),
+                     _mm256_extractf128_si256(max0_256, 1));
+    flat = _mm_max_epu8(max1_256, flat);
+    flat = _mm_subs_epu8(flat, one);
+    flat = _mm_cmpeq_epi8(flat, zero);
+    flat = _mm_and_si128(flat, mask);
+
+    // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+    // flat and wide flat calculations
+    if (0xffff != _mm_movemask_epi8(_mm_cmpeq_epi8(flat, zero))) {
+      const __m256i flat256 =
+          _mm256_insertf128_si256(_mm256_castsi128_si256(flat), flat, 0x1);
+      const __m256i eight = _mm256_set1_epi16(8);
+      const __m256i four = _mm256_set1_epi16(4);
+
+      __m256i p256_4 = _mm256_castpd_si256(
+          _mm256_broadcast_pd((__m128d const *)(s - 5 * p)));
+      __m256i q256_4 = _mm256_castpd_si256(
+          _mm256_broadcast_pd((__m128d const *)(s + 4 * p)));
+      __m256i p256_5 = _mm256_castpd_si256(
+          _mm256_broadcast_pd((__m128d const *)(s - 6 * p)));
+      __m256i q256_5 = _mm256_castpd_si256(
+          _mm256_broadcast_pd((__m128d const *)(s + 5 * p)));
+      __m256i p256_6 = _mm256_castpd_si256(
+          _mm256_broadcast_pd((__m128d const *)(s - 7 * p)));
+      __m256i q256_6 = _mm256_castpd_si256(
+          _mm256_broadcast_pd((__m128d const *)(s + 6 * p)));
+
+      // Derive flat2
+      __m256i p4q4256 = _mm256_blend_epi32(p256_4, q256_4, 0xf0);
+      __m256i p5q5256 = _mm256_blend_epi32(p256_5, q256_5, 0xf0);
+      const __m256i p6q6256 = _mm256_blend_epi32(p256_6, q256_6, 0xf0);
+      const __m256i work1 = _mm256_or_si256(_mm256_subs_epu8(p4q4256, p0q0256),
+                                            _mm256_subs_epu8(p0q0256, p4q4256));
+      const __m256i work2 = _mm256_or_si256(_mm256_subs_epu8(p5q5256, p0q0256),
+                                            _mm256_subs_epu8(p0q0256, p5q5256));
+      const __m256i work3 = _mm256_or_si256(_mm256_subs_epu8(p6q6256, p0q0256),
+                                            _mm256_subs_epu8(p0q0256, p6q6256));
+      __m256i flat2_256 = _mm256_max_epu8(work1, work2);
+      flat2_256 = _mm256_max_epu8(flat2_256, work3);
+      __m128i flat2 = _mm_max_epu8(_mm256_castsi256_si128(flat2_256),
+                                   _mm256_extractf128_si256(flat2_256, 1));
+      flat2 = _mm_subs_epu8(flat2, one);
+      flat2 = _mm_cmpeq_epi8(flat2, zero);
+      flat2 = _mm_and_si128(flat2, flat);  // flat2 & flat & mask
+
+      const __m256i filter =
+          _mm256_load_si256((__m256i const *)filt_loopfilter_avx2);
+
+      p256_3 = _mm256_shuffle_epi8(p256_3, filter);
+      p256_2 = _mm256_shuffle_epi8(p256_2, filter);
+      p256_1 = _mm256_shuffle_epi8(p256_1, filter);
+      p256_0 = _mm256_shuffle_epi8(p256_0, filter);
+      q256_0 = _mm256_shuffle_epi8(q256_0, filter);
+      q256_1 = _mm256_shuffle_epi8(q256_1, filter);
+      q256_2 = _mm256_shuffle_epi8(q256_2, filter);
+      q256_3 = _mm256_shuffle_epi8(q256_3, filter);
+
+      const __m256i p2p1p0 =
+          _mm256_add_epi16(p256_0, _mm256_add_epi16(p256_2, p256_1));
+      const __m256i q2q1q0 =
+          _mm256_add_epi16(q256_0, _mm256_add_epi16(q256_2, q256_1));
+
+      __m256i pixetFilter_p2p1p0 =
+          _mm256_add_epi16(four, _mm256_add_epi16(p2p1p0, q2q1q0));
+      __m256i pixetFilter_q2q1q0 = pixetFilter_p2p1p0;
+
+      // Derive p0 and q0
+      pixetFilter_p2p1p0 = _mm256_add_epi16(pixetFilter_p2p1p0, p256_3);
+      __m256i res_p =
+          _mm256_srli_epi16(_mm256_add_epi16(pixetFilter_p2p1p0, p256_0), 3);
+      pixetFilter_q2q1q0 = _mm256_add_epi16(pixetFilter_q2q1q0, q256_3);
+      __m256i res_q =
+          _mm256_srli_epi16(_mm256_add_epi16(pixetFilter_q2q1q0, q256_0), 3);
+      __m256i flat_p0q0 =
+          _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_q), 0xd8);
+      p0q0256 = _mm256_andnot_si256(flat256, ps0qs0256);
+      flat_p0q0 = _mm256_and_si256(flat256, flat_p0q0);
+      p0q0256 = _mm256_or_si256(flat_p0q0, p0q0256);
+      p0 = _mm256_castsi256_si128(p0q0256);
+      q0 = _mm256_extractf128_si256(p0q0256, 1);
+
+      // Derive p1 and q1
+      __m256i sum_p = _mm256_sub_epi16(p256_3, q256_2);
+      pixetFilter_p2p1p0 = _mm256_add_epi16(pixetFilter_p2p1p0, sum_p);
+      res_p =
+          _mm256_srli_epi16(_mm256_add_epi16(pixetFilter_p2p1p0, p256_1), 3);
+      __m256i sum_q = _mm256_sub_epi16(q256_3, p256_2);
+      pixetFilter_q2q1q0 = _mm256_add_epi16(pixetFilter_q2q1q0, sum_q);
+      res_q =
+          _mm256_srli_epi16(_mm256_add_epi16(pixetFilter_q2q1q0, q256_1), 3);
+      __m256i flat_p1q1 =
+          _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_q), 0xd8);
+      __m256i p1q1256 = _mm256_andnot_si256(flat256, ps1qs1256);
+      flat_p1q1 = _mm256_and_si256(flat256, flat_p1q1);
+      p1q1256 = _mm256_or_si256(flat_p1q1, p1q1256);
+      p1 = _mm256_castsi256_si128(p1q1256);
+      q1 = _mm256_extractf128_si256(p1q1256, 1);
+
+      // Derive p2 and q2
+      sum_p = _mm256_sub_epi16(p256_3, q256_1);
+      pixetFilter_p2p1p0 = _mm256_add_epi16(pixetFilter_p2p1p0, sum_p);
+      res_p =
+          _mm256_srli_epi16(_mm256_add_epi16(pixetFilter_p2p1p0, p256_2), 3);
+      sum_q = _mm256_sub_epi16(q256_3, p256_1);
+      pixetFilter_q2q1q0 = _mm256_add_epi16(pixetFilter_q2q1q0, sum_q);
+      res_q =
+          _mm256_srli_epi16(_mm256_add_epi16(pixetFilter_q2q1q0, q256_2), 3);
+      __m256i flat_p2q2 =
+          _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_q), 0xd8);
+      p2q2256 = _mm256_andnot_si256(flat256, p2q2256);
+      flat_p2q2 = _mm256_and_si256(flat256, flat_p2q2);
+      p2q2256 = _mm256_or_si256(flat_p2q2, p2q2256);
+      p2 = _mm256_castsi256_si128(p2q2256);
+      q2 = _mm256_extractf128_si256(p2q2256, 1);
+      if (0xffff != _mm_movemask_epi8(_mm_cmpeq_epi8(flat2, zero))) {
+        flat2_256 =
+            _mm256_insertf128_si256(_mm256_castsi128_si256(flat2), flat2, 0x1);
+        p256_6 = _mm256_shuffle_epi8(p256_6, filter);
+        p256_5 = _mm256_shuffle_epi8(p256_5, filter);
+        p256_4 = _mm256_shuffle_epi8(p256_4, filter);
+        q256_4 = _mm256_shuffle_epi8(q256_4, filter);
+        q256_5 = _mm256_shuffle_epi8(q256_5, filter);
+        q256_6 = _mm256_shuffle_epi8(q256_6, filter);
+
+        __m256i pixelFilter_p =
+            _mm256_add_epi16(p256_5, _mm256_add_epi16(p256_4, p256_3));
+        __m256i pixelFilter_q =
+            _mm256_add_epi16(q256_5, _mm256_add_epi16(q256_4, q256_3));
+
+        pixelFilter_p = _mm256_add_epi16(pixelFilter_p, p2p1p0);
+        pixelFilter_q = _mm256_add_epi16(pixelFilter_q, q2q1q0);
+
+        pixelFilter_p = _mm256_add_epi16(pixelFilter_p, p256_0);
+        pixelFilter_q = _mm256_add_epi16(pixelFilter_q, q256_0);
+        pixelFilter_p = _mm256_add_epi16(
+            eight, _mm256_add_epi16(pixelFilter_p, pixelFilter_q));
+        pixelFilter_q = pixelFilter_p;
+
+        // Derive p0 and q0
+        pixelFilter_p =
+            _mm256_add_epi16(_mm256_add_epi16(p256_6, p256_1), pixelFilter_p);
+        res_p = _mm256_srli_epi16(pixelFilter_p, 4);
+        pixelFilter_q =
+            _mm256_add_epi16(_mm256_add_epi16(q256_6, q256_1), pixelFilter_q);
+        res_q = _mm256_srli_epi16(pixelFilter_q, 4);
+        __m256i flat2_p0q0 =
+            _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_q), 0xd8);
+        p0q0256 = _mm256_andnot_si256(flat2_256, p0q0256);
+        flat2_p0q0 = _mm256_and_si256(flat2_256, flat2_p0q0);
+        p0q0256 = _mm256_or_si256(flat2_p0q0, p0q0256);
+
+        p0 = _mm256_castsi256_si128(p0q0256);
+        q0 = _mm256_extractf128_si256(p0q0256, 1);
+        _mm_storeu_si128((__m128i *)(s - 1 * p), p0);
+        _mm_storeu_si128((__m128i *)(s - 0 * p), q0);
+
+        // Derive p1 and q1
+        sum_p = _mm256_add_epi16(_mm256_sub_epi16(p256_6, q256_5),
+                                 _mm256_sub_epi16(p256_2, q256_0));
+        pixelFilter_p = _mm256_add_epi16(pixelFilter_p, sum_p);
+        res_p = _mm256_srli_epi16(pixelFilter_p, 4);
+        sum_q = _mm256_add_epi16(_mm256_sub_epi16(q256_6, p256_5),
+                                 _mm256_sub_epi16(q256_2, p256_0));
+        pixelFilter_q = _mm256_add_epi16(pixelFilter_q, sum_q);
+        res_q = _mm256_srli_epi16(pixelFilter_q, 4);
+        __m256i flat2_p1q1 =
+            _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_q), 0xd8);
+        p1q1256 = _mm256_andnot_si256(flat2_256, p1q1256);
+        flat2_p1q1 = _mm256_and_si256(flat2_256, flat2_p1q1);
+        p1q1256 = _mm256_or_si256(flat2_p1q1, p1q1256);
+        p1 = _mm256_castsi256_si128(p1q1256);
+        q1 = _mm256_extractf128_si256(p1q1256, 1);
+        _mm_storeu_si128((__m128i *)(s - 2 * p), p1);
+        _mm_storeu_si128((__m128i *)(s + 1 * p), q1);
+
+        // Derive p2 and q2
+        sum_p = _mm256_add_epi16(_mm256_sub_epi16(p256_6, q256_4),
+                                 _mm256_sub_epi16(p256_3, p256_0));
+        pixelFilter_p = _mm256_add_epi16(pixelFilter_p, sum_p);
+        res_p = _mm256_srli_epi16(pixelFilter_p, 4);
+        sum_q = _mm256_add_epi16(_mm256_sub_epi16(q256_6, p256_4),
+                                 _mm256_sub_epi16(q256_3, q256_0));
+        pixelFilter_q = _mm256_add_epi16(pixelFilter_q, sum_q);
+        res_q = _mm256_srli_epi16(pixelFilter_q, 4);
+        __m256i flat2_p2q2 =
+            _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_q), 0xd8);
+        p2q2256 = _mm256_andnot_si256(flat2_256, p2q2256);
+        flat2_p2q2 = _mm256_and_si256(flat2_256, flat2_p2q2);
+        p2q2256 = _mm256_or_si256(flat2_p2q2, p2q2256);
+        p2 = _mm256_castsi256_si128(p2q2256);
+        q2 = _mm256_extractf128_si256(p2q2256, 1);
+        _mm_storeu_si128((__m128i *)(s - 3 * p), p2);
+        _mm_storeu_si128((__m128i *)(s + 2 * p), q2);
+
+        // Derive p3 and q3
+        sum_p = _mm256_add_epi16(_mm256_sub_epi16(p256_6, q256_3),
+                                 _mm256_sub_epi16(p256_4, p256_1));
+        pixelFilter_p = _mm256_add_epi16(pixelFilter_p, sum_p);
+        res_p = _mm256_srli_epi16(pixelFilter_p, 4);
+        sum_q = _mm256_add_epi16(_mm256_sub_epi16(q256_6, p256_3),
+                                 _mm256_sub_epi16(q256_4, q256_1));
+        pixelFilter_q = _mm256_add_epi16(pixelFilter_q, sum_q);
+        res_q = _mm256_srli_epi16(pixelFilter_q, 4);
+        __m256i flat2_p3q3 =
+            _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_q), 0xd8);
+        p3q3256 = _mm256_andnot_si256(flat2_256, p3q3256);
+        flat2_p3q3 = _mm256_and_si256(flat2_256, flat2_p3q3);
+        p3q3256 = _mm256_or_si256(flat2_p3q3, p3q3256);
+        p3 = _mm256_castsi256_si128(p3q3256);
+        q3 = _mm256_extractf128_si256(p3q3256, 1);
+        _mm_storeu_si128((__m128i *)(s - 4 * p), p3);
+        _mm_storeu_si128((__m128i *)(s + 3 * p), q3);
+
+        // Derive p4 and q4
+        sum_p = _mm256_add_epi16(_mm256_sub_epi16(p256_6, q256_2),
+                                 _mm256_sub_epi16(p256_5, p256_2));
+        pixelFilter_p = _mm256_add_epi16(pixelFilter_p, sum_p);
+        res_p = _mm256_srli_epi16(pixelFilter_p, 4);
+        sum_q = _mm256_add_epi16(_mm256_sub_epi16(q256_6, p256_2),
+                                 _mm256_sub_epi16(q256_5, q256_2));
+        pixelFilter_q = _mm256_add_epi16(pixelFilter_q, sum_q);
+        res_q = _mm256_srli_epi16(pixelFilter_q, 4);
+        __m256i flat2_p4q4 =
+            _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_q), 0xd8);
+        p4q4256 = _mm256_andnot_si256(flat2_256, p4q4256);
+        flat2_p4q4 = _mm256_and_si256(flat2_256, flat2_p4q4);
+        p4q4256 = _mm256_or_si256(flat2_p4q4, p4q4256);
+        _mm_storeu_si128((__m128i *)(s - 5 * p),
+                         _mm256_castsi256_si128(p4q4256));
+        _mm_storeu_si128((__m128i *)(s + 4 * p),
+                         _mm256_extractf128_si256(p4q4256, 1));
+
+        // Derive p5 and q5
+        sum_p = _mm256_add_epi16(_mm256_sub_epi16(p256_6, q256_1),
+                                 _mm256_sub_epi16(p256_6, p256_3));
+        pixelFilter_p = _mm256_add_epi16(pixelFilter_p, sum_p);
+        res_p = _mm256_srli_epi16(pixelFilter_p, 4);
+        sum_q = _mm256_add_epi16(_mm256_sub_epi16(q256_6, p256_1),
+                                 _mm256_sub_epi16(q256_6, q256_3));
+        pixelFilter_q = _mm256_add_epi16(pixelFilter_q, sum_q);
+        res_q = _mm256_srli_epi16(pixelFilter_q, 4);
+        __m256i flat2_p5q5 =
+            _mm256_permute4x64_epi64(_mm256_packus_epi16(res_p, res_q), 0xd8);
+        p5q5256 = _mm256_andnot_si256(flat2_256, p5q5256);
+        flat2_p5q5 = _mm256_and_si256(flat2_256, flat2_p5q5);
+        p5q5256 = _mm256_or_si256(flat2_p5q5, p5q5256);
+        _mm_storeu_si128((__m128i *)(s - 6 * p),
+                         _mm256_castsi256_si128(p5q5256));
+        _mm_storeu_si128((__m128i *)(s + 5 * p),
+                         _mm256_extractf128_si256(p5q5256, 1));
+      } else {
+        _mm_storeu_si128((__m128i *)(s - 3 * p), p2);
+        _mm_storeu_si128((__m128i *)(s - 2 * p), p1);
+        _mm_storeu_si128((__m128i *)(s - 1 * p), p0);
+        _mm_storeu_si128((__m128i *)(s - 0 * p), q0);
+        _mm_storeu_si128((__m128i *)(s + 1 * p), q1);
+        _mm_storeu_si128((__m128i *)(s + 2 * p), q2);
+      }
+    } else {
+      _mm_storeu_si128((__m128i *)(s - 2 * p), ps1);
+      _mm_storeu_si128((__m128i *)(s - 1 * p), ps0);
+      _mm_storeu_si128((__m128i *)(s - 0 * p), qs0);
+      _mm_storeu_si128((__m128i *)(s + 1 * p), qs1);
+    }
+  }
+}
+
+void aom_lpf_vertical_14_quad_avx2(unsigned char *s, int pitch,
+                                   const uint8_t *_blimit0,
+                                   const uint8_t *_limit0,
+                                   const uint8_t *_thresh0) {
+  DECLARE_ALIGNED(16, unsigned char, t_dst[256]);
+
+  // Transpose 16x16
+  trans_store_16x16_lpf_vert14(s - 8, pitch, t_dst, 16, 1);
+
+  // Loop filtering
+  aom_lpf_horizontal_14_quad_avx2(t_dst + 8 * 16, 16, _blimit0, _limit0,
+                                  _thresh0);
+
+  // Transpose back
+  trans_store_16x16_lpf_vert14(t_dst, 16, s - 8, pitch, 0);
+}
diff --git a/third_party/aom/aom_dsp/x86/loopfilter_sse2.c b/third_party/aom/aom_dsp/x86/loopfilter_sse2.c
new file mode 100644
index 0000000000..cdf24c332a
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/loopfilter_sse2.c
@@ -0,0 +1,2973 @@
+/*
+ * Copyright (c) 2016, 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 <emmintrin.h>  // SSE2
+
+#include "config/aom_dsp_rtcd.h"
+
+#include "aom_dsp/x86/synonyms.h"
+#include "aom_ports/mem.h"
+#include "aom_ports/emmintrin_compat.h"
+#include "aom_dsp/x86/lpf_common_sse2.h"
+
+static INLINE __m128i abs_diff(__m128i a, __m128i b) {
+  return _mm_or_si128(_mm_subs_epu8(a, b), _mm_subs_epu8(b, a));
+}
+
+// this function treats its input as 2 parallel 8x4 matrices, transposes each of
+// them to 4x8  independently while flipping the second matrix horizontally.
+// Used for 14 taps pq pairs creation
+static INLINE void transpose_pq_14_sse2(__m128i *x0, __m128i *x1, __m128i *x2,
+                                        __m128i *x3, __m128i *q0p0,
+                                        __m128i *q1p1, __m128i *q2p2,
+                                        __m128i *q3p3, __m128i *q4p4,
+                                        __m128i *q5p5, __m128i *q6p6,
+                                        __m128i *q7p7) {
+  __m128i w0, w1, ww0, ww1, w2, w3, ww2, ww3;
+  w0 = _mm_unpacklo_epi8(
+      *x0, *x1);  // 00 10 01 11 02 12 03 13 04 14 05 15 06 16 07 17
+  w1 = _mm_unpacklo_epi8(
+      *x2, *x3);  // 20 30 21 31 22 32 23 33 24 34 25 35 26 36 27 37
+  w2 = _mm_unpackhi_epi8(
+      *x0, *x1);  // 08 18 09 19 010 110 011 111 012 112 013 113 014 114 015 115
+  w3 = _mm_unpackhi_epi8(
+      *x2, *x3);  // 28 38 29 39 210 310 211 311 212 312 213 313 214 314 215 315
+
+  ww0 = _mm_unpacklo_epi16(
+      w0, w1);  // 00 10 20 30 01 11 21 31        02 12 22 32 03 13 23 33
+  ww1 = _mm_unpackhi_epi16(
+      w0, w1);  // 04 14 24 34 05 15 25 35        06 16 26 36 07 17 27 37
+  ww2 = _mm_unpacklo_epi16(
+      w2, w3);  // 08 18 28 38 09 19 29 39       010 110 210 310 011 111 211 311
+  ww3 = _mm_unpackhi_epi16(
+      w2,
+      w3);  // 012 112 212 312 013 113 213 313  014 114 214 314 015 115 215 315
+
+  *q7p7 = _mm_unpacklo_epi32(
+      ww0,
+      _mm_srli_si128(
+          ww3, 12));  // 00 10 20 30  015 115 215 315  xx xx xx xx xx xx xx xx
+  *q6p6 = _mm_unpackhi_epi32(
+      _mm_slli_si128(ww0, 4),
+      ww3);  // 01 11 21 31  014 114 214 314  xx xx xx xxxx xx xx xx
+  *q5p5 = _mm_unpackhi_epi32(
+      ww0,
+      _mm_slli_si128(
+          ww3, 4));  // 02 12 22 32  013 113 213 313  xx xx xx x xx xx xx xxx
+  *q4p4 = _mm_unpacklo_epi32(
+      _mm_srli_si128(ww0, 12),
+      ww3);  // 03 13 23 33  012 112 212 312 xx xx xx xx xx xx xx xx
+  *q3p3 = _mm_unpacklo_epi32(
+      ww1,
+      _mm_srli_si128(
+          ww2, 12));  // 04 14 24 34  011 111 211 311 xx xx xx xx xx xx xx xx
+  *q2p2 = _mm_unpackhi_epi32(
+      _mm_slli_si128(ww1, 4),
+      ww2);  // 05 15 25 35   010 110 210 310 xx xx xx xx xx xx xx xx
+  *q1p1 = _mm_unpackhi_epi32(
+      ww1,
+      _mm_slli_si128(
+          ww2, 4));  // 06 16 26 36   09 19 29 39     xx xx xx xx xx xx xx xx
+  *q0p0 = _mm_unpacklo_epi32(
+      _mm_srli_si128(ww1, 12),
+      ww2);  // 07 17 27 37  08 18 28 38     xx xx xx xx xx xx xx xx
+}
+
+// this function treats its input as 2 parallel 8x4 matrices, transposes each of
+// them  independently while flipping the second matrix horizontaly  Used for 14
+// taps filter pq pairs inverse
+static INLINE void transpose_pq_14_inv_sse2(__m128i *x0, __m128i *x1,
+                                            __m128i *x2, __m128i *x3,
+                                            __m128i *x4, __m128i *x5,
+                                            __m128i *x6, __m128i *x7,
+                                            __m128i *pq0, __m128i *pq1,
+                                            __m128i *pq2, __m128i *pq3) {
+  __m128i w10, w11, w12, w13;
+  __m128i w0, w1, w2, w3, w4, w5;
+  __m128i d0, d1, d2, d3;
+
+  w0 = _mm_unpacklo_epi8(
+      *x0, *x1);  // p 00 10 01 11 02 12 03 13 04 14 05 15 06 16 07 17
+  w1 = _mm_unpacklo_epi8(
+      *x2, *x3);  // p 20 30 21 31 22 32 23 33 24 34 25 35 26 36 27 37
+  w2 = _mm_unpacklo_epi8(
+      *x4, *x5);  // p 40 50 41 51 42 52 43 53 44 54 45 55 46 56 47 57
+  w3 = _mm_unpacklo_epi8(
+      *x6, *x7);  // p 60 70 61 71 62 72 63 73 64 74 65 75 66 76 67 77
+
+  w4 = _mm_unpacklo_epi16(
+      w0, w1);  // 00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33
+  w5 = _mm_unpacklo_epi16(
+      w2, w3);  // 40 50 60 70 41 51 61 71 42 52 62 72 43 53 63 73
+
+  d0 = _mm_unpacklo_epi32(
+      w4, w5);  // 00 10 20 30 40 50 60 70 01 11 21 31 41 51 61 71
+  d2 = _mm_unpackhi_epi32(
+      w4, w5);  // 02 12 22 32 42 52 62 72 03 13 23 33 43 53 63 73
+
+  w10 = _mm_unpacklo_epi8(
+      *x7, *x6);  // q xx xx xx xx xx xx xx xx 00 10 01 11 02 12 03 13
+  w11 = _mm_unpacklo_epi8(
+      *x5, *x4);  // q  xx xx xx xx xx xx xx xx 20 30 21 31 22 32 23 33
+  w12 = _mm_unpacklo_epi8(
+      *x3, *x2);  // q  xx xx xx xx xx xx xx xx 40 50 41 51 42 52 43 53
+  w13 = _mm_unpacklo_epi8(
+      *x1, *x0);  // q  xx xx xx xx xx xx xx xx 60 70 61 71 62 72 63 73
+
+  w4 = _mm_unpackhi_epi16(
+      w10, w11);  // 00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33
+  w5 = _mm_unpackhi_epi16(
+      w12, w13);  // 40 50 60 70 41 51 61 71 42 52 62 72 43 53 63 73
+
+  d1 = _mm_unpacklo_epi32(
+      w4, w5);  // 00 10 20 30 40 50 60 70 01 11 21 31 41 51 61 71
+  d3 = _mm_unpackhi_epi32(
+      w4, w5);  // 02 12 22 32 42 52 62 72 03 13 23 33 43 53 63 73
+
+  *pq0 = _mm_unpacklo_epi64(d0, d1);  // pq
+  *pq1 = _mm_unpackhi_epi64(d0, d1);  // pq
+  *pq2 = _mm_unpacklo_epi64(d2, d3);  // pq
+  *pq3 = _mm_unpackhi_epi64(d2, d3);  // pq
+}
+
+static AOM_FORCE_INLINE void filter4_sse2(__m128i *p1p0, __m128i *q1q0,
+                                          __m128i *hev, __m128i *mask,
+                                          __m128i *qs1qs0, __m128i *ps1ps0) {
+  __m128i filter, filter2filter1, work;
+  __m128i ps1ps0_work, qs1qs0_work;
+  __m128i hev1;
+  const __m128i t3t4 =
+      _mm_set_epi8(0, 0, 0, 0, 0, 0, 0, 0, 3, 3, 3, 3, 4, 4, 4, 4);
+  const __m128i t80 = _mm_set1_epi8((char)0x80);
+  const __m128i ff = _mm_cmpeq_epi8(t80, t80);
+
+  ps1ps0_work = _mm_xor_si128(*p1p0, t80); /* ^ 0x80 */
+  qs1qs0_work = _mm_xor_si128(*q1q0, t80);
+
+  /* int8_t filter = signed_char_clamp(ps1 - qs1) & hev; */
+  work = _mm_subs_epi8(ps1ps0_work, qs1qs0_work);
+  filter = _mm_and_si128(_mm_srli_si128(work, 4), *hev);
+  /* filter = signed_char_clamp(filter + 3 * (qs0 - ps0)) & mask; */
+  filter = _mm_subs_epi8(filter, work);
+  filter = _mm_subs_epi8(filter, work);
+  filter = _mm_subs_epi8(filter, work);  /* + 3 * (qs0 - ps0) */
+  filter = _mm_and_si128(filter, *mask); /* & mask */
+  filter = _mm_unpacklo_epi32(filter, filter);
+
+  /* filter1 = signed_char_clamp(filter + 4) >> 3; */
+  /* filter2 = signed_char_clamp(filter + 3) >> 3; */
+  filter2filter1 = _mm_adds_epi8(filter, t3t4); /* signed_char_clamp */
+  filter2filter1 =
+      _mm_unpacklo_epi8(filter2filter1, filter2filter1);  // goto 16 bit
+  filter2filter1 = _mm_srai_epi16(filter2filter1, 11);    /* >> 3 */
+  filter2filter1 = _mm_packs_epi16(filter2filter1, filter2filter1);
+
+  /* filter = ROUND_POWER_OF_TWO(filter1, 1) & ~hev; */
+  filter = _mm_subs_epi8(filter2filter1, ff);  /* + 1 */
+  filter = _mm_unpacklo_epi8(filter, filter);  // goto 16 bit
+  filter = _mm_srai_epi16(filter, 9);          /* round */
+  filter = _mm_packs_epi16(filter, filter);
+  filter = _mm_andnot_si128(*hev, filter);
+  filter = _mm_unpacklo_epi32(filter, filter);
+
+  filter2filter1 = _mm_unpacklo_epi32(filter2filter1, filter);
+  hev1 = _mm_srli_si128(filter2filter1, 8);
+  /* signed_char_clamp(qs1 - filter), signed_char_clamp(qs0 - filter1) */
+  qs1qs0_work = _mm_subs_epi8(qs1qs0_work, filter2filter1);
+  /* signed_char_clamp(ps1 + filter), signed_char_clamp(ps0 + filter2) */
+  ps1ps0_work = _mm_adds_epi8(ps1ps0_work, hev1);
+
+  *qs1qs0 = _mm_xor_si128(qs1qs0_work, t80); /* ^ 0x80 */
+  *ps1ps0 = _mm_xor_si128(ps1ps0_work, t80); /* ^ 0x80 */
+}
+
+static AOM_FORCE_INLINE void filter4_dual_sse2(__m128i *p1p0, __m128i *q1q0,
+                                               __m128i *hev, __m128i *mask,
+                                               __m128i *qs1qs0,
+                                               __m128i *ps1ps0) {
+  const __m128i t3t4 =
+      _mm_set_epi8(3, 3, 3, 3, 3, 3, 3, 3, 4, 4, 4, 4, 4, 4, 4, 4);
+  const __m128i t80 = _mm_set1_epi8((char)0x80);
+  __m128i filter, filter2filter1, work;
+  __m128i ps1ps0_work, qs1qs0_work;
+  __m128i hev1;
+  const __m128i ff = _mm_cmpeq_epi8(t80, t80);
+
+  ps1ps0_work = _mm_xor_si128(*p1p0, t80); /* ^ 0x80 */
+  qs1qs0_work = _mm_xor_si128(*q1q0, t80);
+
+  /* int8_t filter = signed_char_clamp(ps1 - qs1) & hev; */
+  work = _mm_subs_epi8(ps1ps0_work, qs1qs0_work);
+  filter = _mm_and_si128(_mm_srli_si128(work, 8), *hev);
+  /* filter = signed_char_clamp(filter + 3 * (qs0 - ps0)) & mask; */
+  filter = _mm_subs_epi8(filter, work);
+  filter = _mm_subs_epi8(filter, work);
+  filter = _mm_subs_epi8(filter, work);  /* + 3 * (qs0 - ps0) */
+  filter = _mm_and_si128(filter, *mask); /* & mask */
+  filter = _mm_unpacklo_epi64(filter, filter);
+
+  /* filter1 = signed_char_clamp(filter + 4) >> 3; */
+  /* filter2 = signed_char_clamp(filter + 3) >> 3; */
+  filter2filter1 = _mm_adds_epi8(filter, t3t4); /* signed_char_clamp */
+  filter = _mm_unpackhi_epi8(filter2filter1, filter2filter1);
+  filter2filter1 = _mm_unpacklo_epi8(filter2filter1, filter2filter1);
+  filter2filter1 = _mm_srai_epi16(filter2filter1, 11); /* >> 3 */
+  filter = _mm_srai_epi16(filter, 11);                 /* >> 3 */
+  filter2filter1 = _mm_packs_epi16(filter2filter1, filter);
+
+  /* filter = ROUND_POWER_OF_TWO(filter1, 1) & ~hev; */
+  filter = _mm_subs_epi8(filter2filter1, ff); /* + 1 */
+  filter = _mm_unpacklo_epi8(filter, filter);
+  filter = _mm_srai_epi16(filter, 9); /* round */
+  filter = _mm_packs_epi16(filter, filter);
+  filter = _mm_andnot_si128(*hev, filter);
+
+  hev1 = _mm_unpackhi_epi64(filter2filter1, filter);
+  filter2filter1 = _mm_unpacklo_epi64(filter2filter1, filter);
+
+  /* signed_char_clamp(qs1 - filter), signed_char_clamp(qs0 - filter1) */
+  qs1qs0_work = _mm_subs_epi8(qs1qs0_work, filter2filter1);
+  /* signed_char_clamp(ps1 + filter), signed_char_clamp(ps0 + filter2) */
+  ps1ps0_work = _mm_adds_epi8(ps1ps0_work, hev1);
+  *qs1qs0 = _mm_xor_si128(qs1qs0_work, t80); /* ^ 0x80 */
+  *ps1ps0 = _mm_xor_si128(ps1ps0_work, t80); /* ^ 0x80 */
+}
+
+static AOM_FORCE_INLINE void lpf_internal_4_sse2(
+    __m128i *p1, __m128i *p0, __m128i *q0, __m128i *q1, __m128i *limit,
+    __m128i *thresh, __m128i *q1q0_out, __m128i *p1p0_out) {
+  __m128i q1p1, q0p0, p1p0, q1q0;
+  __m128i abs_p0q0, abs_p1q1;
+  __m128i mask, flat, hev;
+  const __m128i zero = _mm_setzero_si128();
+
+  q1p1 = _mm_unpacklo_epi32(*p1, *q1);
+  q0p0 = _mm_unpacklo_epi32(*p0, *q0);
+
+  p1p0 = _mm_unpacklo_epi32(q0p0, q1p1);
+  q1q0 = _mm_srli_si128(p1p0, 8);
+
+  /* (abs(q1 - q0), abs(p1 - p0) */
+  flat = abs_diff(q1p1, q0p0);
+  /* abs(p1 - q1), abs(p0 - q0) */
+  __m128i abs_p1q1p0q0 = abs_diff(p1p0, q1q0);
+
+  /* const uint8_t hev = hev_mask(thresh, *op1, *op0, *oq0, *oq1); */
+  flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 4));
+  hev = _mm_unpacklo_epi8(flat, zero);
+
+  hev = _mm_cmpgt_epi16(hev, *thresh);
+  hev = _mm_packs_epi16(hev, hev);
+  hev = _mm_unpacklo_epi32(hev, hev);
+
+  abs_p0q0 = _mm_adds_epu8(abs_p1q1p0q0, abs_p1q1p0q0); /* abs(p0 - q0) * 2 */
+  abs_p1q1 = _mm_srli_si128(abs_p1q1p0q0, 4);           /* abs(p1 - q1) */
+  abs_p1q1 = _mm_unpacklo_epi8(abs_p1q1, abs_p1q1);
+  abs_p1q1 = _mm_srli_epi16(abs_p1q1, 9);
+  abs_p1q1 = _mm_packs_epi16(abs_p1q1, abs_p1q1); /* abs(p1 - q1) / 2 */
+  /* abs(p0 - q0) * 2 + abs(p1 - q1) / 2 */
+
+  mask = _mm_adds_epu8(abs_p0q0, abs_p1q1);
+  mask = _mm_unpacklo_epi32(mask, flat);
+  mask = _mm_subs_epu8(mask, *limit);
+  mask = _mm_cmpeq_epi8(mask, zero);
+  mask = _mm_and_si128(mask, _mm_srli_si128(mask, 4));
+
+  filter4_sse2(&p1p0, &q1q0, &hev, &mask, q1q0_out, p1p0_out);
+}
+
+static AOM_FORCE_INLINE void lpf_internal_4_dual_sse2(
+    __m128i *p1, __m128i *p0, __m128i *q0, __m128i *q1, __m128i *limit,
+    __m128i *thresh, __m128i *q1q0_out, __m128i *p1p0_out) {
+  __m128i q1p1, q0p0, p1p0, q1q0;
+  __m128i abs_p0q0, abs_p1q1;
+  __m128i mask, hev;
+  const __m128i zero = _mm_setzero_si128();
+
+  q1p1 = _mm_unpacklo_epi64(*p1, *q1);
+  q0p0 = _mm_unpacklo_epi64(*p0, *q0);
+
+  p1p0 = _mm_unpacklo_epi64(q0p0, q1p1);
+  q1q0 = _mm_unpackhi_epi64(q0p0, q1p1);
+
+  /* (abs(q1 - q0), abs(p1 - p0) */
+  __m128i flat = abs_diff(q1p1, q0p0);
+  /* abs(p1 - q1), abs(p0 - q0) */
+  const __m128i abs_p1q1p0q0 = abs_diff(p1p0, q1q0);
+
+  /* const uint8_t hev = hev_mask(thresh, *op1, *op0, *oq0, *oq1); */
+  flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 8));
+  hev = _mm_unpacklo_epi8(flat, zero);
+
+  hev = _mm_cmpgt_epi16(hev, *thresh);
+  hev = _mm_packs_epi16(hev, hev);
+
+  /* const int8_t mask = filter_mask2(*limit, *blimit, */
+  /*                                  p1, p0, q0, q1); */
+  abs_p0q0 = _mm_adds_epu8(abs_p1q1p0q0, abs_p1q1p0q0); /* abs(p0 - q0) * 2 */
+  abs_p1q1 = _mm_unpackhi_epi8(abs_p1q1p0q0, abs_p1q1p0q0); /* abs(p1 - q1) */
+  abs_p1q1 = _mm_srli_epi16(abs_p1q1, 9);
+  abs_p1q1 = _mm_packs_epi16(abs_p1q1, abs_p1q1); /* abs(p1 - q1) / 2 */
+  /* abs(p0 - q0) * 2 + abs(p1 - q1) / 2 */
+  mask = _mm_adds_epu8(abs_p0q0, abs_p1q1);
+  mask = _mm_unpacklo_epi64(mask, flat);
+  mask = _mm_subs_epu8(mask, *limit);
+  mask = _mm_cmpeq_epi8(mask, zero);
+  mask = _mm_and_si128(mask, _mm_srli_si128(mask, 8));
+
+  filter4_dual_sse2(&p1p0, &q1q0, &hev, &mask, q1q0_out, p1p0_out);
+}
+
+void aom_lpf_horizontal_4_sse2(uint8_t *s, int p /* pitch */,
+                               const uint8_t *_blimit, const uint8_t *_limit,
+                               const uint8_t *_thresh) {
+  const __m128i zero = _mm_setzero_si128();
+  __m128i limit = _mm_unpacklo_epi32(_mm_loadl_epi64((const __m128i *)_blimit),
+                                     _mm_loadl_epi64((const __m128i *)_limit));
+  __m128i thresh =
+      _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)_thresh), zero);
+
+  __m128i qs1qs0, ps1ps0;
+  __m128i p1, p0, q0, q1;
+
+  p1 = xx_loadl_32(s - 2 * p);
+  p0 = xx_loadl_32(s - 1 * p);
+  q0 = xx_loadl_32(s - 0 * p);
+  q1 = xx_loadl_32(s + 1 * p);
+
+  lpf_internal_4_sse2(&p1, &p0, &q0, &q1, &limit, &thresh, &qs1qs0, &ps1ps0);
+
+  xx_storel_32(s - 1 * p, ps1ps0);
+  xx_storel_32(s - 2 * p, _mm_srli_si128(ps1ps0, 4));
+  xx_storel_32(s + 0 * p, qs1qs0);
+  xx_storel_32(s + 1 * p, _mm_srli_si128(qs1qs0, 4));
+}
+
+void aom_lpf_vertical_4_sse2(uint8_t *s, int p /* pitch */,
+                             const uint8_t *_blimit, const uint8_t *_limit,
+                             const uint8_t *_thresh) {
+  __m128i p1p0, q1q0;
+  __m128i p1, p0, q0, q1;
+
+  const __m128i zero = _mm_setzero_si128();
+  __m128i limit = _mm_unpacklo_epi32(_mm_loadl_epi64((const __m128i *)_blimit),
+                                     _mm_loadl_epi64((const __m128i *)_limit));
+  __m128i thresh =
+      _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)_thresh), zero);
+
+  __m128i x0, x1, x2, x3;
+  __m128i d0, d1, d2, d3;
+  x0 = _mm_loadl_epi64((__m128i *)(s - 2 + 0 * p));
+  x1 = _mm_loadl_epi64((__m128i *)(s - 2 + 1 * p));
+  x2 = _mm_loadl_epi64((__m128i *)(s - 2 + 2 * p));
+  x3 = _mm_loadl_epi64((__m128i *)(s - 2 + 3 * p));
+
+  transpose4x8_8x4_low_sse2(&x0, &x1, &x2, &x3, &p1, &p0, &q0, &q1);
+
+  lpf_internal_4_sse2(&p1, &p0, &q0, &q1, &limit, &thresh, &q1q0, &p1p0);
+
+  // Transpose 8x4 to 4x8
+  p1 = _mm_srli_si128(p1p0, 4);
+  q1 = _mm_srli_si128(q1q0, 4);
+
+  transpose4x8_8x4_low_sse2(&p1, &p1p0, &q1q0, &q1, &d0, &d1, &d2, &d3);
+
+  xx_storel_32(s + 0 * p - 2, d0);
+  xx_storel_32(s + 1 * p - 2, d1);
+  xx_storel_32(s + 2 * p - 2, d2);
+  xx_storel_32(s + 3 * p - 2, d3);
+}
+
+static INLINE void store_buffer_horz_8(__m128i x, int p, int num, uint8_t *s) {
+  xx_storel_32(s - (num + 1) * p, x);
+  xx_storel_32(s + num * p, _mm_srli_si128(x, 4));
+}
+
+static AOM_FORCE_INLINE void lpf_internal_14_dual_sse2(
+    __m128i *q6p6, __m128i *q5p5, __m128i *q4p4, __m128i *q3p3, __m128i *q2p2,
+    __m128i *q1p1, __m128i *q0p0, __m128i *blimit, __m128i *limit,
+    __m128i *thresh) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i one = _mm_set1_epi8(1);
+  __m128i mask, hev, flat, flat2;
+  __m128i qs0ps0, qs1ps1;
+  __m128i p1p0, q1q0, qs1qs0, ps1ps0;
+  __m128i abs_p1p0;
+
+  p1p0 = _mm_unpacklo_epi64(*q0p0, *q1p1);
+  q1q0 = _mm_unpackhi_epi64(*q0p0, *q1p1);
+
+  {
+    __m128i abs_p1q1, abs_p0q0, abs_q1q0;
+    __m128i fe, ff, work;
+    abs_p1p0 = abs_diff(*q1p1, *q0p0);
+    abs_q1q0 = _mm_srli_si128(abs_p1p0, 8);
+    fe = _mm_set1_epi8((char)0xfe);
+    ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0);
+    abs_p0q0 = abs_diff(p1p0, q1q0);
+    abs_p1q1 = _mm_srli_si128(abs_p0q0, 8);
+    abs_p0q0 = _mm_unpacklo_epi64(abs_p0q0, zero);
+
+    flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
+    hev = _mm_subs_epu8(flat, *thresh);
+    hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+    // replicate for the further "merged variables" usage
+    hev = _mm_unpacklo_epi64(hev, hev);
+
+    abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0);
+    abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
+    mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), *blimit);
+    mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
+    // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit) * -1;
+    mask = _mm_max_epu8(abs_p1p0, mask);
+    // mask |= (abs(p1 - p0) > limit) * -1;
+    // mask |= (abs(q1 - q0) > limit) * -1;
+
+    work = _mm_max_epu8(abs_diff(*q2p2, *q1p1), abs_diff(*q3p3, *q2p2));
+    mask = _mm_max_epu8(work, mask);
+    mask = _mm_max_epu8(mask, _mm_srli_si128(mask, 8));
+    mask = _mm_subs_epu8(mask, *limit);
+    mask = _mm_cmpeq_epi8(mask, zero);
+  }
+
+  // lp filter - the same for 6, 8 and 14 versions
+  filter4_dual_sse2(&p1p0, &q1q0, &hev, &mask, &qs1qs0, &ps1ps0);
+  qs0ps0 = _mm_unpacklo_epi64(ps1ps0, qs1qs0);
+  qs1ps1 = _mm_unpackhi_epi64(ps1ps0, qs1qs0);
+  // loopfilter done
+
+  __m128i flat2_q5p5, flat2_q4p4, flat2_q3p3, flat2_q2p2;
+  __m128i flat2_q1p1, flat2_q0p0, flat_q2p2, flat_q1p1, flat_q0p0;
+
+  __m128i work;
+  flat = _mm_max_epu8(abs_diff(*q2p2, *q0p0), abs_diff(*q3p3, *q0p0));
+  flat = _mm_max_epu8(abs_p1p0, flat);
+  flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 8));
+  flat = _mm_subs_epu8(flat, one);
+  flat = _mm_cmpeq_epi8(flat, zero);
+  flat = _mm_and_si128(flat, mask);
+
+  // if flat ==0 then flat2 is zero as well and we don't need any calc below
+  // sse4.1 if (0==_mm_test_all_zeros(flat,ff))
+  if (0xffff != _mm_movemask_epi8(_mm_cmpeq_epi8(flat, zero))) {
+    // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+    // flat and wide flat calculations
+
+    const __m128i eight = _mm_set1_epi16(8);
+    const __m128i four = _mm_set1_epi16(4);
+    __m128i p6_16, p5_16, p4_16, p3_16, p2_16, p1_16, p0_16;
+    __m128i q6_16, q5_16, q4_16, q3_16, q2_16, q1_16, q0_16;
+    __m128i pixelFilter_p, pixelFilter_q;
+    __m128i pixetFilter_p2p1p0, pixetFilter_q2q1q0;
+    __m128i sum_p6, sum_q6;
+    __m128i sum_p3, sum_q3, res_p, res_q;
+
+    p6_16 = _mm_unpacklo_epi8(*q6p6, zero);
+    p5_16 = _mm_unpacklo_epi8(*q5p5, zero);
+    p4_16 = _mm_unpacklo_epi8(*q4p4, zero);
+    p3_16 = _mm_unpacklo_epi8(*q3p3, zero);
+    p2_16 = _mm_unpacklo_epi8(*q2p2, zero);
+    p1_16 = _mm_unpacklo_epi8(*q1p1, zero);
+    p0_16 = _mm_unpacklo_epi8(*q0p0, zero);
+    q0_16 = _mm_unpackhi_epi8(*q0p0, zero);
+    q1_16 = _mm_unpackhi_epi8(*q1p1, zero);
+    q2_16 = _mm_unpackhi_epi8(*q2p2, zero);
+    q3_16 = _mm_unpackhi_epi8(*q3p3, zero);
+    q4_16 = _mm_unpackhi_epi8(*q4p4, zero);
+    q5_16 = _mm_unpackhi_epi8(*q5p5, zero);
+    q6_16 = _mm_unpackhi_epi8(*q6p6, zero);
+    pixelFilter_p = _mm_add_epi16(p5_16, _mm_add_epi16(p4_16, p3_16));
+    pixelFilter_q = _mm_add_epi16(q5_16, _mm_add_epi16(q4_16, q3_16));
+
+    pixetFilter_p2p1p0 = _mm_add_epi16(p0_16, _mm_add_epi16(p2_16, p1_16));
+    pixelFilter_p = _mm_add_epi16(pixelFilter_p, pixetFilter_p2p1p0);
+
+    pixetFilter_q2q1q0 = _mm_add_epi16(q0_16, _mm_add_epi16(q2_16, q1_16));
+    pixelFilter_q = _mm_add_epi16(pixelFilter_q, pixetFilter_q2q1q0);
+    pixelFilter_p =
+        _mm_add_epi16(eight, _mm_add_epi16(pixelFilter_p, pixelFilter_q));
+    pixetFilter_p2p1p0 = _mm_add_epi16(
+        four, _mm_add_epi16(pixetFilter_p2p1p0, pixetFilter_q2q1q0));
+    res_p = _mm_srli_epi16(
+        _mm_add_epi16(pixelFilter_p,
+                      _mm_add_epi16(_mm_add_epi16(p6_16, p0_16),
+                                    _mm_add_epi16(p1_16, q0_16))),
+        4);
+    res_q = _mm_srli_epi16(
+        _mm_add_epi16(pixelFilter_p,
+                      _mm_add_epi16(_mm_add_epi16(q6_16, q0_16),
+                                    _mm_add_epi16(p0_16, q1_16))),
+        4);
+    flat2_q0p0 = _mm_packus_epi16(res_p, res_q);
+
+    res_p = _mm_srli_epi16(
+        _mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(p3_16, p0_16)), 3);
+    res_q = _mm_srli_epi16(
+        _mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(q3_16, q0_16)), 3);
+
+    flat_q0p0 = _mm_packus_epi16(res_p, res_q);
+
+    sum_p6 = _mm_add_epi16(p6_16, p6_16);
+    sum_q6 = _mm_add_epi16(q6_16, q6_16);
+    sum_p3 = _mm_add_epi16(p3_16, p3_16);
+    sum_q3 = _mm_add_epi16(q3_16, q3_16);
+
+    pixelFilter_q = _mm_sub_epi16(pixelFilter_p, p5_16);
+    pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q5_16);
+
+    res_p = _mm_srli_epi16(
+        _mm_add_epi16(
+            pixelFilter_p,
+            _mm_add_epi16(sum_p6,
+                          _mm_add_epi16(p1_16, _mm_add_epi16(p2_16, p0_16)))),
+        4);
+    res_q = _mm_srli_epi16(
+        _mm_add_epi16(
+            pixelFilter_q,
+            _mm_add_epi16(sum_q6,
+                          _mm_add_epi16(q1_16, _mm_add_epi16(q0_16, q2_16)))),
+        4);
+    flat2_q1p1 = _mm_packus_epi16(res_p, res_q);
+
+    pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_p2p1p0, p2_16);
+    pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q2_16);
+    res_p = _mm_srli_epi16(
+        _mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(sum_p3, p1_16)), 3);
+    res_q = _mm_srli_epi16(
+        _mm_add_epi16(pixetFilter_q2q1q0, _mm_add_epi16(sum_q3, q1_16)), 3);
+    flat_q1p1 = _mm_packus_epi16(res_p, res_q);
+
+    pixetFilter_p2p1p0 = _mm_sub_epi16(pixetFilter_p2p1p0, q1_16);
+    pixetFilter_q2q1q0 = _mm_sub_epi16(pixetFilter_q2q1q0, p1_16);
+
+    sum_p3 = _mm_add_epi16(sum_p3, p3_16);
+    sum_q3 = _mm_add_epi16(sum_q3, q3_16);
+
+    res_p = _mm_srli_epi16(
+        _mm_add_epi16(pixetFilter_p2p1p0, _mm_add_epi16(sum_p3, p2_16)), 3);
+    res_q = _mm_srli_epi16(
+        _mm_add_epi16(pixetFilter_q2q1q0, _mm_add_epi16(sum_q3, q2_16)), 3);
+    flat_q2p2 = _mm_packus_epi16(res_p, res_q);
+
+    // work with flat2
+    flat2 = _mm_max_epu8(abs_diff(*q4p4, *q0p0), abs_diff(*q5p5, *q0p0));
+    work = abs_diff(*q6p6, *q0p0);
+    flat2 = _mm_max_epu8(work, flat2);
+    flat2 = _mm_max_epu8(flat2, _mm_srli_si128(flat2, 8));
+    flat2 = _mm_subs_epu8(flat2, one);
+    flat2 = _mm_cmpeq_epi8(flat2, zero);
+    flat2 = _mm_and_si128(flat2, flat);  // flat2 & flat & mask
+
+    // ~~~~~~~~~~ apply flat ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+    flat = _mm_unpacklo_epi64(flat, flat);
+    *q2p2 = _mm_andnot_si128(flat, *q2p2);
+    flat_q2p2 = _mm_and_si128(flat, flat_q2p2);
+    *q2p2 = _mm_or_si128(*q2p2, flat_q2p2);
+
+    qs1ps1 = _mm_andnot_si128(flat, qs1ps1);
+    flat_q1p1 = _mm_and_si128(flat, flat_q1p1);
+    *q1p1 = _mm_or_si128(qs1ps1, flat_q1p1);
+
+    qs0ps0 = _mm_andnot_si128(flat, qs0ps0);
+    flat_q0p0 = _mm_and_si128(flat, flat_q0p0);
+    *q0p0 = _mm_or_si128(qs0ps0, flat_q0p0);
+
+    if (0xffff != _mm_movemask_epi8(_mm_cmpeq_epi8(flat2, zero))) {
+      pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q4_16);
+      pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p4_16);
+
+      sum_p6 = _mm_add_epi16(sum_p6, p6_16);
+      sum_q6 = _mm_add_epi16(sum_q6, q6_16);
+
+      res_p = _mm_srli_epi16(
+          _mm_add_epi16(
+              pixelFilter_p,
+              _mm_add_epi16(sum_p6,
+                            _mm_add_epi16(p2_16, _mm_add_epi16(p3_16, p1_16)))),
+          4);
+      res_q = _mm_srli_epi16(
+          _mm_add_epi16(
+              pixelFilter_q,
+              _mm_add_epi16(sum_q6,
+                            _mm_add_epi16(q2_16, _mm_add_epi16(q1_16, q3_16)))),
+          4);
+      flat2_q2p2 = _mm_packus_epi16(res_p, res_q);
+
+      sum_p6 = _mm_add_epi16(sum_p6, p6_16);
+      sum_q6 = _mm_add_epi16(sum_q6, q6_16);
+
+      pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q3_16);
+      pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p3_16);
+
+      res_p = _mm_srli_epi16(
+          _mm_add_epi16(
+              pixelFilter_p,
+              _mm_add_epi16(sum_p6,
+                            _mm_add_epi16(p3_16, _mm_add_epi16(p4_16, p2_16)))),
+          4);
+      res_q = _mm_srli_epi16(
+          _mm_add_epi16(
+              pixelFilter_q,
+              _mm_add_epi16(sum_q6,
+                            _mm_add_epi16(q3_16, _mm_add_epi16(q2_16, q4_16)))),
+          4);
+      flat2_q3p3 = _mm_packus_epi16(res_p, res_q);
+
+      sum_p6 = _mm_add_epi16(sum_p6, p6_16);
+      sum_q6 = _mm_add_epi16(sum_q6, q6_16);
+
+      pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q2_16);
+      pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p2_16);
+
+      res_p = _mm_srli_epi16(
+          _mm_add_epi16(
+              pixelFilter_p,
+              _mm_add_epi16(sum_p6,
+                            _mm_add_epi16(p4_16, _mm_add_epi16(p5_16, p3_16)))),
+          4);
+      res_q = _mm_srli_epi16(
+          _mm_add_epi16(
+              pixelFilter_q,
+              _mm_add_epi16(sum_q6,
+                            _mm_add_epi16(q4_16, _mm_add_epi16(q3_16, q5_16)))),
+          4);
+      flat2_q4p4 = _mm_packus_epi16(res_p, res_q);
+
+      sum_p6 = _mm_add_epi16(sum_p6, p6_16);
+      sum_q6 = _mm_add_epi16(sum_q6, q6_16);
+      pixelFilter_p = _mm_sub_epi16(pixelFilter_p, q1_16);
+      pixelFilter_q = _mm_sub_epi16(pixelFilter_q, p1_16);
+
+      res_p = _mm_srli_epi16(
+          _mm_add_epi16(
+              pixelFilter_p,
+              _mm_add_epi16(sum_p6,
+                            _mm_add_epi16(p5_16, _mm_add_epi16(p6_16, p4_16)))),
+          4);
+      res_q = _mm_srli_epi16(
+          _mm_add_epi16(
+              pixelFilter_q,
+              _mm_add_epi16(sum_q6,
+                            _mm_add_epi16(q5_16, _mm_add_epi16(q6_16, q4_16)))),
+          4);
+      flat2_q5p5 = _mm_packus_epi16(res_p, res_q);
+
+      // wide flat
+      // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+      flat2 = _mm_unpacklo_epi64(flat2, flat2);
+
+      *q5p5 = _mm_andnot_si128(flat2, *q5p5);
+      flat2_q5p5 = _mm_and_si128(flat2, flat2_q5p5);
+      *q5p5 = _mm_or_si128(*q5p5, flat2_q5p5);
+
+      *q4p4 = _mm_andnot_si128(flat2, *q4p4);
+      flat2_q4p4 = _mm_and_si128(flat2, flat2_q4p4);
+      *q4p4 = _mm_or_si128(*q4p4, flat2_q4p4);
+
+      *q3p3 = _mm_andnot_si128(flat2, *q3p3);
+      flat2_q3p3 = _mm_and_si128(flat2, flat2_q3p3);
+      *q3p3 = _mm_or_si128(*q3p3, flat2_q3p3);
+
+      *q2p2 = _mm_andnot_si128(flat2, *q2p2);
+      flat2_q2p2 = _mm_and_si128(flat2, flat2_q2p2);
+      *q2p2 = _mm_or_si128(*q2p2, flat2_q2p2);
+
+      *q1p1 = _mm_andnot_si128(flat2, *q1p1);
+      flat2_q1p1 = _mm_and_si128(flat2, flat2_q1p1);
+      *q1p1 = _mm_or_si128(*q1p1, flat2_q1p1);
+
+      *q0p0 = _mm_andnot_si128(flat2, *q0p0);
+      flat2_q0p0 = _mm_and_si128(flat2, flat2_q0p0);
+      *q0p0 = _mm_or_si128(*q0p0, flat2_q0p0);
+    }
+  } else {
+    *q0p0 = qs0ps0;
+    *q1p1 = qs1ps1;
+  }
+}
+
+static AOM_FORCE_INLINE void lpf_internal_14_sse2(
+    __m128i *q6p6, __m128i *q5p5, __m128i *q4p4, __m128i *q3p3, __m128i *q2p2,
+    __m128i *q1p1, __m128i *q0p0, __m128i *blimit, __m128i *limit,
+    __m128i *thresh) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i one = _mm_set1_epi8(1);
+  __m128i mask, hev, flat, flat2;
+  __m128i flat2_pq[6], flat_pq[3];
+  __m128i qs0ps0, qs1ps1;
+  __m128i p1p0, q1q0, qs1qs0, ps1ps0;
+  __m128i abs_p1p0;
+
+  p1p0 = _mm_unpacklo_epi32(*q0p0, *q1p1);
+  q1q0 = _mm_srli_si128(p1p0, 8);
+
+  __m128i fe, ff, work;
+  {
+    __m128i abs_p1q1, abs_p0q0, abs_q1q0;
+    abs_p1p0 = abs_diff(*q1p1, *q0p0);
+    abs_q1q0 = _mm_srli_si128(abs_p1p0, 4);
+    fe = _mm_set1_epi8((char)0xfe);
+    ff = _mm_cmpeq_epi8(fe, fe);
+    abs_p0q0 = abs_diff(p1p0, q1q0);
+    abs_p1q1 = _mm_srli_si128(abs_p0q0, 4);
+
+    flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
+
+    hev = _mm_subs_epu8(flat, *thresh);
+    hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+    // replicate for the further "merged variables" usage
+    hev = _mm_unpacklo_epi32(hev, hev);
+
+    abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0);
+    abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
+    mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), *blimit);
+    mask = _mm_unpacklo_epi32(mask, zero);
+    mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
+    // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit) * -1;
+    mask = _mm_max_epu8(abs_p1p0, mask);
+    // mask |= (abs(p1 - p0) > limit) * -1;
+    // mask |= (abs(q1 - q0) > limit) * -1;
+
+    work = _mm_max_epu8(abs_diff(*q2p2, *q1p1), abs_diff(*q3p3, *q2p2));
+    mask = _mm_max_epu8(work, mask);
+    mask = _mm_max_epu8(mask, _mm_srli_si128(mask, 4));
+    mask = _mm_subs_epu8(mask, *limit);
+    mask = _mm_cmpeq_epi8(mask, zero);
+  }
+
+  // lp filter - the same for 6, 8 and 14 versions
+  filter4_sse2(&p1p0, &q1q0, &hev, &mask, &qs1qs0, &ps1ps0);
+  qs0ps0 = _mm_unpacklo_epi32(ps1ps0, qs1qs0);
+  qs1ps1 = _mm_srli_si128(qs0ps0, 8);
+  // loopfilter done
+
+  flat = _mm_max_epu8(abs_diff(*q2p2, *q0p0), abs_diff(*q3p3, *q0p0));
+  flat = _mm_max_epu8(abs_p1p0, flat);
+  flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 4));
+  flat = _mm_subs_epu8(flat, one);
+  flat = _mm_cmpeq_epi8(flat, zero);
+  flat = _mm_and_si128(flat, mask);
+  flat = _mm_unpacklo_epi32(flat, flat);
+  flat = _mm_unpacklo_epi64(flat, flat);
+
+  // if flat ==0 then flat2 is zero as well and we don't need any calc below
+  // sse4.1 if (0==_mm_test_all_zeros(flat,ff))
+  if (0xffff != _mm_movemask_epi8(_mm_cmpeq_epi8(flat, zero))) {
+    // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+    // flat and wide flat calculations
+    __m128i q5_16, q4_16, q3_16, q2_16, q1_16, q0_16;
+    __m128i pq_16[7];
+    const __m128i eight = _mm_set1_epi16(8);
+    const __m128i four = _mm_set1_epi16(4);
+    __m128i sum_p6;
+    __m128i sum_p3;
+
+    pq_16[0] = _mm_unpacklo_epi8(*q0p0, zero);
+    pq_16[1] = _mm_unpacklo_epi8(*q1p1, zero);
+    pq_16[2] = _mm_unpacklo_epi8(*q2p2, zero);
+    pq_16[3] = _mm_unpacklo_epi8(*q3p3, zero);
+    pq_16[4] = _mm_unpacklo_epi8(*q4p4, zero);
+    pq_16[5] = _mm_unpacklo_epi8(*q5p5, zero);
+    pq_16[6] = _mm_unpacklo_epi8(*q6p6, zero);
+    q0_16 = _mm_srli_si128(pq_16[0], 8);
+    q1_16 = _mm_srli_si128(pq_16[1], 8);
+    q2_16 = _mm_srli_si128(pq_16[2], 8);
+    q3_16 = _mm_srli_si128(pq_16[3], 8);
+    q4_16 = _mm_srli_si128(pq_16[4], 8);
+    q5_16 = _mm_srli_si128(pq_16[5], 8);
+
+    __m128i flat_p[3], flat_q[3];
+    __m128i flat2_p[6], flat2_q[6];
+
+    __m128i work0, work0_0, work0_1, sum_p_0;
+    __m128i sum_p = _mm_add_epi16(pq_16[5], _mm_add_epi16(pq_16[4], pq_16[3]));
+    __m128i sum_lp = _mm_add_epi16(pq_16[0], _mm_add_epi16(pq_16[2], pq_16[1]));
+    sum_p = _mm_add_epi16(sum_p, sum_lp);
+
+    __m128i sum_lq = _mm_srli_si128(sum_lp, 8);
+    __m128i sum_q = _mm_srli_si128(sum_p, 8);
+
+    sum_p_0 = _mm_add_epi16(eight, _mm_add_epi16(sum_p, sum_q));
+    sum_lp = _mm_add_epi16(four, _mm_add_epi16(sum_lp, sum_lq));
+
+    flat_p[0] = _mm_add_epi16(sum_lp, _mm_add_epi16(pq_16[3], pq_16[0]));
+    flat_q[0] = _mm_add_epi16(sum_lp, _mm_add_epi16(q3_16, q0_16));
+
+    sum_p6 = _mm_add_epi16(pq_16[6], pq_16[6]);
+    sum_p3 = _mm_add_epi16(pq_16[3], pq_16[3]);
+
+    sum_q = _mm_sub_epi16(sum_p_0, pq_16[5]);
+    sum_p = _mm_sub_epi16(sum_p_0, q5_16);
+
+    work0_0 = _mm_add_epi16(_mm_add_epi16(pq_16[6], pq_16[0]), pq_16[1]);
+    work0_1 = _mm_add_epi16(
+        sum_p6, _mm_add_epi16(pq_16[1], _mm_add_epi16(pq_16[2], pq_16[0])));
+
+    sum_lq = _mm_sub_epi16(sum_lp, pq_16[2]);
+    sum_lp = _mm_sub_epi16(sum_lp, q2_16);
+
+    work0 = _mm_add_epi16(sum_p3, pq_16[1]);
+    flat_p[1] = _mm_add_epi16(sum_lp, work0);
+    flat_q[1] = _mm_add_epi16(sum_lq, _mm_srli_si128(work0, 8));
+
+    flat_pq[0] = _mm_srli_epi16(_mm_unpacklo_epi64(flat_p[0], flat_q[0]), 3);
+    flat_pq[1] = _mm_srli_epi16(_mm_unpacklo_epi64(flat_p[1], flat_q[1]), 3);
+    flat_pq[0] = _mm_packus_epi16(flat_pq[0], flat_pq[0]);
+    flat_pq[1] = _mm_packus_epi16(flat_pq[1], flat_pq[1]);
+
+    sum_lp = _mm_sub_epi16(sum_lp, q1_16);
+    sum_lq = _mm_sub_epi16(sum_lq, pq_16[1]);
+
+    sum_p3 = _mm_add_epi16(sum_p3, pq_16[3]);
+    work0 = _mm_add_epi16(sum_p3, pq_16[2]);
+
+    flat_p[2] = _mm_add_epi16(sum_lp, work0);
+    flat_q[2] = _mm_add_epi16(sum_lq, _mm_srli_si128(work0, 8));
+    flat_pq[2] = _mm_srli_epi16(_mm_unpacklo_epi64(flat_p[2], flat_q[2]), 3);
+    flat_pq[2] = _mm_packus_epi16(flat_pq[2], flat_pq[2]);
+
+    // ~~~~~~~~~~~~~~~~~~~~~~~~~~~ flat 2 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+    flat2 = _mm_max_epu8(abs_diff(*q4p4, *q0p0), abs_diff(*q5p5, *q0p0));
+
+    work = abs_diff(*q6p6, *q0p0);
+    flat2 = _mm_max_epu8(work, flat2);
+    flat2 = _mm_max_epu8(flat2, _mm_srli_si128(flat2, 4));
+    flat2 = _mm_subs_epu8(flat2, one);
+    flat2 = _mm_cmpeq_epi8(flat2, zero);
+    flat2 = _mm_and_si128(flat2, flat);  // flat2 & flat & mask
+    flat2 = _mm_unpacklo_epi32(flat2, flat2);
+
+    // ~~~~~~~~~~ apply flat ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+    qs0ps0 = _mm_andnot_si128(flat, qs0ps0);
+    flat_pq[0] = _mm_and_si128(flat, flat_pq[0]);
+    *q0p0 = _mm_or_si128(qs0ps0, flat_pq[0]);
+
+    qs1ps1 = _mm_andnot_si128(flat, qs1ps1);
+    flat_pq[1] = _mm_and_si128(flat, flat_pq[1]);
+    *q1p1 = _mm_or_si128(qs1ps1, flat_pq[1]);
+
+    *q2p2 = _mm_andnot_si128(flat, *q2p2);
+    flat_pq[2] = _mm_and_si128(flat, flat_pq[2]);
+    *q2p2 = _mm_or_si128(*q2p2, flat_pq[2]);
+
+    if (0xffff != _mm_movemask_epi8(_mm_cmpeq_epi8(flat2, zero))) {
+      flat2_p[0] = _mm_add_epi16(sum_p_0, _mm_add_epi16(work0_0, q0_16));
+      flat2_q[0] = _mm_add_epi16(
+          sum_p_0, _mm_add_epi16(_mm_srli_si128(work0_0, 8), pq_16[0]));
+
+      flat2_p[1] = _mm_add_epi16(sum_p, work0_1);
+      flat2_q[1] = _mm_add_epi16(sum_q, _mm_srli_si128(work0_1, 8));
+
+      flat2_pq[0] =
+          _mm_srli_epi16(_mm_unpacklo_epi64(flat2_p[0], flat2_q[0]), 4);
+      flat2_pq[1] =
+          _mm_srli_epi16(_mm_unpacklo_epi64(flat2_p[1], flat2_q[1]), 4);
+      flat2_pq[0] = _mm_packus_epi16(flat2_pq[0], flat2_pq[0]);
+      flat2_pq[1] = _mm_packus_epi16(flat2_pq[1], flat2_pq[1]);
+
+      sum_p = _mm_sub_epi16(sum_p, q4_16);
+      sum_q = _mm_sub_epi16(sum_q, pq_16[4]);
+
+      sum_p6 = _mm_add_epi16(sum_p6, pq_16[6]);
+      work0 = _mm_add_epi16(
+          sum_p6, _mm_add_epi16(pq_16[2], _mm_add_epi16(pq_16[3], pq_16[1])));
+      flat2_p[2] = _mm_add_epi16(sum_p, work0);
+      flat2_q[2] = _mm_add_epi16(sum_q, _mm_srli_si128(work0, 8));
+      flat2_pq[2] =
+          _mm_srli_epi16(_mm_unpacklo_epi64(flat2_p[2], flat2_q[2]), 4);
+      flat2_pq[2] = _mm_packus_epi16(flat2_pq[2], flat2_pq[2]);
+
+      sum_p6 = _mm_add_epi16(sum_p6, pq_16[6]);
+      sum_p = _mm_sub_epi16(sum_p, q3_16);
+      sum_q = _mm_sub_epi16(sum_q, pq_16[3]);
+
+      work0 = _mm_add_epi16(
+          sum_p6, _mm_add_epi16(pq_16[3], _mm_add_epi16(pq_16[4], pq_16[2])));
+      flat2_p[3] = _mm_add_epi16(sum_p, work0);
+      flat2_q[3] = _mm_add_epi16(sum_q, _mm_srli_si128(work0, 8));
+      flat2_pq[3] =
+          _mm_srli_epi16(_mm_unpacklo_epi64(flat2_p[3], flat2_q[3]), 4);
+      flat2_pq[3] = _mm_packus_epi16(flat2_pq[3], flat2_pq[3]);
+
+      sum_p6 = _mm_add_epi16(sum_p6, pq_16[6]);
+      sum_p = _mm_sub_epi16(sum_p, q2_16);
+      sum_q = _mm_sub_epi16(sum_q, pq_16[2]);
+
+      work0 = _mm_add_epi16(
+          sum_p6, _mm_add_epi16(pq_16[4], _mm_add_epi16(pq_16[5], pq_16[3])));
+      flat2_p[4] = _mm_add_epi16(sum_p, work0);
+      flat2_q[4] = _mm_add_epi16(sum_q, _mm_srli_si128(work0, 8));
+      flat2_pq[4] =
+          _mm_srli_epi16(_mm_unpacklo_epi64(flat2_p[4], flat2_q[4]), 4);
+      flat2_pq[4] = _mm_packus_epi16(flat2_pq[4], flat2_pq[4]);
+
+      sum_p6 = _mm_add_epi16(sum_p6, pq_16[6]);
+      sum_p = _mm_sub_epi16(sum_p, q1_16);
+      sum_q = _mm_sub_epi16(sum_q, pq_16[1]);
+
+      work0 = _mm_add_epi16(
+          sum_p6, _mm_add_epi16(pq_16[5], _mm_add_epi16(pq_16[6], pq_16[4])));
+      flat2_p[5] = _mm_add_epi16(sum_p, work0);
+      flat2_q[5] = _mm_add_epi16(sum_q, _mm_srli_si128(work0, 8));
+      flat2_pq[5] =
+          _mm_srli_epi16(_mm_unpacklo_epi64(flat2_p[5], flat2_q[5]), 4);
+      flat2_pq[5] = _mm_packus_epi16(flat2_pq[5], flat2_pq[5]);
+
+      // wide flat
+      // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+      *q0p0 = _mm_andnot_si128(flat2, *q0p0);
+      flat2_pq[0] = _mm_and_si128(flat2, flat2_pq[0]);
+      *q0p0 = _mm_or_si128(*q0p0, flat2_pq[0]);
+
+      *q1p1 = _mm_andnot_si128(flat2, *q1p1);
+      flat2_pq[1] = _mm_and_si128(flat2, flat2_pq[1]);
+      *q1p1 = _mm_or_si128(*q1p1, flat2_pq[1]);
+
+      *q2p2 = _mm_andnot_si128(flat2, *q2p2);
+      flat2_pq[2] = _mm_and_si128(flat2, flat2_pq[2]);
+      *q2p2 = _mm_or_si128(*q2p2, flat2_pq[2]);
+
+      *q3p3 = _mm_andnot_si128(flat2, *q3p3);
+      flat2_pq[3] = _mm_and_si128(flat2, flat2_pq[3]);
+      *q3p3 = _mm_or_si128(*q3p3, flat2_pq[3]);
+
+      *q4p4 = _mm_andnot_si128(flat2, *q4p4);
+      flat2_pq[4] = _mm_and_si128(flat2, flat2_pq[4]);
+      *q4p4 = _mm_or_si128(*q4p4, flat2_pq[4]);
+
+      *q5p5 = _mm_andnot_si128(flat2, *q5p5);
+      flat2_pq[5] = _mm_and_si128(flat2, flat2_pq[5]);
+      *q5p5 = _mm_or_si128(*q5p5, flat2_pq[5]);
+    }
+  } else {
+    *q0p0 = qs0ps0;
+    *q1p1 = qs1ps1;
+  }
+}
+
+void aom_lpf_horizontal_14_sse2(unsigned char *s, int p,
+                                const unsigned char *_blimit,
+                                const unsigned char *_limit,
+                                const unsigned char *_thresh) {
+  __m128i q6p6, q5p5, q4p4, q3p3, q2p2, q1p1, q0p0;
+  __m128i blimit = _mm_load_si128((const __m128i *)_blimit);
+  __m128i limit = _mm_load_si128((const __m128i *)_limit);
+  __m128i thresh = _mm_load_si128((const __m128i *)_thresh);
+
+  q4p4 = _mm_unpacklo_epi32(xx_loadl_32(s - 5 * p), xx_loadl_32(s + 4 * p));
+  q3p3 = _mm_unpacklo_epi32(xx_loadl_32(s - 4 * p), xx_loadl_32(s + 3 * p));
+  q2p2 = _mm_unpacklo_epi32(xx_loadl_32(s - 3 * p), xx_loadl_32(s + 2 * p));
+  q1p1 = _mm_unpacklo_epi32(xx_loadl_32(s - 2 * p), xx_loadl_32(s + 1 * p));
+
+  q0p0 = _mm_unpacklo_epi32(xx_loadl_32(s - 1 * p), xx_loadl_32(s - 0 * p));
+
+  q5p5 = _mm_unpacklo_epi32(xx_loadl_32(s - 6 * p), xx_loadl_32(s + 5 * p));
+
+  q6p6 = _mm_unpacklo_epi32(xx_loadl_32(s - 7 * p), xx_loadl_32(s + 6 * p));
+
+  lpf_internal_14_sse2(&q6p6, &q5p5, &q4p4, &q3p3, &q2p2, &q1p1, &q0p0, &blimit,
+                       &limit, &thresh);
+
+  store_buffer_horz_8(q0p0, p, 0, s);
+  store_buffer_horz_8(q1p1, p, 1, s);
+  store_buffer_horz_8(q2p2, p, 2, s);
+  store_buffer_horz_8(q3p3, p, 3, s);
+  store_buffer_horz_8(q4p4, p, 4, s);
+  store_buffer_horz_8(q5p5, p, 5, s);
+}
+
+static AOM_FORCE_INLINE void lpf_internal_6_dual_sse2(
+    __m128i *p2, __m128i *q2, __m128i *p1, __m128i *q1, __m128i *p0,
+    __m128i *q0, __m128i *q1q0, __m128i *p1p0, __m128i *blimit, __m128i *limit,
+    __m128i *thresh) {
+  const __m128i zero = _mm_setzero_si128();
+  __m128i mask, hev, flat;
+  __m128i q2p2, q1p1, q0p0, flat_p1p0, flat_q0q1;
+  __m128i p2_16, q2_16, p1_16, q1_16, p0_16, q0_16;
+  __m128i ps1ps0, qs1qs0;
+
+  q2p2 = _mm_unpacklo_epi64(*p2, *q2);
+  q1p1 = _mm_unpacklo_epi64(*p1, *q1);
+  q0p0 = _mm_unpacklo_epi64(*p0, *q0);
+
+  *p1p0 = _mm_unpacklo_epi64(q0p0, q1p1);
+  *q1q0 = _mm_unpackhi_epi64(q0p0, q1p1);
+
+  const __m128i one = _mm_set1_epi8(1);
+  const __m128i fe = _mm_set1_epi8((char)0xfe);
+  const __m128i ff = _mm_cmpeq_epi8(fe, fe);
+
+  {
+    // filter_mask and hev_mask
+    __m128i abs_p1q1, abs_p0q0, abs_q1q0, abs_p1p0, work;
+    abs_p1p0 = abs_diff(q1p1, q0p0);
+    abs_q1q0 = _mm_srli_si128(abs_p1p0, 8);
+
+    abs_p0q0 = abs_diff(*p1p0, *q1q0);
+    abs_p1q1 = _mm_srli_si128(abs_p0q0, 8);
+    abs_p0q0 = _mm_unpacklo_epi64(abs_p0q0, zero);
+
+    // considering sse doesn't have unsigned elements comparison the idea is
+    // to find at least one case when X > limit, it means the corresponding
+    // mask bit is set.
+    // to achieve that we find global max value of all inputs of abs(x-y) or
+    // (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 If it is > limit the mask is set
+    // otherwise - not
+
+    flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
+    hev = _mm_subs_epu8(flat, *thresh);
+    hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+    // replicate for the further "merged variables" usage
+    hev = _mm_unpacklo_epi64(hev, hev);
+
+    abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0);
+    abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
+    mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), *blimit);
+    mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
+    // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit) * -1;
+    mask = _mm_max_epu8(abs_p1p0, mask);
+    // mask |= (abs(p1 - p0) > limit) * -1;
+    // mask |= (abs(q1 - q0) > limit) * -1;
+
+    work = abs_diff(q2p2, q1p1);
+    mask = _mm_max_epu8(work, mask);
+    mask = _mm_max_epu8(mask, _mm_srli_si128(mask, 8));
+    mask = _mm_subs_epu8(mask, *limit);
+    mask = _mm_cmpeq_epi8(mask, zero);
+
+    // lp filter - the same for 6, 8 and 14 versions
+    filter4_dual_sse2(p1p0, q1q0, &hev, &mask, q1q0, p1p0);
+
+    // flat_mask
+    flat = _mm_max_epu8(abs_diff(q2p2, q0p0), abs_p1p0);
+    flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 8));
+    flat = _mm_subs_epu8(flat, one);
+    flat = _mm_cmpeq_epi8(flat, zero);
+    flat = _mm_and_si128(flat, mask);
+    // replicate for the further "merged variables" usage
+    flat = _mm_unpacklo_epi64(flat, flat);
+  }
+
+  // 5 tap filter
+  // need it only if flat !=0
+  if (0xffff != _mm_movemask_epi8(_mm_cmpeq_epi8(flat, zero))) {
+    const __m128i four = _mm_set1_epi16(4);
+    __m128i workp_a, workp_b, workp_shft0, workp_shft1;
+    p2_16 = _mm_unpacklo_epi8(*p2, zero);
+    p1_16 = _mm_unpacklo_epi8(*p1, zero);
+    p0_16 = _mm_unpacklo_epi8(*p0, zero);
+    q0_16 = _mm_unpacklo_epi8(*q0, zero);
+    q1_16 = _mm_unpacklo_epi8(*q1, zero);
+    q2_16 = _mm_unpacklo_epi8(*q2, zero);
+
+    // op1
+    workp_a = _mm_add_epi16(_mm_add_epi16(p0_16, p0_16),
+                            _mm_add_epi16(p1_16, p1_16));  // p0 *2 + p1 * 2
+    workp_a = _mm_add_epi16(_mm_add_epi16(workp_a, four),
+                            p2_16);  // p2 + p0 * 2 + p1 * 2 + 4
+
+    workp_b = _mm_add_epi16(_mm_add_epi16(p2_16, p2_16), q0_16);
+    workp_shft0 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b),
+                                 3);  // p2 * 3 + p1 * 2 + p0 * 2 + q0 + 4
+
+    // op0
+    workp_b = _mm_add_epi16(_mm_add_epi16(q0_16, q0_16), q1_16);  // q0 * 2 + q1
+    workp_a = _mm_add_epi16(workp_a,
+                            workp_b);  // p2 + p0 * 2 + p1 * 2 + q0 * 2 + q1 + 4
+    workp_shft1 = _mm_srli_epi16(workp_a, 3);
+
+    flat_p1p0 = _mm_packus_epi16(workp_shft1, workp_shft0);
+
+    // oq0
+    workp_a = _mm_sub_epi16(_mm_sub_epi16(workp_a, p2_16),
+                            p1_16);  // p0 * 2 + p1  + q0 * 2 + q1 + 4
+    workp_b = _mm_add_epi16(q1_16, q2_16);
+    workp_a = _mm_add_epi16(
+        workp_a, workp_b);  // p0 * 2 + p1  + q0 * 2 + q1 * 2 + q2 + 4
+    workp_shft0 = _mm_srli_epi16(workp_a, 3);
+
+    // oq1
+    workp_a = _mm_sub_epi16(_mm_sub_epi16(workp_a, p1_16),
+                            p0_16);  // p0   + q0 * 2 + q1 * 2 + q2 + 4
+    workp_b = _mm_add_epi16(q2_16, q2_16);
+    workp_shft1 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b),
+                                 3);  // p0  + q0 * 2 + q1 * 2 + q2 * 3 + 4
+
+    flat_q0q1 = _mm_packus_epi16(workp_shft0, workp_shft1);
+
+    qs1qs0 = _mm_andnot_si128(flat, *q1q0);
+    *q1q0 = _mm_and_si128(flat, flat_q0q1);
+    *q1q0 = _mm_or_si128(qs1qs0, *q1q0);
+
+    ps1ps0 = _mm_andnot_si128(flat, *p1p0);
+    *p1p0 = _mm_and_si128(flat, flat_p1p0);
+    *p1p0 = _mm_or_si128(ps1ps0, *p1p0);
+  }
+}
+
+static AOM_FORCE_INLINE void lpf_internal_6_sse2(
+    __m128i *p2, __m128i *q2, __m128i *p1, __m128i *q1, __m128i *p0,
+    __m128i *q0, __m128i *q1q0, __m128i *p1p0, __m128i *blimit, __m128i *limit,
+    __m128i *thresh) {
+  const __m128i zero = _mm_setzero_si128();
+  __m128i mask, hev, flat;
+  __m128i q2p2, q1p1, q0p0, flat_p1p0, flat_q0q1;
+  __m128i pq2_16, q2_16, pq1_16, pq0_16, q0_16;
+  __m128i ps1ps0, qs1qs0;
+
+  q2p2 = _mm_unpacklo_epi32(*p2, *q2);
+  q1p1 = _mm_unpacklo_epi32(*p1, *q1);
+  q0p0 = _mm_unpacklo_epi32(*p0, *q0);
+
+  *p1p0 = _mm_unpacklo_epi32(*p0, *p1);
+  *q1q0 = _mm_unpacklo_epi32(*q0, *q1);
+
+  const __m128i one = _mm_set1_epi8(1);
+  const __m128i fe = _mm_set1_epi8((char)0xfe);
+  const __m128i ff = _mm_cmpeq_epi8(fe, fe);
+  {
+    // filter_mask and hev_mask
+    __m128i abs_p1q1, abs_p0q0, abs_q1q0, abs_p1p0, work;
+    abs_p1p0 = abs_diff(q1p1, q0p0);
+    abs_q1q0 = _mm_srli_si128(abs_p1p0, 4);
+
+    abs_p0q0 = abs_diff(*p1p0, *q1q0);
+    abs_p1q1 = _mm_srli_si128(abs_p0q0, 4);
+
+    // considering sse doesn't have unsigned elements comparison the idea is
+    // to find at least one case when X > limit, it means the corresponding
+    // mask bit is set.
+    // to achieve that we find global max value of all inputs of abs(x-y) or
+    // (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 If it is > limit the mask is set
+    // otherwise - not
+
+    flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
+    hev = _mm_subs_epu8(flat, *thresh);
+    hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+    // replicate for the further "merged variables" usage
+    hev = _mm_unpacklo_epi32(hev, hev);
+
+    abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0);
+    abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
+    mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), *blimit);
+    mask = _mm_unpacklo_epi32(mask, zero);
+    mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
+    // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit) * -1;
+    mask = _mm_max_epu8(abs_p1p0, mask);
+    // mask |= (abs(p1 - p0) > limit) * -1;
+    // mask |= (abs(q1 - q0) > limit) * -1;
+
+    work = abs_diff(q2p2, q1p1);
+    mask = _mm_max_epu8(work, mask);
+    mask = _mm_max_epu8(mask, _mm_srli_si128(mask, 4));
+    mask = _mm_subs_epu8(mask, *limit);
+    mask = _mm_cmpeq_epi8(mask, zero);
+
+    // lp filter - the same for 6, 8 and 14 versions
+    filter4_sse2(p1p0, q1q0, &hev, &mask, q1q0, p1p0);
+
+    // flat_mask
+    flat = _mm_max_epu8(abs_diff(q2p2, q0p0), abs_p1p0);
+    flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 4));
+    flat = _mm_subs_epu8(flat, one);
+    flat = _mm_cmpeq_epi8(flat, zero);
+    flat = _mm_and_si128(flat, mask);
+    // replicate for the further "merged variables" usage
+    flat = _mm_unpacklo_epi32(flat, flat);
+    flat = _mm_unpacklo_epi64(flat, flat);
+  }
+
+  // 5 tap filter
+  // need it only if flat !=0
+  if (0xffff != _mm_movemask_epi8(_mm_cmpeq_epi8(flat, zero))) {
+    const __m128i four = _mm_set1_epi16(4);
+    __m128i workp_a, workp_b, workp_c;
+    __m128i pq0x2_pq1, pq1_pq2;
+    pq2_16 = _mm_unpacklo_epi8(q2p2, zero);
+    pq1_16 = _mm_unpacklo_epi8(q1p1, zero);
+    pq0_16 = _mm_unpacklo_epi8(q0p0, zero);
+    q0_16 = _mm_srli_si128(pq0_16, 8);
+    q2_16 = _mm_srli_si128(pq2_16, 8);
+
+    // op1
+    pq0x2_pq1 =
+        _mm_add_epi16(_mm_add_epi16(pq0_16, pq0_16), pq1_16);  // p0 *2 + p1
+    pq1_pq2 = _mm_add_epi16(pq1_16, pq2_16);                   // p1 + p2
+    workp_a = _mm_add_epi16(_mm_add_epi16(pq0x2_pq1, four),
+                            pq1_pq2);  // p2 + p0 * 2 + p1 * 2 + 4
+
+    workp_b = _mm_add_epi16(_mm_add_epi16(pq2_16, pq2_16), q0_16);
+    workp_b =
+        _mm_add_epi16(workp_a, workp_b);  // p2 * 3 + p1 * 2 + p0 * 2 + q0 + 4
+
+    // op0
+    workp_c = _mm_srli_si128(pq0x2_pq1, 8);  // q0 * 2 + q1
+    workp_a = _mm_add_epi16(workp_a,
+                            workp_c);  // p2 + p0 * 2 + p1 * 2 + q0 * 2 + q1 + 4
+    workp_b = _mm_unpacklo_epi64(workp_a, workp_b);
+    workp_b = _mm_srli_epi16(workp_b, 3);
+
+    flat_p1p0 = _mm_packus_epi16(workp_b, workp_b);
+
+    // oq0
+    workp_a = _mm_sub_epi16(_mm_sub_epi16(workp_a, pq2_16),
+                            pq1_16);  // p0 * 2 + p1  + q0 * 2 + q1 + 4
+    workp_b = _mm_srli_si128(pq1_pq2, 8);
+    workp_a = _mm_add_epi16(
+        workp_a, workp_b);  // p0 * 2 + p1  + q0 * 2 + q1 * 2 + q2 + 4
+    // workp_shft0 = _mm_srli_epi16(workp_a, 3);
+
+    // oq1
+    workp_c = _mm_sub_epi16(_mm_sub_epi16(workp_a, pq1_16),
+                            pq0_16);  // p0   + q0 * 2 + q1 * 2 + q2 + 4
+    workp_b = _mm_add_epi16(q2_16, q2_16);
+    workp_b =
+        _mm_add_epi16(workp_c, workp_b);  // p0  + q0 * 2 + q1 * 2 + q2 * 3 + 4
+
+    workp_a = _mm_unpacklo_epi64(workp_a, workp_b);
+    workp_a = _mm_srli_epi16(workp_a, 3);
+
+    flat_q0q1 = _mm_packus_epi16(workp_a, workp_a);
+
+    qs1qs0 = _mm_andnot_si128(flat, *q1q0);
+    *q1q0 = _mm_and_si128(flat, flat_q0q1);
+    *q1q0 = _mm_or_si128(qs1qs0, *q1q0);
+
+    ps1ps0 = _mm_andnot_si128(flat, *p1p0);
+    *p1p0 = _mm_and_si128(flat, flat_p1p0);
+    *p1p0 = _mm_or_si128(ps1ps0, *p1p0);
+  }
+}
+
+void aom_lpf_horizontal_6_sse2(unsigned char *s, int p,
+                               const unsigned char *_blimit,
+                               const unsigned char *_limit,
+                               const unsigned char *_thresh) {
+  __m128i p2, p1, p0, q0, q1, q2;
+  __m128i p1p0, q1q0;
+  __m128i blimit = _mm_load_si128((__m128i *)_blimit);
+  __m128i limit = _mm_load_si128((__m128i *)_limit);
+  __m128i thresh = _mm_load_si128((__m128i *)_thresh);
+
+  p2 = xx_loadl_32(s - 3 * p);
+  p1 = xx_loadl_32(s - 2 * p);
+  p0 = xx_loadl_32(s - 1 * p);
+  q0 = xx_loadl_32(s - 0 * p);
+  q1 = xx_loadl_32(s + 1 * p);
+  q2 = xx_loadl_32(s + 2 * p);
+
+  lpf_internal_6_sse2(&p2, &q2, &p1, &q1, &p0, &q0, &q1q0, &p1p0, &blimit,
+                      &limit, &thresh);
+
+  xx_storel_32(s - 1 * p, p1p0);
+  xx_storel_32(s - 2 * p, _mm_srli_si128(p1p0, 4));
+  xx_storel_32(s + 0 * p, q1q0);
+  xx_storel_32(s + 1 * p, _mm_srli_si128(q1q0, 4));
+}
+
+void aom_lpf_horizontal_6_dual_sse2(unsigned char *s, int p,
+                                    const unsigned char *_blimit0,
+                                    const unsigned char *_limit0,
+                                    const unsigned char *_thresh0,
+                                    const unsigned char *_blimit1,
+                                    const unsigned char *_limit1,
+                                    const unsigned char *_thresh1) {
+  __m128i blimit = _mm_unpacklo_epi32(_mm_load_si128((__m128i *)_blimit0),
+                                      _mm_load_si128((__m128i *)_blimit1));
+  __m128i limit = _mm_unpacklo_epi32(_mm_load_si128((__m128i *)_limit0),
+                                     _mm_load_si128((__m128i *)_limit1));
+  __m128i thresh = _mm_unpacklo_epi32(_mm_load_si128((__m128i *)_thresh0),
+                                      _mm_load_si128((__m128i *)_thresh1));
+
+  __m128i p2, p1, p0, q0, q1, q2;
+  __m128i p1p0, q1q0;
+
+  p2 = _mm_loadl_epi64((__m128i *)(s - 3 * p));
+  p1 = _mm_loadl_epi64((__m128i *)(s - 2 * p));
+  p0 = _mm_loadl_epi64((__m128i *)(s - 1 * p));
+  q0 = _mm_loadl_epi64((__m128i *)(s - 0 * p));
+  q1 = _mm_loadl_epi64((__m128i *)(s + 1 * p));
+  q2 = _mm_loadl_epi64((__m128i *)(s + 2 * p));
+
+  lpf_internal_6_dual_sse2(&p2, &q2, &p1, &q1, &p0, &q0, &q1q0, &p1p0, &blimit,
+                           &limit, &thresh);
+
+  _mm_storel_epi64((__m128i *)(s - 1 * p), p1p0);
+  _mm_storel_epi64((__m128i *)(s - 2 * p), _mm_srli_si128(p1p0, 8));
+  _mm_storel_epi64((__m128i *)(s + 0 * p), q1q0);
+  _mm_storel_epi64((__m128i *)(s + 1 * p), _mm_srli_si128(q1q0, 8));
+}
+
+static AOM_FORCE_INLINE void lpf_internal_8_sse2(
+    __m128i *p3, __m128i *q3, __m128i *p2, __m128i *q2, __m128i *p1,
+    __m128i *q1, __m128i *p0, __m128i *q0, __m128i *q1q0_out, __m128i *p1p0_out,
+    __m128i *blimit, __m128i *limit, __m128i *thresh) {
+  const __m128i zero = _mm_setzero_si128();
+  __m128i mask, hev, flat;
+  __m128i p2_16, q2_16, p1_16, p0_16, q0_16, q1_16, p3_16, q3_16, q3p3,
+      flat_p1p0, flat_q0q1;
+  __m128i q2p2, q1p1, q0p0;
+  __m128i q1q0, p1p0, ps1ps0, qs1qs0;
+  __m128i work_pq, opq2, pq2;
+
+  q3p3 = _mm_unpacklo_epi32(*p3, *q3);
+  q2p2 = _mm_unpacklo_epi32(*p2, *q2);
+  q1p1 = _mm_unpacklo_epi32(*p1, *q1);
+  q0p0 = _mm_unpacklo_epi32(*p0, *q0);
+
+  p1p0 = _mm_unpacklo_epi32(q0p0, q1p1);  // p1p0 q1q0
+  q1q0 = _mm_srli_si128(p1p0, 8);
+
+  // filter_mask and hev_mask
+
+  // considering sse doesn't have unsigned elements comparison the idea is to
+  // find at least one case when X > limit, it means the corresponding  mask
+  // bit is set.
+  // to achieve that we find global max value of all inputs of abs(x-y) or
+  // (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 If it is > limit the mask is set
+  // otherwise - not
+
+  const __m128i one = _mm_set1_epi8(1);
+  const __m128i fe = _mm_set1_epi8((char)0xfe);
+  const __m128i ff = _mm_cmpeq_epi8(fe, fe);
+  __m128i abs_p1q1, abs_p0q0, abs_q1q0, abs_p1p0, work;
+
+  abs_p1p0 = abs_diff(q1p1, q0p0);
+  abs_q1q0 = _mm_srli_si128(abs_p1p0, 4);
+
+  abs_p0q0 = abs_diff(p1p0, q1q0);
+  abs_p1q1 = _mm_srli_si128(abs_p0q0, 4);
+
+  flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
+  hev = _mm_subs_epu8(flat, *thresh);
+  hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+  // replicate for the further "merged variables" usage
+  hev = _mm_unpacklo_epi32(hev, hev);
+
+  abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0);
+  abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
+  mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), *blimit);
+  mask = _mm_unpacklo_epi32(mask, zero);
+  mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
+  // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit) * -1;
+  mask = _mm_max_epu8(abs_p1p0, mask);
+  // mask |= (abs(p1 - p0) > limit) * -1;
+  // mask |= (abs(q1 - q0) > limit) * -1;
+
+  work = _mm_max_epu8(abs_diff(q2p2, q1p1), abs_diff(q3p3, q2p2));
+
+  mask = _mm_max_epu8(work, mask);
+  mask = _mm_max_epu8(mask, _mm_srli_si128(mask, 4));
+  mask = _mm_subs_epu8(mask, *limit);
+  mask = _mm_cmpeq_epi8(mask, zero);
+
+  // lp filter - the same for 6, 8 and 14 versions
+  filter4_sse2(&p1p0, &q1q0, &hev, &mask, q1q0_out, p1p0_out);
+
+  // flat_mask4
+  flat = _mm_max_epu8(abs_diff(q2p2, q0p0), abs_diff(q3p3, q0p0));
+  flat = _mm_max_epu8(abs_p1p0, flat);
+
+  flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 4));
+  flat = _mm_subs_epu8(flat, one);
+  flat = _mm_cmpeq_epi8(flat, zero);
+  flat = _mm_and_si128(flat, mask);
+  // replicate for the further "merged variables" usage
+  flat = _mm_unpacklo_epi32(flat, flat);
+  flat = _mm_unpacklo_epi64(flat, flat);
+
+  // filter8 need it only if flat !=0
+  if (0xffff != _mm_movemask_epi8(_mm_cmpeq_epi8(flat, zero))) {
+    const __m128i four = _mm_set1_epi16(4);
+    __m128i workp_a, workp_b, workp_c, workp_d, workp_shft1, workp_shft2;
+    p2_16 = _mm_unpacklo_epi8(*p2, zero);
+    p1_16 = _mm_unpacklo_epi8(*p1, zero);
+    p0_16 = _mm_unpacklo_epi8(*p0, zero);
+    q0_16 = _mm_unpacklo_epi8(*q0, zero);
+    q1_16 = _mm_unpacklo_epi8(*q1, zero);
+    q2_16 = _mm_unpacklo_epi8(*q2, zero);
+    p3_16 = _mm_unpacklo_epi8(*p3, zero);
+    q3_16 = _mm_unpacklo_epi8(*q3, zero);
+
+    // op2
+    workp_a =
+        _mm_add_epi16(_mm_add_epi16(p3_16, p3_16), _mm_add_epi16(p2_16, p1_16));
+    workp_a = _mm_add_epi16(_mm_add_epi16(workp_a, four), p0_16);
+    workp_b = _mm_add_epi16(_mm_add_epi16(q0_16, p2_16), p3_16);
+    workp_shft2 = _mm_add_epi16(workp_a, workp_b);
+
+    // op1
+    workp_b = _mm_add_epi16(_mm_add_epi16(q0_16, q1_16), p1_16);
+    workp_c = _mm_add_epi16(workp_a, workp_b);
+    // workp_shft0 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+
+    // op0
+    workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3_16), q2_16);
+    workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p1_16), p0_16);
+    workp_d = _mm_add_epi16(workp_a, workp_b);
+    // workp_shft1 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+
+    workp_c = _mm_unpacklo_epi64(workp_d, workp_c);
+    workp_c = _mm_srli_epi16(workp_c, 3);
+    flat_p1p0 = _mm_packus_epi16(workp_c, workp_c);
+
+    // oq0
+    workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3_16), q3_16);
+    workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p0_16), q0_16);
+    // workp_shft0 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+    workp_c = _mm_add_epi16(workp_a, workp_b);
+
+    // oq1
+    workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p2_16), q3_16);
+    workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q0_16), q1_16);
+    workp_d = _mm_add_epi16(workp_a, workp_b);
+    // workp_shft1 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+
+    workp_c = _mm_unpacklo_epi64(workp_c, workp_d);
+    workp_c = _mm_srli_epi16(workp_c, 3);
+    flat_q0q1 = _mm_packus_epi16(workp_c, workp_c);
+
+    // oq2
+    workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p1_16), q3_16);
+    workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q1_16), q2_16);
+    workp_shft1 = _mm_add_epi16(workp_a, workp_b);
+
+    workp_c = _mm_unpacklo_epi64(workp_shft2, workp_shft1);
+    workp_c = _mm_srli_epi16(workp_c, 3);
+
+    opq2 = _mm_packus_epi16(workp_c, workp_c);
+
+    work_pq = _mm_andnot_si128(flat, q2p2);
+    pq2 = _mm_and_si128(flat, opq2);
+    *p2 = _mm_or_si128(work_pq, pq2);
+    *q2 = _mm_srli_si128(*p2, 4);
+
+    qs1qs0 = _mm_andnot_si128(flat, *q1q0_out);
+    q1q0 = _mm_and_si128(flat, flat_q0q1);
+    *q1q0_out = _mm_or_si128(qs1qs0, q1q0);
+
+    ps1ps0 = _mm_andnot_si128(flat, *p1p0_out);
+    p1p0 = _mm_and_si128(flat, flat_p1p0);
+    *p1p0_out = _mm_or_si128(ps1ps0, p1p0);
+  }
+}
+
+static AOM_FORCE_INLINE void lpf_internal_8_dual_sse2(
+    __m128i *p3, __m128i *q3, __m128i *p2, __m128i *q2, __m128i *p1,
+    __m128i *q1, __m128i *p0, __m128i *q0, __m128i *q1q0_out, __m128i *p1p0_out,
+    __m128i *blimit, __m128i *limit, __m128i *thresh) {
+  const __m128i zero = _mm_setzero_si128();
+  __m128i mask, hev, flat;
+  __m128i p2_16, q2_16, p1_16, p0_16, q0_16, q1_16, p3_16, q3_16, q3p3,
+      flat_p1p0, flat_q0q1;
+  __m128i q2p2, q1p1, q0p0;
+  __m128i q1q0, p1p0, ps1ps0, qs1qs0;
+  __m128i work_pq, opq2, pq2;
+
+  q3p3 = _mm_unpacklo_epi64(*p3, *q3);
+  q2p2 = _mm_unpacklo_epi64(*p2, *q2);
+  q1p1 = _mm_unpacklo_epi64(*p1, *q1);
+  q0p0 = _mm_unpacklo_epi64(*p0, *q0);
+
+  p1p0 = _mm_unpacklo_epi64(q0p0, q1p1);
+  q1q0 = _mm_unpackhi_epi64(q0p0, q1p1);
+
+  {
+    // filter_mask and hev_mask
+
+    // considering sse doesn't have unsigned elements comparison the idea is to
+    // find at least one case when X > limit, it means the corresponding  mask
+    // bit is set.
+    // to achieve that we find global max value of all inputs of abs(x-y) or
+    // (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 If it is > limit the mask is set
+    // otherwise - not
+
+    const __m128i one = _mm_set1_epi8(1);
+    const __m128i fe = _mm_set1_epi8((char)0xfe);
+    const __m128i ff = _mm_cmpeq_epi8(fe, fe);
+    __m128i abs_p1q1, abs_p0q0, abs_q1q0, abs_p1p0, work;
+
+    abs_p1p0 = abs_diff(q1p1, q0p0);
+    abs_q1q0 = _mm_srli_si128(abs_p1p0, 8);
+
+    abs_p0q0 = abs_diff(p1p0, q1q0);
+    abs_p1q1 = _mm_srli_si128(abs_p0q0, 8);
+    abs_p0q0 = _mm_unpacklo_epi64(abs_p0q0, abs_p0q0);
+
+    flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
+    hev = _mm_subs_epu8(flat, *thresh);
+    hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+    // replicate for the further "merged variables" usage
+    hev = _mm_unpacklo_epi64(hev, hev);
+
+    abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0);
+    abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
+    mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), *blimit);
+    mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
+    // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit) * -1;
+    mask = _mm_max_epu8(abs_p1p0, mask);
+    // mask |= (abs(p1 - p0) > limit) * -1;
+    // mask |= (abs(q1 - q0) > limit) * -1;
+
+    work = _mm_max_epu8(abs_diff(q2p2, q1p1), abs_diff(q3p3, q2p2));
+
+    mask = _mm_max_epu8(work, mask);
+    mask = _mm_max_epu8(mask, _mm_srli_si128(mask, 8));
+    mask = _mm_subs_epu8(mask, *limit);
+    mask = _mm_cmpeq_epi8(mask, zero);
+
+    // lp filter - the same for 6, 8 and 14 versions
+    filter4_dual_sse2(&p1p0, &q1q0, &hev, &mask, q1q0_out, p1p0_out);
+
+    // flat_mask4
+    flat = _mm_max_epu8(abs_diff(q2p2, q0p0), abs_diff(q3p3, q0p0));
+    flat = _mm_max_epu8(abs_p1p0, flat);
+
+    flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 8));
+    flat = _mm_subs_epu8(flat, one);
+    flat = _mm_cmpeq_epi8(flat, zero);
+    flat = _mm_and_si128(flat, mask);
+    // replicate for the further "merged variables" usage
+    flat = _mm_unpacklo_epi64(flat, flat);
+  }
+
+  // filter8 need it only if flat !=0
+  if (0xffff != _mm_movemask_epi8(_mm_cmpeq_epi8(flat, zero))) {
+    const __m128i four = _mm_set1_epi16(4);
+
+    __m128i workp_a, workp_b, workp_shft0, workp_shft1, workp_shft2;
+    p2_16 = _mm_unpacklo_epi8(*p2, zero);
+    p1_16 = _mm_unpacklo_epi8(*p1, zero);
+    p0_16 = _mm_unpacklo_epi8(*p0, zero);
+    q0_16 = _mm_unpacklo_epi8(*q0, zero);
+    q1_16 = _mm_unpacklo_epi8(*q1, zero);
+    q2_16 = _mm_unpacklo_epi8(*q2, zero);
+    p3_16 = _mm_unpacklo_epi8(*p3, zero);
+    q3_16 = _mm_unpacklo_epi8(*q3, zero);
+
+    // op2
+    workp_a =
+        _mm_add_epi16(_mm_add_epi16(p3_16, p3_16), _mm_add_epi16(p2_16, p1_16));
+    workp_a = _mm_add_epi16(_mm_add_epi16(workp_a, four), p0_16);
+    workp_b = _mm_add_epi16(_mm_add_epi16(q0_16, p2_16), p3_16);
+    workp_shft2 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+
+    // op1
+    workp_b = _mm_add_epi16(_mm_add_epi16(q0_16, q1_16), p1_16);
+    workp_shft0 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+
+    // op0
+    workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3_16), q2_16);
+    workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p1_16), p0_16);
+    workp_shft1 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+
+    flat_p1p0 = _mm_packus_epi16(workp_shft1, workp_shft0);
+
+    // oq0
+    workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p3_16), q3_16);
+    workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, p0_16), q0_16);
+    workp_shft0 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+
+    // oq1
+    workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p2_16), q3_16);
+    workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q0_16), q1_16);
+    workp_shft1 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+
+    flat_q0q1 = _mm_packus_epi16(workp_shft0, workp_shft1);
+
+    // oq2
+    workp_a = _mm_add_epi16(_mm_sub_epi16(workp_a, p1_16), q3_16);
+    workp_b = _mm_add_epi16(_mm_sub_epi16(workp_b, q1_16), q2_16);
+    workp_shft1 = _mm_srli_epi16(_mm_add_epi16(workp_a, workp_b), 3);
+
+    opq2 = _mm_packus_epi16(workp_shft2, workp_shft1);
+
+    work_pq = _mm_andnot_si128(flat, q2p2);
+    pq2 = _mm_and_si128(flat, opq2);
+    *p2 = _mm_or_si128(work_pq, pq2);
+    *q2 = _mm_srli_si128(*p2, 8);
+
+    qs1qs0 = _mm_andnot_si128(flat, *q1q0_out);
+    q1q0 = _mm_and_si128(flat, flat_q0q1);
+    *q1q0_out = _mm_or_si128(qs1qs0, q1q0);
+
+    ps1ps0 = _mm_andnot_si128(flat, *p1p0_out);
+    p1p0 = _mm_and_si128(flat, flat_p1p0);
+    *p1p0_out = _mm_or_si128(ps1ps0, p1p0);
+  }
+}
+
+void aom_lpf_horizontal_8_sse2(unsigned char *s, int p,
+                               const unsigned char *_blimit,
+                               const unsigned char *_limit,
+                               const unsigned char *_thresh) {
+  __m128i p3, p2, p1, p0, q0, q1, q2, q3;
+  __m128i q1q0, p1p0;
+  __m128i blimit = _mm_load_si128((const __m128i *)_blimit);
+  __m128i limit = _mm_load_si128((const __m128i *)_limit);
+  __m128i thresh = _mm_load_si128((const __m128i *)_thresh);
+
+  p3 = xx_loadl_32(s - 4 * p);
+  p2 = xx_loadl_32(s - 3 * p);
+  p1 = xx_loadl_32(s - 2 * p);
+  p0 = xx_loadl_32(s - 1 * p);
+  q0 = xx_loadl_32(s - 0 * p);
+  q1 = xx_loadl_32(s + 1 * p);
+  q2 = xx_loadl_32(s + 2 * p);
+  q3 = xx_loadl_32(s + 3 * p);
+
+  lpf_internal_8_sse2(&p3, &q3, &p2, &q2, &p1, &q1, &p0, &q0, &q1q0, &p1p0,
+                      &blimit, &limit, &thresh);
+
+  xx_storel_32(s - 1 * p, p1p0);
+  xx_storel_32(s - 2 * p, _mm_srli_si128(p1p0, 4));
+  xx_storel_32(s + 0 * p, q1q0);
+  xx_storel_32(s + 1 * p, _mm_srli_si128(q1q0, 4));
+  xx_storel_32(s - 3 * p, p2);
+  xx_storel_32(s + 2 * p, q2);
+}
+
+void aom_lpf_horizontal_14_dual_sse2(unsigned char *s, int p,
+                                     const unsigned char *_blimit0,
+                                     const unsigned char *_limit0,
+                                     const unsigned char *_thresh0,
+                                     const unsigned char *_blimit1,
+                                     const unsigned char *_limit1,
+                                     const unsigned char *_thresh1) {
+  __m128i q6p6, q5p5, q4p4, q3p3, q2p2, q1p1, q0p0;
+  __m128i blimit =
+      _mm_unpacklo_epi32(_mm_load_si128((const __m128i *)_blimit0),
+                         _mm_load_si128((const __m128i *)_blimit1));
+  __m128i limit = _mm_unpacklo_epi32(_mm_load_si128((const __m128i *)_limit0),
+                                     _mm_load_si128((const __m128i *)_limit1));
+  __m128i thresh =
+      _mm_unpacklo_epi32(_mm_load_si128((const __m128i *)_thresh0),
+                         _mm_load_si128((const __m128i *)_thresh1));
+
+  q4p4 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 5 * p)),
+                            _mm_loadl_epi64((__m128i *)(s + 4 * p)));
+  q3p3 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 4 * p)),
+                            _mm_loadl_epi64((__m128i *)(s + 3 * p)));
+  q2p2 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 3 * p)),
+                            _mm_loadl_epi64((__m128i *)(s + 2 * p)));
+  q1p1 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 2 * p)),
+                            _mm_loadl_epi64((__m128i *)(s + 1 * p)));
+
+  q0p0 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 1 * p)),
+                            _mm_loadl_epi64((__m128i *)(s - 0 * p)));
+
+  q5p5 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 6 * p)),
+                            _mm_loadl_epi64((__m128i *)(s + 5 * p)));
+
+  q6p6 = _mm_unpacklo_epi64(_mm_loadl_epi64((__m128i *)(s - 7 * p)),
+                            _mm_loadl_epi64((__m128i *)(s + 6 * p)));
+
+  lpf_internal_14_dual_sse2(&q6p6, &q5p5, &q4p4, &q3p3, &q2p2, &q1p1, &q0p0,
+                            &blimit, &limit, &thresh);
+
+  _mm_storel_epi64((__m128i *)(s - 1 * p), q0p0);
+  _mm_storel_epi64((__m128i *)(s + 0 * p), _mm_srli_si128(q0p0, 8));
+  _mm_storel_epi64((__m128i *)(s - 2 * p), q1p1);
+  _mm_storel_epi64((__m128i *)(s + 1 * p), _mm_srli_si128(q1p1, 8));
+  _mm_storel_epi64((__m128i *)(s - 3 * p), q2p2);
+  _mm_storel_epi64((__m128i *)(s + 2 * p), _mm_srli_si128(q2p2, 8));
+  _mm_storel_epi64((__m128i *)(s - 4 * p), q3p3);
+  _mm_storel_epi64((__m128i *)(s + 3 * p), _mm_srli_si128(q3p3, 8));
+  _mm_storel_epi64((__m128i *)(s - 5 * p), q4p4);
+  _mm_storel_epi64((__m128i *)(s + 4 * p), _mm_srli_si128(q4p4, 8));
+  _mm_storel_epi64((__m128i *)(s - 6 * p), q5p5);
+  _mm_storel_epi64((__m128i *)(s + 5 * p), _mm_srli_si128(q5p5, 8));
+}
+
+void aom_lpf_horizontal_8_dual_sse2(uint8_t *s, int p, const uint8_t *_blimit0,
+                                    const uint8_t *_limit0,
+                                    const uint8_t *_thresh0,
+                                    const uint8_t *_blimit1,
+                                    const uint8_t *_limit1,
+                                    const uint8_t *_thresh1) {
+  __m128i blimit = _mm_unpacklo_epi32(_mm_load_si128((__m128i *)_blimit0),
+                                      _mm_load_si128((__m128i *)_blimit1));
+  __m128i limit = _mm_unpacklo_epi32(_mm_load_si128((__m128i *)_limit0),
+                                     _mm_load_si128((__m128i *)_limit1));
+  __m128i thresh = _mm_unpacklo_epi32(_mm_load_si128((__m128i *)_thresh0),
+                                      _mm_load_si128((__m128i *)_thresh1));
+
+  __m128i p2, p1, p0, q0, q1, q2, p3, q3;
+  __m128i q1q0, p1p0;
+
+  p3 = _mm_loadl_epi64((__m128i *)(s - 4 * p));
+  p2 = _mm_loadl_epi64((__m128i *)(s - 3 * p));
+  p1 = _mm_loadl_epi64((__m128i *)(s - 2 * p));
+  p0 = _mm_loadl_epi64((__m128i *)(s - 1 * p));
+  q0 = _mm_loadl_epi64((__m128i *)(s - 0 * p));
+  q1 = _mm_loadl_epi64((__m128i *)(s + 1 * p));
+  q2 = _mm_loadl_epi64((__m128i *)(s + 2 * p));
+  q3 = _mm_loadl_epi64((__m128i *)(s + 3 * p));
+
+  lpf_internal_8_dual_sse2(&p3, &q3, &p2, &q2, &p1, &q1, &p0, &q0, &q1q0, &p1p0,
+                           &blimit, &limit, &thresh);
+
+  _mm_storel_epi64((__m128i *)(s - 1 * p), p1p0);
+  _mm_storel_epi64((__m128i *)(s - 2 * p), _mm_srli_si128(p1p0, 8));
+  _mm_storel_epi64((__m128i *)(s + 0 * p), q1q0);
+  _mm_storel_epi64((__m128i *)(s + 1 * p), _mm_srli_si128(q1q0, 8));
+  _mm_storel_epi64((__m128i *)(s - 3 * p), p2);
+  _mm_storel_epi64((__m128i *)(s + 2 * p), q2);
+}
+
+void aom_lpf_horizontal_4_dual_sse2(unsigned char *s, int p,
+                                    const unsigned char *_blimit0,
+                                    const unsigned char *_limit0,
+                                    const unsigned char *_thresh0,
+                                    const unsigned char *_blimit1,
+                                    const unsigned char *_limit1,
+                                    const unsigned char *_thresh1) {
+  __m128i p1, p0, q0, q1;
+  __m128i qs1qs0, ps1ps0;
+
+  p1 = _mm_loadl_epi64((__m128i *)(s - 2 * p));
+  p0 = _mm_loadl_epi64((__m128i *)(s - 1 * p));
+  q0 = _mm_loadl_epi64((__m128i *)(s - 0 * p));
+  q1 = _mm_loadl_epi64((__m128i *)(s + 1 * p));
+
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i blimit =
+      _mm_unpacklo_epi32(_mm_load_si128((const __m128i *)_blimit0),
+                         _mm_load_si128((const __m128i *)_blimit1));
+  const __m128i limit =
+      _mm_unpacklo_epi32(_mm_load_si128((const __m128i *)_limit0),
+                         _mm_load_si128((const __m128i *)_limit1));
+
+  __m128i l = _mm_unpacklo_epi64(blimit, limit);
+
+  __m128i thresh0 =
+      _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)_thresh0), zero);
+
+  __m128i thresh1 =
+      _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)_thresh1), zero);
+
+  __m128i t = _mm_unpacklo_epi64(thresh0, thresh1);
+
+  lpf_internal_4_dual_sse2(&p1, &p0, &q0, &q1, &l, &t, &qs1qs0, &ps1ps0);
+
+  _mm_storel_epi64((__m128i *)(s - 1 * p), ps1ps0);
+  _mm_storel_epi64((__m128i *)(s - 2 * p), _mm_srli_si128(ps1ps0, 8));
+  _mm_storel_epi64((__m128i *)(s + 0 * p), qs1qs0);
+  _mm_storel_epi64((__m128i *)(s + 1 * p), _mm_srli_si128(qs1qs0, 8));
+}
+
+void aom_lpf_vertical_4_dual_sse2(uint8_t *s, int p, const uint8_t *_blimit0,
+                                  const uint8_t *_limit0,
+                                  const uint8_t *_thresh0,
+                                  const uint8_t *_blimit1,
+                                  const uint8_t *_limit1,
+                                  const uint8_t *_thresh1) {
+  __m128i p0, q0, q1, p1;
+  __m128i x0, x1, x2, x3, x4, x5, x6, x7;
+  __m128i d0, d1, d2, d3, d4, d5, d6, d7;
+  __m128i qs1qs0, ps1ps0;
+
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i blimit =
+      _mm_unpacklo_epi32(_mm_load_si128((const __m128i *)_blimit0),
+                         _mm_load_si128((const __m128i *)_blimit1));
+  const __m128i limit =
+      _mm_unpacklo_epi32(_mm_load_si128((const __m128i *)_limit0),
+                         _mm_load_si128((const __m128i *)_limit1));
+
+  __m128i l = _mm_unpacklo_epi64(blimit, limit);
+
+  __m128i thresh0 =
+      _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)_thresh0), zero);
+
+  __m128i thresh1 =
+      _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)_thresh1), zero);
+
+  __m128i t = _mm_unpacklo_epi64(thresh0, thresh1);
+
+  x0 = _mm_loadl_epi64((__m128i *)((s - 2)));
+  x1 = _mm_loadl_epi64((__m128i *)((s - 2) + p));
+  x2 = _mm_loadl_epi64((__m128i *)((s - 2) + 2 * p));
+  x3 = _mm_loadl_epi64((__m128i *)((s - 2) + 3 * p));
+  x4 = _mm_loadl_epi64((__m128i *)((s - 2) + 4 * p));
+  x5 = _mm_loadl_epi64((__m128i *)((s - 2) + 5 * p));
+  x6 = _mm_loadl_epi64((__m128i *)((s - 2) + 6 * p));
+  x7 = _mm_loadl_epi64((__m128i *)((s - 2) + 7 * p));
+
+  transpose8x8_low_sse2(&x0, &x1, &x2, &x3, &x4, &x5, &x6, &x7, &p1, &p0, &q0,
+                        &q1);
+
+  lpf_internal_4_dual_sse2(&p1, &p0, &q0, &q1, &l, &t, &qs1qs0, &ps1ps0);
+
+  p1 = _mm_srli_si128(ps1ps0, 8);
+  q1 = _mm_srli_si128(qs1qs0, 8);
+
+  transpose4x8_8x4_sse2(&p1, &ps1ps0, &qs1qs0, &q1, &d0, &d1, &d2, &d3, &d4,
+                        &d5, &d6, &d7);
+
+  xx_storel_32((s - 2 + 0 * p), d0);
+  xx_storel_32((s - 2 + 1 * p), d1);
+  xx_storel_32((s - 2 + 2 * p), d2);
+  xx_storel_32((s - 2 + 3 * p), d3);
+  xx_storel_32((s - 2 + 4 * p), d4);
+  xx_storel_32((s - 2 + 5 * p), d5);
+  xx_storel_32((s - 2 + 6 * p), d6);
+  xx_storel_32((s - 2 + 7 * p), d7);
+}
+
+void aom_lpf_vertical_6_sse2(unsigned char *s, int p,
+                             const unsigned char *_blimit,
+                             const unsigned char *_limit,
+                             const unsigned char *_thresh) {
+  __m128i d0, d1, d2, d3, d4, d5, d6, d7;
+  __m128i x2, x1, x0, x3;
+  __m128i p0, q0;
+  __m128i p1p0, q1q0;
+  __m128i blimit = _mm_load_si128((__m128i *)_blimit);
+  __m128i limit = _mm_load_si128((__m128i *)_limit);
+  __m128i thresh = _mm_load_si128((__m128i *)_thresh);
+
+  x3 = _mm_loadl_epi64((__m128i *)((s - 3) + 0 * p));
+  x2 = _mm_loadl_epi64((__m128i *)((s - 3) + 1 * p));
+  x1 = _mm_loadl_epi64((__m128i *)((s - 3) + 2 * p));
+  x0 = _mm_loadl_epi64((__m128i *)((s - 3) + 3 * p));
+
+  transpose4x8_8x4_sse2(&x3, &x2, &x1, &x0, &d0, &d1, &d2, &d3, &d4, &d5, &d6,
+                        &d7);
+
+  lpf_internal_6_sse2(&d0, &d5, &d1, &d4, &d2, &d3, &q1q0, &p1p0, &blimit,
+                      &limit, &thresh);
+
+  p0 = _mm_srli_si128(p1p0, 4);
+  q0 = _mm_srli_si128(q1q0, 4);
+
+  transpose4x8_8x4_low_sse2(&p0, &p1p0, &q1q0, &q0, &d0, &d1, &d2, &d3);
+
+  xx_storel_32(s + 0 * p - 2, d0);
+  xx_storel_32(s + 1 * p - 2, d1);
+  xx_storel_32(s + 2 * p - 2, d2);
+  xx_storel_32(s + 3 * p - 2, d3);
+}
+
+void aom_lpf_vertical_6_dual_sse2(uint8_t *s, int p, const uint8_t *_blimit0,
+                                  const uint8_t *_limit0,
+                                  const uint8_t *_thresh0,
+                                  const uint8_t *_blimit1,
+                                  const uint8_t *_limit1,
+                                  const uint8_t *_thresh1) {
+  __m128i blimit = _mm_unpacklo_epi32(_mm_load_si128((__m128i *)_blimit0),
+                                      _mm_load_si128((__m128i *)_blimit1));
+  __m128i limit = _mm_unpacklo_epi32(_mm_load_si128((__m128i *)_limit0),
+                                     _mm_load_si128((__m128i *)_limit1));
+  __m128i thresh = _mm_unpacklo_epi32(_mm_load_si128((__m128i *)_thresh0),
+                                      _mm_load_si128((__m128i *)_thresh1));
+
+  __m128i d0, d1, d2, d3, d4, d5, d6, d7;
+  __m128i x0, x1, x2, x3, x4, x5, x6, x7;
+  __m128i p0, q0;
+  __m128i p1p0, q1q0;
+  __m128i d0d1, d2d3, d4d5, d6d7;
+
+  x0 = _mm_loadl_epi64((__m128i *)((s - 3) + 0 * p));
+  x1 = _mm_loadl_epi64((__m128i *)((s - 3) + 1 * p));
+  x2 = _mm_loadl_epi64((__m128i *)((s - 3) + 2 * p));
+  x3 = _mm_loadl_epi64((__m128i *)((s - 3) + 3 * p));
+  x4 = _mm_loadl_epi64((__m128i *)((s - 3) + 4 * p));
+  x5 = _mm_loadl_epi64((__m128i *)((s - 3) + 5 * p));
+  x6 = _mm_loadl_epi64((__m128i *)((s - 3) + 6 * p));
+  x7 = _mm_loadl_epi64((__m128i *)((s - 3) + 7 * p));
+
+  transpose8x8_sse2(&x0, &x1, &x2, &x3, &x4, &x5, &x6, &x7, &d0d1, &d2d3, &d4d5,
+                    &d6d7);
+
+  d1 = _mm_srli_si128(d0d1, 8);
+  d3 = _mm_srli_si128(d2d3, 8);
+  d5 = _mm_srli_si128(d4d5, 8);
+  d7 = _mm_srli_si128(d6d7, 8);
+
+  lpf_internal_6_dual_sse2(&d0d1, &d5, &d1, &d4d5, &d2d3, &d3, &q1q0, &p1p0,
+                           &blimit, &limit, &thresh);
+
+  p0 = _mm_srli_si128(p1p0, 8);
+  q0 = _mm_srli_si128(q1q0, 8);
+
+  transpose4x8_8x4_sse2(&p0, &p1p0, &q1q0, &q0, &d0, &d1, &d2, &d3, &d4, &d5,
+                        &d6, &d7);
+
+  xx_storel_32((s - 2 + 0 * p), d0);
+  xx_storel_32((s - 2 + 1 * p), d1);
+  xx_storel_32((s - 2 + 2 * p), d2);
+  xx_storel_32((s - 2 + 3 * p), d3);
+  xx_storel_32((s - 2 + 4 * p), d4);
+  xx_storel_32((s - 2 + 5 * p), d5);
+  xx_storel_32((s - 2 + 6 * p), d6);
+  xx_storel_32((s - 2 + 7 * p), d7);
+}
+
+void aom_lpf_vertical_8_sse2(unsigned char *s, int p,
+                             const unsigned char *_blimit,
+                             const unsigned char *_limit,
+                             const unsigned char *_thresh) {
+  __m128i d0, d1, d2, d3, d4, d5, d6, d7;
+
+  __m128i p0, q0;
+  __m128i x2, x1, x0, x3;
+  __m128i q1q0, p1p0;
+  __m128i blimit = _mm_load_si128((const __m128i *)_blimit);
+  __m128i limit = _mm_load_si128((const __m128i *)_limit);
+  __m128i thresh = _mm_load_si128((const __m128i *)_thresh);
+
+  x3 = _mm_loadl_epi64((__m128i *)((s - 4) + 0 * p));
+  x2 = _mm_loadl_epi64((__m128i *)((s - 4) + 1 * p));
+  x1 = _mm_loadl_epi64((__m128i *)((s - 4) + 2 * p));
+  x0 = _mm_loadl_epi64((__m128i *)((s - 4) + 3 * p));
+
+  transpose4x8_8x4_sse2(&x3, &x2, &x1, &x0, &d0, &d1, &d2, &d3, &d4, &d5, &d6,
+                        &d7);
+  // Loop filtering
+  lpf_internal_8_sse2(&d0, &d7, &d1, &d6, &d2, &d5, &d3, &d4, &q1q0, &p1p0,
+                      &blimit, &limit, &thresh);
+
+  p0 = _mm_srli_si128(p1p0, 4);
+  q0 = _mm_srli_si128(q1q0, 4);
+
+  transpose8x8_low_sse2(&d0, &d1, &p0, &p1p0, &q1q0, &q0, &d6, &d7, &d0, &d1,
+                        &d2, &d3);
+
+  _mm_storel_epi64((__m128i *)(s - 4 + 0 * p), d0);
+  _mm_storel_epi64((__m128i *)(s - 4 + 1 * p), d1);
+  _mm_storel_epi64((__m128i *)(s - 4 + 2 * p), d2);
+  _mm_storel_epi64((__m128i *)(s - 4 + 3 * p), d3);
+}
+
+void aom_lpf_vertical_8_dual_sse2(uint8_t *s, int p, const uint8_t *_blimit0,
+                                  const uint8_t *_limit0,
+                                  const uint8_t *_thresh0,
+                                  const uint8_t *_blimit1,
+                                  const uint8_t *_limit1,
+                                  const uint8_t *_thresh1) {
+  __m128i blimit = _mm_unpacklo_epi32(_mm_load_si128((__m128i *)_blimit0),
+                                      _mm_load_si128((__m128i *)_blimit1));
+  __m128i limit = _mm_unpacklo_epi32(_mm_load_si128((__m128i *)_limit0),
+                                     _mm_load_si128((__m128i *)_limit1));
+  __m128i thresh = _mm_unpacklo_epi32(_mm_load_si128((__m128i *)_thresh0),
+                                      _mm_load_si128((__m128i *)_thresh1));
+
+  __m128i x0, x1, x2, x3, x4, x5, x6, x7;
+  __m128i d1, d3, d5, d7;
+  __m128i q1q0, p1p0;
+  __m128i p1, q1;
+  __m128i d0d1, d2d3, d4d5, d6d7;
+
+  x0 = _mm_loadl_epi64((__m128i *)(s - 4 + 0 * p));
+  x1 = _mm_loadl_epi64((__m128i *)(s - 4 + 1 * p));
+  x2 = _mm_loadl_epi64((__m128i *)(s - 4 + 2 * p));
+  x3 = _mm_loadl_epi64((__m128i *)(s - 4 + 3 * p));
+  x4 = _mm_loadl_epi64((__m128i *)(s - 4 + 4 * p));
+  x5 = _mm_loadl_epi64((__m128i *)(s - 4 + 5 * p));
+  x6 = _mm_loadl_epi64((__m128i *)(s - 4 + 6 * p));
+  x7 = _mm_loadl_epi64((__m128i *)(s - 4 + 7 * p));
+
+  transpose8x8_sse2(&x0, &x1, &x2, &x3, &x4, &x5, &x6, &x7, &d0d1, &d2d3, &d4d5,
+                    &d6d7);
+
+  d1 = _mm_srli_si128(d0d1, 8);
+  d3 = _mm_srli_si128(d2d3, 8);
+  d5 = _mm_srli_si128(d4d5, 8);
+  d7 = _mm_srli_si128(d6d7, 8);
+
+  lpf_internal_8_dual_sse2(&d0d1, &d7, &d1, &d6d7, &d2d3, &d5, &d3, &d4d5,
+                           &q1q0, &p1p0, &blimit, &limit, &thresh);
+
+  p1 = _mm_srli_si128(p1p0, 8);
+  q1 = _mm_srli_si128(q1q0, 8);
+
+  transpose8x8_sse2(&d0d1, &d1, &p1, &p1p0, &q1q0, &q1, &d6d7, &d7, &d0d1,
+                    &d2d3, &d4d5, &d6d7);
+
+  _mm_storel_epi64((__m128i *)(s - 4 + 0 * p), d0d1);
+  _mm_storel_epi64((__m128i *)(s - 4 + 1 * p), _mm_srli_si128(d0d1, 8));
+  _mm_storel_epi64((__m128i *)(s - 4 + 2 * p), d2d3);
+  _mm_storel_epi64((__m128i *)(s - 4 + 3 * p), _mm_srli_si128(d2d3, 8));
+  _mm_storel_epi64((__m128i *)(s - 4 + 4 * p), d4d5);
+  _mm_storel_epi64((__m128i *)(s - 4 + 5 * p), _mm_srli_si128(d4d5, 8));
+  _mm_storel_epi64((__m128i *)(s - 4 + 6 * p), d6d7);
+  _mm_storel_epi64((__m128i *)(s - 4 + 7 * p), _mm_srli_si128(d6d7, 8));
+}
+
+void aom_lpf_vertical_14_sse2(unsigned char *s, int p,
+                              const unsigned char *_blimit,
+                              const unsigned char *_limit,
+                              const unsigned char *_thresh) {
+  __m128i q7p7, q6p6, q5p5, q4p4, q3p3, q2p2, q1p1, q0p0;
+  __m128i x6, x5, x4, x3;
+  __m128i pq0, pq1, pq2, pq3;
+  __m128i blimit = _mm_load_si128((__m128i *)_blimit);
+  __m128i limit = _mm_load_si128((__m128i *)_limit);
+  __m128i thresh = _mm_load_si128((__m128i *)_thresh);
+
+  x6 = _mm_loadu_si128((__m128i *)((s - 8) + 0 * p));
+  x5 = _mm_loadu_si128((__m128i *)((s - 8) + 1 * p));
+  x4 = _mm_loadu_si128((__m128i *)((s - 8) + 2 * p));
+  x3 = _mm_loadu_si128((__m128i *)((s - 8) + 3 * p));
+
+  transpose_pq_14_sse2(&x6, &x5, &x4, &x3, &q0p0, &q1p1, &q2p2, &q3p3, &q4p4,
+                       &q5p5, &q6p6, &q7p7);
+
+  lpf_internal_14_sse2(&q6p6, &q5p5, &q4p4, &q3p3, &q2p2, &q1p1, &q0p0, &blimit,
+                       &limit, &thresh);
+
+  transpose_pq_14_inv_sse2(&q7p7, &q6p6, &q5p5, &q4p4, &q3p3, &q2p2, &q1p1,
+                           &q0p0, &pq0, &pq1, &pq2, &pq3);
+  _mm_storeu_si128((__m128i *)(s - 8 + 0 * p), pq0);
+  _mm_storeu_si128((__m128i *)(s - 8 + 1 * p), pq1);
+  _mm_storeu_si128((__m128i *)(s - 8 + 2 * p), pq2);
+  _mm_storeu_si128((__m128i *)(s - 8 + 3 * p), pq3);
+}
+
+void aom_lpf_vertical_14_dual_sse2(
+    unsigned char *s, int p, const uint8_t *_blimit0, const uint8_t *_limit0,
+    const uint8_t *_thresh0, const uint8_t *_blimit1, const uint8_t *_limit1,
+    const uint8_t *_thresh1) {
+  __m128i q6p6, q5p5, q4p4, q3p3, q2p2, q1p1, q0p0;
+  __m128i x7, x6, x5, x4, x3, x2, x1, x0;
+  __m128i d0d1, d2d3, d4d5, d6d7, d8d9, d10d11, d12d13, d14d15;
+  __m128i q0, q1, q2, q3, q7;
+  __m128i p0p1, p2p3, p4p5, p6p7;
+
+  __m128i blimit =
+      _mm_unpacklo_epi32(_mm_load_si128((const __m128i *)_blimit0),
+                         _mm_load_si128((const __m128i *)_blimit1));
+  __m128i limit = _mm_unpacklo_epi32(_mm_load_si128((const __m128i *)_limit0),
+                                     _mm_load_si128((const __m128i *)_limit1));
+  __m128i thresh =
+      _mm_unpacklo_epi32(_mm_load_si128((const __m128i *)_thresh0),
+                         _mm_load_si128((const __m128i *)_thresh1));
+
+  x7 = _mm_loadu_si128((__m128i *)((s - 8) + 0 * p));
+  x6 = _mm_loadu_si128((__m128i *)((s - 8) + 1 * p));
+  x5 = _mm_loadu_si128((__m128i *)((s - 8) + 2 * p));
+  x4 = _mm_loadu_si128((__m128i *)((s - 8) + 3 * p));
+  x3 = _mm_loadu_si128((__m128i *)((s - 8) + 4 * p));
+  x2 = _mm_loadu_si128((__m128i *)((s - 8) + 5 * p));
+  x1 = _mm_loadu_si128((__m128i *)((s - 8) + 6 * p));
+  x0 = _mm_loadu_si128((__m128i *)((s - 8) + 7 * p));
+
+  transpose8x16_16x8_sse2(&x7, &x6, &x5, &x4, &x3, &x2, &x1, &x0, &d0d1, &d2d3,
+                          &d4d5, &d6d7, &d8d9, &d10d11, &d12d13, &d14d15);
+
+  q6p6 = _mm_unpacklo_epi64(d2d3, _mm_srli_si128(d12d13, 8));
+  q5p5 = _mm_unpacklo_epi64(d4d5, _mm_srli_si128(d10d11, 8));
+  q4p4 = _mm_unpacklo_epi64(d6d7, _mm_srli_si128(d8d9, 8));
+  q3p3 = _mm_unpacklo_epi64(d8d9, _mm_srli_si128(d6d7, 8));
+  q2p2 = _mm_unpacklo_epi64(d10d11, _mm_srli_si128(d4d5, 8));
+  q1p1 = _mm_unpacklo_epi64(d12d13, _mm_srli_si128(d2d3, 8));
+  q0p0 = _mm_unpacklo_epi64(d14d15, _mm_srli_si128(d0d1, 8));
+  q7 = _mm_srli_si128(d14d15, 8);
+
+  lpf_internal_14_dual_sse2(&q6p6, &q5p5, &q4p4, &q3p3, &q2p2, &q1p1, &q0p0,
+                            &blimit, &limit, &thresh);
+
+  x0 = _mm_srli_si128(q0p0, 8);
+  x1 = _mm_srli_si128(q1p1, 8);
+  x2 = _mm_srli_si128(q2p2, 8);
+  x3 = _mm_srli_si128(q3p3, 8);
+  x4 = _mm_srli_si128(q4p4, 8);
+  x5 = _mm_srli_si128(q5p5, 8);
+  x6 = _mm_srli_si128(q6p6, 8);
+
+  transpose16x8_8x16_sse2(&d0d1, &q6p6, &q5p5, &q4p4, &q3p3, &q2p2, &q1p1,
+                          &q0p0, &x0, &x1, &x2, &x3, &x4, &x5, &x6, &q7, &p0p1,
+                          &p2p3, &p4p5, &p6p7, &q0, &q1, &q2, &q3);
+
+  _mm_storeu_si128((__m128i *)(s - 8 + 0 * p), p0p1);
+  _mm_storeu_si128((__m128i *)(s - 8 + 1 * p), p2p3);
+  _mm_storeu_si128((__m128i *)(s - 8 + 2 * p), p4p5);
+  _mm_storeu_si128((__m128i *)(s - 8 + 3 * p), p6p7);
+  _mm_storeu_si128((__m128i *)(s - 8 + 4 * p), q0);
+  _mm_storeu_si128((__m128i *)(s - 8 + 5 * p), q1);
+  _mm_storeu_si128((__m128i *)(s - 8 + 6 * p), q2);
+  _mm_storeu_si128((__m128i *)(s - 8 + 7 * p), q3);
+}
+
+static INLINE __m128i filter_add2_sub2(const __m128i *const total,
+                                       const __m128i *const a1,
+                                       const __m128i *const a2,
+                                       const __m128i *const s1,
+                                       const __m128i *const s2) {
+  __m128i x = _mm_add_epi16(*a1, *total);
+  x = _mm_add_epi16(_mm_sub_epi16(x, _mm_add_epi16(*s1, *s2)), *a2);
+  return x;
+}
+
+static INLINE __m128i filter8_mask(const __m128i *const flat,
+                                   const __m128i *const other_filt,
+                                   const __m128i *const f8_lo,
+                                   const __m128i *const f8_hi) {
+  const __m128i f8 =
+      _mm_packus_epi16(_mm_srli_epi16(*f8_lo, 3), _mm_srli_epi16(*f8_hi, 3));
+  const __m128i result = _mm_and_si128(*flat, f8);
+  return _mm_or_si128(_mm_andnot_si128(*flat, *other_filt), result);
+}
+
+static INLINE __m128i filter16_mask(const __m128i *const flat,
+                                    const __m128i *const other_filt,
+                                    const __m128i *const f_lo,
+                                    const __m128i *const f_hi) {
+  const __m128i f =
+      _mm_packus_epi16(_mm_srli_epi16(*f_lo, 4), _mm_srli_epi16(*f_hi, 4));
+  const __m128i result = _mm_and_si128(*flat, f);
+  return _mm_or_si128(_mm_andnot_si128(*flat, *other_filt), result);
+}
+
+void aom_lpf_horizontal_14_quad_sse2(unsigned char *s, int p,
+                                     const unsigned char *_blimit0,
+                                     const unsigned char *_limit0,
+                                     const unsigned char *_thresh0) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i one = _mm_set1_epi8(1);
+  const __m128i blimit_v = _mm_load_si128((const __m128i *)_blimit0);
+  const __m128i limit_v = _mm_load_si128((const __m128i *)_limit0);
+  const __m128i thresh_v = _mm_load_si128((const __m128i *)_thresh0);
+  __m128i mask, hev, flat, flat2;
+  __m128i p6, p5;
+  __m128i p4, p3, p2, p1, p0, q0, q1, q2, q3, q4;
+  __m128i q6, q5;
+
+  __m128i op2, op1, op0, oq0, oq1, oq2;
+
+  __m128i max_abs_p1p0q1q0;
+
+  p6 = _mm_loadu_si128((__m128i *)(s - 7 * p));
+  p5 = _mm_loadu_si128((__m128i *)(s - 6 * p));
+  p4 = _mm_loadu_si128((__m128i *)(s - 5 * p));
+  p3 = _mm_loadu_si128((__m128i *)(s - 4 * p));
+  p2 = _mm_loadu_si128((__m128i *)(s - 3 * p));
+  p1 = _mm_loadu_si128((__m128i *)(s - 2 * p));
+  p0 = _mm_loadu_si128((__m128i *)(s - 1 * p));
+  q0 = _mm_loadu_si128((__m128i *)(s - 0 * p));
+  q1 = _mm_loadu_si128((__m128i *)(s + 1 * p));
+  q2 = _mm_loadu_si128((__m128i *)(s + 2 * p));
+  q3 = _mm_loadu_si128((__m128i *)(s + 3 * p));
+  q4 = _mm_loadu_si128((__m128i *)(s + 4 * p));
+  q5 = _mm_loadu_si128((__m128i *)(s + 5 * p));
+  q6 = _mm_loadu_si128((__m128i *)(s + 6 * p));
+
+  {
+    const __m128i abs_p1p0 = abs_diff(p1, p0);
+    const __m128i abs_q1q0 = abs_diff(q1, q0);
+    const __m128i fe = _mm_set1_epi8((int8_t)0xfe);
+    const __m128i ff = _mm_cmpeq_epi8(zero, zero);
+    __m128i abs_p0q0 = abs_diff(p0, q0);
+    __m128i abs_p1q1 = abs_diff(p1, q1);
+    __m128i work;
+    max_abs_p1p0q1q0 = _mm_max_epu8(abs_p1p0, abs_q1q0);
+
+    abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0);
+    abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
+    mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit_v);
+    mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
+    // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit) * -1;
+    mask = _mm_max_epu8(max_abs_p1p0q1q0, mask);
+    // mask |= (abs(p1 - p0) > limit) * -1;
+    // mask |= (abs(q1 - q0) > limit) * -1;
+    work = _mm_max_epu8(abs_diff(p2, p1), abs_diff(p3, p2));
+    mask = _mm_max_epu8(work, mask);
+    work = _mm_max_epu8(abs_diff(q2, q1), abs_diff(q3, q2));
+    mask = _mm_max_epu8(work, mask);
+    mask = _mm_subs_epu8(mask, limit_v);
+    mask = _mm_cmpeq_epi8(mask, zero);
+  }
+
+  if (0xffff == _mm_movemask_epi8(_mm_cmpeq_epi8(mask, zero))) return;
+
+  {
+    __m128i work;
+    work = _mm_max_epu8(abs_diff(p2, p0), abs_diff(q2, q0));
+    flat = _mm_max_epu8(work, max_abs_p1p0q1q0);
+    work = _mm_max_epu8(abs_diff(p3, p0), abs_diff(q3, q0));
+    flat = _mm_max_epu8(work, flat);
+    work = _mm_max_epu8(abs_diff(p4, p0), abs_diff(q4, q0));
+    flat = _mm_subs_epu8(flat, one);
+    flat = _mm_cmpeq_epi8(flat, zero);
+    flat = _mm_and_si128(flat, mask);
+    flat2 = _mm_max_epu8(abs_diff(p5, p0), abs_diff(q5, q0));
+    flat2 = _mm_max_epu8(work, flat2);
+    work = _mm_max_epu8(abs_diff(p6, p0), abs_diff(q6, q0));
+    flat2 = _mm_max_epu8(work, flat2);
+    flat2 = _mm_subs_epu8(flat2, one);
+    flat2 = _mm_cmpeq_epi8(flat2, zero);
+    flat2 = _mm_and_si128(flat2, flat);  // flat2 & flat & mask
+  }
+
+  // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  // filter4
+  {
+    const __m128i t4 = _mm_set1_epi8(4);
+    const __m128i t3 = _mm_set1_epi8(3);
+    const __m128i t80 = _mm_set1_epi8((int8_t)0x80);
+    const __m128i te0 = _mm_set1_epi8((int8_t)0xe0);
+    const __m128i t1f = _mm_set1_epi8(0x1f);
+    const __m128i t1 = _mm_set1_epi8(0x1);
+    const __m128i t7f = _mm_set1_epi8(0x7f);
+    const __m128i ff = _mm_cmpeq_epi8(t4, t4);
+
+    __m128i filt;
+    __m128i work_a;
+    __m128i filter1, filter2;
+
+    op1 = _mm_xor_si128(p1, t80);
+    op0 = _mm_xor_si128(p0, t80);
+    oq0 = _mm_xor_si128(q0, t80);
+    oq1 = _mm_xor_si128(q1, t80);
+
+    hev = _mm_subs_epu8(max_abs_p1p0q1q0, thresh_v);
+    hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+    filt = _mm_and_si128(_mm_subs_epi8(op1, oq1), hev);
+
+    work_a = _mm_subs_epi8(oq0, op0);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_and_si128(filt, mask);
+    filter1 = _mm_adds_epi8(filt, t4);
+    filter2 = _mm_adds_epi8(filt, t3);
+
+    work_a = _mm_cmpgt_epi8(zero, filter1);
+    filter1 = _mm_srli_epi16(filter1, 3);
+    work_a = _mm_and_si128(work_a, te0);
+    filter1 = _mm_and_si128(filter1, t1f);
+    filter1 = _mm_or_si128(filter1, work_a);
+    oq0 = _mm_xor_si128(_mm_subs_epi8(oq0, filter1), t80);
+
+    work_a = _mm_cmpgt_epi8(zero, filter2);
+    filter2 = _mm_srli_epi16(filter2, 3);
+    work_a = _mm_and_si128(work_a, te0);
+    filter2 = _mm_and_si128(filter2, t1f);
+    filter2 = _mm_or_si128(filter2, work_a);
+    op0 = _mm_xor_si128(_mm_adds_epi8(op0, filter2), t80);
+
+    filt = _mm_adds_epi8(filter1, t1);
+    work_a = _mm_cmpgt_epi8(zero, filt);
+    filt = _mm_srli_epi16(filt, 1);
+    work_a = _mm_and_si128(work_a, t80);
+    filt = _mm_and_si128(filt, t7f);
+    filt = _mm_or_si128(filt, work_a);
+    filt = _mm_andnot_si128(hev, filt);
+    op1 = _mm_xor_si128(_mm_adds_epi8(op1, filt), t80);
+    oq1 = _mm_xor_si128(_mm_subs_epi8(oq1, filt), t80);
+
+    // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+    // filter8
+    if (0xffff != _mm_movemask_epi8(_mm_cmpeq_epi8(flat, zero))) {
+      const __m128i four = _mm_set1_epi16(4);
+      const __m128i p3_lo = _mm_unpacklo_epi8(p3, zero);
+      const __m128i p2_lo = _mm_unpacklo_epi8(p2, zero);
+      const __m128i p1_lo = _mm_unpacklo_epi8(p1, zero);
+      const __m128i p0_lo = _mm_unpacklo_epi8(p0, zero);
+      const __m128i q0_lo = _mm_unpacklo_epi8(q0, zero);
+      const __m128i q1_lo = _mm_unpacklo_epi8(q1, zero);
+      const __m128i q2_lo = _mm_unpacklo_epi8(q2, zero);
+      const __m128i q3_lo = _mm_unpacklo_epi8(q3, zero);
+
+      const __m128i p3_hi = _mm_unpackhi_epi8(p3, zero);
+      const __m128i p2_hi = _mm_unpackhi_epi8(p2, zero);
+      const __m128i p1_hi = _mm_unpackhi_epi8(p1, zero);
+      const __m128i p0_hi = _mm_unpackhi_epi8(p0, zero);
+      const __m128i q0_hi = _mm_unpackhi_epi8(q0, zero);
+      const __m128i q1_hi = _mm_unpackhi_epi8(q1, zero);
+      const __m128i q2_hi = _mm_unpackhi_epi8(q2, zero);
+      const __m128i q3_hi = _mm_unpackhi_epi8(q3, zero);
+      __m128i f8_lo, f8_hi;
+
+      f8_lo = _mm_add_epi16(_mm_add_epi16(p3_lo, four),
+                            _mm_add_epi16(p3_lo, p2_lo));
+      f8_lo = _mm_add_epi16(_mm_add_epi16(p3_lo, f8_lo),
+                            _mm_add_epi16(p2_lo, p1_lo));
+      f8_lo = _mm_add_epi16(_mm_add_epi16(p0_lo, q0_lo), f8_lo);
+
+      f8_hi = _mm_add_epi16(_mm_add_epi16(p3_hi, four),
+                            _mm_add_epi16(p3_hi, p2_hi));
+      f8_hi = _mm_add_epi16(_mm_add_epi16(p3_hi, f8_hi),
+                            _mm_add_epi16(p2_hi, p1_hi));
+      f8_hi = _mm_add_epi16(_mm_add_epi16(p0_hi, q0_hi), f8_hi);
+
+      op2 = filter8_mask(&flat, &p2, &f8_lo, &f8_hi);
+
+      f8_lo = filter_add2_sub2(&f8_lo, &q1_lo, &p1_lo, &p2_lo, &p3_lo);
+      f8_hi = filter_add2_sub2(&f8_hi, &q1_hi, &p1_hi, &p2_hi, &p3_hi);
+      op1 = filter8_mask(&flat, &op1, &f8_lo, &f8_hi);
+
+      f8_lo = filter_add2_sub2(&f8_lo, &q2_lo, &p0_lo, &p1_lo, &p3_lo);
+      f8_hi = filter_add2_sub2(&f8_hi, &q2_hi, &p0_hi, &p1_hi, &p3_hi);
+      op0 = filter8_mask(&flat, &op0, &f8_lo, &f8_hi);
+
+      f8_lo = filter_add2_sub2(&f8_lo, &q3_lo, &q0_lo, &p0_lo, &p3_lo);
+      f8_hi = filter_add2_sub2(&f8_hi, &q3_hi, &q0_hi, &p0_hi, &p3_hi);
+      oq0 = filter8_mask(&flat, &oq0, &f8_lo, &f8_hi);
+
+      f8_lo = filter_add2_sub2(&f8_lo, &q3_lo, &q1_lo, &q0_lo, &p2_lo);
+      f8_hi = filter_add2_sub2(&f8_hi, &q3_hi, &q1_hi, &q0_hi, &p2_hi);
+      oq1 = filter8_mask(&flat, &oq1, &f8_lo, &f8_hi);
+
+      f8_lo = filter_add2_sub2(&f8_lo, &q3_lo, &q2_lo, &q1_lo, &p1_lo);
+      f8_hi = filter_add2_sub2(&f8_hi, &q3_hi, &q2_hi, &q1_hi, &p1_hi);
+      oq2 = filter8_mask(&flat, &q2, &f8_lo, &f8_hi);
+
+      // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+      // wide flat calculations
+      if (0xffff != _mm_movemask_epi8(_mm_cmpeq_epi8(flat2, zero))) {
+        const __m128i eight = _mm_set1_epi16(8);
+        const __m128i p6_lo = _mm_unpacklo_epi8(p6, zero);
+        const __m128i p5_lo = _mm_unpacklo_epi8(p5, zero);
+        const __m128i p4_lo = _mm_unpacklo_epi8(p4, zero);
+        const __m128i q4_lo = _mm_unpacklo_epi8(q4, zero);
+        const __m128i q5_lo = _mm_unpacklo_epi8(q5, zero);
+        const __m128i q6_lo = _mm_unpacklo_epi8(q6, zero);
+
+        const __m128i p6_hi = _mm_unpackhi_epi8(p6, zero);
+        const __m128i p5_hi = _mm_unpackhi_epi8(p5, zero);
+        const __m128i p4_hi = _mm_unpackhi_epi8(p4, zero);
+        const __m128i q4_hi = _mm_unpackhi_epi8(q4, zero);
+        const __m128i q5_hi = _mm_unpackhi_epi8(q5, zero);
+        const __m128i q6_hi = _mm_unpackhi_epi8(q6, zero);
+
+        __m128i f_lo;
+        __m128i f_hi;
+
+        f_lo = _mm_sub_epi16(_mm_slli_epi16(p6_lo, 3), p6_lo);
+        f_lo = _mm_add_epi16(_mm_slli_epi16(p5_lo, 1), f_lo);
+        f_lo = _mm_add_epi16(_mm_slli_epi16(p4_lo, 1), f_lo);
+        f_lo = _mm_add_epi16(_mm_add_epi16(p3_lo, f_lo),
+                             _mm_add_epi16(p2_lo, p1_lo));
+        f_lo = _mm_add_epi16(_mm_add_epi16(p0_lo, q0_lo), f_lo);
+        f_lo = _mm_add_epi16(f_lo, eight);
+
+        f_hi = _mm_sub_epi16(_mm_slli_epi16(p6_hi, 3), p6_hi);
+        f_hi = _mm_add_epi16(_mm_slli_epi16(p5_hi, 1), f_hi);
+        f_hi = _mm_add_epi16(_mm_slli_epi16(p4_hi, 1), f_hi);
+        f_hi = _mm_add_epi16(_mm_add_epi16(p3_hi, f_hi),
+                             _mm_add_epi16(p2_hi, p1_hi));
+        f_hi = _mm_add_epi16(_mm_add_epi16(p0_hi, q0_hi), f_hi);
+        f_hi = _mm_add_epi16(f_hi, eight);
+
+        p5 = filter16_mask(&flat2, &p5, &f_lo, &f_hi);
+        _mm_storeu_si128((__m128i *)(s - 6 * p), p5);
+
+        f_lo = filter_add2_sub2(&f_lo, &q1_lo, &p3_lo, &p6_lo, &p6_lo);
+        f_hi = filter_add2_sub2(&f_hi, &q1_hi, &p3_hi, &p6_hi, &p6_hi);
+        p4 = filter16_mask(&flat2, &p4, &f_lo, &f_hi);
+        _mm_storeu_si128((__m128i *)(s - 5 * p), p4);
+
+        f_lo = filter_add2_sub2(&f_lo, &q2_lo, &p2_lo, &p6_lo, &p5_lo);
+        f_hi = filter_add2_sub2(&f_hi, &q2_hi, &p2_hi, &p6_hi, &p5_hi);
+        p3 = filter16_mask(&flat2, &p3, &f_lo, &f_hi);
+        _mm_storeu_si128((__m128i *)(s - 4 * p), p3);
+
+        f_lo = filter_add2_sub2(&f_lo, &q3_lo, &p1_lo, &p6_lo, &p4_lo);
+        f_hi = filter_add2_sub2(&f_hi, &q3_hi, &p1_hi, &p6_hi, &p4_hi);
+        op2 = filter16_mask(&flat2, &op2, &f_lo, &f_hi);
+        _mm_storeu_si128((__m128i *)(s - 3 * p), op2);
+
+        f_lo = filter_add2_sub2(&f_lo, &q4_lo, &p0_lo, &p6_lo, &p3_lo);
+        f_hi = filter_add2_sub2(&f_hi, &q4_hi, &p0_hi, &p6_hi, &p3_hi);
+        op1 = filter16_mask(&flat2, &op1, &f_lo, &f_hi);
+        _mm_storeu_si128((__m128i *)(s - 2 * p), op1);
+
+        f_lo = filter_add2_sub2(&f_lo, &q5_lo, &q0_lo, &p6_lo, &p2_lo);
+        f_hi = filter_add2_sub2(&f_hi, &q5_hi, &q0_hi, &p6_hi, &p2_hi);
+        op0 = filter16_mask(&flat2, &op0, &f_lo, &f_hi);
+        _mm_storeu_si128((__m128i *)(s - 1 * p), op0);
+
+        f_lo = filter_add2_sub2(&f_lo, &q6_lo, &q1_lo, &p6_lo, &p1_lo);
+        f_hi = filter_add2_sub2(&f_hi, &q6_hi, &q1_hi, &p6_hi, &p1_hi);
+        oq0 = filter16_mask(&flat2, &oq0, &f_lo, &f_hi);
+        _mm_storeu_si128((__m128i *)(s - 0 * p), oq0);
+
+        f_lo = filter_add2_sub2(&f_lo, &q6_lo, &q2_lo, &p5_lo, &p0_lo);
+        f_hi = filter_add2_sub2(&f_hi, &q6_hi, &q2_hi, &p5_hi, &p0_hi);
+        oq1 = filter16_mask(&flat2, &oq1, &f_lo, &f_hi);
+        _mm_storeu_si128((__m128i *)(s + 1 * p), oq1);
+
+        f_lo = filter_add2_sub2(&f_lo, &q6_lo, &q3_lo, &p4_lo, &q0_lo);
+        f_hi = filter_add2_sub2(&f_hi, &q6_hi, &q3_hi, &p4_hi, &q0_hi);
+        oq2 = filter16_mask(&flat2, &oq2, &f_lo, &f_hi);
+        _mm_storeu_si128((__m128i *)(s + 2 * p), oq2);
+
+        f_lo = filter_add2_sub2(&f_lo, &q6_lo, &q4_lo, &p3_lo, &q1_lo);
+        f_hi = filter_add2_sub2(&f_hi, &q6_hi, &q4_hi, &p3_hi, &q1_hi);
+        q3 = filter16_mask(&flat2, &q3, &f_lo, &f_hi);
+        _mm_storeu_si128((__m128i *)(s + 3 * p), q3);
+
+        f_lo = filter_add2_sub2(&f_lo, &q6_lo, &q5_lo, &p2_lo, &q2_lo);
+        f_hi = filter_add2_sub2(&f_hi, &q6_hi, &q5_hi, &p2_hi, &q2_hi);
+        q4 = filter16_mask(&flat2, &q4, &f_lo, &f_hi);
+        _mm_storeu_si128((__m128i *)(s + 4 * p), q4);
+
+        f_lo = filter_add2_sub2(&f_lo, &q6_lo, &q6_lo, &p1_lo, &q3_lo);
+        f_hi = filter_add2_sub2(&f_hi, &q6_hi, &q6_hi, &p1_hi, &q3_hi);
+        q5 = filter16_mask(&flat2, &q5, &f_lo, &f_hi);
+        _mm_storeu_si128((__m128i *)(s + 5 * p), q5);
+      } else {
+        _mm_storeu_si128((__m128i *)(s - 3 * p), op2);
+        _mm_storeu_si128((__m128i *)(s - 2 * p), op1);
+        _mm_storeu_si128((__m128i *)(s - 1 * p), op0);
+        _mm_storeu_si128((__m128i *)(s - 0 * p), oq0);
+        _mm_storeu_si128((__m128i *)(s + 1 * p), oq1);
+        _mm_storeu_si128((__m128i *)(s + 2 * p), oq2);
+      }
+    } else {
+      _mm_storeu_si128((__m128i *)(s - 2 * p), op1);
+      _mm_storeu_si128((__m128i *)(s - 1 * p), op0);
+      _mm_storeu_si128((__m128i *)(s - 0 * p), oq0);
+      _mm_storeu_si128((__m128i *)(s + 1 * p), oq1);
+    }
+  }
+}
+
+void aom_lpf_horizontal_8_quad_sse2(unsigned char *s, int p,
+                                    const unsigned char *_blimit0,
+                                    const unsigned char *_limit0,
+                                    const unsigned char *_thresh0) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i one = _mm_set1_epi8(1);
+  const __m128i blimit_v = _mm_load_si128((const __m128i *)_blimit0);
+  const __m128i limit_v = _mm_load_si128((const __m128i *)_limit0);
+  const __m128i thresh_v = _mm_load_si128((const __m128i *)_thresh0);
+  __m128i mask, hev, flat;
+  __m128i p3, p2, p1, p0, q0, q1, q2, q3;
+
+  __m128i op2, op1, op0, oq0, oq1, oq2;
+
+  __m128i max_abs_p1p0q1q0;
+
+  p3 = _mm_loadu_si128((__m128i *)(s - 4 * p));
+  p2 = _mm_loadu_si128((__m128i *)(s - 3 * p));
+  p1 = _mm_loadu_si128((__m128i *)(s - 2 * p));
+  p0 = _mm_loadu_si128((__m128i *)(s - 1 * p));
+  q0 = _mm_loadu_si128((__m128i *)(s - 0 * p));
+  q1 = _mm_loadu_si128((__m128i *)(s + 1 * p));
+  q2 = _mm_loadu_si128((__m128i *)(s + 2 * p));
+  q3 = _mm_loadu_si128((__m128i *)(s + 3 * p));
+
+  {
+    const __m128i abs_p1p0 = abs_diff(p1, p0);
+    const __m128i abs_q1q0 = abs_diff(q1, q0);
+    const __m128i fe = _mm_set1_epi8((int8_t)0xfe);
+    const __m128i ff = _mm_cmpeq_epi8(zero, zero);
+    __m128i abs_p0q0 = abs_diff(p0, q0);
+    __m128i abs_p1q1 = abs_diff(p1, q1);
+    __m128i work;
+    max_abs_p1p0q1q0 = _mm_max_epu8(abs_p1p0, abs_q1q0);
+
+    abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0);
+    abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
+    mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit_v);
+    mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
+    // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit) * -1;
+    mask = _mm_max_epu8(max_abs_p1p0q1q0, mask);
+    // mask |= (abs(p1 - p0) > limit) * -1;
+    // mask |= (abs(q1 - q0) > limit) * -1;
+    work = _mm_max_epu8(abs_diff(p2, p1), abs_diff(p3, p2));
+    mask = _mm_max_epu8(work, mask);
+    work = _mm_max_epu8(abs_diff(q2, q1), abs_diff(q3, q2));
+    mask = _mm_max_epu8(work, mask);
+    mask = _mm_subs_epu8(mask, limit_v);
+    mask = _mm_cmpeq_epi8(mask, zero);
+  }
+
+  if (0xffff == _mm_movemask_epi8(_mm_cmpeq_epi8(mask, zero))) return;
+
+  {
+    __m128i work;
+    work = _mm_max_epu8(abs_diff(p2, p0), abs_diff(q2, q0));
+    flat = _mm_max_epu8(work, max_abs_p1p0q1q0);
+    work = _mm_max_epu8(abs_diff(p3, p0), abs_diff(q3, q0));
+    flat = _mm_max_epu8(work, flat);
+    flat = _mm_subs_epu8(flat, one);
+    flat = _mm_cmpeq_epi8(flat, zero);
+    flat = _mm_and_si128(flat, mask);
+  }
+
+  // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  // filter4
+  {
+    const __m128i t4 = _mm_set1_epi8(4);
+    const __m128i t3 = _mm_set1_epi8(3);
+    const __m128i t80 = _mm_set1_epi8((int8_t)0x80);
+    const __m128i te0 = _mm_set1_epi8((int8_t)0xe0);
+    const __m128i t1f = _mm_set1_epi8(0x1f);
+    const __m128i t1 = _mm_set1_epi8(0x1);
+    const __m128i t7f = _mm_set1_epi8(0x7f);
+    const __m128i ff = _mm_cmpeq_epi8(t4, t4);
+
+    __m128i filt;
+    __m128i work_a;
+    __m128i filter1, filter2;
+
+    op1 = _mm_xor_si128(p1, t80);
+    op0 = _mm_xor_si128(p0, t80);
+    oq0 = _mm_xor_si128(q0, t80);
+    oq1 = _mm_xor_si128(q1, t80);
+
+    hev = _mm_subs_epu8(max_abs_p1p0q1q0, thresh_v);
+    hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+    filt = _mm_and_si128(_mm_subs_epi8(op1, oq1), hev);
+
+    work_a = _mm_subs_epi8(oq0, op0);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_and_si128(filt, mask);
+    filter1 = _mm_adds_epi8(filt, t4);
+    filter2 = _mm_adds_epi8(filt, t3);
+
+    work_a = _mm_cmpgt_epi8(zero, filter1);
+    filter1 = _mm_srli_epi16(filter1, 3);
+    work_a = _mm_and_si128(work_a, te0);
+    filter1 = _mm_and_si128(filter1, t1f);
+    filter1 = _mm_or_si128(filter1, work_a);
+    oq0 = _mm_xor_si128(_mm_subs_epi8(oq0, filter1), t80);
+
+    work_a = _mm_cmpgt_epi8(zero, filter2);
+    filter2 = _mm_srli_epi16(filter2, 3);
+    work_a = _mm_and_si128(work_a, te0);
+    filter2 = _mm_and_si128(filter2, t1f);
+    filter2 = _mm_or_si128(filter2, work_a);
+    op0 = _mm_xor_si128(_mm_adds_epi8(op0, filter2), t80);
+
+    filt = _mm_adds_epi8(filter1, t1);
+    work_a = _mm_cmpgt_epi8(zero, filt);
+    filt = _mm_srli_epi16(filt, 1);
+    work_a = _mm_and_si128(work_a, t80);
+    filt = _mm_and_si128(filt, t7f);
+    filt = _mm_or_si128(filt, work_a);
+    filt = _mm_andnot_si128(hev, filt);
+    op1 = _mm_xor_si128(_mm_adds_epi8(op1, filt), t80);
+    oq1 = _mm_xor_si128(_mm_subs_epi8(oq1, filt), t80);
+
+    // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+    // filter8
+    if (0xffff != _mm_movemask_epi8(_mm_cmpeq_epi8(flat, zero))) {
+      const __m128i four = _mm_set1_epi16(4);
+      const __m128i p3_lo = _mm_unpacklo_epi8(p3, zero);
+      const __m128i p2_lo = _mm_unpacklo_epi8(p2, zero);
+      const __m128i p1_lo = _mm_unpacklo_epi8(p1, zero);
+      const __m128i p0_lo = _mm_unpacklo_epi8(p0, zero);
+      const __m128i q0_lo = _mm_unpacklo_epi8(q0, zero);
+      const __m128i q1_lo = _mm_unpacklo_epi8(q1, zero);
+      const __m128i q2_lo = _mm_unpacklo_epi8(q2, zero);
+      const __m128i q3_lo = _mm_unpacklo_epi8(q3, zero);
+
+      const __m128i p3_hi = _mm_unpackhi_epi8(p3, zero);
+      const __m128i p2_hi = _mm_unpackhi_epi8(p2, zero);
+      const __m128i p1_hi = _mm_unpackhi_epi8(p1, zero);
+      const __m128i p0_hi = _mm_unpackhi_epi8(p0, zero);
+      const __m128i q0_hi = _mm_unpackhi_epi8(q0, zero);
+      const __m128i q1_hi = _mm_unpackhi_epi8(q1, zero);
+      const __m128i q2_hi = _mm_unpackhi_epi8(q2, zero);
+      const __m128i q3_hi = _mm_unpackhi_epi8(q3, zero);
+      __m128i f8_lo, f8_hi;
+
+      f8_lo = _mm_add_epi16(_mm_add_epi16(p3_lo, four),
+                            _mm_add_epi16(p3_lo, p2_lo));
+      f8_lo = _mm_add_epi16(_mm_add_epi16(p3_lo, f8_lo),
+                            _mm_add_epi16(p2_lo, p1_lo));
+      f8_lo = _mm_add_epi16(_mm_add_epi16(p0_lo, q0_lo), f8_lo);
+
+      f8_hi = _mm_add_epi16(_mm_add_epi16(p3_hi, four),
+                            _mm_add_epi16(p3_hi, p2_hi));
+      f8_hi = _mm_add_epi16(_mm_add_epi16(p3_hi, f8_hi),
+                            _mm_add_epi16(p2_hi, p1_hi));
+      f8_hi = _mm_add_epi16(_mm_add_epi16(p0_hi, q0_hi), f8_hi);
+
+      op2 = filter8_mask(&flat, &p2, &f8_lo, &f8_hi);
+      _mm_storeu_si128((__m128i *)(s - 3 * p), op2);
+
+      f8_lo = filter_add2_sub2(&f8_lo, &q1_lo, &p1_lo, &p2_lo, &p3_lo);
+      f8_hi = filter_add2_sub2(&f8_hi, &q1_hi, &p1_hi, &p2_hi, &p3_hi);
+      op1 = filter8_mask(&flat, &op1, &f8_lo, &f8_hi);
+      _mm_storeu_si128((__m128i *)(s - 2 * p), op1);
+
+      f8_lo = filter_add2_sub2(&f8_lo, &q2_lo, &p0_lo, &p1_lo, &p3_lo);
+      f8_hi = filter_add2_sub2(&f8_hi, &q2_hi, &p0_hi, &p1_hi, &p3_hi);
+      op0 = filter8_mask(&flat, &op0, &f8_lo, &f8_hi);
+      _mm_storeu_si128((__m128i *)(s - 1 * p), op0);
+
+      f8_lo = filter_add2_sub2(&f8_lo, &q3_lo, &q0_lo, &p0_lo, &p3_lo);
+      f8_hi = filter_add2_sub2(&f8_hi, &q3_hi, &q0_hi, &p0_hi, &p3_hi);
+      oq0 = filter8_mask(&flat, &oq0, &f8_lo, &f8_hi);
+      _mm_storeu_si128((__m128i *)(s - 0 * p), oq0);
+
+      f8_lo = filter_add2_sub2(&f8_lo, &q3_lo, &q1_lo, &q0_lo, &p2_lo);
+      f8_hi = filter_add2_sub2(&f8_hi, &q3_hi, &q1_hi, &q0_hi, &p2_hi);
+      oq1 = filter8_mask(&flat, &oq1, &f8_lo, &f8_hi);
+      _mm_storeu_si128((__m128i *)(s + 1 * p), oq1);
+
+      f8_lo = filter_add2_sub2(&f8_lo, &q3_lo, &q2_lo, &q1_lo, &p1_lo);
+      f8_hi = filter_add2_sub2(&f8_hi, &q3_hi, &q2_hi, &q1_hi, &p1_hi);
+      oq2 = filter8_mask(&flat, &q2, &f8_lo, &f8_hi);
+      _mm_storeu_si128((__m128i *)(s + 2 * p), oq2);
+    } else {
+      _mm_storeu_si128((__m128i *)(s - 2 * p), op1);
+      _mm_storeu_si128((__m128i *)(s - 1 * p), op0);
+      _mm_storeu_si128((__m128i *)(s - 0 * p), oq0);
+      _mm_storeu_si128((__m128i *)(s + 1 * p), oq1);
+    }
+  }
+}
+
+void aom_lpf_horizontal_6_quad_sse2(unsigned char *s, int p,
+                                    const unsigned char *_blimit0,
+                                    const unsigned char *_limit0,
+                                    const unsigned char *_thresh0) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i one = _mm_set1_epi8(1);
+  const __m128i blimit_v = _mm_load_si128((const __m128i *)_blimit0);
+  const __m128i limit_v = _mm_load_si128((const __m128i *)_limit0);
+  const __m128i thresh_v = _mm_load_si128((const __m128i *)_thresh0);
+  __m128i mask, hev, flat;
+  __m128i p2, p1, p0, q0, q1, q2;
+
+  __m128i op1, op0, oq0, oq1;
+
+  __m128i max_abs_p1p0q1q0;
+
+  p2 = _mm_loadu_si128((__m128i *)(s - 3 * p));
+  p1 = _mm_loadu_si128((__m128i *)(s - 2 * p));
+  p0 = _mm_loadu_si128((__m128i *)(s - 1 * p));
+  q0 = _mm_loadu_si128((__m128i *)(s - 0 * p));
+  q1 = _mm_loadu_si128((__m128i *)(s + 1 * p));
+  q2 = _mm_loadu_si128((__m128i *)(s + 2 * p));
+
+  {
+    const __m128i abs_p1p0 = abs_diff(p1, p0);
+    const __m128i abs_q1q0 = abs_diff(q1, q0);
+    const __m128i fe = _mm_set1_epi8((int8_t)0xfe);
+    const __m128i ff = _mm_cmpeq_epi8(zero, zero);
+    __m128i abs_p0q0 = abs_diff(p0, q0);
+    __m128i abs_p1q1 = abs_diff(p1, q1);
+    __m128i work;
+    max_abs_p1p0q1q0 = _mm_max_epu8(abs_p1p0, abs_q1q0);
+
+    abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0);
+    abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
+    mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit_v);
+    mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
+    // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit) * -1;
+    mask = _mm_max_epu8(max_abs_p1p0q1q0, mask);
+    // mask |= (abs(p1 - p0) > limit) * -1;
+    // mask |= (abs(q1 - q0) > limit) * -1;
+    work = _mm_max_epu8(abs_diff(p2, p1), abs_diff(q2, q1));
+    mask = _mm_max_epu8(work, mask);
+    mask = _mm_subs_epu8(mask, limit_v);
+    mask = _mm_cmpeq_epi8(mask, zero);
+  }
+
+  if (0xffff == _mm_movemask_epi8(_mm_cmpeq_epi8(mask, zero))) return;
+
+  {
+    __m128i work;
+    work = _mm_max_epu8(abs_diff(p2, p0), abs_diff(q2, q0));
+    flat = _mm_max_epu8(work, max_abs_p1p0q1q0);
+    flat = _mm_subs_epu8(flat, one);
+    flat = _mm_cmpeq_epi8(flat, zero);
+    flat = _mm_and_si128(flat, mask);
+  }
+
+  // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  // filter4
+  {
+    const __m128i t4 = _mm_set1_epi8(4);
+    const __m128i t3 = _mm_set1_epi8(3);
+    const __m128i t80 = _mm_set1_epi8((int8_t)0x80);
+    const __m128i te0 = _mm_set1_epi8((int8_t)0xe0);
+    const __m128i t1f = _mm_set1_epi8(0x1f);
+    const __m128i t1 = _mm_set1_epi8(0x1);
+    const __m128i t7f = _mm_set1_epi8(0x7f);
+    const __m128i ff = _mm_cmpeq_epi8(t4, t4);
+
+    __m128i filt;
+    __m128i work_a;
+    __m128i filter1, filter2;
+
+    op1 = _mm_xor_si128(p1, t80);
+    op0 = _mm_xor_si128(p0, t80);
+    oq0 = _mm_xor_si128(q0, t80);
+    oq1 = _mm_xor_si128(q1, t80);
+
+    hev = _mm_subs_epu8(max_abs_p1p0q1q0, thresh_v);
+    hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+    filt = _mm_and_si128(_mm_subs_epi8(op1, oq1), hev);
+
+    work_a = _mm_subs_epi8(oq0, op0);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_and_si128(filt, mask);
+    filter1 = _mm_adds_epi8(filt, t4);
+    filter2 = _mm_adds_epi8(filt, t3);
+
+    work_a = _mm_cmpgt_epi8(zero, filter1);
+    filter1 = _mm_srli_epi16(filter1, 3);
+    work_a = _mm_and_si128(work_a, te0);
+    filter1 = _mm_and_si128(filter1, t1f);
+    filter1 = _mm_or_si128(filter1, work_a);
+    oq0 = _mm_xor_si128(_mm_subs_epi8(oq0, filter1), t80);
+
+    work_a = _mm_cmpgt_epi8(zero, filter2);
+    filter2 = _mm_srli_epi16(filter2, 3);
+    work_a = _mm_and_si128(work_a, te0);
+    filter2 = _mm_and_si128(filter2, t1f);
+    filter2 = _mm_or_si128(filter2, work_a);
+    op0 = _mm_xor_si128(_mm_adds_epi8(op0, filter2), t80);
+
+    filt = _mm_adds_epi8(filter1, t1);
+    work_a = _mm_cmpgt_epi8(zero, filt);
+    filt = _mm_srli_epi16(filt, 1);
+    work_a = _mm_and_si128(work_a, t80);
+    filt = _mm_and_si128(filt, t7f);
+    filt = _mm_or_si128(filt, work_a);
+    filt = _mm_andnot_si128(hev, filt);
+    op1 = _mm_xor_si128(_mm_adds_epi8(op1, filt), t80);
+    oq1 = _mm_xor_si128(_mm_subs_epi8(oq1, filt), t80);
+
+    // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+    // filter6
+    if (0xffff != _mm_movemask_epi8(_mm_cmpeq_epi8(flat, zero))) {
+      const __m128i four = _mm_set1_epi16(4);
+      const __m128i p2_lo = _mm_unpacklo_epi8(p2, zero);
+      const __m128i p1_lo = _mm_unpacklo_epi8(p1, zero);
+      const __m128i p0_lo = _mm_unpacklo_epi8(p0, zero);
+      const __m128i q0_lo = _mm_unpacklo_epi8(q0, zero);
+      const __m128i q1_lo = _mm_unpacklo_epi8(q1, zero);
+      const __m128i q2_lo = _mm_unpacklo_epi8(q2, zero);
+
+      const __m128i p2_hi = _mm_unpackhi_epi8(p2, zero);
+      const __m128i p1_hi = _mm_unpackhi_epi8(p1, zero);
+      const __m128i p0_hi = _mm_unpackhi_epi8(p0, zero);
+      const __m128i q0_hi = _mm_unpackhi_epi8(q0, zero);
+      const __m128i q1_hi = _mm_unpackhi_epi8(q1, zero);
+      const __m128i q2_hi = _mm_unpackhi_epi8(q2, zero);
+      __m128i f8_lo, f8_hi;
+
+      f8_lo = _mm_add_epi16(_mm_add_epi16(p2_lo, four),
+                            _mm_add_epi16(p2_lo, p2_lo));
+      f8_lo = _mm_add_epi16(_mm_add_epi16(p1_lo, f8_lo),
+                            _mm_add_epi16(p1_lo, p0_lo));
+      f8_lo = _mm_add_epi16(_mm_add_epi16(p0_lo, q0_lo), f8_lo);
+
+      f8_hi = _mm_add_epi16(_mm_add_epi16(p2_hi, four),
+                            _mm_add_epi16(p2_hi, p2_hi));
+      f8_hi = _mm_add_epi16(_mm_add_epi16(p1_hi, f8_hi),
+                            _mm_add_epi16(p1_hi, p0_hi));
+      f8_hi = _mm_add_epi16(_mm_add_epi16(p0_hi, q0_hi), f8_hi);
+
+      op1 = filter8_mask(&flat, &op1, &f8_lo, &f8_hi);
+      _mm_storeu_si128((__m128i *)(s - 2 * p), op1);
+
+      f8_lo = filter_add2_sub2(&f8_lo, &q0_lo, &q1_lo, &p2_lo, &p2_lo);
+      f8_hi = filter_add2_sub2(&f8_hi, &q0_hi, &q1_hi, &p2_hi, &p2_hi);
+      op0 = filter8_mask(&flat, &op0, &f8_lo, &f8_hi);
+      _mm_storeu_si128((__m128i *)(s - 1 * p), op0);
+
+      f8_lo = filter_add2_sub2(&f8_lo, &q1_lo, &q2_lo, &p1_lo, &p2_lo);
+      f8_hi = filter_add2_sub2(&f8_hi, &q1_hi, &q2_hi, &p1_hi, &p2_hi);
+      oq0 = filter8_mask(&flat, &oq0, &f8_lo, &f8_hi);
+      _mm_storeu_si128((__m128i *)(s - 0 * p), oq0);
+
+      f8_lo = filter_add2_sub2(&f8_lo, &q2_lo, &q2_lo, &p0_lo, &p1_lo);
+      f8_hi = filter_add2_sub2(&f8_hi, &q2_hi, &q2_hi, &p0_hi, &p1_hi);
+      oq1 = filter8_mask(&flat, &oq1, &f8_lo, &f8_hi);
+      _mm_storeu_si128((__m128i *)(s + 1 * p), oq1);
+    } else {
+      _mm_storeu_si128((__m128i *)(s - 2 * p), op1);
+      _mm_storeu_si128((__m128i *)(s - 1 * p), op0);
+      _mm_storeu_si128((__m128i *)(s - 0 * p), oq0);
+      _mm_storeu_si128((__m128i *)(s + 1 * p), oq1);
+    }
+  }
+}
+
+void aom_lpf_horizontal_4_quad_sse2(unsigned char *s, int p,
+                                    const unsigned char *_blimit0,
+                                    const unsigned char *_limit0,
+                                    const unsigned char *_thresh0) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i blimit_v = _mm_load_si128((const __m128i *)_blimit0);
+  const __m128i limit_v = _mm_load_si128((const __m128i *)_limit0);
+  const __m128i thresh_v = _mm_load_si128((const __m128i *)_thresh0);
+  __m128i mask, hev;
+  __m128i p1, p0, q0, q1;
+
+  __m128i op1, op0, oq0, oq1;
+
+  __m128i max_abs_p1p0q1q0;
+
+  p1 = _mm_loadu_si128((__m128i *)(s - 2 * p));
+  p0 = _mm_loadu_si128((__m128i *)(s - 1 * p));
+  q0 = _mm_loadu_si128((__m128i *)(s - 0 * p));
+  q1 = _mm_loadu_si128((__m128i *)(s + 1 * p));
+
+  {
+    const __m128i abs_p1p0 = abs_diff(p1, p0);
+    const __m128i abs_q1q0 = abs_diff(q1, q0);
+    const __m128i fe = _mm_set1_epi8((int8_t)0xfe);
+    const __m128i ff = _mm_cmpeq_epi8(zero, zero);
+    __m128i abs_p0q0 = abs_diff(p0, q0);
+    __m128i abs_p1q1 = abs_diff(p1, q1);
+    max_abs_p1p0q1q0 = _mm_max_epu8(abs_p1p0, abs_q1q0);
+
+    abs_p0q0 = _mm_adds_epu8(abs_p0q0, abs_p0q0);
+    abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
+    mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit_v);
+    mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
+    // mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2  > blimit) * -1;
+    mask = _mm_max_epu8(max_abs_p1p0q1q0, mask);
+    // mask |= (abs(p1 - p0) > limit) * -1;
+    // mask |= (abs(q1 - q0) > limit) * -1;
+    mask = _mm_subs_epu8(mask, limit_v);
+    mask = _mm_cmpeq_epi8(mask, zero);
+  }
+
+  if (0xffff == _mm_movemask_epi8(_mm_cmpeq_epi8(mask, zero))) return;
+
+  // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+  // filter4
+  {
+    const __m128i t4 = _mm_set1_epi8(4);
+    const __m128i t3 = _mm_set1_epi8(3);
+    const __m128i t80 = _mm_set1_epi8((int8_t)0x80);
+    const __m128i te0 = _mm_set1_epi8((int8_t)0xe0);
+    const __m128i t1f = _mm_set1_epi8(0x1f);
+    const __m128i t1 = _mm_set1_epi8(0x1);
+    const __m128i t7f = _mm_set1_epi8(0x7f);
+    const __m128i ff = _mm_cmpeq_epi8(t4, t4);
+
+    __m128i filt;
+    __m128i work_a;
+    __m128i filter1, filter2;
+
+    op1 = _mm_xor_si128(p1, t80);
+    op0 = _mm_xor_si128(p0, t80);
+    oq0 = _mm_xor_si128(q0, t80);
+    oq1 = _mm_xor_si128(q1, t80);
+
+    hev = _mm_subs_epu8(max_abs_p1p0q1q0, thresh_v);
+    hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+    filt = _mm_and_si128(_mm_subs_epi8(op1, oq1), hev);
+
+    work_a = _mm_subs_epi8(oq0, op0);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_adds_epi8(filt, work_a);
+    filt = _mm_and_si128(filt, mask);
+    filter1 = _mm_adds_epi8(filt, t4);
+    filter2 = _mm_adds_epi8(filt, t3);
+
+    work_a = _mm_cmpgt_epi8(zero, filter1);
+    filter1 = _mm_srli_epi16(filter1, 3);
+    work_a = _mm_and_si128(work_a, te0);
+    filter1 = _mm_and_si128(filter1, t1f);
+    filter1 = _mm_or_si128(filter1, work_a);
+    oq0 = _mm_xor_si128(_mm_subs_epi8(oq0, filter1), t80);
+
+    work_a = _mm_cmpgt_epi8(zero, filter2);
+    filter2 = _mm_srli_epi16(filter2, 3);
+    work_a = _mm_and_si128(work_a, te0);
+    filter2 = _mm_and_si128(filter2, t1f);
+    filter2 = _mm_or_si128(filter2, work_a);
+    op0 = _mm_xor_si128(_mm_adds_epi8(op0, filter2), t80);
+
+    filt = _mm_adds_epi8(filter1, t1);
+    work_a = _mm_cmpgt_epi8(zero, filt);
+    filt = _mm_srli_epi16(filt, 1);
+    work_a = _mm_and_si128(work_a, t80);
+    filt = _mm_and_si128(filt, t7f);
+    filt = _mm_or_si128(filt, work_a);
+    filt = _mm_andnot_si128(hev, filt);
+    op1 = _mm_xor_si128(_mm_adds_epi8(op1, filt), t80);
+    oq1 = _mm_xor_si128(_mm_subs_epi8(oq1, filt), t80);
+
+    _mm_storeu_si128((__m128i *)(s - 2 * p), op1);
+    _mm_storeu_si128((__m128i *)(s - 1 * p), op0);
+    _mm_storeu_si128((__m128i *)(s - 0 * p), oq0);
+    _mm_storeu_si128((__m128i *)(s + 1 * p), oq1);
+  }
+}
+
+void aom_lpf_vertical_14_quad_sse2(unsigned char *s, int pitch,
+                                   const uint8_t *_blimit0,
+                                   const uint8_t *_limit0,
+                                   const uint8_t *_thresh0) {
+  DECLARE_ALIGNED(16, unsigned char, t_dst[256]);
+
+  // Transpose 16x16
+  transpose_16x8(s - 8, s - 8 + 8 * pitch, pitch, t_dst, 16);
+  transpose_16x8(s, s + 8 * pitch, pitch, t_dst + 8 * 16, 16);
+
+  // Loop filtering
+  aom_lpf_horizontal_14_quad(t_dst + 8 * 16, 16, _blimit0, _limit0, _thresh0);
+
+  // Transpose back
+  transpose_16x8(t_dst, t_dst + 8 * 16, 16, s - 8, pitch);
+  transpose_16x8(t_dst + 8, t_dst + 8 + 8 * 16, 16, s - 8 + 8 * pitch, pitch);
+}
+
+void aom_lpf_vertical_8_quad_sse2(uint8_t *s, int pitch,
+                                  const uint8_t *_blimit0,
+                                  const uint8_t *_limit0,
+                                  const uint8_t *_thresh0) {
+  DECLARE_ALIGNED(16, unsigned char, t_dst[16 * 8]);
+
+  // Transpose 16x8
+  transpose_16x8(s - 4, s - 4 + pitch * 8, pitch, t_dst, 16);
+
+  // Loop filtering
+  aom_lpf_horizontal_8_quad(t_dst + 4 * 16, 16, _blimit0, _limit0, _thresh0);
+
+  // Transpose back
+  transpose_16x8_to_8x16(t_dst, 16, s - 4, pitch);
+}
+
+void aom_lpf_vertical_6_quad_sse2(uint8_t *s, int pitch,
+                                  const uint8_t *_blimit0,
+                                  const uint8_t *_limit0,
+                                  const uint8_t *_thresh0) {
+  DECLARE_ALIGNED(16, unsigned char, t_dst[16 * 8]);
+
+  // Transpose 16x8:: (wxh) 8x16 to 16x8
+  transpose_16x8(s - 4, s - 4 + pitch * 8, pitch, t_dst, 16);
+
+  // Loop filtering
+  aom_lpf_horizontal_6_quad(t_dst + 4 * 16, 16, _blimit0, _limit0, _thresh0);
+
+  // Transpose back:: (wxh) 16x8 to 8x16
+  transpose_16x8_to_8x16(t_dst, 16, s - 4, pitch);
+}
+
+void aom_lpf_vertical_4_quad_sse2(uint8_t *s, int pitch,
+                                  const uint8_t *_blimit0,
+                                  const uint8_t *_limit0,
+                                  const uint8_t *_thresh0) {
+  DECLARE_ALIGNED(16, unsigned char, t_dst[16 * 8]);
+
+  // Transpose 16x8
+  transpose_16x8(s - 4, s - 4 + pitch * 8, pitch, t_dst, 16);
+
+  // Loop filtering
+  aom_lpf_horizontal_4_quad_sse2(t_dst + 4 * 16, 16, _blimit0, _limit0,
+                                 _thresh0);
+
+  // Transpose back
+  transpose_16x8_to_8x16(t_dst, 16, s - 4, pitch);
+}
diff --git a/third_party/aom/aom_dsp/x86/lpf_common_sse2.h b/third_party/aom/aom_dsp/x86/lpf_common_sse2.h
new file mode 100644
index 0000000000..45464e80b1
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/lpf_common_sse2.h
@@ -0,0 +1,721 @@
+/*
+ * Copyright (c) 2017, 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_AOM_DSP_X86_LPF_COMMON_SSE2_H_
+#define AOM_AOM_DSP_X86_LPF_COMMON_SSE2_H_
+
+#include <emmintrin.h>  // SSE2
+
+#include "config/aom_config.h"
+
+#define mm_storelu(dst, v) memcpy((dst), (const char *)&(v), 8)
+#define mm_storehu(dst, v) memcpy((dst), (const char *)&(v) + 8, 8)
+
+static INLINE void highbd_transpose6x6_sse2(__m128i *x0, __m128i *x1,
+                                            __m128i *x2, __m128i *x3,
+                                            __m128i *x4, __m128i *x5,
+                                            __m128i *d0, __m128i *d1,
+                                            __m128i *d2, __m128i *d3,
+                                            __m128i *d4, __m128i *d5) {
+  __m128i w0, w1, w2, w3, w4, w5, ww0;
+
+  // 00 01 02 03 04 05 xx xx
+  // 10 11 12 13 14 15 xx xx
+  // 20 21 22 23 24 25 xx xx
+  // 30 31 32 33 34 35 xx xx
+  // 40 41 42 43 44 45 xx xx
+  // 50 51 52 53 54 55 xx xx
+
+  w0 = _mm_unpacklo_epi16(*x0, *x1);  // 00 10 01 11 02 12 03 13
+  w1 = _mm_unpacklo_epi16(*x2, *x3);  // 20 30 21 31 22 32 23 33
+  w2 = _mm_unpacklo_epi16(*x4, *x5);  // 40 50 41 51 42 52 43 53
+
+  ww0 = _mm_unpacklo_epi32(w0, w1);   // 00 10 20 30 01 11 21 31
+  *d0 = _mm_unpacklo_epi64(ww0, w2);  // 00 10 20 30 40 50 41 51
+  *d1 = _mm_unpackhi_epi64(ww0,
+                           _mm_srli_si128(w2, 4));  // 01 11 21 31 41 51 xx xx
+
+  ww0 = _mm_unpackhi_epi32(w0, w1);  // 02 12 22 32 03 13 23 33
+  *d2 = _mm_unpacklo_epi64(ww0,
+                           _mm_srli_si128(w2, 8));  // 02 12 22 32 42 52 xx xx
+
+  w3 = _mm_unpackhi_epi16(*x0, *x1);  // 04 14 05 15 xx xx xx xx
+  w4 = _mm_unpackhi_epi16(*x2, *x3);  // 24 34 25 35 xx xx xx xx
+  w5 = _mm_unpackhi_epi16(*x4, *x5);  // 44 54 45 55 xx xx xx xx
+
+  *d3 = _mm_unpackhi_epi64(ww0, _mm_srli_si128(w2, 4));  // 03 13 23 33 43 53
+
+  ww0 = _mm_unpacklo_epi32(w3, w4);   //  04 14 24 34 05 15 25 35
+  *d4 = _mm_unpacklo_epi64(ww0, w5);  //  04 14 24 34 44 54 45 55
+  *d5 = _mm_unpackhi_epi64(ww0,
+                           _mm_slli_si128(w5, 4));  // 05 15 25 35 45 55 xx xx
+}
+
+static INLINE void highbd_transpose4x8_8x4_low_sse2(__m128i *x0, __m128i *x1,
+                                                    __m128i *x2, __m128i *x3,
+                                                    __m128i *d0, __m128i *d1,
+                                                    __m128i *d2, __m128i *d3) {
+  __m128i zero = _mm_setzero_si128();
+  __m128i w0, w1, ww0, ww1;
+
+  w0 = _mm_unpacklo_epi16(*x0, *x1);  // 00 10 01 11 02 12 03 13
+  w1 = _mm_unpacklo_epi16(*x2, *x3);  // 20 30 21 31 22 32 23 33
+
+  ww0 = _mm_unpacklo_epi32(w0, w1);  // 00 10 20 30 01 11 21 31
+  ww1 = _mm_unpackhi_epi32(w0, w1);  // 02 12 22 32 03 13 23 33
+
+  *d0 = _mm_unpacklo_epi64(ww0, zero);  // 00 10 20 30 xx xx xx xx
+  *d1 = _mm_unpackhi_epi64(ww0, zero);  // 01 11 21 31 xx xx xx xx
+  *d2 = _mm_unpacklo_epi64(ww1, zero);  // 02 12 22 32 xx xx xx xx
+  *d3 = _mm_unpackhi_epi64(ww1, zero);  // 03 13 23 33 xx xx xx xx
+}
+
+static INLINE void highbd_transpose4x8_8x4_high_sse2(__m128i *x0, __m128i *x1,
+                                                     __m128i *x2, __m128i *x3,
+                                                     __m128i *d4, __m128i *d5,
+                                                     __m128i *d6, __m128i *d7) {
+  __m128i w0, w1, ww2, ww3;
+  __m128i zero = _mm_setzero_si128();
+
+  w0 = _mm_unpackhi_epi16(*x0, *x1);  // 04 14 05 15 06 16 07 17
+  w1 = _mm_unpackhi_epi16(*x2, *x3);  // 24 34 25 35 26 36 27 37
+
+  ww2 = _mm_unpacklo_epi32(w0, w1);  //  04 14 24 34 05 15 25 35
+  ww3 = _mm_unpackhi_epi32(w0, w1);  //  06 16 26 36 07 17 27 37
+
+  *d4 = _mm_unpacklo_epi64(ww2, zero);  // 04 14 24 34 xx xx xx xx
+  *d5 = _mm_unpackhi_epi64(ww2, zero);  // 05 15 25 35 xx xx xx xx
+  *d6 = _mm_unpacklo_epi64(ww3, zero);  // 06 16 26 36 xx xx xx xx
+  *d7 = _mm_unpackhi_epi64(ww3, zero);  // 07 17 27 37 xx xx xx xx
+}
+
+// here in and out pointers (x and d) should be different! we don't store their
+// values inside
+static INLINE void highbd_transpose4x8_8x4_sse2(__m128i *x0, __m128i *x1,
+                                                __m128i *x2, __m128i *x3,
+                                                __m128i *d0, __m128i *d1,
+                                                __m128i *d2, __m128i *d3,
+                                                __m128i *d4, __m128i *d5,
+                                                __m128i *d6, __m128i *d7) {
+  // input
+  // x0 00 01 02 03 04 05 06 07
+  // x1 10 11 12 13 14 15 16 17
+  // x2 20 21 22 23 24 25 26 27
+  // x3 30 31 32 33 34 35 36 37
+  // output
+  // 00 10 20 30 xx xx xx xx
+  // 01 11 21 31 xx xx xx xx
+  // 02 12 22 32 xx xx xx xx
+  // 03 13 23 33 xx xx xx xx
+  // 04 14 24 34 xx xx xx xx
+  // 05 15 25 35 xx xx xx xx
+  // 06 16 26 36 xx xx xx xx
+  // 07 17 27 37 xx xx xx xx
+  highbd_transpose4x8_8x4_low_sse2(x0, x1, x2, x3, d0, d1, d2, d3);
+  highbd_transpose4x8_8x4_high_sse2(x0, x1, x2, x3, d4, d5, d6, d7);
+}
+
+static INLINE void highbd_transpose8x8_low_sse2(__m128i *x0, __m128i *x1,
+                                                __m128i *x2, __m128i *x3,
+                                                __m128i *x4, __m128i *x5,
+                                                __m128i *x6, __m128i *x7,
+                                                __m128i *d0, __m128i *d1,
+                                                __m128i *d2, __m128i *d3) {
+  __m128i w0, w1, w2, w3, ww0, ww1;
+  // x0 00 01 02 03 04 05 06 07
+  // x1 10 11 12 13 14 15 16 17
+  // x2 20 21 22 23 24 25 26 27
+  // x3 30 31 32 33 34 35 36 37
+  // x4 40 41 42 43 44 45 46 47
+  // x5 50 51 52 53 54 55 56 57
+  // x6 60 61 62 63 64 65 66 67
+  // x7 70 71 72 73 74 75 76 77
+
+  w0 = _mm_unpacklo_epi16(*x0, *x1);  // 00 10 01 11 02 12 03 13
+  w1 = _mm_unpacklo_epi16(*x2, *x3);  // 20 30 21 31 22 32 23 33
+  w2 = _mm_unpacklo_epi16(*x4, *x5);  // 40 50 41 51 42 52 43 53
+  w3 = _mm_unpacklo_epi16(*x6, *x7);  // 60 70 61 71 62 72 63 73
+
+  ww0 = _mm_unpacklo_epi32(w0, w1);  // 00 10 20 30 01 11 21 31
+  ww1 = _mm_unpacklo_epi32(w2, w3);  // 40 50 60 70 41 51 61 71
+
+  *d0 = _mm_unpacklo_epi64(ww0, ww1);  // 00 10 20 30 40 50 60 70
+  *d1 = _mm_unpackhi_epi64(ww0, ww1);  // 01 11 21 31 41 51 61 71
+
+  ww0 = _mm_unpackhi_epi32(w0, w1);  // 02 12 22 32 03 13 23 33
+  ww1 = _mm_unpackhi_epi32(w2, w3);  // 42 52 62 72 43 53 63 73
+
+  *d2 = _mm_unpacklo_epi64(ww0, ww1);  // 02 12 22 32 42 52 62 72
+  *d3 = _mm_unpackhi_epi64(ww0, ww1);  // 03 13 23 33 43 53 63 73
+}
+
+static INLINE void highbd_transpose8x8_high_sse2(__m128i *x0, __m128i *x1,
+                                                 __m128i *x2, __m128i *x3,
+                                                 __m128i *x4, __m128i *x5,
+                                                 __m128i *x6, __m128i *x7,
+                                                 __m128i *d4, __m128i *d5,
+                                                 __m128i *d6, __m128i *d7) {
+  __m128i w0, w1, w2, w3, ww0, ww1;
+  // x0 00 01 02 03 04 05 06 07
+  // x1 10 11 12 13 14 15 16 17
+  // x2 20 21 22 23 24 25 26 27
+  // x3 30 31 32 33 34 35 36 37
+  // x4 40 41 42 43 44 45 46 47
+  // x5 50 51 52 53 54 55 56 57
+  // x6 60 61 62 63 64 65 66 67
+  // x7 70 71 72 73 74 75 76 77
+  w0 = _mm_unpackhi_epi16(*x0, *x1);  // 04 14 05 15 06 16 07 17
+  w1 = _mm_unpackhi_epi16(*x2, *x3);  // 24 34 25 35 26 36 27 37
+  w2 = _mm_unpackhi_epi16(*x4, *x5);  // 44 54 45 55 46 56 47 57
+  w3 = _mm_unpackhi_epi16(*x6, *x7);  // 64 74 65 75 66 76 67 77
+
+  ww0 = _mm_unpacklo_epi32(w0, w1);  // 04 14 24 34 05 15 25 35
+  ww1 = _mm_unpacklo_epi32(w2, w3);  // 44 54 64 74 45 55 65 75
+
+  *d4 = _mm_unpacklo_epi64(ww0, ww1);  // 04 14 24 34 44 54 64 74
+  *d5 = _mm_unpackhi_epi64(ww0, ww1);  // 05 15 25 35 45 55 65 75
+
+  ww0 = _mm_unpackhi_epi32(w0, w1);  // 06 16 26 36 07 17 27 37
+  ww1 = _mm_unpackhi_epi32(w2, w3);  // 46 56 66 76 47 57 67 77
+
+  *d6 = _mm_unpacklo_epi64(ww0, ww1);  // 06 16 26 36 46 56 66 76
+  *d7 = _mm_unpackhi_epi64(ww0, ww1);  // 07 17 27 37 47 57 67 77
+}
+
+// here in and out pointers (x and d) should be different! we don't store their
+// values inside
+static INLINE void highbd_transpose8x8_sse2(
+    __m128i *x0, __m128i *x1, __m128i *x2, __m128i *x3, __m128i *x4,
+    __m128i *x5, __m128i *x6, __m128i *x7, __m128i *d0, __m128i *d1,
+    __m128i *d2, __m128i *d3, __m128i *d4, __m128i *d5, __m128i *d6,
+    __m128i *d7) {
+  highbd_transpose8x8_low_sse2(x0, x1, x2, x3, x4, x5, x6, x7, d0, d1, d2, d3);
+  highbd_transpose8x8_high_sse2(x0, x1, x2, x3, x4, x5, x6, x7, d4, d5, d6, d7);
+}
+
+// here in and out pointers (x and d arrays) should be different! we don't store
+// their values inside
+static INLINE void highbd_transpose8x16_sse2(
+    __m128i *x0, __m128i *x1, __m128i *x2, __m128i *x3, __m128i *x4,
+    __m128i *x5, __m128i *x6, __m128i *x7, __m128i *d0, __m128i *d1,
+    __m128i *d2, __m128i *d3, __m128i *d4, __m128i *d5, __m128i *d6,
+    __m128i *d7) {
+  highbd_transpose8x8_sse2(x0, x1, x2, x3, x4, x5, x6, x7, d0, d1, d2, d3, d4,
+                           d5, d6, d7);
+  highbd_transpose8x8_sse2(x0 + 1, x1 + 1, x2 + 1, x3 + 1, x4 + 1, x5 + 1,
+                           x6 + 1, x7 + 1, d0 + 1, d1 + 1, d2 + 1, d3 + 1,
+                           d4 + 1, d5 + 1, d6 + 1, d7 + 1);
+}
+
+// Low bit depth functions
+static INLINE void transpose4x8_8x4_low_sse2(__m128i *x0, __m128i *x1,
+                                             __m128i *x2, __m128i *x3,
+                                             __m128i *d0, __m128i *d1,
+                                             __m128i *d2, __m128i *d3) {
+  // input
+  // x0   00 01 02 03 04 05 06 07 xx xx xx xx xx xx xx xx
+  // x1   10 11 12 13 14 15 16 17 xx xx xx xx xx xx xx xx
+  // x2   20 21 22 23 24 25 26 27 xx xx xx xx xx xx xx xx
+  // x3   30 31 32 33 34 35 36 37 xx xx xx xx xx xx xx xx
+  // output
+  // 00 10 20 30 xx xx xx xx xx xx xx xx xx xx xx xx
+  // 01 11 21 31 xx xx xx xx xx xx xx xx xx xx xx xx
+  // 02 12 22 32 xx xx xx xx xx xx xx xx xx xx xx xx
+  // 03 13 23 33 xx xx xx xx xx xx xx xx xx xx xx xx
+
+  __m128i w0, w1;
+
+  w0 = _mm_unpacklo_epi8(
+      *x0, *x1);  // 00 10 01 11 02 12 03 13 04 14 05 15 06 16 07 17
+  w1 = _mm_unpacklo_epi8(
+      *x2, *x3);  // 20 30 21 31 22 32 23 33 24 34 25 35 26 36 27 37
+
+  *d0 = _mm_unpacklo_epi16(
+      w0, w1);  // 00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33
+
+  *d1 = _mm_srli_si128(*d0,
+                       4);  // 01 11 21 31 xx xx xx xx xx xx xx xx xx xx xx xx
+  *d2 = _mm_srli_si128(*d0,
+                       8);  // 02 12 22 32 xx xx xx xx xx xx xx xx xx xx xx xx
+  *d3 = _mm_srli_si128(*d0,
+                       12);  // 03 13 23 33 xx xx xx xx xx xx xx xx xx xx xx xx
+}
+
+static INLINE void transpose4x8_8x4_sse2(__m128i *x0, __m128i *x1, __m128i *x2,
+                                         __m128i *x3, __m128i *d0, __m128i *d1,
+                                         __m128i *d2, __m128i *d3, __m128i *d4,
+                                         __m128i *d5, __m128i *d6,
+                                         __m128i *d7) {
+  // input
+  // x0   00 01 02 03 04 05 06 07 xx xx xx xx xx xx xx xx
+  // x1   10 11 12 13 14 15 16 17 xx xx xx xx xx xx xx xx
+  // x2   20 21 22 23 24 25 26 27 xx xx xx xx xx xx xx xx
+  // x3   30 31 32 33 34 35 36 37 xx xx xx xx xx xx xx xx
+  // output
+  // 00 10 20 30 xx xx xx xx xx xx xx xx xx xx xx xx
+  // 01 11 21 31 xx xx xx xx xx xx xx xx xx xx xx xx
+  // 02 12 22 32 xx xx xx xx xx xx xx xx xx xx xx xx
+  // 03 13 23 33 xx xx xx xx xx xx xx xx xx xx xx xx
+  // 04 14 24 34 xx xx xx xx xx xx xx xx xx xx xx xx
+  // 05 15 25 35 xx xx xx xx xx xx xx xx xx xx xx xx
+  // 06 16 26 36 xx xx xx xx xx xx xx xx xx xx xx xx
+  // 07 17 27 37 xx xx xx xx xx xx xx xx xx xx xx xx
+
+  __m128i w0, w1, ww0, ww1;
+
+  w0 = _mm_unpacklo_epi8(
+      *x0, *x1);  // 00 10 01 11 02 12 03 13 04 14 05 15 06 16 07 17
+  w1 = _mm_unpacklo_epi8(
+      *x2, *x3);  // 20 30 21 31 22 32 23 33 24 34 25 35 26 36 27 37
+
+  ww0 = _mm_unpacklo_epi16(
+      w0, w1);  // 00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33
+  ww1 = _mm_unpackhi_epi16(
+      w0, w1);  // 04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37
+
+  *d0 = ww0;  // 00 10 20 30 xx xx xx xx xx xx xx xx xx xx xx xx
+  *d1 = _mm_srli_si128(ww0,
+                       4);  // 01 11 21 31 xx xx xx xx xx xx xx xx xx xx xx xx
+  *d2 = _mm_srli_si128(ww0,
+                       8);  // 02 12 22 32 xx xx xx xx xx xx xx xx xx xx xx xx
+  *d3 = _mm_srli_si128(ww0,
+                       12);  // 03 13 23 33 xx xx xx xx xx xx xx xx xx xx xx xx
+
+  *d4 = ww1;  // 04 14 24 34 xx xx xx xx xx xx xx xx xx xx xx xx
+  *d5 = _mm_srli_si128(ww1,
+                       4);  // 05 15 25 35 xx xx xx xx xx xx xx xx xx xx xx xx
+  *d6 = _mm_srli_si128(ww1,
+                       8);  // 06 16 26 36 xx xx xx xx xx xx xx xx xx xx xx xx
+  *d7 = _mm_srli_si128(ww1,
+                       12);  // 07 17 27 37 xx xx xx xx xx xx xx xx xx xx xx xx
+}
+
+static INLINE void transpose8x8_low_sse2(__m128i *x0, __m128i *x1, __m128i *x2,
+                                         __m128i *x3, __m128i *x4, __m128i *x5,
+                                         __m128i *x6, __m128i *x7, __m128i *d0,
+                                         __m128i *d1, __m128i *d2,
+                                         __m128i *d3) {
+  // input
+  // x0 00 01 02 03 04 05 06 07
+  // x1 10 11 12 13 14 15 16 17
+  // x2 20 21 22 23 24 25 26 27
+  // x3 30 31 32 33 34 35 36 37
+  // x4 40 41 42 43 44 45 46 47
+  // x5  50 51 52 53 54 55 56 57
+  // x6  60 61 62 63 64 65 66 67
+  // x7 70 71 72 73 74 75 76 77
+  // output
+  // d0 00 10 20 30 40 50 60 70 xx xx xx xx xx xx xx
+  // d1 01 11 21 31 41 51 61 71 xx xx xx xx xx xx xx xx
+  // d2 02 12 22 32 42 52 62 72 xx xx xx xx xx xx xx xx
+  // d3 03 13 23 33 43 53 63 73 xx xx xx xx xx xx xx xx
+
+  __m128i w0, w1, w2, w3, w4, w5;
+
+  w0 = _mm_unpacklo_epi8(
+      *x0, *x1);  // 00 10 01 11 02 12 03 13 04 14 05 15 06 16 07 17
+
+  w1 = _mm_unpacklo_epi8(
+      *x2, *x3);  // 20 30 21 31 22 32 23 33 24 34 25 35 26 36 27 37
+
+  w2 = _mm_unpacklo_epi8(
+      *x4, *x5);  // 40 50 41 51 42 52 43 53 44 54 45 55 46 56 47 57
+
+  w3 = _mm_unpacklo_epi8(
+      *x6, *x7);  // 60 70 61 71 62 72 63 73 64 74 65 75 66 76 67 77
+
+  w4 = _mm_unpacklo_epi16(
+      w0, w1);  // 00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33
+  w5 = _mm_unpacklo_epi16(
+      w2, w3);  // 40 50 60 70 41 51 61 71 42 52 62 72 43 53 63 73
+
+  *d0 = _mm_unpacklo_epi32(
+      w4, w5);  // 00 10 20 30 40 50 60 70 01 11 21 31 41 51 61 71
+  *d1 = _mm_srli_si128(*d0, 8);
+  *d2 = _mm_unpackhi_epi32(
+      w4, w5);  // 02 12 22 32 42 52 62 72 03 13 23 33 43 53 63 73
+  *d3 = _mm_srli_si128(*d2, 8);
+}
+
+static INLINE void transpose8x8_sse2(__m128i *x0, __m128i *x1, __m128i *x2,
+                                     __m128i *x3, __m128i *x4, __m128i *x5,
+                                     __m128i *x6, __m128i *x7, __m128i *d0d1,
+                                     __m128i *d2d3, __m128i *d4d5,
+                                     __m128i *d6d7) {
+  __m128i w0, w1, w2, w3, w4, w5, w6, w7;
+  // x0 00 01 02 03 04 05 06 07
+  // x1 10 11 12 13 14 15 16 17
+  w0 = _mm_unpacklo_epi8(
+      *x0, *x1);  // 00 10 01 11 02 12 03 13 04 14 05 15 06 16 07 17
+
+  // x2 20 21 22 23 24 25 26 27
+  // x3 30 31 32 33 34 35 36 37
+  w1 = _mm_unpacklo_epi8(
+      *x2, *x3);  // 20 30 21 31 22 32 23 33 24 34 25 35 26 36 27 37
+
+  // x4 40 41 42 43 44 45 46 47
+  // x5  50 51 52 53 54 55 56 57
+  w2 = _mm_unpacklo_epi8(
+      *x4, *x5);  // 40 50 41 51 42 52 43 53 44 54 45 55 46 56 47 57
+
+  // x6  60 61 62 63 64 65 66 67
+  // x7 70 71 72 73 74 75 76 77
+  w3 = _mm_unpacklo_epi8(
+      *x6, *x7);  // 60 70 61 71 62 72 63 73 64 74 65 75 66 76 67 77
+
+  w4 = _mm_unpacklo_epi16(
+      w0, w1);  // 00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33
+  w5 = _mm_unpacklo_epi16(
+      w2, w3);  // 40 50 60 70 41 51 61 71 42 52 62 72 43 53 63 73
+
+  *d0d1 = _mm_unpacklo_epi32(
+      w4, w5);  // 00 10 20 30 40 50 60 70 01 11 21 31 41 51 61 71
+  *d2d3 = _mm_unpackhi_epi32(
+      w4, w5);  // 02 12 22 32 42 52 62 72 03 13 23 33 43 53 63 73
+
+  w6 = _mm_unpackhi_epi16(
+      w0, w1);  // 04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37
+  w7 = _mm_unpackhi_epi16(
+      w2, w3);  // 44 54 64 74 45 55 65 75 46 56 66 76 47 57 67 77
+
+  *d4d5 = _mm_unpacklo_epi32(
+      w6, w7);  // 04 14 24 34 44 54 64 74 05 15 25 35 45 55 65 75
+  *d6d7 = _mm_unpackhi_epi32(
+      w6, w7);  // 06 16 26 36 46 56 66 76 07 17 27 37 47 57 67 77
+}
+
+static INLINE void transpose16x8_8x16_sse2(
+    __m128i *x0, __m128i *x1, __m128i *x2, __m128i *x3, __m128i *x4,
+    __m128i *x5, __m128i *x6, __m128i *x7, __m128i *x8, __m128i *x9,
+    __m128i *x10, __m128i *x11, __m128i *x12, __m128i *x13, __m128i *x14,
+    __m128i *x15, __m128i *d0, __m128i *d1, __m128i *d2, __m128i *d3,
+    __m128i *d4, __m128i *d5, __m128i *d6, __m128i *d7) {
+  __m128i w0, w1, w2, w3, w4, w5, w6, w7, w8, w9;
+  __m128i w10, w11, w12, w13, w14, w15;
+
+  w0 = _mm_unpacklo_epi8(*x0, *x1);
+  w1 = _mm_unpacklo_epi8(*x2, *x3);
+  w2 = _mm_unpacklo_epi8(*x4, *x5);
+  w3 = _mm_unpacklo_epi8(*x6, *x7);
+
+  w8 = _mm_unpacklo_epi8(*x8, *x9);
+  w9 = _mm_unpacklo_epi8(*x10, *x11);
+  w10 = _mm_unpacklo_epi8(*x12, *x13);
+  w11 = _mm_unpacklo_epi8(*x14, *x15);
+
+  w4 = _mm_unpacklo_epi16(w0, w1);
+  w5 = _mm_unpacklo_epi16(w2, w3);
+  w12 = _mm_unpacklo_epi16(w8, w9);
+  w13 = _mm_unpacklo_epi16(w10, w11);
+
+  w6 = _mm_unpacklo_epi32(w4, w5);
+  w7 = _mm_unpackhi_epi32(w4, w5);
+  w14 = _mm_unpacklo_epi32(w12, w13);
+  w15 = _mm_unpackhi_epi32(w12, w13);
+
+  // Store first 4-line result
+  *d0 = _mm_unpacklo_epi64(w6, w14);
+  *d1 = _mm_unpackhi_epi64(w6, w14);
+  *d2 = _mm_unpacklo_epi64(w7, w15);
+  *d3 = _mm_unpackhi_epi64(w7, w15);
+
+  w4 = _mm_unpackhi_epi16(w0, w1);
+  w5 = _mm_unpackhi_epi16(w2, w3);
+  w12 = _mm_unpackhi_epi16(w8, w9);
+  w13 = _mm_unpackhi_epi16(w10, w11);
+
+  w6 = _mm_unpacklo_epi32(w4, w5);
+  w7 = _mm_unpackhi_epi32(w4, w5);
+  w14 = _mm_unpacklo_epi32(w12, w13);
+  w15 = _mm_unpackhi_epi32(w12, w13);
+
+  // Store second 4-line result
+  *d4 = _mm_unpacklo_epi64(w6, w14);
+  *d5 = _mm_unpackhi_epi64(w6, w14);
+  *d6 = _mm_unpacklo_epi64(w7, w15);
+  *d7 = _mm_unpackhi_epi64(w7, w15);
+}
+
+static INLINE void transpose8x16_16x8_sse2(
+    __m128i *x0, __m128i *x1, __m128i *x2, __m128i *x3, __m128i *x4,
+    __m128i *x5, __m128i *x6, __m128i *x7, __m128i *d0d1, __m128i *d2d3,
+    __m128i *d4d5, __m128i *d6d7, __m128i *d8d9, __m128i *d10d11,
+    __m128i *d12d13, __m128i *d14d15) {
+  __m128i w0, w1, w2, w3, w4, w5, w6, w7, w8, w9;
+  __m128i w10, w11, w12, w13, w14, w15;
+
+  w0 = _mm_unpacklo_epi8(*x0, *x1);
+  w1 = _mm_unpacklo_epi8(*x2, *x3);
+  w2 = _mm_unpacklo_epi8(*x4, *x5);
+  w3 = _mm_unpacklo_epi8(*x6, *x7);
+
+  w8 = _mm_unpackhi_epi8(*x0, *x1);
+  w9 = _mm_unpackhi_epi8(*x2, *x3);
+  w10 = _mm_unpackhi_epi8(*x4, *x5);
+  w11 = _mm_unpackhi_epi8(*x6, *x7);
+
+  w4 = _mm_unpacklo_epi16(w0, w1);
+  w5 = _mm_unpacklo_epi16(w2, w3);
+  w12 = _mm_unpacklo_epi16(w8, w9);
+  w13 = _mm_unpacklo_epi16(w10, w11);
+
+  w6 = _mm_unpacklo_epi32(w4, w5);
+  w7 = _mm_unpackhi_epi32(w4, w5);
+  w14 = _mm_unpacklo_epi32(w12, w13);
+  w15 = _mm_unpackhi_epi32(w12, w13);
+
+  // Store first 4-line result
+  *d0d1 = _mm_unpacklo_epi64(w6, w14);
+  *d2d3 = _mm_unpackhi_epi64(w6, w14);
+  *d4d5 = _mm_unpacklo_epi64(w7, w15);
+  *d6d7 = _mm_unpackhi_epi64(w7, w15);
+
+  w4 = _mm_unpackhi_epi16(w0, w1);
+  w5 = _mm_unpackhi_epi16(w2, w3);
+  w12 = _mm_unpackhi_epi16(w8, w9);
+  w13 = _mm_unpackhi_epi16(w10, w11);
+
+  w6 = _mm_unpacklo_epi32(w4, w5);
+  w7 = _mm_unpackhi_epi32(w4, w5);
+  w14 = _mm_unpacklo_epi32(w12, w13);
+  w15 = _mm_unpackhi_epi32(w12, w13);
+
+  // Store second 4-line result
+  *d8d9 = _mm_unpacklo_epi64(w6, w14);
+  *d10d11 = _mm_unpackhi_epi64(w6, w14);
+  *d12d13 = _mm_unpacklo_epi64(w7, w15);
+  *d14d15 = _mm_unpackhi_epi64(w7, w15);
+}
+
+static INLINE void transpose_16x8(unsigned char *in0, unsigned char *in1,
+                                  int in_p, unsigned char *out, int out_p) {
+  __m128i x0, x1, x2, x3, x4, x5, x6, x7;
+  __m128i x8, x9, x10, x11, x12, x13, x14, x15;
+
+  x0 = _mm_loadl_epi64((__m128i *)in0);
+  x1 = _mm_loadl_epi64((__m128i *)(in0 + in_p));
+  x0 = _mm_unpacklo_epi8(x0, x1);
+
+  x2 = _mm_loadl_epi64((__m128i *)(in0 + 2 * in_p));
+  x3 = _mm_loadl_epi64((__m128i *)(in0 + 3 * in_p));
+  x1 = _mm_unpacklo_epi8(x2, x3);
+
+  x4 = _mm_loadl_epi64((__m128i *)(in0 + 4 * in_p));
+  x5 = _mm_loadl_epi64((__m128i *)(in0 + 5 * in_p));
+  x2 = _mm_unpacklo_epi8(x4, x5);
+
+  x6 = _mm_loadl_epi64((__m128i *)(in0 + 6 * in_p));
+  x7 = _mm_loadl_epi64((__m128i *)(in0 + 7 * in_p));
+  x3 = _mm_unpacklo_epi8(x6, x7);
+  x4 = _mm_unpacklo_epi16(x0, x1);
+
+  x8 = _mm_loadl_epi64((__m128i *)in1);
+  x9 = _mm_loadl_epi64((__m128i *)(in1 + in_p));
+  x8 = _mm_unpacklo_epi8(x8, x9);
+  x5 = _mm_unpacklo_epi16(x2, x3);
+
+  x10 = _mm_loadl_epi64((__m128i *)(in1 + 2 * in_p));
+  x11 = _mm_loadl_epi64((__m128i *)(in1 + 3 * in_p));
+  x9 = _mm_unpacklo_epi8(x10, x11);
+
+  x12 = _mm_loadl_epi64((__m128i *)(in1 + 4 * in_p));
+  x13 = _mm_loadl_epi64((__m128i *)(in1 + 5 * in_p));
+  x10 = _mm_unpacklo_epi8(x12, x13);
+  x12 = _mm_unpacklo_epi16(x8, x9);
+
+  x14 = _mm_loadl_epi64((__m128i *)(in1 + 6 * in_p));
+  x15 = _mm_loadl_epi64((__m128i *)(in1 + 7 * in_p));
+  x11 = _mm_unpacklo_epi8(x14, x15);
+  x13 = _mm_unpacklo_epi16(x10, x11);
+
+  x6 = _mm_unpacklo_epi32(x4, x5);
+  x7 = _mm_unpackhi_epi32(x4, x5);
+  x14 = _mm_unpacklo_epi32(x12, x13);
+  x15 = _mm_unpackhi_epi32(x12, x13);
+
+  // Store first 4-line result
+  _mm_storeu_si128((__m128i *)out, _mm_unpacklo_epi64(x6, x14));
+  _mm_storeu_si128((__m128i *)(out + out_p), _mm_unpackhi_epi64(x6, x14));
+  _mm_storeu_si128((__m128i *)(out + 2 * out_p), _mm_unpacklo_epi64(x7, x15));
+  _mm_storeu_si128((__m128i *)(out + 3 * out_p), _mm_unpackhi_epi64(x7, x15));
+
+  x4 = _mm_unpackhi_epi16(x0, x1);
+  x5 = _mm_unpackhi_epi16(x2, x3);
+  x12 = _mm_unpackhi_epi16(x8, x9);
+  x13 = _mm_unpackhi_epi16(x10, x11);
+
+  x6 = _mm_unpacklo_epi32(x4, x5);
+  x7 = _mm_unpackhi_epi32(x4, x5);
+  x14 = _mm_unpacklo_epi32(x12, x13);
+  x15 = _mm_unpackhi_epi32(x12, x13);
+
+  // Store second 4-line result
+  _mm_storeu_si128((__m128i *)(out + 4 * out_p), _mm_unpacklo_epi64(x6, x14));
+  _mm_storeu_si128((__m128i *)(out + 5 * out_p), _mm_unpackhi_epi64(x6, x14));
+  _mm_storeu_si128((__m128i *)(out + 6 * out_p), _mm_unpacklo_epi64(x7, x15));
+  _mm_storeu_si128((__m128i *)(out + 7 * out_p), _mm_unpackhi_epi64(x7, x15));
+}
+
+static INLINE void transpose_16x8_to_8x16(unsigned char *src, int in_p,
+                                          unsigned char *dst, int out_p) {
+  // a0 b0 c0 d0 e0 f0 g0 h0 A0 B0 C0 D0 E0 F0 G0 H0
+  // a1 b1 c1 d1 e1 f1 g1 h1 A1 B1 C1 D1 E1 F1 G1 H1
+  // a2 b2 c2 d2 e2 f2 g2 h2 A2 B2 C2 D2 E2 F2 G2 H2
+  // a3 b3 c3 d3 e3 f3 g3 h3 A3 B3 C3 D3 E3 F3 G3 H3
+  // a4 b4 c4 d4 e4 f4 g4 h4 A4 B4 C4 D4 E4 F4 G4 H4
+  // a5 b5 c5 d5 e5 f5 g5 h5 A5 B5 C5 D5 E5 F5 G5 H5
+  // a6 b6 c6 d6 e6 f6 g6 h6 A6 B6 C6 D6 E6 F6 G6 H6
+  // a7 b7 c7 d7 e7 f7 g7 h7 A7 B7 C7 D7 E7 F7 G7 H7
+  const __m128i x0 = _mm_loadu_si128((__m128i *)(src));
+  const __m128i x1 = _mm_loadu_si128((__m128i *)(src + (1 * in_p)));
+  const __m128i x2 = _mm_loadu_si128((__m128i *)(src + (2 * in_p)));
+  const __m128i x3 = _mm_loadu_si128((__m128i *)(src + (3 * in_p)));
+  const __m128i x4 = _mm_loadu_si128((__m128i *)(src + (4 * in_p)));
+  const __m128i x5 = _mm_loadu_si128((__m128i *)(src + (5 * in_p)));
+  const __m128i x6 = _mm_loadu_si128((__m128i *)(src + (6 * in_p)));
+  const __m128i x7 = _mm_loadu_si128((__m128i *)(src + (7 * in_p)));
+
+  // a0 a1 b0 b1 c0 c1 d0 d1 A0 A1 B0 B1 C0 C1 D0 D1
+  // e0 e1 f0 f1 g0 g1 h0 h1 E0 E1 F0 F1 G0 G1 H0 H1
+  // a2 a3 b2 b3 c2 c3 d2 d3 A2 A3 B2 B3 C2 C3 D2 D3
+  // e2 e3 f2 f3 g2 g3 h2 h3 E2 E3 F2 F3 G2 G3 H2 H3
+  // a4 a5 b4 b5 c4 c5 d4 d5 A4 A5 B4 B5 C4 C5 D4 D5
+  // e4 e5 f4 f5 g4 g5 h4 h5 E4 E5 F4 F5 G4 G5 H4 H5
+  // a6 a7 b6 b7 c6 c7 d6 d7 A6 A7 B6 B7 C6 C7 D6 D7
+  // e6 e7 f6 f7 g6 g7 h6 h7 E6 E7 F6 F7 G6 G7 H6 H7
+  const __m128i x_s10 = _mm_unpacklo_epi8(x0, x1);
+  const __m128i x_s11 = _mm_unpackhi_epi8(x0, x1);
+  const __m128i x_s12 = _mm_unpacklo_epi8(x2, x3);
+  const __m128i x_s13 = _mm_unpackhi_epi8(x2, x3);
+  const __m128i x_s14 = _mm_unpacklo_epi8(x4, x5);
+  const __m128i x_s15 = _mm_unpackhi_epi8(x4, x5);
+  const __m128i x_s16 = _mm_unpacklo_epi8(x6, x7);
+  const __m128i x_s17 = _mm_unpackhi_epi8(x6, x7);
+
+  // a0 a1 a2 a3 b0 b1 b2 b3 | A0 A1 A2 A3 B0 B1 B2 B3
+  // c0 c1 c2 c3 d0 d1 d2 d3 | C0 C1 C2 C3 D0 D1 D2 D3
+  // e0 e1 e2 e3 f0 f1 f2 f3 | E0 E1 E2 E3 F0 F1 F2 F3
+  // g0 g1 g2 g3 h0 h1 h2 h3 | G0 G1 G2 G3 H0 H1 H2 H3
+  // a4 a5 a6 a7 b4 b5 b6 b7 | A4 A5 A6 A7 B4 B5 B6 B7
+  // c4 c5 c6 c7 d4 d5 d6 d7 | C4 C5 C6 C7 D4 D5 D6 D7
+  // e4 e5 e6 e7 f4 f5 f6 f7 | E4 E5 E6 E7 F4 F5 F6 F7
+  // g4 g5 g6 g7 h4 h5 h6 h7 | G4 G5 G6 G7 H4 H5 H6 H7
+  const __m128i x_s20 = _mm_unpacklo_epi16(x_s10, x_s12);
+  const __m128i x_s21 = _mm_unpackhi_epi16(x_s10, x_s12);
+  const __m128i x_s22 = _mm_unpacklo_epi16(x_s11, x_s13);
+  const __m128i x_s23 = _mm_unpackhi_epi16(x_s11, x_s13);
+  const __m128i x_s24 = _mm_unpacklo_epi16(x_s14, x_s16);
+  const __m128i x_s25 = _mm_unpackhi_epi16(x_s14, x_s16);
+  const __m128i x_s26 = _mm_unpacklo_epi16(x_s15, x_s17);
+  const __m128i x_s27 = _mm_unpackhi_epi16(x_s15, x_s17);
+
+  // a0 a1 a2 a3 a4 a5 a6 a7 | A0 A1 A2 A3 A4 A5 A6 A7
+  // b0 b1 b2 b3 b4 b5 b6 b7 | B0 B1 B2 B3 B4 B5 B6 B7
+  // c0 c1 c2 c3 c4 c5 c6 c7 | C0 C1 C2 C3 C4 C5 C6 C7
+  // d0 d1 d2 d3 d4 d5 d6 d7 | D0 D1 D2 D3 D4 D5 D6 D7
+  // e0 e1 e2 e3 e4 e5 e6 e7 | E0 E1 E2 E3 E4 E5 E6 E7
+  // f0 f1 f2 f3 f4 f5 f6 f7 | F0 F1 F2 F3 F4 F5 F6 F7
+  // g0 g1 g2 g3 g4 g5 g6 g7 | G0 G1 G2 G3 G4 G5 G6 G7
+  // h0 h1 h2 h3 h4 h5 h6 h7 | H0 H1 H2 H3 H4 H5 H6 H7
+  const __m128i x_s30 = _mm_unpacklo_epi32(x_s20, x_s24);
+  const __m128i x_s31 = _mm_unpackhi_epi32(x_s20, x_s24);
+  const __m128i x_s32 = _mm_unpacklo_epi32(x_s21, x_s25);
+  const __m128i x_s33 = _mm_unpackhi_epi32(x_s21, x_s25);
+  const __m128i x_s34 = _mm_unpacklo_epi32(x_s22, x_s26);
+  const __m128i x_s35 = _mm_unpackhi_epi32(x_s22, x_s26);
+  const __m128i x_s36 = _mm_unpacklo_epi32(x_s23, x_s27);
+  const __m128i x_s37 = _mm_unpackhi_epi32(x_s23, x_s27);
+
+  mm_storelu(dst, x_s30);
+  mm_storehu(dst + (1 * out_p), x_s30);
+  mm_storelu(dst + (2 * out_p), x_s31);
+  mm_storehu(dst + (3 * out_p), x_s31);
+  mm_storelu(dst + (4 * out_p), x_s32);
+  mm_storehu(dst + (5 * out_p), x_s32);
+  mm_storelu(dst + (6 * out_p), x_s33);
+  mm_storehu(dst + (7 * out_p), x_s33);
+  mm_storelu(dst + (8 * out_p), x_s34);
+  mm_storehu(dst + (9 * out_p), x_s34);
+  mm_storelu(dst + (10 * out_p), x_s35);
+  mm_storehu(dst + (11 * out_p), x_s35);
+  mm_storelu(dst + (12 * out_p), x_s36);
+  mm_storehu(dst + (13 * out_p), x_s36);
+  mm_storelu(dst + (14 * out_p), x_s37);
+  mm_storehu(dst + (15 * out_p), x_s37);
+}
+
+static INLINE void transpose_8xn(unsigned char *src[], int in_p,
+                                 unsigned char *dst[], int out_p,
+                                 int num_8x8_to_transpose) {
+  int idx8x8 = 0;
+  __m128i x0, x1, x2, x3, x4, x5, x6, x7;
+  do {
+    unsigned char *in = src[idx8x8];
+    unsigned char *out = dst[idx8x8];
+
+    x0 =
+        _mm_loadl_epi64((__m128i *)(in + 0 * in_p));  // 00 01 02 03 04 05 06 07
+    x1 =
+        _mm_loadl_epi64((__m128i *)(in + 1 * in_p));  // 10 11 12 13 14 15 16 17
+    // 00 10 01 11 02 12 03 13 04 14 05 15 06 16 07 17
+    x0 = _mm_unpacklo_epi8(x0, x1);
+
+    x2 =
+        _mm_loadl_epi64((__m128i *)(in + 2 * in_p));  // 20 21 22 23 24 25 26 27
+    x3 =
+        _mm_loadl_epi64((__m128i *)(in + 3 * in_p));  // 30 31 32 33 34 35 36 37
+    // 20 30 21 31 22 32 23 33 24 34 25 35 26 36 27 37
+    x1 = _mm_unpacklo_epi8(x2, x3);
+
+    x4 =
+        _mm_loadl_epi64((__m128i *)(in + 4 * in_p));  // 40 41 42 43 44 45 46 47
+    x5 =
+        _mm_loadl_epi64((__m128i *)(in + 5 * in_p));  // 50 51 52 53 54 55 56 57
+    // 40 50 41 51 42 52 43 53 44 54 45 55 46 56 47 57
+    x2 = _mm_unpacklo_epi8(x4, x5);
+
+    x6 =
+        _mm_loadl_epi64((__m128i *)(in + 6 * in_p));  // 60 61 62 63 64 65 66 67
+    x7 =
+        _mm_loadl_epi64((__m128i *)(in + 7 * in_p));  // 70 71 72 73 74 75 76 77
+    // 60 70 61 71 62 72 63 73 64 74 65 75 66 76 67 77
+    x3 = _mm_unpacklo_epi8(x6, x7);
+
+    // 00 10 20 30 01 11 21 31 02 12 22 32 03 13 23 33
+    x4 = _mm_unpacklo_epi16(x0, x1);
+    // 40 50 60 70 41 51 61 71 42 52 62 72 43 53 63 73
+    x5 = _mm_unpacklo_epi16(x2, x3);
+    // 00 10 20 30 40 50 60 70 01 11 21 31 41 51 61 71
+    x6 = _mm_unpacklo_epi32(x4, x5);
+    mm_storelu(out + 0 * out_p, x6);  // 00 10 20 30 40 50 60 70
+    mm_storehu(out + 1 * out_p, x6);  // 01 11 21 31 41 51 61 71
+    // 02 12 22 32 42 52 62 72 03 13 23 33 43 53 63 73
+    x7 = _mm_unpackhi_epi32(x4, x5);
+    mm_storelu(out + 2 * out_p, x7);  // 02 12 22 32 42 52 62 72
+    mm_storehu(out + 3 * out_p, x7);  // 03 13 23 33 43 53 63 73
+
+    // 04 14 24 34 05 15 25 35 06 16 26 36 07 17 27 37
+    x4 = _mm_unpackhi_epi16(x0, x1);
+    // 44 54 64 74 45 55 65 75 46 56 66 76 47 57 67 77
+    x5 = _mm_unpackhi_epi16(x2, x3);
+    // 04 14 24 34 44 54 64 74 05 15 25 35 45 55 65 75
+    x6 = _mm_unpacklo_epi32(x4, x5);
+    mm_storelu(out + 4 * out_p, x6);  // 04 14 24 34 44 54 64 74
+    mm_storehu(out + 5 * out_p, x6);  // 05 15 25 35 45 55 65 75
+    // 06 16 26 36 46 56 66 76 07 17 27 37 47 57 67 77
+    x7 = _mm_unpackhi_epi32(x4, x5);
+
+    mm_storelu(out + 6 * out_p, x7);  // 06 16 26 36 46 56 66 76
+    mm_storehu(out + 7 * out_p, x7);  // 07 17 27 37 47 57 67 77
+  } while (++idx8x8 < num_8x8_to_transpose);
+}
+
+#endif  // AOM_AOM_DSP_X86_LPF_COMMON_SSE2_H_
diff --git a/third_party/aom/aom_dsp/x86/masked_sad4d_ssse3.c b/third_party/aom/aom_dsp/x86/masked_sad4d_ssse3.c
new file mode 100644
index 0000000000..799ce9ef44
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/masked_sad4d_ssse3.c
@@ -0,0 +1,266 @@
+/*
+ * Copyright (c) 2020, 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"
+
+#define MASK_SAD16XH_ONE_REF(idx)                             \
+  a = _mm_loadu_si128((const __m128i *)&ref##idx[x]);         \
+  data_l = _mm_unpacklo_epi8(a, b);                           \
+  mask_l = _mm_unpacklo_epi8(m, m_inv);                       \
+  pred_l = _mm_maddubs_epi16(data_l, mask_l);                 \
+  pred_l = xx_roundn_epu16(pred_l, AOM_BLEND_A64_ROUND_BITS); \
+                                                              \
+  data_r = _mm_unpackhi_epi8(a, b);                           \
+  mask_r = _mm_unpackhi_epi8(m, m_inv);                       \
+  pred_r = _mm_maddubs_epi16(data_r, mask_r);                 \
+  pred_r = xx_roundn_epu16(pred_r, AOM_BLEND_A64_ROUND_BITS); \
+                                                              \
+  pred = _mm_packus_epi16(pred_l, pred_r);                    \
+  res##idx = _mm_add_epi32(res##idx, _mm_sad_epu8(pred, src));
+
+static INLINE void masked_sadx4d_ssse3(const uint8_t *src_ptr, int src_stride,
+                                       const uint8_t *a_ptr[4], int a_stride,
+                                       const uint8_t *b_ptr, int b_stride,
+                                       const uint8_t *m_ptr, int m_stride,
+                                       int width, int height, int inv_mask,
+                                       unsigned sad_array[4]) {
+  int x, y;
+  __m128i a;
+  __m128i data_l, data_r, mask_l, mask_r, pred_l, pred_r, pred;
+  const __m128i mask_max = _mm_set1_epi8((1 << AOM_BLEND_A64_ROUND_BITS));
+  __m128i res0 = _mm_setzero_si128();
+  __m128i res1 = _mm_setzero_si128();
+  __m128i res2 = _mm_setzero_si128();
+  __m128i res3 = _mm_setzero_si128();
+  const uint8_t *ref0 = a_ptr[0];
+  const uint8_t *ref1 = a_ptr[1];
+  const uint8_t *ref2 = a_ptr[2];
+  const uint8_t *ref3 = a_ptr[3];
+
+  for (y = 0; y < height; y++) {
+    for (x = 0; x < width; x += 16) {
+      const __m128i src = _mm_loadu_si128((const __m128i *)&src_ptr[x]);
+      const __m128i b = _mm_loadu_si128((const __m128i *)&b_ptr[x]);
+      const __m128i m_copy = _mm_loadu_si128((const __m128i *)&m_ptr[x]);
+      __m128i m_inv = _mm_sub_epi8(mask_max, m_copy);
+      __m128i m = inv_mask ? m_inv : m_copy;
+      m_inv = inv_mask ? m_copy : m_inv;
+
+      MASK_SAD16XH_ONE_REF(0)
+      MASK_SAD16XH_ONE_REF(1)
+      MASK_SAD16XH_ONE_REF(2)
+      MASK_SAD16XH_ONE_REF(3)
+    }
+
+    src_ptr += src_stride;
+    ref0 += a_stride;
+    ref1 += a_stride;
+    ref2 += a_stride;
+    ref3 += a_stride;
+    b_ptr += b_stride;
+    m_ptr += m_stride;
+  }
+  res0 = _mm_add_epi32(_mm_unpacklo_epi32(res0, res1),
+                       _mm_unpackhi_epi32(res0, res1));
+  res2 = _mm_add_epi32(_mm_unpacklo_epi32(res2, res3),
+                       _mm_unpackhi_epi32(res2, res3));
+
+  res0 = _mm_unpacklo_epi64(res0, res2);
+  _mm_storeu_si128((__m128i *)sad_array, res0);
+}
+
+#define MASK_SAD8XH_ONE_REF(idx)                                               \
+  const __m128i a##idx##0 = _mm_loadl_epi64((__m128i *)ref##idx);              \
+  const __m128i a##idx##1 = _mm_loadl_epi64((__m128i *)(ref##idx + a_stride)); \
+  data_l = _mm_unpacklo_epi8(a##idx##0, b0);                                   \
+  mask_l = _mm_unpacklo_epi8(m, m_inv);                                        \
+  pred_l = _mm_maddubs_epi16(data_l, mask_l);                                  \
+  pred_l = xx_roundn_epu16(pred_l, AOM_BLEND_A64_ROUND_BITS);                  \
+                                                                               \
+  data_r = _mm_unpacklo_epi8(a##idx##1, b1);                                   \
+  mask_r = _mm_unpackhi_epi8(m, m_inv);                                        \
+  pred_r = _mm_maddubs_epi16(data_r, mask_r);                                  \
+  pred_r = xx_roundn_epu16(pred_r, AOM_BLEND_A64_ROUND_BITS);                  \
+                                                                               \
+  pred = _mm_packus_epi16(pred_l, pred_r);                                     \
+  res##idx = _mm_add_epi32(res##idx, _mm_sad_epu8(pred, src));
+
+void aom_masked_sad8xhx4d_ssse3(const uint8_t *src_ptr, int src_stride,
+                                const uint8_t *ref_array[4], int a_stride,
+                                const uint8_t *b_ptr, int b_stride,
+                                const uint8_t *m_ptr, int m_stride, int height,
+                                int inv_mask, unsigned sad_array[4]) {
+  const uint8_t *ref0 = ref_array[0];
+  const uint8_t *ref1 = ref_array[1];
+  const uint8_t *ref2 = ref_array[2];
+  const uint8_t *ref3 = ref_array[3];
+  __m128i data_l, data_r, pred_l, pred_r, mask_l, mask_r, pred;
+  __m128i res0 = _mm_setzero_si128();
+  __m128i res1 = _mm_setzero_si128();
+  __m128i res2 = _mm_setzero_si128();
+  __m128i res3 = _mm_setzero_si128();
+  const __m128i mask_max = _mm_set1_epi8((1 << AOM_BLEND_A64_ROUND_BITS));
+
+  for (int y = 0; y < height; y += 2) {
+    const __m128i src = _mm_unpacklo_epi64(
+        _mm_loadl_epi64((const __m128i *)src_ptr),
+        _mm_loadl_epi64((const __m128i *)(src_ptr + src_stride)));
+    const __m128i b0 = _mm_loadl_epi64((__m128i *)b_ptr);
+    const __m128i b1 = _mm_loadl_epi64((__m128i *)(b_ptr + b_stride));
+    const __m128i m0 = _mm_loadl_epi64((__m128i *)m_ptr);
+    const __m128i m1 = _mm_loadl_epi64((__m128i *)(m_ptr + m_stride));
+    __m128i m_copy = _mm_unpacklo_epi64(m0, m1);
+    __m128i m_inv = _mm_sub_epi8(mask_max, m_copy);
+    __m128i m = inv_mask ? m_inv : m_copy;
+    m_inv = inv_mask ? m_copy : m_inv;
+
+    MASK_SAD8XH_ONE_REF(0)
+    MASK_SAD8XH_ONE_REF(1)
+    MASK_SAD8XH_ONE_REF(2)
+    MASK_SAD8XH_ONE_REF(3)
+
+    ref0 += 2 * a_stride;
+    ref1 += 2 * a_stride;
+    ref2 += 2 * a_stride;
+    ref3 += 2 * a_stride;
+    src_ptr += 2 * src_stride;
+    b_ptr += 2 * b_stride;
+    m_ptr += 2 * m_stride;
+  }
+  res0 = _mm_add_epi32(_mm_unpacklo_epi32(res0, res1),
+                       _mm_unpackhi_epi32(res0, res1));
+  res2 = _mm_add_epi32(_mm_unpacklo_epi32(res2, res3),
+                       _mm_unpackhi_epi32(res2, res3));
+  res0 = _mm_unpacklo_epi64(res0, res2);
+  _mm_storeu_si128((__m128i *)sad_array, res0);
+}
+
+#define MASK_SAD4XH_ONE_REF(idx)                                          \
+  a = _mm_unpacklo_epi32(_mm_cvtsi32_si128(*(int *)ref##idx),             \
+                         _mm_cvtsi32_si128(*(int *)&ref##idx[a_stride])); \
+  data = _mm_unpacklo_epi8(a, b);                                         \
+  mask = _mm_unpacklo_epi8(m, m_inv);                                     \
+  pred = _mm_maddubs_epi16(data, mask);                                   \
+  pred = xx_roundn_epu16(pred, AOM_BLEND_A64_ROUND_BITS);                 \
+                                                                          \
+  pred = _mm_packus_epi16(pred, _mm_setzero_si128());                     \
+  res##idx = _mm_add_epi32(res##idx, _mm_sad_epu8(pred, src));
+
+void aom_masked_sad4xhx4d_ssse3(const uint8_t *src_ptr, int src_stride,
+                                const uint8_t *ref_array[4], int a_stride,
+                                const uint8_t *b_ptr, int b_stride,
+                                const uint8_t *m_ptr, int m_stride, int height,
+                                int inv_mask, unsigned sad_array[4]) {
+  const uint8_t *ref0 = ref_array[0];
+  const uint8_t *ref1 = ref_array[1];
+  const uint8_t *ref2 = ref_array[2];
+  const uint8_t *ref3 = ref_array[3];
+  __m128i data, pred, mask;
+  __m128i res0 = _mm_setzero_si128();
+  __m128i res1 = _mm_setzero_si128();
+  __m128i res2 = _mm_setzero_si128();
+  __m128i res3 = _mm_setzero_si128();
+  __m128i a;
+  const __m128i mask_max = _mm_set1_epi8((1 << AOM_BLEND_A64_ROUND_BITS));
+
+  for (int y = 0; y < height; y += 2) {
+    const __m128i src =
+        _mm_unpacklo_epi32(_mm_cvtsi32_si128(*(int *)src_ptr),
+                           _mm_cvtsi32_si128(*(int *)&src_ptr[src_stride]));
+    const __m128i b =
+        _mm_unpacklo_epi32(_mm_cvtsi32_si128(*(int *)b_ptr),
+                           _mm_cvtsi32_si128(*(int *)&b_ptr[b_stride]));
+    const __m128i m_copy =
+        _mm_unpacklo_epi32(_mm_cvtsi32_si128(*(int *)m_ptr),
+                           _mm_cvtsi32_si128(*(int *)&m_ptr[m_stride]));
+
+    __m128i m_inv = _mm_sub_epi8(mask_max, m_copy);
+    __m128i m = inv_mask ? m_inv : m_copy;
+    m_inv = inv_mask ? m_copy : m_inv;
+
+    MASK_SAD4XH_ONE_REF(0)
+    MASK_SAD4XH_ONE_REF(1)
+    MASK_SAD4XH_ONE_REF(2)
+    MASK_SAD4XH_ONE_REF(3)
+
+    ref0 += 2 * a_stride;
+    ref1 += 2 * a_stride;
+    ref2 += 2 * a_stride;
+    ref3 += 2 * a_stride;
+    src_ptr += 2 * src_stride;
+    b_ptr += 2 * b_stride;
+    m_ptr += 2 * m_stride;
+  }
+  res0 = _mm_unpacklo_epi32(res0, res1);
+  res2 = _mm_unpacklo_epi32(res2, res3);
+  res0 = _mm_unpacklo_epi64(res0, res2);
+  _mm_storeu_si128((__m128i *)sad_array, res0);
+}
+
+#define MASKSADMXN_SSSE3(m, n)                                                 \
+  void aom_masked_sad##m##x##n##x4d_ssse3(                                     \
+      const uint8_t *src, int src_stride, const uint8_t *ref[4],               \
+      int ref_stride, const uint8_t *second_pred, const uint8_t *msk,          \
+      int msk_stride, int inv_mask, unsigned sad_array[4]) {                   \
+    masked_sadx4d_ssse3(src, src_stride, ref, ref_stride, second_pred, m, msk, \
+                        msk_stride, m, n, inv_mask, sad_array);                \
+  }
+
+#define MASKSAD8XN_SSSE3(n)                                                   \
+  void aom_masked_sad8x##n##x4d_ssse3(                                        \
+      const uint8_t *src, int src_stride, const uint8_t *ref[4],              \
+      int ref_stride, const uint8_t *second_pred, const uint8_t *msk,         \
+      int msk_stride, int inv_mask, unsigned sad_array[4]) {                  \
+    aom_masked_sad8xhx4d_ssse3(src, src_stride, ref, ref_stride, second_pred, \
+                               8, msk, msk_stride, n, inv_mask, sad_array);   \
+  }
+
+#define MASKSAD4XN_SSSE3(n)                                                   \
+  void aom_masked_sad4x##n##x4d_ssse3(                                        \
+      const uint8_t *src, int src_stride, const uint8_t *ref[4],              \
+      int ref_stride, const uint8_t *second_pred, const uint8_t *msk,         \
+      int msk_stride, int inv_mask, unsigned sad_array[4]) {                  \
+    aom_masked_sad4xhx4d_ssse3(src, src_stride, ref, ref_stride, second_pred, \
+                               4, msk, msk_stride, n, inv_mask, sad_array);   \
+  }
+
+MASKSADMXN_SSSE3(128, 128)
+MASKSADMXN_SSSE3(128, 64)
+MASKSADMXN_SSSE3(64, 128)
+MASKSADMXN_SSSE3(64, 64)
+MASKSADMXN_SSSE3(64, 32)
+MASKSADMXN_SSSE3(32, 64)
+MASKSADMXN_SSSE3(32, 32)
+MASKSADMXN_SSSE3(32, 16)
+MASKSADMXN_SSSE3(16, 32)
+MASKSADMXN_SSSE3(16, 16)
+MASKSADMXN_SSSE3(16, 8)
+MASKSAD8XN_SSSE3(16)
+MASKSAD8XN_SSSE3(8)
+MASKSAD8XN_SSSE3(4)
+MASKSAD4XN_SSSE3(8)
+MASKSAD4XN_SSSE3(4)
+MASKSAD4XN_SSSE3(16)
+MASKSADMXN_SSSE3(16, 4)
+MASKSAD8XN_SSSE3(32)
+MASKSADMXN_SSSE3(32, 8)
+MASKSADMXN_SSSE3(16, 64)
+MASKSADMXN_SSSE3(64, 16)
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 0000000000..2c022555b5
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/masked_sad_intrin_avx2.c
@@ -0,0 +1,389 @@
+/*
+ * Copyright (c) 2018, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <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;
+}
+
+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;
+}
+
+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;
+}
+
+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;
+}
+
+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
new file mode 100644
index 0000000000..df3a8764e3
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/masked_sad_intrin_ssse3.c
@@ -0,0 +1,400 @@
+/*
+ * Copyright (c) 2017, 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"
+
+// For width a multiple of 16
+static INLINE unsigned int masked_sad_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 width, 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, \
+      const uint8_t *second_pred, const uint8_t *msk, int msk_stride,         \
+      int invert_mask) {                                                      \
+    if (!invert_mask)                                                         \
+      return masked_sad_ssse3(src, src_stride, ref, ref_stride, second_pred,  \
+                              m, msk, msk_stride, m, n);                      \
+    else                                                                      \
+      return masked_sad_ssse3(src, src_stride, second_pred, m, ref,           \
+                              ref_stride, msk, msk_stride, m, n);             \
+  }
+
+#define MASKSAD8XN_SSSE3(n)                                                   \
+  unsigned int aom_masked_sad8x##n##_ssse3(                                   \
+      const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, \
+      const uint8_t *second_pred, const uint8_t *msk, int msk_stride,         \
+      int invert_mask) {                                                      \
+    if (!invert_mask)                                                         \
+      return aom_masked_sad8xh_ssse3(src, src_stride, ref, ref_stride,        \
+                                     second_pred, 8, msk, msk_stride, n);     \
+    else                                                                      \
+      return aom_masked_sad8xh_ssse3(src, src_stride, second_pred, 8, ref,    \
+                                     ref_stride, msk, msk_stride, n);         \
+  }
+
+#define MASKSAD4XN_SSSE3(n)                                                   \
+  unsigned int aom_masked_sad4x##n##_ssse3(                                   \
+      const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, \
+      const uint8_t *second_pred, const uint8_t *msk, int msk_stride,         \
+      int invert_mask) {                                                      \
+    if (!invert_mask)                                                         \
+      return aom_masked_sad4xh_ssse3(src, src_stride, ref, ref_stride,        \
+                                     second_pred, 4, msk, msk_stride, n);     \
+    else                                                                      \
+      return aom_masked_sad4xh_ssse3(src, src_stride, second_pred, 4, ref,    \
+                                     ref_stride, msk, msk_stride, n);         \
+  }
+
+MASKSADMXN_SSSE3(128, 128)
+MASKSADMXN_SSSE3(128, 64)
+MASKSADMXN_SSSE3(64, 128)
+MASKSADMXN_SSSE3(64, 64)
+MASKSADMXN_SSSE3(64, 32)
+MASKSADMXN_SSSE3(32, 64)
+MASKSADMXN_SSSE3(32, 32)
+MASKSADMXN_SSSE3(32, 16)
+MASKSADMXN_SSSE3(16, 32)
+MASKSADMXN_SSSE3(16, 16)
+MASKSADMXN_SSSE3(16, 8)
+MASKSAD8XN_SSSE3(16)
+MASKSAD8XN_SSSE3(8)
+MASKSAD8XN_SSSE3(4)
+MASKSAD4XN_SSSE3(8)
+MASKSAD4XN_SSSE3(4)
+MASKSAD4XN_SSSE3(16)
+MASKSADMXN_SSSE3(16, 4)
+MASKSAD8XN_SSSE3(32)
+MASKSADMXN_SSSE3(32, 8)
+MASKSADMXN_SSSE3(16, 64)
+MASKSADMXN_SSSE3(64, 16)
+
+static INLINE unsigned int masked_sad_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 width, int height) {
+  int x, y;
+  __m128i res = _mm_setzero_si128();
+  const __m128i mask_max = _mm_set1_epi8((1 << AOM_BLEND_A64_ROUND_BITS));
+
+  for (y = 0; y < height; y++) {
+    for (x = 0; x < width; x += 16) {
+      const __m128i src = _mm_loadu_si128((const __m128i *)&src_ptr[x]);
+      const __m128i a = _mm_loadu_si128((const __m128i *)&a_ptr[x]);
+      const __m128i b = _mm_loadu_si128((const __m128i *)&b_ptr[x]);
+      const __m128i m = _mm_loadu_si128((const __m128i *)&m_ptr[x]);
+      const __m128i m_inv = _mm_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 __m128i data_l = _mm_unpacklo_epi8(a, b);
+      const __m128i mask_l = _mm_unpacklo_epi8(m, m_inv);
+      __m128i pred_l = _mm_maddubs_epi16(data_l, mask_l);
+      pred_l = xx_roundn_epu16(pred_l, AOM_BLEND_A64_ROUND_BITS);
+
+      const __m128i data_r = _mm_unpackhi_epi8(a, b);
+      const __m128i mask_r = _mm_unpackhi_epi8(m, m_inv);
+      __m128i pred_r = _mm_maddubs_epi16(data_r, mask_r);
+      pred_r = xx_roundn_epu16(pred_r, AOM_BLEND_A64_ROUND_BITS);
+
+      const __m128i pred = _mm_packus_epi16(pred_l, pred_r);
+      res = _mm_add_epi32(res, _mm_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'.
+  unsigned int sad = (unsigned int)(_mm_cvtsi128_si32(res) +
+                                    _mm_cvtsi128_si32(_mm_srli_si128(res, 8)));
+  return sad;
+}
+
+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));
+
+  for (y = 0; y < height; y += 2) {
+    const __m128i src = _mm_unpacklo_epi64(
+        _mm_loadl_epi64((const __m128i *)src_ptr),
+        _mm_loadl_epi64((const __m128i *)&src_ptr[src_stride]));
+    const __m128i a0 = _mm_loadl_epi64((const __m128i *)a_ptr);
+    const __m128i a1 = _mm_loadl_epi64((const __m128i *)&a_ptr[a_stride]);
+    const __m128i b0 = _mm_loadl_epi64((const __m128i *)b_ptr);
+    const __m128i b1 = _mm_loadl_epi64((const __m128i *)&b_ptr[b_stride]);
+    const __m128i m =
+        _mm_unpacklo_epi64(_mm_loadl_epi64((const __m128i *)m_ptr),
+                           _mm_loadl_epi64((const __m128i *)&m_ptr[m_stride]));
+    const __m128i m_inv = _mm_sub_epi8(mask_max, m);
+
+    const __m128i data_l = _mm_unpacklo_epi8(a0, b0);
+    const __m128i mask_l = _mm_unpacklo_epi8(m, m_inv);
+    __m128i pred_l = _mm_maddubs_epi16(data_l, mask_l);
+    pred_l = xx_roundn_epu16(pred_l, AOM_BLEND_A64_ROUND_BITS);
+
+    const __m128i data_r = _mm_unpacklo_epi8(a1, b1);
+    const __m128i mask_r = _mm_unpackhi_epi8(m, m_inv);
+    __m128i pred_r = _mm_maddubs_epi16(data_r, mask_r);
+    pred_r = xx_roundn_epu16(pred_r, AOM_BLEND_A64_ROUND_BITS);
+
+    const __m128i pred = _mm_packus_epi16(pred_l, pred_r);
+    res = _mm_add_epi32(res, _mm_sad_epu8(pred, src));
+
+    src_ptr += src_stride * 2;
+    a_ptr += a_stride * 2;
+    b_ptr += b_stride * 2;
+    m_ptr += m_stride * 2;
+  }
+  unsigned int sad = (unsigned int)(_mm_cvtsi128_si32(res) +
+                                    _mm_cvtsi128_si32(_mm_srli_si128(res, 8)));
+  return sad;
+}
+
+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));
+
+  for (y = 0; y < height; y += 2) {
+    // Load two rows at a time, this seems to be a bit faster
+    // than four rows at a time in this case.
+    const __m128i src =
+        _mm_unpacklo_epi32(_mm_cvtsi32_si128(*(int *)src_ptr),
+                           _mm_cvtsi32_si128(*(int *)&src_ptr[src_stride]));
+    const __m128i a =
+        _mm_unpacklo_epi32(_mm_cvtsi32_si128(*(int *)a_ptr),
+                           _mm_cvtsi32_si128(*(int *)&a_ptr[a_stride]));
+    const __m128i b =
+        _mm_unpacklo_epi32(_mm_cvtsi32_si128(*(int *)b_ptr),
+                           _mm_cvtsi32_si128(*(int *)&b_ptr[b_stride]));
+    const __m128i m =
+        _mm_unpacklo_epi32(_mm_cvtsi32_si128(*(int *)m_ptr),
+                           _mm_cvtsi32_si128(*(int *)&m_ptr[m_stride]));
+    const __m128i m_inv = _mm_sub_epi8(mask_max, m);
+
+    const __m128i data = _mm_unpacklo_epi8(a, b);
+    const __m128i mask = _mm_unpacklo_epi8(m, m_inv);
+    __m128i pred_16bit = _mm_maddubs_epi16(data, mask);
+    pred_16bit = xx_roundn_epu16(pred_16bit, AOM_BLEND_A64_ROUND_BITS);
+
+    const __m128i pred = _mm_packus_epi16(pred_16bit, _mm_setzero_si128());
+    res = _mm_add_epi32(res, _mm_sad_epu8(pred, src));
+
+    src_ptr += src_stride * 2;
+    a_ptr += a_stride * 2;
+    b_ptr += b_stride * 2;
+    m_ptr += m_stride * 2;
+  }
+  // At this point, the SAD is stored in lane 0 of 'res'
+  return (unsigned int)_mm_cvtsi128_si32(res);
+}
+
+// For width a multiple of 8
+static INLINE unsigned int highbd_masked_sad_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 width, 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,               \
+      int ref_stride, const uint8_t *second_pred8, const uint8_t *msk,        \
+      int msk_stride, int invert_mask) {                                      \
+    if (!invert_mask)                                                         \
+      return highbd_masked_sad_ssse3(src8, src_stride, ref8, ref_stride,      \
+                                     second_pred8, m, msk, msk_stride, m, n); \
+    else                                                                      \
+      return highbd_masked_sad_ssse3(src8, src_stride, second_pred8, m, ref8, \
+                                     ref_stride, msk, msk_stride, m, n);      \
+  }
+
+#define HIGHBD_MASKSAD4XN_SSSE3(n)                                             \
+  unsigned int aom_highbd_masked_sad4x##n##_ssse3(                             \
+      const uint8_t *src8, int src_stride, const uint8_t *ref8,                \
+      int ref_stride, const uint8_t *second_pred8, const uint8_t *msk,         \
+      int msk_stride, int invert_mask) {                                       \
+    if (!invert_mask)                                                          \
+      return aom_highbd_masked_sad4xh_ssse3(src8, src_stride, ref8,            \
+                                            ref_stride, second_pred8, 4, msk,  \
+                                            msk_stride, n);                    \
+    else                                                                       \
+      return aom_highbd_masked_sad4xh_ssse3(src8, src_stride, second_pred8, 4, \
+                                            ref8, ref_stride, msk, msk_stride, \
+                                            n);                                \
+  }
+
+HIGHBD_MASKSADMXN_SSSE3(128, 128)
+HIGHBD_MASKSADMXN_SSSE3(128, 64)
+HIGHBD_MASKSADMXN_SSSE3(64, 128)
+HIGHBD_MASKSADMXN_SSSE3(64, 64)
+HIGHBD_MASKSADMXN_SSSE3(64, 32)
+HIGHBD_MASKSADMXN_SSSE3(32, 64)
+HIGHBD_MASKSADMXN_SSSE3(32, 32)
+HIGHBD_MASKSADMXN_SSSE3(32, 16)
+HIGHBD_MASKSADMXN_SSSE3(16, 32)
+HIGHBD_MASKSADMXN_SSSE3(16, 16)
+HIGHBD_MASKSADMXN_SSSE3(16, 8)
+HIGHBD_MASKSADMXN_SSSE3(8, 16)
+HIGHBD_MASKSADMXN_SSSE3(8, 8)
+HIGHBD_MASKSADMXN_SSSE3(8, 4)
+HIGHBD_MASKSAD4XN_SSSE3(8)
+HIGHBD_MASKSAD4XN_SSSE3(4)
+HIGHBD_MASKSAD4XN_SSSE3(16)
+HIGHBD_MASKSADMXN_SSSE3(16, 4)
+HIGHBD_MASKSADMXN_SSSE3(8, 32)
+HIGHBD_MASKSADMXN_SSSE3(32, 8)
+HIGHBD_MASKSADMXN_SSSE3(16, 64)
+HIGHBD_MASKSADMXN_SSSE3(64, 16)
+
+static INLINE unsigned int highbd_masked_sad_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 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;
+  __m128i res = _mm_setzero_si128();
+  const __m128i mask_max = _mm_set1_epi16((1 << AOM_BLEND_A64_ROUND_BITS));
+  const __m128i round_const =
+      _mm_set1_epi32((1 << AOM_BLEND_A64_ROUND_BITS) >> 1);
+  const __m128i one = _mm_set1_epi16(1);
+
+  for (y = 0; y < height; y++) {
+    for (x = 0; x < width; x += 8) {
+      const __m128i src = _mm_loadu_si128((const __m128i *)&src_ptr[x]);
+      const __m128i a = _mm_loadu_si128((const __m128i *)&a_ptr[x]);
+      const __m128i b = _mm_loadu_si128((const __m128i *)&b_ptr[x]);
+      // Zero-extend mask to 16 bits
+      const __m128i m = _mm_unpacklo_epi8(
+          _mm_loadl_epi64((const __m128i *)&m_ptr[x]), _mm_setzero_si128());
+      const __m128i m_inv = _mm_sub_epi16(mask_max, m);
+
+      const __m128i data_l = _mm_unpacklo_epi16(a, b);
+      const __m128i mask_l = _mm_unpacklo_epi16(m, m_inv);
+      __m128i pred_l = _mm_madd_epi16(data_l, mask_l);
+      pred_l = _mm_srai_epi32(_mm_add_epi32(pred_l, round_const),
+                              AOM_BLEND_A64_ROUND_BITS);
+
+      const __m128i data_r = _mm_unpackhi_epi16(a, b);
+      const __m128i mask_r = _mm_unpackhi_epi16(m, m_inv);
+      __m128i pred_r = _mm_madd_epi16(data_r, mask_r);
+      pred_r = _mm_srai_epi32(_mm_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 __m128i pred = _mm_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 __m128i diff = _mm_abs_epi16(_mm_sub_epi16(pred, src));
+      res = _mm_add_epi32(res, _mm_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 = _mm_hadd_epi32(res, res);
+  res = _mm_hadd_epi32(res, res);
+  int sad = _mm_cvtsi128_si32(res);
+  return sad;
+}
+
+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);
+  int y;
+  __m128i res = _mm_setzero_si128();
+  const __m128i mask_max = _mm_set1_epi16((1 << AOM_BLEND_A64_ROUND_BITS));
+  const __m128i round_const =
+      _mm_set1_epi32((1 << AOM_BLEND_A64_ROUND_BITS) >> 1);
+  const __m128i one = _mm_set1_epi16(1);
+
+  for (y = 0; y < height; y += 2) {
+    const __m128i src = _mm_unpacklo_epi64(
+        _mm_loadl_epi64((const __m128i *)src_ptr),
+        _mm_loadl_epi64((const __m128i *)&src_ptr[src_stride]));
+    const __m128i a =
+        _mm_unpacklo_epi64(_mm_loadl_epi64((const __m128i *)a_ptr),
+                           _mm_loadl_epi64((const __m128i *)&a_ptr[a_stride]));
+    const __m128i b =
+        _mm_unpacklo_epi64(_mm_loadl_epi64((const __m128i *)b_ptr),
+                           _mm_loadl_epi64((const __m128i *)&b_ptr[b_stride]));
+    // Zero-extend mask to 16 bits
+    const __m128i m = _mm_unpacklo_epi8(
+        _mm_unpacklo_epi32(_mm_cvtsi32_si128(*(const int *)m_ptr),
+                           _mm_cvtsi32_si128(*(const int *)&m_ptr[m_stride])),
+        _mm_setzero_si128());
+    const __m128i m_inv = _mm_sub_epi16(mask_max, m);
+
+    const __m128i data_l = _mm_unpacklo_epi16(a, b);
+    const __m128i mask_l = _mm_unpacklo_epi16(m, m_inv);
+    __m128i pred_l = _mm_madd_epi16(data_l, mask_l);
+    pred_l = _mm_srai_epi32(_mm_add_epi32(pred_l, round_const),
+                            AOM_BLEND_A64_ROUND_BITS);
+
+    const __m128i data_r = _mm_unpackhi_epi16(a, b);
+    const __m128i mask_r = _mm_unpackhi_epi16(m, m_inv);
+    __m128i pred_r = _mm_madd_epi16(data_r, mask_r);
+    pred_r = _mm_srai_epi32(_mm_add_epi32(pred_r, round_const),
+                            AOM_BLEND_A64_ROUND_BITS);
+
+    const __m128i pred = _mm_packs_epi32(pred_l, pred_r);
+    const __m128i diff = _mm_abs_epi16(_mm_sub_epi16(pred, src));
+    res = _mm_add_epi32(res, _mm_madd_epi16(diff, one));
+
+    src_ptr += src_stride * 2;
+    a_ptr += a_stride * 2;
+    b_ptr += b_stride * 2;
+    m_ptr += m_stride * 2;
+  }
+  res = _mm_hadd_epi32(res, res);
+  res = _mm_hadd_epi32(res, res);
+  int sad = _mm_cvtsi128_si32(res);
+  return sad;
+}
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 0000000000..cffbd9672c
--- /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_AOM_DSP_X86_MASKED_SAD_INTRIN_SSSE3_H_
+#define AOM_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  // AOM_AOM_DSP_X86_MASKED_SAD_INTRIN_SSSE3_H_
diff --git a/third_party/aom/aom_dsp/x86/masked_variance_intrin_ssse3.c b/third_party/aom/aom_dsp/x86/masked_variance_intrin_ssse3.c
new file mode 100644
index 0000000000..0bf383fffd
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/masked_variance_intrin_ssse3.c
@@ -0,0 +1,1067 @@
+/*
+ * Copyright (c) 2017, 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 <stdlib.h>
+#include <string.h>
+#include <tmmintrin.h>
+
+#include "config/aom_config.h"
+#include "config/aom_dsp_rtcd.h"
+
+#include "aom/aom_integer.h"
+#include "aom_dsp/aom_filter.h"
+#include "aom_dsp/blend.h"
+#include "aom_dsp/x86/masked_variance_intrin_ssse3.h"
+#include "aom_dsp/x86/synonyms.h"
+#include "aom_ports/mem.h"
+
+// For width a multiple of 16
+static void bilinear_filter(const uint8_t *src, int src_stride, int xoffset,
+                            int yoffset, uint8_t *dst, int w, int h);
+
+static void bilinear_filter8xh(const uint8_t *src, int src_stride, int xoffset,
+                               int yoffset, uint8_t *dst, int h);
+
+static void bilinear_filter4xh(const uint8_t *src, int src_stride, int xoffset,
+                               int yoffset, uint8_t *dst, int h);
+
+// For width a multiple of 16
+static void masked_variance(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, unsigned int *sse, int *sum_);
+
+static void masked_variance8xh(const uint8_t *src_ptr, int src_stride,
+                               const uint8_t *a_ptr, const uint8_t *b_ptr,
+                               const uint8_t *m_ptr, int m_stride, int height,
+                               unsigned int *sse, int *sum_);
+
+static void masked_variance4xh(const uint8_t *src_ptr, int src_stride,
+                               const uint8_t *a_ptr, const uint8_t *b_ptr,
+                               const uint8_t *m_ptr, int m_stride, int height,
+                               unsigned int *sse, int *sum_);
+
+#define MASK_SUBPIX_VAR_SSSE3(W, H)                                   \
+  unsigned int aom_masked_sub_pixel_variance##W##x##H##_ssse3(        \
+      const uint8_t *src, int src_stride, int xoffset, int yoffset,   \
+      const uint8_t *ref, int ref_stride, const uint8_t *second_pred, \
+      const uint8_t *msk, int msk_stride, int invert_mask,            \
+      unsigned int *sse) {                                            \
+    int sum;                                                          \
+    uint8_t temp[(H + 1) * W];                                        \
+                                                                      \
+    bilinear_filter(src, src_stride, xoffset, yoffset, temp, W, H);   \
+                                                                      \
+    if (!invert_mask)                                                 \
+      masked_variance(ref, ref_stride, temp, W, second_pred, W, msk,  \
+                      msk_stride, W, H, sse, &sum);                   \
+    else                                                              \
+      masked_variance(ref, ref_stride, second_pred, W, temp, W, msk,  \
+                      msk_stride, W, H, sse, &sum);                   \
+    return *sse - (uint32_t)(((int64_t)sum * sum) / (W * H));         \
+  }
+
+#define MASK_SUBPIX_VAR8XH_SSSE3(H)                                           \
+  unsigned int aom_masked_sub_pixel_variance8x##H##_ssse3(                    \
+      const uint8_t *src, int src_stride, int xoffset, int yoffset,           \
+      const uint8_t *ref, int ref_stride, const uint8_t *second_pred,         \
+      const uint8_t *msk, int msk_stride, int invert_mask,                    \
+      unsigned int *sse) {                                                    \
+    int sum;                                                                  \
+    uint8_t temp[(H + 1) * 8];                                                \
+                                                                              \
+    bilinear_filter8xh(src, src_stride, xoffset, yoffset, temp, H);           \
+                                                                              \
+    if (!invert_mask)                                                         \
+      masked_variance8xh(ref, ref_stride, temp, second_pred, msk, msk_stride, \
+                         H, sse, &sum);                                       \
+    else                                                                      \
+      masked_variance8xh(ref, ref_stride, second_pred, temp, msk, msk_stride, \
+                         H, sse, &sum);                                       \
+    return *sse - (uint32_t)(((int64_t)sum * sum) / (8 * H));                 \
+  }
+
+#define MASK_SUBPIX_VAR4XH_SSSE3(H)                                           \
+  unsigned int aom_masked_sub_pixel_variance4x##H##_ssse3(                    \
+      const uint8_t *src, int src_stride, int xoffset, int yoffset,           \
+      const uint8_t *ref, int ref_stride, const uint8_t *second_pred,         \
+      const uint8_t *msk, int msk_stride, int invert_mask,                    \
+      unsigned int *sse) {                                                    \
+    int sum;                                                                  \
+    uint8_t temp[(H + 1) * 4];                                                \
+                                                                              \
+    bilinear_filter4xh(src, src_stride, xoffset, yoffset, temp, H);           \
+                                                                              \
+    if (!invert_mask)                                                         \
+      masked_variance4xh(ref, ref_stride, temp, second_pred, msk, msk_stride, \
+                         H, sse, &sum);                                       \
+    else                                                                      \
+      masked_variance4xh(ref, ref_stride, second_pred, temp, msk, msk_stride, \
+                         H, sse, &sum);                                       \
+    return *sse - (uint32_t)(((int64_t)sum * sum) / (4 * H));                 \
+  }
+
+MASK_SUBPIX_VAR_SSSE3(128, 128)
+MASK_SUBPIX_VAR_SSSE3(128, 64)
+MASK_SUBPIX_VAR_SSSE3(64, 128)
+MASK_SUBPIX_VAR_SSSE3(64, 64)
+MASK_SUBPIX_VAR_SSSE3(64, 32)
+MASK_SUBPIX_VAR_SSSE3(32, 64)
+MASK_SUBPIX_VAR_SSSE3(32, 32)
+MASK_SUBPIX_VAR_SSSE3(32, 16)
+MASK_SUBPIX_VAR_SSSE3(16, 32)
+MASK_SUBPIX_VAR_SSSE3(16, 16)
+MASK_SUBPIX_VAR_SSSE3(16, 8)
+MASK_SUBPIX_VAR8XH_SSSE3(16)
+MASK_SUBPIX_VAR8XH_SSSE3(8)
+MASK_SUBPIX_VAR8XH_SSSE3(4)
+MASK_SUBPIX_VAR4XH_SSSE3(8)
+MASK_SUBPIX_VAR4XH_SSSE3(4)
+MASK_SUBPIX_VAR4XH_SSSE3(16)
+MASK_SUBPIX_VAR_SSSE3(16, 4)
+MASK_SUBPIX_VAR8XH_SSSE3(32)
+MASK_SUBPIX_VAR_SSSE3(32, 8)
+MASK_SUBPIX_VAR_SSSE3(64, 16)
+MASK_SUBPIX_VAR_SSSE3(16, 64)
+
+static INLINE __m128i filter_block(const __m128i a, const __m128i b,
+                                   const __m128i filter) {
+  __m128i v0 = _mm_unpacklo_epi8(a, b);
+  v0 = _mm_maddubs_epi16(v0, filter);
+  v0 = xx_roundn_epu16(v0, FILTER_BITS);
+
+  __m128i v1 = _mm_unpackhi_epi8(a, b);
+  v1 = _mm_maddubs_epi16(v1, filter);
+  v1 = xx_roundn_epu16(v1, FILTER_BITS);
+
+  return _mm_packus_epi16(v0, v1);
+}
+
+static void bilinear_filter(const uint8_t *src, int src_stride, int xoffset,
+                            int yoffset, uint8_t *dst, int w, int h) {
+  int i, j;
+  // Horizontal filter
+  if (xoffset == 0) {
+    uint8_t *b = dst;
+    for (i = 0; i < h + 1; ++i) {
+      for (j = 0; j < w; j += 16) {
+        __m128i x = _mm_loadu_si128((__m128i *)&src[j]);
+        _mm_storeu_si128((__m128i *)&b[j], x);
+      }
+      src += src_stride;
+      b += w;
+    }
+  } else if (xoffset == 4) {
+    uint8_t *b = dst;
+    for (i = 0; i < h + 1; ++i) {
+      for (j = 0; j < w; j += 16) {
+        __m128i x = _mm_loadu_si128((__m128i *)&src[j]);
+        __m128i y = _mm_loadu_si128((__m128i *)&src[j + 16]);
+        __m128i z = _mm_alignr_epi8(y, x, 1);
+        _mm_storeu_si128((__m128i *)&b[j], _mm_avg_epu8(x, z));
+      }
+      src += src_stride;
+      b += w;
+    }
+  } else {
+    uint8_t *b = dst;
+    const uint8_t *hfilter = bilinear_filters_2t[xoffset];
+    const __m128i hfilter_vec = _mm_set1_epi16(hfilter[0] | (hfilter[1] << 8));
+    for (i = 0; i < h + 1; ++i) {
+      for (j = 0; j < w; j += 16) {
+        const __m128i x = _mm_loadu_si128((__m128i *)&src[j]);
+        const __m128i y = _mm_loadu_si128((__m128i *)&src[j + 16]);
+        const __m128i z = _mm_alignr_epi8(y, x, 1);
+        const __m128i res = filter_block(x, z, hfilter_vec);
+        _mm_storeu_si128((__m128i *)&b[j], res);
+      }
+
+      src += src_stride;
+      b += w;
+    }
+  }
+
+  // Vertical filter
+  if (yoffset == 0) {
+    // The data is already in 'dst', so no need to filter
+  } else if (yoffset == 4) {
+    for (i = 0; i < h; ++i) {
+      for (j = 0; j < w; j += 16) {
+        __m128i x = _mm_loadu_si128((__m128i *)&dst[j]);
+        __m128i y = _mm_loadu_si128((__m128i *)&dst[j + w]);
+        _mm_storeu_si128((__m128i *)&dst[j], _mm_avg_epu8(x, y));
+      }
+      dst += w;
+    }
+  } else {
+    const uint8_t *vfilter = bilinear_filters_2t[yoffset];
+    const __m128i vfilter_vec = _mm_set1_epi16(vfilter[0] | (vfilter[1] << 8));
+    for (i = 0; i < h; ++i) {
+      for (j = 0; j < w; j += 16) {
+        const __m128i x = _mm_loadu_si128((__m128i *)&dst[j]);
+        const __m128i y = _mm_loadu_si128((__m128i *)&dst[j + w]);
+        const __m128i res = filter_block(x, y, vfilter_vec);
+        _mm_storeu_si128((__m128i *)&dst[j], res);
+      }
+
+      dst += w;
+    }
+  }
+}
+
+static INLINE __m128i filter_block_2rows(const __m128i *a0, const __m128i *b0,
+                                         const __m128i *a1, const __m128i *b1,
+                                         const __m128i *filter) {
+  __m128i v0 = _mm_unpacklo_epi8(*a0, *b0);
+  v0 = _mm_maddubs_epi16(v0, *filter);
+  v0 = xx_roundn_epu16(v0, FILTER_BITS);
+
+  __m128i v1 = _mm_unpacklo_epi8(*a1, *b1);
+  v1 = _mm_maddubs_epi16(v1, *filter);
+  v1 = xx_roundn_epu16(v1, FILTER_BITS);
+
+  return _mm_packus_epi16(v0, v1);
+}
+
+static void bilinear_filter8xh(const uint8_t *src, int src_stride, int xoffset,
+                               int yoffset, uint8_t *dst, int h) {
+  int i;
+  // Horizontal filter
+  if (xoffset == 0) {
+    uint8_t *b = dst;
+    for (i = 0; i < h + 1; ++i) {
+      __m128i x = _mm_loadl_epi64((__m128i *)src);
+      _mm_storel_epi64((__m128i *)b, x);
+      src += src_stride;
+      b += 8;
+    }
+  } else if (xoffset == 4) {
+    uint8_t *b = dst;
+    for (i = 0; i < h + 1; ++i) {
+      __m128i x = _mm_loadu_si128((__m128i *)src);
+      __m128i z = _mm_srli_si128(x, 1);
+      _mm_storel_epi64((__m128i *)b, _mm_avg_epu8(x, z));
+      src += src_stride;
+      b += 8;
+    }
+  } else {
+    uint8_t *b = dst;
+    const uint8_t *hfilter = bilinear_filters_2t[xoffset];
+    const __m128i hfilter_vec = _mm_set1_epi16(hfilter[0] | (hfilter[1] << 8));
+    for (i = 0; i < h; i += 2) {
+      const __m128i x0 = _mm_loadu_si128((__m128i *)src);
+      const __m128i z0 = _mm_srli_si128(x0, 1);
+      const __m128i x1 = _mm_loadu_si128((__m128i *)&src[src_stride]);
+      const __m128i z1 = _mm_srli_si128(x1, 1);
+      const __m128i res = filter_block_2rows(&x0, &z0, &x1, &z1, &hfilter_vec);
+      _mm_storeu_si128((__m128i *)b, res);
+
+      src += src_stride * 2;
+      b += 16;
+    }
+    // Handle i = h separately
+    const __m128i x0 = _mm_loadu_si128((__m128i *)src);
+    const __m128i z0 = _mm_srli_si128(x0, 1);
+
+    __m128i v0 = _mm_unpacklo_epi8(x0, z0);
+    v0 = _mm_maddubs_epi16(v0, hfilter_vec);
+    v0 = xx_roundn_epu16(v0, FILTER_BITS);
+
+    _mm_storel_epi64((__m128i *)b, _mm_packus_epi16(v0, v0));
+  }
+
+  // Vertical filter
+  if (yoffset == 0) {
+    // The data is already in 'dst', so no need to filter
+  } else if (yoffset == 4) {
+    for (i = 0; i < h; ++i) {
+      __m128i x = _mm_loadl_epi64((__m128i *)dst);
+      __m128i y = _mm_loadl_epi64((__m128i *)&dst[8]);
+      _mm_storel_epi64((__m128i *)dst, _mm_avg_epu8(x, y));
+      dst += 8;
+    }
+  } else {
+    const uint8_t *vfilter = bilinear_filters_2t[yoffset];
+    const __m128i vfilter_vec = _mm_set1_epi16(vfilter[0] | (vfilter[1] << 8));
+    for (i = 0; i < h; i += 2) {
+      const __m128i x = _mm_loadl_epi64((__m128i *)dst);
+      const __m128i y = _mm_loadl_epi64((__m128i *)&dst[8]);
+      const __m128i z = _mm_loadl_epi64((__m128i *)&dst[16]);
+      const __m128i res = filter_block_2rows(&x, &y, &y, &z, &vfilter_vec);
+      _mm_storeu_si128((__m128i *)dst, res);
+
+      dst += 16;
+    }
+  }
+}
+
+static void bilinear_filter4xh(const uint8_t *src, int src_stride, int xoffset,
+                               int yoffset, uint8_t *dst, int h) {
+  int i;
+  // Horizontal filter
+  if (xoffset == 0) {
+    uint8_t *b = dst;
+    for (i = 0; i < h + 1; ++i) {
+      __m128i x = xx_loadl_32((__m128i *)src);
+      xx_storel_32(b, x);
+      src += src_stride;
+      b += 4;
+    }
+  } else if (xoffset == 4) {
+    uint8_t *b = dst;
+    for (i = 0; i < h + 1; ++i) {
+      __m128i x = _mm_loadl_epi64((__m128i *)src);
+      __m128i z = _mm_srli_si128(x, 1);
+      xx_storel_32(b, _mm_avg_epu8(x, z));
+      src += src_stride;
+      b += 4;
+    }
+  } else {
+    uint8_t *b = dst;
+    const uint8_t *hfilter = bilinear_filters_2t[xoffset];
+    const __m128i hfilter_vec = _mm_set1_epi16(hfilter[0] | (hfilter[1] << 8));
+    for (i = 0; i < h; i += 4) {
+      const __m128i x0 = _mm_loadl_epi64((__m128i *)src);
+      const __m128i z0 = _mm_srli_si128(x0, 1);
+      const __m128i x1 = _mm_loadl_epi64((__m128i *)&src[src_stride]);
+      const __m128i z1 = _mm_srli_si128(x1, 1);
+      const __m128i x2 = _mm_loadl_epi64((__m128i *)&src[src_stride * 2]);
+      const __m128i z2 = _mm_srli_si128(x2, 1);
+      const __m128i x3 = _mm_loadl_epi64((__m128i *)&src[src_stride * 3]);
+      const __m128i z3 = _mm_srli_si128(x3, 1);
+
+      const __m128i a0 = _mm_unpacklo_epi32(x0, x1);
+      const __m128i b0 = _mm_unpacklo_epi32(z0, z1);
+      const __m128i a1 = _mm_unpacklo_epi32(x2, x3);
+      const __m128i b1 = _mm_unpacklo_epi32(z2, z3);
+      const __m128i res = filter_block_2rows(&a0, &b0, &a1, &b1, &hfilter_vec);
+      _mm_storeu_si128((__m128i *)b, res);
+
+      src += src_stride * 4;
+      b += 16;
+    }
+    // Handle i = h separately
+    const __m128i x = _mm_loadl_epi64((__m128i *)src);
+    const __m128i z = _mm_srli_si128(x, 1);
+
+    __m128i v0 = _mm_unpacklo_epi8(x, z);
+    v0 = _mm_maddubs_epi16(v0, hfilter_vec);
+    v0 = xx_roundn_epu16(v0, FILTER_BITS);
+
+    xx_storel_32(b, _mm_packus_epi16(v0, v0));
+  }
+
+  // Vertical filter
+  if (yoffset == 0) {
+    // The data is already in 'dst', so no need to filter
+  } else if (yoffset == 4) {
+    for (i = 0; i < h; ++i) {
+      __m128i x = xx_loadl_32((__m128i *)dst);
+      __m128i y = xx_loadl_32((__m128i *)&dst[4]);
+      xx_storel_32(dst, _mm_avg_epu8(x, y));
+      dst += 4;
+    }
+  } else {
+    const uint8_t *vfilter = bilinear_filters_2t[yoffset];
+    const __m128i vfilter_vec = _mm_set1_epi16(vfilter[0] | (vfilter[1] << 8));
+    for (i = 0; i < h; i += 4) {
+      const __m128i a = xx_loadl_32((__m128i *)dst);
+      const __m128i b = xx_loadl_32((__m128i *)&dst[4]);
+      const __m128i c = xx_loadl_32((__m128i *)&dst[8]);
+      const __m128i d = xx_loadl_32((__m128i *)&dst[12]);
+      const __m128i e = xx_loadl_32((__m128i *)&dst[16]);
+
+      const __m128i a0 = _mm_unpacklo_epi32(a, b);
+      const __m128i b0 = _mm_unpacklo_epi32(b, c);
+      const __m128i a1 = _mm_unpacklo_epi32(c, d);
+      const __m128i b1 = _mm_unpacklo_epi32(d, e);
+      const __m128i res = filter_block_2rows(&a0, &b0, &a1, &b1, &vfilter_vec);
+      _mm_storeu_si128((__m128i *)dst, res);
+
+      dst += 16;
+    }
+  }
+}
+
+static INLINE void accumulate_block(const __m128i *src, const __m128i *a,
+                                    const __m128i *b, const __m128i *m,
+                                    __m128i *sum, __m128i *sum_sq) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i one = _mm_set1_epi16(1);
+  const __m128i mask_max = _mm_set1_epi8((1 << AOM_BLEND_A64_ROUND_BITS));
+  const __m128i m_inv = _mm_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 __m128i data_l = _mm_unpacklo_epi8(*a, *b);
+  const __m128i mask_l = _mm_unpacklo_epi8(*m, m_inv);
+  __m128i pred_l = _mm_maddubs_epi16(data_l, mask_l);
+  pred_l = xx_roundn_epu16(pred_l, AOM_BLEND_A64_ROUND_BITS);
+
+  const __m128i data_r = _mm_unpackhi_epi8(*a, *b);
+  const __m128i mask_r = _mm_unpackhi_epi8(*m, m_inv);
+  __m128i pred_r = _mm_maddubs_epi16(data_r, mask_r);
+  pred_r = xx_roundn_epu16(pred_r, AOM_BLEND_A64_ROUND_BITS);
+
+  const __m128i src_l = _mm_unpacklo_epi8(*src, zero);
+  const __m128i src_r = _mm_unpackhi_epi8(*src, zero);
+  const __m128i diff_l = _mm_sub_epi16(pred_l, src_l);
+  const __m128i diff_r = _mm_sub_epi16(pred_r, src_r);
+
+  // Update partial sums and partial sums of squares
+  *sum =
+      _mm_add_epi32(*sum, _mm_madd_epi16(_mm_add_epi16(diff_l, diff_r), one));
+  *sum_sq =
+      _mm_add_epi32(*sum_sq, _mm_add_epi32(_mm_madd_epi16(diff_l, diff_l),
+                                           _mm_madd_epi16(diff_r, diff_r)));
+}
+
+static void masked_variance(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, unsigned int *sse, int *sum_) {
+  int x, y;
+  __m128i sum = _mm_setzero_si128(), sum_sq = _mm_setzero_si128();
+
+  for (y = 0; y < height; y++) {
+    for (x = 0; x < width; x += 16) {
+      const __m128i src = _mm_loadu_si128((const __m128i *)&src_ptr[x]);
+      const __m128i a = _mm_loadu_si128((const __m128i *)&a_ptr[x]);
+      const __m128i b = _mm_loadu_si128((const __m128i *)&b_ptr[x]);
+      const __m128i m = _mm_loadu_si128((const __m128i *)&m_ptr[x]);
+      accumulate_block(&src, &a, &b, &m, &sum, &sum_sq);
+    }
+
+    src_ptr += src_stride;
+    a_ptr += a_stride;
+    b_ptr += b_stride;
+    m_ptr += m_stride;
+  }
+  // Reduce down to a single sum and sum of squares
+  sum = _mm_hadd_epi32(sum, sum_sq);
+  sum = _mm_hadd_epi32(sum, sum);
+  *sum_ = _mm_cvtsi128_si32(sum);
+  *sse = (unsigned int)_mm_cvtsi128_si32(_mm_srli_si128(sum, 4));
+}
+
+static void masked_variance8xh(const uint8_t *src_ptr, int src_stride,
+                               const uint8_t *a_ptr, const uint8_t *b_ptr,
+                               const uint8_t *m_ptr, int m_stride, int height,
+                               unsigned int *sse, int *sum_) {
+  int y;
+  __m128i sum = _mm_setzero_si128(), sum_sq = _mm_setzero_si128();
+
+  for (y = 0; y < height; y += 2) {
+    __m128i src = _mm_unpacklo_epi64(
+        _mm_loadl_epi64((const __m128i *)src_ptr),
+        _mm_loadl_epi64((const __m128i *)&src_ptr[src_stride]));
+    const __m128i a = _mm_loadu_si128((const __m128i *)a_ptr);
+    const __m128i b = _mm_loadu_si128((const __m128i *)b_ptr);
+    const __m128i m =
+        _mm_unpacklo_epi64(_mm_loadl_epi64((const __m128i *)m_ptr),
+                           _mm_loadl_epi64((const __m128i *)&m_ptr[m_stride]));
+    accumulate_block(&src, &a, &b, &m, &sum, &sum_sq);
+
+    src_ptr += src_stride * 2;
+    a_ptr += 16;
+    b_ptr += 16;
+    m_ptr += m_stride * 2;
+  }
+  // Reduce down to a single sum and sum of squares
+  sum = _mm_hadd_epi32(sum, sum_sq);
+  sum = _mm_hadd_epi32(sum, sum);
+  *sum_ = _mm_cvtsi128_si32(sum);
+  *sse = (unsigned int)_mm_cvtsi128_si32(_mm_srli_si128(sum, 4));
+}
+
+static void masked_variance4xh(const uint8_t *src_ptr, int src_stride,
+                               const uint8_t *a_ptr, const uint8_t *b_ptr,
+                               const uint8_t *m_ptr, int m_stride, int height,
+                               unsigned int *sse, int *sum_) {
+  int y;
+  __m128i sum = _mm_setzero_si128(), sum_sq = _mm_setzero_si128();
+
+  for (y = 0; y < height; y += 4) {
+    // Load four rows at a time
+    __m128i src = _mm_setr_epi32(*(int *)src_ptr, *(int *)&src_ptr[src_stride],
+                                 *(int *)&src_ptr[src_stride * 2],
+                                 *(int *)&src_ptr[src_stride * 3]);
+    const __m128i a = _mm_loadu_si128((const __m128i *)a_ptr);
+    const __m128i b = _mm_loadu_si128((const __m128i *)b_ptr);
+    const __m128i m = _mm_setr_epi32(*(int *)m_ptr, *(int *)&m_ptr[m_stride],
+                                     *(int *)&m_ptr[m_stride * 2],
+                                     *(int *)&m_ptr[m_stride * 3]);
+    accumulate_block(&src, &a, &b, &m, &sum, &sum_sq);
+
+    src_ptr += src_stride * 4;
+    a_ptr += 16;
+    b_ptr += 16;
+    m_ptr += m_stride * 4;
+  }
+  // Reduce down to a single sum and sum of squares
+  sum = _mm_hadd_epi32(sum, sum_sq);
+  sum = _mm_hadd_epi32(sum, sum);
+  *sum_ = _mm_cvtsi128_si32(sum);
+  *sse = (unsigned int)_mm_cvtsi128_si32(_mm_srli_si128(sum, 4));
+}
+
+#if CONFIG_AV1_HIGHBITDEPTH
+// For width a multiple of 8
+static void highbd_bilinear_filter(const uint16_t *src, int src_stride,
+                                   int xoffset, int yoffset, uint16_t *dst,
+                                   int w, int h);
+
+static void highbd_bilinear_filter4xh(const uint16_t *src, int src_stride,
+                                      int xoffset, int yoffset, uint16_t *dst,
+                                      int h);
+
+// For width a multiple of 8
+static void highbd_masked_variance(const uint16_t *src_ptr, int src_stride,
+                                   const uint16_t *a_ptr, int a_stride,
+                                   const uint16_t *b_ptr, int b_stride,
+                                   const uint8_t *m_ptr, int m_stride,
+                                   int width, int height, uint64_t *sse,
+                                   int *sum_);
+
+static void highbd_masked_variance4xh(const uint16_t *src_ptr, int src_stride,
+                                      const uint16_t *a_ptr,
+                                      const uint16_t *b_ptr,
+                                      const uint8_t *m_ptr, int m_stride,
+                                      int height, int *sse, int *sum_);
+
+#define HIGHBD_MASK_SUBPIX_VAR_SSSE3(W, H)                                  \
+  unsigned int aom_highbd_8_masked_sub_pixel_variance##W##x##H##_ssse3(     \
+      const uint8_t *src8, int src_stride, int xoffset, int yoffset,        \
+      const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8,     \
+      const uint8_t *msk, int msk_stride, int invert_mask, uint32_t *sse) { \
+    uint64_t sse64;                                                         \
+    int sum;                                                                \
+    uint16_t temp[(H + 1) * W];                                             \
+    const uint16_t *src = CONVERT_TO_SHORTPTR(src8);                        \
+    const uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);                        \
+    const uint16_t *second_pred = CONVERT_TO_SHORTPTR(second_pred8);        \
+                                                                            \
+    highbd_bilinear_filter(src, src_stride, xoffset, yoffset, temp, W, H);  \
+                                                                            \
+    if (!invert_mask)                                                       \
+      highbd_masked_variance(ref, ref_stride, temp, W, second_pred, W, msk, \
+                             msk_stride, W, H, &sse64, &sum);               \
+    else                                                                    \
+      highbd_masked_variance(ref, ref_stride, second_pred, W, temp, W, msk, \
+                             msk_stride, W, H, &sse64, &sum);               \
+    *sse = (uint32_t)sse64;                                                 \
+    return *sse - (uint32_t)(((int64_t)sum * sum) / (W * H));               \
+  }                                                                         \
+  unsigned int aom_highbd_10_masked_sub_pixel_variance##W##x##H##_ssse3(    \
+      const uint8_t *src8, int src_stride, int xoffset, int yoffset,        \
+      const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8,     \
+      const uint8_t *msk, int msk_stride, int invert_mask, uint32_t *sse) { \
+    uint64_t sse64;                                                         \
+    int sum;                                                                \
+    int64_t var;                                                            \
+    uint16_t temp[(H + 1) * W];                                             \
+    const uint16_t *src = CONVERT_TO_SHORTPTR(src8);                        \
+    const uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);                        \
+    const uint16_t *second_pred = CONVERT_TO_SHORTPTR(second_pred8);        \
+                                                                            \
+    highbd_bilinear_filter(src, src_stride, xoffset, yoffset, temp, W, H);  \
+                                                                            \
+    if (!invert_mask)                                                       \
+      highbd_masked_variance(ref, ref_stride, temp, W, second_pred, W, msk, \
+                             msk_stride, W, H, &sse64, &sum);               \
+    else                                                                    \
+      highbd_masked_variance(ref, ref_stride, second_pred, W, temp, W, msk, \
+                             msk_stride, W, H, &sse64, &sum);               \
+    *sse = (uint32_t)ROUND_POWER_OF_TWO(sse64, 4);                          \
+    sum = ROUND_POWER_OF_TWO(sum, 2);                                       \
+    var = (int64_t)(*sse) - (((int64_t)sum * sum) / (W * H));               \
+    return (var >= 0) ? (uint32_t)var : 0;                                  \
+  }                                                                         \
+  unsigned int aom_highbd_12_masked_sub_pixel_variance##W##x##H##_ssse3(    \
+      const uint8_t *src8, int src_stride, int xoffset, int yoffset,        \
+      const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8,     \
+      const uint8_t *msk, int msk_stride, int invert_mask, uint32_t *sse) { \
+    uint64_t sse64;                                                         \
+    int sum;                                                                \
+    int64_t var;                                                            \
+    uint16_t temp[(H + 1) * W];                                             \
+    const uint16_t *src = CONVERT_TO_SHORTPTR(src8);                        \
+    const uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);                        \
+    const uint16_t *second_pred = CONVERT_TO_SHORTPTR(second_pred8);        \
+                                                                            \
+    highbd_bilinear_filter(src, src_stride, xoffset, yoffset, temp, W, H);  \
+                                                                            \
+    if (!invert_mask)                                                       \
+      highbd_masked_variance(ref, ref_stride, temp, W, second_pred, W, msk, \
+                             msk_stride, W, H, &sse64, &sum);               \
+    else                                                                    \
+      highbd_masked_variance(ref, ref_stride, second_pred, W, temp, W, msk, \
+                             msk_stride, W, H, &sse64, &sum);               \
+    *sse = (uint32_t)ROUND_POWER_OF_TWO(sse64, 8);                          \
+    sum = ROUND_POWER_OF_TWO(sum, 4);                                       \
+    var = (int64_t)(*sse) - (((int64_t)sum * sum) / (W * H));               \
+    return (var >= 0) ? (uint32_t)var : 0;                                  \
+  }
+
+#define HIGHBD_MASK_SUBPIX_VAR4XH_SSSE3(H)                                  \
+  unsigned int aom_highbd_8_masked_sub_pixel_variance4x##H##_ssse3(         \
+      const uint8_t *src8, int src_stride, int xoffset, int yoffset,        \
+      const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8,     \
+      const uint8_t *msk, int msk_stride, int invert_mask, uint32_t *sse) { \
+    int sse_;                                                               \
+    int sum;                                                                \
+    uint16_t temp[(H + 1) * 4];                                             \
+    const uint16_t *src = CONVERT_TO_SHORTPTR(src8);                        \
+    const uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);                        \
+    const uint16_t *second_pred = CONVERT_TO_SHORTPTR(second_pred8);        \
+                                                                            \
+    highbd_bilinear_filter4xh(src, src_stride, xoffset, yoffset, temp, H);  \
+                                                                            \
+    if (!invert_mask)                                                       \
+      highbd_masked_variance4xh(ref, ref_stride, temp, second_pred, msk,    \
+                                msk_stride, H, &sse_, &sum);                \
+    else                                                                    \
+      highbd_masked_variance4xh(ref, ref_stride, second_pred, temp, msk,    \
+                                msk_stride, H, &sse_, &sum);                \
+    *sse = (uint32_t)sse_;                                                  \
+    return *sse - (uint32_t)(((int64_t)sum * sum) / (4 * H));               \
+  }                                                                         \
+  unsigned int aom_highbd_10_masked_sub_pixel_variance4x##H##_ssse3(        \
+      const uint8_t *src8, int src_stride, int xoffset, int yoffset,        \
+      const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8,     \
+      const uint8_t *msk, int msk_stride, int invert_mask, uint32_t *sse) { \
+    int sse_;                                                               \
+    int sum;                                                                \
+    int64_t var;                                                            \
+    uint16_t temp[(H + 1) * 4];                                             \
+    const uint16_t *src = CONVERT_TO_SHORTPTR(src8);                        \
+    const uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);                        \
+    const uint16_t *second_pred = CONVERT_TO_SHORTPTR(second_pred8);        \
+                                                                            \
+    highbd_bilinear_filter4xh(src, src_stride, xoffset, yoffset, temp, H);  \
+                                                                            \
+    if (!invert_mask)                                                       \
+      highbd_masked_variance4xh(ref, ref_stride, temp, second_pred, msk,    \
+                                msk_stride, H, &sse_, &sum);                \
+    else                                                                    \
+      highbd_masked_variance4xh(ref, ref_stride, second_pred, temp, msk,    \
+                                msk_stride, H, &sse_, &sum);                \
+    *sse = (uint32_t)ROUND_POWER_OF_TWO(sse_, 4);                           \
+    sum = ROUND_POWER_OF_TWO(sum, 2);                                       \
+    var = (int64_t)(*sse) - (((int64_t)sum * sum) / (4 * H));               \
+    return (var >= 0) ? (uint32_t)var : 0;                                  \
+  }                                                                         \
+  unsigned int aom_highbd_12_masked_sub_pixel_variance4x##H##_ssse3(        \
+      const uint8_t *src8, int src_stride, int xoffset, int yoffset,        \
+      const uint8_t *ref8, int ref_stride, const uint8_t *second_pred8,     \
+      const uint8_t *msk, int msk_stride, int invert_mask, uint32_t *sse) { \
+    int sse_;                                                               \
+    int sum;                                                                \
+    int64_t var;                                                            \
+    uint16_t temp[(H + 1) * 4];                                             \
+    const uint16_t *src = CONVERT_TO_SHORTPTR(src8);                        \
+    const uint16_t *ref = CONVERT_TO_SHORTPTR(ref8);                        \
+    const uint16_t *second_pred = CONVERT_TO_SHORTPTR(second_pred8);        \
+                                                                            \
+    highbd_bilinear_filter4xh(src, src_stride, xoffset, yoffset, temp, H);  \
+                                                                            \
+    if (!invert_mask)                                                       \
+      highbd_masked_variance4xh(ref, ref_stride, temp, second_pred, msk,    \
+                                msk_stride, H, &sse_, &sum);                \
+    else                                                                    \
+      highbd_masked_variance4xh(ref, ref_stride, second_pred, temp, msk,    \
+                                msk_stride, H, &sse_, &sum);                \
+    *sse = (uint32_t)ROUND_POWER_OF_TWO(sse_, 8);                           \
+    sum = ROUND_POWER_OF_TWO(sum, 4);                                       \
+    var = (int64_t)(*sse) - (((int64_t)sum * sum) / (4 * H));               \
+    return (var >= 0) ? (uint32_t)var : 0;                                  \
+  }
+
+HIGHBD_MASK_SUBPIX_VAR_SSSE3(128, 128)
+HIGHBD_MASK_SUBPIX_VAR_SSSE3(128, 64)
+HIGHBD_MASK_SUBPIX_VAR_SSSE3(64, 128)
+HIGHBD_MASK_SUBPIX_VAR_SSSE3(64, 64)
+HIGHBD_MASK_SUBPIX_VAR_SSSE3(64, 32)
+HIGHBD_MASK_SUBPIX_VAR_SSSE3(32, 64)
+HIGHBD_MASK_SUBPIX_VAR_SSSE3(32, 32)
+HIGHBD_MASK_SUBPIX_VAR_SSSE3(32, 16)
+HIGHBD_MASK_SUBPIX_VAR_SSSE3(16, 32)
+HIGHBD_MASK_SUBPIX_VAR_SSSE3(16, 16)
+HIGHBD_MASK_SUBPIX_VAR_SSSE3(16, 8)
+HIGHBD_MASK_SUBPIX_VAR_SSSE3(8, 16)
+HIGHBD_MASK_SUBPIX_VAR_SSSE3(8, 8)
+HIGHBD_MASK_SUBPIX_VAR_SSSE3(8, 4)
+HIGHBD_MASK_SUBPIX_VAR4XH_SSSE3(8)
+HIGHBD_MASK_SUBPIX_VAR4XH_SSSE3(4)
+HIGHBD_MASK_SUBPIX_VAR4XH_SSSE3(16)
+HIGHBD_MASK_SUBPIX_VAR_SSSE3(16, 4)
+HIGHBD_MASK_SUBPIX_VAR_SSSE3(8, 32)
+HIGHBD_MASK_SUBPIX_VAR_SSSE3(32, 8)
+HIGHBD_MASK_SUBPIX_VAR_SSSE3(16, 64)
+HIGHBD_MASK_SUBPIX_VAR_SSSE3(64, 16)
+
+static INLINE __m128i highbd_filter_block(const __m128i a, const __m128i b,
+                                          const __m128i filter) {
+  __m128i v0 = _mm_unpacklo_epi16(a, b);
+  v0 = _mm_madd_epi16(v0, filter);
+  v0 = xx_roundn_epu32(v0, FILTER_BITS);
+
+  __m128i v1 = _mm_unpackhi_epi16(a, b);
+  v1 = _mm_madd_epi16(v1, filter);
+  v1 = xx_roundn_epu32(v1, FILTER_BITS);
+
+  return _mm_packs_epi32(v0, v1);
+}
+
+static void highbd_bilinear_filter(const uint16_t *src, int src_stride,
+                                   int xoffset, int yoffset, uint16_t *dst,
+                                   int w, int h) {
+  int i, j;
+  // Horizontal filter
+  if (xoffset == 0) {
+    uint16_t *b = dst;
+    for (i = 0; i < h + 1; ++i) {
+      for (j = 0; j < w; j += 8) {
+        __m128i x = _mm_loadu_si128((__m128i *)&src[j]);
+        _mm_storeu_si128((__m128i *)&b[j], x);
+      }
+      src += src_stride;
+      b += w;
+    }
+  } else if (xoffset == 4) {
+    uint16_t *b = dst;
+    for (i = 0; i < h + 1; ++i) {
+      for (j = 0; j < w; j += 8) {
+        __m128i x = _mm_loadu_si128((__m128i *)&src[j]);
+        __m128i y = _mm_loadu_si128((__m128i *)&src[j + 8]);
+        __m128i z = _mm_alignr_epi8(y, x, 2);
+        _mm_storeu_si128((__m128i *)&b[j], _mm_avg_epu16(x, z));
+      }
+      src += src_stride;
+      b += w;
+    }
+  } else {
+    uint16_t *b = dst;
+    const uint8_t *hfilter = bilinear_filters_2t[xoffset];
+    const __m128i hfilter_vec = _mm_set1_epi32(hfilter[0] | (hfilter[1] << 16));
+    for (i = 0; i < h + 1; ++i) {
+      for (j = 0; j < w; j += 8) {
+        const __m128i x = _mm_loadu_si128((__m128i *)&src[j]);
+        const __m128i y = _mm_loadu_si128((__m128i *)&src[j + 8]);
+        const __m128i z = _mm_alignr_epi8(y, x, 2);
+        const __m128i res = highbd_filter_block(x, z, hfilter_vec);
+        _mm_storeu_si128((__m128i *)&b[j], res);
+      }
+
+      src += src_stride;
+      b += w;
+    }
+  }
+
+  // Vertical filter
+  if (yoffset == 0) {
+    // The data is already in 'dst', so no need to filter
+  } else if (yoffset == 4) {
+    for (i = 0; i < h; ++i) {
+      for (j = 0; j < w; j += 8) {
+        __m128i x = _mm_loadu_si128((__m128i *)&dst[j]);
+        __m128i y = _mm_loadu_si128((__m128i *)&dst[j + w]);
+        _mm_storeu_si128((__m128i *)&dst[j], _mm_avg_epu16(x, y));
+      }
+      dst += w;
+    }
+  } else {
+    const uint8_t *vfilter = bilinear_filters_2t[yoffset];
+    const __m128i vfilter_vec = _mm_set1_epi32(vfilter[0] | (vfilter[1] << 16));
+    for (i = 0; i < h; ++i) {
+      for (j = 0; j < w; j += 8) {
+        const __m128i x = _mm_loadu_si128((__m128i *)&dst[j]);
+        const __m128i y = _mm_loadu_si128((__m128i *)&dst[j + w]);
+        const __m128i res = highbd_filter_block(x, y, vfilter_vec);
+        _mm_storeu_si128((__m128i *)&dst[j], res);
+      }
+
+      dst += w;
+    }
+  }
+}
+
+static INLINE __m128i highbd_filter_block_2rows(const __m128i *a0,
+                                                const __m128i *b0,
+                                                const __m128i *a1,
+                                                const __m128i *b1,
+                                                const __m128i *filter) {
+  __m128i v0 = _mm_unpacklo_epi16(*a0, *b0);
+  v0 = _mm_madd_epi16(v0, *filter);
+  v0 = xx_roundn_epu32(v0, FILTER_BITS);
+
+  __m128i v1 = _mm_unpacklo_epi16(*a1, *b1);
+  v1 = _mm_madd_epi16(v1, *filter);
+  v1 = xx_roundn_epu32(v1, FILTER_BITS);
+
+  return _mm_packs_epi32(v0, v1);
+}
+
+static void highbd_bilinear_filter4xh(const uint16_t *src, int src_stride,
+                                      int xoffset, int yoffset, uint16_t *dst,
+                                      int h) {
+  int i;
+  // Horizontal filter
+  if (xoffset == 0) {
+    uint16_t *b = dst;
+    for (i = 0; i < h + 1; ++i) {
+      __m128i x = _mm_loadl_epi64((__m128i *)src);
+      _mm_storel_epi64((__m128i *)b, x);
+      src += src_stride;
+      b += 4;
+    }
+  } else if (xoffset == 4) {
+    uint16_t *b = dst;
+    for (i = 0; i < h + 1; ++i) {
+      __m128i x = _mm_loadu_si128((__m128i *)src);
+      __m128i z = _mm_srli_si128(x, 2);
+      _mm_storel_epi64((__m128i *)b, _mm_avg_epu16(x, z));
+      src += src_stride;
+      b += 4;
+    }
+  } else {
+    uint16_t *b = dst;
+    const uint8_t *hfilter = bilinear_filters_2t[xoffset];
+    const __m128i hfilter_vec = _mm_set1_epi32(hfilter[0] | (hfilter[1] << 16));
+    for (i = 0; i < h; i += 2) {
+      const __m128i x0 = _mm_loadu_si128((__m128i *)src);
+      const __m128i z0 = _mm_srli_si128(x0, 2);
+      const __m128i x1 = _mm_loadu_si128((__m128i *)&src[src_stride]);
+      const __m128i z1 = _mm_srli_si128(x1, 2);
+      const __m128i res =
+          highbd_filter_block_2rows(&x0, &z0, &x1, &z1, &hfilter_vec);
+      _mm_storeu_si128((__m128i *)b, res);
+
+      src += src_stride * 2;
+      b += 8;
+    }
+    // Process i = h separately
+    __m128i x = _mm_loadu_si128((__m128i *)src);
+    __m128i z = _mm_srli_si128(x, 2);
+
+    __m128i v0 = _mm_unpacklo_epi16(x, z);
+    v0 = _mm_madd_epi16(v0, hfilter_vec);
+    v0 = xx_roundn_epu32(v0, FILTER_BITS);
+
+    _mm_storel_epi64((__m128i *)b, _mm_packs_epi32(v0, v0));
+  }
+
+  // Vertical filter
+  if (yoffset == 0) {
+    // The data is already in 'dst', so no need to filter
+  } else if (yoffset == 4) {
+    for (i = 0; i < h; ++i) {
+      __m128i x = _mm_loadl_epi64((__m128i *)dst);
+      __m128i y = _mm_loadl_epi64((__m128i *)&dst[4]);
+      _mm_storel_epi64((__m128i *)dst, _mm_avg_epu16(x, y));
+      dst += 4;
+    }
+  } else {
+    const uint8_t *vfilter = bilinear_filters_2t[yoffset];
+    const __m128i vfilter_vec = _mm_set1_epi32(vfilter[0] | (vfilter[1] << 16));
+    for (i = 0; i < h; i += 2) {
+      const __m128i x = _mm_loadl_epi64((__m128i *)dst);
+      const __m128i y = _mm_loadl_epi64((__m128i *)&dst[4]);
+      const __m128i z = _mm_loadl_epi64((__m128i *)&dst[8]);
+      const __m128i res =
+          highbd_filter_block_2rows(&x, &y, &y, &z, &vfilter_vec);
+      _mm_storeu_si128((__m128i *)dst, res);
+
+      dst += 8;
+    }
+  }
+}
+
+static void highbd_masked_variance(const uint16_t *src_ptr, int src_stride,
+                                   const uint16_t *a_ptr, int a_stride,
+                                   const uint16_t *b_ptr, int b_stride,
+                                   const uint8_t *m_ptr, int m_stride,
+                                   int width, int height, uint64_t *sse,
+                                   int *sum_) {
+  int x, y;
+  // Note on bit widths:
+  // The maximum value of 'sum' is (2^12 - 1) * 128 * 128 =~ 2^26,
+  // so this can be kept as four 32-bit values.
+  // But the maximum value of 'sum_sq' is (2^12 - 1)^2 * 128 * 128 =~ 2^38,
+  // so this must be stored as two 64-bit values.
+  __m128i sum = _mm_setzero_si128(), sum_sq = _mm_setzero_si128();
+  const __m128i mask_max = _mm_set1_epi16((1 << AOM_BLEND_A64_ROUND_BITS));
+  const __m128i round_const =
+      _mm_set1_epi32((1 << AOM_BLEND_A64_ROUND_BITS) >> 1);
+  const __m128i zero = _mm_setzero_si128();
+
+  for (y = 0; y < height; y++) {
+    for (x = 0; x < width; x += 8) {
+      const __m128i src = _mm_loadu_si128((const __m128i *)&src_ptr[x]);
+      const __m128i a = _mm_loadu_si128((const __m128i *)&a_ptr[x]);
+      const __m128i b = _mm_loadu_si128((const __m128i *)&b_ptr[x]);
+      const __m128i m =
+          _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)&m_ptr[x]), zero);
+      const __m128i m_inv = _mm_sub_epi16(mask_max, m);
+
+      // Calculate 8 predicted pixels.
+      const __m128i data_l = _mm_unpacklo_epi16(a, b);
+      const __m128i mask_l = _mm_unpacklo_epi16(m, m_inv);
+      __m128i pred_l = _mm_madd_epi16(data_l, mask_l);
+      pred_l = _mm_srai_epi32(_mm_add_epi32(pred_l, round_const),
+                              AOM_BLEND_A64_ROUND_BITS);
+
+      const __m128i data_r = _mm_unpackhi_epi16(a, b);
+      const __m128i mask_r = _mm_unpackhi_epi16(m, m_inv);
+      __m128i pred_r = _mm_madd_epi16(data_r, mask_r);
+      pred_r = _mm_srai_epi32(_mm_add_epi32(pred_r, round_const),
+                              AOM_BLEND_A64_ROUND_BITS);
+
+      const __m128i src_l = _mm_unpacklo_epi16(src, zero);
+      const __m128i src_r = _mm_unpackhi_epi16(src, zero);
+      __m128i diff_l = _mm_sub_epi32(pred_l, src_l);
+      __m128i diff_r = _mm_sub_epi32(pred_r, src_r);
+
+      // Update partial sums and partial sums of squares
+      sum = _mm_add_epi32(sum, _mm_add_epi32(diff_l, diff_r));
+      // A trick: Now each entry of diff_l and diff_r is stored in a 32-bit
+      // field, but the range of values is only [-(2^12 - 1), 2^12 - 1].
+      // So we can re-pack into 16-bit fields and use _mm_madd_epi16
+      // to calculate the squares and partially sum them.
+      const __m128i tmp = _mm_packs_epi32(diff_l, diff_r);
+      const __m128i prod = _mm_madd_epi16(tmp, tmp);
+      // Then we want to sign-extend to 64 bits and accumulate
+      const __m128i sign = _mm_srai_epi32(prod, 31);
+      const __m128i tmp_0 = _mm_unpacklo_epi32(prod, sign);
+      const __m128i tmp_1 = _mm_unpackhi_epi32(prod, sign);
+      sum_sq = _mm_add_epi64(sum_sq, _mm_add_epi64(tmp_0, tmp_1));
+    }
+
+    src_ptr += src_stride;
+    a_ptr += a_stride;
+    b_ptr += b_stride;
+    m_ptr += m_stride;
+  }
+  // Reduce down to a single sum and sum of squares
+  sum = _mm_hadd_epi32(sum, zero);
+  sum = _mm_hadd_epi32(sum, zero);
+  *sum_ = _mm_cvtsi128_si32(sum);
+  sum_sq = _mm_add_epi64(sum_sq, _mm_srli_si128(sum_sq, 8));
+  _mm_storel_epi64((__m128i *)sse, sum_sq);
+}
+
+static void highbd_masked_variance4xh(const uint16_t *src_ptr, int src_stride,
+                                      const uint16_t *a_ptr,
+                                      const uint16_t *b_ptr,
+                                      const uint8_t *m_ptr, int m_stride,
+                                      int height, int *sse, int *sum_) {
+  int y;
+  // Note: For this function, h <= 8 (or maybe 16 if we add 4:1 partitions).
+  // So the maximum value of sum is (2^12 - 1) * 4 * 16 =~ 2^18
+  // and the maximum value of sum_sq is (2^12 - 1)^2 * 4 * 16 =~ 2^30.
+  // So we can safely pack sum_sq into 32-bit fields, which is slightly more
+  // convenient.
+  __m128i sum = _mm_setzero_si128(), sum_sq = _mm_setzero_si128();
+  const __m128i mask_max = _mm_set1_epi16((1 << AOM_BLEND_A64_ROUND_BITS));
+  const __m128i round_const =
+      _mm_set1_epi32((1 << AOM_BLEND_A64_ROUND_BITS) >> 1);
+  const __m128i zero = _mm_setzero_si128();
+
+  for (y = 0; y < height; y += 2) {
+    __m128i src = _mm_unpacklo_epi64(
+        _mm_loadl_epi64((const __m128i *)src_ptr),
+        _mm_loadl_epi64((const __m128i *)&src_ptr[src_stride]));
+    const __m128i a = _mm_loadu_si128((const __m128i *)a_ptr);
+    const __m128i b = _mm_loadu_si128((const __m128i *)b_ptr);
+    const __m128i m = _mm_unpacklo_epi8(
+        _mm_unpacklo_epi32(_mm_cvtsi32_si128(*(const int *)m_ptr),
+                           _mm_cvtsi32_si128(*(const int *)&m_ptr[m_stride])),
+        zero);
+    const __m128i m_inv = _mm_sub_epi16(mask_max, m);
+
+    const __m128i data_l = _mm_unpacklo_epi16(a, b);
+    const __m128i mask_l = _mm_unpacklo_epi16(m, m_inv);
+    __m128i pred_l = _mm_madd_epi16(data_l, mask_l);
+    pred_l = _mm_srai_epi32(_mm_add_epi32(pred_l, round_const),
+                            AOM_BLEND_A64_ROUND_BITS);
+
+    const __m128i data_r = _mm_unpackhi_epi16(a, b);
+    const __m128i mask_r = _mm_unpackhi_epi16(m, m_inv);
+    __m128i pred_r = _mm_madd_epi16(data_r, mask_r);
+    pred_r = _mm_srai_epi32(_mm_add_epi32(pred_r, round_const),
+                            AOM_BLEND_A64_ROUND_BITS);
+
+    const __m128i src_l = _mm_unpacklo_epi16(src, zero);
+    const __m128i src_r = _mm_unpackhi_epi16(src, zero);
+    __m128i diff_l = _mm_sub_epi32(pred_l, src_l);
+    __m128i diff_r = _mm_sub_epi32(pred_r, src_r);
+
+    // Update partial sums and partial sums of squares
+    sum = _mm_add_epi32(sum, _mm_add_epi32(diff_l, diff_r));
+    const __m128i tmp = _mm_packs_epi32(diff_l, diff_r);
+    const __m128i prod = _mm_madd_epi16(tmp, tmp);
+    sum_sq = _mm_add_epi32(sum_sq, prod);
+
+    src_ptr += src_stride * 2;
+    a_ptr += 8;
+    b_ptr += 8;
+    m_ptr += m_stride * 2;
+  }
+  // Reduce down to a single sum and sum of squares
+  sum = _mm_hadd_epi32(sum, sum_sq);
+  sum = _mm_hadd_epi32(sum, zero);
+  *sum_ = _mm_cvtsi128_si32(sum);
+  *sse = (unsigned int)_mm_cvtsi128_si32(_mm_srli_si128(sum, 4));
+}
+#endif  // CONFIG_AV1_HIGHBITDEPTH
+
+void aom_comp_mask_pred_ssse3(uint8_t *comp_pred, const uint8_t *pred,
+                              int width, int height, const uint8_t *ref,
+                              int ref_stride, const uint8_t *mask,
+                              int mask_stride, int invert_mask) {
+  const uint8_t *src0 = invert_mask ? pred : ref;
+  const uint8_t *src1 = invert_mask ? ref : pred;
+  const int stride0 = invert_mask ? width : ref_stride;
+  const int stride1 = invert_mask ? ref_stride : width;
+  assert(height % 2 == 0);
+  int i = 0;
+  if (width == 8) {
+    comp_mask_pred_8_ssse3(comp_pred, height, src0, stride0, src1, stride1,
+                           mask, mask_stride);
+  } else if (width == 16) {
+    do {
+      comp_mask_pred_16_ssse3(src0, src1, mask, comp_pred);
+      comp_mask_pred_16_ssse3(src0 + stride0, src1 + stride1,
+                              mask + mask_stride, comp_pred + width);
+      comp_pred += (width << 1);
+      src0 += (stride0 << 1);
+      src1 += (stride1 << 1);
+      mask += (mask_stride << 1);
+      i += 2;
+    } while (i < height);
+  } else {
+    do {
+      for (int x = 0; x < width; x += 32) {
+        comp_mask_pred_16_ssse3(src0 + x, src1 + x, mask + x, comp_pred);
+        comp_mask_pred_16_ssse3(src0 + x + 16, src1 + x + 16, mask + x + 16,
+                                comp_pred + 16);
+        comp_pred += 32;
+      }
+      src0 += (stride0);
+      src1 += (stride1);
+      mask += (mask_stride);
+      i += 1;
+    } while (i < height);
+  }
+}
diff --git a/third_party/aom/aom_dsp/x86/masked_variance_intrin_ssse3.h b/third_party/aom/aom_dsp/x86/masked_variance_intrin_ssse3.h
new file mode 100644
index 0000000000..4faa098ace
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/masked_variance_intrin_ssse3.h
@@ -0,0 +1,92 @@
+/*
+ * 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_AOM_DSP_X86_MASKED_VARIANCE_INTRIN_SSSE3_H_
+#define AOM_AOM_DSP_X86_MASKED_VARIANCE_INTRIN_SSSE3_H_
+
+#include <stdlib.h>
+#include <string.h>
+#include <tmmintrin.h>
+
+#include "config/aom_config.h"
+#include "config/aom_dsp_rtcd.h"
+
+#include "aom_dsp/blend.h"
+
+static INLINE void comp_mask_pred_16_ssse3(const uint8_t *src0,
+                                           const uint8_t *src1,
+                                           const uint8_t *mask, uint8_t *dst) {
+  const __m128i alpha_max = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA);
+  const __m128i round_offset =
+      _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS));
+
+  const __m128i sA0 = _mm_lddqu_si128((const __m128i *)(src0));
+  const __m128i sA1 = _mm_lddqu_si128((const __m128i *)(src1));
+  const __m128i aA = _mm_load_si128((const __m128i *)(mask));
+
+  const __m128i maA = _mm_sub_epi8(alpha_max, aA);
+
+  const __m128i ssAL = _mm_unpacklo_epi8(sA0, sA1);
+  const __m128i aaAL = _mm_unpacklo_epi8(aA, maA);
+  const __m128i ssAH = _mm_unpackhi_epi8(sA0, sA1);
+  const __m128i aaAH = _mm_unpackhi_epi8(aA, maA);
+
+  const __m128i blendAL = _mm_maddubs_epi16(ssAL, aaAL);
+  const __m128i blendAH = _mm_maddubs_epi16(ssAH, aaAH);
+
+  const __m128i roundAL = _mm_mulhrs_epi16(blendAL, round_offset);
+  const __m128i roundAH = _mm_mulhrs_epi16(blendAH, round_offset);
+  _mm_store_si128((__m128i *)dst, _mm_packus_epi16(roundAL, roundAH));
+}
+
+static INLINE void comp_mask_pred_8_ssse3(uint8_t *comp_pred, int height,
+                                          const uint8_t *src0, int stride0,
+                                          const uint8_t *src1, int stride1,
+                                          const uint8_t *mask,
+                                          int mask_stride) {
+  int i = 0;
+  const __m128i alpha_max = _mm_set1_epi8(AOM_BLEND_A64_MAX_ALPHA);
+  const __m128i round_offset =
+      _mm_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS));
+  do {
+    // odd line A
+    const __m128i sA0 = _mm_loadl_epi64((const __m128i *)(src0));
+    const __m128i sA1 = _mm_loadl_epi64((const __m128i *)(src1));
+    const __m128i aA = _mm_loadl_epi64((const __m128i *)(mask));
+    // even line B
+    const __m128i sB0 = _mm_loadl_epi64((const __m128i *)(src0 + stride0));
+    const __m128i sB1 = _mm_loadl_epi64((const __m128i *)(src1 + stride1));
+    const __m128i a = _mm_castps_si128(_mm_loadh_pi(
+        _mm_castsi128_ps(aA), (const __m64 *)(mask + mask_stride)));
+
+    const __m128i ssA = _mm_unpacklo_epi8(sA0, sA1);
+    const __m128i ssB = _mm_unpacklo_epi8(sB0, sB1);
+
+    const __m128i ma = _mm_sub_epi8(alpha_max, a);
+    const __m128i aaA = _mm_unpacklo_epi8(a, ma);
+    const __m128i aaB = _mm_unpackhi_epi8(a, ma);
+
+    const __m128i blendA = _mm_maddubs_epi16(ssA, aaA);
+    const __m128i blendB = _mm_maddubs_epi16(ssB, aaB);
+    const __m128i roundA = _mm_mulhrs_epi16(blendA, round_offset);
+    const __m128i roundB = _mm_mulhrs_epi16(blendB, round_offset);
+    const __m128i round = _mm_packus_epi16(roundA, roundB);
+    // comp_pred's stride == width == 8
+    _mm_store_si128((__m128i *)(comp_pred), round);
+    comp_pred += (8 << 1);
+    src0 += (stride0 << 1);
+    src1 += (stride1 << 1);
+    mask += (mask_stride << 1);
+    i += 2;
+  } while (i < height);
+}
+
+#endif  // AOM_AOM_DSP_X86_MASKED_VARIANCE_INTRIN_SSSE3_H_
diff --git a/third_party/aom/aom_dsp/x86/mem_sse2.h b/third_party/aom/aom_dsp/x86/mem_sse2.h
new file mode 100644
index 0000000000..085a572cb1
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/mem_sse2.h
@@ -0,0 +1,167 @@
+/*
+ * Copyright (c) 2017, 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_AOM_DSP_X86_MEM_SSE2_H_
+#define AOM_AOM_DSP_X86_MEM_SSE2_H_
+
+#include <emmintrin.h>  // SSE2
+#include <string.h>
+
+#include "config/aom_config.h"
+
+#include "aom/aom_integer.h"
+
+static INLINE int16_t loadu_int16(const void *src) {
+  int16_t v;
+  memcpy(&v, src, sizeof(v));
+  return v;
+}
+
+static INLINE int32_t loadu_int32(const void *src) {
+  int32_t v;
+  memcpy(&v, src, sizeof(v));
+  return v;
+}
+
+static INLINE int64_t loadu_int64(const void *src) {
+  int64_t v;
+  memcpy(&v, src, sizeof(v));
+  return v;
+}
+
+static INLINE void _mm_storeh_epi64(__m128i *const d, const __m128i s) {
+  _mm_storeh_pi((__m64 *)d, _mm_castsi128_ps(s));
+}
+
+static INLINE __m128i loadh_epi64(const void *const src, const __m128i s) {
+  return _mm_castps_si128(
+      _mm_loadh_pi(_mm_castsi128_ps(s), (const __m64 *)src));
+}
+
+static INLINE __m128i load_8bit_4x4_to_1_reg_sse2(const void *const src,
+                                                  const int byte_stride) {
+  return _mm_setr_epi32(loadu_int32((int8_t *)src + 0 * byte_stride),
+                        loadu_int32((int8_t *)src + 1 * byte_stride),
+                        loadu_int32((int8_t *)src + 2 * byte_stride),
+                        loadu_int32((int8_t *)src + 3 * byte_stride));
+}
+
+static INLINE __m128i load_8bit_8x2_to_1_reg_sse2(const void *const src,
+                                                  const int byte_stride) {
+  __m128i dst;
+  dst = _mm_loadl_epi64((__m128i *)((int8_t *)src + 0 * byte_stride));
+  dst = loadh_epi64((int8_t *)src + 1 * byte_stride, dst);
+  return dst;
+}
+
+static INLINE void store_8bit_8x4_from_16x2(const __m128i *const s,
+                                            uint8_t *const d,
+                                            const ptrdiff_t stride) {
+  _mm_storel_epi64((__m128i *)(d + 0 * stride), s[0]);
+  _mm_storeh_epi64((__m128i *)(d + 1 * stride), s[0]);
+  _mm_storel_epi64((__m128i *)(d + 2 * stride), s[1]);
+  _mm_storeh_epi64((__m128i *)(d + 3 * stride), s[1]);
+}
+
+static INLINE void store_8bit_4x4(const __m128i *const s, uint8_t *const d,
+                                  const ptrdiff_t stride) {
+  *(int *)(d + 0 * stride) = _mm_cvtsi128_si32(s[0]);
+  *(int *)(d + 1 * stride) = _mm_cvtsi128_si32(s[1]);
+  *(int *)(d + 2 * stride) = _mm_cvtsi128_si32(s[2]);
+  *(int *)(d + 3 * stride) = _mm_cvtsi128_si32(s[3]);
+}
+
+static INLINE void store_8bit_4x4_sse2(const __m128i s, uint8_t *const d,
+                                       const ptrdiff_t stride) {
+  __m128i ss[4];
+
+  ss[0] = s;
+  ss[1] = _mm_srli_si128(s, 4);
+  ss[2] = _mm_srli_si128(s, 8);
+  ss[3] = _mm_srli_si128(s, 12);
+  store_8bit_4x4(ss, d, stride);
+}
+
+static INLINE void load_8bit_4x4(const uint8_t *const s, const ptrdiff_t stride,
+                                 __m128i *const d) {
+  d[0] = _mm_cvtsi32_si128(*(const int *)(s + 0 * stride));
+  d[1] = _mm_cvtsi32_si128(*(const int *)(s + 1 * stride));
+  d[2] = _mm_cvtsi32_si128(*(const int *)(s + 2 * stride));
+  d[3] = _mm_cvtsi32_si128(*(const int *)(s + 3 * stride));
+}
+
+static INLINE void load_8bit_4x8(const uint8_t *const s, const ptrdiff_t stride,
+                                 __m128i *const d) {
+  load_8bit_4x4(s + 0 * stride, stride, &d[0]);
+  load_8bit_4x4(s + 4 * stride, stride, &d[4]);
+}
+
+static INLINE void load_8bit_8x4(const uint8_t *const s, const ptrdiff_t stride,
+                                 __m128i *const d) {
+  d[0] = _mm_loadl_epi64((const __m128i *)(s + 0 * stride));
+  d[1] = _mm_loadl_epi64((const __m128i *)(s + 1 * stride));
+  d[2] = _mm_loadl_epi64((const __m128i *)(s + 2 * stride));
+  d[3] = _mm_loadl_epi64((const __m128i *)(s + 3 * stride));
+}
+
+static INLINE void loadu_8bit_16x4(const uint8_t *const s,
+                                   const ptrdiff_t stride, __m128i *const d) {
+  d[0] = _mm_loadu_si128((const __m128i *)(s + 0 * stride));
+  d[1] = _mm_loadu_si128((const __m128i *)(s + 1 * stride));
+  d[2] = _mm_loadu_si128((const __m128i *)(s + 2 * stride));
+  d[3] = _mm_loadu_si128((const __m128i *)(s + 3 * stride));
+}
+
+static INLINE void load_8bit_8x8(const uint8_t *const s, const ptrdiff_t stride,
+                                 __m128i *const d) {
+  load_8bit_8x4(s + 0 * stride, stride, &d[0]);
+  load_8bit_8x4(s + 4 * stride, stride, &d[4]);
+}
+
+static INLINE void load_8bit_16x8(const uint8_t *const s,
+                                  const ptrdiff_t stride, __m128i *const d) {
+  d[0] = _mm_load_si128((const __m128i *)(s + 0 * stride));
+  d[1] = _mm_load_si128((const __m128i *)(s + 1 * stride));
+  d[2] = _mm_load_si128((const __m128i *)(s + 2 * stride));
+  d[3] = _mm_load_si128((const __m128i *)(s + 3 * stride));
+  d[4] = _mm_load_si128((const __m128i *)(s + 4 * stride));
+  d[5] = _mm_load_si128((const __m128i *)(s + 5 * stride));
+  d[6] = _mm_load_si128((const __m128i *)(s + 6 * stride));
+  d[7] = _mm_load_si128((const __m128i *)(s + 7 * stride));
+}
+
+static INLINE void loadu_8bit_16x8(const uint8_t *const s,
+                                   const ptrdiff_t stride, __m128i *const d) {
+  loadu_8bit_16x4(s + 0 * stride, stride, &d[0]);
+  loadu_8bit_16x4(s + 4 * stride, stride, &d[4]);
+}
+
+static INLINE void store_8bit_8x8(const __m128i *const s, uint8_t *const d,
+                                  const ptrdiff_t stride) {
+  _mm_storel_epi64((__m128i *)(d + 0 * stride), s[0]);
+  _mm_storel_epi64((__m128i *)(d + 1 * stride), s[1]);
+  _mm_storel_epi64((__m128i *)(d + 2 * stride), s[2]);
+  _mm_storel_epi64((__m128i *)(d + 3 * stride), s[3]);
+  _mm_storel_epi64((__m128i *)(d + 4 * stride), s[4]);
+  _mm_storel_epi64((__m128i *)(d + 5 * stride), s[5]);
+  _mm_storel_epi64((__m128i *)(d + 6 * stride), s[6]);
+  _mm_storel_epi64((__m128i *)(d + 7 * stride), s[7]);
+}
+
+static INLINE void storeu_8bit_16x4(const __m128i *const s, uint8_t *const d,
+                                    const ptrdiff_t stride) {
+  _mm_storeu_si128((__m128i *)(d + 0 * stride), s[0]);
+  _mm_storeu_si128((__m128i *)(d + 1 * stride), s[1]);
+  _mm_storeu_si128((__m128i *)(d + 2 * stride), s[2]);
+  _mm_storeu_si128((__m128i *)(d + 3 * stride), s[3]);
+}
+
+#endif  // AOM_AOM_DSP_X86_MEM_SSE2_H_
diff --git a/third_party/aom/aom_dsp/x86/obmc_intrinsic_sse4.h b/third_party/aom/aom_dsp/x86/obmc_intrinsic_sse4.h
new file mode 100644
index 0000000000..210f466b6f
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/obmc_intrinsic_sse4.h
@@ -0,0 +1,58 @@
+/*
+ * 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_AOM_DSP_X86_OBMC_INTRINSIC_SSE4_H_
+#define AOM_AOM_DSP_X86_OBMC_INTRINSIC_SSE4_H_
+
+#include <smmintrin.h>
+
+#include "aom_dsp/x86/obmc_intrinsic_ssse3.h"
+
+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,
+                                    const int h) {
+  const int pre_step = pre_stride - 4;
+  int n = 0;
+  __m128i v_sum_d = _mm_setzero_si128();
+  __m128i v_sse_d = _mm_setzero_si128();
+
+  assert(IS_POWER_OF_TWO(h));
+
+  do {
+    const __m128i v_p_b = _mm_cvtsi32_si128(*(const int *)(pre + n));
+    const __m128i v_m_d = _mm_load_si128((const __m128i *)(mask + n));
+    const __m128i v_w_d = _mm_load_si128((const __m128i *)(wsrc + n));
+
+    const __m128i v_p_d = _mm_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 __m128i v_pm_d = _mm_madd_epi16(v_p_d, v_m_d);
+
+    const __m128i v_diff_d = _mm_sub_epi32(v_w_d, v_pm_d);
+    const __m128i v_rdiff_d = xx_roundn_epi32(v_diff_d, 12);
+    const __m128i v_sqrdiff_d = _mm_mullo_epi32(v_rdiff_d, v_rdiff_d);
+
+    v_sum_d = _mm_add_epi32(v_sum_d, v_rdiff_d);
+    v_sse_d = _mm_add_epi32(v_sse_d, v_sqrdiff_d);
+
+    n += 4;
+
+    if (n % 4 == 0) pre += pre_step;
+  } while (n < 4 * h);
+
+  *sum = xx_hsum_epi32_si32(v_sum_d);
+  *sse = xx_hsum_epi32_si32(v_sse_d);
+}
+
+#endif  // AOM_AOM_DSP_X86_OBMC_INTRINSIC_SSE4_H_
diff --git a/third_party/aom/aom_dsp/x86/obmc_intrinsic_ssse3.h b/third_party/aom/aom_dsp/x86/obmc_intrinsic_ssse3.h
new file mode 100644
index 0000000000..27398ffd62
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/obmc_intrinsic_ssse3.h
@@ -0,0 +1,54 @@
+/*
+ * Copyright (c) 2017, 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_AOM_DSP_X86_OBMC_INTRINSIC_SSSE3_H_
+#define AOM_AOM_DSP_X86_OBMC_INTRINSIC_SSSE3_H_
+
+#include <immintrin.h>
+
+#include "config/aom_config.h"
+
+static INLINE int32_t xx_hsum_epi32_si32(__m128i v_d) {
+  v_d = _mm_hadd_epi32(v_d, v_d);
+  v_d = _mm_hadd_epi32(v_d, v_d);
+  return _mm_cvtsi128_si32(v_d);
+}
+
+static INLINE int64_t xx_hsum_epi64_si64(__m128i v_q) {
+  v_q = _mm_add_epi64(v_q, _mm_srli_si128(v_q, 8));
+#if AOM_ARCH_X86_64
+  return _mm_cvtsi128_si64(v_q);
+#else
+  {
+    int64_t tmp;
+    _mm_storel_epi64((__m128i *)&tmp, v_q);
+    return tmp;
+  }
+#endif
+}
+
+static INLINE int64_t xx_hsum_epi32_si64(__m128i v_d) {
+  const __m128i v_sign_d = _mm_cmplt_epi32(v_d, _mm_setzero_si128());
+  const __m128i v_0_q = _mm_unpacklo_epi32(v_d, v_sign_d);
+  const __m128i v_1_q = _mm_unpackhi_epi32(v_d, v_sign_d);
+  return xx_hsum_epi64_si64(_mm_add_epi64(v_0_q, v_1_q));
+}
+
+// This is equivalent to ROUND_POWER_OF_TWO_SIGNED(v_val_d, bits)
+static INLINE __m128i xx_roundn_epi32(__m128i v_val_d, int bits) {
+  const __m128i v_bias_d = _mm_set1_epi32((1 << bits) >> 1);
+  const __m128i v_sign_d = _mm_srai_epi32(v_val_d, 31);
+  const __m128i v_tmp_d =
+      _mm_add_epi32(_mm_add_epi32(v_val_d, v_bias_d), v_sign_d);
+  return _mm_srai_epi32(v_tmp_d, bits);
+}
+
+#endif  // AOM_AOM_DSP_X86_OBMC_INTRINSIC_SSSE3_H_
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 0000000000..9d1b7d4968
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/obmc_sad_avx2.c
@@ -0,0 +1,271 @@
+/*
+ * 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 "config/aom_dsp_rtcd.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_sad_sse4.c b/third_party/aom/aom_dsp/x86/obmc_sad_sse4.c
new file mode 100644
index 0000000000..542572c761
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/obmc_sad_sse4.c
@@ -0,0 +1,269 @@
+/*
+ * Copyright (c) 2016, 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 "config/aom_dsp_rtcd.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 AOM_FORCE_INLINE unsigned int obmc_sad_w4(const uint8_t *pre,
+                                                 const int pre_stride,
+                                                 const int32_t *wsrc,
+                                                 const int32_t *mask,
+                                                 const int height) {
+  const int pre_step = pre_stride - 4;
+  int n = 0;
+  __m128i v_sad_d = _mm_setzero_si128();
+
+  do {
+    const __m128i v_p_b = xx_loadl_32(pre + n);
+    const __m128i v_m_d = xx_load_128(mask + n);
+    const __m128i v_w_d = xx_load_128(wsrc + n);
+
+    const __m128i v_p_d = _mm_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 __m128i v_pm_d = _mm_madd_epi16(v_p_d, v_m_d);
+
+    const __m128i v_diff_d = _mm_sub_epi32(v_w_d, v_pm_d);
+    const __m128i v_absdiff_d = _mm_abs_epi32(v_diff_d);
+
+    // Rounded absolute difference
+    const __m128i v_rad_d = xx_roundn_epu32(v_absdiff_d, 12);
+
+    v_sad_d = _mm_add_epi32(v_sad_d, v_rad_d);
+
+    n += 4;
+
+    if (n % 4 == 0) pre += pre_step;
+  } while (n < 4 * height);
+
+  return xx_hsum_epi32_si32(v_sad_d);
+}
+
+static AOM_FORCE_INLINE unsigned int obmc_sad_w8n(
+    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;
+  __m128i v_sad_d = _mm_setzero_si128();
+
+  assert(width >= 8);
+  assert(IS_POWER_OF_TWO(width));
+
+  do {
+    const __m128i v_p1_b = xx_loadl_32(pre + n + 4);
+    const __m128i v_m1_d = xx_load_128(mask + n + 4);
+    const __m128i v_w1_d = xx_load_128(wsrc + n + 4);
+    const __m128i v_p0_b = xx_loadl_32(pre + n);
+    const __m128i v_m0_d = xx_load_128(mask + n);
+    const __m128i v_w0_d = xx_load_128(wsrc + n);
+
+    const __m128i v_p0_d = _mm_cvtepu8_epi32(v_p0_b);
+    const __m128i v_p1_d = _mm_cvtepu8_epi32(v_p1_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 __m128i v_pm0_d = _mm_madd_epi16(v_p0_d, v_m0_d);
+    const __m128i v_pm1_d = _mm_madd_epi16(v_p1_d, v_m1_d);
+
+    const __m128i v_diff0_d = _mm_sub_epi32(v_w0_d, v_pm0_d);
+    const __m128i v_diff1_d = _mm_sub_epi32(v_w1_d, v_pm1_d);
+    const __m128i v_absdiff0_d = _mm_abs_epi32(v_diff0_d);
+    const __m128i v_absdiff1_d = _mm_abs_epi32(v_diff1_d);
+
+    // Rounded absolute difference
+    const __m128i v_rad0_d = xx_roundn_epu32(v_absdiff0_d, 12);
+    const __m128i v_rad1_d = xx_roundn_epu32(v_absdiff1_d, 12);
+
+    v_sad_d = _mm_add_epi32(v_sad_d, v_rad0_d);
+    v_sad_d = _mm_add_epi32(v_sad_d, v_rad1_d);
+
+    n += 8;
+
+    if (n % width == 0) pre += pre_step;
+  } while (n < width * height);
+
+  return xx_hsum_epi32_si32(v_sad_d);
+}
+
+#define OBMCSADWXH(w, h)                                       \
+  unsigned int aom_obmc_sad##w##x##h##_sse4_1(                 \
+      const uint8_t *pre, int pre_stride, const int32_t *wsrc, \
+      const int32_t *msk) {                                    \
+    if (w == 4) {                                              \
+      return obmc_sad_w4(pre, pre_stride, wsrc, msk, h);       \
+    } else {                                                   \
+      return obmc_sad_w8n(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 AOM_FORCE_INLINE unsigned int hbd_obmc_sad_w4(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);
+  const int pre_step = pre_stride - 4;
+  int n = 0;
+  __m128i v_sad_d = _mm_setzero_si128();
+
+  do {
+    const __m128i v_p_w = xx_loadl_64(pre + n);
+    const __m128i v_m_d = xx_load_128(mask + n);
+    const __m128i v_w_d = xx_load_128(wsrc + n);
+
+    const __m128i v_p_d = _mm_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 __m128i v_pm_d = _mm_madd_epi16(v_p_d, v_m_d);
+
+    const __m128i v_diff_d = _mm_sub_epi32(v_w_d, v_pm_d);
+    const __m128i v_absdiff_d = _mm_abs_epi32(v_diff_d);
+
+    // Rounded absolute difference
+    const __m128i v_rad_d = xx_roundn_epu32(v_absdiff_d, 12);
+
+    v_sad_d = _mm_add_epi32(v_sad_d, v_rad_d);
+
+    n += 4;
+
+    if (n % 4 == 0) pre += pre_step;
+  } while (n < 4 * height);
+
+  return xx_hsum_epi32_si32(v_sad_d);
+}
+
+static AOM_FORCE_INLINE unsigned int hbd_obmc_sad_w8n(
+    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;
+  __m128i v_sad_d = _mm_setzero_si128();
+
+  assert(width >= 8);
+  assert(IS_POWER_OF_TWO(width));
+
+  do {
+    const __m128i v_p1_w = xx_loadl_64(pre + n + 4);
+    const __m128i v_m1_d = xx_load_128(mask + n + 4);
+    const __m128i v_w1_d = xx_load_128(wsrc + n + 4);
+    const __m128i v_p0_w = xx_loadl_64(pre + n);
+    const __m128i v_m0_d = xx_load_128(mask + n);
+    const __m128i v_w0_d = xx_load_128(wsrc + n);
+
+    const __m128i v_p0_d = _mm_cvtepu16_epi32(v_p0_w);
+    const __m128i v_p1_d = _mm_cvtepu16_epi32(v_p1_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 __m128i v_pm0_d = _mm_madd_epi16(v_p0_d, v_m0_d);
+    const __m128i v_pm1_d = _mm_madd_epi16(v_p1_d, v_m1_d);
+
+    const __m128i v_diff0_d = _mm_sub_epi32(v_w0_d, v_pm0_d);
+    const __m128i v_diff1_d = _mm_sub_epi32(v_w1_d, v_pm1_d);
+    const __m128i v_absdiff0_d = _mm_abs_epi32(v_diff0_d);
+    const __m128i v_absdiff1_d = _mm_abs_epi32(v_diff1_d);
+
+    // Rounded absolute difference
+    const __m128i v_rad0_d = xx_roundn_epu32(v_absdiff0_d, 12);
+    const __m128i v_rad1_d = xx_roundn_epu32(v_absdiff1_d, 12);
+
+    v_sad_d = _mm_add_epi32(v_sad_d, v_rad0_d);
+    v_sad_d = _mm_add_epi32(v_sad_d, v_rad1_d);
+
+    n += 8;
+
+    if (n % width == 0) pre += pre_step;
+  } while (n < width * height);
+
+  return xx_hsum_epi32_si32(v_sad_d);
+}
+
+#define HBD_OBMCSADWXH(w, h)                                      \
+  unsigned int aom_highbd_obmc_sad##w##x##h##_sse4_1(             \
+      const uint8_t *pre, int pre_stride, const int32_t *wsrc,    \
+      const int32_t *mask) {                                      \
+    if (w == 4) {                                                 \
+      return hbd_obmc_sad_w4(pre, pre_stride, wsrc, mask, h);     \
+    } else {                                                      \
+      return hbd_obmc_sad_w8n(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_avx2.c b/third_party/aom/aom_dsp/x86/obmc_variance_avx2.c
new file mode 100644
index 0000000000..c23d8c4eb0
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/obmc_variance_avx2.c
@@ -0,0 +1,191 @@
+/*
+ * 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 "config/aom_dsp_rtcd.h"
+
+#include "aom_ports/mem.h"
+#include "aom/aom_integer.h"
+
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/aom_filter.h"
+#include "aom_dsp/x86/obmc_intrinsic_sse4.h"
+
+////////////////////////////////////////////////////////////////////////////////
+// 8 bit
+////////////////////////////////////////////////////////////////////////////////
+
+static INLINE void obmc_variance_w8n(const uint8_t *pre, const int pre_stride,
+                                     const int32_t *wsrc, const int32_t *mask,
+                                     unsigned int *const sse, int *const sum,
+                                     const int w, const int h) {
+  int n = 0, width, height = h;
+  __m128i v_sum_d = _mm_setzero_si128();
+  __m128i v_sse_d = _mm_setzero_si128();
+  const __m256i v_bias_d = _mm256_set1_epi32((1 << 12) >> 1);
+  __m128i v_d;
+  const uint8_t *pre_temp;
+  assert(w >= 8);
+  assert(IS_POWER_OF_TWO(w));
+  assert(IS_POWER_OF_TWO(h));
+  do {
+    width = w;
+    pre_temp = pre;
+    do {
+      const __m128i v_p_b = _mm_loadl_epi64((const __m128i *)pre_temp);
+      const __m256i v_m_d = _mm256_loadu_si256((__m256i const *)(mask + n));
+      const __m256i v_w_d = _mm256_loadu_si256((__m256i const *)(wsrc + n));
+      const __m256i v_p0_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_p0_d, v_m_d);
+      const __m256i v_diff0_d = _mm256_sub_epi32(v_w_d, v_pm_d);
+
+      const __m256i v_sign_d = _mm256_srai_epi32(v_diff0_d, 31);
+      const __m256i v_tmp_d =
+          _mm256_add_epi32(_mm256_add_epi32(v_diff0_d, v_bias_d), v_sign_d);
+      const __m256i v_rdiff0_d = _mm256_srai_epi32(v_tmp_d, 12);
+      const __m128i v_rdiff_d = _mm256_castsi256_si128(v_rdiff0_d);
+      const __m128i v_rdiff1_d = _mm256_extracti128_si256(v_rdiff0_d, 1);
+
+      const __m128i v_rdiff01_w = _mm_packs_epi32(v_rdiff_d, v_rdiff1_d);
+      const __m128i v_sqrdiff_d = _mm_madd_epi16(v_rdiff01_w, v_rdiff01_w);
+
+      v_sum_d = _mm_add_epi32(v_sum_d, v_rdiff_d);
+      v_sum_d = _mm_add_epi32(v_sum_d, v_rdiff1_d);
+      v_sse_d = _mm_add_epi32(v_sse_d, v_sqrdiff_d);
+
+      pre_temp += 8;
+      n += 8;
+      width -= 8;
+    } while (width > 0);
+    pre += pre_stride;
+    height -= 1;
+  } while (height > 0);
+  v_d = _mm_hadd_epi32(v_sum_d, v_sse_d);
+  v_d = _mm_hadd_epi32(v_d, v_d);
+  *sum = _mm_cvtsi128_si32(v_d);
+  *sse = (unsigned int)_mm_cvtsi128_si32(_mm_srli_si128(v_d, 4));
+}
+
+static INLINE void obmc_variance_w16n(const uint8_t *pre, const int pre_stride,
+                                      const int32_t *wsrc, const int32_t *mask,
+                                      unsigned int *const sse, int *const sum,
+                                      const int w, const int h) {
+  int n = 0, width, height = h;
+  __m256i v_d;
+  __m128i res0;
+  const uint8_t *pre_temp;
+  const __m256i v_bias_d = _mm256_set1_epi32((1 << 12) >> 1);
+  __m256i v_sum_d = _mm256_setzero_si256();
+  __m256i v_sse_d = _mm256_setzero_si256();
+
+  assert(w >= 16);
+  assert(IS_POWER_OF_TWO(w));
+  assert(IS_POWER_OF_TWO(h));
+  do {
+    width = w;
+    pre_temp = pre;
+    do {
+      const __m128i v_p_b = _mm_loadu_si128((__m128i *)pre_temp);
+      const __m256i v_m0_d = _mm256_loadu_si256((__m256i const *)(mask + n));
+      const __m256i v_w0_d = _mm256_loadu_si256((__m256i const *)(wsrc + n));
+      const __m256i v_m1_d =
+          _mm256_loadu_si256((__m256i const *)(mask + n + 8));
+      const __m256i v_w1_d =
+          _mm256_loadu_si256((__m256i const *)(wsrc + n + 8));
+
+      const __m256i v_p0_d = _mm256_cvtepu8_epi32(v_p_b);
+      const __m256i v_p1_d = _mm256_cvtepu8_epi32(_mm_srli_si128(v_p_b, 8));
+
+      const __m256i v_pm0_d = _mm256_madd_epi16(v_p0_d, v_m0_d);
+      const __m256i v_pm1_d = _mm256_madd_epi16(v_p1_d, v_m1_d);
+
+      const __m256i v_diff0_d = _mm256_sub_epi32(v_w0_d, v_pm0_d);
+      const __m256i v_diff1_d = _mm256_sub_epi32(v_w1_d, v_pm1_d);
+
+      const __m256i v_sign0_d = _mm256_srai_epi32(v_diff0_d, 31);
+      const __m256i v_sign1_d = _mm256_srai_epi32(v_diff1_d, 31);
+
+      const __m256i v_tmp0_d =
+          _mm256_add_epi32(_mm256_add_epi32(v_diff0_d, v_bias_d), v_sign0_d);
+      const __m256i v_tmp1_d =
+          _mm256_add_epi32(_mm256_add_epi32(v_diff1_d, v_bias_d), v_sign1_d);
+
+      const __m256i v_rdiff0_d = _mm256_srai_epi32(v_tmp0_d, 12);
+      const __m256i v_rdiff2_d = _mm256_srai_epi32(v_tmp1_d, 12);
+
+      const __m256i v_rdiff1_d = _mm256_add_epi32(v_rdiff0_d, v_rdiff2_d);
+      const __m256i v_rdiff01_w = _mm256_packs_epi32(v_rdiff0_d, v_rdiff2_d);
+      const __m256i v_sqrdiff_d = _mm256_madd_epi16(v_rdiff01_w, v_rdiff01_w);
+
+      v_sum_d = _mm256_add_epi32(v_sum_d, v_rdiff1_d);
+      v_sse_d = _mm256_add_epi32(v_sse_d, v_sqrdiff_d);
+
+      pre_temp += 16;
+      n += 16;
+      width -= 16;
+    } while (width > 0);
+    pre += pre_stride;
+    height -= 1;
+  } while (height > 0);
+
+  v_d = _mm256_hadd_epi32(v_sum_d, v_sse_d);
+  v_d = _mm256_hadd_epi32(v_d, v_d);
+  res0 = _mm256_castsi256_si128(v_d);
+  res0 = _mm_add_epi32(res0, _mm256_extractf128_si256(v_d, 1));
+  *sum = _mm_cvtsi128_si32(res0);
+  *sse = (unsigned int)_mm_cvtsi128_si32(_mm_srli_si128(res0, 4));
+}
+
+#define OBMCVARWXH(W, H)                                                \
+  unsigned int aom_obmc_variance##W##x##H##_avx2(                       \
+      const uint8_t *pre, int pre_stride, const int32_t *wsrc,          \
+      const int32_t *mask, unsigned int *sse) {                         \
+    int sum;                                                            \
+    if (W == 4) {                                                       \
+      obmc_variance_w4(pre, pre_stride, wsrc, mask, sse, &sum, H);      \
+    } else if (W == 8) {                                                \
+      obmc_variance_w8n(pre, pre_stride, wsrc, mask, sse, &sum, W, H);  \
+    } else {                                                            \
+      obmc_variance_w16n(pre, pre_stride, wsrc, mask, sse, &sum, W, H); \
+    }                                                                   \
+                                                                        \
+    return *sse - (unsigned int)(((int64_t)sum * sum) / (W * H));       \
+  }
+
+OBMCVARWXH(128, 128)
+OBMCVARWXH(128, 64)
+OBMCVARWXH(64, 128)
+OBMCVARWXH(64, 64)
+OBMCVARWXH(64, 32)
+OBMCVARWXH(32, 64)
+OBMCVARWXH(32, 32)
+OBMCVARWXH(32, 16)
+OBMCVARWXH(16, 32)
+OBMCVARWXH(16, 16)
+OBMCVARWXH(16, 8)
+OBMCVARWXH(8, 16)
+OBMCVARWXH(8, 8)
+OBMCVARWXH(8, 4)
+OBMCVARWXH(4, 8)
+OBMCVARWXH(4, 4)
+OBMCVARWXH(4, 16)
+OBMCVARWXH(16, 4)
+OBMCVARWXH(8, 32)
+OBMCVARWXH(32, 8)
+OBMCVARWXH(16, 64)
+OBMCVARWXH(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
new file mode 100644
index 0000000000..89b050eb20
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/obmc_variance_sse4.c
@@ -0,0 +1,382 @@
+/*
+ * Copyright (c) 2016, 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 "config/aom_dsp_rtcd.h"
+
+#include "aom_ports/mem.h"
+#include "aom/aom_integer.h"
+
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/aom_filter.h"
+#include "aom_dsp/x86/obmc_intrinsic_sse4.h"
+#include "aom_dsp/x86/synonyms.h"
+
+////////////////////////////////////////////////////////////////////////////////
+// 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_w8n(const uint8_t *pre, const int pre_stride,
+                                     const int32_t *wsrc, const int32_t *mask,
+                                     unsigned int *const sse, int *const sum,
+                                     const int w, const int h) {
+  const int pre_step = pre_stride - w;
+  int n = 0;
+  __m128i v_sum_d = _mm_setzero_si128();
+  __m128i v_sse_d = _mm_setzero_si128();
+
+  assert(w >= 8);
+  assert(IS_POWER_OF_TWO(w));
+  assert(IS_POWER_OF_TWO(h));
+
+  do {
+    const __m128i v_p1_b = xx_loadl_32(pre + n + 4);
+    const __m128i v_m1_d = xx_load_128(mask + n + 4);
+    const __m128i v_w1_d = xx_load_128(wsrc + n + 4);
+    const __m128i v_p0_b = xx_loadl_32(pre + n);
+    const __m128i v_m0_d = xx_load_128(mask + n);
+    const __m128i v_w0_d = xx_load_128(wsrc + n);
+
+    const __m128i v_p0_d = _mm_cvtepu8_epi32(v_p0_b);
+    const __m128i v_p1_d = _mm_cvtepu8_epi32(v_p1_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 __m128i v_pm0_d = _mm_madd_epi16(v_p0_d, v_m0_d);
+    const __m128i v_pm1_d = _mm_madd_epi16(v_p1_d, v_m1_d);
+
+    const __m128i v_diff0_d = _mm_sub_epi32(v_w0_d, v_pm0_d);
+    const __m128i v_diff1_d = _mm_sub_epi32(v_w1_d, v_pm1_d);
+
+    const __m128i v_rdiff0_d = xx_roundn_epi32(v_diff0_d, 12);
+    const __m128i v_rdiff1_d = xx_roundn_epi32(v_diff1_d, 12);
+    const __m128i v_rdiff01_w = _mm_packs_epi32(v_rdiff0_d, v_rdiff1_d);
+    const __m128i v_sqrdiff_d = _mm_madd_epi16(v_rdiff01_w, v_rdiff01_w);
+
+    v_sum_d = _mm_add_epi32(v_sum_d, v_rdiff0_d);
+    v_sum_d = _mm_add_epi32(v_sum_d, v_rdiff1_d);
+    v_sse_d = _mm_add_epi32(v_sse_d, v_sqrdiff_d);
+
+    n += 8;
+
+    if (n % w == 0) pre += pre_step;
+  } while (n < w * h);
+
+  *sum = xx_hsum_epi32_si32(v_sum_d);
+  *sse = xx_hsum_epi32_si32(v_sse_d);
+}
+
+#define OBMCVARWXH(W, H)                                               \
+  unsigned int aom_obmc_variance##W##x##H##_sse4_1(                    \
+      const uint8_t *pre, int pre_stride, const int32_t *wsrc,         \
+      const int32_t *mask, unsigned int *sse) {                        \
+    int sum;                                                           \
+    if (W == 4) {                                                      \
+      obmc_variance_w4(pre, pre_stride, wsrc, mask, sse, &sum, H);     \
+    } else {                                                           \
+      obmc_variance_w8n(pre, pre_stride, wsrc, mask, sse, &sum, W, H); \
+    }                                                                  \
+    return *sse - (unsigned int)(((int64_t)sum * sum) / (W * H));      \
+  }
+
+OBMCVARWXH(128, 128)
+OBMCVARWXH(128, 64)
+OBMCVARWXH(64, 128)
+OBMCVARWXH(64, 64)
+OBMCVARWXH(64, 32)
+OBMCVARWXH(32, 64)
+OBMCVARWXH(32, 32)
+OBMCVARWXH(32, 16)
+OBMCVARWXH(16, 32)
+OBMCVARWXH(16, 16)
+OBMCVARWXH(16, 8)
+OBMCVARWXH(8, 16)
+OBMCVARWXH(8, 8)
+OBMCVARWXH(8, 4)
+OBMCVARWXH(4, 8)
+OBMCVARWXH(4, 4)
+OBMCVARWXH(4, 16)
+OBMCVARWXH(16, 4)
+OBMCVARWXH(8, 32)
+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
+////////////////////////////////////////////////////////////////////////////////
+#if CONFIG_AV1_HIGHBITDEPTH
+static INLINE void hbd_obmc_variance_w4(
+    const uint8_t *pre8, const int pre_stride, const int32_t *wsrc,
+    const int32_t *mask, uint64_t *const sse, int64_t *const sum, const int h) {
+  const uint16_t *pre = CONVERT_TO_SHORTPTR(pre8);
+  const int pre_step = pre_stride - 4;
+  int n = 0;
+  __m128i v_sum_d = _mm_setzero_si128();
+  __m128i v_sse_d = _mm_setzero_si128();
+
+  assert(IS_POWER_OF_TWO(h));
+
+  do {
+    const __m128i v_p_w = xx_loadl_64(pre + n);
+    const __m128i v_m_d = xx_load_128(mask + n);
+    const __m128i v_w_d = xx_load_128(wsrc + n);
+
+    const __m128i v_p_d = _mm_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 __m128i v_pm_d = _mm_madd_epi16(v_p_d, v_m_d);
+
+    const __m128i v_diff_d = _mm_sub_epi32(v_w_d, v_pm_d);
+    const __m128i v_rdiff_d = xx_roundn_epi32(v_diff_d, 12);
+    const __m128i v_sqrdiff_d = _mm_mullo_epi32(v_rdiff_d, v_rdiff_d);
+
+    v_sum_d = _mm_add_epi32(v_sum_d, v_rdiff_d);
+    v_sse_d = _mm_add_epi32(v_sse_d, v_sqrdiff_d);
+
+    n += 4;
+
+    if (n % 4 == 0) pre += pre_step;
+  } while (n < 4 * h);
+
+  *sum = xx_hsum_epi32_si32(v_sum_d);
+  *sse = xx_hsum_epi32_si32(v_sse_d);
+}
+
+static INLINE void hbd_obmc_variance_w8n(
+    const uint8_t *pre8, const int pre_stride, const int32_t *wsrc,
+    const int32_t *mask, uint64_t *const sse, int64_t *const sum, const int w,
+    const int h) {
+  const uint16_t *pre = CONVERT_TO_SHORTPTR(pre8);
+  const int pre_step = pre_stride - w;
+  int n = 0;
+  __m128i v_sum_d = _mm_setzero_si128();
+  __m128i v_sse_d = _mm_setzero_si128();
+
+  assert(w >= 8);
+  assert(IS_POWER_OF_TWO(w));
+  assert(IS_POWER_OF_TWO(h));
+
+  do {
+    const __m128i v_p1_w = xx_loadl_64(pre + n + 4);
+    const __m128i v_m1_d = xx_load_128(mask + n + 4);
+    const __m128i v_w1_d = xx_load_128(wsrc + n + 4);
+    const __m128i v_p0_w = xx_loadl_64(pre + n);
+    const __m128i v_m0_d = xx_load_128(mask + n);
+    const __m128i v_w0_d = xx_load_128(wsrc + n);
+
+    const __m128i v_p0_d = _mm_cvtepu16_epi32(v_p0_w);
+    const __m128i v_p1_d = _mm_cvtepu16_epi32(v_p1_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 __m128i v_pm0_d = _mm_madd_epi16(v_p0_d, v_m0_d);
+    const __m128i v_pm1_d = _mm_madd_epi16(v_p1_d, v_m1_d);
+
+    const __m128i v_diff0_d = _mm_sub_epi32(v_w0_d, v_pm0_d);
+    const __m128i v_diff1_d = _mm_sub_epi32(v_w1_d, v_pm1_d);
+
+    const __m128i v_rdiff0_d = xx_roundn_epi32(v_diff0_d, 12);
+    const __m128i v_rdiff1_d = xx_roundn_epi32(v_diff1_d, 12);
+    const __m128i v_rdiff01_w = _mm_packs_epi32(v_rdiff0_d, v_rdiff1_d);
+    const __m128i v_sqrdiff_d = _mm_madd_epi16(v_rdiff01_w, v_rdiff01_w);
+
+    v_sum_d = _mm_add_epi32(v_sum_d, v_rdiff0_d);
+    v_sum_d = _mm_add_epi32(v_sum_d, v_rdiff1_d);
+    v_sse_d = _mm_add_epi32(v_sse_d, v_sqrdiff_d);
+
+    n += 8;
+
+    if (n % w == 0) pre += pre_step;
+  } while (n < w * h);
+
+  *sum += xx_hsum_epi32_si64(v_sum_d);
+  *sse += xx_hsum_epi32_si64(v_sse_d);
+}
+
+static INLINE void highbd_8_obmc_variance(const uint8_t *pre8, int pre_stride,
+                                          const int32_t *wsrc,
+                                          const int32_t *mask, int w, int h,
+                                          unsigned int *sse, int *sum) {
+  int64_t sum64 = 0;
+  uint64_t sse64 = 0;
+  if (w == 4) {
+    hbd_obmc_variance_w4(pre8, pre_stride, wsrc, mask, &sse64, &sum64, h);
+  } else {
+    hbd_obmc_variance_w8n(pre8, pre_stride, wsrc, mask, &sse64, &sum64, w, h);
+  }
+  *sum = (int)sum64;
+  *sse = (unsigned int)sse64;
+}
+
+static INLINE void highbd_10_obmc_variance(const uint8_t *pre8, int pre_stride,
+                                           const int32_t *wsrc,
+                                           const int32_t *mask, int w, int h,
+                                           unsigned int *sse, int *sum) {
+  int64_t sum64 = 0;
+  uint64_t sse64 = 0;
+  if (w == 4) {
+    hbd_obmc_variance_w4(pre8, pre_stride, wsrc, mask, &sse64, &sum64, h);
+  } else if (w < 128 || h < 128) {
+    hbd_obmc_variance_w8n(pre8, pre_stride, wsrc, mask, &sse64, &sum64, w, h);
+  } else {
+    assert(w == 128 && h == 128);
+
+    do {
+      hbd_obmc_variance_w8n(pre8, pre_stride, wsrc, mask, &sse64, &sum64, w,
+                            64);
+      pre8 += 64 * pre_stride;
+      wsrc += 64 * w;
+      mask += 64 * w;
+      h -= 64;
+    } while (h > 0);
+  }
+  *sum = (int)ROUND_POWER_OF_TWO(sum64, 2);
+  *sse = (unsigned int)ROUND_POWER_OF_TWO(sse64, 4);
+}
+
+static INLINE void highbd_12_obmc_variance(const uint8_t *pre8, int pre_stride,
+                                           const int32_t *wsrc,
+                                           const int32_t *mask, int w, int h,
+                                           unsigned int *sse, int *sum) {
+  int64_t sum64 = 0;
+  uint64_t sse64 = 0;
+  int max_pel_allowed_per_ovf = 512;
+  if (w == 4) {
+    hbd_obmc_variance_w4(pre8, pre_stride, wsrc, mask, &sse64, &sum64, h);
+  } else if (w * h <= max_pel_allowed_per_ovf) {
+    hbd_obmc_variance_w8n(pre8, pre_stride, wsrc, mask, &sse64, &sum64, w, h);
+  } else {
+    int h_per_ovf = max_pel_allowed_per_ovf / w;
+
+    assert(max_pel_allowed_per_ovf % w == 0);
+    do {
+      hbd_obmc_variance_w8n(pre8, pre_stride, wsrc, mask, &sse64, &sum64, w,
+                            h_per_ovf);
+      pre8 += h_per_ovf * pre_stride;
+      wsrc += h_per_ovf * w;
+      mask += h_per_ovf * w;
+      h -= h_per_ovf;
+    } while (h > 0);
+  }
+  *sum = (int)ROUND_POWER_OF_TWO(sum64, 4);
+  *sse = (unsigned int)ROUND_POWER_OF_TWO(sse64, 8);
+}
+
+#define HBD_OBMCVARWXH(W, H)                                               \
+  unsigned int aom_highbd_8_obmc_variance##W##x##H##_sse4_1(               \
+      const uint8_t *pre, int pre_stride, const int32_t *wsrc,             \
+      const int32_t *mask, unsigned int *sse) {                            \
+    int sum;                                                               \
+    highbd_8_obmc_variance(pre, pre_stride, wsrc, mask, W, H, sse, &sum);  \
+    return *sse - (unsigned int)(((int64_t)sum * sum) / (W * H));          \
+  }                                                                        \
+                                                                           \
+  unsigned int aom_highbd_10_obmc_variance##W##x##H##_sse4_1(              \
+      const uint8_t *pre, int pre_stride, const int32_t *wsrc,             \
+      const int32_t *mask, unsigned int *sse) {                            \
+    int sum;                                                               \
+    int64_t var;                                                           \
+    highbd_10_obmc_variance(pre, pre_stride, wsrc, mask, W, H, sse, &sum); \
+    var = (int64_t)(*sse) - (((int64_t)sum * sum) / (W * H));              \
+    return (var >= 0) ? (uint32_t)var : 0;                                 \
+  }                                                                        \
+                                                                           \
+  unsigned int aom_highbd_12_obmc_variance##W##x##H##_sse4_1(              \
+      const uint8_t *pre, int pre_stride, const int32_t *wsrc,             \
+      const int32_t *mask, unsigned int *sse) {                            \
+    int sum;                                                               \
+    int64_t var;                                                           \
+    highbd_12_obmc_variance(pre, pre_stride, wsrc, mask, W, H, sse, &sum); \
+    var = (int64_t)(*sse) - (((int64_t)sum * sum) / (W * H));              \
+    return (var >= 0) ? (uint32_t)var : 0;                                 \
+  }
+
+HBD_OBMCVARWXH(128, 128)
+HBD_OBMCVARWXH(128, 64)
+HBD_OBMCVARWXH(64, 128)
+HBD_OBMCVARWXH(64, 64)
+HBD_OBMCVARWXH(64, 32)
+HBD_OBMCVARWXH(32, 64)
+HBD_OBMCVARWXH(32, 32)
+HBD_OBMCVARWXH(32, 16)
+HBD_OBMCVARWXH(16, 32)
+HBD_OBMCVARWXH(16, 16)
+HBD_OBMCVARWXH(16, 8)
+HBD_OBMCVARWXH(8, 16)
+HBD_OBMCVARWXH(8, 8)
+HBD_OBMCVARWXH(8, 4)
+HBD_OBMCVARWXH(4, 8)
+HBD_OBMCVARWXH(4, 4)
+HBD_OBMCVARWXH(4, 16)
+HBD_OBMCVARWXH(16, 4)
+HBD_OBMCVARWXH(8, 32)
+HBD_OBMCVARWXH(32, 8)
+HBD_OBMCVARWXH(16, 64)
+HBD_OBMCVARWXH(64, 16)
+#endif  // CONFIG_AV1_HIGHBITDEPTH
diff --git a/third_party/aom/aom_dsp/x86/quantize_avx2.c b/third_party/aom/aom_dsp/x86/quantize_avx2.c
new file mode 100644
index 0000000000..b808d46778
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/quantize_avx2.c
@@ -0,0 +1,274 @@
+/*
+ * Copyright (c) 2022, 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"
+#include "aom/aom_integer.h"
+#include "aom_dsp/x86/quantize_x86.h"
+
+static INLINE void load_b_values_avx2(const int16_t *zbin_ptr, __m256i *zbin,
+                                      const int16_t *round_ptr, __m256i *round,
+                                      const int16_t *quant_ptr, __m256i *quant,
+                                      const int16_t *dequant_ptr,
+                                      __m256i *dequant,
+                                      const int16_t *shift_ptr, __m256i *shift,
+                                      int log_scale) {
+  *zbin = _mm256_castsi128_si256(_mm_load_si128((const __m128i *)zbin_ptr));
+  *zbin = _mm256_permute4x64_epi64(*zbin, 0x54);
+  if (log_scale > 0) {
+    const __m256i rnd = _mm256_set1_epi16((int16_t)(1 << (log_scale - 1)));
+    *zbin = _mm256_add_epi16(*zbin, rnd);
+    *zbin = _mm256_srai_epi16(*zbin, log_scale);
+  }
+  // Subtracting 1 here eliminates a _mm256_cmpeq_epi16() instruction when
+  // calculating the zbin mask. (See quantize_b_logscale{0,1,2}_16)
+  *zbin = _mm256_sub_epi16(*zbin, _mm256_set1_epi16(1));
+
+  *round = _mm256_castsi128_si256(_mm_load_si128((const __m128i *)round_ptr));
+  *round = _mm256_permute4x64_epi64(*round, 0x54);
+  if (log_scale > 0) {
+    const __m256i rnd = _mm256_set1_epi16((int16_t)(1 << (log_scale - 1)));
+    *round = _mm256_add_epi16(*round, rnd);
+    *round = _mm256_srai_epi16(*round, log_scale);
+  }
+
+  *quant = _mm256_castsi128_si256(_mm_load_si128((const __m128i *)quant_ptr));
+  *quant = _mm256_permute4x64_epi64(*quant, 0x54);
+  *dequant =
+      _mm256_castsi128_si256(_mm_load_si128((const __m128i *)dequant_ptr));
+  *dequant = _mm256_permute4x64_epi64(*dequant, 0x54);
+  *shift = _mm256_castsi128_si256(_mm_load_si128((const __m128i *)shift_ptr));
+  *shift = _mm256_permute4x64_epi64(*shift, 0x54);
+}
+
+static INLINE __m256i load_coefficients_avx2(const tran_low_t *coeff_ptr) {
+  const __m256i coeff1 = _mm256_load_si256((__m256i *)coeff_ptr);
+  const __m256i coeff2 = _mm256_load_si256((__m256i *)(coeff_ptr + 8));
+  return _mm256_packs_epi32(coeff1, coeff2);
+}
+
+static INLINE void store_coefficients_avx2(__m256i coeff_vals,
+                                           tran_low_t *coeff_ptr) {
+  __m256i coeff_sign = _mm256_srai_epi16(coeff_vals, 15);
+  __m256i coeff_vals_lo = _mm256_unpacklo_epi16(coeff_vals, coeff_sign);
+  __m256i coeff_vals_hi = _mm256_unpackhi_epi16(coeff_vals, coeff_sign);
+  _mm256_store_si256((__m256i *)coeff_ptr, coeff_vals_lo);
+  _mm256_store_si256((__m256i *)(coeff_ptr + 8), coeff_vals_hi);
+}
+
+static AOM_FORCE_INLINE __m256i quantize_b_logscale0_16(
+    const tran_low_t *coeff_ptr, tran_low_t *qcoeff_ptr,
+    tran_low_t *dqcoeff_ptr, __m256i *v_quant, __m256i *v_dequant,
+    __m256i *v_round, __m256i *v_zbin, __m256i *v_quant_shift) {
+  const __m256i v_coeff = load_coefficients_avx2(coeff_ptr);
+  const __m256i v_abs_coeff = _mm256_abs_epi16(v_coeff);
+  const __m256i v_zbin_mask = _mm256_cmpgt_epi16(v_abs_coeff, *v_zbin);
+
+  if (_mm256_movemask_epi8(v_zbin_mask) == 0) {
+    _mm256_store_si256((__m256i *)qcoeff_ptr, _mm256_setzero_si256());
+    _mm256_store_si256((__m256i *)dqcoeff_ptr, _mm256_setzero_si256());
+    _mm256_store_si256((__m256i *)(qcoeff_ptr + 8), _mm256_setzero_si256());
+    _mm256_store_si256((__m256i *)(dqcoeff_ptr + 8), _mm256_setzero_si256());
+    return _mm256_setzero_si256();
+  }
+
+  // tmp = v_zbin_mask ? (int64_t)abs_coeff + log_scaled_round : 0
+  const __m256i v_tmp_rnd =
+      _mm256_and_si256(_mm256_adds_epi16(v_abs_coeff, *v_round), v_zbin_mask);
+  //  tmp32 = (int)(((((tmp * quant_ptr[rc != 0]) >> 16) + tmp) *
+  //                 quant_shift_ptr[rc != 0]) >>
+  //                (16 - log_scale + AOM_QM_BITS));
+  const __m256i v_tmp32_a = _mm256_mulhi_epi16(v_tmp_rnd, *v_quant);
+  const __m256i v_tmp32_b = _mm256_add_epi16(v_tmp32_a, v_tmp_rnd);
+  const __m256i v_tmp32 = _mm256_mulhi_epi16(v_tmp32_b, *v_quant_shift);
+  const __m256i v_nz_mask = _mm256_cmpgt_epi16(v_tmp32, _mm256_setzero_si256());
+  const __m256i v_qcoeff = _mm256_sign_epi16(v_tmp32, v_coeff);
+  const __m256i v_dqcoeff = _mm256_mullo_epi16(v_qcoeff, *v_dequant);
+  store_coefficients_avx2(v_qcoeff, qcoeff_ptr);
+  store_coefficients_avx2(v_dqcoeff, dqcoeff_ptr);
+  return v_nz_mask;
+}
+
+static INLINE __m256i get_max_lane_eob(const int16_t *iscan, __m256i v_eobmax,
+                                       __m256i v_mask) {
+  const __m256i v_iscan = _mm256_loadu_si256((const __m256i *)iscan);
+  const __m256i v_iscan_perm = _mm256_permute4x64_epi64(v_iscan, 0xD8);
+  const __m256i v_iscan_plus1 = _mm256_sub_epi16(v_iscan_perm, v_mask);
+  const __m256i v_nz_iscan = _mm256_and_si256(v_iscan_plus1, v_mask);
+  return _mm256_max_epi16(v_eobmax, v_nz_iscan);
+}
+
+static INLINE int16_t accumulate_eob256(__m256i eob256) {
+  const __m128i eob_lo = _mm256_castsi256_si128(eob256);
+  const __m128i eob_hi = _mm256_extractf128_si256(eob256, 1);
+  __m128i eob = _mm_max_epi16(eob_lo, eob_hi);
+  __m128i eob_shuffled = _mm_shuffle_epi32(eob, 0xe);
+  eob = _mm_max_epi16(eob, eob_shuffled);
+  eob_shuffled = _mm_shufflelo_epi16(eob, 0xe);
+  eob = _mm_max_epi16(eob, eob_shuffled);
+  eob_shuffled = _mm_shufflelo_epi16(eob, 0x1);
+  eob = _mm_max_epi16(eob, eob_shuffled);
+  return _mm_extract_epi16(eob, 1);
+}
+
+void aom_quantize_b_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                         const int16_t *zbin_ptr, const int16_t *round_ptr,
+                         const int16_t *quant_ptr,
+                         const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
+                         tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr,
+                         uint16_t *eob_ptr, const int16_t *scan,
+                         const int16_t *iscan) {
+  (void)scan;
+  __m256i v_zbin, v_round, v_quant, v_dequant, v_quant_shift;
+  __m256i v_eobmax = _mm256_setzero_si256();
+
+  load_b_values_avx2(zbin_ptr, &v_zbin, round_ptr, &v_round, quant_ptr,
+                     &v_quant, dequant_ptr, &v_dequant, quant_shift_ptr,
+                     &v_quant_shift, 0);
+
+  // Do DC and first 15 AC.
+  __m256i v_nz_mask =
+      quantize_b_logscale0_16(coeff_ptr, qcoeff_ptr, dqcoeff_ptr, &v_quant,
+                              &v_dequant, &v_round, &v_zbin, &v_quant_shift);
+
+  v_eobmax = get_max_lane_eob(iscan, v_eobmax, v_nz_mask);
+
+  v_round = _mm256_unpackhi_epi64(v_round, v_round);
+  v_quant = _mm256_unpackhi_epi64(v_quant, v_quant);
+  v_dequant = _mm256_unpackhi_epi64(v_dequant, v_dequant);
+  v_quant_shift = _mm256_unpackhi_epi64(v_quant_shift, v_quant_shift);
+  v_zbin = _mm256_unpackhi_epi64(v_zbin, v_zbin);
+
+  for (intptr_t count = n_coeffs - 16; count > 0; count -= 16) {
+    coeff_ptr += 16;
+    qcoeff_ptr += 16;
+    dqcoeff_ptr += 16;
+    iscan += 16;
+    v_nz_mask =
+        quantize_b_logscale0_16(coeff_ptr, qcoeff_ptr, dqcoeff_ptr, &v_quant,
+                                &v_dequant, &v_round, &v_zbin, &v_quant_shift);
+
+    v_eobmax = get_max_lane_eob(iscan, v_eobmax, v_nz_mask);
+  }
+
+  *eob_ptr = accumulate_eob256(v_eobmax);
+}
+
+static AOM_FORCE_INLINE __m256i quantize_b_logscale_16(
+    const tran_low_t *coeff_ptr, tran_low_t *qcoeff_ptr,
+    tran_low_t *dqcoeff_ptr, __m256i *v_quant, __m256i *v_dequant,
+    __m256i *v_round, __m256i *v_zbin, __m256i *v_quant_shift, int log_scale) {
+  const __m256i v_coeff = load_coefficients_avx2(coeff_ptr);
+  const __m256i v_abs_coeff = _mm256_abs_epi16(v_coeff);
+  const __m256i v_zbin_mask = _mm256_cmpgt_epi16(v_abs_coeff, *v_zbin);
+
+  if (_mm256_movemask_epi8(v_zbin_mask) == 0) {
+    _mm256_store_si256((__m256i *)qcoeff_ptr, _mm256_setzero_si256());
+    _mm256_store_si256((__m256i *)dqcoeff_ptr, _mm256_setzero_si256());
+    _mm256_store_si256((__m256i *)(qcoeff_ptr + 8), _mm256_setzero_si256());
+    _mm256_store_si256((__m256i *)(dqcoeff_ptr + 8), _mm256_setzero_si256());
+    return _mm256_setzero_si256();
+  }
+
+  // tmp = v_zbin_mask ? (int64_t)abs_coeff + log_scaled_round : 0
+  const __m256i v_tmp_rnd =
+      _mm256_and_si256(_mm256_adds_epi16(v_abs_coeff, *v_round), v_zbin_mask);
+  //  tmp32 = (int)(((((tmp * quant_ptr[rc != 0]) >> 16) + tmp) *
+  //                 quant_shift_ptr[rc != 0]) >>
+  //                (16 - log_scale + AOM_QM_BITS));
+  const __m256i v_tmp32_a = _mm256_mulhi_epi16(v_tmp_rnd, *v_quant);
+  const __m256i v_tmp32_b = _mm256_add_epi16(v_tmp32_a, v_tmp_rnd);
+  const __m256i v_tmp32_hi = _mm256_slli_epi16(
+      _mm256_mulhi_epi16(v_tmp32_b, *v_quant_shift), log_scale);
+  const __m256i v_tmp32_lo = _mm256_srli_epi16(
+      _mm256_mullo_epi16(v_tmp32_b, *v_quant_shift), 16 - log_scale);
+  const __m256i v_tmp32 = _mm256_or_si256(v_tmp32_hi, v_tmp32_lo);
+  const __m256i v_dqcoeff_hi = _mm256_slli_epi16(
+      _mm256_mulhi_epi16(v_tmp32, *v_dequant), 16 - log_scale);
+  const __m256i v_dqcoeff_lo =
+      _mm256_srli_epi16(_mm256_mullo_epi16(v_tmp32, *v_dequant), log_scale);
+  const __m256i v_dqcoeff =
+      _mm256_sign_epi16(_mm256_or_si256(v_dqcoeff_hi, v_dqcoeff_lo), v_coeff);
+  const __m256i v_qcoeff = _mm256_sign_epi16(v_tmp32, v_coeff);
+  const __m256i v_nz_mask = _mm256_cmpgt_epi16(v_tmp32, _mm256_setzero_si256());
+  store_coefficients_avx2(v_qcoeff, qcoeff_ptr);
+  store_coefficients_avx2(v_dqcoeff, dqcoeff_ptr);
+  return v_nz_mask;
+}
+
+static AOM_FORCE_INLINE void quantize_b_no_qmatrix_avx2(
+    const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr,
+    const int16_t *round_ptr, const int16_t *quant_ptr,
+    const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
+    tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
+    const int16_t *iscan, int log_scale) {
+  __m256i v_zbin, v_round, v_quant, v_dequant, v_quant_shift;
+  __m256i v_eobmax = _mm256_setzero_si256();
+
+  load_b_values_avx2(zbin_ptr, &v_zbin, round_ptr, &v_round, quant_ptr,
+                     &v_quant, dequant_ptr, &v_dequant, quant_shift_ptr,
+                     &v_quant_shift, log_scale);
+
+  // Do DC and first 15 AC.
+  __m256i v_nz_mask = quantize_b_logscale_16(
+      coeff_ptr, qcoeff_ptr, dqcoeff_ptr, &v_quant, &v_dequant, &v_round,
+      &v_zbin, &v_quant_shift, log_scale);
+
+  v_eobmax = get_max_lane_eob(iscan, v_eobmax, v_nz_mask);
+
+  v_round = _mm256_unpackhi_epi64(v_round, v_round);
+  v_quant = _mm256_unpackhi_epi64(v_quant, v_quant);
+  v_dequant = _mm256_unpackhi_epi64(v_dequant, v_dequant);
+  v_quant_shift = _mm256_unpackhi_epi64(v_quant_shift, v_quant_shift);
+  v_zbin = _mm256_unpackhi_epi64(v_zbin, v_zbin);
+
+  for (intptr_t count = n_coeffs - 16; count > 0; count -= 16) {
+    coeff_ptr += 16;
+    qcoeff_ptr += 16;
+    dqcoeff_ptr += 16;
+    iscan += 16;
+    v_nz_mask = quantize_b_logscale_16(coeff_ptr, qcoeff_ptr, dqcoeff_ptr,
+                                       &v_quant, &v_dequant, &v_round, &v_zbin,
+                                       &v_quant_shift, log_scale);
+
+    v_eobmax = get_max_lane_eob(iscan, v_eobmax, v_nz_mask);
+  }
+
+  *eob_ptr = accumulate_eob256(v_eobmax);
+}
+
+void aom_quantize_b_32x32_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                               const int16_t *zbin_ptr,
+                               const int16_t *round_ptr,
+                               const int16_t *quant_ptr,
+                               const int16_t *quant_shift_ptr,
+                               tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                               const int16_t *dequant_ptr, uint16_t *eob_ptr,
+                               const int16_t *scan, const int16_t *iscan) {
+  (void)scan;
+  quantize_b_no_qmatrix_avx2(coeff_ptr, n_coeffs, zbin_ptr, round_ptr,
+                             quant_ptr, quant_shift_ptr, qcoeff_ptr,
+                             dqcoeff_ptr, dequant_ptr, eob_ptr, iscan, 1);
+}
+
+void aom_quantize_b_64x64_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                               const int16_t *zbin_ptr,
+                               const int16_t *round_ptr,
+                               const int16_t *quant_ptr,
+                               const int16_t *quant_shift_ptr,
+                               tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                               const int16_t *dequant_ptr, uint16_t *eob_ptr,
+                               const int16_t *scan, const int16_t *iscan) {
+  (void)scan;
+  quantize_b_no_qmatrix_avx2(coeff_ptr, n_coeffs, zbin_ptr, round_ptr,
+                             quant_ptr, quant_shift_ptr, qcoeff_ptr,
+                             dqcoeff_ptr, dequant_ptr, eob_ptr, iscan, 2);
+}
diff --git a/third_party/aom/aom_dsp/x86/quantize_sse2.c b/third_party/aom/aom_dsp/x86/quantize_sse2.c
new file mode 100644
index 0000000000..ebef1fbac2
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/quantize_sse2.c
@@ -0,0 +1,125 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <emmintrin.h>
+#include <xmmintrin.h>
+
+#include "config/aom_dsp_rtcd.h"
+
+#include "aom/aom_integer.h"
+#include "aom_dsp/x86/quantize_x86.h"
+
+void aom_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                         const int16_t *zbin_ptr, const int16_t *round_ptr,
+                         const int16_t *quant_ptr,
+                         const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
+                         tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr,
+                         uint16_t *eob_ptr, const int16_t *scan_ptr,
+                         const int16_t *iscan_ptr) {
+  const __m128i zero = _mm_setzero_si128();
+  int index = 16;
+
+  __m128i zbin, round, quant, dequant, shift;
+  __m128i coeff0, coeff1, coeff0_sign, coeff1_sign;
+  __m128i qcoeff0, qcoeff1;
+  __m128i cmp_mask0, cmp_mask1;
+  __m128i eob, eob0;
+
+  (void)scan_ptr;
+
+  // Setup global values.
+  load_b_values(zbin_ptr, &zbin, round_ptr, &round, quant_ptr, &quant,
+                dequant_ptr, &dequant, quant_shift_ptr, &shift);
+
+  // Do DC and first 15 AC.
+  coeff0 = load_coefficients(coeff_ptr);
+  coeff1 = load_coefficients(coeff_ptr + 8);
+
+  // Poor man's abs().
+  coeff0_sign = _mm_srai_epi16(coeff0, 15);
+  coeff1_sign = _mm_srai_epi16(coeff1, 15);
+  qcoeff0 = invert_sign_sse2(coeff0, coeff0_sign);
+  qcoeff1 = invert_sign_sse2(coeff1, coeff1_sign);
+
+  cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin);
+  zbin = _mm_unpackhi_epi64(zbin, zbin);  // Switch DC to AC
+  cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);
+
+  calculate_qcoeff(&qcoeff0, round, quant, shift);
+
+  round = _mm_unpackhi_epi64(round, round);
+  quant = _mm_unpackhi_epi64(quant, quant);
+  shift = _mm_unpackhi_epi64(shift, shift);
+
+  calculate_qcoeff(&qcoeff1, round, quant, shift);
+
+  // Reinsert signs
+  qcoeff0 = invert_sign_sse2(qcoeff0, coeff0_sign);
+  qcoeff1 = invert_sign_sse2(qcoeff1, coeff1_sign);
+
+  // Mask out zbin threshold coeffs
+  qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
+  qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
+
+  store_coefficients(qcoeff0, qcoeff_ptr);
+  store_coefficients(qcoeff1, qcoeff_ptr + 8);
+
+  coeff0 = calculate_dqcoeff(qcoeff0, dequant);
+  dequant = _mm_unpackhi_epi64(dequant, dequant);
+  coeff1 = calculate_dqcoeff(qcoeff1, dequant);
+
+  store_coefficients(coeff0, dqcoeff_ptr);
+  store_coefficients(coeff1, dqcoeff_ptr + 8);
+
+  eob =
+      scan_for_eob(&coeff0, &coeff1, cmp_mask0, cmp_mask1, iscan_ptr, 0, zero);
+
+  // AC only loop.
+  while (index < n_coeffs) {
+    coeff0 = load_coefficients(coeff_ptr + index);
+    coeff1 = load_coefficients(coeff_ptr + index + 8);
+
+    coeff0_sign = _mm_srai_epi16(coeff0, 15);
+    coeff1_sign = _mm_srai_epi16(coeff1, 15);
+    qcoeff0 = invert_sign_sse2(coeff0, coeff0_sign);
+    qcoeff1 = invert_sign_sse2(coeff1, coeff1_sign);
+
+    cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin);
+    cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);
+
+    calculate_qcoeff(&qcoeff0, round, quant, shift);
+    calculate_qcoeff(&qcoeff1, round, quant, shift);
+
+    qcoeff0 = invert_sign_sse2(qcoeff0, coeff0_sign);
+    qcoeff1 = invert_sign_sse2(qcoeff1, coeff1_sign);
+
+    qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
+    qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
+
+    store_coefficients(qcoeff0, qcoeff_ptr + index);
+    store_coefficients(qcoeff1, qcoeff_ptr + index + 8);
+
+    coeff0 = calculate_dqcoeff(qcoeff0, dequant);
+    coeff1 = calculate_dqcoeff(qcoeff1, dequant);
+
+    store_coefficients(coeff0, dqcoeff_ptr + index);
+    store_coefficients(coeff1, dqcoeff_ptr + index + 8);
+
+    eob0 = scan_for_eob(&coeff0, &coeff1, cmp_mask0, cmp_mask1, iscan_ptr,
+                        index, zero);
+    eob = _mm_max_epi16(eob, eob0);
+
+    index += 16;
+  }
+
+  *eob_ptr = accumulate_eob(eob);
+}
diff --git a/third_party/aom/aom_dsp/x86/quantize_ssse3.c b/third_party/aom/aom_dsp/x86/quantize_ssse3.c
new file mode 100644
index 0000000000..25980a055a
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/quantize_ssse3.c
@@ -0,0 +1,192 @@
+/*
+ * Copyright (c) 2019, 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 <tmmintrin.h>
+#include <emmintrin.h>
+#include <xmmintrin.h>
+
+#include "config/aom_dsp_rtcd.h"
+
+#include "aom/aom_integer.h"
+#include "aom_dsp/x86/quantize_x86.h"
+
+static INLINE void calculate_qcoeff_64x64(__m128i *coeff, const __m128i round,
+                                          const __m128i quant,
+                                          const __m128i *shift) {
+  __m128i tmp, qcoeff, tmp1;
+  qcoeff = _mm_adds_epi16(*coeff, round);
+  tmp = _mm_mulhi_epi16(qcoeff, quant);
+  qcoeff = _mm_add_epi16(tmp, qcoeff);
+  tmp = _mm_mullo_epi16(qcoeff, *shift);
+  tmp = _mm_srli_epi16(tmp, 14);
+  tmp1 = _mm_mulhi_epi16(qcoeff, *shift);
+  tmp1 = _mm_slli_epi16(tmp1, 2);
+  *coeff = _mm_or_si128(tmp, tmp1);
+}
+
+static INLINE void calculate_dqcoeff_and_store_64x64(const __m128i qcoeff,
+                                                     const __m128i dequant,
+                                                     const __m128i zero,
+                                                     tran_low_t *dqcoeff) {
+  // Un-sign to bias rounding like C.
+  const __m128i coeff = _mm_abs_epi16(qcoeff);
+
+  const __m128i sign_0 = _mm_unpacklo_epi16(zero, qcoeff);
+  const __m128i sign_1 = _mm_unpackhi_epi16(zero, qcoeff);
+
+  const __m128i low = _mm_mullo_epi16(coeff, dequant);
+  const __m128i high = _mm_mulhi_epi16(coeff, dequant);
+  __m128i dqcoeff32_0 = _mm_unpacklo_epi16(low, high);
+  __m128i dqcoeff32_1 = _mm_unpackhi_epi16(low, high);
+
+  // "Divide" by 4.
+  dqcoeff32_0 = _mm_srli_epi32(dqcoeff32_0, 2);
+  dqcoeff32_1 = _mm_srli_epi32(dqcoeff32_1, 2);
+
+  dqcoeff32_0 = _mm_sign_epi32(dqcoeff32_0, sign_0);
+  dqcoeff32_1 = _mm_sign_epi32(dqcoeff32_1, sign_1);
+
+  _mm_store_si128((__m128i *)(dqcoeff), dqcoeff32_0);
+  _mm_store_si128((__m128i *)(dqcoeff + 4), dqcoeff32_1);
+}
+
+void aom_quantize_b_64x64_ssse3(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
+                                const int16_t *zbin_ptr,
+                                const int16_t *round_ptr,
+                                const int16_t *quant_ptr,
+                                const int16_t *quant_shift_ptr,
+                                tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
+                                const int16_t *dequant_ptr, uint16_t *eob_ptr,
+                                const int16_t *scan, const int16_t *iscan) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i one = _mm_set1_epi16(1);
+  const __m128i two = _mm_set1_epi16(2);
+  int index;
+
+  __m128i zbin, round, quant, dequant, shift;
+  __m128i coeff0, coeff1, qcoeff0, qcoeff1;
+  __m128i cmp_mask0, cmp_mask1, all_zero;
+  __m128i eob = zero, eob0;
+
+  (void)scan;
+  (void)n_coeffs;
+
+  // Setup global values.
+  zbin = _mm_load_si128((const __m128i *)zbin_ptr);
+  round = _mm_load_si128((const __m128i *)round_ptr);
+  quant = _mm_load_si128((const __m128i *)quant_ptr);
+  dequant = _mm_load_si128((const __m128i *)dequant_ptr);
+  shift = _mm_load_si128((const __m128i *)quant_shift_ptr);
+
+  // Shift with rounding.
+  zbin = _mm_add_epi16(zbin, two);
+  round = _mm_add_epi16(round, two);
+  zbin = _mm_srli_epi16(zbin, 2);
+  round = _mm_srli_epi16(round, 2);
+  zbin = _mm_sub_epi16(zbin, one);
+  // Do DC and first 15 AC.
+  coeff0 = load_coefficients(coeff_ptr);
+  coeff1 = load_coefficients(coeff_ptr + 8);
+
+  qcoeff0 = _mm_abs_epi16(coeff0);
+  qcoeff1 = _mm_abs_epi16(coeff1);
+
+  cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin);
+  zbin = _mm_unpackhi_epi64(zbin, zbin);
+  cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);
+  all_zero = _mm_or_si128(cmp_mask0, cmp_mask1);
+  if (_mm_movemask_epi8(all_zero) == 0) {
+    _mm_store_si128((__m128i *)(qcoeff_ptr), zero);
+    _mm_store_si128((__m128i *)(qcoeff_ptr + 4), zero);
+    _mm_store_si128((__m128i *)(qcoeff_ptr + 8), zero);
+    _mm_store_si128((__m128i *)(qcoeff_ptr + 12), zero);
+    _mm_store_si128((__m128i *)(dqcoeff_ptr), zero);
+    _mm_store_si128((__m128i *)(dqcoeff_ptr + 4), zero);
+    _mm_store_si128((__m128i *)(dqcoeff_ptr + 8), zero);
+    _mm_store_si128((__m128i *)(dqcoeff_ptr + 12), zero);
+    round = _mm_unpackhi_epi64(round, round);
+    quant = _mm_unpackhi_epi64(quant, quant);
+    shift = _mm_unpackhi_epi64(shift, shift);
+    dequant = _mm_unpackhi_epi64(dequant, dequant);
+  } else {
+    calculate_qcoeff_64x64(&qcoeff0, round, quant, &shift);
+    round = _mm_unpackhi_epi64(round, round);
+    quant = _mm_unpackhi_epi64(quant, quant);
+    shift = _mm_unpackhi_epi64(shift, shift);
+    calculate_qcoeff_64x64(&qcoeff1, round, quant, &shift);
+
+    // Reinsert signs.
+    qcoeff0 = _mm_sign_epi16(qcoeff0, coeff0);
+    qcoeff1 = _mm_sign_epi16(qcoeff1, coeff1);
+
+    // Mask out zbin threshold coeffs.
+    qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
+    qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
+
+    store_coefficients(qcoeff0, qcoeff_ptr);
+    store_coefficients(qcoeff1, qcoeff_ptr + 8);
+
+    calculate_dqcoeff_and_store_64x64(qcoeff0, dequant, zero, dqcoeff_ptr);
+    dequant = _mm_unpackhi_epi64(dequant, dequant);
+    calculate_dqcoeff_and_store_64x64(qcoeff1, dequant, zero, dqcoeff_ptr + 8);
+
+    eob =
+        scan_for_eob(&qcoeff0, &qcoeff1, cmp_mask0, cmp_mask1, iscan, 0, zero);
+  }
+
+  // AC only loop.
+  for (index = 16; index < 1024; index += 16) {
+    coeff0 = load_coefficients(coeff_ptr + index);
+    coeff1 = load_coefficients(coeff_ptr + index + 8);
+
+    qcoeff0 = _mm_abs_epi16(coeff0);
+    qcoeff1 = _mm_abs_epi16(coeff1);
+
+    cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin);
+    cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);
+
+    all_zero = _mm_or_si128(cmp_mask0, cmp_mask1);
+    if (_mm_movemask_epi8(all_zero) == 0) {
+      _mm_store_si128((__m128i *)(qcoeff_ptr + index), zero);
+      _mm_store_si128((__m128i *)(qcoeff_ptr + index + 4), zero);
+      _mm_store_si128((__m128i *)(qcoeff_ptr + index + 8), zero);
+      _mm_store_si128((__m128i *)(qcoeff_ptr + index + 12), zero);
+      _mm_store_si128((__m128i *)(dqcoeff_ptr + index), zero);
+      _mm_store_si128((__m128i *)(dqcoeff_ptr + index + 4), zero);
+      _mm_store_si128((__m128i *)(dqcoeff_ptr + index + 8), zero);
+      _mm_store_si128((__m128i *)(dqcoeff_ptr + index + 12), zero);
+      continue;
+    }
+    calculate_qcoeff_64x64(&qcoeff0, round, quant, &shift);
+    calculate_qcoeff_64x64(&qcoeff1, round, quant, &shift);
+
+    qcoeff0 = _mm_sign_epi16(qcoeff0, coeff0);
+    qcoeff1 = _mm_sign_epi16(qcoeff1, coeff1);
+
+    qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
+    qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
+
+    store_coefficients(qcoeff0, qcoeff_ptr + index);
+    store_coefficients(qcoeff1, qcoeff_ptr + index + 8);
+
+    calculate_dqcoeff_and_store_64x64(qcoeff0, dequant, zero,
+                                      dqcoeff_ptr + index);
+    calculate_dqcoeff_and_store_64x64(qcoeff1, dequant, zero,
+                                      dqcoeff_ptr + 8 + index);
+
+    eob0 = scan_for_eob(&qcoeff0, &qcoeff1, cmp_mask0, cmp_mask1, iscan, index,
+                        zero);
+    eob = _mm_max_epi16(eob, eob0);
+  }
+
+  *eob_ptr = accumulate_eob(eob);
+}
diff --git a/third_party/aom/aom_dsp/x86/quantize_ssse3_x86_64.asm b/third_party/aom/aom_dsp/x86/quantize_ssse3_x86_64.asm
new file mode 100644
index 0000000000..fa616a6f1a
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/quantize_ssse3_x86_64.asm
@@ -0,0 +1,302 @@
+;
+; Copyright (c) 2016, 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 "third_party/x86inc/x86inc.asm"
+
+SECTION_RODATA
+pw_1: times 8 dw 1
+
+SECTION .text
+
+%macro QUANTIZE_FN 2
+cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, zbin, round, quant, \
+                                shift, qcoeff, dqcoeff, dequant, \
+                                eob, scan, iscan
+
+  ; actual quantize loop - setup pointers, rounders, etc.
+  movifnidn                   coeffq, coeffmp
+  movifnidn                  ncoeffq, ncoeffmp
+  movifnidn                    zbinq, zbinmp
+  movifnidn                   roundq, roundmp
+  movifnidn                   quantq, quantmp
+  movifnidn                 dequantq, dequantmp
+  mova                            m0, [zbinq]              ; m0 = zbin
+  mova                            m1, [roundq]             ; m1 = round
+  mova                            m2, [quantq]             ; m2 = quant
+%ifidn %1, b_32x32
+  pcmpeqw                         m5, m5
+  psrlw                           m5, 15
+  paddw                           m0, m5
+  paddw                           m1, m5
+  psrlw                           m0, 1                    ; m0 = (m0 + 1) / 2
+  psrlw                           m1, 1                    ; m1 = (m1 + 1) / 2
+%endif
+  mova                            m3, [dequantq]           ; m3 = dequant
+  mov                             r2, shiftmp
+  psubw                           m0, [GLOBAL(pw_1)]
+  mova                            m4, [r2]                 ; m4 = shift
+  mov                             r3, qcoeffmp
+  mov                             r4, dqcoeffmp
+  mov                             r5, iscanmp
+  pxor                            m5, m5                   ; m5 = dedicated zero
+  DEFINE_ARGS coeff, ncoeff, d1, qcoeff, dqcoeff, iscan, d2, d3, d4, eob
+  lea                         coeffq, [  coeffq+ncoeffq*4]
+  lea                        qcoeffq, [ qcoeffq+ncoeffq*4]
+  lea                       dqcoeffq, [dqcoeffq+ncoeffq*4]
+  lea                         iscanq, [  iscanq+ncoeffq*2]
+  neg                        ncoeffq
+
+  ; get DC and first 15 AC coeffs
+  ; coeff stored as 32bit numbers & require 16bit numbers
+  mova                            m9, [  coeffq+ncoeffq*4+ 0]
+  packssdw                        m9, [  coeffq+ncoeffq*4+16]
+  mova                           m10, [  coeffq+ncoeffq*4+32]
+  packssdw                       m10, [  coeffq+ncoeffq*4+48]
+  pabsw                           m6, m9                   ; m6 = abs(m9)
+  pabsw                          m11, m10                  ; m11 = abs(m10)
+  pcmpgtw                         m7, m6, m0               ; m7 = c[i] >= zbin
+  punpckhqdq                      m0, m0
+  pcmpgtw                        m12, m11, m0              ; m12 = c[i] >= zbin
+  paddsw                          m6, m1                   ; m6 += round
+  punpckhqdq                      m1, m1
+  paddsw                         m11, m1                   ; m11 += round
+  pmulhw                          m8, m6, m2               ; m8 = m6*q>>16
+  punpckhqdq                      m2, m2
+  pmulhw                         m13, m11, m2              ; m13 = m11*q>>16
+  paddw                           m8, m6                   ; m8 += m6
+  paddw                          m13, m11                  ; m13 += m11
+  %ifidn %1, b_32x32
+  pmullw                          m5, m8, m4               ; store the lower 16 bits of m8*qsh
+  %endif
+  pmulhw                          m8, m4                   ; m8 = m8*qsh>>16
+  %ifidn %1, b_32x32
+  psllw                           m8, 1
+  psrlw                           m5, 15
+  por                             m8, m5
+  %endif
+  punpckhqdq                      m4, m4
+  %ifidn %1, b_32x32
+  pmullw                          m5, m13, m4              ; store the lower 16 bits of m13*qsh
+  %endif
+  pmulhw                         m13, m4                   ; m13 = m13*qsh>>16
+  %ifidn %1, b_32x32
+  psllw                          m13, 1
+  psrlw                           m5, 15
+  por                            m13, m5
+  pxor                            m5, m5                   ; reset m5 to zero register
+  %endif
+  psignw                          m8, m9                   ; m8 = reinsert sign
+  psignw                         m13, m10                  ; m13 = reinsert sign
+  pand                            m8, m7
+  pand                           m13, m12
+
+  ; store 16bit numbers as 32bit numbers in array pointed to by qcoeff
+  mova                           m11, m8
+  mova                            m6, m8
+  pcmpgtw                         m5, m8
+  punpcklwd                      m11, m5
+  punpckhwd                       m6, m5
+  mova        [qcoeffq+ncoeffq*4+ 0], m11
+  mova        [qcoeffq+ncoeffq*4+16], m6
+  pxor                            m5, m5
+  mova                           m11, m13
+  mova                            m6, m13
+  pcmpgtw                         m5, m13
+  punpcklwd                      m11, m5
+  punpckhwd                       m6, m5
+  mova        [qcoeffq+ncoeffq*4+32], m11
+  mova        [qcoeffq+ncoeffq*4+48], m6
+  pxor                            m5, m5             ; reset m5 to zero register
+
+%ifidn %1, b_32x32
+  pabsw                           m8, m8
+  pabsw                          m13, m13
+%endif
+  pmullw                          m8, m3                   ; dqc[i] = qc[i] * q
+  punpckhqdq                      m3, m3
+  pmullw                         m13, m3                   ; dqc[i] = qc[i] * q
+%ifidn %1, b_32x32
+  psrlw                           m8, 1
+  psrlw                          m13, 1
+  psignw                          m8, m9
+  psignw                         m13, m10
+%endif
+  ; store 16bit numbers as 32bit numbers in array pointed to by dqcoeff
+  mova                            m11, m8
+  mova                            m6, m8
+  pcmpgtw                         m5, m8
+  punpcklwd                      m11, m5
+  punpckhwd                       m6, m5
+  mova       [dqcoeffq+ncoeffq*4+ 0], m11
+  mova       [dqcoeffq+ncoeffq*4+16], m6
+  pxor                            m5, m5
+  mova                           m11, m13
+  mova                            m6, m13
+  pcmpgtw                         m5, m13
+  punpcklwd                      m11, m5
+  punpckhwd                       m6, m5
+  mova       [dqcoeffq+ncoeffq*4+32], m11
+  mova       [dqcoeffq+ncoeffq*4+48], m6
+  pxor                            m5, m5             ; reset m5 to zero register
+  pcmpeqw                         m8, m5                   ; m8 = c[i] == 0
+  pcmpeqw                        m13, m5                   ; m13 = c[i] == 0
+  mova                            m6, [  iscanq+ncoeffq*2+ 0] ; m6 = scan[i]
+  mova                           m11, [  iscanq+ncoeffq*2+16] ; m11 = scan[i]
+  psubw                           m6, m7                   ; m6 = scan[i] + 1
+  psubw                          m11, m12                  ; m11 = scan[i] + 1
+  pandn                           m8, m6                   ; m8 = max(eob)
+  pandn                          m13, m11                  ; m13 = max(eob)
+  pmaxsw                          m8, m13
+  add                        ncoeffq, mmsize
+  jz .accumulate_eob
+
+.ac_only_loop:
+  ; pack coeff from 32bit to 16bit array
+  mova                            m9, [  coeffq+ncoeffq*4+ 0]
+  packssdw                        m9, [  coeffq+ncoeffq*4+16]
+  mova                           m10, [  coeffq+ncoeffq*4+32]
+  packssdw                       m10, [  coeffq+ncoeffq*4+48]
+
+  pabsw                           m6, m9                   ; m6 = abs(m9)
+  pabsw                          m11, m10                  ; m11 = abs(m10)
+  pcmpgtw                         m7, m6, m0               ; m7 = c[i] >= zbin
+  pcmpgtw                        m12, m11, m0              ; m12 = c[i] >= zbin
+%ifidn %1, b_32x32
+  pmovmskb                       r6d, m7
+  pmovmskb                       r2d, m12
+  or                              r6, r2
+  jz .skip_iter
+%endif
+  paddsw                          m6, m1                   ; m6 += round
+  paddsw                         m11, m1                   ; m11 += round
+  pmulhw                         m14, m6, m2               ; m14 = m6*q>>16
+  pmulhw                         m13, m11, m2              ; m13 = m11*q>>16
+  paddw                          m14, m6                   ; m14 += m6
+  paddw                          m13, m11                  ; m13 += m11
+  %ifidn %1, b_32x32
+  pmullw                          m5, m14, m4              ; store the lower 16 bits of m14*qsh
+  %endif
+  pmulhw                         m14, m4                   ; m14 = m14*qsh>>16
+  %ifidn %1, b_32x32
+  psllw                          m14, 1
+  psrlw                           m5, 15
+  por                            m14, m5
+  pmullw                          m5, m13, m4              ; store the lower 16 bits of m13*qsh
+  %endif
+  pmulhw                         m13, m4                   ; m13 = m13*qsh>>16
+  %ifidn %1, b_32x32
+  psllw                          m13, 1
+  psrlw                           m5, 15
+  por                            m13, m5
+  pxor                            m5, m5                   ; reset m5 to zero register
+  %endif
+  psignw                         m14, m9                   ; m14 = reinsert sign
+  psignw                         m13, m10                  ; m13 = reinsert sign
+  pand                           m14, m7
+  pand                           m13, m12
+
+  ; store 16bit numbers as 32bit numbers in array pointed to by qcoeff
+  pxor                           m11, m11
+  mova                           m11, m14
+  mova                            m6, m14
+  pcmpgtw                         m5, m14
+  punpcklwd                      m11, m5
+  punpckhwd                       m6, m5
+  mova        [qcoeffq+ncoeffq*4+ 0], m11
+  mova        [qcoeffq+ncoeffq*4+16], m6
+  pxor                            m5, m5
+  mova                           m11, m13
+  mova                            m6, m13
+  pcmpgtw                         m5, m13
+  punpcklwd                      m11, m5
+  punpckhwd                       m6, m5
+  mova        [qcoeffq+ncoeffq*4+32], m11
+  mova        [qcoeffq+ncoeffq*4+48], m6
+  pxor                            m5, m5             ; reset m5 to zero register
+
+%ifidn %1, b_32x32
+  pabsw                          m14, m14
+  pabsw                          m13, m13
+%endif
+  pmullw                         m14, m3                   ; dqc[i] = qc[i] * q
+  pmullw                         m13, m3                   ; dqc[i] = qc[i] * q
+%ifidn %1, b_32x32
+  psrlw                          m14, 1
+  psrlw                          m13, 1
+  psignw                         m14, m9
+  psignw                         m13, m10
+%endif
+
+  ; store 16bit numbers as 32bit numbers in array pointed to by dqcoeff
+  mova                           m11, m14
+  mova                            m6, m14
+  pcmpgtw                         m5, m14
+  punpcklwd                      m11, m5
+  punpckhwd                       m6, m5
+  mova       [dqcoeffq+ncoeffq*4+ 0], m11
+  mova       [dqcoeffq+ncoeffq*4+16], m6
+  pxor                            m5, m5
+  mova                           m11, m13
+  mova                            m6, m13
+  pcmpgtw                         m5, m13
+  punpcklwd                      m11, m5
+  punpckhwd                       m6, m5
+  mova       [dqcoeffq+ncoeffq*4+32], m11
+  mova       [dqcoeffq+ncoeffq*4+48], m6
+  pxor                            m5, m5
+
+  pcmpeqw                        m14, m5                   ; m14 = c[i] == 0
+  pcmpeqw                        m13, m5                   ; m13 = c[i] == 0
+  mova                            m6, [  iscanq+ncoeffq*2+ 0] ; m6 = scan[i]
+  mova                           m11, [  iscanq+ncoeffq*2+16] ; m11 = scan[i]
+  psubw                           m6, m7                   ; m6 = scan[i] + 1
+  psubw                          m11, m12                  ; m11 = scan[i] + 1
+  pandn                          m14, m6                   ; m14 = max(eob)
+  pandn                          m13, m11                  ; m13 = max(eob)
+  pmaxsw                          m8, m14
+  pmaxsw                          m8, m13
+  add                        ncoeffq, mmsize
+  jl .ac_only_loop
+
+%ifidn %1, b_32x32
+  jmp .accumulate_eob
+.skip_iter:
+  mova        [qcoeffq+ncoeffq*4+ 0], m5
+  mova        [qcoeffq+ncoeffq*4+16], m5
+  mova        [qcoeffq+ncoeffq*4+32], m5
+  mova        [qcoeffq+ncoeffq*4+48], m5
+  mova       [dqcoeffq+ncoeffq*4+ 0], m5
+  mova       [dqcoeffq+ncoeffq*4+16], m5
+  mova       [dqcoeffq+ncoeffq*4+32], m5
+  mova       [dqcoeffq+ncoeffq*4+48], m5
+  add                        ncoeffq, mmsize
+  jl .ac_only_loop
+%endif
+
+.accumulate_eob:
+  ; horizontally accumulate/max eobs and write into [eob] memory pointer
+  mov                             r2, eobmp
+  pshufd                          m7, m8, 0xe
+  pmaxsw                          m8, m7
+  pshuflw                         m7, m8, 0xe
+  pmaxsw                          m8, m7
+  pshuflw                         m7, m8, 0x1
+  pmaxsw                          m8, m7
+  pextrw                          r6, m8, 0
+  mov                             [r2], r6
+  RET
+%endmacro
+
+INIT_XMM ssse3
+QUANTIZE_FN b, 9
+QUANTIZE_FN b_32x32, 9
diff --git a/third_party/aom/aom_dsp/x86/quantize_x86.h b/third_party/aom/aom_dsp/x86/quantize_x86.h
new file mode 100644
index 0000000000..5b040a278a
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/quantize_x86.h
@@ -0,0 +1,202 @@
+/*
+ *  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 <emmintrin.h>
+
+#include "aom/aom_integer.h"
+
+static INLINE void load_b_values(const int16_t *zbin_ptr, __m128i *zbin,
+                                 const int16_t *round_ptr, __m128i *round,
+                                 const int16_t *quant_ptr, __m128i *quant,
+                                 const int16_t *dequant_ptr, __m128i *dequant,
+                                 const int16_t *shift_ptr, __m128i *shift) {
+  *zbin = _mm_load_si128((const __m128i *)zbin_ptr);
+  *round = _mm_load_si128((const __m128i *)round_ptr);
+  *quant = _mm_load_si128((const __m128i *)quant_ptr);
+  *zbin = _mm_sub_epi16(*zbin, _mm_set1_epi16(1));
+  *dequant = _mm_load_si128((const __m128i *)dequant_ptr);
+  *shift = _mm_load_si128((const __m128i *)shift_ptr);
+}
+
+// With ssse3 and later abs() and sign() are preferred.
+static INLINE __m128i invert_sign_sse2(__m128i a, __m128i sign) {
+  a = _mm_xor_si128(a, sign);
+  return _mm_sub_epi16(a, sign);
+}
+
+static INLINE __m128i invert_sign_32_sse2(__m128i a, __m128i sign) {
+  a = _mm_xor_si128(a, sign);
+  return _mm_sub_epi32(a, sign);
+}
+
+static INLINE void calculate_qcoeff(__m128i *coeff, const __m128i round,
+                                    const __m128i quant, const __m128i shift) {
+  __m128i tmp, qcoeff;
+  qcoeff = _mm_adds_epi16(*coeff, round);
+  tmp = _mm_mulhi_epi16(qcoeff, quant);
+  qcoeff = _mm_add_epi16(tmp, qcoeff);
+  *coeff = _mm_mulhi_epi16(qcoeff, shift);
+}
+
+static INLINE void calculate_qcoeff_log_scale(__m128i *coeff,
+                                              const __m128i round,
+                                              const __m128i quant,
+                                              const __m128i *shift,
+                                              const int *log_scale) {
+  __m128i tmp, tmp1, qcoeff;
+  qcoeff = _mm_adds_epi16(*coeff, round);
+  tmp = _mm_mulhi_epi16(qcoeff, quant);
+  qcoeff = _mm_add_epi16(tmp, qcoeff);
+  tmp = _mm_mullo_epi16(qcoeff, *shift);
+  tmp = _mm_srli_epi16(tmp, (16 - *log_scale));
+  tmp1 = _mm_mulhi_epi16(qcoeff, *shift);
+  tmp1 = _mm_slli_epi16(tmp1, *log_scale);
+  *coeff = _mm_or_si128(tmp, tmp1);
+}
+
+static INLINE __m128i calculate_dqcoeff(__m128i qcoeff, __m128i dequant) {
+  return _mm_mullo_epi16(qcoeff, dequant);
+}
+
+static INLINE void calculate_dqcoeff_and_store_log_scale(__m128i qcoeff,
+                                                         __m128i dequant,
+                                                         const __m128i zero,
+                                                         tran_low_t *dqcoeff,
+                                                         const int *log_scale) {
+  // calculate abs
+  __m128i coeff_sign = _mm_srai_epi16(qcoeff, 15);
+  __m128i coeff = invert_sign_sse2(qcoeff, coeff_sign);
+
+  const __m128i sign_0 = _mm_unpacklo_epi16(coeff_sign, zero);
+  const __m128i sign_1 = _mm_unpackhi_epi16(coeff_sign, zero);
+
+  const __m128i low = _mm_mullo_epi16(coeff, dequant);
+  const __m128i high = _mm_mulhi_epi16(coeff, dequant);
+  __m128i dqcoeff32_0 = _mm_unpacklo_epi16(low, high);
+  __m128i dqcoeff32_1 = _mm_unpackhi_epi16(low, high);
+
+  dqcoeff32_0 = _mm_srli_epi32(dqcoeff32_0, *log_scale);
+  dqcoeff32_1 = _mm_srli_epi32(dqcoeff32_1, *log_scale);
+
+  dqcoeff32_0 = invert_sign_32_sse2(dqcoeff32_0, sign_0);
+  dqcoeff32_1 = invert_sign_32_sse2(dqcoeff32_1, sign_1);
+
+  _mm_store_si128((__m128i *)(dqcoeff), dqcoeff32_0);
+  _mm_store_si128((__m128i *)(dqcoeff + 4), dqcoeff32_1);
+}
+
+// Scan 16 values for eob reference in scan_ptr. Use masks (-1) from comparing
+// to zbin to add 1 to the index in 'scan'.
+static INLINE __m128i scan_for_eob(__m128i *coeff0, __m128i *coeff1,
+                                   const __m128i zbin_mask0,
+                                   const __m128i zbin_mask1,
+                                   const int16_t *scan_ptr, const int index,
+                                   const __m128i zero) {
+  const __m128i zero_coeff0 = _mm_cmpeq_epi16(*coeff0, zero);
+  const __m128i zero_coeff1 = _mm_cmpeq_epi16(*coeff1, zero);
+  __m128i scan0 = _mm_load_si128((const __m128i *)(scan_ptr + index));
+  __m128i scan1 = _mm_load_si128((const __m128i *)(scan_ptr + index + 8));
+  __m128i eob0, eob1;
+  // Add one to convert from indices to counts
+  scan0 = _mm_sub_epi16(scan0, zbin_mask0);
+  scan1 = _mm_sub_epi16(scan1, zbin_mask1);
+  eob0 = _mm_andnot_si128(zero_coeff0, scan0);
+  eob1 = _mm_andnot_si128(zero_coeff1, scan1);
+  return _mm_max_epi16(eob0, eob1);
+}
+
+static INLINE int16_t accumulate_eob(__m128i eob) {
+  __m128i eob_shuffled;
+  eob_shuffled = _mm_shuffle_epi32(eob, 0xe);
+  eob = _mm_max_epi16(eob, eob_shuffled);
+  eob_shuffled = _mm_shufflelo_epi16(eob, 0xe);
+  eob = _mm_max_epi16(eob, eob_shuffled);
+  eob_shuffled = _mm_shufflelo_epi16(eob, 0x1);
+  eob = _mm_max_epi16(eob, eob_shuffled);
+  return _mm_extract_epi16(eob, 1);
+}
+
+static INLINE __m128i load_coefficients(const tran_low_t *coeff_ptr) {
+  assert(sizeof(tran_low_t) == 4);
+  const __m128i coeff1 = _mm_load_si128((__m128i *)(coeff_ptr));
+  const __m128i coeff2 = _mm_load_si128((__m128i *)(coeff_ptr + 4));
+  return _mm_packs_epi32(coeff1, coeff2);
+}
+
+static INLINE void store_coefficients(__m128i coeff_vals,
+                                      tran_low_t *coeff_ptr) {
+  assert(sizeof(tran_low_t) == 4);
+
+  __m128i one = _mm_set1_epi16(1);
+  __m128i coeff_vals_hi = _mm_mulhi_epi16(coeff_vals, one);
+  __m128i coeff_vals_lo = _mm_mullo_epi16(coeff_vals, one);
+  __m128i coeff_vals_1 = _mm_unpacklo_epi16(coeff_vals_lo, coeff_vals_hi);
+  __m128i coeff_vals_2 = _mm_unpackhi_epi16(coeff_vals_lo, coeff_vals_hi);
+  _mm_store_si128((__m128i *)(coeff_ptr), coeff_vals_1);
+  _mm_store_si128((__m128i *)(coeff_ptr + 4), coeff_vals_2);
+}
+
+static INLINE void update_mask1(__m128i *cmp_mask0, __m128i *cmp_mask1,
+                                const int16_t *iscan_ptr, int *is_found,
+                                __m128i *mask) {
+  __m128i all_zero;
+  __m128i temp_mask = _mm_setzero_si128();
+  all_zero = _mm_or_si128(*cmp_mask0, *cmp_mask1);
+  if (_mm_movemask_epi8(all_zero)) {
+    __m128i iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr));
+    __m128i mask0 = _mm_and_si128(*cmp_mask0, iscan0);
+    __m128i iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + 8));
+    __m128i mask1 = _mm_and_si128(*cmp_mask1, iscan1);
+    temp_mask = _mm_max_epi16(mask0, mask1);
+    *is_found = 1;
+  }
+  *mask = _mm_max_epi16(temp_mask, *mask);
+}
+
+static INLINE void update_mask0(__m128i *qcoeff0, __m128i *qcoeff1,
+                                __m128i *threshold, const int16_t *iscan_ptr,
+                                int *is_found, __m128i *mask) {
+  __m128i zero = _mm_setzero_si128();
+  __m128i coeff[4], cmp_mask0, cmp_mask1, cmp_mask2, cmp_mask3;
+
+  coeff[0] = _mm_unpacklo_epi16(*qcoeff0, zero);
+  coeff[1] = _mm_unpackhi_epi16(*qcoeff0, zero);
+  coeff[2] = _mm_unpacklo_epi16(*qcoeff1, zero);
+  coeff[3] = _mm_unpackhi_epi16(*qcoeff1, zero);
+
+  coeff[0] = _mm_slli_epi32(coeff[0], AOM_QM_BITS);
+  cmp_mask0 = _mm_cmpgt_epi32(coeff[0], threshold[0]);
+  coeff[1] = _mm_slli_epi32(coeff[1], AOM_QM_BITS);
+  cmp_mask1 = _mm_cmpgt_epi32(coeff[1], threshold[1]);
+  coeff[2] = _mm_slli_epi32(coeff[2], AOM_QM_BITS);
+  cmp_mask2 = _mm_cmpgt_epi32(coeff[2], threshold[1]);
+  coeff[3] = _mm_slli_epi32(coeff[3], AOM_QM_BITS);
+  cmp_mask3 = _mm_cmpgt_epi32(coeff[3], threshold[1]);
+
+  cmp_mask0 = _mm_packs_epi32(cmp_mask0, cmp_mask1);
+  cmp_mask1 = _mm_packs_epi32(cmp_mask2, cmp_mask3);
+
+  update_mask1(&cmp_mask0, &cmp_mask1, iscan_ptr, is_found, mask);
+}
+
+static INLINE int calculate_non_zero_count(__m128i mask) {
+  __m128i mask0, mask1;
+  int non_zero_count = 0;
+  mask0 = _mm_unpackhi_epi64(mask, mask);
+  mask1 = _mm_max_epi16(mask0, mask);
+  mask0 = _mm_shuffle_epi32(mask1, 1);
+  mask0 = _mm_max_epi16(mask0, mask1);
+  mask1 = _mm_srli_epi32(mask0, 16);
+  mask0 = _mm_max_epi16(mask0, mask1);
+  non_zero_count = _mm_extract_epi16(mask0, 0) + 1;
+
+  return non_zero_count;
+}
diff --git a/third_party/aom/aom_dsp/x86/sad4d_avx2.c b/third_party/aom/aom_dsp/x86/sad4d_avx2.c
new file mode 100644
index 0000000000..0fea6ddfd3
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/sad4d_avx2.c
@@ -0,0 +1,326 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+#include <immintrin.h>  // AVX2
+
+#include "config/aom_dsp_rtcd.h"
+
+#include "aom/aom_integer.h"
+#include "aom_dsp/x86/synonyms_avx2.h"
+
+static AOM_FORCE_INLINE void aggregate_and_store_sum(uint32_t res[4],
+                                                     const __m256i *sum_ref0,
+                                                     const __m256i *sum_ref1,
+                                                     const __m256i *sum_ref2,
+                                                     const __m256i *sum_ref3) {
+  // In sum_ref-i the result is saved in the first 4 bytes and the other 4
+  // bytes are zeroed.
+  // merge sum_ref0 and sum_ref1 also sum_ref2 and sum_ref3
+  // 0, 0, 1, 1
+  __m256i sum_ref01 = _mm256_castps_si256(_mm256_shuffle_ps(
+      _mm256_castsi256_ps(*sum_ref0), _mm256_castsi256_ps(*sum_ref1),
+      _MM_SHUFFLE(2, 0, 2, 0)));
+  // 2, 2, 3, 3
+  __m256i sum_ref23 = _mm256_castps_si256(_mm256_shuffle_ps(
+      _mm256_castsi256_ps(*sum_ref2), _mm256_castsi256_ps(*sum_ref3),
+      _MM_SHUFFLE(2, 0, 2, 0)));
+
+  // sum adjacent 32 bit integers
+  __m256i sum_ref0123 = _mm256_hadd_epi32(sum_ref01, sum_ref23);
+
+  // add the low 128 bit to the high 128 bit
+  __m128i sum = _mm_add_epi32(_mm256_castsi256_si128(sum_ref0123),
+                              _mm256_extractf128_si256(sum_ref0123, 1));
+
+  _mm_storeu_si128((__m128i *)(res), sum);
+}
+
+static AOM_FORCE_INLINE void aom_sadMxNx4d_avx2(
+    int M, int N, const uint8_t *src, int src_stride,
+    const uint8_t *const ref[4], int ref_stride, uint32_t res[4]) {
+  __m256i src_reg, ref0_reg, ref1_reg, ref2_reg, ref3_reg;
+  __m256i sum_ref0, sum_ref1, sum_ref2, sum_ref3;
+  int i, j;
+  const uint8_t *ref0, *ref1, *ref2, *ref3;
+
+  ref0 = ref[0];
+  ref1 = ref[1];
+  ref2 = ref[2];
+  ref3 = ref[3];
+  sum_ref0 = _mm256_setzero_si256();
+  sum_ref2 = _mm256_setzero_si256();
+  sum_ref1 = _mm256_setzero_si256();
+  sum_ref3 = _mm256_setzero_si256();
+
+  for (i = 0; i < N; i++) {
+    for (j = 0; j < M; j += 32) {
+      // load src and all refs
+      src_reg = _mm256_loadu_si256((const __m256i *)(src + j));
+      ref0_reg = _mm256_loadu_si256((const __m256i *)(ref0 + j));
+      ref1_reg = _mm256_loadu_si256((const __m256i *)(ref1 + j));
+      ref2_reg = _mm256_loadu_si256((const __m256i *)(ref2 + j));
+      ref3_reg = _mm256_loadu_si256((const __m256i *)(ref3 + j));
+
+      // sum of the absolute differences between every ref-i to src
+      ref0_reg = _mm256_sad_epu8(ref0_reg, src_reg);
+      ref1_reg = _mm256_sad_epu8(ref1_reg, src_reg);
+      ref2_reg = _mm256_sad_epu8(ref2_reg, src_reg);
+      ref3_reg = _mm256_sad_epu8(ref3_reg, src_reg);
+      // sum every ref-i
+      sum_ref0 = _mm256_add_epi32(sum_ref0, ref0_reg);
+      sum_ref1 = _mm256_add_epi32(sum_ref1, ref1_reg);
+      sum_ref2 = _mm256_add_epi32(sum_ref2, ref2_reg);
+      sum_ref3 = _mm256_add_epi32(sum_ref3, ref3_reg);
+    }
+    src += src_stride;
+    ref0 += ref_stride;
+    ref1 += ref_stride;
+    ref2 += ref_stride;
+    ref3 += ref_stride;
+  }
+
+  aggregate_and_store_sum(res, &sum_ref0, &sum_ref1, &sum_ref2, &sum_ref3);
+}
+
+static AOM_FORCE_INLINE void aom_sadMxNx3d_avx2(
+    int M, int N, const uint8_t *src, int src_stride,
+    const uint8_t *const ref[4], int ref_stride, uint32_t res[4]) {
+  __m256i src_reg, ref0_reg, ref1_reg, ref2_reg;
+  __m256i sum_ref0, sum_ref1, sum_ref2;
+  int i, j;
+  const uint8_t *ref0, *ref1, *ref2;
+  const __m256i zero = _mm256_setzero_si256();
+
+  ref0 = ref[0];
+  ref1 = ref[1];
+  ref2 = ref[2];
+  sum_ref0 = _mm256_setzero_si256();
+  sum_ref2 = _mm256_setzero_si256();
+  sum_ref1 = _mm256_setzero_si256();
+
+  for (i = 0; i < N; i++) {
+    for (j = 0; j < M; j += 32) {
+      // load src and all refs
+      src_reg = _mm256_loadu_si256((const __m256i *)(src + j));
+      ref0_reg = _mm256_loadu_si256((const __m256i *)(ref0 + j));
+      ref1_reg = _mm256_loadu_si256((const __m256i *)(ref1 + j));
+      ref2_reg = _mm256_loadu_si256((const __m256i *)(ref2 + j));
+
+      // sum of the absolute differences between every ref-i to src
+      ref0_reg = _mm256_sad_epu8(ref0_reg, src_reg);
+      ref1_reg = _mm256_sad_epu8(ref1_reg, src_reg);
+      ref2_reg = _mm256_sad_epu8(ref2_reg, src_reg);
+      // sum every ref-i
+      sum_ref0 = _mm256_add_epi32(sum_ref0, ref0_reg);
+      sum_ref1 = _mm256_add_epi32(sum_ref1, ref1_reg);
+      sum_ref2 = _mm256_add_epi32(sum_ref2, ref2_reg);
+    }
+    src += src_stride;
+    ref0 += ref_stride;
+    ref1 += ref_stride;
+    ref2 += ref_stride;
+  }
+  aggregate_and_store_sum(res, &sum_ref0, &sum_ref1, &sum_ref2, &zero);
+}
+
+#define SADMXN_AVX2(m, n)                                                      \
+  void aom_sad##m##x##n##x4d_avx2(const uint8_t *src, int src_stride,          \
+                                  const uint8_t *const ref[4], int ref_stride, \
+                                  uint32_t res[4]) {                           \
+    aom_sadMxNx4d_avx2(m, n, src, src_stride, ref, ref_stride, res);           \
+  }                                                                            \
+  void aom_sad##m##x##n##x3d_avx2(const uint8_t *src, int src_stride,          \
+                                  const uint8_t *const ref[4], int ref_stride, \
+                                  uint32_t res[4]) {                           \
+    aom_sadMxNx3d_avx2(m, n, src, src_stride, ref, ref_stride, res);           \
+  }
+
+SADMXN_AVX2(32, 8)
+SADMXN_AVX2(32, 16)
+SADMXN_AVX2(32, 32)
+SADMXN_AVX2(32, 64)
+
+SADMXN_AVX2(64, 16)
+SADMXN_AVX2(64, 32)
+SADMXN_AVX2(64, 64)
+SADMXN_AVX2(64, 128)
+
+SADMXN_AVX2(128, 64)
+SADMXN_AVX2(128, 128)
+
+#define SAD_SKIP_MXN_AVX2(m, n)                                             \
+  void aom_sad_skip_##m##x##n##x4d_avx2(const uint8_t *src, int src_stride, \
+                                        const uint8_t *const ref[4],        \
+                                        int ref_stride, uint32_t res[4]) {  \
+    aom_sadMxNx4d_avx2(m, ((n) >> 1), src, 2 * src_stride, ref,             \
+                       2 * ref_stride, res);                                \
+    res[0] <<= 1;                                                           \
+    res[1] <<= 1;                                                           \
+    res[2] <<= 1;                                                           \
+    res[3] <<= 1;                                                           \
+  }
+
+SAD_SKIP_MXN_AVX2(32, 8)
+SAD_SKIP_MXN_AVX2(32, 16)
+SAD_SKIP_MXN_AVX2(32, 32)
+SAD_SKIP_MXN_AVX2(32, 64)
+
+SAD_SKIP_MXN_AVX2(64, 16)
+SAD_SKIP_MXN_AVX2(64, 32)
+SAD_SKIP_MXN_AVX2(64, 64)
+SAD_SKIP_MXN_AVX2(64, 128)
+
+SAD_SKIP_MXN_AVX2(128, 64)
+SAD_SKIP_MXN_AVX2(128, 128)
+
+static AOM_FORCE_INLINE void aom_sad16xNx3d_avx2(int N, const uint8_t *src,
+                                                 int src_stride,
+                                                 const uint8_t *const ref[4],
+                                                 int ref_stride,
+                                                 uint32_t res[4]) {
+  __m256i src_reg, ref0_reg, ref1_reg, ref2_reg;
+  __m256i sum_ref0, sum_ref1, sum_ref2;
+  const uint8_t *ref0, *ref1, *ref2;
+  const __m256i zero = _mm256_setzero_si256();
+  assert(N % 2 == 0);
+
+  ref0 = ref[0];
+  ref1 = ref[1];
+  ref2 = ref[2];
+  sum_ref0 = _mm256_setzero_si256();
+  sum_ref2 = _mm256_setzero_si256();
+  sum_ref1 = _mm256_setzero_si256();
+
+  for (int i = 0; i < N; i += 2) {
+    // load src and all refs
+    src_reg = yy_loadu2_128(src + src_stride, src);
+    ref0_reg = yy_loadu2_128(ref0 + ref_stride, ref0);
+    ref1_reg = yy_loadu2_128(ref1 + ref_stride, ref1);
+    ref2_reg = yy_loadu2_128(ref2 + ref_stride, ref2);
+
+    // sum of the absolute differences between every ref-i to src
+    ref0_reg = _mm256_sad_epu8(ref0_reg, src_reg);
+    ref1_reg = _mm256_sad_epu8(ref1_reg, src_reg);
+    ref2_reg = _mm256_sad_epu8(ref2_reg, src_reg);
+
+    // sum every ref-i
+    sum_ref0 = _mm256_add_epi32(sum_ref0, ref0_reg);
+    sum_ref1 = _mm256_add_epi32(sum_ref1, ref1_reg);
+    sum_ref2 = _mm256_add_epi32(sum_ref2, ref2_reg);
+
+    src += 2 * src_stride;
+    ref0 += 2 * ref_stride;
+    ref1 += 2 * ref_stride;
+    ref2 += 2 * ref_stride;
+  }
+
+  aggregate_and_store_sum(res, &sum_ref0, &sum_ref1, &sum_ref2, &zero);
+}
+
+static AOM_FORCE_INLINE void aom_sad16xNx4d_avx2(int N, const uint8_t *src,
+                                                 int src_stride,
+                                                 const uint8_t *const ref[4],
+                                                 int ref_stride,
+                                                 uint32_t res[4]) {
+  __m256i src_reg, ref0_reg, ref1_reg, ref2_reg, ref3_reg;
+  __m256i sum_ref0, sum_ref1, sum_ref2, sum_ref3;
+  const uint8_t *ref0, *ref1, *ref2, *ref3;
+  assert(N % 2 == 0);
+
+  ref0 = ref[0];
+  ref1 = ref[1];
+  ref2 = ref[2];
+  ref3 = ref[3];
+
+  sum_ref0 = _mm256_setzero_si256();
+  sum_ref2 = _mm256_setzero_si256();
+  sum_ref1 = _mm256_setzero_si256();
+  sum_ref3 = _mm256_setzero_si256();
+
+  for (int i = 0; i < N; i += 2) {
+    // load src and all refs
+    src_reg = yy_loadu2_128(src + src_stride, src);
+    ref0_reg = yy_loadu2_128(ref0 + ref_stride, ref0);
+    ref1_reg = yy_loadu2_128(ref1 + ref_stride, ref1);
+    ref2_reg = yy_loadu2_128(ref2 + ref_stride, ref2);
+    ref3_reg = yy_loadu2_128(ref3 + ref_stride, ref3);
+
+    // sum of the absolute differences between every ref-i to src
+    ref0_reg = _mm256_sad_epu8(ref0_reg, src_reg);
+    ref1_reg = _mm256_sad_epu8(ref1_reg, src_reg);
+    ref2_reg = _mm256_sad_epu8(ref2_reg, src_reg);
+    ref3_reg = _mm256_sad_epu8(ref3_reg, src_reg);
+
+    // sum every ref-i
+    sum_ref0 = _mm256_add_epi32(sum_ref0, ref0_reg);
+    sum_ref1 = _mm256_add_epi32(sum_ref1, ref1_reg);
+    sum_ref2 = _mm256_add_epi32(sum_ref2, ref2_reg);
+    sum_ref3 = _mm256_add_epi32(sum_ref3, ref3_reg);
+
+    src += 2 * src_stride;
+    ref0 += 2 * ref_stride;
+    ref1 += 2 * ref_stride;
+    ref2 += 2 * ref_stride;
+    ref3 += 2 * ref_stride;
+  }
+
+  aggregate_and_store_sum(res, &sum_ref0, &sum_ref1, &sum_ref2, &sum_ref3);
+}
+
+#define SAD16XNX3_AVX2(n)                                                   \
+  void aom_sad16x##n##x3d_avx2(const uint8_t *src, int src_stride,          \
+                               const uint8_t *const ref[4], int ref_stride, \
+                               uint32_t res[4]) {                           \
+    aom_sad16xNx3d_avx2(n, src, src_stride, ref, ref_stride, res);          \
+  }
+#define SAD16XNX4_AVX2(n)                                                   \
+  void aom_sad16x##n##x4d_avx2(const uint8_t *src, int src_stride,          \
+                               const uint8_t *const ref[4], int ref_stride, \
+                               uint32_t res[4]) {                           \
+    aom_sad16xNx4d_avx2(n, src, src_stride, ref, ref_stride, res);          \
+  }
+
+SAD16XNX4_AVX2(32)
+SAD16XNX4_AVX2(16)
+SAD16XNX4_AVX2(8)
+
+SAD16XNX3_AVX2(32)
+SAD16XNX3_AVX2(16)
+SAD16XNX3_AVX2(8)
+
+#if !CONFIG_REALTIME_ONLY
+SAD16XNX3_AVX2(64)
+SAD16XNX3_AVX2(4)
+
+SAD16XNX4_AVX2(64)
+SAD16XNX4_AVX2(4)
+
+#endif  // !CONFIG_REALTIME_ONLY
+
+#define SAD_SKIP_16XN_AVX2(n)                                                 \
+  void aom_sad_skip_16x##n##x4d_avx2(const uint8_t *src, int src_stride,      \
+                                     const uint8_t *const ref[4],             \
+                                     int ref_stride, uint32_t res[4]) {       \
+    aom_sad16xNx4d_avx2(((n) >> 1), src, 2 * src_stride, ref, 2 * ref_stride, \
+                        res);                                                 \
+    res[0] <<= 1;                                                             \
+    res[1] <<= 1;                                                             \
+    res[2] <<= 1;                                                             \
+    res[3] <<= 1;                                                             \
+  }
+
+SAD_SKIP_16XN_AVX2(32)
+SAD_SKIP_16XN_AVX2(16)
+SAD_SKIP_16XN_AVX2(8)
+
+#if !CONFIG_REALTIME_ONLY
+SAD_SKIP_16XN_AVX2(64)
+SAD_SKIP_16XN_AVX2(4)
+#endif  // !CONFIG_REALTIME_ONLY
diff --git a/third_party/aom/aom_dsp/x86/sad4d_sse2.asm b/third_party/aom/aom_dsp/x86/sad4d_sse2.asm
new file mode 100644
index 0000000000..6edad99516
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/sad4d_sse2.asm
@@ -0,0 +1,437 @@
+;
+; Copyright (c) 2016, 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 "third_party/x86inc/x86inc.asm"
+
+SECTION .text
+
+; 'spill_src_stride' affect a lot how the code works.
+;
+; When 'spill_src_stride' is false, the 'src_strideq' resides in
+; register, [srcq + src_strideq + offset] is allowed, so we can simply
+; use such form to access src memory and don't bother to update 'srcq'
+; at each line. We only update 'srcq' each two-lines using a compact
+; LEA instruction like [srcq+src_strideq*2].
+;
+; When 'spill_src_stride' is true, the 'src_strideq' resides in memory.
+; we cannot use above form to access memory, we have to update
+; 'srcq' at each line break. As we process two parts (first,second)
+; together in each macro function, the second part may also sit
+; in the next line, which means we also need to possibly add
+; one 'src_strideq' to 'srcq' before processing second part.
+
+%macro HANDLE_SECOND_OFFSET 0
+  %if spill_src_stride
+    %define second_offset 0
+    add srcq, src_strideq
+  %else
+    %define second_offset (src_strideq)
+  %endif
+%endmacro
+
+; This is specically designed to handle when src_strideq is a
+; memory position, under such case, we can not accomplish
+; complex address calculation using LEA, and fall back to
+; using simple ADD instruction at each line ending.
+%macro ADVANCE_END_OF_TWO_LINES 0
+  %if spill_src_stride
+    add srcq, src_strideq
+  %else
+    lea                 srcq, [srcq+src_strideq*2]
+  %endif
+
+; note: ref_stride is never spilled when processing two lines
+  lea                ref1q, [ref1q+ref_strideq*2]
+  lea                ref2q, [ref2q+ref_strideq*2]
+  lea                ref3q, [ref3q+ref_strideq*2]
+  lea                ref4q, [ref4q+ref_strideq*2]
+%endmacro
+
+; PROCESS_4x2x4 first
+%macro PROCESS_4x2x4 1
+  movd                  m0, [srcq]
+  HANDLE_SECOND_OFFSET
+%if %1 == 1
+  movd                  m6, [ref1q]
+  movd                  m4, [ref2q]
+  movd                  m7, [ref3q]
+  movd                  m5, [ref4q]
+
+  movd                  m1, [srcq + second_offset]
+  movd                  m2, [ref1q+ref_strideq]
+  punpckldq             m0, m1
+  punpckldq             m6, m2
+  movd                  m1, [ref2q+ref_strideq]
+  movd                  m2, [ref3q+ref_strideq]
+  movd                  m3, [ref4q+ref_strideq]
+  punpckldq             m4, m1
+  punpckldq             m7, m2
+  punpckldq             m5, m3
+  movlhps               m0, m0
+  movlhps               m6, m4
+  movlhps               m7, m5
+  psadbw                m6, m0
+  psadbw                m7, m0
+%else
+  movd                  m1, [ref1q]
+  movd                  m5, [ref1q+ref_strideq]
+  movd                  m2, [ref2q]
+  movd                  m4, [ref2q+ref_strideq]
+  punpckldq             m1, m5
+  punpckldq             m2, m4
+  movd                  m3, [ref3q]
+  movd                  m5, [ref3q+ref_strideq]
+  punpckldq             m3, m5
+  movd                  m4, [ref4q]
+  movd                  m5, [ref4q+ref_strideq]
+  punpckldq             m4, m5
+  movd                  m5, [srcq + second_offset]
+  punpckldq             m0, m5
+  movlhps               m0, m0
+  movlhps               m1, m2
+  movlhps               m3, m4
+  psadbw                m1, m0
+  psadbw                m3, m0
+  paddd                 m6, m1
+  paddd                 m7, m3
+%endif
+%endmacro
+
+; PROCESS_8x2x4 first
+%macro PROCESS_8x2x4 1
+  movh                  m0, [srcq]
+  HANDLE_SECOND_OFFSET
+%if %1 == 1
+  movh                  m4, [ref1q]
+  movh                  m5, [ref2q]
+  movh                  m6, [ref3q]
+  movh                  m7, [ref4q]
+  movhps                m0, [srcq + second_offset]
+  movhps                m4, [ref1q+ref_strideq]
+  movhps                m5, [ref2q+ref_strideq]
+  movhps                m6, [ref3q+ref_strideq]
+  movhps                m7, [ref4q+ref_strideq]
+  psadbw                m4, m0
+  psadbw                m5, m0
+  psadbw                m6, m0
+  psadbw                m7, m0
+%else
+  movh                  m1, [ref1q]
+  movh                  m2, [ref2q]
+  movhps                m0, [srcq + second_offset]
+  movhps                m1, [ref1q+ref_strideq]
+  movhps                m2, [ref2q+ref_strideq]
+  psadbw                m1, m0
+  psadbw                m2, m0
+  paddd                 m4, m1
+  paddd                 m5, m2
+
+  movh                  m1, [ref3q]
+  movhps                m1, [ref3q+ref_strideq]
+  movh                  m2, [ref4q]
+  movhps                m2, [ref4q+ref_strideq]
+  psadbw                m1, m0
+  psadbw                m2, m0
+  paddd                 m6, m1
+  paddd                 m7, m2
+%endif
+%endmacro
+
+; PROCESS_FIRST_MMSIZE
+%macro PROCESS_FIRST_MMSIZE 0
+  mova                  m0, [srcq]
+  movu                  m4, [ref1q]
+  movu                  m5, [ref2q]
+  movu                  m6, [ref3q]
+  movu                  m7, [ref4q]
+  psadbw                m4, m0
+  psadbw                m5, m0
+  psadbw                m6, m0
+  psadbw                m7, m0
+%endmacro
+
+; PROCESS_16x1x4 offset
+%macro PROCESS_16x1x4 1
+  mova                  m0, [srcq + %1]
+  movu                  m1, [ref1q + ref_offsetq + %1]
+  movu                  m2, [ref2q + ref_offsetq + %1]
+  psadbw                m1, m0
+  psadbw                m2, m0
+  paddd                 m4, m1
+  paddd                 m5, m2
+
+  movu                  m1, [ref3q + ref_offsetq + %1]
+  movu                  m2, [ref4q + ref_offsetq + %1]
+  psadbw                m1, m0
+  psadbw                m2, m0
+  paddd                 m6, m1
+  paddd                 m7, m2
+%endmacro
+
+; void aom_sadNxNx4d_sse2(uint8_t *src,    int src_stride,
+;                         uint8_t *ref[4], int ref_stride,
+;                         uint32_t res[4]);
+; Macro Arguments:
+;   1: Width
+;   2: Height
+;   3: If 0, then normal sad, else skip rows
+%macro SADNXN4D 2-3 0
+
+%define spill_src_stride 0
+%define spill_ref_stride 0
+%define spill_cnt 0
+
+; Whether a shared offset should be used instead of adding strides to
+; each reference array. With this option, only one line will be processed
+; per loop iteration.
+%define use_ref_offset (%1 >= mmsize)
+
+; Remove loops in the 4x4 and 8x4 case
+%define use_loop (use_ref_offset || %2 > 4)
+
+%if %3 == 1  ; skip rows
+%if AOM_ARCH_X86_64
+%if use_ref_offset
+cglobal sad_skip_%1x%2x4d, 5, 10, 8, src, src_stride, ref1, ref_stride, res, \
+                                     ref2, ref3, ref4, cnt, ref_offset
+%elif use_loop
+cglobal sad_skip_%1x%2x4d, 5, 9, 8, src, src_stride, ref1, ref_stride, res, \
+                                    ref2, ref3, ref4, cnt
+%else
+cglobal sad_skip_%1x%2x4d, 5, 8, 8, src, src_stride, ref1, ref_stride, res, \
+                                    ref2, ref3, ref4
+%endif
+%else
+%if use_ref_offset
+cglobal sad_skip_%1x%2x4d, 4, 7, 8, src, ref_offset, ref1, cnt, ref2, ref3, \
+                                    ref4
+%define spill_src_stride 1
+%define spill_ref_stride 1
+%elif use_loop
+cglobal sad_skip_%1x%2x4d, 4, 7, 8, src, cnt, ref1, ref_stride, ref2, \
+                                    ref3, ref4
+%define spill_src_stride 1
+%else
+cglobal sad_skip_%1x%2x4d, 4, 7, 8, src, src_stride, ref1, ref_stride, ref2, \
+                                    ref3, ref4
+%endif
+%endif
+%else ; normal sad
+%if AOM_ARCH_X86_64
+%if use_ref_offset
+cglobal sad%1x%2x4d, 5, 10, 8, src, src_stride, ref1, ref_stride, res, ref2, \
+                               ref3, ref4, cnt, ref_offset
+%elif use_loop
+cglobal sad%1x%2x4d, 5, 9, 8, src, src_stride, ref1, ref_stride, res, ref2, \
+                              ref3, ref4, cnt
+%else
+cglobal sad%1x%2x4d, 5, 8, 8, src, src_stride, ref1, ref_stride, res, ref2, \
+                              ref3, ref4
+%endif
+%else
+%if use_ref_offset
+cglobal sad%1x%2x4d, 4, 7, 8, src, ref_offset, ref1, cnt, ref2, ref3, ref4
+  %define spill_src_stride 1
+  %define spill_ref_stride 1
+%elif use_loop
+cglobal sad%1x%2x4d, 4, 7, 8, src, cnt, ref1, ref_stride, ref2, ref3, ref4
+  %define spill_src_stride 1
+%else
+cglobal sad%1x%2x4d, 4, 7, 8, src, src_stride, ref1, ref_stride, ref2, ref3, \
+                              ref4
+%endif
+%endif
+%endif
+
+%if spill_src_stride
+  %define src_strideq r1mp
+  %define src_strided r1mp
+%endif
+%if spill_ref_stride
+  %define ref_strideq r3mp
+  %define ref_strided r3mp
+%endif
+
+%if spill_cnt
+  SUB                  rsp, 4
+  %define cntd word [rsp]
+%endif
+
+%if %3 == 1
+  sal          src_strided, 1
+  sal          ref_strided, 1
+%endif
+  movsxdifnidn src_strideq, src_strided
+  movsxdifnidn ref_strideq, ref_strided
+
+  mov                ref2q, [ref1q+gprsize*1]
+  mov                ref3q, [ref1q+gprsize*2]
+  mov                ref4q, [ref1q+gprsize*3]
+  mov                ref1q, [ref1q+gprsize*0]
+
+; Is the loop for this wxh in another function?
+; If so, we jump into that function for the loop and returning
+%define external_loop (use_ref_offset && %1 > mmsize && %1 != %2)
+
+%if use_ref_offset
+  PROCESS_FIRST_MMSIZE
+%if %1 > mmsize
+  mov          ref_offsetq, 0
+  mov                 cntd, %2 >> %3
+; Jump part way into the loop for the square version of this width
+%if %3 == 1
+  jmp mangle(private_prefix %+ _sad_skip_%1x%1x4d %+ SUFFIX).midloop
+%else
+  jmp mangle(private_prefix %+ _sad%1x%1x4d %+ SUFFIX).midloop
+%endif
+%else
+  mov          ref_offsetq, ref_strideq
+  add                 srcq, src_strideq
+  mov                 cntd, (%2 >> %3) - 1
+%endif
+%if external_loop == 0
+.loop:
+; Unrolled horizontal loop
+%assign h_offset 0
+%rep %1/mmsize
+  PROCESS_16x1x4 h_offset
+%if h_offset == 0
+; The first row of the first column is done outside the loop and jumps here
+.midloop:
+%endif
+%assign h_offset h_offset+mmsize
+%endrep
+
+  add                 srcq, src_strideq
+  add          ref_offsetq, ref_strideq
+  sub                 cntd, 1
+  jnz .loop
+%endif
+%else
+  PROCESS_%1x2x4 1
+  ADVANCE_END_OF_TWO_LINES
+%if use_loop
+  mov                 cntd, (%2/2 >> %3) - 1
+.loop:
+%endif
+  PROCESS_%1x2x4 0
+%if use_loop
+  ADVANCE_END_OF_TWO_LINES
+  sub                 cntd, 1
+  jnz .loop
+%endif
+%endif
+
+%if spill_cnt
+; Undo stack allocation for cnt
+  ADD                  rsp, 4
+%endif
+
+%if external_loop == 0
+%if %3 == 0
+  %define resultq r4
+  %define resultmp r4mp
+%endif
+
+; Undo modifications on parameters on the stack
+%if %3 == 1
+%if spill_src_stride
+  shr          src_strided, 1
+%endif
+%if spill_ref_stride
+  shr          ref_strided, 1
+%endif
+%endif
+
+%if %1 > 4
+  pslldq                m5, 4
+  pslldq                m7, 4
+  por                   m4, m5
+  por                   m6, m7
+  mova                  m5, m4
+  mova                  m7, m6
+  punpcklqdq            m4, m6
+  punpckhqdq            m5, m7
+  paddd                 m4, m5
+%if %3 == 1
+  pslld                 m4, 1
+%endif
+  movifnidn             resultq, resultmp
+  movu                [resultq], m4
+  RET
+%else
+  pshufd            m6, m6, 0x08
+  pshufd            m7, m7, 0x08
+%if %3 == 1
+  pslld                 m6, 1
+  pslld                 m7, 1
+%endif
+  movifnidn             resultq, resultmp
+  movq              [resultq+0], m6
+  movq              [resultq+8], m7
+  RET
+%endif
+%endif ; external_loop == 0
+%endmacro
+
+INIT_XMM sse2
+SADNXN4D 128, 128
+SADNXN4D 128,  64
+SADNXN4D  64, 128
+SADNXN4D  64,  64
+SADNXN4D  64,  32
+SADNXN4D  32,  64
+SADNXN4D  32,  32
+SADNXN4D  32,  16
+SADNXN4D  16,  32
+SADNXN4D  16,  16
+SADNXN4D  16,   8
+SADNXN4D   8,  16
+SADNXN4D   8,   8
+SADNXN4D   8,   4
+SADNXN4D   4,   8
+SADNXN4D   4,   4
+%if CONFIG_REALTIME_ONLY==0
+SADNXN4D   4,  16
+SADNXN4D  16,   4
+SADNXN4D   8,  32
+SADNXN4D  32,   8
+SADNXN4D  16,  64
+SADNXN4D  64,  16
+%endif
+SADNXN4D 128, 128, 1
+SADNXN4D 128,  64, 1
+SADNXN4D  64, 128, 1
+SADNXN4D  64,  64, 1
+SADNXN4D  64,  32, 1
+SADNXN4D  32,  64, 1
+SADNXN4D  32,  32, 1
+SADNXN4D  32,  16, 1
+SADNXN4D  16,  32, 1
+SADNXN4D  16,  16, 1
+SADNXN4D  16,   8, 1
+SADNXN4D   8,  16, 1
+SADNXN4D   8,   8, 1
+SADNXN4D   4,   8, 1
+%if CONFIG_REALTIME_ONLY==0
+SADNXN4D   4,  16, 1
+SADNXN4D   8,  32, 1
+SADNXN4D  32,   8, 1
+SADNXN4D  16,  64, 1
+SADNXN4D  64,  16, 1
+%endif
+
+; Different assembly is needed when the height gets subsampled to 2
+; SADNXN4D 16,  4, 1
+; SADNXN4D  8,  4, 1
+; SADNXN4D  4,  4, 1
diff --git a/third_party/aom/aom_dsp/x86/sad_avx2.c b/third_party/aom/aom_dsp/x86/sad_avx2.c
new file mode 100644
index 0000000000..24cea76b37
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/sad_avx2.c
@@ -0,0 +1,219 @@
+/*
+ * Copyright (c) 2016, 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"
+
+#include "aom_ports/mem.h"
+
+static INLINE unsigned int sad64xh_avx2(const uint8_t *src_ptr, int src_stride,
+                                        const uint8_t *ref_ptr, int ref_stride,
+                                        int h) {
+  int i;
+  __m256i sad1_reg, sad2_reg, ref1_reg, ref2_reg;
+  __m256i sum_sad = _mm256_setzero_si256();
+  __m256i sum_sad_h;
+  __m128i sum_sad128;
+  for (i = 0; i < h; i++) {
+    ref1_reg = _mm256_loadu_si256((__m256i const *)ref_ptr);
+    ref2_reg = _mm256_loadu_si256((__m256i const *)(ref_ptr + 32));
+    sad1_reg =
+        _mm256_sad_epu8(ref1_reg, _mm256_loadu_si256((__m256i const *)src_ptr));
+    sad2_reg = _mm256_sad_epu8(
+        ref2_reg, _mm256_loadu_si256((__m256i const *)(src_ptr + 32)));
+    sum_sad = _mm256_add_epi32(sum_sad, _mm256_add_epi32(sad1_reg, sad2_reg));
+    ref_ptr += ref_stride;
+    src_ptr += src_stride;
+  }
+  sum_sad_h = _mm256_srli_si256(sum_sad, 8);
+  sum_sad = _mm256_add_epi32(sum_sad, sum_sad_h);
+  sum_sad128 = _mm256_extracti128_si256(sum_sad, 1);
+  sum_sad128 = _mm_add_epi32(_mm256_castsi256_si128(sum_sad), sum_sad128);
+  unsigned int res = (unsigned int)_mm_cvtsi128_si32(sum_sad128);
+  _mm256_zeroupper();
+  return res;
+}
+
+static INLINE unsigned int sad32xh_avx2(const uint8_t *src_ptr, int src_stride,
+                                        const uint8_t *ref_ptr, int ref_stride,
+                                        int h) {
+  int i;
+  __m256i sad1_reg, sad2_reg, ref1_reg, ref2_reg;
+  __m256i sum_sad = _mm256_setzero_si256();
+  __m256i sum_sad_h;
+  __m128i sum_sad128;
+  int ref2_stride = ref_stride << 1;
+  int src2_stride = src_stride << 1;
+  int max = h >> 1;
+  for (i = 0; i < max; i++) {
+    ref1_reg = _mm256_loadu_si256((__m256i const *)ref_ptr);
+    ref2_reg = _mm256_loadu_si256((__m256i const *)(ref_ptr + ref_stride));
+    sad1_reg =
+        _mm256_sad_epu8(ref1_reg, _mm256_loadu_si256((__m256i const *)src_ptr));
+    sad2_reg = _mm256_sad_epu8(
+        ref2_reg, _mm256_loadu_si256((__m256i const *)(src_ptr + src_stride)));
+    sum_sad = _mm256_add_epi32(sum_sad, _mm256_add_epi32(sad1_reg, sad2_reg));
+    ref_ptr += ref2_stride;
+    src_ptr += src2_stride;
+  }
+  sum_sad_h = _mm256_srli_si256(sum_sad, 8);
+  sum_sad = _mm256_add_epi32(sum_sad, sum_sad_h);
+  sum_sad128 = _mm256_extracti128_si256(sum_sad, 1);
+  sum_sad128 = _mm_add_epi32(_mm256_castsi256_si128(sum_sad), sum_sad128);
+  unsigned int res = (unsigned int)_mm_cvtsi128_si32(sum_sad128);
+  _mm256_zeroupper();
+  return res;
+}
+
+#define FSAD64_H(h)                                                           \
+  unsigned int aom_sad64x##h##_avx2(const uint8_t *src_ptr, int src_stride,   \
+                                    const uint8_t *ref_ptr, int ref_stride) { \
+    return sad64xh_avx2(src_ptr, src_stride, ref_ptr, ref_stride, h);         \
+  }
+
+#define FSADS64_H(h)                                                          \
+  unsigned int aom_sad_skip_64x##h##_avx2(                                    \
+      const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr,         \
+      int ref_stride) {                                                       \
+    return 2 * sad64xh_avx2(src_ptr, src_stride * 2, ref_ptr, ref_stride * 2, \
+                            h / 2);                                           \
+  }
+
+#define FSAD32_H(h)                                                           \
+  unsigned int aom_sad32x##h##_avx2(const uint8_t *src_ptr, int src_stride,   \
+                                    const uint8_t *ref_ptr, int ref_stride) { \
+    return sad32xh_avx2(src_ptr, src_stride, ref_ptr, ref_stride, h);         \
+  }
+
+#define FSADS32_H(h)                                                          \
+  unsigned int aom_sad_skip_32x##h##_avx2(                                    \
+      const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr,         \
+      int ref_stride) {                                                       \
+    return 2 * sad32xh_avx2(src_ptr, src_stride * 2, ref_ptr, ref_stride * 2, \
+                            h / 2);                                           \
+  }
+
+#define FSAD64  \
+  FSAD64_H(64)  \
+  FSAD64_H(32)  \
+  FSADS64_H(64) \
+  FSADS64_H(32)
+
+#define FSAD32  \
+  FSAD32_H(64)  \
+  FSAD32_H(32)  \
+  FSAD32_H(16)  \
+  FSADS32_H(64) \
+  FSADS32_H(32) \
+  FSADS32_H(16)
+
+/* clang-format off */
+FSAD64
+FSAD32
+/* clang-format on */
+
+#undef FSAD64
+#undef FSAD32
+#undef FSAD64_H
+#undef FSAD32_H
+
+#define FSADAVG64_H(h)                                                        \
+  unsigned int aom_sad64x##h##_avg_avx2(                                      \
+      const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr,         \
+      int ref_stride, const uint8_t *second_pred) {                           \
+    int i;                                                                    \
+    __m256i sad1_reg, sad2_reg, ref1_reg, ref2_reg;                           \
+    __m256i sum_sad = _mm256_setzero_si256();                                 \
+    __m256i sum_sad_h;                                                        \
+    __m128i sum_sad128;                                                       \
+    for (i = 0; i < h; i++) {                                                 \
+      ref1_reg = _mm256_loadu_si256((__m256i const *)ref_ptr);                \
+      ref2_reg = _mm256_loadu_si256((__m256i const *)(ref_ptr + 32));         \
+      ref1_reg = _mm256_avg_epu8(                                             \
+          ref1_reg, _mm256_loadu_si256((__m256i const *)second_pred));        \
+      ref2_reg = _mm256_avg_epu8(                                             \
+          ref2_reg, _mm256_loadu_si256((__m256i const *)(second_pred + 32))); \
+      sad1_reg = _mm256_sad_epu8(                                             \
+          ref1_reg, _mm256_loadu_si256((__m256i const *)src_ptr));            \
+      sad2_reg = _mm256_sad_epu8(                                             \
+          ref2_reg, _mm256_loadu_si256((__m256i const *)(src_ptr + 32)));     \
+      sum_sad =                                                               \
+          _mm256_add_epi32(sum_sad, _mm256_add_epi32(sad1_reg, sad2_reg));    \
+      ref_ptr += ref_stride;                                                  \
+      src_ptr += src_stride;                                                  \
+      second_pred += 64;                                                      \
+    }                                                                         \
+    sum_sad_h = _mm256_srli_si256(sum_sad, 8);                                \
+    sum_sad = _mm256_add_epi32(sum_sad, sum_sad_h);                           \
+    sum_sad128 = _mm256_extracti128_si256(sum_sad, 1);                        \
+    sum_sad128 = _mm_add_epi32(_mm256_castsi256_si128(sum_sad), sum_sad128);  \
+    unsigned int res = (unsigned int)_mm_cvtsi128_si32(sum_sad128);           \
+    _mm256_zeroupper();                                                       \
+    return res;                                                               \
+  }
+
+#define FSADAVG32_H(h)                                                        \
+  unsigned int aom_sad32x##h##_avg_avx2(                                      \
+      const uint8_t *src_ptr, int src_stride, const uint8_t *ref_ptr,         \
+      int ref_stride, const uint8_t *second_pred) {                           \
+    int i;                                                                    \
+    __m256i sad1_reg, sad2_reg, ref1_reg, ref2_reg;                           \
+    __m256i sum_sad = _mm256_setzero_si256();                                 \
+    __m256i sum_sad_h;                                                        \
+    __m128i sum_sad128;                                                       \
+    int ref2_stride = ref_stride << 1;                                        \
+    int src2_stride = src_stride << 1;                                        \
+    int max = h >> 1;                                                         \
+    for (i = 0; i < max; i++) {                                               \
+      ref1_reg = _mm256_loadu_si256((__m256i const *)ref_ptr);                \
+      ref2_reg = _mm256_loadu_si256((__m256i const *)(ref_ptr + ref_stride)); \
+      ref1_reg = _mm256_avg_epu8(                                             \
+          ref1_reg, _mm256_loadu_si256((__m256i const *)second_pred));        \
+      ref2_reg = _mm256_avg_epu8(                                             \
+          ref2_reg, _mm256_loadu_si256((__m256i const *)(second_pred + 32))); \
+      sad1_reg = _mm256_sad_epu8(                                             \
+          ref1_reg, _mm256_loadu_si256((__m256i const *)src_ptr));            \
+      sad2_reg = _mm256_sad_epu8(                                             \
+          ref2_reg,                                                           \
+          _mm256_loadu_si256((__m256i const *)(src_ptr + src_stride)));       \
+      sum_sad =                                                               \
+          _mm256_add_epi32(sum_sad, _mm256_add_epi32(sad1_reg, sad2_reg));    \
+      ref_ptr += ref2_stride;                                                 \
+      src_ptr += src2_stride;                                                 \
+      second_pred += 64;                                                      \
+    }                                                                         \
+    sum_sad_h = _mm256_srli_si256(sum_sad, 8);                                \
+    sum_sad = _mm256_add_epi32(sum_sad, sum_sad_h);                           \
+    sum_sad128 = _mm256_extracti128_si256(sum_sad, 1);                        \
+    sum_sad128 = _mm_add_epi32(_mm256_castsi256_si128(sum_sad), sum_sad128);  \
+    unsigned int res = (unsigned int)_mm_cvtsi128_si32(sum_sad128);           \
+    _mm256_zeroupper();                                                       \
+    return res;                                                               \
+  }
+
+#define FSADAVG64 \
+  FSADAVG64_H(64) \
+  FSADAVG64_H(32)
+
+#define FSADAVG32 \
+  FSADAVG32_H(64) \
+  FSADAVG32_H(32) \
+  FSADAVG32_H(16)
+
+/* clang-format off */
+FSADAVG64
+FSADAVG32
+/* clang-format on */
+
+#undef FSADAVG64
+#undef FSADAVG32
+#undef FSADAVG64_H
+#undef FSADAVG32_H
diff --git a/third_party/aom/aom_dsp/x86/sad_impl_avx2.c b/third_party/aom/aom_dsp/x86/sad_impl_avx2.c
new file mode 100644
index 0000000000..c5da6e9ab3
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/sad_impl_avx2.c
@@ -0,0 +1,181 @@
+/*
+ * Copyright (c) 2016, 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 unsigned int sad32x32(const uint8_t *src_ptr, int src_stride,
+                             const uint8_t *ref_ptr, int ref_stride) {
+  __m256i s1, s2, r1, r2;
+  __m256i sum = _mm256_setzero_si256();
+  __m128i sum_i128;
+  int i;
+
+  for (i = 0; i < 16; ++i) {
+    r1 = _mm256_loadu_si256((__m256i const *)ref_ptr);
+    r2 = _mm256_loadu_si256((__m256i const *)(ref_ptr + ref_stride));
+    s1 = _mm256_sad_epu8(r1, _mm256_loadu_si256((__m256i const *)src_ptr));
+    s2 = _mm256_sad_epu8(
+        r2, _mm256_loadu_si256((__m256i const *)(src_ptr + src_stride)));
+    sum = _mm256_add_epi32(sum, _mm256_add_epi32(s1, s2));
+    ref_ptr += ref_stride << 1;
+    src_ptr += src_stride << 1;
+  }
+
+  sum = _mm256_add_epi32(sum, _mm256_srli_si256(sum, 8));
+  sum_i128 = _mm_add_epi32(_mm256_extracti128_si256(sum, 1),
+                           _mm256_castsi256_si128(sum));
+  return (unsigned int)_mm_cvtsi128_si32(sum_i128);
+}
+
+static unsigned int sad64x32(const uint8_t *src_ptr, int src_stride,
+                             const uint8_t *ref_ptr, int ref_stride) {
+  unsigned int half_width = 32;
+  uint32_t sum = sad32x32(src_ptr, src_stride, ref_ptr, ref_stride);
+  src_ptr += half_width;
+  ref_ptr += half_width;
+  sum += sad32x32(src_ptr, src_stride, ref_ptr, ref_stride);
+  return sum;
+}
+
+static unsigned int sad64x64(const uint8_t *src_ptr, int src_stride,
+                             const uint8_t *ref_ptr, int ref_stride) {
+  uint32_t sum = sad64x32(src_ptr, src_stride, ref_ptr, ref_stride);
+  src_ptr += src_stride << 5;
+  ref_ptr += ref_stride << 5;
+  sum += sad64x32(src_ptr, src_stride, ref_ptr, ref_stride);
+  return sum;
+}
+
+unsigned int aom_sad128x64_avx2(const uint8_t *src_ptr, int src_stride,
+                                const uint8_t *ref_ptr, int ref_stride) {
+  unsigned int half_width = 64;
+  uint32_t sum = sad64x64(src_ptr, src_stride, ref_ptr, ref_stride);
+  src_ptr += half_width;
+  ref_ptr += half_width;
+  sum += sad64x64(src_ptr, src_stride, ref_ptr, ref_stride);
+  return sum;
+}
+
+unsigned int aom_sad64x128_avx2(const uint8_t *src_ptr, int src_stride,
+                                const uint8_t *ref_ptr, int ref_stride) {
+  uint32_t sum = sad64x64(src_ptr, src_stride, ref_ptr, ref_stride);
+  src_ptr += src_stride << 6;
+  ref_ptr += ref_stride << 6;
+  sum += sad64x64(src_ptr, src_stride, ref_ptr, ref_stride);
+  return sum;
+}
+
+unsigned int aom_sad128x128_avx2(const uint8_t *src_ptr, int src_stride,
+                                 const uint8_t *ref_ptr, int ref_stride) {
+  uint32_t sum = aom_sad128x64_avx2(src_ptr, src_stride, ref_ptr, ref_stride);
+  src_ptr += src_stride << 6;
+  ref_ptr += ref_stride << 6;
+  sum += aom_sad128x64_avx2(src_ptr, src_stride, ref_ptr, ref_stride);
+  return sum;
+}
+
+unsigned int aom_sad_skip_128x64_avx2(const uint8_t *src_ptr, int src_stride,
+                                      const uint8_t *ref_ptr, int ref_stride) {
+  const uint32_t half_width = 64;
+  uint32_t sum = sad64x32(src_ptr, src_stride * 2, ref_ptr, ref_stride * 2);
+  src_ptr += half_width;
+  ref_ptr += half_width;
+  sum += sad64x32(src_ptr, src_stride * 2, ref_ptr, ref_stride * 2);
+  return 2 * sum;
+}
+
+unsigned int aom_sad_skip_64x128_avx2(const uint8_t *src_ptr, int src_stride,
+                                      const uint8_t *ref_ptr, int ref_stride) {
+  const uint32_t sum =
+      sad64x64(src_ptr, 2 * src_stride, ref_ptr, 2 * ref_stride);
+  return 2 * sum;
+}
+
+unsigned int aom_sad_skip_128x128_avx2(const uint8_t *src_ptr, int src_stride,
+                                       const uint8_t *ref_ptr, int ref_stride) {
+  const uint32_t sum =
+      aom_sad128x64_avx2(src_ptr, 2 * src_stride, ref_ptr, 2 * ref_stride);
+  return 2 * sum;
+}
+
+static unsigned int sad_w64_avg_avx2(const uint8_t *src_ptr, int src_stride,
+                                     const uint8_t *ref_ptr, int ref_stride,
+                                     const int h, const uint8_t *second_pred,
+                                     const int second_pred_stride) {
+  int i;
+  __m256i sad1_reg, sad2_reg, ref1_reg, ref2_reg;
+  __m256i sum_sad = _mm256_setzero_si256();
+  __m256i sum_sad_h;
+  __m128i sum_sad128;
+  for (i = 0; i < h; i++) {
+    ref1_reg = _mm256_loadu_si256((__m256i const *)ref_ptr);
+    ref2_reg = _mm256_loadu_si256((__m256i const *)(ref_ptr + 32));
+    ref1_reg = _mm256_avg_epu8(
+        ref1_reg, _mm256_loadu_si256((__m256i const *)second_pred));
+    ref2_reg = _mm256_avg_epu8(
+        ref2_reg, _mm256_loadu_si256((__m256i const *)(second_pred + 32)));
+    sad1_reg =
+        _mm256_sad_epu8(ref1_reg, _mm256_loadu_si256((__m256i const *)src_ptr));
+    sad2_reg = _mm256_sad_epu8(
+        ref2_reg, _mm256_loadu_si256((__m256i const *)(src_ptr + 32)));
+    sum_sad = _mm256_add_epi32(sum_sad, _mm256_add_epi32(sad1_reg, sad2_reg));
+    ref_ptr += ref_stride;
+    src_ptr += src_stride;
+    second_pred += second_pred_stride;
+  }
+  sum_sad_h = _mm256_srli_si256(sum_sad, 8);
+  sum_sad = _mm256_add_epi32(sum_sad, sum_sad_h);
+  sum_sad128 = _mm256_extracti128_si256(sum_sad, 1);
+  sum_sad128 = _mm_add_epi32(_mm256_castsi256_si128(sum_sad), sum_sad128);
+  return (unsigned int)_mm_cvtsi128_si32(sum_sad128);
+}
+
+unsigned int aom_sad64x128_avg_avx2(const uint8_t *src_ptr, int src_stride,
+                                    const uint8_t *ref_ptr, int ref_stride,
+                                    const uint8_t *second_pred) {
+  uint32_t sum = sad_w64_avg_avx2(src_ptr, src_stride, ref_ptr, ref_stride, 64,
+                                  second_pred, 64);
+  src_ptr += src_stride << 6;
+  ref_ptr += ref_stride << 6;
+  second_pred += 64 << 6;
+  sum += sad_w64_avg_avx2(src_ptr, src_stride, ref_ptr, ref_stride, 64,
+                          second_pred, 64);
+  return sum;
+}
+
+unsigned int aom_sad128x64_avg_avx2(const uint8_t *src_ptr, int src_stride,
+                                    const uint8_t *ref_ptr, int ref_stride,
+                                    const uint8_t *second_pred) {
+  unsigned int half_width = 64;
+  uint32_t sum = sad_w64_avg_avx2(src_ptr, src_stride, ref_ptr, ref_stride, 64,
+                                  second_pred, 128);
+  src_ptr += half_width;
+  ref_ptr += half_width;
+  second_pred += half_width;
+  sum += sad_w64_avg_avx2(src_ptr, src_stride, ref_ptr, ref_stride, 64,
+                          second_pred, 128);
+  return sum;
+}
+
+unsigned int aom_sad128x128_avg_avx2(const uint8_t *src_ptr, int src_stride,
+                                     const uint8_t *ref_ptr, int ref_stride,
+                                     const uint8_t *second_pred) {
+  uint32_t sum = aom_sad128x64_avg_avx2(src_ptr, src_stride, ref_ptr,
+                                        ref_stride, second_pred);
+  src_ptr += src_stride << 6;
+  ref_ptr += ref_stride << 6;
+  second_pred += 128 << 6;
+  sum += aom_sad128x64_avg_avx2(src_ptr, src_stride, ref_ptr, ref_stride,
+                                second_pred);
+  return sum;
+}
diff --git a/third_party/aom/aom_dsp/x86/sad_sse2.asm b/third_party/aom/aom_dsp/x86/sad_sse2.asm
new file mode 100644
index 0000000000..dbe8ca3161
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/sad_sse2.asm
@@ -0,0 +1,432 @@
+;
+; Copyright (c) 2016, 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 "third_party/x86inc/x86inc.asm"
+
+SECTION .text
+
+; Macro Arguments
+; Arg 1: Width
+; Arg 2: Height
+; Arg 3: Number of general purpose registers: 5 for 32-bit build, 6 for 64-bit
+; Arg 4: Type of function: if 0, normal sad; if 1, avg; if 2, skip rows
+%macro SAD_FN 4
+%if %4 == 0 ; normal sad
+%if %3 == 5
+cglobal sad%1x%2, 4, %3, 5, src, src_stride, ref, ref_stride, n_rows
+%else ; %3 == 7
+cglobal sad%1x%2, 4, %3, 6, src, src_stride, ref, ref_stride, \
+                            src_stride3, ref_stride3, n_rows
+%endif ; %3 == 5/7
+
+%elif %4 == 2 ; skip
+%if %3 == 5
+cglobal sad_skip_%1x%2, 4, %3, 5, src, src_stride, ref, ref_stride, n_rows
+%else ; %3 == 7
+cglobal sad_skip_%1x%2, 4, %3, 6, src, src_stride, ref, ref_stride, \
+                            src_stride3, ref_stride3, n_rows
+%endif ; %3 == 5/7
+
+%else
+%if %3 == 5
+cglobal sad%1x%2_avg, 5, 1 + %3, 5, src, src_stride, ref, ref_stride, \
+                                    second_pred, n_rows
+%else ; %3 == 7
+cglobal sad%1x%2_avg, 5, AOM_ARCH_X86_64 + %3, 6, src, src_stride, \
+                                              ref, ref_stride, \
+                                              second_pred, \
+                                              src_stride3, ref_stride3
+%if AOM_ARCH_X86_64
+%define n_rowsd r7d
+%else ; x86-32
+%define n_rowsd dword r0m
+%endif ; x86-32/64
+%endif ; %3 == 5/7
+%endif ; sad/avg/skip
+%if %4 == 2; skip rows so double the stride
+lea           src_strided, [src_strided*2]
+lea           ref_strided, [ref_strided*2]
+%endif ; %4 skip
+  movsxdifnidn src_strideq, src_strided
+  movsxdifnidn ref_strideq, ref_strided
+%if %3 == 7
+  lea         src_stride3q, [src_strideq*3]
+  lea         ref_stride3q, [ref_strideq*3]
+%endif ; %3 == 7
+%endmacro
+
+; unsigned int aom_sad128x128_sse2(uint8_t *src, int src_stride,
+;                                  uint8_t *ref, int ref_stride);
+%macro SAD128XN 1-2 0
+  SAD_FN 128, %1, 5, %2
+%if %2 == 2
+  mov              n_rowsd, %1/2
+%else
+  mov              n_rowsd, %1
+%endif
+  pxor                  m0, m0
+
+.loop:
+  movu                  m1, [refq]
+  movu                  m2, [refq+16]
+  movu                  m3, [refq+32]
+  movu                  m4, [refq+48]
+%if %2 == 1
+  pavgb                 m1, [second_predq+mmsize*0]
+  pavgb                 m2, [second_predq+mmsize*1]
+  pavgb                 m3, [second_predq+mmsize*2]
+  pavgb                 m4, [second_predq+mmsize*3]
+%endif
+  psadbw                m1, [srcq]
+  psadbw                m2, [srcq+16]
+  psadbw                m3, [srcq+32]
+  psadbw                m4, [srcq+48]
+
+  paddd                 m1, m2
+  paddd                 m3, m4
+  paddd                 m0, m1
+  paddd                 m0, m3
+
+  movu                  m1, [refq+64]
+  movu                  m2, [refq+80]
+  movu                  m3, [refq+96]
+  movu                  m4, [refq+112]
+%if %2 == 1
+  pavgb                 m1, [second_predq+mmsize*4]
+  pavgb                 m2, [second_predq+mmsize*5]
+  pavgb                 m3, [second_predq+mmsize*6]
+  pavgb                 m4, [second_predq+mmsize*7]
+  lea         second_predq, [second_predq+mmsize*8]
+%endif
+  psadbw                m1, [srcq+64]
+  psadbw                m2, [srcq+80]
+  psadbw                m3, [srcq+96]
+  psadbw                m4, [srcq+112]
+
+  add                 refq, ref_strideq
+  add                 srcq, src_strideq
+
+  paddd                 m1, m2
+  paddd                 m3, m4
+  paddd                 m0, m1
+  paddd                 m0, m3
+
+  sub              n_rowsd, 1
+  jg .loop
+
+  movhlps               m1, m0
+  paddd                 m0, m1
+%if %2 == 2 ; we skipped rows, so now we need to double the sad
+  pslld                 m0, 1
+%endif
+  movd                 eax, m0
+  RET
+%endmacro
+
+INIT_XMM sse2
+SAD128XN 128     ; sad128x128_sse2
+SAD128XN 128, 1  ; sad128x128_avg_sse2
+SAD128XN 128, 2  ; sad128x128_skip_sse2
+SAD128XN 64      ; sad128x64_sse2
+SAD128XN 64, 1   ; sad128x64_avg_sse2
+SAD128XN 64, 2   ; sad128x64_skip_sse2
+
+
+; unsigned int aom_sad64x64_sse2(uint8_t *src, int src_stride,
+;                                uint8_t *ref, int ref_stride);
+%macro SAD64XN 1-2 0
+  SAD_FN 64, %1, 5, %2
+%if %2 == 2
+  mov              n_rowsd, %1/2
+%else
+  mov              n_rowsd, %1
+%endif
+  pxor                  m0, m0
+.loop:
+  movu                  m1, [refq]
+  movu                  m2, [refq+16]
+  movu                  m3, [refq+32]
+  movu                  m4, [refq+48]
+%if %2 == 1
+  pavgb                 m1, [second_predq+mmsize*0]
+  pavgb                 m2, [second_predq+mmsize*1]
+  pavgb                 m3, [second_predq+mmsize*2]
+  pavgb                 m4, [second_predq+mmsize*3]
+  lea         second_predq, [second_predq+mmsize*4]
+%endif
+  psadbw                m1, [srcq]
+  psadbw                m2, [srcq+16]
+  psadbw                m3, [srcq+32]
+  psadbw                m4, [srcq+48]
+  paddd                 m1, m2
+  paddd                 m3, m4
+  add                 refq, ref_strideq
+  paddd                 m0, m1
+  add                 srcq, src_strideq
+  paddd                 m0, m3
+  dec              n_rowsd
+  jg .loop
+
+  movhlps               m1, m0
+  paddd                 m0, m1
+%if %2 == 2 ; we skipped rows, so now we need to double the sad
+  pslld                 m0, 1
+%endif
+  movd                 eax, m0
+  RET
+%endmacro
+
+INIT_XMM sse2
+SAD64XN 128     ; sad64x128_sse2
+SAD64XN  64     ; sad64x64_sse2
+SAD64XN  32     ; sad64x32_sse2
+SAD64XN  16     ; sad64x16_sse2
+SAD64XN 128, 1  ; sad64x128_avg_sse2
+SAD64XN  64, 1  ; sad64x64_avg_sse2
+SAD64XN  32, 1  ; sad64x32_avg_sse2
+SAD64XN  16, 1  ; sad64x16_avg_sse2
+SAD64XN 128, 2  ; sad64x128_skip_sse2
+SAD64XN  64, 2  ; sad64x64_skip_sse2
+SAD64XN  32, 2  ; sad64x32_skip_sse2
+SAD64XN  16, 2  ; sad64x16_skip_sse2
+
+; unsigned int aom_sad32x32_sse2(uint8_t *src, int src_stride,
+;                                uint8_t *ref, int ref_stride);
+%macro SAD32XN 1-2 0
+  SAD_FN 32, %1, 5, %2
+%if %2 == 2
+  mov              n_rowsd, %1/4
+%else
+  mov              n_rowsd, %1/2
+%endif
+  pxor                  m0, m0
+.loop:
+  movu                  m1, [refq]
+  movu                  m2, [refq+16]
+  movu                  m3, [refq+ref_strideq]
+  movu                  m4, [refq+ref_strideq+16]
+%if %2 == 1
+  pavgb                 m1, [second_predq+mmsize*0]
+  pavgb                 m2, [second_predq+mmsize*1]
+  pavgb                 m3, [second_predq+mmsize*2]
+  pavgb                 m4, [second_predq+mmsize*3]
+  lea         second_predq, [second_predq+mmsize*4]
+%endif
+  psadbw                m1, [srcq]
+  psadbw                m2, [srcq+16]
+  psadbw                m3, [srcq+src_strideq]
+  psadbw                m4, [srcq+src_strideq+16]
+  paddd                 m1, m2
+  paddd                 m3, m4
+  lea                 refq, [refq+ref_strideq*2]
+  paddd                 m0, m1
+  lea                 srcq, [srcq+src_strideq*2]
+  paddd                 m0, m3
+  dec              n_rowsd
+  jg .loop
+
+  movhlps               m1, m0
+  paddd                 m0, m1
+%if %2 == 2 ; we skipped rows, so now we need to double the sad
+  pslld                 m0, 1
+%endif
+  movd                 eax, m0
+  RET
+%endmacro
+
+INIT_XMM sse2
+SAD32XN 64    ; sad32x64_sse2
+SAD32XN 32    ; sad32x32_sse2
+SAD32XN 16    ; sad32x16_sse2
+SAD32XN  8    ; sad_32x8_sse2
+SAD32XN 64, 1 ; sad32x64_avg_sse2
+SAD32XN 32, 1 ; sad32x32_avg_sse2
+SAD32XN 16, 1 ; sad32x16_avg_sse2
+SAD32XN  8, 1 ; sad_32x8_avg_sse2
+SAD32XN 64, 2 ; sad32x64_skip_sse2
+SAD32XN 32, 2 ; sad32x32_skip_sse2
+SAD32XN 16, 2 ; sad32x16_skip_sse2
+SAD32XN  8, 2 ; sad_32x8_skip_sse2
+
+; unsigned int aom_sad16x{8,16}_sse2(uint8_t *src, int src_stride,
+;                                    uint8_t *ref, int ref_stride);
+%macro SAD16XN 1-2 0
+  SAD_FN 16, %1, 7, %2
+%if %2 == 2
+  mov              n_rowsd, %1/8
+%else
+  mov              n_rowsd, %1/4
+%endif
+  pxor                  m0, m0
+
+.loop:
+  movu                  m1, [refq]
+  movu                  m2, [refq+ref_strideq]
+  movu                  m3, [refq+ref_strideq*2]
+  movu                  m4, [refq+ref_stride3q]
+%if %2 == 1
+  pavgb                 m1, [second_predq+mmsize*0]
+  pavgb                 m2, [second_predq+mmsize*1]
+  pavgb                 m3, [second_predq+mmsize*2]
+  pavgb                 m4, [second_predq+mmsize*3]
+  lea         second_predq, [second_predq+mmsize*4]
+%endif
+  psadbw                m1, [srcq]
+  psadbw                m2, [srcq+src_strideq]
+  psadbw                m3, [srcq+src_strideq*2]
+  psadbw                m4, [srcq+src_stride3q]
+  paddd                 m1, m2
+  paddd                 m3, m4
+  lea                 refq, [refq+ref_strideq*4]
+  paddd                 m0, m1
+  lea                 srcq, [srcq+src_strideq*4]
+  paddd                 m0, m3
+  dec              n_rowsd
+  jg .loop
+
+  movhlps               m1, m0
+  paddd                 m0, m1
+%if %2 == 2 ; we skipped rows, so now we need to double the sad
+  pslld                 m0, 1
+%endif
+  movd                 eax, m0
+  RET
+%endmacro
+
+INIT_XMM sse2
+SAD16XN 64    ; sad_16x64_sse2
+SAD16XN 32    ; sad16x32_sse2
+SAD16XN 16    ; sad16x16_sse2
+SAD16XN  8    ; sad16x8_sse2
+SAD16XN  4    ; sad_16x4_sse2
+SAD16XN 64, 1 ; sad_16x64_avg_sse2
+SAD16XN 32, 1 ; sad16x32_avg_sse2
+SAD16XN 16, 1 ; sad16x16_avg_sse2
+SAD16XN  8, 1 ; sad16x8_avg_sse2
+SAD16XN  4, 1 ; sad_16x4_avg_sse2
+SAD16XN 64, 2 ; sad_16x64_skip_sse2
+SAD16XN 32, 2 ; sad16x32_skip_sse2
+SAD16XN 16, 2 ; sad16x16_skip_sse2
+SAD16XN  8, 2 ; sad16x8_skip_sse2
+
+; unsigned int aom_sad8x{8,16}_sse2(uint8_t *src, int src_stride,
+;                                   uint8_t *ref, int ref_stride);
+%macro SAD8XN 1-2 0
+  SAD_FN 8, %1, 7, %2
+%if %2 == 2
+  mov              n_rowsd, %1/8
+%else
+  mov              n_rowsd, %1/4
+%endif
+  pxor                  m0, m0
+
+.loop:
+  movh                  m1, [refq]
+  movhps                m1, [refq+ref_strideq]
+  movh                  m2, [refq+ref_strideq*2]
+  movhps                m2, [refq+ref_stride3q]
+%if %2 == 1
+  pavgb                 m1, [second_predq+mmsize*0]
+  pavgb                 m2, [second_predq+mmsize*1]
+  lea         second_predq, [second_predq+mmsize*2]
+%endif
+  movh                  m3, [srcq]
+  movhps                m3, [srcq+src_strideq]
+  movh                  m4, [srcq+src_strideq*2]
+  movhps                m4, [srcq+src_stride3q]
+  psadbw                m1, m3
+  psadbw                m2, m4
+  lea                 refq, [refq+ref_strideq*4]
+  paddd                 m0, m1
+  lea                 srcq, [srcq+src_strideq*4]
+  paddd                 m0, m2
+  dec              n_rowsd
+  jg .loop
+
+  movhlps               m1, m0
+  paddd                 m0, m1
+%if %2 == 2 ; we skipped rows, so now we need to double the sad
+  pslld                 m0, 1
+%endif
+  movd                 eax, m0
+  RET
+%endmacro
+
+INIT_XMM sse2
+SAD8XN 32    ; sad_8x32_sse2
+SAD8XN 16    ; sad8x16_sse2
+SAD8XN  8    ; sad8x8_sse2
+SAD8XN  4    ; sad8x4_sse2
+SAD8XN 32, 1 ; sad_8x32_avg_sse2
+SAD8XN 16, 1 ; sad8x16_avg_sse2
+SAD8XN  8, 1 ; sad8x8_avg_sse2
+SAD8XN  4, 1 ; sad8x4_avg_sse2
+SAD8XN 32, 2 ; sad_8x32_skip_sse2
+SAD8XN 16, 2 ; sad8x16_skip_sse2
+SAD8XN  8, 2 ; sad8x8_skip_sse2
+
+; unsigned int aom_sad4x{4, 8}_sse2(uint8_t *src, int src_stride,
+;                                   uint8_t *ref, int ref_stride);
+%macro SAD4XN 1-2 0
+  SAD_FN 4, %1, 7, %2
+%if %2 == 2
+  mov              n_rowsd, %1/8
+%else
+  mov              n_rowsd, %1/4
+%endif
+  pxor                  m0, m0
+
+.loop:
+  movd                  m1, [refq]
+  movd                  m2, [refq+ref_strideq]
+  movd                  m3, [refq+ref_strideq*2]
+  movd                  m4, [refq+ref_stride3q]
+  punpckldq             m1, m2
+  punpckldq             m3, m4
+  movlhps               m1, m3
+%if %2 == 1
+  pavgb                 m1, [second_predq+mmsize*0]
+  lea         second_predq, [second_predq+mmsize*1]
+%endif
+  movd                  m2, [srcq]
+  movd                  m5, [srcq+src_strideq]
+  movd                  m4, [srcq+src_strideq*2]
+  movd                  m3, [srcq+src_stride3q]
+  punpckldq             m2, m5
+  punpckldq             m4, m3
+  movlhps               m2, m4
+  psadbw                m1, m2
+  lea                 refq, [refq+ref_strideq*4]
+  paddd                 m0, m1
+  lea                 srcq, [srcq+src_strideq*4]
+  dec              n_rowsd
+  jg .loop
+
+  movhlps               m1, m0
+  paddd                 m0, m1
+%if %2 == 2 ; we skipped rows, so now we need to double the sad
+  pslld                 m0, 1
+%endif
+  movd                 eax, m0
+  RET
+%endmacro
+
+INIT_XMM sse2
+SAD4XN 16 ; sad_4x16_sse2
+SAD4XN  8 ; sad4x8_sse
+SAD4XN  4 ; sad4x4_sse
+SAD4XN 16, 1 ; sad_4x16_avg_sse2
+SAD4XN  8, 1 ; sad4x8_avg_sse
+SAD4XN  4, 1 ; sad4x4_avg_sse
+SAD4XN 16, 2 ; sad_4x16_skip_sse2
+SAD4XN  8, 2 ; sad4x8_skip_sse
diff --git a/third_party/aom/aom_dsp/x86/sse_avx2.c b/third_party/aom/aom_dsp/x86/sse_avx2.c
new file mode 100644
index 0000000000..c5a5f5c234
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/sse_avx2.c
@@ -0,0 +1,389 @@
+/*
+ * Copyright (c) 2018, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <smmintrin.h>
+#include <immintrin.h>
+
+#include "config/aom_config.h"
+#include "config/aom_dsp_rtcd.h"
+
+#include "aom_ports/mem.h"
+#include "aom_dsp/x86/synonyms.h"
+#include "aom_dsp/x86/synonyms_avx2.h"
+
+static INLINE void sse_w32_avx2(__m256i *sum, const uint8_t *a,
+                                const uint8_t *b) {
+  const __m256i v_a0 = yy_loadu_256(a);
+  const __m256i v_b0 = yy_loadu_256(b);
+  const __m256i zero = _mm256_setzero_si256();
+  const __m256i v_a00_w = _mm256_unpacklo_epi8(v_a0, zero);
+  const __m256i v_a01_w = _mm256_unpackhi_epi8(v_a0, zero);
+  const __m256i v_b00_w = _mm256_unpacklo_epi8(v_b0, zero);
+  const __m256i v_b01_w = _mm256_unpackhi_epi8(v_b0, zero);
+  const __m256i v_d00_w = _mm256_sub_epi16(v_a00_w, v_b00_w);
+  const __m256i v_d01_w = _mm256_sub_epi16(v_a01_w, v_b01_w);
+  *sum = _mm256_add_epi32(*sum, _mm256_madd_epi16(v_d00_w, v_d00_w));
+  *sum = _mm256_add_epi32(*sum, _mm256_madd_epi16(v_d01_w, v_d01_w));
+}
+
+static INLINE int64_t summary_all_avx2(const __m256i *sum_all) {
+  int64_t sum;
+  __m256i zero = _mm256_setzero_si256();
+  const __m256i sum0_4x64 = _mm256_unpacklo_epi32(*sum_all, zero);
+  const __m256i sum1_4x64 = _mm256_unpackhi_epi32(*sum_all, zero);
+  const __m256i sum_4x64 = _mm256_add_epi64(sum0_4x64, sum1_4x64);
+  const __m128i sum_2x64 = _mm_add_epi64(_mm256_castsi256_si128(sum_4x64),
+                                         _mm256_extracti128_si256(sum_4x64, 1));
+  const __m128i sum_1x64 = _mm_add_epi64(sum_2x64, _mm_srli_si128(sum_2x64, 8));
+  xx_storel_64(&sum, sum_1x64);
+  return sum;
+}
+
+#if CONFIG_AV1_HIGHBITDEPTH
+static INLINE void summary_32_avx2(const __m256i *sum32, __m256i *sum) {
+  const __m256i sum0_4x64 =
+      _mm256_cvtepu32_epi64(_mm256_castsi256_si128(*sum32));
+  const __m256i sum1_4x64 =
+      _mm256_cvtepu32_epi64(_mm256_extracti128_si256(*sum32, 1));
+  const __m256i sum_4x64 = _mm256_add_epi64(sum0_4x64, sum1_4x64);
+  *sum = _mm256_add_epi64(*sum, sum_4x64);
+}
+
+static INLINE int64_t summary_4x64_avx2(const __m256i sum_4x64) {
+  int64_t sum;
+  const __m128i sum_2x64 = _mm_add_epi64(_mm256_castsi256_si128(sum_4x64),
+                                         _mm256_extracti128_si256(sum_4x64, 1));
+  const __m128i sum_1x64 = _mm_add_epi64(sum_2x64, _mm_srli_si128(sum_2x64, 8));
+
+  xx_storel_64(&sum, sum_1x64);
+  return sum;
+}
+#endif
+
+static INLINE void sse_w4x4_avx2(const uint8_t *a, int a_stride,
+                                 const uint8_t *b, int b_stride, __m256i *sum) {
+  const __m128i v_a0 = xx_loadl_32(a);
+  const __m128i v_a1 = xx_loadl_32(a + a_stride);
+  const __m128i v_a2 = xx_loadl_32(a + a_stride * 2);
+  const __m128i v_a3 = xx_loadl_32(a + a_stride * 3);
+  const __m128i v_b0 = xx_loadl_32(b);
+  const __m128i v_b1 = xx_loadl_32(b + b_stride);
+  const __m128i v_b2 = xx_loadl_32(b + b_stride * 2);
+  const __m128i v_b3 = xx_loadl_32(b + b_stride * 3);
+  const __m128i v_a0123 = _mm_unpacklo_epi64(_mm_unpacklo_epi32(v_a0, v_a1),
+                                             _mm_unpacklo_epi32(v_a2, v_a3));
+  const __m128i v_b0123 = _mm_unpacklo_epi64(_mm_unpacklo_epi32(v_b0, v_b1),
+                                             _mm_unpacklo_epi32(v_b2, v_b3));
+  const __m256i v_a_w = _mm256_cvtepu8_epi16(v_a0123);
+  const __m256i v_b_w = _mm256_cvtepu8_epi16(v_b0123);
+  const __m256i v_d_w = _mm256_sub_epi16(v_a_w, v_b_w);
+  *sum = _mm256_add_epi32(*sum, _mm256_madd_epi16(v_d_w, v_d_w));
+}
+
+static INLINE void sse_w8x2_avx2(const uint8_t *a, int a_stride,
+                                 const uint8_t *b, int b_stride, __m256i *sum) {
+  const __m128i v_a0 = xx_loadl_64(a);
+  const __m128i v_a1 = xx_loadl_64(a + a_stride);
+  const __m128i v_b0 = xx_loadl_64(b);
+  const __m128i v_b1 = xx_loadl_64(b + b_stride);
+  const __m256i v_a_w = _mm256_cvtepu8_epi16(_mm_unpacklo_epi64(v_a0, v_a1));
+  const __m256i v_b_w = _mm256_cvtepu8_epi16(_mm_unpacklo_epi64(v_b0, v_b1));
+  const __m256i v_d_w = _mm256_sub_epi16(v_a_w, v_b_w);
+  *sum = _mm256_add_epi32(*sum, _mm256_madd_epi16(v_d_w, v_d_w));
+}
+
+int64_t aom_sse_avx2(const uint8_t *a, int a_stride, const uint8_t *b,
+                     int b_stride, int width, int height) {
+  int32_t y = 0;
+  int64_t sse = 0;
+  __m256i sum = _mm256_setzero_si256();
+  __m256i zero = _mm256_setzero_si256();
+  switch (width) {
+    case 4:
+      do {
+        sse_w4x4_avx2(a, a_stride, b, b_stride, &sum);
+        a += a_stride << 2;
+        b += b_stride << 2;
+        y += 4;
+      } while (y < height);
+      sse = summary_all_avx2(&sum);
+      break;
+    case 8:
+      do {
+        sse_w8x2_avx2(a, a_stride, b, b_stride, &sum);
+        a += a_stride << 1;
+        b += b_stride << 1;
+        y += 2;
+      } while (y < height);
+      sse = summary_all_avx2(&sum);
+      break;
+    case 16:
+      do {
+        const __m128i v_a0 = xx_loadu_128(a);
+        const __m128i v_a1 = xx_loadu_128(a + a_stride);
+        const __m128i v_b0 = xx_loadu_128(b);
+        const __m128i v_b1 = xx_loadu_128(b + b_stride);
+        const __m256i v_a =
+            _mm256_insertf128_si256(_mm256_castsi128_si256(v_a0), v_a1, 0x01);
+        const __m256i v_b =
+            _mm256_insertf128_si256(_mm256_castsi128_si256(v_b0), v_b1, 0x01);
+        const __m256i v_al = _mm256_unpacklo_epi8(v_a, zero);
+        const __m256i v_au = _mm256_unpackhi_epi8(v_a, zero);
+        const __m256i v_bl = _mm256_unpacklo_epi8(v_b, zero);
+        const __m256i v_bu = _mm256_unpackhi_epi8(v_b, zero);
+        const __m256i v_asub = _mm256_sub_epi16(v_al, v_bl);
+        const __m256i v_bsub = _mm256_sub_epi16(v_au, v_bu);
+        const __m256i temp =
+            _mm256_add_epi32(_mm256_madd_epi16(v_asub, v_asub),
+                             _mm256_madd_epi16(v_bsub, v_bsub));
+        sum = _mm256_add_epi32(sum, temp);
+        a += a_stride << 1;
+        b += b_stride << 1;
+        y += 2;
+      } while (y < height);
+      sse = summary_all_avx2(&sum);
+      break;
+    case 32:
+      do {
+        sse_w32_avx2(&sum, a, b);
+        a += a_stride;
+        b += b_stride;
+        y += 1;
+      } while (y < height);
+      sse = summary_all_avx2(&sum);
+      break;
+    case 64:
+      do {
+        sse_w32_avx2(&sum, a, b);
+        sse_w32_avx2(&sum, a + 32, b + 32);
+        a += a_stride;
+        b += b_stride;
+        y += 1;
+      } while (y < height);
+      sse = summary_all_avx2(&sum);
+      break;
+    case 128:
+      do {
+        sse_w32_avx2(&sum, a, b);
+        sse_w32_avx2(&sum, a + 32, b + 32);
+        sse_w32_avx2(&sum, a + 64, b + 64);
+        sse_w32_avx2(&sum, a + 96, b + 96);
+        a += a_stride;
+        b += b_stride;
+        y += 1;
+      } while (y < height);
+      sse = summary_all_avx2(&sum);
+      break;
+    default:
+      if ((width & 0x07) == 0) {
+        do {
+          int i = 0;
+          do {
+            sse_w8x2_avx2(a + i, a_stride, b + i, b_stride, &sum);
+            i += 8;
+          } while (i < width);
+          a += a_stride << 1;
+          b += b_stride << 1;
+          y += 2;
+        } while (y < height);
+      } else {
+        do {
+          int i = 0;
+          do {
+            sse_w8x2_avx2(a + i, a_stride, b + i, b_stride, &sum);
+            const uint8_t *a2 = a + i + (a_stride << 1);
+            const uint8_t *b2 = b + i + (b_stride << 1);
+            sse_w8x2_avx2(a2, a_stride, b2, b_stride, &sum);
+            i += 8;
+          } while (i + 4 < width);
+          sse_w4x4_avx2(a + i, a_stride, b + i, b_stride, &sum);
+          a += a_stride << 2;
+          b += b_stride << 2;
+          y += 4;
+        } while (y < height);
+      }
+      sse = summary_all_avx2(&sum);
+      break;
+  }
+
+  return sse;
+}
+
+#if CONFIG_AV1_HIGHBITDEPTH
+static INLINE void highbd_sse_w16_avx2(__m256i *sum, const uint16_t *a,
+                                       const uint16_t *b) {
+  const __m256i v_a_w = yy_loadu_256(a);
+  const __m256i v_b_w = yy_loadu_256(b);
+  const __m256i v_d_w = _mm256_sub_epi16(v_a_w, v_b_w);
+  *sum = _mm256_add_epi32(*sum, _mm256_madd_epi16(v_d_w, v_d_w));
+}
+
+static INLINE void highbd_sse_w4x4_avx2(__m256i *sum, const uint16_t *a,
+                                        int a_stride, const uint16_t *b,
+                                        int b_stride) {
+  const __m128i v_a0 = xx_loadl_64(a);
+  const __m128i v_a1 = xx_loadl_64(a + a_stride);
+  const __m128i v_a2 = xx_loadl_64(a + a_stride * 2);
+  const __m128i v_a3 = xx_loadl_64(a + a_stride * 3);
+  const __m128i v_b0 = xx_loadl_64(b);
+  const __m128i v_b1 = xx_loadl_64(b + b_stride);
+  const __m128i v_b2 = xx_loadl_64(b + b_stride * 2);
+  const __m128i v_b3 = xx_loadl_64(b + b_stride * 3);
+  const __m256i v_a_w = yy_set_m128i(_mm_unpacklo_epi64(v_a0, v_a1),
+                                     _mm_unpacklo_epi64(v_a2, v_a3));
+  const __m256i v_b_w = yy_set_m128i(_mm_unpacklo_epi64(v_b0, v_b1),
+                                     _mm_unpacklo_epi64(v_b2, v_b3));
+  const __m256i v_d_w = _mm256_sub_epi16(v_a_w, v_b_w);
+  *sum = _mm256_add_epi32(*sum, _mm256_madd_epi16(v_d_w, v_d_w));
+}
+
+static INLINE void highbd_sse_w8x2_avx2(__m256i *sum, const uint16_t *a,
+                                        int a_stride, const uint16_t *b,
+                                        int b_stride) {
+  const __m256i v_a_w = yy_loadu2_128(a + a_stride, a);
+  const __m256i v_b_w = yy_loadu2_128(b + b_stride, b);
+  const __m256i v_d_w = _mm256_sub_epi16(v_a_w, v_b_w);
+  *sum = _mm256_add_epi32(*sum, _mm256_madd_epi16(v_d_w, v_d_w));
+}
+
+int64_t aom_highbd_sse_avx2(const uint8_t *a8, int a_stride, const uint8_t *b8,
+                            int b_stride, int width, int height) {
+  int32_t y = 0;
+  int64_t sse = 0;
+  uint16_t *a = CONVERT_TO_SHORTPTR(a8);
+  uint16_t *b = CONVERT_TO_SHORTPTR(b8);
+  __m256i sum = _mm256_setzero_si256();
+  switch (width) {
+    case 4:
+      do {
+        highbd_sse_w4x4_avx2(&sum, a, a_stride, b, b_stride);
+        a += a_stride << 2;
+        b += b_stride << 2;
+        y += 4;
+      } while (y < height);
+      sse = summary_all_avx2(&sum);
+      break;
+    case 8:
+      do {
+        highbd_sse_w8x2_avx2(&sum, a, a_stride, b, b_stride);
+        a += a_stride << 1;
+        b += b_stride << 1;
+        y += 2;
+      } while (y < height);
+      sse = summary_all_avx2(&sum);
+      break;
+    case 16:
+      do {
+        highbd_sse_w16_avx2(&sum, a, b);
+        a += a_stride;
+        b += b_stride;
+        y += 1;
+      } while (y < height);
+      sse = summary_all_avx2(&sum);
+      break;
+    case 32:
+      do {
+        int l = 0;
+        __m256i sum32 = _mm256_setzero_si256();
+        do {
+          highbd_sse_w16_avx2(&sum32, a, b);
+          highbd_sse_w16_avx2(&sum32, a + 16, b + 16);
+          a += a_stride;
+          b += b_stride;
+          l += 1;
+        } while (l < 64 && l < (height - y));
+        summary_32_avx2(&sum32, &sum);
+        y += 64;
+      } while (y < height);
+      sse = summary_4x64_avx2(sum);
+      break;
+    case 64:
+      do {
+        int l = 0;
+        __m256i sum32 = _mm256_setzero_si256();
+        do {
+          highbd_sse_w16_avx2(&sum32, a, b);
+          highbd_sse_w16_avx2(&sum32, a + 16 * 1, b + 16 * 1);
+          highbd_sse_w16_avx2(&sum32, a + 16 * 2, b + 16 * 2);
+          highbd_sse_w16_avx2(&sum32, a + 16 * 3, b + 16 * 3);
+          a += a_stride;
+          b += b_stride;
+          l += 1;
+        } while (l < 32 && l < (height - y));
+        summary_32_avx2(&sum32, &sum);
+        y += 32;
+      } while (y < height);
+      sse = summary_4x64_avx2(sum);
+      break;
+    case 128:
+      do {
+        int l = 0;
+        __m256i sum32 = _mm256_setzero_si256();
+        do {
+          highbd_sse_w16_avx2(&sum32, a, b);
+          highbd_sse_w16_avx2(&sum32, a + 16 * 1, b + 16 * 1);
+          highbd_sse_w16_avx2(&sum32, a + 16 * 2, b + 16 * 2);
+          highbd_sse_w16_avx2(&sum32, a + 16 * 3, b + 16 * 3);
+          highbd_sse_w16_avx2(&sum32, a + 16 * 4, b + 16 * 4);
+          highbd_sse_w16_avx2(&sum32, a + 16 * 5, b + 16 * 5);
+          highbd_sse_w16_avx2(&sum32, a + 16 * 6, b + 16 * 6);
+          highbd_sse_w16_avx2(&sum32, a + 16 * 7, b + 16 * 7);
+          a += a_stride;
+          b += b_stride;
+          l += 1;
+        } while (l < 16 && l < (height - y));
+        summary_32_avx2(&sum32, &sum);
+        y += 16;
+      } while (y < height);
+      sse = summary_4x64_avx2(sum);
+      break;
+    default:
+      if (width & 0x7) {
+        do {
+          int i = 0;
+          __m256i sum32 = _mm256_setzero_si256();
+          do {
+            highbd_sse_w8x2_avx2(&sum32, a + i, a_stride, b + i, b_stride);
+            const uint16_t *a2 = a + i + (a_stride << 1);
+            const uint16_t *b2 = b + i + (b_stride << 1);
+            highbd_sse_w8x2_avx2(&sum32, a2, a_stride, b2, b_stride);
+            i += 8;
+          } while (i + 4 < width);
+          highbd_sse_w4x4_avx2(&sum32, a + i, a_stride, b + i, b_stride);
+          summary_32_avx2(&sum32, &sum);
+          a += a_stride << 2;
+          b += b_stride << 2;
+          y += 4;
+        } while (y < height);
+      } else {
+        do {
+          int l = 0;
+          __m256i sum32 = _mm256_setzero_si256();
+          do {
+            int i = 0;
+            do {
+              highbd_sse_w8x2_avx2(&sum32, a + i, a_stride, b + i, b_stride);
+              i += 8;
+            } while (i < width);
+            a += a_stride << 1;
+            b += b_stride << 1;
+            l += 2;
+          } while (l < 8 && l < (height - y));
+          summary_32_avx2(&sum32, &sum);
+          y += 8;
+        } while (y < height);
+      }
+      sse = summary_4x64_avx2(sum);
+      break;
+  }
+  return sse;
+}
+#endif  // CONFIG_AV1_HIGHBITDEPTH
diff --git a/third_party/aom/aom_dsp/x86/sse_sse4.c b/third_party/aom/aom_dsp/x86/sse_sse4.c
new file mode 100644
index 0000000000..7e74554d75
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/sse_sse4.c
@@ -0,0 +1,355 @@
+/*
+ * 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 <smmintrin.h>
+
+#include "config/aom_config.h"
+#include "config/aom_dsp_rtcd.h"
+
+#include "aom_ports/mem.h"
+#include "aom/aom_integer.h"
+#include "aom_dsp/x86/synonyms.h"
+
+static INLINE int64_t summary_all_sse4(const __m128i *sum_all) {
+  int64_t sum;
+  const __m128i sum0 = _mm_cvtepu32_epi64(*sum_all);
+  const __m128i sum1 = _mm_cvtepu32_epi64(_mm_srli_si128(*sum_all, 8));
+  const __m128i sum_2x64 = _mm_add_epi64(sum0, sum1);
+  const __m128i sum_1x64 = _mm_add_epi64(sum_2x64, _mm_srli_si128(sum_2x64, 8));
+  xx_storel_64(&sum, sum_1x64);
+  return sum;
+}
+
+#if CONFIG_AV1_HIGHBITDEPTH
+static INLINE void summary_32_sse4(const __m128i *sum32, __m128i *sum64) {
+  const __m128i sum0 = _mm_cvtepu32_epi64(*sum32);
+  const __m128i sum1 = _mm_cvtepu32_epi64(_mm_srli_si128(*sum32, 8));
+  *sum64 = _mm_add_epi64(sum0, *sum64);
+  *sum64 = _mm_add_epi64(sum1, *sum64);
+}
+#endif
+
+static INLINE void sse_w16_sse4_1(__m128i *sum, const uint8_t *a,
+                                  const uint8_t *b) {
+  const __m128i v_a0 = xx_loadu_128(a);
+  const __m128i v_b0 = xx_loadu_128(b);
+  const __m128i v_a00_w = _mm_cvtepu8_epi16(v_a0);
+  const __m128i v_a01_w = _mm_cvtepu8_epi16(_mm_srli_si128(v_a0, 8));
+  const __m128i v_b00_w = _mm_cvtepu8_epi16(v_b0);
+  const __m128i v_b01_w = _mm_cvtepu8_epi16(_mm_srli_si128(v_b0, 8));
+  const __m128i v_d00_w = _mm_sub_epi16(v_a00_w, v_b00_w);
+  const __m128i v_d01_w = _mm_sub_epi16(v_a01_w, v_b01_w);
+  *sum = _mm_add_epi32(*sum, _mm_madd_epi16(v_d00_w, v_d00_w));
+  *sum = _mm_add_epi32(*sum, _mm_madd_epi16(v_d01_w, v_d01_w));
+}
+
+static INLINE void sse4x2_sse4_1(const uint8_t *a, int a_stride,
+                                 const uint8_t *b, int b_stride, __m128i *sum) {
+  const __m128i v_a0 = xx_loadl_32(a);
+  const __m128i v_a1 = xx_loadl_32(a + a_stride);
+  const __m128i v_b0 = xx_loadl_32(b);
+  const __m128i v_b1 = xx_loadl_32(b + b_stride);
+  const __m128i v_a_w = _mm_cvtepu8_epi16(_mm_unpacklo_epi32(v_a0, v_a1));
+  const __m128i v_b_w = _mm_cvtepu8_epi16(_mm_unpacklo_epi32(v_b0, v_b1));
+  const __m128i v_d_w = _mm_sub_epi16(v_a_w, v_b_w);
+  *sum = _mm_add_epi32(*sum, _mm_madd_epi16(v_d_w, v_d_w));
+}
+
+static INLINE void sse8_sse4_1(const uint8_t *a, const uint8_t *b,
+                               __m128i *sum) {
+  const __m128i v_a0 = xx_loadl_64(a);
+  const __m128i v_b0 = xx_loadl_64(b);
+  const __m128i v_a_w = _mm_cvtepu8_epi16(v_a0);
+  const __m128i v_b_w = _mm_cvtepu8_epi16(v_b0);
+  const __m128i v_d_w = _mm_sub_epi16(v_a_w, v_b_w);
+  *sum = _mm_add_epi32(*sum, _mm_madd_epi16(v_d_w, v_d_w));
+}
+
+int64_t aom_sse_sse4_1(const uint8_t *a, int a_stride, const uint8_t *b,
+                       int b_stride, int width, int height) {
+  int y = 0;
+  int64_t sse = 0;
+  __m128i sum = _mm_setzero_si128();
+  switch (width) {
+    case 4:
+      do {
+        sse4x2_sse4_1(a, a_stride, b, b_stride, &sum);
+        a += a_stride << 1;
+        b += b_stride << 1;
+        y += 2;
+      } while (y < height);
+      sse = summary_all_sse4(&sum);
+      break;
+    case 8:
+      do {
+        sse8_sse4_1(a, b, &sum);
+        a += a_stride;
+        b += b_stride;
+        y += 1;
+      } while (y < height);
+      sse = summary_all_sse4(&sum);
+      break;
+    case 16:
+      do {
+        sse_w16_sse4_1(&sum, a, b);
+        a += a_stride;
+        b += b_stride;
+        y += 1;
+      } while (y < height);
+      sse = summary_all_sse4(&sum);
+      break;
+    case 32:
+      do {
+        sse_w16_sse4_1(&sum, a, b);
+        sse_w16_sse4_1(&sum, a + 16, b + 16);
+        a += a_stride;
+        b += b_stride;
+        y += 1;
+      } while (y < height);
+      sse = summary_all_sse4(&sum);
+      break;
+    case 64:
+      do {
+        sse_w16_sse4_1(&sum, a, b);
+        sse_w16_sse4_1(&sum, a + 16 * 1, b + 16 * 1);
+        sse_w16_sse4_1(&sum, a + 16 * 2, b + 16 * 2);
+        sse_w16_sse4_1(&sum, a + 16 * 3, b + 16 * 3);
+        a += a_stride;
+        b += b_stride;
+        y += 1;
+      } while (y < height);
+      sse = summary_all_sse4(&sum);
+      break;
+    case 128:
+      do {
+        sse_w16_sse4_1(&sum, a, b);
+        sse_w16_sse4_1(&sum, a + 16 * 1, b + 16 * 1);
+        sse_w16_sse4_1(&sum, a + 16 * 2, b + 16 * 2);
+        sse_w16_sse4_1(&sum, a + 16 * 3, b + 16 * 3);
+        sse_w16_sse4_1(&sum, a + 16 * 4, b + 16 * 4);
+        sse_w16_sse4_1(&sum, a + 16 * 5, b + 16 * 5);
+        sse_w16_sse4_1(&sum, a + 16 * 6, b + 16 * 6);
+        sse_w16_sse4_1(&sum, a + 16 * 7, b + 16 * 7);
+        a += a_stride;
+        b += b_stride;
+        y += 1;
+      } while (y < height);
+      sse = summary_all_sse4(&sum);
+      break;
+    default:
+      if (width & 0x07) {
+        do {
+          int i = 0;
+          do {
+            sse8_sse4_1(a + i, b + i, &sum);
+            sse8_sse4_1(a + i + a_stride, b + i + b_stride, &sum);
+            i += 8;
+          } while (i + 4 < width);
+          sse4x2_sse4_1(a + i, a_stride, b + i, b_stride, &sum);
+          a += (a_stride << 1);
+          b += (b_stride << 1);
+          y += 2;
+        } while (y < height);
+      } else {
+        do {
+          int i = 0;
+          do {
+            sse8_sse4_1(a + i, b + i, &sum);
+            i += 8;
+          } while (i < width);
+          a += a_stride;
+          b += b_stride;
+          y += 1;
+        } while (y < height);
+      }
+      sse = summary_all_sse4(&sum);
+      break;
+  }
+
+  return sse;
+}
+
+#if CONFIG_AV1_HIGHBITDEPTH
+static INLINE void highbd_sse_w4x2_sse4_1(__m128i *sum, const uint16_t *a,
+                                          int a_stride, const uint16_t *b,
+                                          int b_stride) {
+  const __m128i v_a0 = xx_loadl_64(a);
+  const __m128i v_a1 = xx_loadl_64(a + a_stride);
+  const __m128i v_b0 = xx_loadl_64(b);
+  const __m128i v_b1 = xx_loadl_64(b + b_stride);
+  const __m128i v_a_w = _mm_unpacklo_epi64(v_a0, v_a1);
+  const __m128i v_b_w = _mm_unpacklo_epi64(v_b0, v_b1);
+  const __m128i v_d_w = _mm_sub_epi16(v_a_w, v_b_w);
+  *sum = _mm_add_epi32(*sum, _mm_madd_epi16(v_d_w, v_d_w));
+}
+
+static INLINE void highbd_sse_w8_sse4_1(__m128i *sum, const uint16_t *a,
+                                        const uint16_t *b) {
+  const __m128i v_a_w = xx_loadu_128(a);
+  const __m128i v_b_w = xx_loadu_128(b);
+  const __m128i v_d_w = _mm_sub_epi16(v_a_w, v_b_w);
+  *sum = _mm_add_epi32(*sum, _mm_madd_epi16(v_d_w, v_d_w));
+}
+
+int64_t aom_highbd_sse_sse4_1(const uint8_t *a8, int a_stride,
+                              const uint8_t *b8, int b_stride, int width,
+                              int height) {
+  int32_t y = 0;
+  int64_t sse = 0;
+  uint16_t *a = CONVERT_TO_SHORTPTR(a8);
+  uint16_t *b = CONVERT_TO_SHORTPTR(b8);
+  __m128i sum = _mm_setzero_si128();
+  switch (width) {
+    case 4:
+      do {
+        highbd_sse_w4x2_sse4_1(&sum, a, a_stride, b, b_stride);
+        a += a_stride << 1;
+        b += b_stride << 1;
+        y += 2;
+      } while (y < height);
+      sse = summary_all_sse4(&sum);
+      break;
+    case 8:
+      do {
+        highbd_sse_w8_sse4_1(&sum, a, b);
+        a += a_stride;
+        b += b_stride;
+        y += 1;
+      } while (y < height);
+      sse = summary_all_sse4(&sum);
+      break;
+    case 16:
+      do {
+        int l = 0;
+        __m128i sum32 = _mm_setzero_si128();
+        do {
+          highbd_sse_w8_sse4_1(&sum32, a, b);
+          highbd_sse_w8_sse4_1(&sum32, a + 8, b + 8);
+          a += a_stride;
+          b += b_stride;
+          l += 1;
+        } while (l < 64 && l < (height - y));
+        summary_32_sse4(&sum32, &sum);
+        y += 64;
+      } while (y < height);
+      xx_storel_64(&sse, _mm_add_epi64(sum, _mm_srli_si128(sum, 8)));
+      break;
+    case 32:
+      do {
+        int l = 0;
+        __m128i sum32 = _mm_setzero_si128();
+        do {
+          highbd_sse_w8_sse4_1(&sum32, a, b);
+          highbd_sse_w8_sse4_1(&sum32, a + 8 * 1, b + 8 * 1);
+          highbd_sse_w8_sse4_1(&sum32, a + 8 * 2, b + 8 * 2);
+          highbd_sse_w8_sse4_1(&sum32, a + 8 * 3, b + 8 * 3);
+          a += a_stride;
+          b += b_stride;
+          l += 1;
+        } while (l < 32 && l < (height - y));
+        summary_32_sse4(&sum32, &sum);
+        y += 32;
+      } while (y < height);
+      xx_storel_64(&sse, _mm_add_epi64(sum, _mm_srli_si128(sum, 8)));
+      break;
+    case 64:
+      do {
+        int l = 0;
+        __m128i sum32 = _mm_setzero_si128();
+        do {
+          highbd_sse_w8_sse4_1(&sum32, a, b);
+          highbd_sse_w8_sse4_1(&sum32, a + 8 * 1, b + 8 * 1);
+          highbd_sse_w8_sse4_1(&sum32, a + 8 * 2, b + 8 * 2);
+          highbd_sse_w8_sse4_1(&sum32, a + 8 * 3, b + 8 * 3);
+          highbd_sse_w8_sse4_1(&sum32, a + 8 * 4, b + 8 * 4);
+          highbd_sse_w8_sse4_1(&sum32, a + 8 * 5, b + 8 * 5);
+          highbd_sse_w8_sse4_1(&sum32, a + 8 * 6, b + 8 * 6);
+          highbd_sse_w8_sse4_1(&sum32, a + 8 * 7, b + 8 * 7);
+          a += a_stride;
+          b += b_stride;
+          l += 1;
+        } while (l < 16 && l < (height - y));
+        summary_32_sse4(&sum32, &sum);
+        y += 16;
+      } while (y < height);
+      xx_storel_64(&sse, _mm_add_epi64(sum, _mm_srli_si128(sum, 8)));
+      break;
+    case 128:
+      do {
+        int l = 0;
+        __m128i sum32 = _mm_setzero_si128();
+        do {
+          highbd_sse_w8_sse4_1(&sum32, a, b);
+          highbd_sse_w8_sse4_1(&sum32, a + 8 * 1, b + 8 * 1);
+          highbd_sse_w8_sse4_1(&sum32, a + 8 * 2, b + 8 * 2);
+          highbd_sse_w8_sse4_1(&sum32, a + 8 * 3, b + 8 * 3);
+          highbd_sse_w8_sse4_1(&sum32, a + 8 * 4, b + 8 * 4);
+          highbd_sse_w8_sse4_1(&sum32, a + 8 * 5, b + 8 * 5);
+          highbd_sse_w8_sse4_1(&sum32, a + 8 * 6, b + 8 * 6);
+          highbd_sse_w8_sse4_1(&sum32, a + 8 * 7, b + 8 * 7);
+          highbd_sse_w8_sse4_1(&sum32, a + 8 * 8, b + 8 * 8);
+          highbd_sse_w8_sse4_1(&sum32, a + 8 * 9, b + 8 * 9);
+          highbd_sse_w8_sse4_1(&sum32, a + 8 * 10, b + 8 * 10);
+          highbd_sse_w8_sse4_1(&sum32, a + 8 * 11, b + 8 * 11);
+          highbd_sse_w8_sse4_1(&sum32, a + 8 * 12, b + 8 * 12);
+          highbd_sse_w8_sse4_1(&sum32, a + 8 * 13, b + 8 * 13);
+          highbd_sse_w8_sse4_1(&sum32, a + 8 * 14, b + 8 * 14);
+          highbd_sse_w8_sse4_1(&sum32, a + 8 * 15, b + 8 * 15);
+          a += a_stride;
+          b += b_stride;
+          l += 1;
+        } while (l < 8 && l < (height - y));
+        summary_32_sse4(&sum32, &sum);
+        y += 8;
+      } while (y < height);
+      xx_storel_64(&sse, _mm_add_epi64(sum, _mm_srli_si128(sum, 8)));
+      break;
+    default:
+      if (width & 0x7) {
+        do {
+          __m128i sum32 = _mm_setzero_si128();
+          int i = 0;
+          do {
+            highbd_sse_w8_sse4_1(&sum32, a + i, b + i);
+            highbd_sse_w8_sse4_1(&sum32, a + i + a_stride, b + i + b_stride);
+            i += 8;
+          } while (i + 4 < width);
+          highbd_sse_w4x2_sse4_1(&sum32, a + i, a_stride, b + i, b_stride);
+          a += (a_stride << 1);
+          b += (b_stride << 1);
+          y += 2;
+          summary_32_sse4(&sum32, &sum);
+        } while (y < height);
+      } else {
+        do {
+          int l = 0;
+          __m128i sum32 = _mm_setzero_si128();
+          do {
+            int i = 0;
+            do {
+              highbd_sse_w8_sse4_1(&sum32, a + i, b + i);
+              i += 8;
+            } while (i < width);
+            a += a_stride;
+            b += b_stride;
+            l += 1;
+          } while (l < 8 && l < (height - y));
+          summary_32_sse4(&sum32, &sum);
+          y += 8;
+        } while (y < height);
+      }
+      xx_storel_64(&sse, _mm_add_epi64(sum, _mm_srli_si128(sum, 8)));
+      break;
+  }
+  return sse;
+}
+#endif  // CONFIG_AV1_HIGHBITDEPTH
diff --git a/third_party/aom/aom_dsp/x86/ssim_sse2_x86_64.asm b/third_party/aom/aom_dsp/x86/ssim_sse2_x86_64.asm
new file mode 100644
index 0000000000..49bc655336
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/ssim_sse2_x86_64.asm
@@ -0,0 +1,222 @@
+;
+; Copyright (c) 2016, 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 "aom_ports/x86_abi_support.asm"
+
+; tabulate_ssim - sums sum_s,sum_r,sum_sq_s,sum_sq_r, sum_sxr
+%macro TABULATE_SSIM 0
+        paddusw         xmm15, xmm3  ; sum_s
+        paddusw         xmm14, xmm4  ; sum_r
+        movdqa          xmm1, xmm3
+        pmaddwd         xmm1, xmm1
+        paddd           xmm13, xmm1 ; sum_sq_s
+        movdqa          xmm2, xmm4
+        pmaddwd         xmm2, xmm2
+        paddd           xmm12, xmm2 ; sum_sq_r
+        pmaddwd         xmm3, xmm4
+        paddd           xmm11, xmm3  ; sum_sxr
+%endmacro
+
+; Sum across the register %1 starting with q words
+%macro SUM_ACROSS_Q 1
+        movdqa          xmm2,%1
+        punpckldq       %1,xmm0
+        punpckhdq       xmm2,xmm0
+        paddq           %1,xmm2
+        movdqa          xmm2,%1
+        punpcklqdq      %1,xmm0
+        punpckhqdq      xmm2,xmm0
+        paddq           %1,xmm2
+%endmacro
+
+; Sum across the register %1 starting with q words
+%macro SUM_ACROSS_W 1
+        movdqa          xmm1, %1
+        punpcklwd       %1,xmm0
+        punpckhwd       xmm1,xmm0
+        paddd           %1, xmm1
+        SUM_ACROSS_Q    %1
+%endmacro
+
+SECTION .text
+
+;void ssim_parms_sse2(
+;    unsigned char *s,
+;    int sp,
+;    unsigned char *r,
+;    int rp
+;    uint32_t *sum_s,
+;    uint32_t *sum_r,
+;    uint32_t *sum_sq_s,
+;    uint32_t *sum_sq_r,
+;    uint32_t *sum_sxr);
+;
+; TODO: Use parm passing through structure, probably don't need the pxors
+; ( calling app will initialize to 0 ) could easily fit everything in sse2
+; without too much hastle, and can probably do better estimates with psadw
+; or pavgb At this point this is just meant to be first pass for calculating
+; all the parms needed for 16x16 ssim so we can play with dssim as distortion
+; in mode selection code.
+globalsym(aom_ssim_parms_16x16_sse2)
+sym(aom_ssim_parms_16x16_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 9
+    SAVE_XMM 15
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    mov             rsi,        arg(0) ;s
+    mov             rcx,        arg(1) ;sp
+    mov             rdi,        arg(2) ;r
+    mov             rax,        arg(3) ;rp
+
+    pxor            xmm0, xmm0
+    pxor            xmm15,xmm15  ;sum_s
+    pxor            xmm14,xmm14  ;sum_r
+    pxor            xmm13,xmm13  ;sum_sq_s
+    pxor            xmm12,xmm12  ;sum_sq_r
+    pxor            xmm11,xmm11  ;sum_sxr
+
+    mov             rdx, 16      ;row counter
+.NextRow:
+
+    ;grab source and reference pixels
+    movdqu          xmm5, [rsi]
+    movdqu          xmm6, [rdi]
+    movdqa          xmm3, xmm5
+    movdqa          xmm4, xmm6
+    punpckhbw       xmm3, xmm0 ; high_s
+    punpckhbw       xmm4, xmm0 ; high_r
+
+    TABULATE_SSIM
+
+    movdqa          xmm3, xmm5
+    movdqa          xmm4, xmm6
+    punpcklbw       xmm3, xmm0 ; low_s
+    punpcklbw       xmm4, xmm0 ; low_r
+
+    TABULATE_SSIM
+
+    add             rsi, rcx   ; next s row
+    add             rdi, rax   ; next r row
+
+    dec             rdx        ; counter
+    jnz .NextRow
+
+    SUM_ACROSS_W    xmm15
+    SUM_ACROSS_W    xmm14
+    SUM_ACROSS_Q    xmm13
+    SUM_ACROSS_Q    xmm12
+    SUM_ACROSS_Q    xmm11
+
+    mov             rdi,arg(4)
+    movd            [rdi], xmm15;
+    mov             rdi,arg(5)
+    movd            [rdi], xmm14;
+    mov             rdi,arg(6)
+    movd            [rdi], xmm13;
+    mov             rdi,arg(7)
+    movd            [rdi], xmm12;
+    mov             rdi,arg(8)
+    movd            [rdi], xmm11;
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void ssim_parms_sse2(
+;    unsigned char *s,
+;    int sp,
+;    unsigned char *r,
+;    int rp
+;    uint32_t *sum_s,
+;    uint32_t *sum_r,
+;    uint32_t *sum_sq_s,
+;    uint32_t *sum_sq_r,
+;    uint32_t *sum_sxr);
+;
+; TODO: Use parm passing through structure, probably don't need the pxors
+; ( calling app will initialize to 0 ) could easily fit everything in sse2
+; without too much hastle, and can probably do better estimates with psadw
+; or pavgb At this point this is just meant to be first pass for calculating
+; all the parms needed for 16x16 ssim so we can play with dssim as distortion
+; in mode selection code.
+globalsym(aom_ssim_parms_8x8_sse2)
+sym(aom_ssim_parms_8x8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 9
+    SAVE_XMM 15
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    mov             rsi,        arg(0) ;s
+    mov             rcx,        arg(1) ;sp
+    mov             rdi,        arg(2) ;r
+    mov             rax,        arg(3) ;rp
+
+    pxor            xmm0, xmm0
+    pxor            xmm15,xmm15  ;sum_s
+    pxor            xmm14,xmm14  ;sum_r
+    pxor            xmm13,xmm13  ;sum_sq_s
+    pxor            xmm12,xmm12  ;sum_sq_r
+    pxor            xmm11,xmm11  ;sum_sxr
+
+    mov             rdx, 8      ;row counter
+.NextRow:
+
+    ;grab source and reference pixels
+    movq            xmm3, [rsi]
+    movq            xmm4, [rdi]
+    punpcklbw       xmm3, xmm0 ; low_s
+    punpcklbw       xmm4, xmm0 ; low_r
+
+    TABULATE_SSIM
+
+    add             rsi, rcx   ; next s row
+    add             rdi, rax   ; next r row
+
+    dec             rdx        ; counter
+    jnz .NextRow
+
+    SUM_ACROSS_W    xmm15
+    SUM_ACROSS_W    xmm14
+    SUM_ACROSS_Q    xmm13
+    SUM_ACROSS_Q    xmm12
+    SUM_ACROSS_Q    xmm11
+
+    mov             rdi,arg(4)
+    movd            [rdi], xmm15;
+    mov             rdi,arg(5)
+    movd            [rdi], xmm14;
+    mov             rdi,arg(6)
+    movd            [rdi], xmm13;
+    mov             rdi,arg(7)
+    movd            [rdi], xmm12;
+    mov             rdi,arg(8)
+    movd            [rdi], xmm11;
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
diff --git a/third_party/aom/aom_dsp/x86/subpel_variance_sse2.asm b/third_party/aom/aom_dsp/x86/subpel_variance_sse2.asm
new file mode 100644
index 0000000000..d1d8373456
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/subpel_variance_sse2.asm
@@ -0,0 +1,1470 @@
+;
+; Copyright (c) 2016, 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 "third_party/x86inc/x86inc.asm"
+
+SECTION_RODATA
+pw_8: times  8 dw  8
+bilin_filter_m_sse2: times  8 dw 16
+                     times  8 dw  0
+                     times  8 dw 14
+                     times  8 dw  2
+                     times  8 dw 12
+                     times  8 dw  4
+                     times  8 dw 10
+                     times  8 dw  6
+                     times 16 dw  8
+                     times  8 dw  6
+                     times  8 dw 10
+                     times  8 dw  4
+                     times  8 dw 12
+                     times  8 dw  2
+                     times  8 dw 14
+
+bilin_filter_m_ssse3: times  8 db 16,  0
+                      times  8 db 14,  2
+                      times  8 db 12,  4
+                      times  8 db 10,  6
+                      times 16 db  8
+                      times  8 db  6, 10
+                      times  8 db  4, 12
+                      times  8 db  2, 14
+
+SECTION .text
+
+; int aom_sub_pixel_varianceNxh(const uint8_t *src, ptrdiff_t src_stride,
+;                               int x_offset, int y_offset,
+;                               const uint8_t *dst, ptrdiff_t dst_stride,
+;                               int height, unsigned int *sse);
+;
+; This function returns the SE and stores SSE in the given pointer.
+
+%macro SUM_SSE 6 ; src1, dst1, src2, dst2, sum, sse
+  psubw                %3, %4
+  psubw                %1, %2
+  paddw                %5, %3
+  pmaddwd              %3, %3
+  paddw                %5, %1
+  pmaddwd              %1, %1
+  paddd                %6, %3
+  paddd                %6, %1
+%endmacro
+
+%macro STORE_AND_RET 1
+%if %1 > 4
+  ; if H=64 and W=16, we have 8 words of each 2(1bit)x64(6bit)x9bit=16bit
+  ; in m6, i.e. it _exactly_ fits in a signed word per word in the xmm reg.
+  ; We have to sign-extend it before adding the words within the register
+  ; and outputing to a dword.
+  pcmpgtw              m5, m6           ; mask for 0 > x
+  movhlps              m3, m7
+  punpcklwd            m4, m6, m5
+  punpckhwd            m6, m5           ; sign-extend m6 word->dword
+  paddd                m7, m3
+  paddd                m6, m4
+  pshufd               m3, m7, 0x1
+  movhlps              m4, m6
+  paddd                m7, m3
+  paddd                m6, m4
+  mov                  r1, ssem         ; r1 = unsigned int *sse
+  pshufd               m4, m6, 0x1
+  movd               [r1], m7           ; store sse
+  paddd                m6, m4
+  movd               raxd, m6           ; store sum as return value
+%else ; 4xh
+  pshuflw              m4, m6, 0xe
+  pshuflw              m3, m7, 0xe
+  paddw                m6, m4
+  paddd                m7, m3
+  pcmpgtw              m5, m6           ; mask for 0 > x
+  mov                  r1, ssem         ; r1 = unsigned int *sse
+  punpcklwd            m6, m5           ; sign-extend m6 word->dword
+  movd               [r1], m7           ; store sse
+  pshuflw              m4, m6, 0xe
+  paddd                m6, m4
+  movd               raxd, m6           ; store sum as return value
+%endif
+  RET
+%endmacro
+
+%macro INC_SRC_BY_SRC_STRIDE  0
+%if AOM_ARCH_X86=1 && CONFIG_PIC=1
+  add                srcq, src_stridemp
+%else
+  add                srcq, src_strideq
+%endif
+%endmacro
+
+%macro SUBPEL_VARIANCE 1-2 0 ; W
+%if cpuflag(ssse3)
+%define bilin_filter_m bilin_filter_m_ssse3
+%define filter_idx_shift 4
+%else
+%define bilin_filter_m bilin_filter_m_sse2
+%define filter_idx_shift 5
+%endif
+; FIXME(rbultje) only bilinear filters use >8 registers, and ssse3 only uses
+; 11, not 13, if the registers are ordered correctly. May make a minor speed
+; difference on Win64
+
+%if AOM_ARCH_X86_64
+  %if %2 == 1 ; avg
+    cglobal sub_pixel_avg_variance%1xh, 9, 10, 13, src, src_stride, \
+                                        x_offset, y_offset, dst, dst_stride, \
+                                        sec, sec_stride, height, sse
+    %define sec_str sec_strideq
+  %else
+    cglobal sub_pixel_variance%1xh, 7, 8, 13, src, src_stride, \
+                                    x_offset, y_offset, dst, dst_stride, \
+                                    height, sse
+  %endif
+  %define block_height heightd
+  %define bilin_filter sseq
+%else
+  %if CONFIG_PIC=1
+    %if %2 == 1 ; avg
+      cglobal sub_pixel_avg_variance%1xh, 7, 7, 13, src, src_stride, \
+                                          x_offset, y_offset, dst, dst_stride, \
+                                          sec, sec_stride, height, sse
+      %define block_height dword heightm
+      %define sec_str sec_stridemp
+    %else
+      cglobal sub_pixel_variance%1xh, 7, 7, 13, src, src_stride, \
+                                      x_offset, y_offset, dst, dst_stride, \
+                                      height, sse
+      %define block_height heightd
+    %endif
+
+    ; reuse argument stack space
+    %define g_bilin_filterm x_offsetm
+    %define g_pw_8m y_offsetm
+
+    ;Store bilin_filter and pw_8 location in stack
+    %if GET_GOT_DEFINED == 1
+      GET_GOT eax
+      add esp, 4                ; restore esp
+    %endif
+
+    lea ecx, [GLOBAL(bilin_filter_m)]
+    mov g_bilin_filterm, ecx
+
+    lea ecx, [GLOBAL(pw_8)]
+    mov g_pw_8m, ecx
+
+    LOAD_IF_USED 0, 1         ; load eax, ecx back
+  %else
+    %if %2 == 1 ; avg
+      cglobal sub_pixel_avg_variance%1xh, 7, 7, 13, src, src_stride, \
+                                          x_offset, y_offset, \
+                                          dst, dst_stride, sec, sec_stride, \
+                                          height, sse
+      %define block_height dword heightm
+      %define sec_str sec_stridemp
+    %else
+      cglobal sub_pixel_variance%1xh, 7, 7, 13, src, src_stride, \
+                                      x_offset, y_offset, dst, dst_stride, \
+                                      height, sse
+      %define block_height heightd
+    %endif
+    %define bilin_filter bilin_filter_m
+  %endif
+%endif
+
+%if %1 == 4
+  %define movx movd
+%else
+  %define movx movh
+%endif
+
+  ASSERT               %1 <= 16         ; m6 overflows if w > 16
+  pxor                 m6, m6           ; sum
+  pxor                 m7, m7           ; sse
+  ; FIXME(rbultje) if both filters are bilinear, we don't actually use m5; we
+  ; could perhaps use it for something more productive then
+  pxor                 m5, m5           ; dedicated zero register
+%if %1 < 16
+  sar                   block_height, 1
+%if %2 == 1 ; avg
+  shl             sec_str, 1
+%endif
+%endif
+
+  ; FIXME(rbultje) replace by jumptable?
+  test          x_offsetd, x_offsetd
+  jnz .x_nonzero
+  ; x_offset == 0
+  test          y_offsetd, y_offsetd
+  jnz .x_zero_y_nonzero
+
+  ; x_offset == 0 && y_offset == 0
+.x_zero_y_zero_loop:
+%if %1 == 16
+  movu                 m0, [srcq]
+  mova                 m1, [dstq]
+%if %2 == 1 ; avg
+  pavgb                m0, [secq]
+  punpckhbw            m3, m1, m5
+  punpcklbw            m1, m5
+%endif
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+
+%if %2 == 0 ; !avg
+  punpckhbw            m3, m1, m5
+  punpcklbw            m1, m5
+%endif
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  add                srcq, src_strideq
+  add                dstq, dst_strideq
+%else ; %1 < 16
+  movx                 m0, [srcq]
+%if %2 == 1 ; avg
+%if %1 > 4
+  movhps               m0, [srcq+src_strideq]
+%else ; 4xh
+  movx                 m1, [srcq+src_strideq]
+  punpckldq            m0, m1
+%endif
+%else ; !avg
+  movx                 m2, [srcq+src_strideq]
+%endif
+
+  movx                 m1, [dstq]
+  movx                 m3, [dstq+dst_strideq]
+
+%if %2 == 1 ; avg
+%if %1 > 4
+  pavgb                m0, [secq]
+%else
+  movh                 m2, [secq]
+  pavgb                m0, m2
+%endif
+  punpcklbw            m3, m5
+  punpcklbw            m1, m5
+%if %1 > 4
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%else ; 4xh
+  punpcklbw            m0, m5
+  movhlps              m2, m0
+%endif
+%else ; !avg
+  punpcklbw            m0, m5
+  punpcklbw            m2, m5
+  punpcklbw            m3, m5
+  punpcklbw            m1, m5
+%endif
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  lea                srcq, [srcq+src_strideq*2]
+  lea                dstq, [dstq+dst_strideq*2]
+%endif
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+  dec                   block_height
+  jg .x_zero_y_zero_loop
+  STORE_AND_RET %1
+
+.x_zero_y_nonzero:
+  cmp           y_offsetd, 4
+  jne .x_zero_y_nonhalf
+
+  ; x_offset == 0 && y_offset == 0.5
+.x_zero_y_half_loop:
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m4, [srcq+src_strideq]
+  mova                 m1, [dstq]
+  pavgb                m0, m4
+  punpckhbw            m3, m1, m5
+%if %2 == 1 ; avg
+  pavgb                m0, [secq]
+%endif
+  punpcklbw            m1, m5
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  add                srcq, src_strideq
+  add                dstq, dst_strideq
+%else ; %1 < 16
+  movx                 m0, [srcq]
+  movx                 m2, [srcq+src_strideq]
+%if %2 == 1 ; avg
+%if %1 > 4
+  movhps               m2, [srcq+src_strideq*2]
+%else ; 4xh
+  movx                 m1, [srcq+src_strideq*2]
+  punpckldq            m2, m1
+%endif
+  movx                 m1, [dstq]
+%if %1 > 4
+  movlhps              m0, m2
+%else ; 4xh
+  punpckldq            m0, m2
+%endif
+  movx                 m3, [dstq+dst_strideq]
+  pavgb                m0, m2
+  punpcklbw            m1, m5
+%if %1 > 4
+  pavgb                m0, [secq]
+  punpcklbw            m3, m5
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%else ; 4xh
+  movh                 m4, [secq]
+  pavgb                m0, m4
+  punpcklbw            m3, m5
+  punpcklbw            m0, m5
+  movhlps              m2, m0
+%endif
+%else ; !avg
+  movx                 m4, [srcq+src_strideq*2]
+  movx                 m1, [dstq]
+  pavgb                m0, m2
+  movx                 m3, [dstq+dst_strideq]
+  pavgb                m2, m4
+  punpcklbw            m0, m5
+  punpcklbw            m2, m5
+  punpcklbw            m3, m5
+  punpcklbw            m1, m5
+%endif
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  lea                srcq, [srcq+src_strideq*2]
+  lea                dstq, [dstq+dst_strideq*2]
+%endif
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+  dec                   block_height
+  jg .x_zero_y_half_loop
+  STORE_AND_RET %1
+
+.x_zero_y_nonhalf:
+  ; x_offset == 0 && y_offset == bilin interpolation
+%if AOM_ARCH_X86_64
+  lea        bilin_filter, [GLOBAL(bilin_filter_m)]
+%endif
+  shl           y_offsetd, filter_idx_shift
+%if AOM_ARCH_X86_64 && %1 > 4
+  mova                 m8, [bilin_filter+y_offsetq]
+%if notcpuflag(ssse3) ; FIXME(rbultje) don't scatter registers on x86-64
+  mova                 m9, [bilin_filter+y_offsetq+16]
+%endif
+  mova                m10, [GLOBAL(pw_8)]
+%define filter_y_a m8
+%define filter_y_b m9
+%define filter_rnd m10
+%else ; x86-32 or mmx
+%if AOM_ARCH_X86=1 && CONFIG_PIC=1
+; x_offset == 0, reuse x_offset reg
+%define tempq x_offsetq
+  add y_offsetq, g_bilin_filterm
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+  mov tempq, g_pw_8m
+%define filter_rnd [tempq]
+%else
+  add           y_offsetq, bilin_filter
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+%define filter_rnd [GLOBAL(pw_8)]
+%endif
+%endif
+
+.x_zero_y_other_loop:
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m4, [srcq+src_strideq]
+  mova                 m1, [dstq]
+%if cpuflag(ssse3)
+  punpckhbw            m2, m0, m4
+  punpcklbw            m0, m4
+  pmaddubsw            m2, filter_y_a
+  pmaddubsw            m0, filter_y_a
+  paddw                m2, filter_rnd
+  paddw                m0, filter_rnd
+%else
+  punpckhbw            m2, m0, m5
+  punpckhbw            m3, m4, m5
+  punpcklbw            m0, m5
+  punpcklbw            m4, m5
+  ; FIXME(rbultje) instead of out=((num-x)*in1+x*in2+rnd)>>log2(num), we can
+  ; also do out=in1+(((num-x)*(in2-in1)+rnd)>>log2(num)). Total number of
+  ; instructions is the same (5), but it is 1 mul instead of 2, so might be
+  ; slightly faster because of pmullw latency. It would also cut our rodata
+  ; tables in half for this function, and save 1-2 registers on x86-64.
+  pmullw               m2, filter_y_a
+  pmullw               m3, filter_y_b
+  paddw                m2, filter_rnd
+  pmullw               m0, filter_y_a
+  pmullw               m4, filter_y_b
+  paddw                m0, filter_rnd
+  paddw                m2, m3
+  paddw                m0, m4
+%endif
+  psraw                m2, 4
+  psraw                m0, 4
+%if %2 == 1 ; avg
+  ; FIXME(rbultje) pipeline
+  packuswb             m0, m2
+  pavgb                m0, [secq]
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%endif
+  punpckhbw            m3, m1, m5
+  punpcklbw            m1, m5
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  add                srcq, src_strideq
+  add                dstq, dst_strideq
+%else ; %1 < 16
+  movx                 m0, [srcq]
+  movx                 m2, [srcq+src_strideq]
+  movx                 m4, [srcq+src_strideq*2]
+  movx                 m3, [dstq+dst_strideq]
+%if cpuflag(ssse3)
+  movx                 m1, [dstq]
+  punpcklbw            m0, m2
+  punpcklbw            m2, m4
+  pmaddubsw            m0, filter_y_a
+  pmaddubsw            m2, filter_y_a
+  punpcklbw            m3, m5
+  paddw                m2, filter_rnd
+  paddw                m0, filter_rnd
+%else
+  punpcklbw            m0, m5
+  punpcklbw            m2, m5
+  punpcklbw            m4, m5
+  pmullw               m0, filter_y_a
+  pmullw               m1, m2, filter_y_b
+  punpcklbw            m3, m5
+  paddw                m0, filter_rnd
+  pmullw               m2, filter_y_a
+  pmullw               m4, filter_y_b
+  paddw                m0, m1
+  paddw                m2, filter_rnd
+  movx                 m1, [dstq]
+  paddw                m2, m4
+%endif
+  psraw                m0, 4
+  psraw                m2, 4
+%if %2 == 1 ; avg
+  ; FIXME(rbultje) pipeline
+%if %1 == 4
+  movlhps              m0, m2
+%endif
+  packuswb             m0, m2
+%if %1 > 4
+  pavgb                m0, [secq]
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%else ; 4xh
+  movh                 m2, [secq]
+  pavgb                m0, m2
+  punpcklbw            m0, m5
+  movhlps              m2, m0
+%endif
+%endif
+  punpcklbw            m1, m5
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  lea                srcq, [srcq+src_strideq*2]
+  lea                dstq, [dstq+dst_strideq*2]
+%endif
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+  dec                   block_height
+  jg .x_zero_y_other_loop
+%undef filter_y_a
+%undef filter_y_b
+%undef filter_rnd
+  STORE_AND_RET %1
+
+.x_nonzero:
+  cmp           x_offsetd, 4
+  jne .x_nonhalf
+  ; x_offset == 0.5
+  test          y_offsetd, y_offsetd
+  jnz .x_half_y_nonzero
+
+  ; x_offset == 0.5 && y_offset == 0
+.x_half_y_zero_loop:
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m4, [srcq+1]
+  mova                 m1, [dstq]
+  pavgb                m0, m4
+  punpckhbw            m3, m1, m5
+%if %2 == 1 ; avg
+  pavgb                m0, [secq]
+%endif
+  punpcklbw            m1, m5
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  add                srcq, src_strideq
+  add                dstq, dst_strideq
+%else ; %1 < 16
+  movx                 m0, [srcq]
+  movx                 m4, [srcq+1]
+%if %2 == 1 ; avg
+%if %1 > 4
+  movhps               m0, [srcq+src_strideq]
+  movhps               m4, [srcq+src_strideq+1]
+%else ; 4xh
+  movx                 m1, [srcq+src_strideq]
+  punpckldq            m0, m1
+  movx                 m2, [srcq+src_strideq+1]
+  punpckldq            m4, m2
+%endif
+  movx                 m1, [dstq]
+  movx                 m3, [dstq+dst_strideq]
+  pavgb                m0, m4
+  punpcklbw            m3, m5
+%if %1 > 4
+  pavgb                m0, [secq]
+  punpcklbw            m1, m5
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%else ; 4xh
+  movh                 m2, [secq]
+  pavgb                m0, m2
+  punpcklbw            m1, m5
+  punpcklbw            m0, m5
+  movhlps              m2, m0
+%endif
+%else ; !avg
+  movx                 m2, [srcq+src_strideq]
+  movx                 m1, [dstq]
+  pavgb                m0, m4
+  movx                 m4, [srcq+src_strideq+1]
+  movx                 m3, [dstq+dst_strideq]
+  pavgb                m2, m4
+  punpcklbw            m0, m5
+  punpcklbw            m2, m5
+  punpcklbw            m3, m5
+  punpcklbw            m1, m5
+%endif
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  lea                srcq, [srcq+src_strideq*2]
+  lea                dstq, [dstq+dst_strideq*2]
+%endif
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+  dec                   block_height
+  jg .x_half_y_zero_loop
+  STORE_AND_RET %1
+
+.x_half_y_nonzero:
+  cmp           y_offsetd, 4
+  jne .x_half_y_nonhalf
+
+  ; x_offset == 0.5 && y_offset == 0.5
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m3, [srcq+1]
+  add                srcq, src_strideq
+  pavgb                m0, m3
+.x_half_y_half_loop:
+  movu                 m4, [srcq]
+  movu                 m3, [srcq+1]
+  mova                 m1, [dstq]
+  pavgb                m4, m3
+  punpckhbw            m3, m1, m5
+  pavgb                m0, m4
+%if %2 == 1 ; avg
+  punpcklbw            m1, m5
+  pavgb                m0, [secq]
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%else
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+  punpcklbw            m1, m5
+%endif
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+  mova                 m0, m4
+
+  add                srcq, src_strideq
+  add                dstq, dst_strideq
+%else ; %1 < 16
+  movx                 m0, [srcq]
+  movx                 m3, [srcq+1]
+  add                srcq, src_strideq
+  pavgb                m0, m3
+.x_half_y_half_loop:
+  movx                 m2, [srcq]
+  movx                 m3, [srcq+1]
+%if %2 == 1 ; avg
+%if %1 > 4
+  movhps               m2, [srcq+src_strideq]
+  movhps               m3, [srcq+src_strideq+1]
+%else
+  movx                 m1, [srcq+src_strideq]
+  punpckldq            m2, m1
+  movx                 m1, [srcq+src_strideq+1]
+  punpckldq            m3, m1
+%endif
+  pavgb                m2, m3
+%if %1 > 4
+  movlhps              m0, m2
+  movhlps              m4, m2
+%else ; 4xh
+  punpckldq            m0, m2
+  pshuflw              m4, m2, 0xe
+%endif
+  movx                 m1, [dstq]
+  pavgb                m0, m2
+  movx                 m3, [dstq+dst_strideq]
+%if %1 > 4
+  pavgb                m0, [secq]
+%else
+  movh                 m2, [secq]
+  pavgb                m0, m2
+%endif
+  punpcklbw            m3, m5
+  punpcklbw            m1, m5
+%if %1 > 4
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%else
+  punpcklbw            m0, m5
+  movhlps              m2, m0
+%endif
+%else ; !avg
+  movx                 m4, [srcq+src_strideq]
+  movx                 m1, [srcq+src_strideq+1]
+  pavgb                m2, m3
+  pavgb                m4, m1
+  pavgb                m0, m2
+  pavgb                m2, m4
+  movx                 m1, [dstq]
+  movx                 m3, [dstq+dst_strideq]
+  punpcklbw            m0, m5
+  punpcklbw            m2, m5
+  punpcklbw            m3, m5
+  punpcklbw            m1, m5
+%endif
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+  mova                 m0, m4
+
+  lea                srcq, [srcq+src_strideq*2]
+  lea                dstq, [dstq+dst_strideq*2]
+%endif
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+  dec                   block_height
+  jg .x_half_y_half_loop
+  STORE_AND_RET %1
+
+.x_half_y_nonhalf:
+  ; x_offset == 0.5 && y_offset == bilin interpolation
+%if AOM_ARCH_X86_64
+  lea        bilin_filter, [GLOBAL(bilin_filter_m)]
+%endif
+  shl           y_offsetd, filter_idx_shift
+%if AOM_ARCH_X86_64 && %1 > 4
+  mova                 m8, [bilin_filter+y_offsetq]
+%if notcpuflag(ssse3) ; FIXME(rbultje) don't scatter registers on x86-64
+  mova                 m9, [bilin_filter+y_offsetq+16]
+%endif
+  mova                m10, [GLOBAL(pw_8)]
+%define filter_y_a m8
+%define filter_y_b m9
+%define filter_rnd m10
+%else  ;x86_32
+%if AOM_ARCH_X86=1 && CONFIG_PIC=1
+; x_offset == 0.5. We can reuse x_offset reg
+%define tempq x_offsetq
+  add y_offsetq, g_bilin_filterm
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+  mov tempq, g_pw_8m
+%define filter_rnd [tempq]
+%else
+  add           y_offsetq, bilin_filter
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+%define filter_rnd [GLOBAL(pw_8)]
+%endif
+%endif
+
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m3, [srcq+1]
+  add                srcq, src_strideq
+  pavgb                m0, m3
+.x_half_y_other_loop:
+  movu                 m4, [srcq]
+  movu                 m2, [srcq+1]
+  mova                 m1, [dstq]
+  pavgb                m4, m2
+%if cpuflag(ssse3)
+  punpckhbw            m2, m0, m4
+  punpcklbw            m0, m4
+  pmaddubsw            m2, filter_y_a
+  pmaddubsw            m0, filter_y_a
+  paddw                m2, filter_rnd
+  paddw                m0, filter_rnd
+  psraw                m2, 4
+%else
+  punpckhbw            m2, m0, m5
+  punpckhbw            m3, m4, m5
+  pmullw               m2, filter_y_a
+  pmullw               m3, filter_y_b
+  paddw                m2, filter_rnd
+  punpcklbw            m0, m5
+  paddw                m2, m3
+  punpcklbw            m3, m4, m5
+  pmullw               m0, filter_y_a
+  pmullw               m3, filter_y_b
+  paddw                m0, filter_rnd
+  psraw                m2, 4
+  paddw                m0, m3
+%endif
+  punpckhbw            m3, m1, m5
+  psraw                m0, 4
+%if %2 == 1 ; avg
+  ; FIXME(rbultje) pipeline
+  packuswb             m0, m2
+  pavgb                m0, [secq]
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%endif
+  punpcklbw            m1, m5
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+  mova                 m0, m4
+
+  add                srcq, src_strideq
+  add                dstq, dst_strideq
+%else ; %1 < 16
+  movx                 m0, [srcq]
+  movx                 m3, [srcq+1]
+  add                srcq, src_strideq
+  pavgb                m0, m3
+%if notcpuflag(ssse3)
+  punpcklbw            m0, m5
+%endif
+.x_half_y_other_loop:
+  movx                 m2, [srcq]
+  movx                 m1, [srcq+1]
+  movx                 m4, [srcq+src_strideq]
+  movx                 m3, [srcq+src_strideq+1]
+  pavgb                m2, m1
+  pavgb                m4, m3
+  movx                 m3, [dstq+dst_strideq]
+%if cpuflag(ssse3)
+  movx                 m1, [dstq]
+  punpcklbw            m0, m2
+  punpcklbw            m2, m4
+  pmaddubsw            m0, filter_y_a
+  pmaddubsw            m2, filter_y_a
+  punpcklbw            m3, m5
+  paddw                m0, filter_rnd
+  paddw                m2, filter_rnd
+%else
+  punpcklbw            m2, m5
+  punpcklbw            m4, m5
+  pmullw               m0, filter_y_a
+  pmullw               m1, m2, filter_y_b
+  punpcklbw            m3, m5
+  paddw                m0, filter_rnd
+  pmullw               m2, filter_y_a
+  paddw                m0, m1
+  pmullw               m1, m4, filter_y_b
+  paddw                m2, filter_rnd
+  paddw                m2, m1
+  movx                 m1, [dstq]
+%endif
+  psraw                m0, 4
+  psraw                m2, 4
+%if %2 == 1 ; avg
+  ; FIXME(rbultje) pipeline
+%if %1 == 4
+  movlhps              m0, m2
+%endif
+  packuswb             m0, m2
+%if %1 > 4
+  pavgb                m0, [secq]
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%else
+  movh                 m2, [secq]
+  pavgb                m0, m2
+  punpcklbw            m0, m5
+  movhlps              m2, m0
+%endif
+%endif
+  punpcklbw            m1, m5
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+  mova                 m0, m4
+
+  lea                srcq, [srcq+src_strideq*2]
+  lea                dstq, [dstq+dst_strideq*2]
+%endif
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+  dec                   block_height
+  jg .x_half_y_other_loop
+%undef filter_y_a
+%undef filter_y_b
+%undef filter_rnd
+  STORE_AND_RET %1
+
+.x_nonhalf:
+  test          y_offsetd, y_offsetd
+  jnz .x_nonhalf_y_nonzero
+
+  ; x_offset == bilin interpolation && y_offset == 0
+%if AOM_ARCH_X86_64
+  lea        bilin_filter, [GLOBAL(bilin_filter_m)]
+%endif
+  shl           x_offsetd, filter_idx_shift
+%if AOM_ARCH_X86_64 && %1 > 4
+  mova                 m8, [bilin_filter+x_offsetq]
+%if notcpuflag(ssse3) ; FIXME(rbultje) don't scatter registers on x86-64
+  mova                 m9, [bilin_filter+x_offsetq+16]
+%endif
+  mova                m10, [GLOBAL(pw_8)]
+%define filter_x_a m8
+%define filter_x_b m9
+%define filter_rnd m10
+%else    ; x86-32
+%if AOM_ARCH_X86=1 && CONFIG_PIC=1
+;y_offset == 0. We can reuse y_offset reg.
+%define tempq y_offsetq
+  add x_offsetq, g_bilin_filterm
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+  mov tempq, g_pw_8m
+%define filter_rnd [tempq]
+%else
+  add           x_offsetq, bilin_filter
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+%define filter_rnd [GLOBAL(pw_8)]
+%endif
+%endif
+
+.x_other_y_zero_loop:
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m4, [srcq+1]
+  mova                 m1, [dstq]
+%if cpuflag(ssse3)
+  punpckhbw            m2, m0, m4
+  punpcklbw            m0, m4
+  pmaddubsw            m2, filter_x_a
+  pmaddubsw            m0, filter_x_a
+  paddw                m2, filter_rnd
+  paddw                m0, filter_rnd
+%else
+  punpckhbw            m2, m0, m5
+  punpckhbw            m3, m4, m5
+  punpcklbw            m0, m5
+  punpcklbw            m4, m5
+  pmullw               m2, filter_x_a
+  pmullw               m3, filter_x_b
+  paddw                m2, filter_rnd
+  pmullw               m0, filter_x_a
+  pmullw               m4, filter_x_b
+  paddw                m0, filter_rnd
+  paddw                m2, m3
+  paddw                m0, m4
+%endif
+  psraw                m2, 4
+  psraw                m0, 4
+%if %2 == 1 ; avg
+  ; FIXME(rbultje) pipeline
+  packuswb             m0, m2
+  pavgb                m0, [secq]
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%endif
+  punpckhbw            m3, m1, m5
+  punpcklbw            m1, m5
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  add                srcq, src_strideq
+  add                dstq, dst_strideq
+%else ; %1 < 16
+  movx                 m0, [srcq]
+  movx                 m1, [srcq+1]
+  movx                 m2, [srcq+src_strideq]
+  movx                 m4, [srcq+src_strideq+1]
+  movx                 m3, [dstq+dst_strideq]
+%if cpuflag(ssse3)
+  punpcklbw            m0, m1
+  movx                 m1, [dstq]
+  punpcklbw            m2, m4
+  pmaddubsw            m0, filter_x_a
+  pmaddubsw            m2, filter_x_a
+  punpcklbw            m3, m5
+  paddw                m0, filter_rnd
+  paddw                m2, filter_rnd
+%else
+  punpcklbw            m0, m5
+  punpcklbw            m1, m5
+  punpcklbw            m2, m5
+  punpcklbw            m4, m5
+  pmullw               m0, filter_x_a
+  pmullw               m1, filter_x_b
+  punpcklbw            m3, m5
+  paddw                m0, filter_rnd
+  pmullw               m2, filter_x_a
+  pmullw               m4, filter_x_b
+  paddw                m0, m1
+  paddw                m2, filter_rnd
+  movx                 m1, [dstq]
+  paddw                m2, m4
+%endif
+  psraw                m0, 4
+  psraw                m2, 4
+%if %2 == 1 ; avg
+  ; FIXME(rbultje) pipeline
+%if %1 == 4
+  movlhps              m0, m2
+%endif
+  packuswb             m0, m2
+%if %1 > 4
+  pavgb                m0, [secq]
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%else
+  movh                 m2, [secq]
+  pavgb                m0, m2
+  punpcklbw            m0, m5
+  movhlps              m2, m0
+%endif
+%endif
+  punpcklbw            m1, m5
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  lea                srcq, [srcq+src_strideq*2]
+  lea                dstq, [dstq+dst_strideq*2]
+%endif
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+  dec                   block_height
+  jg .x_other_y_zero_loop
+%undef filter_x_a
+%undef filter_x_b
+%undef filter_rnd
+  STORE_AND_RET %1
+
+.x_nonhalf_y_nonzero:
+  cmp           y_offsetd, 4
+  jne .x_nonhalf_y_nonhalf
+
+  ; x_offset == bilin interpolation && y_offset == 0.5
+%if AOM_ARCH_X86_64
+  lea        bilin_filter, [GLOBAL(bilin_filter_m)]
+%endif
+  shl           x_offsetd, filter_idx_shift
+%if AOM_ARCH_X86_64 && %1 > 4
+  mova                 m8, [bilin_filter+x_offsetq]
+%if notcpuflag(ssse3) ; FIXME(rbultje) don't scatter registers on x86-64
+  mova                 m9, [bilin_filter+x_offsetq+16]
+%endif
+  mova                m10, [GLOBAL(pw_8)]
+%define filter_x_a m8
+%define filter_x_b m9
+%define filter_rnd m10
+%else    ; x86-32
+%if AOM_ARCH_X86=1 && CONFIG_PIC=1
+; y_offset == 0.5. We can reuse y_offset reg.
+%define tempq y_offsetq
+  add x_offsetq, g_bilin_filterm
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+  mov tempq, g_pw_8m
+%define filter_rnd [tempq]
+%else
+  add           x_offsetq, bilin_filter
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+%define filter_rnd [GLOBAL(pw_8)]
+%endif
+%endif
+
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq+1]
+%if cpuflag(ssse3)
+  punpckhbw            m2, m0, m1
+  punpcklbw            m0, m1
+  pmaddubsw            m2, filter_x_a
+  pmaddubsw            m0, filter_x_a
+  paddw                m2, filter_rnd
+  paddw                m0, filter_rnd
+%else
+  punpckhbw            m2, m0, m5
+  punpckhbw            m3, m1, m5
+  punpcklbw            m0, m5
+  punpcklbw            m1, m5
+  pmullw               m0, filter_x_a
+  pmullw               m1, filter_x_b
+  paddw                m0, filter_rnd
+  pmullw               m2, filter_x_a
+  pmullw               m3, filter_x_b
+  paddw                m2, filter_rnd
+  paddw                m0, m1
+  paddw                m2, m3
+%endif
+  psraw                m0, 4
+  psraw                m2, 4
+  add                srcq, src_strideq
+  packuswb             m0, m2
+.x_other_y_half_loop:
+  movu                 m4, [srcq]
+  movu                 m3, [srcq+1]
+%if cpuflag(ssse3)
+  mova                 m1, [dstq]
+  punpckhbw            m2, m4, m3
+  punpcklbw            m4, m3
+  pmaddubsw            m2, filter_x_a
+  pmaddubsw            m4, filter_x_a
+  paddw                m2, filter_rnd
+  paddw                m4, filter_rnd
+  psraw                m2, 4
+  psraw                m4, 4
+  packuswb             m4, m2
+  pavgb                m0, m4
+  punpckhbw            m3, m1, m5
+  punpcklbw            m1, m5
+%else
+  punpckhbw            m2, m4, m5
+  punpckhbw            m1, m3, m5
+  punpcklbw            m4, m5
+  punpcklbw            m3, m5
+  pmullw               m4, filter_x_a
+  pmullw               m3, filter_x_b
+  paddw                m4, filter_rnd
+  pmullw               m2, filter_x_a
+  pmullw               m1, filter_x_b
+  paddw                m2, filter_rnd
+  paddw                m4, m3
+  paddw                m2, m1
+  mova                 m1, [dstq]
+  psraw                m4, 4
+  psraw                m2, 4
+  punpckhbw            m3, m1, m5
+  ; FIXME(rbultje) the repeated pack/unpack here around m0/m2 is because we
+  ; have a 1-register shortage to be able to store the backup of the bilin
+  ; filtered second line as words as cache for the next line. Packing into
+  ; a byte costs 1 pack and 2 unpacks, but saves a register.
+  packuswb             m4, m2
+  punpcklbw            m1, m5
+  pavgb                m0, m4
+%endif
+%if %2 == 1 ; avg
+  ; FIXME(rbultje) pipeline
+  pavgb                m0, [secq]
+%endif
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+  mova                 m0, m4
+
+  add                srcq, src_strideq
+  add                dstq, dst_strideq
+%else ; %1 < 16
+  movx                 m0, [srcq]
+  movx                 m1, [srcq+1]
+%if cpuflag(ssse3)
+  punpcklbw            m0, m1
+  pmaddubsw            m0, filter_x_a
+  paddw                m0, filter_rnd
+%else
+  punpcklbw            m0, m5
+  punpcklbw            m1, m5
+  pmullw               m0, filter_x_a
+  pmullw               m1, filter_x_b
+  paddw                m0, filter_rnd
+  paddw                m0, m1
+%endif
+  add                srcq, src_strideq
+  psraw                m0, 4
+.x_other_y_half_loop:
+  movx                 m2, [srcq]
+  movx                 m1, [srcq+1]
+  movx                 m4, [srcq+src_strideq]
+  movx                 m3, [srcq+src_strideq+1]
+%if cpuflag(ssse3)
+  punpcklbw            m2, m1
+  punpcklbw            m4, m3
+  pmaddubsw            m2, filter_x_a
+  pmaddubsw            m4, filter_x_a
+  movx                 m1, [dstq]
+  movx                 m3, [dstq+dst_strideq]
+  paddw                m2, filter_rnd
+  paddw                m4, filter_rnd
+%else
+  punpcklbw            m2, m5
+  punpcklbw            m1, m5
+  punpcklbw            m4, m5
+  punpcklbw            m3, m5
+  pmullw               m2, filter_x_a
+  pmullw               m1, filter_x_b
+  paddw                m2, filter_rnd
+  pmullw               m4, filter_x_a
+  pmullw               m3, filter_x_b
+  paddw                m4, filter_rnd
+  paddw                m2, m1
+  movx                 m1, [dstq]
+  paddw                m4, m3
+  movx                 m3, [dstq+dst_strideq]
+%endif
+  psraw                m2, 4
+  psraw                m4, 4
+  pavgw                m0, m2
+  pavgw                m2, m4
+%if %2 == 1 ; avg
+  ; FIXME(rbultje) pipeline - also consider going to bytes here
+%if %1 == 4
+  movlhps              m0, m2
+%endif
+  packuswb             m0, m2
+%if %1 > 4
+  pavgb                m0, [secq]
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%else
+  movh                 m2, [secq]
+  pavgb                m0, m2
+  punpcklbw            m0, m5
+  movhlps              m2, m0
+%endif
+%endif
+  punpcklbw            m3, m5
+  punpcklbw            m1, m5
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+  mova                 m0, m4
+
+  lea                srcq, [srcq+src_strideq*2]
+  lea                dstq, [dstq+dst_strideq*2]
+%endif
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+  dec                   block_height
+  jg .x_other_y_half_loop
+%undef filter_x_a
+%undef filter_x_b
+%undef filter_rnd
+  STORE_AND_RET %1
+
+.x_nonhalf_y_nonhalf:
+%if AOM_ARCH_X86_64
+  lea        bilin_filter, [GLOBAL(bilin_filter_m)]
+%endif
+  shl           x_offsetd, filter_idx_shift
+  shl           y_offsetd, filter_idx_shift
+%if AOM_ARCH_X86_64 && %1 > 4
+  mova                 m8, [bilin_filter+x_offsetq]
+%if notcpuflag(ssse3) ; FIXME(rbultje) don't scatter registers on x86-64
+  mova                 m9, [bilin_filter+x_offsetq+16]
+%endif
+  mova                m10, [bilin_filter+y_offsetq]
+%if notcpuflag(ssse3) ; FIXME(rbultje) don't scatter registers on x86-64
+  mova                m11, [bilin_filter+y_offsetq+16]
+%endif
+  mova                m12, [GLOBAL(pw_8)]
+%define filter_x_a m8
+%define filter_x_b m9
+%define filter_y_a m10
+%define filter_y_b m11
+%define filter_rnd m12
+%else   ; x86-32
+%if AOM_ARCH_X86=1 && CONFIG_PIC=1
+; In this case, there is NO unused register. Used src_stride register. Later,
+; src_stride has to be loaded from stack when it is needed.
+%define tempq src_strideq
+  mov tempq, g_bilin_filterm
+  add           x_offsetq, tempq
+  add           y_offsetq, tempq
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+
+  mov tempq, g_pw_8m
+%define filter_rnd [tempq]
+%else
+  add           x_offsetq, bilin_filter
+  add           y_offsetq, bilin_filter
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+%define filter_rnd [GLOBAL(pw_8)]
+%endif
+%endif
+
+  ; x_offset == bilin interpolation && y_offset == bilin interpolation
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq+1]
+%if cpuflag(ssse3)
+  punpckhbw            m2, m0, m1
+  punpcklbw            m0, m1
+  pmaddubsw            m2, filter_x_a
+  pmaddubsw            m0, filter_x_a
+  paddw                m2, filter_rnd
+  paddw                m0, filter_rnd
+%else
+  punpckhbw            m2, m0, m5
+  punpckhbw            m3, m1, m5
+  punpcklbw            m0, m5
+  punpcklbw            m1, m5
+  pmullw               m0, filter_x_a
+  pmullw               m1, filter_x_b
+  paddw                m0, filter_rnd
+  pmullw               m2, filter_x_a
+  pmullw               m3, filter_x_b
+  paddw                m2, filter_rnd
+  paddw                m0, m1
+  paddw                m2, m3
+%endif
+  psraw                m0, 4
+  psraw                m2, 4
+
+  INC_SRC_BY_SRC_STRIDE
+
+  packuswb             m0, m2
+.x_other_y_other_loop:
+%if cpuflag(ssse3)
+  movu                 m4, [srcq]
+  movu                 m3, [srcq+1]
+  mova                 m1, [dstq]
+  punpckhbw            m2, m4, m3
+  punpcklbw            m4, m3
+  pmaddubsw            m2, filter_x_a
+  pmaddubsw            m4, filter_x_a
+  punpckhbw            m3, m1, m5
+  paddw                m2, filter_rnd
+  paddw                m4, filter_rnd
+  psraw                m2, 4
+  psraw                m4, 4
+  packuswb             m4, m2
+  punpckhbw            m2, m0, m4
+  punpcklbw            m0, m4
+  pmaddubsw            m2, filter_y_a
+  pmaddubsw            m0, filter_y_a
+  punpcklbw            m1, m5
+  paddw                m2, filter_rnd
+  paddw                m0, filter_rnd
+  psraw                m2, 4
+  psraw                m0, 4
+%else
+  movu                 m3, [srcq]
+  movu                 m4, [srcq+1]
+  punpckhbw            m1, m3, m5
+  punpckhbw            m2, m4, m5
+  punpcklbw            m3, m5
+  punpcklbw            m4, m5
+  pmullw               m3, filter_x_a
+  pmullw               m4, filter_x_b
+  paddw                m3, filter_rnd
+  pmullw               m1, filter_x_a
+  pmullw               m2, filter_x_b
+  paddw                m1, filter_rnd
+  paddw                m3, m4
+  paddw                m1, m2
+  psraw                m3, 4
+  psraw                m1, 4
+  packuswb             m4, m3, m1
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+  pmullw               m2, filter_y_a
+  pmullw               m1, filter_y_b
+  paddw                m2, filter_rnd
+  pmullw               m0, filter_y_a
+  pmullw               m3, filter_y_b
+  paddw                m2, m1
+  mova                 m1, [dstq]
+  paddw                m0, filter_rnd
+  psraw                m2, 4
+  paddw                m0, m3
+  punpckhbw            m3, m1, m5
+  psraw                m0, 4
+  punpcklbw            m1, m5
+%endif
+%if %2 == 1 ; avg
+  ; FIXME(rbultje) pipeline
+  packuswb             m0, m2
+  pavgb                m0, [secq]
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%endif
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+  mova                 m0, m4
+
+  INC_SRC_BY_SRC_STRIDE
+  add                dstq, dst_strideq
+%else ; %1 < 16
+  movx                 m0, [srcq]
+  movx                 m1, [srcq+1]
+%if cpuflag(ssse3)
+  punpcklbw            m0, m1
+  pmaddubsw            m0, filter_x_a
+  paddw                m0, filter_rnd
+%else
+  punpcklbw            m0, m5
+  punpcklbw            m1, m5
+  pmullw               m0, filter_x_a
+  pmullw               m1, filter_x_b
+  paddw                m0, filter_rnd
+  paddw                m0, m1
+%endif
+  psraw                m0, 4
+%if cpuflag(ssse3)
+  packuswb             m0, m0
+%endif
+
+  INC_SRC_BY_SRC_STRIDE
+
+.x_other_y_other_loop:
+  movx                 m2, [srcq]
+  movx                 m1, [srcq+1]
+
+  INC_SRC_BY_SRC_STRIDE
+  movx                 m4, [srcq]
+  movx                 m3, [srcq+1]
+
+%if cpuflag(ssse3)
+  punpcklbw            m2, m1
+  punpcklbw            m4, m3
+  pmaddubsw            m2, filter_x_a
+  pmaddubsw            m4, filter_x_a
+  movx                 m3, [dstq+dst_strideq]
+  movx                 m1, [dstq]
+  paddw                m2, filter_rnd
+  paddw                m4, filter_rnd
+  psraw                m2, 4
+  psraw                m4, 4
+  packuswb             m2, m2
+  packuswb             m4, m4
+  punpcklbw            m0, m2
+  punpcklbw            m2, m4
+  pmaddubsw            m0, filter_y_a
+  pmaddubsw            m2, filter_y_a
+  punpcklbw            m3, m5
+  paddw                m0, filter_rnd
+  paddw                m2, filter_rnd
+  psraw                m0, 4
+  psraw                m2, 4
+  punpcklbw            m1, m5
+%else
+  punpcklbw            m2, m5
+  punpcklbw            m1, m5
+  punpcklbw            m4, m5
+  punpcklbw            m3, m5
+  pmullw               m2, filter_x_a
+  pmullw               m1, filter_x_b
+  paddw                m2, filter_rnd
+  pmullw               m4, filter_x_a
+  pmullw               m3, filter_x_b
+  paddw                m4, filter_rnd
+  paddw                m2, m1
+  paddw                m4, m3
+  psraw                m2, 4
+  psraw                m4, 4
+  pmullw               m0, filter_y_a
+  pmullw               m3, m2, filter_y_b
+  paddw                m0, filter_rnd
+  pmullw               m2, filter_y_a
+  pmullw               m1, m4, filter_y_b
+  paddw                m2, filter_rnd
+  paddw                m0, m3
+  movx                 m3, [dstq+dst_strideq]
+  paddw                m2, m1
+  movx                 m1, [dstq]
+  psraw                m0, 4
+  psraw                m2, 4
+  punpcklbw            m3, m5
+  punpcklbw            m1, m5
+%endif
+%if %2 == 1 ; avg
+  ; FIXME(rbultje) pipeline
+%if %1 == 4
+  movlhps              m0, m2
+%endif
+  packuswb             m0, m2
+%if %1 > 4
+  pavgb                m0, [secq]
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%else
+  movh                 m2, [secq]
+  pavgb                m0, m2
+  punpcklbw            m0, m5
+  movhlps              m2, m0
+%endif
+%endif
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+  mova                 m0, m4
+
+  INC_SRC_BY_SRC_STRIDE
+  lea                dstq, [dstq+dst_strideq*2]
+%endif
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+  dec                   block_height
+  jg .x_other_y_other_loop
+%undef filter_x_a
+%undef filter_x_b
+%undef filter_y_a
+%undef filter_y_b
+%undef filter_rnd
+%undef movx
+  STORE_AND_RET %1
+%endmacro
+
+; FIXME(rbultje) the non-bilinear versions (i.e. x=0,8&&y=0,8) are identical
+; between the ssse3 and non-ssse3 version. It may make sense to merge their
+; code in the sense that the ssse3 version would jump to the appropriate
+; location in the sse/2 version, rather than duplicating that code in the
+; binary.
+
+INIT_XMM sse2
+SUBPEL_VARIANCE  4
+SUBPEL_VARIANCE  8
+SUBPEL_VARIANCE 16
+
+INIT_XMM ssse3
+SUBPEL_VARIANCE  4
+SUBPEL_VARIANCE  8
+SUBPEL_VARIANCE 16
+
+INIT_XMM sse2
+SUBPEL_VARIANCE  4, 1
+SUBPEL_VARIANCE  8, 1
+SUBPEL_VARIANCE 16, 1
+
+INIT_XMM ssse3
+SUBPEL_VARIANCE  4, 1
+SUBPEL_VARIANCE  8, 1
+SUBPEL_VARIANCE 16, 1
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 0000000000..b4c5cc7c7b
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/subtract_avx2.c
@@ -0,0 +1,109 @@
+/*
+ * 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 set_one_minusone = _mm256_set1_epi32((int)0xff01ff01);
+  __m256i diff0 = _mm256_unpacklo_epi8(s, p);
+  __m256i diff1 = _mm256_unpackhi_epi8(s, p);
+  diff0 = _mm256_maddubs_epi16(diff0, set_one_minusone);
+  diff1 = _mm256_maddubs_epi16(diff1, set_one_minusone);
+  _mm256_store_si256((__m256i *)(diff_ptr),
+                     _mm256_permute2x128_si256(diff0, diff1, 0x20));
+  _mm256_store_si256((__m256i *)(diff_ptr + 16),
+                     _mm256_permute2x128_si256(diff0, diff1, 0x31));
+}
+
+static INLINE void 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 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 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 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:
+      subtract_block_16xn_avx2(rows, diff_ptr, diff_stride, src_ptr, src_stride,
+                               pred_ptr, pred_stride);
+      break;
+    case 32:
+      subtract_block_32xn_avx2(rows, diff_ptr, diff_stride, src_ptr, src_stride,
+                               pred_ptr, pred_stride);
+      break;
+    case 64:
+      subtract_block_64xn_avx2(rows, diff_ptr, diff_stride, src_ptr, src_stride,
+                               pred_ptr, pred_stride);
+      break;
+    case 128:
+      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/subtract_sse2.asm b/third_party/aom/aom_dsp/x86/subtract_sse2.asm
new file mode 100644
index 0000000000..fd508c0916
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/subtract_sse2.asm
@@ -0,0 +1,147 @@
+;
+; Copyright (c) 2016, 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 "third_party/x86inc/x86inc.asm"
+
+SECTION .text
+
+; void aom_subtract_block(int rows, int cols,
+;                         int16_t *diff, ptrdiff_t diff_stride,
+;                         const uint8_t *src, ptrdiff_t src_stride,
+;                         const uint8_t *pred, ptrdiff_t pred_stride)
+
+INIT_XMM sse2
+cglobal subtract_block, 7, 7, 8, \
+                        rows, cols, diff, diff_stride, src, src_stride, \
+                        pred, pred_stride
+%define pred_str colsq
+  pxor                  m7, m7         ; dedicated zero register
+  cmp                colsd, 4
+  je .case_4
+  cmp                colsd, 8
+  je .case_8
+  cmp                colsd, 16
+  je .case_16
+  cmp                colsd, 32
+  je .case_32
+  cmp                colsd, 64
+  je .case_64
+
+%macro loop16 6
+  mova                  m0, [srcq+%1]
+  mova                  m4, [srcq+%2]
+  movu                  m1, [predq+%3]
+  movu                  m5, [predq+%4]
+  punpckhbw             m2, m0, m7
+  punpckhbw             m3, m1, m7
+  punpcklbw             m0, m7
+  punpcklbw             m1, m7
+  psubw                 m2, m3
+  psubw                 m0, m1
+  punpckhbw             m1, m4, m7
+  punpckhbw             m3, m5, m7
+  punpcklbw             m4, m7
+  punpcklbw             m5, m7
+  psubw                 m1, m3
+  psubw                 m4, m5
+  mova [diffq+mmsize*0+%5], m0
+  mova [diffq+mmsize*1+%5], m2
+  mova [diffq+mmsize*0+%6], m4
+  mova [diffq+mmsize*1+%6], m1
+%endmacro
+
+  mov             pred_str, pred_stridemp
+.loop_128:
+  loop16 0*mmsize, 1*mmsize, 0*mmsize, 1*mmsize,  0*mmsize,  2*mmsize
+  loop16 2*mmsize, 3*mmsize, 2*mmsize, 3*mmsize,  4*mmsize,  6*mmsize
+  loop16 4*mmsize, 5*mmsize, 4*mmsize, 5*mmsize,  8*mmsize, 10*mmsize
+  loop16 6*mmsize, 7*mmsize, 6*mmsize, 7*mmsize, 12*mmsize, 14*mmsize
+  lea                diffq, [diffq+diff_strideq*2]
+  add                predq, pred_str
+  add                 srcq, src_strideq
+  sub                rowsd, 1
+  jnz .loop_128
+  RET
+
+.case_64:
+  mov             pred_str, pred_stridemp
+.loop_64:
+  loop16 0*mmsize, 1*mmsize, 0*mmsize, 1*mmsize, 0*mmsize, 2*mmsize
+  loop16 2*mmsize, 3*mmsize, 2*mmsize, 3*mmsize, 4*mmsize, 6*mmsize
+  lea                diffq, [diffq+diff_strideq*2]
+  add                predq, pred_str
+  add                 srcq, src_strideq
+  dec                rowsd
+  jg .loop_64
+  RET
+
+.case_32:
+  mov             pred_str, pred_stridemp
+.loop_32:
+  loop16 0, mmsize, 0, mmsize, 0, 2*mmsize
+  lea                diffq, [diffq+diff_strideq*2]
+  add                predq, pred_str
+  add                 srcq, src_strideq
+  dec                rowsd
+  jg .loop_32
+  RET
+
+.case_16:
+  mov             pred_str, pred_stridemp
+.loop_16:
+  loop16 0, src_strideq, 0, pred_str, 0, diff_strideq*2
+  lea                diffq, [diffq+diff_strideq*4]
+  lea                predq, [predq+pred_str*2]
+  lea                 srcq, [srcq+src_strideq*2]
+  sub                rowsd, 2
+  jg .loop_16
+  RET
+
+%macro loop_h 0
+  movh                  m0, [srcq]
+  movh                  m2, [srcq+src_strideq]
+  movh                  m1, [predq]
+  movh                  m3, [predq+pred_str]
+  punpcklbw             m0, m7
+  punpcklbw             m1, m7
+  punpcklbw             m2, m7
+  punpcklbw             m3, m7
+  psubw                 m0, m1
+  psubw                 m2, m3
+  mova             [diffq], m0
+  mova [diffq+diff_strideq*2], m2
+%endmacro
+
+.case_8:
+  mov             pred_str, pred_stridemp
+.loop_8:
+  loop_h
+  lea                diffq, [diffq+diff_strideq*4]
+  lea                 srcq, [srcq+src_strideq*2]
+  lea                predq, [predq+pred_str*2]
+  sub                rowsd, 2
+  jg .loop_8
+  RET
+
+INIT_MMX
+.case_4:
+  mov             pred_str, pred_stridemp
+.loop_4:
+  loop_h
+  lea                diffq, [diffq+diff_strideq*4]
+  lea                 srcq, [srcq+src_strideq*2]
+  lea                predq, [predq+pred_str*2]
+  sub                rowsd, 2
+  jg .loop_4
+  emms
+  RET
diff --git a/third_party/aom/aom_dsp/x86/sum_squares_avx2.c b/third_party/aom/aom_dsp/x86/sum_squares_avx2.c
new file mode 100644
index 0000000000..89b9b824bf
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/sum_squares_avx2.c
@@ -0,0 +1,326 @@
+/*
+ * 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 <smmintrin.h>
+
+#include "aom_dsp/x86/synonyms.h"
+#include "aom_dsp/x86/synonyms_avx2.h"
+#include "aom_dsp/x86/sum_squares_sse2.h"
+#include "config/aom_dsp_rtcd.h"
+
+static uint64_t aom_sum_squares_2d_i16_nxn_avx2(const int16_t *src, int stride,
+                                                int width, int height) {
+  uint64_t result;
+  __m256i v_acc_q = _mm256_setzero_si256();
+  const __m256i v_zext_mask_q = yy_set1_64_from_32i(~0);
+  for (int col = 0; col < height; col += 4) {
+    __m256i v_acc_d = _mm256_setzero_si256();
+    for (int row = 0; row < width; row += 16) {
+      const int16_t *tempsrc = src + row;
+      const __m256i v_val_0_w =
+          _mm256_loadu_si256((const __m256i *)(tempsrc + 0 * stride));
+      const __m256i v_val_1_w =
+          _mm256_loadu_si256((const __m256i *)(tempsrc + 1 * stride));
+      const __m256i v_val_2_w =
+          _mm256_loadu_si256((const __m256i *)(tempsrc + 2 * stride));
+      const __m256i v_val_3_w =
+          _mm256_loadu_si256((const __m256i *)(tempsrc + 3 * stride));
+
+      const __m256i v_sq_0_d = _mm256_madd_epi16(v_val_0_w, v_val_0_w);
+      const __m256i v_sq_1_d = _mm256_madd_epi16(v_val_1_w, v_val_1_w);
+      const __m256i v_sq_2_d = _mm256_madd_epi16(v_val_2_w, v_val_2_w);
+      const __m256i v_sq_3_d = _mm256_madd_epi16(v_val_3_w, v_val_3_w);
+
+      const __m256i v_sum_01_d = _mm256_add_epi32(v_sq_0_d, v_sq_1_d);
+      const __m256i v_sum_23_d = _mm256_add_epi32(v_sq_2_d, v_sq_3_d);
+      const __m256i v_sum_0123_d = _mm256_add_epi32(v_sum_01_d, v_sum_23_d);
+
+      v_acc_d = _mm256_add_epi32(v_acc_d, v_sum_0123_d);
+    }
+    v_acc_q =
+        _mm256_add_epi64(v_acc_q, _mm256_and_si256(v_acc_d, v_zext_mask_q));
+    v_acc_q = _mm256_add_epi64(v_acc_q, _mm256_srli_epi64(v_acc_d, 32));
+    src += 4 * stride;
+  }
+  __m128i lower_64_2_Value = _mm256_castsi256_si128(v_acc_q);
+  __m128i higher_64_2_Value = _mm256_extracti128_si256(v_acc_q, 1);
+  __m128i result_64_2_int = _mm_add_epi64(lower_64_2_Value, higher_64_2_Value);
+
+  result_64_2_int = _mm_add_epi64(
+      result_64_2_int, _mm_unpackhi_epi64(result_64_2_int, result_64_2_int));
+
+  xx_storel_64(&result, result_64_2_int);
+
+  return result;
+}
+
+uint64_t aom_sum_squares_2d_i16_avx2(const int16_t *src, int stride, int width,
+                                     int height) {
+  if (LIKELY(width == 4 && height == 4)) {
+    return aom_sum_squares_2d_i16_4x4_sse2(src, stride);
+  } else if (LIKELY(width == 4 && (height & 3) == 0)) {
+    return aom_sum_squares_2d_i16_4xn_sse2(src, stride, height);
+  } else if (LIKELY(width == 8 && (height & 3) == 0)) {
+    return aom_sum_squares_2d_i16_nxn_sse2(src, stride, width, height);
+  } else if (LIKELY(((width & 15) == 0) && ((height & 3) == 0))) {
+    return aom_sum_squares_2d_i16_nxn_avx2(src, stride, width, height);
+  } else {
+    return aom_sum_squares_2d_i16_c(src, stride, width, height);
+  }
+}
+
+static uint64_t aom_sum_sse_2d_i16_nxn_avx2(const int16_t *src, int stride,
+                                            int width, int height, int *sum) {
+  uint64_t result;
+  const __m256i zero_reg = _mm256_setzero_si256();
+  const __m256i one_reg = _mm256_set1_epi16(1);
+
+  __m256i v_sse_total = zero_reg;
+  __m256i v_sum_total = zero_reg;
+
+  for (int col = 0; col < height; col += 4) {
+    __m256i v_sse_row = zero_reg;
+    for (int row = 0; row < width; row += 16) {
+      const int16_t *tempsrc = src + row;
+      const __m256i v_val_0_w =
+          _mm256_loadu_si256((const __m256i *)(tempsrc + 0 * stride));
+      const __m256i v_val_1_w =
+          _mm256_loadu_si256((const __m256i *)(tempsrc + 1 * stride));
+      const __m256i v_val_2_w =
+          _mm256_loadu_si256((const __m256i *)(tempsrc + 2 * stride));
+      const __m256i v_val_3_w =
+          _mm256_loadu_si256((const __m256i *)(tempsrc + 3 * stride));
+
+      const __m256i v_sum_01 = _mm256_add_epi16(v_val_0_w, v_val_1_w);
+      const __m256i v_sum_23 = _mm256_add_epi16(v_val_2_w, v_val_3_w);
+      __m256i v_sum_0123 = _mm256_add_epi16(v_sum_01, v_sum_23);
+      v_sum_0123 = _mm256_madd_epi16(v_sum_0123, one_reg);
+      v_sum_total = _mm256_add_epi32(v_sum_total, v_sum_0123);
+
+      const __m256i v_sq_0_d = _mm256_madd_epi16(v_val_0_w, v_val_0_w);
+      const __m256i v_sq_1_d = _mm256_madd_epi16(v_val_1_w, v_val_1_w);
+      const __m256i v_sq_2_d = _mm256_madd_epi16(v_val_2_w, v_val_2_w);
+      const __m256i v_sq_3_d = _mm256_madd_epi16(v_val_3_w, v_val_3_w);
+      const __m256i v_sq_01_d = _mm256_add_epi32(v_sq_0_d, v_sq_1_d);
+      const __m256i v_sq_23_d = _mm256_add_epi32(v_sq_2_d, v_sq_3_d);
+      const __m256i v_sq_0123_d = _mm256_add_epi32(v_sq_01_d, v_sq_23_d);
+      v_sse_row = _mm256_add_epi32(v_sse_row, v_sq_0123_d);
+    }
+    const __m256i v_sse_row_low = _mm256_unpacklo_epi32(v_sse_row, zero_reg);
+    const __m256i v_sse_row_hi = _mm256_unpackhi_epi32(v_sse_row, zero_reg);
+    v_sse_row = _mm256_add_epi64(v_sse_row_low, v_sse_row_hi);
+    v_sse_total = _mm256_add_epi64(v_sse_total, v_sse_row);
+    src += 4 * stride;
+  }
+
+  const __m128i v_sum_total_low = _mm256_castsi256_si128(v_sum_total);
+  const __m128i v_sum_total_hi = _mm256_extracti128_si256(v_sum_total, 1);
+  __m128i sum_128bit = _mm_add_epi32(v_sum_total_hi, v_sum_total_low);
+  sum_128bit = _mm_add_epi32(sum_128bit, _mm_srli_si128(sum_128bit, 8));
+  sum_128bit = _mm_add_epi32(sum_128bit, _mm_srli_si128(sum_128bit, 4));
+  *sum += _mm_cvtsi128_si32(sum_128bit);
+
+  __m128i v_sse_total_lo = _mm256_castsi256_si128(v_sse_total);
+  __m128i v_sse_total_hi = _mm256_extracti128_si256(v_sse_total, 1);
+  __m128i sse_128bit = _mm_add_epi64(v_sse_total_lo, v_sse_total_hi);
+
+  sse_128bit =
+      _mm_add_epi64(sse_128bit, _mm_unpackhi_epi64(sse_128bit, sse_128bit));
+
+  xx_storel_64(&result, sse_128bit);
+
+  return result;
+}
+
+uint64_t aom_sum_sse_2d_i16_avx2(const int16_t *src, int src_stride, int width,
+                                 int height, int *sum) {
+  if (LIKELY(width == 4 && height == 4)) {
+    return aom_sum_sse_2d_i16_4x4_sse2(src, src_stride, sum);
+  } else if (LIKELY(width == 4 && (height & 3) == 0)) {
+    return aom_sum_sse_2d_i16_4xn_sse2(src, src_stride, height, sum);
+  } else if (LIKELY(width == 8 && (height & 3) == 0)) {
+    return aom_sum_sse_2d_i16_nxn_sse2(src, src_stride, width, height, sum);
+  } else if (LIKELY(((width & 15) == 0) && ((height & 3) == 0))) {
+    return aom_sum_sse_2d_i16_nxn_avx2(src, src_stride, width, height, sum);
+  } else {
+    return aom_sum_sse_2d_i16_c(src, src_stride, width, height, sum);
+  }
+}
+
+// Accumulate sum of 16-bit elements in the vector
+static AOM_INLINE int32_t mm256_accumulate_epi16(__m256i vec_a) {
+  __m128i vtmp1 = _mm256_extracti128_si256(vec_a, 1);
+  __m128i vtmp2 = _mm256_castsi256_si128(vec_a);
+  vtmp1 = _mm_add_epi16(vtmp1, vtmp2);
+  vtmp2 = _mm_srli_si128(vtmp1, 8);
+  vtmp1 = _mm_add_epi16(vtmp1, vtmp2);
+  vtmp2 = _mm_srli_si128(vtmp1, 4);
+  vtmp1 = _mm_add_epi16(vtmp1, vtmp2);
+  vtmp2 = _mm_srli_si128(vtmp1, 2);
+  vtmp1 = _mm_add_epi16(vtmp1, vtmp2);
+  return _mm_extract_epi16(vtmp1, 0);
+}
+
+// Accumulate sum of 32-bit elements in the vector
+static AOM_INLINE int32_t mm256_accumulate_epi32(__m256i vec_a) {
+  __m128i vtmp1 = _mm256_extracti128_si256(vec_a, 1);
+  __m128i vtmp2 = _mm256_castsi256_si128(vec_a);
+  vtmp1 = _mm_add_epi32(vtmp1, vtmp2);
+  vtmp2 = _mm_srli_si128(vtmp1, 8);
+  vtmp1 = _mm_add_epi32(vtmp1, vtmp2);
+  vtmp2 = _mm_srli_si128(vtmp1, 4);
+  vtmp1 = _mm_add_epi32(vtmp1, vtmp2);
+  return _mm_cvtsi128_si32(vtmp1);
+}
+
+uint64_t aom_var_2d_u8_avx2(uint8_t *src, int src_stride, int width,
+                            int height) {
+  uint8_t *srcp;
+  uint64_t s = 0, ss = 0;
+  __m256i vzero = _mm256_setzero_si256();
+  __m256i v_acc_sum = vzero;
+  __m256i v_acc_sqs = vzero;
+  int i, j;
+
+  // Process 32 elements in a row
+  for (i = 0; i < width - 31; i += 32) {
+    srcp = src + i;
+    // Process 8 columns at a time
+    for (j = 0; j < height - 7; j += 8) {
+      __m256i vsrc[8];
+      for (int k = 0; k < 8; k++) {
+        vsrc[k] = _mm256_loadu_si256((__m256i *)srcp);
+        srcp += src_stride;
+      }
+      for (int k = 0; k < 8; k++) {
+        __m256i vsrc0 = _mm256_unpacklo_epi8(vsrc[k], vzero);
+        __m256i vsrc1 = _mm256_unpackhi_epi8(vsrc[k], vzero);
+        v_acc_sum = _mm256_add_epi16(v_acc_sum, vsrc0);
+        v_acc_sum = _mm256_add_epi16(v_acc_sum, vsrc1);
+
+        __m256i vsqs0 = _mm256_madd_epi16(vsrc0, vsrc0);
+        __m256i vsqs1 = _mm256_madd_epi16(vsrc1, vsrc1);
+        v_acc_sqs = _mm256_add_epi32(v_acc_sqs, vsqs0);
+        v_acc_sqs = _mm256_add_epi32(v_acc_sqs, vsqs1);
+      }
+
+      // Update total sum and clear the vectors
+      s += mm256_accumulate_epi16(v_acc_sum);
+      ss += mm256_accumulate_epi32(v_acc_sqs);
+      v_acc_sum = vzero;
+      v_acc_sqs = vzero;
+    }
+
+    // Process remaining rows (height not a multiple of 8)
+    for (; j < height; j++) {
+      __m256i vsrc = _mm256_loadu_si256((__m256i *)srcp);
+      __m256i vsrc0 = _mm256_unpacklo_epi8(vsrc, vzero);
+      __m256i vsrc1 = _mm256_unpackhi_epi8(vsrc, vzero);
+      v_acc_sum = _mm256_add_epi16(v_acc_sum, vsrc0);
+      v_acc_sum = _mm256_add_epi16(v_acc_sum, vsrc1);
+
+      __m256i vsqs0 = _mm256_madd_epi16(vsrc0, vsrc0);
+      __m256i vsqs1 = _mm256_madd_epi16(vsrc1, vsrc1);
+      v_acc_sqs = _mm256_add_epi32(v_acc_sqs, vsqs0);
+      v_acc_sqs = _mm256_add_epi32(v_acc_sqs, vsqs1);
+
+      srcp += src_stride;
+    }
+
+    // Update total sum and clear the vectors
+    s += mm256_accumulate_epi16(v_acc_sum);
+    ss += mm256_accumulate_epi32(v_acc_sqs);
+    v_acc_sum = vzero;
+    v_acc_sqs = vzero;
+  }
+
+  // Process the remaining area using C
+  srcp = src;
+  for (int k = 0; k < height; k++) {
+    for (int m = i; m < width; m++) {
+      uint8_t val = srcp[m];
+      s += val;
+      ss += val * val;
+    }
+    srcp += src_stride;
+  }
+  return (ss - s * s / (width * height));
+}
+
+uint64_t aom_var_2d_u16_avx2(uint8_t *src, int src_stride, int width,
+                             int height) {
+  uint16_t *srcp1 = CONVERT_TO_SHORTPTR(src), *srcp;
+  uint64_t s = 0, ss = 0;
+  __m256i vzero = _mm256_setzero_si256();
+  __m256i v_acc_sum = vzero;
+  __m256i v_acc_sqs = vzero;
+  int i, j;
+
+  // Process 16 elements in a row
+  for (i = 0; i < width - 15; i += 16) {
+    srcp = srcp1 + i;
+    // Process 8 columns at a time
+    for (j = 0; j < height - 8; j += 8) {
+      __m256i vsrc[8];
+      for (int k = 0; k < 8; k++) {
+        vsrc[k] = _mm256_loadu_si256((__m256i *)srcp);
+        srcp += src_stride;
+      }
+      for (int k = 0; k < 8; k++) {
+        __m256i vsrc0 = _mm256_unpacklo_epi16(vsrc[k], vzero);
+        __m256i vsrc1 = _mm256_unpackhi_epi16(vsrc[k], vzero);
+        v_acc_sum = _mm256_add_epi32(vsrc0, v_acc_sum);
+        v_acc_sum = _mm256_add_epi32(vsrc1, v_acc_sum);
+
+        __m256i vsqs0 = _mm256_madd_epi16(vsrc[k], vsrc[k]);
+        v_acc_sqs = _mm256_add_epi32(v_acc_sqs, vsqs0);
+      }
+
+      // Update total sum and clear the vectors
+      s += mm256_accumulate_epi32(v_acc_sum);
+      ss += mm256_accumulate_epi32(v_acc_sqs);
+      v_acc_sum = vzero;
+      v_acc_sqs = vzero;
+    }
+
+    // Process remaining rows (height not a multiple of 8)
+    for (; j < height; j++) {
+      __m256i vsrc = _mm256_loadu_si256((__m256i *)srcp);
+      __m256i vsrc0 = _mm256_unpacklo_epi16(vsrc, vzero);
+      __m256i vsrc1 = _mm256_unpackhi_epi16(vsrc, vzero);
+      v_acc_sum = _mm256_add_epi32(vsrc0, v_acc_sum);
+      v_acc_sum = _mm256_add_epi32(vsrc1, v_acc_sum);
+
+      __m256i vsqs0 = _mm256_madd_epi16(vsrc, vsrc);
+      v_acc_sqs = _mm256_add_epi32(v_acc_sqs, vsqs0);
+      srcp += src_stride;
+    }
+
+    // Update total sum and clear the vectors
+    s += mm256_accumulate_epi32(v_acc_sum);
+    ss += mm256_accumulate_epi32(v_acc_sqs);
+    v_acc_sum = vzero;
+    v_acc_sqs = vzero;
+  }
+
+  // Process the remaining area using C
+  srcp = srcp1;
+  for (int k = 0; k < height; k++) {
+    for (int m = i; m < width; m++) {
+      uint16_t val = srcp[m];
+      s += val;
+      ss += val * val;
+    }
+    srcp += src_stride;
+  }
+  return (ss - s * s / (width * height));
+}
diff --git a/third_party/aom/aom_dsp/x86/sum_squares_sse2.c b/third_party/aom/aom_dsp/x86/sum_squares_sse2.c
new file mode 100644
index 0000000000..cf3ed98974
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/sum_squares_sse2.c
@@ -0,0 +1,478 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <emmintrin.h>
+#include <stdio.h>
+
+#include "aom_dsp/x86/synonyms.h"
+#include "aom_dsp/x86/sum_squares_sse2.h"
+#include "config/aom_dsp_rtcd.h"
+
+static INLINE __m128i xx_loadh_64(__m128i a, const void *b) {
+  const __m128d ad = _mm_castsi128_pd(a);
+  return _mm_castpd_si128(_mm_loadh_pd(ad, (double *)b));
+}
+
+static INLINE uint64_t xx_cvtsi128_si64(__m128i a) {
+#if AOM_ARCH_X86_64
+  return (uint64_t)_mm_cvtsi128_si64(a);
+#else
+  {
+    uint64_t tmp;
+    _mm_storel_epi64((__m128i *)&tmp, a);
+    return tmp;
+  }
+#endif
+}
+
+static INLINE __m128i sum_squares_i16_4x4_sse2(const int16_t *src, int stride) {
+  const __m128i v_val_0_w = xx_loadl_64(src + 0 * stride);
+  const __m128i v_val_2_w = xx_loadl_64(src + 2 * stride);
+  const __m128i v_val_01_w = xx_loadh_64(v_val_0_w, src + 1 * stride);
+  const __m128i v_val_23_w = xx_loadh_64(v_val_2_w, src + 3 * stride);
+  const __m128i v_sq_01_d = _mm_madd_epi16(v_val_01_w, v_val_01_w);
+  const __m128i v_sq_23_d = _mm_madd_epi16(v_val_23_w, v_val_23_w);
+
+  return _mm_add_epi32(v_sq_01_d, v_sq_23_d);
+}
+
+uint64_t aom_sum_squares_2d_i16_4x4_sse2(const int16_t *src, int stride) {
+  const __m128i v_sum_0123_d = sum_squares_i16_4x4_sse2(src, stride);
+  __m128i v_sum_d =
+      _mm_add_epi32(v_sum_0123_d, _mm_srli_epi64(v_sum_0123_d, 32));
+  v_sum_d = _mm_add_epi32(v_sum_d, _mm_srli_si128(v_sum_d, 8));
+  return (uint64_t)_mm_cvtsi128_si32(v_sum_d);
+}
+
+uint64_t aom_sum_sse_2d_i16_4x4_sse2(const int16_t *src, int stride, int *sum) {
+  const __m128i one_reg = _mm_set1_epi16(1);
+  const __m128i v_val_0_w = xx_loadl_64(src + 0 * stride);
+  const __m128i v_val_2_w = xx_loadl_64(src + 2 * stride);
+  __m128i v_val_01_w = xx_loadh_64(v_val_0_w, src + 1 * stride);
+  __m128i v_val_23_w = xx_loadh_64(v_val_2_w, src + 3 * stride);
+
+  __m128i v_sum_0123_d = _mm_add_epi16(v_val_01_w, v_val_23_w);
+  v_sum_0123_d = _mm_madd_epi16(v_sum_0123_d, one_reg);
+  v_sum_0123_d = _mm_add_epi32(v_sum_0123_d, _mm_srli_si128(v_sum_0123_d, 8));
+  v_sum_0123_d = _mm_add_epi32(v_sum_0123_d, _mm_srli_si128(v_sum_0123_d, 4));
+  *sum = _mm_cvtsi128_si32(v_sum_0123_d);
+
+  const __m128i v_sq_01_d = _mm_madd_epi16(v_val_01_w, v_val_01_w);
+  const __m128i v_sq_23_d = _mm_madd_epi16(v_val_23_w, v_val_23_w);
+  __m128i v_sq_0123_d = _mm_add_epi32(v_sq_01_d, v_sq_23_d);
+  v_sq_0123_d = _mm_add_epi32(v_sq_0123_d, _mm_srli_si128(v_sq_0123_d, 8));
+  v_sq_0123_d = _mm_add_epi32(v_sq_0123_d, _mm_srli_si128(v_sq_0123_d, 4));
+  return (uint64_t)_mm_cvtsi128_si32(v_sq_0123_d);
+}
+
+uint64_t aom_sum_squares_2d_i16_4xn_sse2(const int16_t *src, int stride,
+                                         int height) {
+  int r = 0;
+  __m128i v_acc_q = _mm_setzero_si128();
+  do {
+    const __m128i v_acc_d = sum_squares_i16_4x4_sse2(src, stride);
+    v_acc_q = _mm_add_epi32(v_acc_q, v_acc_d);
+    src += stride << 2;
+    r += 4;
+  } while (r < height);
+  const __m128i v_zext_mask_q = xx_set1_64_from_32i(~0);
+  __m128i v_acc_64 = _mm_add_epi64(_mm_srli_epi64(v_acc_q, 32),
+                                   _mm_and_si128(v_acc_q, v_zext_mask_q));
+  v_acc_64 = _mm_add_epi64(v_acc_64, _mm_srli_si128(v_acc_64, 8));
+  return xx_cvtsi128_si64(v_acc_64);
+}
+
+uint64_t aom_sum_sse_2d_i16_4xn_sse2(const int16_t *src, int stride, int height,
+                                     int *sum) {
+  int r = 0;
+  uint64_t sse = 0;
+  do {
+    int curr_sum = 0;
+    sse += aom_sum_sse_2d_i16_4x4_sse2(src, stride, &curr_sum);
+    *sum += curr_sum;
+    src += stride << 2;
+    r += 4;
+  } while (r < height);
+  return sse;
+}
+
+#ifdef __GNUC__
+// This prevents GCC/Clang from inlining this function into
+// aom_sum_squares_2d_i16_sse2, which in turn saves some stack
+// maintenance instructions in the common case of 4x4.
+__attribute__((noinline))
+#endif
+uint64_t
+aom_sum_squares_2d_i16_nxn_sse2(const int16_t *src, int stride, int width,
+                                int height) {
+  int r = 0;
+
+  const __m128i v_zext_mask_q = xx_set1_64_from_32i(~0);
+  __m128i v_acc_q = _mm_setzero_si128();
+
+  do {
+    __m128i v_acc_d = _mm_setzero_si128();
+    int c = 0;
+    do {
+      const int16_t *b = src + c;
+
+      const __m128i v_val_0_w = xx_load_128(b + 0 * stride);
+      const __m128i v_val_1_w = xx_load_128(b + 1 * stride);
+      const __m128i v_val_2_w = xx_load_128(b + 2 * stride);
+      const __m128i v_val_3_w = xx_load_128(b + 3 * stride);
+
+      const __m128i v_sq_0_d = _mm_madd_epi16(v_val_0_w, v_val_0_w);
+      const __m128i v_sq_1_d = _mm_madd_epi16(v_val_1_w, v_val_1_w);
+      const __m128i v_sq_2_d = _mm_madd_epi16(v_val_2_w, v_val_2_w);
+      const __m128i v_sq_3_d = _mm_madd_epi16(v_val_3_w, v_val_3_w);
+
+      const __m128i v_sum_01_d = _mm_add_epi32(v_sq_0_d, v_sq_1_d);
+      const __m128i v_sum_23_d = _mm_add_epi32(v_sq_2_d, v_sq_3_d);
+
+      const __m128i v_sum_0123_d = _mm_add_epi32(v_sum_01_d, v_sum_23_d);
+
+      v_acc_d = _mm_add_epi32(v_acc_d, v_sum_0123_d);
+      c += 8;
+    } while (c < width);
+
+    v_acc_q = _mm_add_epi64(v_acc_q, _mm_and_si128(v_acc_d, v_zext_mask_q));
+    v_acc_q = _mm_add_epi64(v_acc_q, _mm_srli_epi64(v_acc_d, 32));
+
+    src += 4 * stride;
+    r += 4;
+  } while (r < height);
+
+  v_acc_q = _mm_add_epi64(v_acc_q, _mm_srli_si128(v_acc_q, 8));
+  return xx_cvtsi128_si64(v_acc_q);
+}
+
+#ifdef __GNUC__
+// This prevents GCC/Clang from inlining this function into
+// aom_sum_sse_2d_i16_nxn_sse2, which in turn saves some stack
+// maintenance instructions in the common case of 4x4.
+__attribute__((noinline))
+#endif
+uint64_t
+aom_sum_sse_2d_i16_nxn_sse2(const int16_t *src, int stride, int width,
+                            int height, int *sum) {
+  int r = 0;
+  uint64_t result;
+  const __m128i zero_reg = _mm_setzero_si128();
+  const __m128i one_reg = _mm_set1_epi16(1);
+
+  __m128i v_sse_total = zero_reg;
+  __m128i v_sum_total = zero_reg;
+
+  do {
+    int c = 0;
+    __m128i v_sse_row = zero_reg;
+    do {
+      const int16_t *b = src + c;
+
+      __m128i v_val_0_w = xx_load_128(b + 0 * stride);
+      __m128i v_val_1_w = xx_load_128(b + 1 * stride);
+      __m128i v_val_2_w = xx_load_128(b + 2 * stride);
+      __m128i v_val_3_w = xx_load_128(b + 3 * stride);
+
+      const __m128i v_sq_0_d = _mm_madd_epi16(v_val_0_w, v_val_0_w);
+      const __m128i v_sq_1_d = _mm_madd_epi16(v_val_1_w, v_val_1_w);
+      const __m128i v_sq_2_d = _mm_madd_epi16(v_val_2_w, v_val_2_w);
+      const __m128i v_sq_3_d = _mm_madd_epi16(v_val_3_w, v_val_3_w);
+      const __m128i v_sq_01_d = _mm_add_epi32(v_sq_0_d, v_sq_1_d);
+      const __m128i v_sq_23_d = _mm_add_epi32(v_sq_2_d, v_sq_3_d);
+      const __m128i v_sq_0123_d = _mm_add_epi32(v_sq_01_d, v_sq_23_d);
+      v_sse_row = _mm_add_epi32(v_sse_row, v_sq_0123_d);
+
+      const __m128i v_sum_01 = _mm_add_epi16(v_val_0_w, v_val_1_w);
+      const __m128i v_sum_23 = _mm_add_epi16(v_val_2_w, v_val_3_w);
+      __m128i v_sum_0123_d = _mm_add_epi16(v_sum_01, v_sum_23);
+      v_sum_0123_d = _mm_madd_epi16(v_sum_0123_d, one_reg);
+      v_sum_total = _mm_add_epi32(v_sum_total, v_sum_0123_d);
+
+      c += 8;
+    } while (c < width);
+
+    const __m128i v_sse_row_low = _mm_unpacklo_epi32(v_sse_row, zero_reg);
+    const __m128i v_sse_row_hi = _mm_unpackhi_epi32(v_sse_row, zero_reg);
+    v_sse_row = _mm_add_epi64(v_sse_row_low, v_sse_row_hi);
+    v_sse_total = _mm_add_epi64(v_sse_total, v_sse_row);
+    src += 4 * stride;
+    r += 4;
+  } while (r < height);
+
+  v_sum_total = _mm_add_epi32(v_sum_total, _mm_srli_si128(v_sum_total, 8));
+  v_sum_total = _mm_add_epi32(v_sum_total, _mm_srli_si128(v_sum_total, 4));
+  *sum += _mm_cvtsi128_si32(v_sum_total);
+
+  v_sse_total = _mm_add_epi64(v_sse_total, _mm_srli_si128(v_sse_total, 8));
+  xx_storel_64(&result, v_sse_total);
+  return result;
+}
+
+uint64_t aom_sum_squares_2d_i16_sse2(const int16_t *src, int stride, int width,
+                                     int height) {
+  // 4 elements per row only requires half an XMM register, so this
+  // must be a special case, but also note that over 75% of all calls
+  // are with size == 4, so it is also the common case.
+  if (LIKELY(width == 4 && height == 4)) {
+    return aom_sum_squares_2d_i16_4x4_sse2(src, stride);
+  } else if (LIKELY(width == 4 && (height & 3) == 0)) {
+    return aom_sum_squares_2d_i16_4xn_sse2(src, stride, height);
+  } else if (LIKELY((width & 7) == 0 && (height & 3) == 0)) {
+    // Generic case
+    return aom_sum_squares_2d_i16_nxn_sse2(src, stride, width, height);
+  } else {
+    return aom_sum_squares_2d_i16_c(src, stride, width, height);
+  }
+}
+
+uint64_t aom_sum_sse_2d_i16_sse2(const int16_t *src, int src_stride, int width,
+                                 int height, int *sum) {
+  if (LIKELY(width == 4 && height == 4)) {
+    return aom_sum_sse_2d_i16_4x4_sse2(src, src_stride, sum);
+  } else if (LIKELY(width == 4 && (height & 3) == 0)) {
+    return aom_sum_sse_2d_i16_4xn_sse2(src, src_stride, height, sum);
+  } else if (LIKELY((width & 7) == 0 && (height & 3) == 0)) {
+    // Generic case
+    return aom_sum_sse_2d_i16_nxn_sse2(src, src_stride, width, height, sum);
+  } else {
+    return aom_sum_sse_2d_i16_c(src, src_stride, width, height, sum);
+  }
+}
+
+//////////////////////////////////////////////////////////////////////////////
+// 1D version
+//////////////////////////////////////////////////////////////////////////////
+
+static uint64_t aom_sum_squares_i16_64n_sse2(const int16_t *src, uint32_t n) {
+  const __m128i v_zext_mask_q = xx_set1_64_from_32i(~0);
+  __m128i v_acc0_q = _mm_setzero_si128();
+  __m128i v_acc1_q = _mm_setzero_si128();
+
+  const int16_t *const end = src + n;
+
+  assert(n % 64 == 0);
+
+  while (src < end) {
+    const __m128i v_val_0_w = xx_load_128(src);
+    const __m128i v_val_1_w = xx_load_128(src + 8);
+    const __m128i v_val_2_w = xx_load_128(src + 16);
+    const __m128i v_val_3_w = xx_load_128(src + 24);
+    const __m128i v_val_4_w = xx_load_128(src + 32);
+    const __m128i v_val_5_w = xx_load_128(src + 40);
+    const __m128i v_val_6_w = xx_load_128(src + 48);
+    const __m128i v_val_7_w = xx_load_128(src + 56);
+
+    const __m128i v_sq_0_d = _mm_madd_epi16(v_val_0_w, v_val_0_w);
+    const __m128i v_sq_1_d = _mm_madd_epi16(v_val_1_w, v_val_1_w);
+    const __m128i v_sq_2_d = _mm_madd_epi16(v_val_2_w, v_val_2_w);
+    const __m128i v_sq_3_d = _mm_madd_epi16(v_val_3_w, v_val_3_w);
+    const __m128i v_sq_4_d = _mm_madd_epi16(v_val_4_w, v_val_4_w);
+    const __m128i v_sq_5_d = _mm_madd_epi16(v_val_5_w, v_val_5_w);
+    const __m128i v_sq_6_d = _mm_madd_epi16(v_val_6_w, v_val_6_w);
+    const __m128i v_sq_7_d = _mm_madd_epi16(v_val_7_w, v_val_7_w);
+
+    const __m128i v_sum_01_d = _mm_add_epi32(v_sq_0_d, v_sq_1_d);
+    const __m128i v_sum_23_d = _mm_add_epi32(v_sq_2_d, v_sq_3_d);
+    const __m128i v_sum_45_d = _mm_add_epi32(v_sq_4_d, v_sq_5_d);
+    const __m128i v_sum_67_d = _mm_add_epi32(v_sq_6_d, v_sq_7_d);
+
+    const __m128i v_sum_0123_d = _mm_add_epi32(v_sum_01_d, v_sum_23_d);
+    const __m128i v_sum_4567_d = _mm_add_epi32(v_sum_45_d, v_sum_67_d);
+
+    const __m128i v_sum_d = _mm_add_epi32(v_sum_0123_d, v_sum_4567_d);
+
+    v_acc0_q = _mm_add_epi64(v_acc0_q, _mm_and_si128(v_sum_d, v_zext_mask_q));
+    v_acc1_q = _mm_add_epi64(v_acc1_q, _mm_srli_epi64(v_sum_d, 32));
+
+    src += 64;
+  }
+
+  v_acc0_q = _mm_add_epi64(v_acc0_q, v_acc1_q);
+  v_acc0_q = _mm_add_epi64(v_acc0_q, _mm_srli_si128(v_acc0_q, 8));
+  return xx_cvtsi128_si64(v_acc0_q);
+}
+
+uint64_t aom_sum_squares_i16_sse2(const int16_t *src, uint32_t n) {
+  if (n % 64 == 0) {
+    return aom_sum_squares_i16_64n_sse2(src, n);
+  } else if (n > 64) {
+    const uint32_t k = n & ~63u;
+    return aom_sum_squares_i16_64n_sse2(src, k) +
+           aom_sum_squares_i16_c(src + k, n - k);
+  } else {
+    return aom_sum_squares_i16_c(src, n);
+  }
+}
+
+// Accumulate sum of 16-bit elements in the vector
+static AOM_INLINE int32_t mm_accumulate_epi16(__m128i vec_a) {
+  __m128i vtmp = _mm_srli_si128(vec_a, 8);
+  vec_a = _mm_add_epi16(vec_a, vtmp);
+  vtmp = _mm_srli_si128(vec_a, 4);
+  vec_a = _mm_add_epi16(vec_a, vtmp);
+  vtmp = _mm_srli_si128(vec_a, 2);
+  vec_a = _mm_add_epi16(vec_a, vtmp);
+  return _mm_extract_epi16(vec_a, 0);
+}
+
+// Accumulate sum of 32-bit elements in the vector
+static AOM_INLINE int32_t mm_accumulate_epi32(__m128i vec_a) {
+  __m128i vtmp = _mm_srli_si128(vec_a, 8);
+  vec_a = _mm_add_epi32(vec_a, vtmp);
+  vtmp = _mm_srli_si128(vec_a, 4);
+  vec_a = _mm_add_epi32(vec_a, vtmp);
+  return _mm_cvtsi128_si32(vec_a);
+}
+
+uint64_t aom_var_2d_u8_sse2(uint8_t *src, int src_stride, int width,
+                            int height) {
+  uint8_t *srcp;
+  uint64_t s = 0, ss = 0;
+  __m128i vzero = _mm_setzero_si128();
+  __m128i v_acc_sum = vzero;
+  __m128i v_acc_sqs = vzero;
+  int i, j;
+
+  // Process 16 elements in a row
+  for (i = 0; i < width - 15; i += 16) {
+    srcp = src + i;
+    // Process 8 columns at a time
+    for (j = 0; j < height - 7; j += 8) {
+      __m128i vsrc[8];
+      for (int k = 0; k < 8; k++) {
+        vsrc[k] = _mm_loadu_si128((__m128i *)srcp);
+        srcp += src_stride;
+      }
+      for (int k = 0; k < 8; k++) {
+        __m128i vsrc0 = _mm_unpacklo_epi8(vsrc[k], vzero);
+        __m128i vsrc1 = _mm_unpackhi_epi8(vsrc[k], vzero);
+        v_acc_sum = _mm_add_epi16(v_acc_sum, vsrc0);
+        v_acc_sum = _mm_add_epi16(v_acc_sum, vsrc1);
+
+        __m128i vsqs0 = _mm_madd_epi16(vsrc0, vsrc0);
+        __m128i vsqs1 = _mm_madd_epi16(vsrc1, vsrc1);
+        v_acc_sqs = _mm_add_epi32(v_acc_sqs, vsqs0);
+        v_acc_sqs = _mm_add_epi32(v_acc_sqs, vsqs1);
+      }
+
+      // Update total sum and clear the vectors
+      s += mm_accumulate_epi16(v_acc_sum);
+      ss += mm_accumulate_epi32(v_acc_sqs);
+      v_acc_sum = vzero;
+      v_acc_sqs = vzero;
+    }
+
+    // Process remaining rows (height not a multiple of 8)
+    for (; j < height; j++) {
+      __m128i vsrc = _mm_loadu_si128((__m128i *)srcp);
+      __m128i vsrc0 = _mm_unpacklo_epi8(vsrc, vzero);
+      __m128i vsrc1 = _mm_unpackhi_epi8(vsrc, vzero);
+      v_acc_sum = _mm_add_epi16(v_acc_sum, vsrc0);
+      v_acc_sum = _mm_add_epi16(v_acc_sum, vsrc1);
+
+      __m128i vsqs0 = _mm_madd_epi16(vsrc0, vsrc0);
+      __m128i vsqs1 = _mm_madd_epi16(vsrc1, vsrc1);
+      v_acc_sqs = _mm_add_epi32(v_acc_sqs, vsqs0);
+      v_acc_sqs = _mm_add_epi32(v_acc_sqs, vsqs1);
+
+      srcp += src_stride;
+    }
+
+    // Update total sum and clear the vectors
+    s += mm_accumulate_epi16(v_acc_sum);
+    ss += mm_accumulate_epi32(v_acc_sqs);
+    v_acc_sum = vzero;
+    v_acc_sqs = vzero;
+  }
+
+  // Process the remaining area using C
+  srcp = src;
+  for (int k = 0; k < height; k++) {
+    for (int m = i; m < width; m++) {
+      uint8_t val = srcp[m];
+      s += val;
+      ss += val * val;
+    }
+    srcp += src_stride;
+  }
+  return (ss - s * s / (width * height));
+}
+
+uint64_t aom_var_2d_u16_sse2(uint8_t *src, int src_stride, int width,
+                             int height) {
+  uint16_t *srcp1 = CONVERT_TO_SHORTPTR(src), *srcp;
+  uint64_t s = 0, ss = 0;
+  __m128i vzero = _mm_setzero_si128();
+  __m128i v_acc_sum = vzero;
+  __m128i v_acc_sqs = vzero;
+  int i, j;
+
+  // Process 8 elements in a row
+  for (i = 0; i < width - 8; i += 8) {
+    srcp = srcp1 + i;
+    // Process 8 columns at a time
+    for (j = 0; j < height - 8; j += 8) {
+      __m128i vsrc[8];
+      for (int k = 0; k < 8; k++) {
+        vsrc[k] = _mm_loadu_si128((__m128i *)srcp);
+        srcp += src_stride;
+      }
+      for (int k = 0; k < 8; k++) {
+        __m128i vsrc0 = _mm_unpacklo_epi16(vsrc[k], vzero);
+        __m128i vsrc1 = _mm_unpackhi_epi16(vsrc[k], vzero);
+        v_acc_sum = _mm_add_epi32(vsrc0, v_acc_sum);
+        v_acc_sum = _mm_add_epi32(vsrc1, v_acc_sum);
+
+        __m128i vsqs0 = _mm_madd_epi16(vsrc[k], vsrc[k]);
+        v_acc_sqs = _mm_add_epi32(v_acc_sqs, vsqs0);
+      }
+
+      // Update total sum and clear the vectors
+      s += mm_accumulate_epi32(v_acc_sum);
+      ss += mm_accumulate_epi32(v_acc_sqs);
+      v_acc_sum = vzero;
+      v_acc_sqs = vzero;
+    }
+
+    // Process remaining rows (height not a multiple of 8)
+    for (; j < height; j++) {
+      __m128i vsrc = _mm_loadu_si128((__m128i *)srcp);
+      __m128i vsrc0 = _mm_unpacklo_epi16(vsrc, vzero);
+      __m128i vsrc1 = _mm_unpackhi_epi16(vsrc, vzero);
+      v_acc_sum = _mm_add_epi32(vsrc0, v_acc_sum);
+      v_acc_sum = _mm_add_epi32(vsrc1, v_acc_sum);
+
+      __m128i vsqs0 = _mm_madd_epi16(vsrc, vsrc);
+      v_acc_sqs = _mm_add_epi32(v_acc_sqs, vsqs0);
+      srcp += src_stride;
+    }
+
+    // Update total sum and clear the vectors
+    s += mm_accumulate_epi32(v_acc_sum);
+    ss += mm_accumulate_epi32(v_acc_sqs);
+    v_acc_sum = vzero;
+    v_acc_sqs = vzero;
+  }
+
+  // Process the remaining area using C
+  srcp = srcp1;
+  for (int k = 0; k < height; k++) {
+    for (int m = i; m < width; m++) {
+      uint16_t val = srcp[m];
+      s += val;
+      ss += val * val;
+    }
+    srcp += src_stride;
+  }
+  return (ss - s * s / (width * height));
+}
diff --git a/third_party/aom/aom_dsp/x86/sum_squares_sse2.h b/third_party/aom/aom_dsp/x86/sum_squares_sse2.h
new file mode 100644
index 0000000000..5ed3f2c7bf
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/sum_squares_sse2.h
@@ -0,0 +1,28 @@
+/*
+ * 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_SUM_SQUARES_SSE2_H_
+#define AOM_DSP_X86_SUM_SQUARES_SSE2_H_
+
+uint64_t aom_sum_squares_2d_i16_nxn_sse2(const int16_t *src, int stride,
+                                         int width, int height);
+
+uint64_t aom_sum_squares_2d_i16_4xn_sse2(const int16_t *src, int stride,
+                                         int height);
+uint64_t aom_sum_squares_2d_i16_4x4_sse2(const int16_t *src, int stride);
+
+uint64_t aom_sum_sse_2d_i16_4x4_sse2(const int16_t *src, int stride, int *sum);
+uint64_t aom_sum_sse_2d_i16_4xn_sse2(const int16_t *src, int stride, int height,
+                                     int *sum);
+uint64_t aom_sum_sse_2d_i16_nxn_sse2(const int16_t *src, int stride, int width,
+                                     int height, int *sum);
+
+#endif  // AOM_DSP_X86_SUM_SQUARES_SSE2_H_
diff --git a/third_party/aom/aom_dsp/x86/synonyms.h b/third_party/aom/aom_dsp/x86/synonyms.h
new file mode 100644
index 0000000000..6744ec51d0
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/synonyms.h
@@ -0,0 +1,134 @@
+/*
+ * Copyright (c) 2016, 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_AOM_DSP_X86_SYNONYMS_H_
+#define AOM_AOM_DSP_X86_SYNONYMS_H_
+
+#include <immintrin.h>
+#include <string.h>
+
+#include "config/aom_config.h"
+
+#include "aom/aom_integer.h"
+
+/**
+ * Various reusable shorthands for x86 SIMD intrinsics.
+ *
+ * Intrinsics prefixed with xx_ operate on or return 128bit XMM registers.
+ * Intrinsics prefixed with yy_ operate on or return 256bit YMM registers.
+ */
+
+// Loads and stores to do away with the tedium of casting the address
+// to the right type.
+static INLINE __m128i xx_loadl_32(const void *a) {
+  int val;
+  memcpy(&val, a, sizeof(val));
+  return _mm_cvtsi32_si128(val);
+}
+
+static INLINE __m128i xx_loadl_64(const void *a) {
+  return _mm_loadl_epi64((const __m128i *)a);
+}
+
+static INLINE __m128i xx_load_128(const void *a) {
+  return _mm_load_si128((const __m128i *)a);
+}
+
+static INLINE __m128i xx_loadu_128(const void *a) {
+  return _mm_loadu_si128((const __m128i *)a);
+}
+
+static INLINE void xx_storel_32(void *const a, const __m128i v) {
+  const int val = _mm_cvtsi128_si32(v);
+  memcpy(a, &val, sizeof(val));
+}
+
+static INLINE void xx_storel_64(void *const a, const __m128i v) {
+  _mm_storel_epi64((__m128i *)a, v);
+}
+
+static INLINE void xx_store_128(void *const a, const __m128i v) {
+  _mm_store_si128((__m128i *)a, v);
+}
+
+static INLINE void xx_storeu_128(void *const a, const __m128i v) {
+  _mm_storeu_si128((__m128i *)a, v);
+}
+
+// The _mm_set_epi64x() intrinsic is undefined for some Visual Studio
+// compilers. The following function is equivalent to _mm_set_epi64x()
+// acting on 32-bit integers.
+static INLINE __m128i xx_set_64_from_32i(int32_t e1, int32_t e0) {
+#if defined(_MSC_VER) && _MSC_VER < 1900
+  return _mm_set_epi32(0, e1, 0, e0);
+#else
+  return _mm_set_epi64x((uint32_t)e1, (uint32_t)e0);
+#endif
+}
+
+// The _mm_set1_epi64x() intrinsic is undefined for some Visual Studio
+// compilers. The following function is equivalent to _mm_set1_epi64x()
+// acting on a 32-bit integer.
+static INLINE __m128i xx_set1_64_from_32i(int32_t a) {
+#if defined(_MSC_VER) && _MSC_VER < 1900
+  return _mm_set_epi32(0, a, 0, a);
+#else
+  return _mm_set1_epi64x((uint32_t)a);
+#endif
+}
+
+// Fill an SSE register using an interleaved pair of values, ie. set the
+// 8 channels to {a, b, a, b, a, b, a, b}, using the same channel ordering
+// as when a register is stored to / loaded from memory.
+//
+// This is useful for rearranging filter kernels for use with the _mm_madd_epi16
+// instruction
+static INLINE __m128i xx_set2_epi16(int16_t a, int16_t b) {
+  return _mm_setr_epi16(a, b, a, b, a, b, a, b);
+}
+
+static INLINE __m128i xx_round_epu16(__m128i v_val_w) {
+  return _mm_avg_epu16(v_val_w, _mm_setzero_si128());
+}
+
+static INLINE __m128i xx_roundn_epu16(__m128i v_val_w, int bits) {
+  const __m128i v_s_w = _mm_srli_epi16(v_val_w, bits - 1);
+  return _mm_avg_epu16(v_s_w, _mm_setzero_si128());
+}
+
+static INLINE __m128i xx_roundn_epu32(__m128i v_val_d, int bits) {
+  const __m128i v_bias_d = _mm_set1_epi32((1 << bits) >> 1);
+  const __m128i v_tmp_d = _mm_add_epi32(v_val_d, v_bias_d);
+  return _mm_srli_epi32(v_tmp_d, bits);
+}
+
+static INLINE __m128i xx_roundn_epi16_unsigned(__m128i v_val_d, int bits) {
+  const __m128i v_bias_d = _mm_set1_epi16((1 << bits) >> 1);
+  const __m128i v_tmp_d = _mm_add_epi16(v_val_d, v_bias_d);
+  return _mm_srai_epi16(v_tmp_d, bits);
+}
+
+// This is equivalent to ROUND_POWER_OF_TWO(v_val_d, bits)
+static INLINE __m128i xx_roundn_epi32_unsigned(__m128i v_val_d, int bits) {
+  const __m128i v_bias_d = _mm_set1_epi32((1 << bits) >> 1);
+  const __m128i v_tmp_d = _mm_add_epi32(v_val_d, v_bias_d);
+  return _mm_srai_epi32(v_tmp_d, bits);
+}
+
+static INLINE __m128i xx_roundn_epi16(__m128i v_val_d, int bits) {
+  const __m128i v_bias_d = _mm_set1_epi16((1 << bits) >> 1);
+  const __m128i v_sign_d = _mm_srai_epi16(v_val_d, 15);
+  const __m128i v_tmp_d =
+      _mm_add_epi16(_mm_add_epi16(v_val_d, v_bias_d), v_sign_d);
+  return _mm_srai_epi16(v_tmp_d, bits);
+}
+
+#endif  // AOM_AOM_DSP_X86_SYNONYMS_H_
diff --git a/third_party/aom/aom_dsp/x86/synonyms_avx2.h b/third_party/aom/aom_dsp/x86/synonyms_avx2.h
new file mode 100644
index 0000000000..b729e5f410
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/synonyms_avx2.h
@@ -0,0 +1,79 @@
+/*
+ * 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_AOM_DSP_X86_SYNONYMS_AVX2_H_
+#define AOM_AOM_DSP_X86_SYNONYMS_AVX2_H_
+
+#include <immintrin.h>
+
+#include "config/aom_config.h"
+
+#include "aom/aom_integer.h"
+
+/**
+ * Various reusable shorthands for x86 SIMD intrinsics.
+ *
+ * Intrinsics prefixed with xx_ operate on or return 128bit XMM registers.
+ * Intrinsics prefixed with yy_ operate on or return 256bit YMM registers.
+ */
+
+// Loads and stores to do away with the tedium of casting the address
+// to the right type.
+static INLINE __m256i yy_load_256(const void *a) {
+  return _mm256_load_si256((const __m256i *)a);
+}
+
+static INLINE __m256i yy_loadu_256(const void *a) {
+  return _mm256_loadu_si256((const __m256i *)a);
+}
+
+static INLINE void yy_store_256(void *const a, const __m256i v) {
+  _mm256_store_si256((__m256i *)a, v);
+}
+
+static INLINE void yy_storeu_256(void *const a, const __m256i v) {
+  _mm256_storeu_si256((__m256i *)a, v);
+}
+
+// The _mm256_set1_epi64x() intrinsic is undefined for some Visual Studio
+// compilers. The following function is equivalent to _mm256_set1_epi64x()
+// acting on a 32-bit integer.
+static INLINE __m256i yy_set1_64_from_32i(int32_t a) {
+#if defined(_MSC_VER) && defined(_M_IX86) && _MSC_VER < 1900
+  return _mm256_set_epi32(0, a, 0, a, 0, a, 0, a);
+#else
+  return _mm256_set1_epi64x((uint32_t)a);
+#endif
+}
+
+// Some compilers don't have _mm256_set_m128i defined in immintrin.h. We
+// therefore define an equivalent function using a different intrinsic.
+// ([ hi ], [ lo ]) -> [ hi ][ lo ]
+static INLINE __m256i yy_set_m128i(__m128i hi, __m128i lo) {
+  return _mm256_insertf128_si256(_mm256_castsi128_si256(lo), hi, 1);
+}
+
+static INLINE __m256i yy_loadu2_128(const void *hi, const void *lo) {
+  __m128i mhi = _mm_loadu_si128((const __m128i *)(hi));
+  __m128i mlo = _mm_loadu_si128((const __m128i *)(lo));
+  return yy_set_m128i(mhi, mlo);
+}
+
+static INLINE void yy_storeu2_128(void *hi, void *lo, const __m256i a) {
+  _mm_storeu_si128((__m128i *)hi, _mm256_extracti128_si256(a, 1));
+  _mm_storeu_si128((__m128i *)lo, _mm256_castsi256_si128(a));
+}
+
+static INLINE __m256i yy_roundn_epu16(__m256i v_val_w, int bits) {
+  const __m256i v_s_w = _mm256_srli_epi16(v_val_w, bits - 1);
+  return _mm256_avg_epu16(v_s_w, _mm256_setzero_si256());
+}
+#endif  // AOM_AOM_DSP_X86_SYNONYMS_AVX2_H_
diff --git a/third_party/aom/aom_dsp/x86/transpose_sse2.h b/third_party/aom/aom_dsp/x86/transpose_sse2.h
new file mode 100644
index 0000000000..9dab750f44
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/transpose_sse2.h
@@ -0,0 +1,424 @@
+/*
+ * 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_AOM_DSP_X86_TRANSPOSE_SSE2_H_
+#define AOM_AOM_DSP_X86_TRANSPOSE_SSE2_H_
+
+#include <emmintrin.h>  // SSE2
+
+#include "config/aom_config.h"
+
+static INLINE __m128i transpose_8bit_4x4(const __m128i *const in) {
+  // Unpack 8 bit elements. Goes from:
+  // in[0]: 00 01 02 03
+  // in[1]: 10 11 12 13
+  // in[2]: 20 21 22 23
+  // in[3]: 30 31 32 33
+  // to:
+  // a0:    00 10 01 11  02 12 03 13
+  // a1:    20 30 21 31  22 32 23 33
+  const __m128i a0 = _mm_unpacklo_epi8(in[0], in[1]);
+  const __m128i a1 = _mm_unpacklo_epi8(in[2], in[3]);
+
+  // Unpack 16 bit elements resulting in:
+  // 00 10 20 30  01 11 21 31  02 12 22 32  03 13 23 33
+  return _mm_unpacklo_epi16(a0, a1);
+}
+
+static INLINE void transpose_8bit_8x8(const __m128i *const in,
+                                      __m128i *const out) {
+  // Unpack 8 bit elements. Goes from:
+  // in[0]: 00 01 02 03 04 05 06 07
+  // in[1]: 10 11 12 13 14 15 16 17
+  // in[2]: 20 21 22 23 24 25 26 27
+  // in[3]: 30 31 32 33 34 35 36 37
+  // in[4]: 40 41 42 43 44 45 46 47
+  // in[5]: 50 51 52 53 54 55 56 57
+  // in[6]: 60 61 62 63 64 65 66 67
+  // in[7]: 70 71 72 73 74 75 76 77
+  // 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
+  const __m128i a0 = _mm_unpacklo_epi8(in[0], in[1]);
+  const __m128i a1 = _mm_unpacklo_epi8(in[2], in[3]);
+  const __m128i a2 = _mm_unpacklo_epi8(in[4], in[5]);
+  const __m128i a3 = _mm_unpacklo_epi8(in[6], in[7]);
+
+  // Unpack 16 bit elements resulting in:
+  // b0: 00 10 20 30 01 11 21 31  02 12 22 32 03 13 23 33
+  // b1: 40 50 60 70 41 51 61 71  42 52 62 72 43 53 63 73
+  // b2: 04 14 24 34 05 15 25 35  06 16 26 36 07 17 27 37
+  // b3: 44 54 64 74 45 55 65 75  46 56 66 76 47 57 67 77
+  const __m128i b0 = _mm_unpacklo_epi16(a0, a1);
+  const __m128i b1 = _mm_unpackhi_epi16(a0, a1);
+  const __m128i b2 = _mm_unpacklo_epi16(a2, a3);
+  const __m128i b3 = _mm_unpackhi_epi16(a2, a3);
+
+  // Unpack 32 bit elements resulting in:
+  // c0: 00 10 20 30 40 50 60 70  01 11 21 31 41 51 61 71
+  // c1: 02 12 22 32 42 52 62 72  03 13 23 33 43 53 63 73
+  // c2: 04 14 24 34 44 54 64 74  05 15 25 35 45 55 65 75
+  // c3: 06 16 26 36 46 56 66 76  07 17 27 37 47 57 67 77
+  const __m128i c0 = _mm_unpacklo_epi32(b0, b2);
+  const __m128i c1 = _mm_unpackhi_epi32(b0, b2);
+  const __m128i c2 = _mm_unpacklo_epi32(b1, b3);
+  const __m128i c3 = _mm_unpackhi_epi32(b1, b3);
+
+  // Unpack 64 bit elements resulting in:
+  // out[0]: 00 10 20 30 40 50 60 70
+  // out[1]: 01 11 21 31 41 51 61 71
+  // out[2]: 02 12 22 32 42 52 62 72
+  // out[3]: 03 13 23 33 43 53 63 73
+  // out[4]: 04 14 24 34 44 54 64 74
+  // out[5]: 05 15 25 35 45 55 65 75
+  // out[6]: 06 16 26 36 46 56 66 76
+  // out[7]: 07 17 27 37 47 57 67 77
+  out[0] = _mm_unpacklo_epi64(c0, c0);
+  out[1] = _mm_unpackhi_epi64(c0, c0);
+  out[2] = _mm_unpacklo_epi64(c1, c1);
+  out[3] = _mm_unpackhi_epi64(c1, c1);
+  out[4] = _mm_unpacklo_epi64(c2, c2);
+  out[5] = _mm_unpackhi_epi64(c2, c2);
+  out[6] = _mm_unpacklo_epi64(c3, c3);
+  out[7] = _mm_unpackhi_epi64(c3, c3);
+}
+
+static INLINE void transpose_16bit_4x4(const __m128i *const in,
+                                       __m128i *const out) {
+  // Unpack 16 bit elements. Goes from:
+  // in[0]: 00 01 02 03  XX XX XX XX
+  // in[1]: 10 11 12 13  XX XX XX XX
+  // in[2]: 20 21 22 23  XX XX XX XX
+  // in[3]: 30 31 32 33  XX XX XX XX
+  // to:
+  // a0:    00 10 01 11  02 12 03 13
+  // a1:    20 30 21 31  22 32 23 33
+  const __m128i a0 = _mm_unpacklo_epi16(in[0], in[1]);
+  const __m128i a1 = _mm_unpacklo_epi16(in[2], in[3]);
+
+  // Unpack 32 bit elements resulting in:
+  // out[0]: 00 10 20 30  01 11 21 31
+  // out[1]: 01 11 21 31  __ __ __ __
+  // out[2]: 02 12 22 32  03 13 23 33
+  // out[3]: 03 13 23 33  __ __ __ __
+  //
+  // Note: The high 64 bits of the output registers are shown for informational
+  // purposes only. Callers should only use the low 64 bits of the output
+  // registers. "__" indicates zeros.
+  out[0] = _mm_unpacklo_epi32(a0, a1);
+  out[1] = _mm_srli_si128(out[0], 8);
+  out[2] = _mm_unpackhi_epi32(a0, a1);
+  out[3] = _mm_srli_si128(out[2], 8);
+}
+
+static INLINE void transpose_16bit_4x8(const __m128i *const in,
+                                       __m128i *const out) {
+  // Unpack 16 bit elements. Goes from:
+  // in[0]: 00 01 02 03  XX XX XX XX
+  // in[1]: 10 11 12 13  XX XX XX XX
+  // in[2]: 20 21 22 23  XX XX XX XX
+  // in[3]: 30 31 32 33  XX XX XX XX
+  // in[4]: 40 41 42 43  XX XX XX XX
+  // in[5]: 50 51 52 53  XX XX XX XX
+  // in[6]: 60 61 62 63  XX XX XX XX
+  // in[7]: 70 71 72 73  XX XX XX XX
+  // to:
+  // a0:    00 10 01 11  02 12 03 13
+  // a1:    20 30 21 31  22 32 23 33
+  // a2:    40 50 41 51  42 52 43 53
+  // a3:    60 70 61 71  62 72 63 73
+  const __m128i a0 = _mm_unpacklo_epi16(in[0], in[1]);
+  const __m128i a1 = _mm_unpacklo_epi16(in[2], in[3]);
+  const __m128i a2 = _mm_unpacklo_epi16(in[4], in[5]);
+  const __m128i a3 = _mm_unpacklo_epi16(in[6], in[7]);
+
+  // Unpack 32 bit elements resulting in:
+  // b0: 00 10 20 30  01 11 21 31
+  // b1: 40 50 60 70  41 51 61 71
+  // b2: 02 12 22 32  03 13 23 33
+  // b3: 42 52 62 72  43 53 63 73
+  const __m128i b0 = _mm_unpacklo_epi32(a0, a1);
+  const __m128i b1 = _mm_unpacklo_epi32(a2, a3);
+  const __m128i b2 = _mm_unpackhi_epi32(a0, a1);
+  const __m128i b3 = _mm_unpackhi_epi32(a2, a3);
+
+  // Unpack 64 bit elements resulting in:
+  // out[0]: 00 10 20 30  40 50 60 70
+  // out[1]: 01 11 21 31  41 51 61 71
+  // out[2]: 02 12 22 32  42 52 62 72
+  // out[3]: 03 13 23 33  43 53 63 73
+  out[0] = _mm_unpacklo_epi64(b0, b1);
+  out[1] = _mm_unpackhi_epi64(b0, b1);
+  out[2] = _mm_unpacklo_epi64(b2, b3);
+  out[3] = _mm_unpackhi_epi64(b2, b3);
+}
+
+static INLINE void transpose_16bit_8x4(const __m128i *const in,
+                                       __m128i *const out) {
+  // Unpack 16 bit elements. Goes from:
+  // in[0]: 00 01 02 03  04 05 06 07
+  // in[1]: 10 11 12 13  14 15 16 17
+  // in[2]: 20 21 22 23  24 25 26 27
+  // in[3]: 30 31 32 33  34 35 36 37
+
+  // to:
+  // a0:    00 10 01 11  02 12 03 13
+  // a1:    20 30 21 31  22 32 23 33
+  // a4:    04 14 05 15  06 16 07 17
+  // a5:    24 34 25 35  26 36 27 37
+  const __m128i a0 = _mm_unpacklo_epi16(in[0], in[1]);
+  const __m128i a1 = _mm_unpacklo_epi16(in[2], in[3]);
+  const __m128i a4 = _mm_unpackhi_epi16(in[0], in[1]);
+  const __m128i a5 = _mm_unpackhi_epi16(in[2], in[3]);
+
+  // Unpack 32 bit elements resulting in:
+  // b0: 00 10 20 30  01 11 21 31
+  // b2: 04 14 24 34  05 15 25 35
+  // b4: 02 12 22 32  03 13 23 33
+  // b6: 06 16 26 36  07 17 27 37
+  const __m128i b0 = _mm_unpacklo_epi32(a0, a1);
+  const __m128i b2 = _mm_unpacklo_epi32(a4, a5);
+  const __m128i b4 = _mm_unpackhi_epi32(a0, a1);
+  const __m128i b6 = _mm_unpackhi_epi32(a4, a5);
+
+  // Unpack 64 bit elements resulting in:
+  // out[0]: 00 10 20 30  XX XX XX XX
+  // out[1]: 01 11 21 31  XX XX XX XX
+  // out[2]: 02 12 22 32  XX XX XX XX
+  // out[3]: 03 13 23 33  XX XX XX XX
+  // out[4]: 04 14 24 34  XX XX XX XX
+  // out[5]: 05 15 25 35  XX XX XX XX
+  // out[6]: 06 16 26 36  XX XX XX XX
+  // out[7]: 07 17 27 37  XX XX XX XX
+  const __m128i zeros = _mm_setzero_si128();
+  out[0] = _mm_unpacklo_epi64(b0, zeros);
+  out[1] = _mm_unpackhi_epi64(b0, zeros);
+  out[2] = _mm_unpacklo_epi64(b4, zeros);
+  out[3] = _mm_unpackhi_epi64(b4, zeros);
+  out[4] = _mm_unpacklo_epi64(b2, zeros);
+  out[5] = _mm_unpackhi_epi64(b2, zeros);
+  out[6] = _mm_unpacklo_epi64(b6, zeros);
+  out[7] = _mm_unpackhi_epi64(b6, zeros);
+}
+
+static INLINE void transpose_16bit_8x8(const __m128i *const in,
+                                       __m128i *const out) {
+  // Unpack 16 bit elements. Goes from:
+  // in[0]: 00 01 02 03  04 05 06 07
+  // in[1]: 10 11 12 13  14 15 16 17
+  // in[2]: 20 21 22 23  24 25 26 27
+  // in[3]: 30 31 32 33  34 35 36 37
+  // in[4]: 40 41 42 43  44 45 46 47
+  // in[5]: 50 51 52 53  54 55 56 57
+  // in[6]: 60 61 62 63  64 65 66 67
+  // in[7]: 70 71 72 73  74 75 76 77
+  // to:
+  // a0:    00 10 01 11  02 12 03 13
+  // a1:    20 30 21 31  22 32 23 33
+  // a2:    40 50 41 51  42 52 43 53
+  // a3:    60 70 61 71  62 72 63 73
+  // a4:    04 14 05 15  06 16 07 17
+  // a5:    24 34 25 35  26 36 27 37
+  // a6:    44 54 45 55  46 56 47 57
+  // a7:    64 74 65 75  66 76 67 77
+  const __m128i a0 = _mm_unpacklo_epi16(in[0], in[1]);
+  const __m128i a1 = _mm_unpacklo_epi16(in[2], in[3]);
+  const __m128i a2 = _mm_unpacklo_epi16(in[4], in[5]);
+  const __m128i a3 = _mm_unpacklo_epi16(in[6], in[7]);
+  const __m128i a4 = _mm_unpackhi_epi16(in[0], in[1]);
+  const __m128i a5 = _mm_unpackhi_epi16(in[2], in[3]);
+  const __m128i a6 = _mm_unpackhi_epi16(in[4], in[5]);
+  const __m128i a7 = _mm_unpackhi_epi16(in[6], in[7]);
+
+  // Unpack 32 bit elements resulting in:
+  // b0: 00 10 20 30  01 11 21 31
+  // b1: 40 50 60 70  41 51 61 71
+  // b2: 04 14 24 34  05 15 25 35
+  // b3: 44 54 64 74  45 55 65 75
+  // b4: 02 12 22 32  03 13 23 33
+  // b5: 42 52 62 72  43 53 63 73
+  // b6: 06 16 26 36  07 17 27 37
+  // b7: 46 56 66 76  47 57 67 77
+  const __m128i b0 = _mm_unpacklo_epi32(a0, a1);
+  const __m128i b1 = _mm_unpacklo_epi32(a2, a3);
+  const __m128i b2 = _mm_unpacklo_epi32(a4, a5);
+  const __m128i b3 = _mm_unpacklo_epi32(a6, a7);
+  const __m128i b4 = _mm_unpackhi_epi32(a0, a1);
+  const __m128i b5 = _mm_unpackhi_epi32(a2, a3);
+  const __m128i b6 = _mm_unpackhi_epi32(a4, a5);
+  const __m128i b7 = _mm_unpackhi_epi32(a6, a7);
+
+  // Unpack 64 bit elements resulting in:
+  // out[0]: 00 10 20 30  40 50 60 70
+  // out[1]: 01 11 21 31  41 51 61 71
+  // out[2]: 02 12 22 32  42 52 62 72
+  // out[3]: 03 13 23 33  43 53 63 73
+  // out[4]: 04 14 24 34  44 54 64 74
+  // out[5]: 05 15 25 35  45 55 65 75
+  // out[6]: 06 16 26 36  46 56 66 76
+  // out[7]: 07 17 27 37  47 57 67 77
+  out[0] = _mm_unpacklo_epi64(b0, b1);
+  out[1] = _mm_unpackhi_epi64(b0, b1);
+  out[2] = _mm_unpacklo_epi64(b4, b5);
+  out[3] = _mm_unpackhi_epi64(b4, b5);
+  out[4] = _mm_unpacklo_epi64(b2, b3);
+  out[5] = _mm_unpackhi_epi64(b2, b3);
+  out[6] = _mm_unpacklo_epi64(b6, b7);
+  out[7] = _mm_unpackhi_epi64(b6, b7);
+}
+
+// Transpose in-place
+static INLINE void transpose_16bit_16x16(__m128i *const left,
+                                         __m128i *const right) {
+  __m128i tbuf[8];
+  transpose_16bit_8x8(left, left);
+  transpose_16bit_8x8(right, tbuf);
+  transpose_16bit_8x8(left + 8, right);
+  transpose_16bit_8x8(right + 8, right + 8);
+
+  left[8] = tbuf[0];
+  left[9] = tbuf[1];
+  left[10] = tbuf[2];
+  left[11] = tbuf[3];
+  left[12] = tbuf[4];
+  left[13] = tbuf[5];
+  left[14] = tbuf[6];
+  left[15] = tbuf[7];
+}
+
+static INLINE void transpose_32bit_4x4(const __m128i *const in,
+                                       __m128i *const out) {
+  // Unpack 32 bit elements. Goes from:
+  // in[0]: 00 01 02 03
+  // in[1]: 10 11 12 13
+  // in[2]: 20 21 22 23
+  // in[3]: 30 31 32 33
+  // to:
+  // a0:    00 10 01 11
+  // a1:    20 30 21 31
+  // a2:    02 12 03 13
+  // a3:    22 32 23 33
+
+  const __m128i a0 = _mm_unpacklo_epi32(in[0], in[1]);
+  const __m128i a1 = _mm_unpacklo_epi32(in[2], in[3]);
+  const __m128i a2 = _mm_unpackhi_epi32(in[0], in[1]);
+  const __m128i a3 = _mm_unpackhi_epi32(in[2], in[3]);
+
+  // Unpack 64 bit elements resulting in:
+  // out[0]: 00 10 20 30
+  // out[1]: 01 11 21 31
+  // out[2]: 02 12 22 32
+  // out[3]: 03 13 23 33
+  out[0] = _mm_unpacklo_epi64(a0, a1);
+  out[1] = _mm_unpackhi_epi64(a0, a1);
+  out[2] = _mm_unpacklo_epi64(a2, a3);
+  out[3] = _mm_unpackhi_epi64(a2, a3);
+}
+
+static INLINE void transpose_32bit_4x4x2(const __m128i *const in,
+                                         __m128i *const out) {
+  // Unpack 32 bit elements. Goes from:
+  // in[0]: 00 01 02 03
+  // in[1]: 10 11 12 13
+  // in[2]: 20 21 22 23
+  // in[3]: 30 31 32 33
+  // in[4]: 04 05 06 07
+  // in[5]: 14 15 16 17
+  // in[6]: 24 25 26 27
+  // in[7]: 34 35 36 37
+  // to:
+  // a0:    00 10 01 11
+  // a1:    20 30 21 31
+  // a2:    02 12 03 13
+  // a3:    22 32 23 33
+  // a4:    04 14 05 15
+  // a5:    24 34 25 35
+  // a6:    06 16 07 17
+  // a7:    26 36 27 37
+  const __m128i a0 = _mm_unpacklo_epi32(in[0], in[1]);
+  const __m128i a1 = _mm_unpacklo_epi32(in[2], in[3]);
+  const __m128i a2 = _mm_unpackhi_epi32(in[0], in[1]);
+  const __m128i a3 = _mm_unpackhi_epi32(in[2], in[3]);
+  const __m128i a4 = _mm_unpacklo_epi32(in[4], in[5]);
+  const __m128i a5 = _mm_unpacklo_epi32(in[6], in[7]);
+  const __m128i a6 = _mm_unpackhi_epi32(in[4], in[5]);
+  const __m128i a7 = _mm_unpackhi_epi32(in[6], in[7]);
+
+  // Unpack 64 bit elements resulting in:
+  // out[0]: 00 10 20 30
+  // out[1]: 01 11 21 31
+  // out[2]: 02 12 22 32
+  // out[3]: 03 13 23 33
+  // out[4]: 04 14 24 34
+  // out[5]: 05 15 25 35
+  // out[6]: 06 16 26 36
+  // out[7]: 07 17 27 37
+  out[0] = _mm_unpacklo_epi64(a0, a1);
+  out[1] = _mm_unpackhi_epi64(a0, a1);
+  out[2] = _mm_unpacklo_epi64(a2, a3);
+  out[3] = _mm_unpackhi_epi64(a2, a3);
+  out[4] = _mm_unpacklo_epi64(a4, a5);
+  out[5] = _mm_unpackhi_epi64(a4, a5);
+  out[6] = _mm_unpacklo_epi64(a6, a7);
+  out[7] = _mm_unpackhi_epi64(a6, a7);
+}
+
+static INLINE void transpose_32bit_8x4(const __m128i *const in,
+                                       __m128i *const out) {
+  // Unpack 32 bit elements. Goes from:
+  // in[0]: 00 01 02 03
+  // in[1]: 04 05 06 07
+  // in[2]: 10 11 12 13
+  // in[3]: 14 15 16 17
+  // in[4]: 20 21 22 23
+  // in[5]: 24 25 26 27
+  // in[6]: 30 31 32 33
+  // in[7]: 34 35 36 37
+  // to:
+  // a0: 00 10 01 11
+  // a1: 20 30 21 31
+  // a2: 02 12 03 13
+  // a3: 22 32 23 33
+  // a4: 04 14 05 15
+  // a5: 24 34 25 35
+  // a6: 06 16 07 17
+  // a7: 26 36 27 37
+  const __m128i a0 = _mm_unpacklo_epi32(in[0], in[2]);
+  const __m128i a1 = _mm_unpacklo_epi32(in[4], in[6]);
+  const __m128i a2 = _mm_unpackhi_epi32(in[0], in[2]);
+  const __m128i a3 = _mm_unpackhi_epi32(in[4], in[6]);
+  const __m128i a4 = _mm_unpacklo_epi32(in[1], in[3]);
+  const __m128i a5 = _mm_unpacklo_epi32(in[5], in[7]);
+  const __m128i a6 = _mm_unpackhi_epi32(in[1], in[3]);
+  const __m128i a7 = _mm_unpackhi_epi32(in[5], in[7]);
+
+  // Unpack 64 bit elements resulting in:
+  // out[0]: 00 10 20 30
+  // out[1]: 01 11 21 31
+  // out[2]: 02 12 22 32
+  // out[3]: 03 13 23 33
+  // out[4]: 04 14 24 34
+  // out[5]: 05 15 25 35
+  // out[6]: 06 16 26 36
+  // out[7]: 07 17 27 37
+  out[0] = _mm_unpacklo_epi64(a0, a1);
+  out[1] = _mm_unpackhi_epi64(a0, a1);
+  out[2] = _mm_unpacklo_epi64(a2, a3);
+  out[3] = _mm_unpackhi_epi64(a2, a3);
+  out[4] = _mm_unpacklo_epi64(a4, a5);
+  out[5] = _mm_unpackhi_epi64(a4, a5);
+  out[6] = _mm_unpacklo_epi64(a6, a7);
+  out[7] = _mm_unpackhi_epi64(a6, a7);
+}
+
+#endif  // AOM_AOM_DSP_X86_TRANSPOSE_SSE2_H_
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 0000000000..4105250bc0
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/txfm_common_avx2.h
@@ -0,0 +1,357 @@
+/*
+ * 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_AOM_DSP_X86_TXFM_COMMON_AVX2_H_
+#define AOM_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
+
+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 transpose2_8x8_avx2(const __m256i *const in,
+                                       __m256i *const out) {
+  __m256i t[16], u[16];
+  // (1st, 2nd) ==> (lo, hi)
+  //   (0, 1)   ==>  (0, 1)
+  //   (2, 3)   ==>  (2, 3)
+  //   (4, 5)   ==>  (4, 5)
+  //   (6, 7)   ==>  (6, 7)
+  for (int i = 0; i < 4; i++) {
+    t[2 * i] = _mm256_unpacklo_epi16(in[2 * i], in[2 * i + 1]);
+    t[2 * i + 1] = _mm256_unpackhi_epi16(in[2 * i], in[2 * i + 1]);
+  }
+
+  // (1st, 2nd) ==> (lo, hi)
+  //   (0, 2)   ==>  (0, 2)
+  //   (1, 3)   ==>  (1, 3)
+  //   (4, 6)   ==>  (4, 6)
+  //   (5, 7)   ==>  (5, 7)
+  for (int i = 0; i < 2; i++) {
+    u[i] = _mm256_unpacklo_epi32(t[i], t[i + 2]);
+    u[i + 2] = _mm256_unpackhi_epi32(t[i], t[i + 2]);
+
+    u[i + 4] = _mm256_unpacklo_epi32(t[i + 4], t[i + 6]);
+    u[i + 6] = _mm256_unpackhi_epi32(t[i + 4], t[i + 6]);
+  }
+
+  // (1st, 2nd) ==> (lo, hi)
+  //   (0, 4)   ==>  (0, 1)
+  //   (1, 5)   ==>  (4, 5)
+  //   (2, 6)   ==>  (2, 3)
+  //   (3, 7)   ==>  (6, 7)
+  for (int i = 0; i < 2; i++) {
+    out[2 * i] = _mm256_unpacklo_epi64(u[2 * i], u[2 * i + 4]);
+    out[2 * i + 1] = _mm256_unpackhi_epi64(u[2 * i], u[2 * i + 4]);
+
+    out[2 * i + 4] = _mm256_unpacklo_epi64(u[2 * i + 1], u[2 * i + 5]);
+    out[2 * i + 5] = _mm256_unpackhi_epi64(u[2 * i + 1], u[2 * i + 5]);
+  }
+}
+
+static INLINE void transpose_16bit_16x16_avx2(const __m256i *const in,
+                                              __m256i *const out) {
+  __m256i t[16];
+
+#define LOADL(idx)                                                            \
+  t[idx] = _mm256_castsi128_si256(_mm_load_si128((__m128i const *)&in[idx])); \
+  t[idx] = _mm256_inserti128_si256(                                           \
+      t[idx], _mm_load_si128((__m128i const *)&in[idx + 8]), 1);
+
+#define LOADR(idx)                                                           \
+  t[8 + idx] =                                                               \
+      _mm256_castsi128_si256(_mm_load_si128((__m128i const *)&in[idx] + 1)); \
+  t[8 + idx] = _mm256_inserti128_si256(                                      \
+      t[8 + idx], _mm_load_si128((__m128i const *)&in[idx + 8] + 1), 1);
+
+  // load left 8x16
+  LOADL(0)
+  LOADL(1)
+  LOADL(2)
+  LOADL(3)
+  LOADL(4)
+  LOADL(5)
+  LOADL(6)
+  LOADL(7)
+
+  // load right 8x16
+  LOADR(0)
+  LOADR(1)
+  LOADR(2)
+  LOADR(3)
+  LOADR(4)
+  LOADR(5)
+  LOADR(6)
+  LOADR(7)
+
+  // get the top 16x8 result
+  transpose2_8x8_avx2(t, out);
+  // get the bottom 16x8 result
+  transpose2_8x8_avx2(&t[8], &out[8]);
+}
+
+static INLINE void transpose_16bit_16x8_avx2(const __m256i *const in,
+                                             __m256i *const out) {
+  const __m256i a0 = _mm256_unpacklo_epi16(in[0], in[1]);
+  const __m256i a1 = _mm256_unpacklo_epi16(in[2], in[3]);
+  const __m256i a2 = _mm256_unpacklo_epi16(in[4], in[5]);
+  const __m256i a3 = _mm256_unpacklo_epi16(in[6], in[7]);
+  const __m256i a4 = _mm256_unpackhi_epi16(in[0], in[1]);
+  const __m256i a5 = _mm256_unpackhi_epi16(in[2], in[3]);
+  const __m256i a6 = _mm256_unpackhi_epi16(in[4], in[5]);
+  const __m256i a7 = _mm256_unpackhi_epi16(in[6], in[7]);
+
+  const __m256i b0 = _mm256_unpacklo_epi32(a0, a1);
+  const __m256i b1 = _mm256_unpacklo_epi32(a2, a3);
+  const __m256i b2 = _mm256_unpacklo_epi32(a4, a5);
+  const __m256i b3 = _mm256_unpacklo_epi32(a6, a7);
+  const __m256i b4 = _mm256_unpackhi_epi32(a0, a1);
+  const __m256i b5 = _mm256_unpackhi_epi32(a2, a3);
+  const __m256i b6 = _mm256_unpackhi_epi32(a4, a5);
+  const __m256i b7 = _mm256_unpackhi_epi32(a6, a7);
+
+  out[0] = _mm256_unpacklo_epi64(b0, b1);
+  out[1] = _mm256_unpackhi_epi64(b0, b1);
+  out[2] = _mm256_unpacklo_epi64(b4, b5);
+  out[3] = _mm256_unpackhi_epi64(b4, b5);
+  out[4] = _mm256_unpacklo_epi64(b2, b3);
+  out[5] = _mm256_unpackhi_epi64(b2, b3);
+  out[6] = _mm256_unpacklo_epi64(b6, b7);
+  out[7] = _mm256_unpackhi_epi64(b6, b7);
+}
+
+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);
+    }
+  }
+}
+
+static INLINE __m256i round_shift_32_avx2(__m256i vec, int bit) {
+  __m256i tmp, round;
+  round = _mm256_set1_epi32(1 << (bit - 1));
+  tmp = _mm256_add_epi32(vec, round);
+  return _mm256_srai_epi32(tmp, bit);
+}
+
+static INLINE void round_shift_array_32_avx2(__m256i *input, __m256i *output,
+                                             const int size, const int bit) {
+  if (bit > 0) {
+    int i;
+    for (i = 0; i < size; i++) {
+      output[i] = round_shift_32_avx2(input[i], bit);
+    }
+  } else {
+    int i;
+    for (i = 0; i < size; i++) {
+      output[i] = _mm256_slli_epi32(input[i], -bit);
+    }
+  }
+}
+
+static INLINE void round_shift_rect_array_32_avx2(__m256i *input,
+                                                  __m256i *output,
+                                                  const int size, const int bit,
+                                                  const int val) {
+  const __m256i sqrt2 = _mm256_set1_epi32(val);
+  if (bit > 0) {
+    int i;
+    for (i = 0; i < size; i++) {
+      const __m256i r0 = round_shift_32_avx2(input[i], bit);
+      const __m256i r1 = _mm256_mullo_epi32(sqrt2, r0);
+      output[i] = round_shift_32_avx2(r1, NewSqrt2Bits);
+    }
+  } else {
+    int i;
+    for (i = 0; i < size; i++) {
+      const __m256i r0 = _mm256_slli_epi32(input[i], -bit);
+      const __m256i r1 = _mm256_mullo_epi32(sqrt2, r0);
+      output[i] = round_shift_32_avx2(r1, NewSqrt2Bits);
+    }
+  }
+}
+
+static INLINE __m256i scale_round_avx2(const __m256i a, const int scale) {
+  const __m256i scale_rounding =
+      pair_set_w16_epi16(scale, 1 << (NewSqrt2Bits - 1));
+  const __m256i b = _mm256_madd_epi16(a, scale_rounding);
+  return _mm256_srai_epi32(b, NewSqrt2Bits);
+}
+
+static INLINE void store_rect_16bit_to_32bit_w8_avx2(const __m256i a,
+                                                     int32_t *const b) {
+  const __m256i one = _mm256_set1_epi16(1);
+  const __m256i a_lo = _mm256_unpacklo_epi16(a, one);
+  const __m256i a_hi = _mm256_unpackhi_epi16(a, one);
+  const __m256i b_lo = scale_round_avx2(a_lo, NewSqrt2);
+  const __m256i b_hi = scale_round_avx2(a_hi, NewSqrt2);
+  const __m256i temp = _mm256_permute2f128_si256(b_lo, b_hi, 0x31);
+  _mm_store_si128((__m128i *)b, _mm256_castsi256_si128(b_lo));
+  _mm_store_si128((__m128i *)(b + 4), _mm256_castsi256_si128(b_hi));
+  _mm256_store_si256((__m256i *)(b + 64), temp);
+}
+
+static INLINE void store_rect_buffer_16bit_to_32bit_w8_avx2(
+    const __m256i *const in, int32_t *const out, const int stride,
+    const int out_size) {
+  for (int i = 0; i < out_size; ++i) {
+    store_rect_16bit_to_32bit_w8_avx2(in[i], out + i * stride);
+  }
+}
+
+static INLINE void pack_reg(const __m128i *in1, const __m128i *in2,
+                            __m256i *out) {
+  out[0] = _mm256_insertf128_si256(_mm256_castsi128_si256(in1[0]), in2[0], 0x1);
+  out[1] = _mm256_insertf128_si256(_mm256_castsi128_si256(in1[1]), in2[1], 0x1);
+  out[2] = _mm256_insertf128_si256(_mm256_castsi128_si256(in1[2]), in2[2], 0x1);
+  out[3] = _mm256_insertf128_si256(_mm256_castsi128_si256(in1[3]), in2[3], 0x1);
+  out[4] = _mm256_insertf128_si256(_mm256_castsi128_si256(in1[4]), in2[4], 0x1);
+  out[5] = _mm256_insertf128_si256(_mm256_castsi128_si256(in1[5]), in2[5], 0x1);
+  out[6] = _mm256_insertf128_si256(_mm256_castsi128_si256(in1[6]), in2[6], 0x1);
+  out[7] = _mm256_insertf128_si256(_mm256_castsi128_si256(in1[7]), in2[7], 0x1);
+}
+
+static INLINE void extract_reg(const __m256i *in, __m128i *out1) {
+  out1[0] = _mm256_castsi256_si128(in[0]);
+  out1[1] = _mm256_castsi256_si128(in[1]);
+  out1[2] = _mm256_castsi256_si128(in[2]);
+  out1[3] = _mm256_castsi256_si128(in[3]);
+  out1[4] = _mm256_castsi256_si128(in[4]);
+  out1[5] = _mm256_castsi256_si128(in[5]);
+  out1[6] = _mm256_castsi256_si128(in[6]);
+  out1[7] = _mm256_castsi256_si128(in[7]);
+
+  out1[8] = _mm256_extracti128_si256(in[0], 0x01);
+  out1[9] = _mm256_extracti128_si256(in[1], 0x01);
+  out1[10] = _mm256_extracti128_si256(in[2], 0x01);
+  out1[11] = _mm256_extracti128_si256(in[3], 0x01);
+  out1[12] = _mm256_extracti128_si256(in[4], 0x01);
+  out1[13] = _mm256_extracti128_si256(in[5], 0x01);
+  out1[14] = _mm256_extracti128_si256(in[6], 0x01);
+  out1[15] = _mm256_extracti128_si256(in[7], 0x01);
+}
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif  // AOM_AOM_DSP_X86_TXFM_COMMON_AVX2_H_
diff --git a/third_party/aom/aom_dsp/x86/txfm_common_sse2.h b/third_party/aom/aom_dsp/x86/txfm_common_sse2.h
new file mode 100644
index 0000000000..9c99eb93bd
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/txfm_common_sse2.h
@@ -0,0 +1,33 @@
+/*
+ * Copyright (c) 2016, 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_AOM_DSP_X86_TXFM_COMMON_SSE2_H_
+#define AOM_AOM_DSP_X86_TXFM_COMMON_SSE2_H_
+
+#include <emmintrin.h>
+#include "aom/aom_integer.h"
+#include "aom_dsp/x86/synonyms.h"
+
+#define pair_set_epi16(a, b) \
+  _mm_set1_epi32((int32_t)(((uint16_t)(a)) | (((uint32_t)(b)) << 16)))
+
+// Reverse the 8 16 bit words in __m128i
+static INLINE __m128i mm_reverse_epi16(const __m128i x) {
+  const __m128i a = _mm_shufflelo_epi16(x, 0x1b);
+  const __m128i b = _mm_shufflehi_epi16(a, 0x1b);
+  return _mm_shuffle_epi32(b, 0x4e);
+}
+
+#define octa_set_epi16(a, b, c, d, e, f, g, h)                           \
+  _mm_setr_epi16((int16_t)(a), (int16_t)(b), (int16_t)(c), (int16_t)(d), \
+                 (int16_t)(e), (int16_t)(f), (int16_t)(g), (int16_t)(h))
+
+#endif  // AOM_AOM_DSP_X86_TXFM_COMMON_SSE2_H_
diff --git a/third_party/aom/aom_dsp/x86/variance_avx2.c b/third_party/aom/aom_dsp/x86/variance_avx2.c
new file mode 100644
index 0000000000..046d6f10f8
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/variance_avx2.c
@@ -0,0 +1,961 @@
+/*
+ * Copyright (c) 2016, 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"
+
+#include "aom_dsp/x86/masked_variance_intrin_ssse3.h"
+#include "aom_dsp/x86/synonyms.h"
+
+static INLINE __m128i mm256_add_hi_lo_epi16(const __m256i val) {
+  return _mm_add_epi16(_mm256_castsi256_si128(val),
+                       _mm256_extractf128_si256(val, 1));
+}
+
+static INLINE __m128i mm256_add_hi_lo_epi32(const __m256i val) {
+  return _mm_add_epi32(_mm256_castsi256_si128(val),
+                       _mm256_extractf128_si256(val, 1));
+}
+
+static INLINE void variance_kernel_avx2(const __m256i src, const __m256i ref,
+                                        __m256i *const sse,
+                                        __m256i *const sum) {
+  const __m256i adj_sub = _mm256_set1_epi16((short)0xff01);  // (1,-1)
+
+  // unpack into pairs of source and reference values
+  const __m256i src_ref0 = _mm256_unpacklo_epi8(src, ref);
+  const __m256i src_ref1 = _mm256_unpackhi_epi8(src, ref);
+
+  // subtract adjacent elements using src*1 + ref*-1
+  const __m256i diff0 = _mm256_maddubs_epi16(src_ref0, adj_sub);
+  const __m256i diff1 = _mm256_maddubs_epi16(src_ref1, adj_sub);
+  const __m256i madd0 = _mm256_madd_epi16(diff0, diff0);
+  const __m256i madd1 = _mm256_madd_epi16(diff1, diff1);
+
+  // add to the running totals
+  *sum = _mm256_add_epi16(*sum, _mm256_add_epi16(diff0, diff1));
+  *sse = _mm256_add_epi32(*sse, _mm256_add_epi32(madd0, madd1));
+}
+
+static INLINE int variance_final_from_32bit_sum_avx2(__m256i vsse, __m128i vsum,
+                                                     unsigned int *const sse) {
+  // extract the low lane and add it to the high lane
+  const __m128i sse_reg_128 = mm256_add_hi_lo_epi32(vsse);
+
+  // unpack sse and sum registers and add
+  const __m128i sse_sum_lo = _mm_unpacklo_epi32(sse_reg_128, vsum);
+  const __m128i sse_sum_hi = _mm_unpackhi_epi32(sse_reg_128, vsum);
+  const __m128i sse_sum = _mm_add_epi32(sse_sum_lo, sse_sum_hi);
+
+  // perform the final summation and extract the results
+  const __m128i res = _mm_add_epi32(sse_sum, _mm_srli_si128(sse_sum, 8));
+  *((int *)sse) = _mm_cvtsi128_si32(res);
+  return _mm_extract_epi32(res, 1);
+}
+
+// handle pixels (<= 512)
+static INLINE int variance_final_512_avx2(__m256i vsse, __m256i vsum,
+                                          unsigned int *const sse) {
+  // extract the low lane and add it to the high lane
+  const __m128i vsum_128 = mm256_add_hi_lo_epi16(vsum);
+  const __m128i vsum_64 = _mm_add_epi16(vsum_128, _mm_srli_si128(vsum_128, 8));
+  const __m128i sum_int32 = _mm_cvtepi16_epi32(vsum_64);
+  return variance_final_from_32bit_sum_avx2(vsse, sum_int32, sse);
+}
+
+// handle 1024 pixels (32x32, 16x64, 64x16)
+static INLINE int variance_final_1024_avx2(__m256i vsse, __m256i vsum,
+                                           unsigned int *const sse) {
+  // extract the low lane and add it to the high lane
+  const __m128i vsum_128 = mm256_add_hi_lo_epi16(vsum);
+  const __m128i vsum_64 =
+      _mm_add_epi32(_mm_cvtepi16_epi32(vsum_128),
+                    _mm_cvtepi16_epi32(_mm_srli_si128(vsum_128, 8)));
+  return variance_final_from_32bit_sum_avx2(vsse, vsum_64, sse);
+}
+
+static INLINE __m256i sum_to_32bit_avx2(const __m256i sum) {
+  const __m256i sum_lo = _mm256_cvtepi16_epi32(_mm256_castsi256_si128(sum));
+  const __m256i sum_hi =
+      _mm256_cvtepi16_epi32(_mm256_extractf128_si256(sum, 1));
+  return _mm256_add_epi32(sum_lo, sum_hi);
+}
+
+// handle 2048 pixels (32x64, 64x32)
+static INLINE int variance_final_2048_avx2(__m256i vsse, __m256i vsum,
+                                           unsigned int *const sse) {
+  vsum = sum_to_32bit_avx2(vsum);
+  const __m128i vsum_128 = mm256_add_hi_lo_epi32(vsum);
+  return variance_final_from_32bit_sum_avx2(vsse, vsum_128, sse);
+}
+
+static INLINE void variance16_kernel_avx2(
+    const uint8_t *const src, const int src_stride, const uint8_t *const ref,
+    const int ref_stride, __m256i *const sse, __m256i *const sum) {
+  const __m128i s0 = _mm_loadu_si128((__m128i const *)(src + 0 * src_stride));
+  const __m128i s1 = _mm_loadu_si128((__m128i const *)(src + 1 * src_stride));
+  const __m128i r0 = _mm_loadu_si128((__m128i const *)(ref + 0 * ref_stride));
+  const __m128i r1 = _mm_loadu_si128((__m128i const *)(ref + 1 * ref_stride));
+  const __m256i s = _mm256_inserti128_si256(_mm256_castsi128_si256(s0), s1, 1);
+  const __m256i r = _mm256_inserti128_si256(_mm256_castsi128_si256(r0), r1, 1);
+  variance_kernel_avx2(s, r, sse, sum);
+}
+
+static INLINE void variance32_kernel_avx2(const uint8_t *const src,
+                                          const uint8_t *const ref,
+                                          __m256i *const sse,
+                                          __m256i *const sum) {
+  const __m256i s = _mm256_loadu_si256((__m256i const *)(src));
+  const __m256i r = _mm256_loadu_si256((__m256i const *)(ref));
+  variance_kernel_avx2(s, r, sse, sum);
+}
+
+static INLINE void variance16_avx2(const uint8_t *src, const int src_stride,
+                                   const uint8_t *ref, const int ref_stride,
+                                   const int h, __m256i *const vsse,
+                                   __m256i *const vsum) {
+  *vsum = _mm256_setzero_si256();
+
+  for (int i = 0; i < h; i += 2) {
+    variance16_kernel_avx2(src, src_stride, ref, ref_stride, vsse, vsum);
+    src += 2 * src_stride;
+    ref += 2 * ref_stride;
+  }
+}
+
+static INLINE void variance32_avx2(const uint8_t *src, const int src_stride,
+                                   const uint8_t *ref, const int ref_stride,
+                                   const int h, __m256i *const vsse,
+                                   __m256i *const vsum) {
+  *vsum = _mm256_setzero_si256();
+
+  for (int i = 0; i < h; i++) {
+    variance32_kernel_avx2(src, ref, vsse, vsum);
+    src += src_stride;
+    ref += ref_stride;
+  }
+}
+
+static INLINE void variance64_avx2(const uint8_t *src, const int src_stride,
+                                   const uint8_t *ref, const int ref_stride,
+                                   const int h, __m256i *const vsse,
+                                   __m256i *const vsum) {
+  *vsum = _mm256_setzero_si256();
+
+  for (int i = 0; i < h; i++) {
+    variance32_kernel_avx2(src + 0, ref + 0, vsse, vsum);
+    variance32_kernel_avx2(src + 32, ref + 32, vsse, vsum);
+    src += src_stride;
+    ref += ref_stride;
+  }
+}
+
+static INLINE void variance128_avx2(const uint8_t *src, const int src_stride,
+                                    const uint8_t *ref, const int ref_stride,
+                                    const int h, __m256i *const vsse,
+                                    __m256i *const vsum) {
+  *vsum = _mm256_setzero_si256();
+
+  for (int i = 0; i < h; i++) {
+    variance32_kernel_avx2(src + 0, ref + 0, vsse, vsum);
+    variance32_kernel_avx2(src + 32, ref + 32, vsse, vsum);
+    variance32_kernel_avx2(src + 64, ref + 64, vsse, vsum);
+    variance32_kernel_avx2(src + 96, ref + 96, vsse, vsum);
+    src += src_stride;
+    ref += ref_stride;
+  }
+}
+
+#define AOM_VAR_NO_LOOP_AVX2(bw, bh, bits, max_pixel)                         \
+  unsigned int aom_variance##bw##x##bh##_avx2(                                \
+      const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, \
+      unsigned int *sse) {                                                    \
+    __m256i vsse = _mm256_setzero_si256();                                    \
+    __m256i vsum;                                                             \
+    variance##bw##_avx2(src, src_stride, ref, ref_stride, bh, &vsse, &vsum);  \
+    const int sum = variance_final_##max_pixel##_avx2(vsse, vsum, sse);       \
+    return *sse - (uint32_t)(((int64_t)sum * sum) >> bits);                   \
+  }
+
+AOM_VAR_NO_LOOP_AVX2(16, 8, 7, 512)
+AOM_VAR_NO_LOOP_AVX2(16, 16, 8, 512)
+AOM_VAR_NO_LOOP_AVX2(16, 32, 9, 512)
+
+AOM_VAR_NO_LOOP_AVX2(32, 16, 9, 512)
+AOM_VAR_NO_LOOP_AVX2(32, 32, 10, 1024)
+AOM_VAR_NO_LOOP_AVX2(32, 64, 11, 2048)
+
+AOM_VAR_NO_LOOP_AVX2(64, 32, 11, 2048)
+
+#if !CONFIG_REALTIME_ONLY
+AOM_VAR_NO_LOOP_AVX2(64, 16, 10, 1024)
+AOM_VAR_NO_LOOP_AVX2(32, 8, 8, 512)
+AOM_VAR_NO_LOOP_AVX2(16, 64, 10, 1024)
+AOM_VAR_NO_LOOP_AVX2(16, 4, 6, 512)
+#endif
+
+#define AOM_VAR_LOOP_AVX2(bw, bh, bits, uh)                                   \
+  unsigned int aom_variance##bw##x##bh##_avx2(                                \
+      const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, \
+      unsigned int *sse) {                                                    \
+    __m256i vsse = _mm256_setzero_si256();                                    \
+    __m256i vsum = _mm256_setzero_si256();                                    \
+    for (int i = 0; i < (bh / uh); i++) {                                     \
+      __m256i vsum16;                                                         \
+      variance##bw##_avx2(src, src_stride, ref, ref_stride, uh, &vsse,        \
+                          &vsum16);                                           \
+      vsum = _mm256_add_epi32(vsum, sum_to_32bit_avx2(vsum16));               \
+      src += uh * src_stride;                                                 \
+      ref += uh * ref_stride;                                                 \
+    }                                                                         \
+    const __m128i vsum_128 = mm256_add_hi_lo_epi32(vsum);                     \
+    const int sum = variance_final_from_32bit_sum_avx2(vsse, vsum_128, sse);  \
+    return *sse - (unsigned int)(((int64_t)sum * sum) >> bits);               \
+  }
+
+AOM_VAR_LOOP_AVX2(64, 64, 12, 32)    // 64x32 * ( 64/32)
+AOM_VAR_LOOP_AVX2(64, 128, 13, 32)   // 64x32 * (128/32)
+AOM_VAR_LOOP_AVX2(128, 64, 13, 16)   // 128x16 * ( 64/16)
+AOM_VAR_LOOP_AVX2(128, 128, 14, 16)  // 128x16 * (128/16)
+
+unsigned int aom_mse16x16_avx2(const uint8_t *src, int src_stride,
+                               const uint8_t *ref, int ref_stride,
+                               unsigned int *sse) {
+  aom_variance16x16_avx2(src, src_stride, ref, ref_stride, sse);
+  return *sse;
+}
+
+static INLINE __m256i mm256_loadu2(const uint8_t *p0, const uint8_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 __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) {
+  const __m256i alpha_max = _mm256_set1_epi8(AOM_BLEND_A64_MAX_ALPHA);
+  const int16_t round_bits = 15 - AOM_BLEND_A64_ROUND_BITS;
+  const __m256i round_offset = _mm256_set1_epi16(1 << (round_bits));
+
+  const __m256i ma = _mm256_sub_epi8(alpha_max, a);
+
+  const __m256i ssAL = _mm256_unpacklo_epi8(s0, s1);
+  const __m256i aaAL = _mm256_unpacklo_epi8(a, ma);
+  const __m256i ssAH = _mm256_unpackhi_epi8(s0, s1);
+  const __m256i aaAH = _mm256_unpackhi_epi8(a, ma);
+
+  const __m256i blendAL = _mm256_maddubs_epi16(ssAL, aaAL);
+  const __m256i blendAH = _mm256_maddubs_epi16(ssAH, aaAH);
+  const __m256i roundAL = _mm256_mulhrs_epi16(blendAL, round_offset);
+  const __m256i roundAH = _mm256_mulhrs_epi16(blendAH, round_offset);
+
+  const __m256i roundA = _mm256_packus_epi16(roundAL, roundAH);
+  _mm256_storeu_si256((__m256i *)(comp_pred), roundA);
+}
+
+void aom_comp_avg_pred_avx2(uint8_t *comp_pred, const uint8_t *pred, int width,
+                            int height, const uint8_t *ref, int ref_stride) {
+  int row = 0;
+  if (width == 8) {
+    do {
+      const __m256i pred_0123 = _mm256_loadu_si256((const __m256i *)(pred));
+      const __m128i ref_0 = _mm_loadl_epi64((const __m128i *)(ref));
+      const __m128i ref_1 =
+          _mm_loadl_epi64((const __m128i *)(ref + ref_stride));
+      const __m128i ref_2 =
+          _mm_loadl_epi64((const __m128i *)(ref + 2 * ref_stride));
+      const __m128i ref_3 =
+          _mm_loadl_epi64((const __m128i *)(ref + 3 * ref_stride));
+      const __m128i ref_01 = _mm_unpacklo_epi64(ref_0, ref_1);
+      const __m128i ref_23 = _mm_unpacklo_epi64(ref_2, ref_3);
+
+      const __m256i ref_0123 =
+          _mm256_inserti128_si256(_mm256_castsi128_si256(ref_01), ref_23, 1);
+      const __m256i average = _mm256_avg_epu8(pred_0123, ref_0123);
+      _mm256_storeu_si256((__m256i *)(comp_pred), average);
+
+      row += 4;
+      pred += 32;
+      comp_pred += 32;
+      ref += 4 * ref_stride;
+    } while (row < height);
+  } else if (width == 16) {
+    do {
+      const __m256i pred_0 = _mm256_loadu_si256((const __m256i *)(pred));
+      const __m256i pred_1 = _mm256_loadu_si256((const __m256i *)(pred + 32));
+      const __m256i tmp0 =
+          _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(ref)));
+      const __m256i ref_0 = _mm256_inserti128_si256(
+          tmp0, _mm_loadu_si128((const __m128i *)(ref + ref_stride)), 1);
+      const __m256i tmp1 = _mm256_castsi128_si256(
+          _mm_loadu_si128((const __m128i *)(ref + 2 * ref_stride)));
+      const __m256i ref_1 = _mm256_inserti128_si256(
+          tmp1, _mm_loadu_si128((const __m128i *)(ref + 3 * ref_stride)), 1);
+      const __m256i average_0 = _mm256_avg_epu8(pred_0, ref_0);
+      const __m256i average_1 = _mm256_avg_epu8(pred_1, ref_1);
+      _mm256_storeu_si256((__m256i *)(comp_pred), average_0);
+      _mm256_storeu_si256((__m256i *)(comp_pred + 32), average_1);
+
+      row += 4;
+      pred += 64;
+      comp_pred += 64;
+      ref += 4 * ref_stride;
+    } while (row < height);
+  } else if (width == 32) {
+    do {
+      const __m256i pred_0 = _mm256_loadu_si256((const __m256i *)(pred));
+      const __m256i pred_1 = _mm256_loadu_si256((const __m256i *)(pred + 32));
+      const __m256i ref_0 = _mm256_loadu_si256((const __m256i *)(ref));
+      const __m256i ref_1 =
+          _mm256_loadu_si256((const __m256i *)(ref + ref_stride));
+      const __m256i average_0 = _mm256_avg_epu8(pred_0, ref_0);
+      const __m256i average_1 = _mm256_avg_epu8(pred_1, ref_1);
+      _mm256_storeu_si256((__m256i *)(comp_pred), average_0);
+      _mm256_storeu_si256((__m256i *)(comp_pred + 32), average_1);
+
+      row += 2;
+      pred += 64;
+      comp_pred += 64;
+      ref += 2 * ref_stride;
+    } while (row < height);
+  } else if (width % 64 == 0) {
+    do {
+      for (int x = 0; x < width; x += 64) {
+        const __m256i pred_0 = _mm256_loadu_si256((const __m256i *)(pred + x));
+        const __m256i pred_1 =
+            _mm256_loadu_si256((const __m256i *)(pred + x + 32));
+        const __m256i ref_0 = _mm256_loadu_si256((const __m256i *)(ref + x));
+        const __m256i ref_1 =
+            _mm256_loadu_si256((const __m256i *)(ref + x + 32));
+        const __m256i average_0 = _mm256_avg_epu8(pred_0, ref_0);
+        const __m256i average_1 = _mm256_avg_epu8(pred_1, ref_1);
+        _mm256_storeu_si256((__m256i *)(comp_pred + x), average_0);
+        _mm256_storeu_si256((__m256i *)(comp_pred + x + 32), average_1);
+      }
+      row++;
+      pred += width;
+      comp_pred += width;
+      ref += ref_stride;
+    } while (row < height);
+  } else {
+    aom_comp_avg_pred_c(comp_pred, pred, width, height, ref, ref_stride);
+  }
+}
+
+void aom_comp_mask_pred_avx2(uint8_t *comp_pred, const uint8_t *pred, int width,
+                             int height, const uint8_t *ref, int ref_stride,
+                             const uint8_t *mask, int mask_stride,
+                             int invert_mask) {
+  int i = 0;
+  const uint8_t *src0 = invert_mask ? pred : ref;
+  const uint8_t *src1 = invert_mask ? ref : pred;
+  const int stride0 = invert_mask ? width : ref_stride;
+  const int stride1 = invert_mask ? ref_stride : width;
+  if (width == 8) {
+    comp_mask_pred_8_ssse3(comp_pred, height, src0, stride0, src1, stride1,
+                           mask, mask_stride);
+  } else if (width == 16) {
+    do {
+      const __m256i sA0 = mm256_loadu2(src0 + stride0, src0);
+      const __m256i sA1 = mm256_loadu2(src1 + stride1, src1);
+      const __m256i aA = mm256_loadu2(mask + mask_stride, mask);
+      src0 += (stride0 << 1);
+      src1 += (stride1 << 1);
+      mask += (mask_stride << 1);
+      const __m256i sB0 = mm256_loadu2(src0 + stride0, src0);
+      const __m256i sB1 = mm256_loadu2(src1 + stride1, src1);
+      const __m256i aB = mm256_loadu2(mask + mask_stride, mask);
+      src0 += (stride0 << 1);
+      src1 += (stride1 << 1);
+      mask += (mask_stride << 1);
+      // comp_pred's stride == width == 16
+      comp_mask_pred_line_avx2(sA0, sA1, aA, comp_pred);
+      comp_mask_pred_line_avx2(sB0, sB1, aB, comp_pred + 32);
+      comp_pred += (16 << 2);
+      i += 4;
+    } while (i < height);
+  } else {
+    do {
+      for (int x = 0; x < width; x += 32) {
+        const __m256i sA0 = _mm256_lddqu_si256((const __m256i *)(src0 + x));
+        const __m256i sA1 = _mm256_lddqu_si256((const __m256i *)(src1 + x));
+        const __m256i aA = _mm256_lddqu_si256((const __m256i *)(mask + x));
+
+        comp_mask_pred_line_avx2(sA0, sA1, aA, comp_pred);
+        comp_pred += 32;
+      }
+      src0 += stride0;
+      src1 += stride1;
+      mask += mask_stride;
+      i++;
+    } 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(uint8_t *comp_pred8, 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);
+  uint16_t *comp_pred = CONVERT_TO_SHORTPTR(comp_pred8);
+  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 {
+    do {
+      for (int x = 0; x < width; x += 32) {
+        const __m256i s0 = _mm256_loadu_si256((const __m256i *)(src0 + x));
+        const __m256i s2 = _mm256_loadu_si256((const __m256i *)(src0 + x + 16));
+        const __m256i s1 = _mm256_loadu_si256((const __m256i *)(src1 + x));
+        const __m256i s3 = _mm256_loadu_si256((const __m256i *)(src1 + x + 16));
+
+        const __m256i m01_16 =
+            _mm256_cvtepu8_epi16(_mm_loadu_si128((const __m128i *)(mask + x)));
+        const __m256i m23_16 = _mm256_cvtepu8_epi16(
+            _mm_loadu_si128((const __m128i *)(mask + x + 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);
+
+        comp_pred += 32;
+      }
+      src0 += stride0;
+      src1 += stride1;
+      mask += mask_stride;
+      i += 1;
+    } while (i < height);
+  }
+}
+
+uint64_t aom_mse_4xh_16bit_avx2(uint8_t *dst, int dstride, uint16_t *src,
+                                int sstride, int h) {
+  uint64_t sum = 0;
+  __m128i dst0_4x8, dst1_4x8, dst2_4x8, dst3_4x8, dst_16x8;
+  __m128i src0_4x16, src1_4x16, src2_4x16, src3_4x16;
+  __m256i src0_8x16, src1_8x16, dst_16x16, src_16x16;
+  __m256i res0_4x64, res1_4x64;
+  __m256i sub_result;
+  const __m256i zeros = _mm256_broadcastsi128_si256(_mm_setzero_si128());
+  __m256i square_result = _mm256_broadcastsi128_si256(_mm_setzero_si128());
+  for (int i = 0; i < h; i += 4) {
+    dst0_4x8 = _mm_cvtsi32_si128(*(int const *)(&dst[(i + 0) * dstride]));
+    dst1_4x8 = _mm_cvtsi32_si128(*(int const *)(&dst[(i + 1) * dstride]));
+    dst2_4x8 = _mm_cvtsi32_si128(*(int const *)(&dst[(i + 2) * dstride]));
+    dst3_4x8 = _mm_cvtsi32_si128(*(int const *)(&dst[(i + 3) * dstride]));
+    dst_16x8 = _mm_unpacklo_epi64(_mm_unpacklo_epi32(dst0_4x8, dst1_4x8),
+                                  _mm_unpacklo_epi32(dst2_4x8, dst3_4x8));
+    dst_16x16 = _mm256_cvtepu8_epi16(dst_16x8);
+
+    src0_4x16 = _mm_loadl_epi64((__m128i const *)(&src[(i + 0) * sstride]));
+    src1_4x16 = _mm_loadl_epi64((__m128i const *)(&src[(i + 1) * sstride]));
+    src2_4x16 = _mm_loadl_epi64((__m128i const *)(&src[(i + 2) * sstride]));
+    src3_4x16 = _mm_loadl_epi64((__m128i const *)(&src[(i + 3) * sstride]));
+    src0_8x16 =
+        _mm256_castsi128_si256(_mm_unpacklo_epi64(src0_4x16, src1_4x16));
+    src1_8x16 =
+        _mm256_castsi128_si256(_mm_unpacklo_epi64(src2_4x16, src3_4x16));
+    src_16x16 = _mm256_permute2x128_si256(src0_8x16, src1_8x16, 0x20);
+
+    // r15 r14 r13------------r1 r0  - 16 bit
+    sub_result = _mm256_abs_epi16(_mm256_sub_epi16(src_16x16, dst_16x16));
+
+    // s7 s6 s5 s4 s3 s2 s1 s0 - 32bit
+    src_16x16 = _mm256_madd_epi16(sub_result, sub_result);
+
+    // accumulation of result
+    square_result = _mm256_add_epi32(square_result, src_16x16);
+  }
+
+  // s5 s4 s1 s0  - 64bit
+  res0_4x64 = _mm256_unpacklo_epi32(square_result, zeros);
+  // s7 s6 s3 s2 - 64bit
+  res1_4x64 = _mm256_unpackhi_epi32(square_result, zeros);
+  // r3 r2 r1 r0 - 64bit
+  res0_4x64 = _mm256_add_epi64(res0_4x64, res1_4x64);
+  // r1+r3 r2+r0 - 64bit
+  const __m128i sum_1x64 =
+      _mm_add_epi64(_mm256_castsi256_si128(res0_4x64),
+                    _mm256_extracti128_si256(res0_4x64, 1));
+  xx_storel_64(&sum, _mm_add_epi64(sum_1x64, _mm_srli_si128(sum_1x64, 8)));
+  return sum;
+}
+
+// Compute mse of four consecutive 4x4 blocks.
+// In src buffer, each 4x4 block in a 32x32 filter block is stored sequentially.
+// Hence src_blk_stride is same as block width. Whereas dst buffer is a frame
+// buffer, thus dstride is a frame level stride.
+uint64_t aom_mse_4xh_quad_16bit_avx2(uint8_t *dst, int dstride, uint16_t *src,
+                                     int src_blk_stride, int h) {
+  uint64_t sum = 0;
+  __m128i dst0_16x8, dst1_16x8, dst2_16x8, dst3_16x8;
+  __m256i dst0_16x16, dst1_16x16, dst2_16x16, dst3_16x16;
+  __m256i res0_4x64, res1_4x64;
+  __m256i sub_result_0, sub_result_1, sub_result_2, sub_result_3;
+  const __m256i zeros = _mm256_broadcastsi128_si256(_mm_setzero_si128());
+  __m256i square_result = zeros;
+  uint16_t *src_temp = src;
+
+  for (int i = 0; i < h; i += 4) {
+    dst0_16x8 = _mm_loadu_si128((__m128i *)(&dst[(i + 0) * dstride]));
+    dst1_16x8 = _mm_loadu_si128((__m128i *)(&dst[(i + 1) * dstride]));
+    dst2_16x8 = _mm_loadu_si128((__m128i *)(&dst[(i + 2) * dstride]));
+    dst3_16x8 = _mm_loadu_si128((__m128i *)(&dst[(i + 3) * dstride]));
+
+    // row0 of 1st,2nd, 3rd and 4th 4x4 blocks- d00 d10 d20 d30
+    dst0_16x16 = _mm256_cvtepu8_epi16(dst0_16x8);
+    // row1 of 1st,2nd, 3rd and 4th 4x4 blocks - d01 d11 d21 d31
+    dst1_16x16 = _mm256_cvtepu8_epi16(dst1_16x8);
+    // row2 of 1st,2nd, 3rd and 4th 4x4 blocks - d02 d12 d22 d32
+    dst2_16x16 = _mm256_cvtepu8_epi16(dst2_16x8);
+    // row3 of 1st,2nd, 3rd and 4th 4x4 blocks - d03 d13 d23 d33
+    dst3_16x16 = _mm256_cvtepu8_epi16(dst3_16x8);
+
+    // All rows of 1st 4x4 block - r00 r01 r02 r03
+    __m256i src0_16x16 = _mm256_loadu_si256((__m256i const *)(&src_temp[0]));
+    // All rows of 2nd 4x4 block - r10 r11 r12 r13
+    __m256i src1_16x16 =
+        _mm256_loadu_si256((__m256i const *)(&src_temp[src_blk_stride]));
+    // All rows of 3rd 4x4 block - r20 r21 r22 r23
+    __m256i src2_16x16 =
+        _mm256_loadu_si256((__m256i const *)(&src_temp[2 * src_blk_stride]));
+    // All rows of 4th 4x4 block - r30 r31 r32 r33
+    __m256i src3_16x16 =
+        _mm256_loadu_si256((__m256i const *)(&src_temp[3 * src_blk_stride]));
+
+    // r00 r10 r02 r12
+    __m256i tmp0_16x16 = _mm256_unpacklo_epi64(src0_16x16, src1_16x16);
+    // r01 r11 r03 r13
+    __m256i tmp1_16x16 = _mm256_unpackhi_epi64(src0_16x16, src1_16x16);
+    // r20 r30 r22 r32
+    __m256i tmp2_16x16 = _mm256_unpacklo_epi64(src2_16x16, src3_16x16);
+    // r21 r31 r23 r33
+    __m256i tmp3_16x16 = _mm256_unpackhi_epi64(src2_16x16, src3_16x16);
+
+    // r00 r10 r20 r30
+    src0_16x16 = _mm256_permute2f128_si256(tmp0_16x16, tmp2_16x16, 0x20);
+    // r01 r11 r21 r31
+    src1_16x16 = _mm256_permute2f128_si256(tmp1_16x16, tmp3_16x16, 0x20);
+    // r02 r12 r22 r32
+    src2_16x16 = _mm256_permute2f128_si256(tmp0_16x16, tmp2_16x16, 0x31);
+    // r03 r13 r23 r33
+    src3_16x16 = _mm256_permute2f128_si256(tmp1_16x16, tmp3_16x16, 0x31);
+
+    // r15 r14 r13------------r1 r0  - 16 bit
+    sub_result_0 = _mm256_abs_epi16(_mm256_sub_epi16(src0_16x16, dst0_16x16));
+    sub_result_1 = _mm256_abs_epi16(_mm256_sub_epi16(src1_16x16, dst1_16x16));
+    sub_result_2 = _mm256_abs_epi16(_mm256_sub_epi16(src2_16x16, dst2_16x16));
+    sub_result_3 = _mm256_abs_epi16(_mm256_sub_epi16(src3_16x16, dst3_16x16));
+
+    // s7 s6 s5 s4 s3 s2 s1 s0    - 32bit
+    src0_16x16 = _mm256_madd_epi16(sub_result_0, sub_result_0);
+    src1_16x16 = _mm256_madd_epi16(sub_result_1, sub_result_1);
+    src2_16x16 = _mm256_madd_epi16(sub_result_2, sub_result_2);
+    src3_16x16 = _mm256_madd_epi16(sub_result_3, sub_result_3);
+
+    // accumulation of result
+    src0_16x16 = _mm256_add_epi32(src0_16x16, src1_16x16);
+    src2_16x16 = _mm256_add_epi32(src2_16x16, src3_16x16);
+    const __m256i square_result_0 = _mm256_add_epi32(src0_16x16, src2_16x16);
+    square_result = _mm256_add_epi32(square_result, square_result_0);
+    src_temp += 16;
+  }
+
+  // s5 s4 s1 s0  - 64bit
+  res0_4x64 = _mm256_unpacklo_epi32(square_result, zeros);
+  // s7  s6  s3  s2 - 64bit
+  res1_4x64 = _mm256_unpackhi_epi32(square_result, zeros);
+  // r3 r2 r1 r0 - 64bit
+  res0_4x64 = _mm256_add_epi64(res0_4x64, res1_4x64);
+  // r1+r3 r2+r0 - 64bit
+  const __m128i sum_1x64 =
+      _mm_add_epi64(_mm256_castsi256_si128(res0_4x64),
+                    _mm256_extracti128_si256(res0_4x64, 1));
+  xx_storel_64(&sum, _mm_add_epi64(sum_1x64, _mm_srli_si128(sum_1x64, 8)));
+  return sum;
+}
+
+uint64_t aom_mse_8xh_16bit_avx2(uint8_t *dst, int dstride, uint16_t *src,
+                                int sstride, int h) {
+  uint64_t sum = 0;
+  __m128i dst0_8x8, dst1_8x8, dst3_16x8;
+  __m256i src0_8x16, src1_8x16, src_16x16, dst_16x16;
+  __m256i res0_4x64, res1_4x64;
+  __m256i sub_result;
+  const __m256i zeros = _mm256_broadcastsi128_si256(_mm_setzero_si128());
+  __m256i square_result = _mm256_broadcastsi128_si256(_mm_setzero_si128());
+
+  for (int i = 0; i < h; i += 2) {
+    dst0_8x8 = _mm_loadl_epi64((__m128i const *)(&dst[(i + 0) * dstride]));
+    dst1_8x8 = _mm_loadl_epi64((__m128i const *)(&dst[(i + 1) * dstride]));
+    dst3_16x8 = _mm_unpacklo_epi64(dst0_8x8, dst1_8x8);
+    dst_16x16 = _mm256_cvtepu8_epi16(dst3_16x8);
+
+    src0_8x16 =
+        _mm256_castsi128_si256(_mm_loadu_si128((__m128i *)&src[i * sstride]));
+    src1_8x16 = _mm256_castsi128_si256(
+        _mm_loadu_si128((__m128i *)&src[(i + 1) * sstride]));
+    src_16x16 = _mm256_permute2x128_si256(src0_8x16, src1_8x16, 0x20);
+
+    // r15 r14 r13 - - - r1 r0 - 16 bit
+    sub_result = _mm256_abs_epi16(_mm256_sub_epi16(src_16x16, dst_16x16));
+
+    // s7 s6 s5 s4 s3 s2 s1 s0 - 32bit
+    src_16x16 = _mm256_madd_epi16(sub_result, sub_result);
+
+    // accumulation of result
+    square_result = _mm256_add_epi32(square_result, src_16x16);
+  }
+
+  // s5 s4 s1 s0  - 64bit
+  res0_4x64 = _mm256_unpacklo_epi32(square_result, zeros);
+  // s7 s6 s3 s2 - 64bit
+  res1_4x64 = _mm256_unpackhi_epi32(square_result, zeros);
+  // r3 r2 r1 r0 - 64bit
+  res0_4x64 = _mm256_add_epi64(res0_4x64, res1_4x64);
+  // r1+r3 r2+r0 - 64bit
+  const __m128i sum_1x64 =
+      _mm_add_epi64(_mm256_castsi256_si128(res0_4x64),
+                    _mm256_extracti128_si256(res0_4x64, 1));
+  xx_storel_64(&sum, _mm_add_epi64(sum_1x64, _mm_srli_si128(sum_1x64, 8)));
+  return sum;
+}
+
+// Compute mse of two consecutive 8x8 blocks.
+// In src buffer, each 8x8 block in a 64x64 filter block is stored sequentially.
+// Hence src_blk_stride is same as block width. Whereas dst buffer is a frame
+// buffer, thus dstride is a frame level stride.
+uint64_t aom_mse_8xh_dual_16bit_avx2(uint8_t *dst, int dstride, uint16_t *src,
+                                     int src_blk_stride, int h) {
+  uint64_t sum = 0;
+  __m128i dst0_16x8, dst1_16x8;
+  __m256i dst0_16x16, dst1_16x16;
+  __m256i res0_4x64, res1_4x64;
+  __m256i sub_result_0, sub_result_1;
+  const __m256i zeros = _mm256_broadcastsi128_si256(_mm_setzero_si128());
+  __m256i square_result = zeros;
+  uint16_t *src_temp = src;
+
+  for (int i = 0; i < h; i += 2) {
+    dst0_16x8 = _mm_loadu_si128((__m128i *)(&dst[(i + 0) * dstride]));
+    dst1_16x8 = _mm_loadu_si128((__m128i *)(&dst[(i + 1) * dstride]));
+
+    // row0 of 1st and 2nd 8x8 block - d00 d10
+    dst0_16x16 = _mm256_cvtepu8_epi16(dst0_16x8);
+    // row1 of 1st and 2nd 8x8 block - d01 d11
+    dst1_16x16 = _mm256_cvtepu8_epi16(dst1_16x8);
+
+    // 2 rows of 1st 8x8 block - r00 r01
+    __m256i src0_16x16 = _mm256_loadu_si256((__m256i const *)(&src_temp[0]));
+    // 2 rows of 2nd 8x8 block - r10 r11
+    __m256i src1_16x16 =
+        _mm256_loadu_si256((__m256i const *)(&src_temp[src_blk_stride]));
+    // r00 r10 - 128bit
+    __m256i tmp0_16x16 =
+        _mm256_permute2f128_si256(src0_16x16, src1_16x16, 0x20);
+    // r01 r11 - 128bit
+    __m256i tmp1_16x16 =
+        _mm256_permute2f128_si256(src0_16x16, src1_16x16, 0x31);
+
+    // r15 r14 r13------------r1 r0 - 16 bit
+    sub_result_0 = _mm256_abs_epi16(_mm256_sub_epi16(tmp0_16x16, dst0_16x16));
+    sub_result_1 = _mm256_abs_epi16(_mm256_sub_epi16(tmp1_16x16, dst1_16x16));
+
+    // s7 s6 s5 s4 s3 s2 s1 s0 - 32bit each
+    src0_16x16 = _mm256_madd_epi16(sub_result_0, sub_result_0);
+    src1_16x16 = _mm256_madd_epi16(sub_result_1, sub_result_1);
+
+    // accumulation of result
+    src0_16x16 = _mm256_add_epi32(src0_16x16, src1_16x16);
+    square_result = _mm256_add_epi32(square_result, src0_16x16);
+    src_temp += 16;
+  }
+
+  // s5 s4 s1 s0  - 64bit
+  res0_4x64 = _mm256_unpacklo_epi32(square_result, zeros);
+  // s7 s6 s3 s2 - 64bit
+  res1_4x64 = _mm256_unpackhi_epi32(square_result, zeros);
+  // r3 r2 r1 r0 - 64bit
+  res0_4x64 = _mm256_add_epi64(res0_4x64, res1_4x64);
+  // r1+r3 r2+r0 - 64bit
+  const __m128i sum_1x64 =
+      _mm_add_epi64(_mm256_castsi256_si128(res0_4x64),
+                    _mm256_extracti128_si256(res0_4x64, 1));
+  xx_storel_64(&sum, _mm_add_epi64(sum_1x64, _mm_srli_si128(sum_1x64, 8)));
+  return sum;
+}
+
+uint64_t aom_mse_wxh_16bit_avx2(uint8_t *dst, int dstride, uint16_t *src,
+                                int sstride, int w, int h) {
+  assert((w == 8 || w == 4) && (h == 8 || h == 4) &&
+         "w=8/4 and h=8/4 must be satisfied");
+  switch (w) {
+    case 4: return aom_mse_4xh_16bit_avx2(dst, dstride, src, sstride, h);
+    case 8: return aom_mse_8xh_16bit_avx2(dst, dstride, src, sstride, h);
+    default: assert(0 && "unsupported width"); return -1;
+  }
+}
+
+// Computes mse of two 8x8 or four 4x4 consecutive blocks. Luma plane uses 8x8
+// block and Chroma uses 4x4 block. In src buffer, each block in a filter block
+// is stored sequentially. Hence src_blk_stride is same as block width. Whereas
+// dst buffer is a frame buffer, thus dstride is a frame level stride.
+uint64_t aom_mse_16xh_16bit_avx2(uint8_t *dst, int dstride, uint16_t *src,
+                                 int w, int h) {
+  assert((w == 8 || w == 4) && (h == 8 || h == 4) &&
+         "w=8/4 and h=8/4 must be satisfied");
+  switch (w) {
+    case 4: return aom_mse_4xh_quad_16bit_avx2(dst, dstride, src, w * h, h);
+    case 8: return aom_mse_8xh_dual_16bit_avx2(dst, dstride, src, w * h, h);
+    default: assert(0 && "unsupported width"); return -1;
+  }
+}
+
+static INLINE void calc_sum_sse_wd32_avx2(const uint8_t *src,
+                                          const uint8_t *ref,
+                                          __m256i set_one_minusone,
+                                          __m256i sse_8x16[2],
+                                          __m256i sum_8x16[2]) {
+  const __m256i s00_256 = _mm256_loadu_si256((__m256i const *)(src));
+  const __m256i r00_256 = _mm256_loadu_si256((__m256i const *)(ref));
+
+  const __m256i u_low_256 = _mm256_unpacklo_epi8(s00_256, r00_256);
+  const __m256i u_high_256 = _mm256_unpackhi_epi8(s00_256, r00_256);
+
+  const __m256i diff0 = _mm256_maddubs_epi16(u_low_256, set_one_minusone);
+  const __m256i diff1 = _mm256_maddubs_epi16(u_high_256, set_one_minusone);
+
+  sse_8x16[0] = _mm256_add_epi32(sse_8x16[0], _mm256_madd_epi16(diff0, diff0));
+  sse_8x16[1] = _mm256_add_epi32(sse_8x16[1], _mm256_madd_epi16(diff1, diff1));
+  sum_8x16[0] = _mm256_add_epi16(sum_8x16[0], diff0);
+  sum_8x16[1] = _mm256_add_epi16(sum_8x16[1], diff1);
+}
+
+static INLINE __m256i calc_sum_sse_order(__m256i *sse_hx16, __m256i *sum_hx16,
+                                         unsigned int *tot_sse, int *tot_sum) {
+  // s00 s01 s10 s11 s20 s21 s30 s31
+  const __m256i sse_results = _mm256_hadd_epi32(sse_hx16[0], sse_hx16[1]);
+  // d00 d01 d02 d03 | d10 d11 d12 d13 | d20 d21 d22 d23 | d30 d31 d32 d33
+  const __m256i sum_result_r0 = _mm256_hadd_epi16(sum_hx16[0], sum_hx16[1]);
+  // d00 d01 d10 d11 | d00 d02 d10 d11 | d20 d21 d30 d31 | d20 d21 d30 d31
+  const __m256i sum_result_1 = _mm256_hadd_epi16(sum_result_r0, sum_result_r0);
+  // d00 d01 d10 d11 d20 d21 d30 d31 | X
+  const __m256i sum_result_3 = _mm256_permute4x64_epi64(sum_result_1, 0x08);
+  // d00 d01 d10 d11 d20 d21 d30 d31
+  const __m256i sum_results =
+      _mm256_cvtepi16_epi32(_mm256_castsi256_si128(sum_result_3));
+
+  // Add sum & sse registers appropriately to get total sum & sse separately.
+  // s0 s1 d0 d1 s2 s3 d2 d3
+  const __m256i sum_sse_add = _mm256_hadd_epi32(sse_results, sum_results);
+  // s0 s1 s2 s3 d0 d1 d2 d3
+  const __m256i sum_sse_order_add = _mm256_permute4x64_epi64(sum_sse_add, 0xd8);
+  // s0+s1 s2+s3 s0+s1 s2+s3 d0+d1 d2+d3 d0+d1 d2+d3
+  const __m256i sum_sse_order_add_1 =
+      _mm256_hadd_epi32(sum_sse_order_add, sum_sse_order_add);
+  // s0 x x x | d0 x x x
+  const __m256i sum_sse_order_add_final =
+      _mm256_hadd_epi32(sum_sse_order_add_1, sum_sse_order_add_1);
+  // s0
+  const uint32_t first_value =
+      (uint32_t)_mm256_extract_epi32(sum_sse_order_add_final, 0);
+  *tot_sse += first_value;
+  // d0
+  const int second_value = _mm256_extract_epi32(sum_sse_order_add_final, 4);
+  *tot_sum += second_value;
+  return sum_sse_order_add;
+}
+
+static INLINE void get_var_sse_sum_8x8_quad_avx2(
+    const uint8_t *src, int src_stride, const uint8_t *ref,
+    const int ref_stride, const int h, uint32_t *sse8x8, int *sum8x8,
+    unsigned int *tot_sse, int *tot_sum, uint32_t *var8x8) {
+  assert(h <= 128);  // May overflow for larger height.
+  __m256i sse_8x16[2], sum_8x16[2];
+  sum_8x16[0] = _mm256_setzero_si256();
+  sse_8x16[0] = _mm256_setzero_si256();
+  sum_8x16[1] = sum_8x16[0];
+  sse_8x16[1] = sse_8x16[0];
+  const __m256i set_one_minusone = _mm256_set1_epi16((short)0xff01);
+
+  for (int i = 0; i < h; i++) {
+    // Process 8x32 block of one row.
+    calc_sum_sse_wd32_avx2(src, ref, set_one_minusone, sse_8x16, sum_8x16);
+    src += src_stride;
+    ref += ref_stride;
+  }
+
+  const __m256i sum_sse_order_add =
+      calc_sum_sse_order(sse_8x16, sum_8x16, tot_sse, tot_sum);
+
+  // s0 s1 s2 s3
+  _mm_storeu_si128((__m128i *)sse8x8,
+                   _mm256_castsi256_si128(sum_sse_order_add));
+  // d0 d1 d2 d3
+  const __m128i sum_temp8x8 = _mm256_extractf128_si256(sum_sse_order_add, 1);
+  _mm_storeu_si128((__m128i *)sum8x8, sum_temp8x8);
+
+  // (d0xd0 >> 6)=f0 (d1xd1 >> 6)=f1 (d2xd2 >> 6)=f2 (d3xd3 >> 6)=f3
+  const __m128i mull_results =
+      _mm_srli_epi32(_mm_mullo_epi32(sum_temp8x8, sum_temp8x8), 6);
+  // s0-f0=v0 s1-f1=v1 s2-f2=v2 s3-f3=v3
+  const __m128i variance_8x8 =
+      _mm_sub_epi32(_mm256_castsi256_si128(sum_sse_order_add), mull_results);
+  // v0 v1 v2 v3
+  _mm_storeu_si128((__m128i *)var8x8, variance_8x8);
+}
+
+static INLINE void get_var_sse_sum_16x16_dual_avx2(
+    const uint8_t *src, int src_stride, const uint8_t *ref,
+    const int ref_stride, const int h, uint32_t *sse16x16,
+    unsigned int *tot_sse, int *tot_sum, uint32_t *var16x16) {
+  assert(h <= 128);  // May overflow for larger height.
+  __m256i sse_16x16[2], sum_16x16[2];
+  sum_16x16[0] = _mm256_setzero_si256();
+  sse_16x16[0] = _mm256_setzero_si256();
+  sum_16x16[1] = sum_16x16[0];
+  sse_16x16[1] = sse_16x16[0];
+  const __m256i set_one_minusone = _mm256_set1_epi16((short)0xff01);
+
+  for (int i = 0; i < h; i++) {
+    // Process 16x32 block of one row.
+    calc_sum_sse_wd32_avx2(src, ref, set_one_minusone, sse_16x16, sum_16x16);
+    src += src_stride;
+    ref += ref_stride;
+  }
+
+  const __m256i sum_sse_order_add =
+      calc_sum_sse_order(sse_16x16, sum_16x16, tot_sse, tot_sum);
+
+  const __m256i sum_sse_order_add_1 =
+      _mm256_hadd_epi32(sum_sse_order_add, sum_sse_order_add);
+
+  // s0+s1 s2+s3 x x
+  _mm_storel_epi64((__m128i *)sse16x16,
+                   _mm256_castsi256_si128(sum_sse_order_add_1));
+
+  // d0+d1 d2+d3 x x
+  const __m128i sum_temp16x16 =
+      _mm256_extractf128_si256(sum_sse_order_add_1, 1);
+
+  // (d0xd0 >> 6)=f0 (d1xd1 >> 6)=f1 (d2xd2 >> 6)=f2 (d3xd3 >> 6)=f3
+  const __m128i mull_results =
+      _mm_srli_epi32(_mm_mullo_epi32(sum_temp16x16, sum_temp16x16), 8);
+
+  // s0-f0=v0 s1-f1=v1 s2-f2=v2 s3-f3=v3
+  const __m128i variance_16x16 =
+      _mm_sub_epi32(_mm256_castsi256_si128(sum_sse_order_add_1), mull_results);
+
+  // v0 v1 v2 v3
+  _mm_storel_epi64((__m128i *)var16x16, variance_16x16);
+}
+
+void aom_get_var_sse_sum_8x8_quad_avx2(const uint8_t *src_ptr,
+                                       int source_stride,
+                                       const uint8_t *ref_ptr, int ref_stride,
+                                       uint32_t *sse8x8, int *sum8x8,
+                                       unsigned int *tot_sse, int *tot_sum,
+                                       uint32_t *var8x8) {
+  get_var_sse_sum_8x8_quad_avx2(src_ptr, source_stride, ref_ptr, ref_stride, 8,
+                                sse8x8, sum8x8, tot_sse, tot_sum, var8x8);
+}
+
+void aom_get_var_sse_sum_16x16_dual_avx2(const uint8_t *src_ptr,
+                                         int source_stride,
+                                         const uint8_t *ref_ptr, int ref_stride,
+                                         uint32_t *sse16x16,
+                                         unsigned int *tot_sse, int *tot_sum,
+                                         uint32_t *var16x16) {
+  get_var_sse_sum_16x16_dual_avx2(src_ptr, source_stride, ref_ptr, ref_stride,
+                                  16, sse16x16, tot_sse, tot_sum, var16x16);
+}
diff --git a/third_party/aom/aom_dsp/x86/variance_impl_avx2.c b/third_party/aom/aom_dsp/x86/variance_impl_avx2.c
new file mode 100644
index 0000000000..9e9e70ea01
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/variance_impl_avx2.c
@@ -0,0 +1,924 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <immintrin.h>  // AVX2
+
+#include "config/aom_dsp_rtcd.h"
+
+#include "aom_ports/mem.h"
+
+/* clang-format off */
+DECLARE_ALIGNED(32, static const uint8_t, bilinear_filters_avx2[512]) = {
+  16,  0, 16,  0, 16,  0, 16,  0, 16,  0, 16,  0, 16,  0, 16,  0,
+  16,  0, 16,  0, 16,  0, 16,  0, 16,  0, 16,  0, 16,  0, 16,  0,
+  14,  2, 14,  2, 14,  2, 14,  2, 14,  2, 14,  2, 14,  2, 14,  2,
+  14,  2, 14,  2, 14,  2, 14,  2, 14,  2, 14,  2, 14,  2, 14,  2,
+  12,  4, 12,  4, 12,  4, 12,  4, 12,  4, 12,  4, 12,  4, 12,  4,
+  12,  4, 12,  4, 12,  4, 12,  4, 12,  4, 12,  4, 12,  4, 12,  4,
+  10,  6, 10,  6, 10,  6, 10,  6, 10,  6, 10,  6, 10,  6, 10,  6,
+  10,  6, 10,  6, 10,  6, 10,  6, 10,  6, 10,  6, 10,  6, 10,  6,
+   8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,
+   8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,  8,
+   6, 10,  6, 10,  6, 10,  6, 10,  6, 10,  6, 10,  6, 10,  6, 10,
+   6, 10,  6, 10,  6, 10,  6, 10,  6, 10,  6, 10,  6, 10,  6, 10,
+   4, 12,  4, 12,  4, 12,  4, 12,  4, 12,  4, 12,  4, 12,  4, 12,
+   4, 12,  4, 12,  4, 12,  4, 12,  4, 12,  4, 12,  4, 12,  4, 12,
+   2, 14,  2, 14,  2, 14,  2, 14,  2, 14,  2, 14,  2, 14,  2, 14,
+   2, 14,  2, 14,  2, 14,  2, 14,  2, 14,  2, 14,  2, 14,  2, 14,
+};
+/* clang-format on */
+
+#define FILTER_SRC(filter)                               \
+  /* filter the source */                                \
+  exp_src_lo = _mm256_maddubs_epi16(exp_src_lo, filter); \
+  exp_src_hi = _mm256_maddubs_epi16(exp_src_hi, filter); \
+                                                         \
+  /* add 8 to source */                                  \
+  exp_src_lo = _mm256_add_epi16(exp_src_lo, pw8);        \
+  exp_src_hi = _mm256_add_epi16(exp_src_hi, pw8);        \
+                                                         \
+  /* divide source by 16 */                              \
+  exp_src_lo = _mm256_srai_epi16(exp_src_lo, 4);         \
+  exp_src_hi = _mm256_srai_epi16(exp_src_hi, 4);
+
+#define MERGE_WITH_SRC(src_reg, reg)               \
+  exp_src_lo = _mm256_unpacklo_epi8(src_reg, reg); \
+  exp_src_hi = _mm256_unpackhi_epi8(src_reg, reg);
+
+#define LOAD_SRC_DST                                    \
+  /* load source and destination */                     \
+  src_reg = _mm256_loadu_si256((__m256i const *)(src)); \
+  dst_reg = _mm256_loadu_si256((__m256i const *)(dst));
+
+#define AVG_NEXT_SRC(src_reg, size_stride)                                 \
+  src_next_reg = _mm256_loadu_si256((__m256i const *)(src + size_stride)); \
+  /* average between current and next stride source */                     \
+  src_reg = _mm256_avg_epu8(src_reg, src_next_reg);
+
+#define MERGE_NEXT_SRC(src_reg, size_stride)                               \
+  src_next_reg = _mm256_loadu_si256((__m256i const *)(src + size_stride)); \
+  MERGE_WITH_SRC(src_reg, src_next_reg)
+
+#define CALC_SUM_SSE_INSIDE_LOOP                          \
+  /* expand each byte to 2 bytes */                       \
+  exp_dst_lo = _mm256_unpacklo_epi8(dst_reg, zero_reg);   \
+  exp_dst_hi = _mm256_unpackhi_epi8(dst_reg, zero_reg);   \
+  /* source - dest */                                     \
+  exp_src_lo = _mm256_sub_epi16(exp_src_lo, exp_dst_lo);  \
+  exp_src_hi = _mm256_sub_epi16(exp_src_hi, exp_dst_hi);  \
+  /* caculate sum */                                      \
+  sum_reg = _mm256_add_epi16(sum_reg, exp_src_lo);        \
+  exp_src_lo = _mm256_madd_epi16(exp_src_lo, exp_src_lo); \
+  sum_reg = _mm256_add_epi16(sum_reg, exp_src_hi);        \
+  exp_src_hi = _mm256_madd_epi16(exp_src_hi, exp_src_hi); \
+  /* calculate sse */                                     \
+  sse_reg = _mm256_add_epi32(sse_reg, exp_src_lo);        \
+  sse_reg = _mm256_add_epi32(sse_reg, exp_src_hi);
+
+// final calculation to sum and sse
+#define CALC_SUM_AND_SSE                                                   \
+  res_cmp = _mm256_cmpgt_epi16(zero_reg, sum_reg);                         \
+  sse_reg_hi = _mm256_srli_si256(sse_reg, 8);                              \
+  sum_reg_lo = _mm256_unpacklo_epi16(sum_reg, res_cmp);                    \
+  sum_reg_hi = _mm256_unpackhi_epi16(sum_reg, res_cmp);                    \
+  sse_reg = _mm256_add_epi32(sse_reg, sse_reg_hi);                         \
+  sum_reg = _mm256_add_epi32(sum_reg_lo, sum_reg_hi);                      \
+                                                                           \
+  sse_reg_hi = _mm256_srli_si256(sse_reg, 4);                              \
+  sum_reg_hi = _mm256_srli_si256(sum_reg, 8);                              \
+                                                                           \
+  sse_reg = _mm256_add_epi32(sse_reg, sse_reg_hi);                         \
+  sum_reg = _mm256_add_epi32(sum_reg, sum_reg_hi);                         \
+  *((int *)sse) = _mm_cvtsi128_si32(_mm256_castsi256_si128(sse_reg)) +     \
+                  _mm_cvtsi128_si32(_mm256_extractf128_si256(sse_reg, 1)); \
+  sum_reg_hi = _mm256_srli_si256(sum_reg, 4);                              \
+  sum_reg = _mm256_add_epi32(sum_reg, sum_reg_hi);                         \
+  sum = _mm_cvtsi128_si32(_mm256_castsi256_si128(sum_reg)) +               \
+        _mm_cvtsi128_si32(_mm256_extractf128_si256(sum_reg, 1));
+
+// Functions related to sub pixel variance width 16
+#define LOAD_SRC_DST_INSERT(src_stride, dst_stride)              \
+  /* load source and destination of 2 rows and insert*/          \
+  src_reg = _mm256_inserti128_si256(                             \
+      _mm256_castsi128_si256(_mm_loadu_si128((__m128i *)(src))), \
+      _mm_loadu_si128((__m128i *)(src + src_stride)), 1);        \
+  dst_reg = _mm256_inserti128_si256(                             \
+      _mm256_castsi128_si256(_mm_loadu_si128((__m128i *)(dst))), \
+      _mm_loadu_si128((__m128i *)(dst + dst_stride)), 1);
+
+#define AVG_NEXT_SRC_INSERT(src_reg, size_stride)                              \
+  src_next_reg = _mm256_inserti128_si256(                                      \
+      _mm256_castsi128_si256(_mm_loadu_si128((__m128i *)(src + size_stride))), \
+      _mm_loadu_si128((__m128i *)(src + (size_stride << 1))), 1);              \
+  /* average between current and next stride source */                         \
+  src_reg = _mm256_avg_epu8(src_reg, src_next_reg);
+
+#define MERGE_NEXT_SRC_INSERT(src_reg, size_stride)                            \
+  src_next_reg = _mm256_inserti128_si256(                                      \
+      _mm256_castsi128_si256(_mm_loadu_si128((__m128i *)(src + size_stride))), \
+      _mm_loadu_si128((__m128i *)(src + (src_stride + size_stride))), 1);      \
+  MERGE_WITH_SRC(src_reg, src_next_reg)
+
+#define LOAD_SRC_NEXT_BYTE_INSERT                                    \
+  /* load source and another source from next row   */               \
+  src_reg = _mm256_inserti128_si256(                                 \
+      _mm256_castsi128_si256(_mm_loadu_si128((__m128i *)(src))),     \
+      _mm_loadu_si128((__m128i *)(src + src_stride)), 1);            \
+  /* load source and next row source from 1 byte onwards   */        \
+  src_next_reg = _mm256_inserti128_si256(                            \
+      _mm256_castsi128_si256(_mm_loadu_si128((__m128i *)(src + 1))), \
+      _mm_loadu_si128((__m128i *)(src + src_stride + 1)), 1);
+
+#define LOAD_DST_INSERT                                          \
+  dst_reg = _mm256_inserti128_si256(                             \
+      _mm256_castsi128_si256(_mm_loadu_si128((__m128i *)(dst))), \
+      _mm_loadu_si128((__m128i *)(dst + dst_stride)), 1);
+
+#define LOAD_SRC_MERGE_128BIT(filter)                        \
+  __m128i src_reg_0 = _mm_loadu_si128((__m128i *)(src));     \
+  __m128i src_reg_1 = _mm_loadu_si128((__m128i *)(src + 1)); \
+  __m128i src_lo = _mm_unpacklo_epi8(src_reg_0, src_reg_1);  \
+  __m128i src_hi = _mm_unpackhi_epi8(src_reg_0, src_reg_1);  \
+  __m128i filter_128bit = _mm256_castsi256_si128(filter);    \
+  __m128i pw8_128bit = _mm256_castsi256_si128(pw8);
+
+#define FILTER_SRC_128BIT(filter)             \
+  /* filter the source */                     \
+  src_lo = _mm_maddubs_epi16(src_lo, filter); \
+  src_hi = _mm_maddubs_epi16(src_hi, filter); \
+                                              \
+  /* add 8 to source */                       \
+  src_lo = _mm_add_epi16(src_lo, pw8_128bit); \
+  src_hi = _mm_add_epi16(src_hi, pw8_128bit); \
+                                              \
+  /* divide source by 16 */                   \
+  src_lo = _mm_srai_epi16(src_lo, 4);         \
+  src_hi = _mm_srai_epi16(src_hi, 4);
+
+// TODO(chiyotsai@google.com): These variance functions are macro-fied so we
+// don't have to manually optimize the individual for-loops. We could save some
+// binary size by optimizing the loops more carefully without duplicating the
+// codes with a macro.
+#define MAKE_SUB_PIXEL_VAR_32XH(height, log2height)                           \
+  static AOM_INLINE int aom_sub_pixel_variance32x##height##_imp_avx2(         \
+      const uint8_t *src, int src_stride, int x_offset, int y_offset,         \
+      const uint8_t *dst, int dst_stride, unsigned int *sse) {                \
+    __m256i src_reg, dst_reg, exp_src_lo, exp_src_hi, exp_dst_lo, exp_dst_hi; \
+    __m256i sse_reg, sum_reg, sse_reg_hi, res_cmp, sum_reg_lo, sum_reg_hi;    \
+    __m256i zero_reg;                                                         \
+    int i, sum;                                                               \
+    sum_reg = _mm256_setzero_si256();                                         \
+    sse_reg = _mm256_setzero_si256();                                         \
+    zero_reg = _mm256_setzero_si256();                                        \
+                                                                              \
+    /* x_offset = 0 and y_offset = 0 */                                       \
+    if (x_offset == 0) {                                                      \
+      if (y_offset == 0) {                                                    \
+        for (i = 0; i < height; i++) {                                        \
+          LOAD_SRC_DST                                                        \
+          /* expend each byte to 2 bytes */                                   \
+          MERGE_WITH_SRC(src_reg, zero_reg)                                   \
+          CALC_SUM_SSE_INSIDE_LOOP                                            \
+          src += src_stride;                                                  \
+          dst += dst_stride;                                                  \
+        }                                                                     \
+        /* x_offset = 0 and y_offset = 4 */                                   \
+      } else if (y_offset == 4) {                                             \
+        __m256i src_next_reg;                                                 \
+        for (i = 0; i < height; i++) {                                        \
+          LOAD_SRC_DST                                                        \
+          AVG_NEXT_SRC(src_reg, src_stride)                                   \
+          /* expend each byte to 2 bytes */                                   \
+          MERGE_WITH_SRC(src_reg, zero_reg)                                   \
+          CALC_SUM_SSE_INSIDE_LOOP                                            \
+          src += src_stride;                                                  \
+          dst += dst_stride;                                                  \
+        }                                                                     \
+        /* x_offset = 0 and y_offset = bilin interpolation */                 \
+      } else {                                                                \
+        __m256i filter, pw8, src_next_reg;                                    \
+                                                                              \
+        y_offset <<= 5;                                                       \
+        filter = _mm256_load_si256(                                           \
+            (__m256i const *)(bilinear_filters_avx2 + y_offset));             \
+        pw8 = _mm256_set1_epi16(8);                                           \
+        for (i = 0; i < height; i++) {                                        \
+          LOAD_SRC_DST                                                        \
+          MERGE_NEXT_SRC(src_reg, src_stride)                                 \
+          FILTER_SRC(filter)                                                  \
+          CALC_SUM_SSE_INSIDE_LOOP                                            \
+          src += src_stride;                                                  \
+          dst += dst_stride;                                                  \
+        }                                                                     \
+      }                                                                       \
+      /* x_offset = 4  and y_offset = 0 */                                    \
+    } else if (x_offset == 4) {                                               \
+      if (y_offset == 0) {                                                    \
+        __m256i src_next_reg;                                                 \
+        for (i = 0; i < height; i++) {                                        \
+          LOAD_SRC_DST                                                        \
+          AVG_NEXT_SRC(src_reg, 1)                                            \
+          /* expand each byte to 2 bytes */                                   \
+          MERGE_WITH_SRC(src_reg, zero_reg)                                   \
+          CALC_SUM_SSE_INSIDE_LOOP                                            \
+          src += src_stride;                                                  \
+          dst += dst_stride;                                                  \
+        }                                                                     \
+        /* x_offset = 4  and y_offset = 4 */                                  \
+      } else if (y_offset == 4) {                                             \
+        __m256i src_next_reg, src_avg;                                        \
+        /* load source and another source starting from the next */           \
+        /* following byte */                                                  \
+        src_reg = _mm256_loadu_si256((__m256i const *)(src));                 \
+        AVG_NEXT_SRC(src_reg, 1)                                              \
+        for (i = 0; i < height; i++) {                                        \
+          src_avg = src_reg;                                                  \
+          src += src_stride;                                                  \
+          LOAD_SRC_DST                                                        \
+          AVG_NEXT_SRC(src_reg, 1)                                            \
+          /* average between previous average to current average */           \
+          src_avg = _mm256_avg_epu8(src_avg, src_reg);                        \
+          /* expand each byte to 2 bytes */                                   \
+          MERGE_WITH_SRC(src_avg, zero_reg)                                   \
+          /* save current source average */                                   \
+          CALC_SUM_SSE_INSIDE_LOOP                                            \
+          dst += dst_stride;                                                  \
+        }                                                                     \
+        /* x_offset = 4  and y_offset = bilin interpolation */                \
+      } else {                                                                \
+        __m256i filter, pw8, src_next_reg, src_avg;                           \
+        y_offset <<= 5;                                                       \
+        filter = _mm256_load_si256(                                           \
+            (__m256i const *)(bilinear_filters_avx2 + y_offset));             \
+        pw8 = _mm256_set1_epi16(8);                                           \
+        /* load source and another source starting from the next */           \
+        /* following byte */                                                  \
+        src_reg = _mm256_loadu_si256((__m256i const *)(src));                 \
+        AVG_NEXT_SRC(src_reg, 1)                                              \
+        for (i = 0; i < height; i++) {                                        \
+          /* save current source average */                                   \
+          src_avg = src_reg;                                                  \
+          src += src_stride;                                                  \
+          LOAD_SRC_DST                                                        \
+          AVG_NEXT_SRC(src_reg, 1)                                            \
+          MERGE_WITH_SRC(src_avg, src_reg)                                    \
+          FILTER_SRC(filter)                                                  \
+          CALC_SUM_SSE_INSIDE_LOOP                                            \
+          dst += dst_stride;                                                  \
+        }                                                                     \
+      }                                                                       \
+      /* x_offset = bilin interpolation and y_offset = 0 */                   \
+    } else {                                                                  \
+      if (y_offset == 0) {                                                    \
+        __m256i filter, pw8, src_next_reg;                                    \
+        x_offset <<= 5;                                                       \
+        filter = _mm256_load_si256(                                           \
+            (__m256i const *)(bilinear_filters_avx2 + x_offset));             \
+        pw8 = _mm256_set1_epi16(8);                                           \
+        for (i = 0; i < height; i++) {                                        \
+          LOAD_SRC_DST                                                        \
+          MERGE_NEXT_SRC(src_reg, 1)                                          \
+          FILTER_SRC(filter)                                                  \
+          CALC_SUM_SSE_INSIDE_LOOP                                            \
+          src += src_stride;                                                  \
+          dst += dst_stride;                                                  \
+        }                                                                     \
+        /* x_offset = bilin interpolation and y_offset = 4 */                 \
+      } else if (y_offset == 4) {                                             \
+        __m256i filter, pw8, src_next_reg, src_pack;                          \
+        x_offset <<= 5;                                                       \
+        filter = _mm256_load_si256(                                           \
+            (__m256i const *)(bilinear_filters_avx2 + x_offset));             \
+        pw8 = _mm256_set1_epi16(8);                                           \
+        src_reg = _mm256_loadu_si256((__m256i const *)(src));                 \
+        MERGE_NEXT_SRC(src_reg, 1)                                            \
+        FILTER_SRC(filter)                                                    \
+        /* convert each 16 bit to 8 bit to each low and high lane source */   \
+        src_pack = _mm256_packus_epi16(exp_src_lo, exp_src_hi);               \
+        for (i = 0; i < height; i++) {                                        \
+          src += src_stride;                                                  \
+          LOAD_SRC_DST                                                        \
+          MERGE_NEXT_SRC(src_reg, 1)                                          \
+          FILTER_SRC(filter)                                                  \
+          src_reg = _mm256_packus_epi16(exp_src_lo, exp_src_hi);              \
+          /* average between previous pack to the current */                  \
+          src_pack = _mm256_avg_epu8(src_pack, src_reg);                      \
+          MERGE_WITH_SRC(src_pack, zero_reg)                                  \
+          CALC_SUM_SSE_INSIDE_LOOP                                            \
+          src_pack = src_reg;                                                 \
+          dst += dst_stride;                                                  \
+        }                                                                     \
+        /* x_offset = bilin interpolation and y_offset = bilin interpolation  \
+         */                                                                   \
+      } else {                                                                \
+        __m256i xfilter, yfilter, pw8, src_next_reg, src_pack;                \
+        x_offset <<= 5;                                                       \
+        xfilter = _mm256_load_si256(                                          \
+            (__m256i const *)(bilinear_filters_avx2 + x_offset));             \
+        y_offset <<= 5;                                                       \
+        yfilter = _mm256_load_si256(                                          \
+            (__m256i const *)(bilinear_filters_avx2 + y_offset));             \
+        pw8 = _mm256_set1_epi16(8);                                           \
+        /* load source and another source starting from the next */           \
+        /* following byte */                                                  \
+        src_reg = _mm256_loadu_si256((__m256i const *)(src));                 \
+        MERGE_NEXT_SRC(src_reg, 1)                                            \
+                                                                              \
+        FILTER_SRC(xfilter)                                                   \
+        /* convert each 16 bit to 8 bit to each low and high lane source */   \
+        src_pack = _mm256_packus_epi16(exp_src_lo, exp_src_hi);               \
+        for (i = 0; i < height; i++) {                                        \
+          src += src_stride;                                                  \
+          LOAD_SRC_DST                                                        \
+          MERGE_NEXT_SRC(src_reg, 1)                                          \
+          FILTER_SRC(xfilter)                                                 \
+          src_reg = _mm256_packus_epi16(exp_src_lo, exp_src_hi);              \
+          /* merge previous pack to current pack source */                    \
+          MERGE_WITH_SRC(src_pack, src_reg)                                   \
+          /* filter the source */                                             \
+          FILTER_SRC(yfilter)                                                 \
+          src_pack = src_reg;                                                 \
+          CALC_SUM_SSE_INSIDE_LOOP                                            \
+          dst += dst_stride;                                                  \
+        }                                                                     \
+      }                                                                       \
+    }                                                                         \
+    CALC_SUM_AND_SSE                                                          \
+    _mm256_zeroupper();                                                       \
+    return sum;                                                               \
+  }                                                                           \
+  unsigned int aom_sub_pixel_variance32x##height##_avx2(                      \
+      const uint8_t *src, int src_stride, int x_offset, int y_offset,         \
+      const uint8_t *dst, int dst_stride, unsigned int *sse) {                \
+    const int sum = aom_sub_pixel_variance32x##height##_imp_avx2(             \
+        src, src_stride, x_offset, y_offset, dst, dst_stride, sse);           \
+    return *sse - (unsigned int)(((int64_t)sum * sum) >> (5 + log2height));   \
+  }
+
+MAKE_SUB_PIXEL_VAR_32XH(64, 6)
+MAKE_SUB_PIXEL_VAR_32XH(32, 5)
+MAKE_SUB_PIXEL_VAR_32XH(16, 4)
+
+#define AOM_SUB_PIXEL_VAR_AVX2(w, h, wf, hf, wlog2, hlog2)                \
+  unsigned int aom_sub_pixel_variance##w##x##h##_avx2(                    \
+      const uint8_t *src, int src_stride, int x_offset, int y_offset,     \
+      const uint8_t *dst, int dst_stride, unsigned int *sse_ptr) {        \
+    unsigned int sse = 0;                                                 \
+    int se = 0;                                                           \
+    for (int i = 0; i < (w / wf); ++i) {                                  \
+      const uint8_t *src_ptr = src;                                       \
+      const uint8_t *dst_ptr = dst;                                       \
+      for (int j = 0; j < (h / hf); ++j) {                                \
+        unsigned int sse2;                                                \
+        const int se2 = aom_sub_pixel_variance##wf##x##hf##_imp_avx2(     \
+            src_ptr, src_stride, x_offset, y_offset, dst_ptr, dst_stride, \
+            &sse2);                                                       \
+        dst_ptr += hf * dst_stride;                                       \
+        src_ptr += hf * src_stride;                                       \
+        se += se2;                                                        \
+        sse += sse2;                                                      \
+      }                                                                   \
+      src += wf;                                                          \
+      dst += wf;                                                          \
+    }                                                                     \
+    *sse_ptr = sse;                                                       \
+    return sse - (unsigned int)(((int64_t)se * se) >> (wlog2 + hlog2));   \
+  }
+
+// Note: hf = AOMMIN(h, 64) to avoid overflow in helper by capping height.
+AOM_SUB_PIXEL_VAR_AVX2(128, 128, 32, 64, 7, 7)
+AOM_SUB_PIXEL_VAR_AVX2(128, 64, 32, 64, 7, 6)
+AOM_SUB_PIXEL_VAR_AVX2(64, 128, 32, 64, 6, 7)
+AOM_SUB_PIXEL_VAR_AVX2(64, 64, 32, 64, 6, 6)
+AOM_SUB_PIXEL_VAR_AVX2(64, 32, 32, 32, 6, 5)
+
+#define MAKE_SUB_PIXEL_VAR_16XH(height, log2height)                           \
+  unsigned int aom_sub_pixel_variance16x##height##_avx2(                      \
+      const uint8_t *src, int src_stride, int x_offset, int y_offset,         \
+      const uint8_t *dst, int dst_stride, unsigned int *sse) {                \
+    __m256i src_reg, dst_reg, exp_src_lo, exp_src_hi, exp_dst_lo, exp_dst_hi; \
+    __m256i sse_reg, sum_reg, sse_reg_hi, res_cmp, sum_reg_lo, sum_reg_hi;    \
+    __m256i zero_reg;                                                         \
+    int i, sum;                                                               \
+    sum_reg = _mm256_setzero_si256();                                         \
+    sse_reg = _mm256_setzero_si256();                                         \
+    zero_reg = _mm256_setzero_si256();                                        \
+                                                                              \
+    /* x_offset = 0 and y_offset = 0 */                                       \
+    if (x_offset == 0) {                                                      \
+      if (y_offset == 0) {                                                    \
+        for (i = 0; i < height; i += 2) {                                     \
+          LOAD_SRC_DST_INSERT(src_stride, dst_stride)                         \
+          /* expend each byte to 2 bytes */                                   \
+          MERGE_WITH_SRC(src_reg, zero_reg)                                   \
+          CALC_SUM_SSE_INSIDE_LOOP                                            \
+          src += (src_stride << 1);                                           \
+          dst += (dst_stride << 1);                                           \
+        }                                                                     \
+        /* x_offset = 0 and y_offset = 4 */                                   \
+      } else if (y_offset == 4) {                                             \
+        __m256i src_next_reg;                                                 \
+        for (i = 0; i < height; i += 2) {                                     \
+          LOAD_SRC_DST_INSERT(src_stride, dst_stride)                         \
+          AVG_NEXT_SRC_INSERT(src_reg, src_stride)                            \
+          /* expend each byte to 2 bytes */                                   \
+          MERGE_WITH_SRC(src_reg, zero_reg)                                   \
+          CALC_SUM_SSE_INSIDE_LOOP                                            \
+          src += (src_stride << 1);                                           \
+          dst += (dst_stride << 1);                                           \
+        }                                                                     \
+        /* x_offset = 0 and y_offset = bilin interpolation */                 \
+      } else {                                                                \
+        __m256i filter, pw8, src_next_reg;                                    \
+        y_offset <<= 5;                                                       \
+        filter = _mm256_load_si256(                                           \
+            (__m256i const *)(bilinear_filters_avx2 + y_offset));             \
+        pw8 = _mm256_set1_epi16(8);                                           \
+        for (i = 0; i < height; i += 2) {                                     \
+          LOAD_SRC_DST_INSERT(src_stride, dst_stride)                         \
+          MERGE_NEXT_SRC_INSERT(src_reg, src_stride)                          \
+          FILTER_SRC(filter)                                                  \
+          CALC_SUM_SSE_INSIDE_LOOP                                            \
+          src += (src_stride << 1);                                           \
+          dst += (dst_stride << 1);                                           \
+        }                                                                     \
+      }                                                                       \
+      /* x_offset = 4  and y_offset = 0 */                                    \
+    } else if (x_offset == 4) {                                               \
+      if (y_offset == 0) {                                                    \
+        __m256i src_next_reg;                                                 \
+        for (i = 0; i < height; i += 2) {                                     \
+          LOAD_SRC_NEXT_BYTE_INSERT                                           \
+          LOAD_DST_INSERT                                                     \
+          /* average between current and next stride source */                \
+          src_reg = _mm256_avg_epu8(src_reg, src_next_reg);                   \
+          /* expand each byte to 2 bytes */                                   \
+          MERGE_WITH_SRC(src_reg, zero_reg)                                   \
+          CALC_SUM_SSE_INSIDE_LOOP                                            \
+          src += (src_stride << 1);                                           \
+          dst += (dst_stride << 1);                                           \
+        }                                                                     \
+        /* x_offset = 4  and y_offset = 4 */                                  \
+      } else if (y_offset == 4) {                                             \
+        __m256i src_next_reg, src_avg, src_temp;                              \
+        /* load and insert source and next row source */                      \
+        LOAD_SRC_NEXT_BYTE_INSERT                                             \
+        src_avg = _mm256_avg_epu8(src_reg, src_next_reg);                     \
+        src += src_stride << 1;                                               \
+        for (i = 0; i < height - 2; i += 2) {                                 \
+          LOAD_SRC_NEXT_BYTE_INSERT                                           \
+          src_next_reg = _mm256_avg_epu8(src_reg, src_next_reg);              \
+          src_temp = _mm256_permute2x128_si256(src_avg, src_next_reg, 0x21);  \
+          src_temp = _mm256_avg_epu8(src_avg, src_temp);                      \
+          LOAD_DST_INSERT                                                     \
+          /* expand each byte to 2 bytes */                                   \
+          MERGE_WITH_SRC(src_temp, zero_reg)                                  \
+          /* save current source average */                                   \
+          src_avg = src_next_reg;                                             \
+          CALC_SUM_SSE_INSIDE_LOOP                                            \
+          dst += dst_stride << 1;                                             \
+          src += src_stride << 1;                                             \
+        }                                                                     \
+        /* last 2 rows processing happens here */                             \
+        __m128i src_reg_0 = _mm_loadu_si128((__m128i *)(src));                \
+        __m128i src_reg_1 = _mm_loadu_si128((__m128i *)(src + 1));            \
+        src_reg_0 = _mm_avg_epu8(src_reg_0, src_reg_1);                       \
+        src_next_reg = _mm256_permute2x128_si256(                             \
+            src_avg, _mm256_castsi128_si256(src_reg_0), 0x21);                \
+        LOAD_DST_INSERT                                                       \
+        src_avg = _mm256_avg_epu8(src_avg, src_next_reg);                     \
+        MERGE_WITH_SRC(src_avg, zero_reg)                                     \
+        CALC_SUM_SSE_INSIDE_LOOP                                              \
+      } else {                                                                \
+        /* x_offset = 4  and y_offset = bilin interpolation */                \
+        __m256i filter, pw8, src_next_reg, src_avg, src_temp;                 \
+        y_offset <<= 5;                                                       \
+        filter = _mm256_load_si256(                                           \
+            (__m256i const *)(bilinear_filters_avx2 + y_offset));             \
+        pw8 = _mm256_set1_epi16(8);                                           \
+        /* load and insert source and next row source */                      \
+        LOAD_SRC_NEXT_BYTE_INSERT                                             \
+        src_avg = _mm256_avg_epu8(src_reg, src_next_reg);                     \
+        src += src_stride << 1;                                               \
+        for (i = 0; i < height - 2; i += 2) {                                 \
+          LOAD_SRC_NEXT_BYTE_INSERT                                           \
+          src_next_reg = _mm256_avg_epu8(src_reg, src_next_reg);              \
+          src_temp = _mm256_permute2x128_si256(src_avg, src_next_reg, 0x21);  \
+          LOAD_DST_INSERT                                                     \
+          MERGE_WITH_SRC(src_avg, src_temp)                                   \
+          /* save current source average */                                   \
+          src_avg = src_next_reg;                                             \
+          FILTER_SRC(filter)                                                  \
+          CALC_SUM_SSE_INSIDE_LOOP                                            \
+          dst += dst_stride << 1;                                             \
+          src += src_stride << 1;                                             \
+        }                                                                     \
+        /* last 2 rows processing happens here */                             \
+        __m128i src_reg_0 = _mm_loadu_si128((__m128i *)(src));                \
+        __m128i src_reg_1 = _mm_loadu_si128((__m128i *)(src + 1));            \
+        src_reg_0 = _mm_avg_epu8(src_reg_0, src_reg_1);                       \
+        src_next_reg = _mm256_permute2x128_si256(                             \
+            src_avg, _mm256_castsi128_si256(src_reg_0), 0x21);                \
+        LOAD_DST_INSERT                                                       \
+        MERGE_WITH_SRC(src_avg, src_next_reg)                                 \
+        FILTER_SRC(filter)                                                    \
+        CALC_SUM_SSE_INSIDE_LOOP                                              \
+      }                                                                       \
+      /* x_offset = bilin interpolation and y_offset = 0 */                   \
+    } else {                                                                  \
+      if (y_offset == 0) {                                                    \
+        __m256i filter, pw8, src_next_reg;                                    \
+        x_offset <<= 5;                                                       \
+        filter = _mm256_load_si256(                                           \
+            (__m256i const *)(bilinear_filters_avx2 + x_offset));             \
+        pw8 = _mm256_set1_epi16(8);                                           \
+        for (i = 0; i < height; i += 2) {                                     \
+          LOAD_SRC_DST_INSERT(src_stride, dst_stride)                         \
+          MERGE_NEXT_SRC_INSERT(src_reg, 1)                                   \
+          FILTER_SRC(filter)                                                  \
+          CALC_SUM_SSE_INSIDE_LOOP                                            \
+          src += (src_stride << 1);                                           \
+          dst += (dst_stride << 1);                                           \
+        }                                                                     \
+        /* x_offset = bilin interpolation and y_offset = 4 */                 \
+      } else if (y_offset == 4) {                                             \
+        __m256i filter, pw8, src_next_reg, src_pack;                          \
+        x_offset <<= 5;                                                       \
+        filter = _mm256_load_si256(                                           \
+            (__m256i const *)(bilinear_filters_avx2 + x_offset));             \
+        pw8 = _mm256_set1_epi16(8);                                           \
+        /* load and insert source and next row source */                      \
+        LOAD_SRC_NEXT_BYTE_INSERT                                             \
+        MERGE_WITH_SRC(src_reg, src_next_reg)                                 \
+        FILTER_SRC(filter)                                                    \
+        /* convert each 16 bit to 8 bit to each low and high lane source */   \
+        src_pack = _mm256_packus_epi16(exp_src_lo, exp_src_hi);               \
+        src += src_stride << 1;                                               \
+        for (i = 0; i < height - 2; i += 2) {                                 \
+          LOAD_SRC_NEXT_BYTE_INSERT                                           \
+          LOAD_DST_INSERT                                                     \
+          MERGE_WITH_SRC(src_reg, src_next_reg)                               \
+          FILTER_SRC(filter)                                                  \
+          src_reg = _mm256_packus_epi16(exp_src_lo, exp_src_hi);              \
+          src_next_reg = _mm256_permute2x128_si256(src_pack, src_reg, 0x21);  \
+          /* average between previous pack to the current */                  \
+          src_pack = _mm256_avg_epu8(src_pack, src_next_reg);                 \
+          MERGE_WITH_SRC(src_pack, zero_reg)                                  \
+          CALC_SUM_SSE_INSIDE_LOOP                                            \
+          src_pack = src_reg;                                                 \
+          src += src_stride << 1;                                             \
+          dst += dst_stride << 1;                                             \
+        }                                                                     \
+        /* last 2 rows processing happens here */                             \
+        LOAD_SRC_MERGE_128BIT(filter)                                         \
+        LOAD_DST_INSERT                                                       \
+        FILTER_SRC_128BIT(filter_128bit)                                      \
+        src_reg_0 = _mm_packus_epi16(src_lo, src_hi);                         \
+        src_next_reg = _mm256_permute2x128_si256(                             \
+            src_pack, _mm256_castsi128_si256(src_reg_0), 0x21);               \
+        /* average between previous pack to the current */                    \
+        src_pack = _mm256_avg_epu8(src_pack, src_next_reg);                   \
+        MERGE_WITH_SRC(src_pack, zero_reg)                                    \
+        CALC_SUM_SSE_INSIDE_LOOP                                              \
+      } else {                                                                \
+        /* x_offset = bilin interpolation and y_offset = bilin interpolation  \
+         */                                                                   \
+        __m256i xfilter, yfilter, pw8, src_next_reg, src_pack;                \
+        x_offset <<= 5;                                                       \
+        xfilter = _mm256_load_si256(                                          \
+            (__m256i const *)(bilinear_filters_avx2 + x_offset));             \
+        y_offset <<= 5;                                                       \
+        yfilter = _mm256_load_si256(                                          \
+            (__m256i const *)(bilinear_filters_avx2 + y_offset));             \
+        pw8 = _mm256_set1_epi16(8);                                           \
+        /* load and insert source and next row source */                      \
+        LOAD_SRC_NEXT_BYTE_INSERT                                             \
+        MERGE_WITH_SRC(src_reg, src_next_reg)                                 \
+        FILTER_SRC(xfilter)                                                   \
+        /* convert each 16 bit to 8 bit to each low and high lane source */   \
+        src_pack = _mm256_packus_epi16(exp_src_lo, exp_src_hi);               \
+        src += src_stride << 1;                                               \
+        for (i = 0; i < height - 2; i += 2) {                                 \
+          LOAD_SRC_NEXT_BYTE_INSERT                                           \
+          LOAD_DST_INSERT                                                     \
+          MERGE_WITH_SRC(src_reg, src_next_reg)                               \
+          FILTER_SRC(xfilter)                                                 \
+          src_reg = _mm256_packus_epi16(exp_src_lo, exp_src_hi);              \
+          src_next_reg = _mm256_permute2x128_si256(src_pack, src_reg, 0x21);  \
+          /* average between previous pack to the current */                  \
+          MERGE_WITH_SRC(src_pack, src_next_reg)                              \
+          /* filter the source */                                             \
+          FILTER_SRC(yfilter)                                                 \
+          src_pack = src_reg;                                                 \
+          CALC_SUM_SSE_INSIDE_LOOP                                            \
+          src += src_stride << 1;                                             \
+          dst += dst_stride << 1;                                             \
+        }                                                                     \
+        /* last 2 rows processing happens here */                             \
+        LOAD_SRC_MERGE_128BIT(xfilter)                                        \
+        LOAD_DST_INSERT                                                       \
+        FILTER_SRC_128BIT(filter_128bit)                                      \
+        src_reg_0 = _mm_packus_epi16(src_lo, src_hi);                         \
+        src_next_reg = _mm256_permute2x128_si256(                             \
+            src_pack, _mm256_castsi128_si256(src_reg_0), 0x21);               \
+        MERGE_WITH_SRC(src_pack, src_next_reg)                                \
+        FILTER_SRC(yfilter)                                                   \
+        CALC_SUM_SSE_INSIDE_LOOP                                              \
+      }                                                                       \
+    }                                                                         \
+    CALC_SUM_AND_SSE                                                          \
+    _mm256_zeroupper();                                                       \
+    return *sse - (unsigned int)(((int64_t)sum * sum) >> (4 + log2height));   \
+  }
+
+MAKE_SUB_PIXEL_VAR_16XH(32, 5)
+MAKE_SUB_PIXEL_VAR_16XH(16, 4)
+MAKE_SUB_PIXEL_VAR_16XH(8, 3)
+#if !CONFIG_REALTIME_ONLY
+MAKE_SUB_PIXEL_VAR_16XH(64, 6)
+MAKE_SUB_PIXEL_VAR_16XH(4, 2)
+#endif
+
+#define MAKE_SUB_PIXEL_AVG_VAR_32XH(height, log2height)                       \
+  int aom_sub_pixel_avg_variance32x##height##_imp_avx2(                       \
+      const uint8_t *src, int src_stride, int x_offset, int y_offset,         \
+      const uint8_t *dst, int dst_stride, const uint8_t *sec, int sec_stride, \
+      unsigned int *sse) {                                                    \
+    __m256i sec_reg;                                                          \
+    __m256i src_reg, dst_reg, exp_src_lo, exp_src_hi, exp_dst_lo, exp_dst_hi; \
+    __m256i sse_reg, sum_reg, sse_reg_hi, res_cmp, sum_reg_lo, sum_reg_hi;    \
+    __m256i zero_reg;                                                         \
+    int i, sum;                                                               \
+    sum_reg = _mm256_setzero_si256();                                         \
+    sse_reg = _mm256_setzero_si256();                                         \
+    zero_reg = _mm256_setzero_si256();                                        \
+                                                                              \
+    /* x_offset = 0 and y_offset = 0 */                                       \
+    if (x_offset == 0) {                                                      \
+      if (y_offset == 0) {                                                    \
+        for (i = 0; i < height; i++) {                                        \
+          LOAD_SRC_DST                                                        \
+          sec_reg = _mm256_loadu_si256((__m256i const *)(sec));               \
+          src_reg = _mm256_avg_epu8(src_reg, sec_reg);                        \
+          sec += sec_stride;                                                  \
+          /* expend each byte to 2 bytes */                                   \
+          MERGE_WITH_SRC(src_reg, zero_reg)                                   \
+          CALC_SUM_SSE_INSIDE_LOOP                                            \
+          src += src_stride;                                                  \
+          dst += dst_stride;                                                  \
+        }                                                                     \
+      } else if (y_offset == 4) {                                             \
+        __m256i src_next_reg;                                                 \
+        for (i = 0; i < height; i++) {                                        \
+          LOAD_SRC_DST                                                        \
+          AVG_NEXT_SRC(src_reg, src_stride)                                   \
+          sec_reg = _mm256_loadu_si256((__m256i const *)(sec));               \
+          src_reg = _mm256_avg_epu8(src_reg, sec_reg);                        \
+          sec += sec_stride;                                                  \
+          /* expend each byte to 2 bytes */                                   \
+          MERGE_WITH_SRC(src_reg, zero_reg)                                   \
+          CALC_SUM_SSE_INSIDE_LOOP                                            \
+          src += src_stride;                                                  \
+          dst += dst_stride;                                                  \
+        }                                                                     \
+        /* x_offset = 0 and y_offset = bilin interpolation */                 \
+      } else {                                                                \
+        __m256i filter, pw8, src_next_reg;                                    \
+                                                                              \
+        y_offset <<= 5;                                                       \
+        filter = _mm256_load_si256(                                           \
+            (__m256i const *)(bilinear_filters_avx2 + y_offset));             \
+        pw8 = _mm256_set1_epi16(8);                                           \
+        for (i = 0; i < height; i++) {                                        \
+          LOAD_SRC_DST                                                        \
+          MERGE_NEXT_SRC(src_reg, src_stride)                                 \
+          FILTER_SRC(filter)                                                  \
+          src_reg = _mm256_packus_epi16(exp_src_lo, exp_src_hi);              \
+          sec_reg = _mm256_loadu_si256((__m256i const *)(sec));               \
+          src_reg = _mm256_avg_epu8(src_reg, sec_reg);                        \
+          sec += sec_stride;                                                  \
+          MERGE_WITH_SRC(src_reg, zero_reg)                                   \
+          CALC_SUM_SSE_INSIDE_LOOP                                            \
+          src += src_stride;                                                  \
+          dst += dst_stride;                                                  \
+        }                                                                     \
+      }                                                                       \
+      /* x_offset = 4  and y_offset = 0 */                                    \
+    } else if (x_offset == 4) {                                               \
+      if (y_offset == 0) {                                                    \
+        __m256i src_next_reg;                                                 \
+        for (i = 0; i < height; i++) {                                        \
+          LOAD_SRC_DST                                                        \
+          AVG_NEXT_SRC(src_reg, 1)                                            \
+          sec_reg = _mm256_loadu_si256((__m256i const *)(sec));               \
+          src_reg = _mm256_avg_epu8(src_reg, sec_reg);                        \
+          sec += sec_stride;                                                  \
+          /* expand each byte to 2 bytes */                                   \
+          MERGE_WITH_SRC(src_reg, zero_reg)                                   \
+          CALC_SUM_SSE_INSIDE_LOOP                                            \
+          src += src_stride;                                                  \
+          dst += dst_stride;                                                  \
+        }                                                                     \
+        /* x_offset = 4  and y_offset = 4 */                                  \
+      } else if (y_offset == 4) {                                             \
+        __m256i src_next_reg, src_avg;                                        \
+        /* load source and another source starting from the next */           \
+        /* following byte */                                                  \
+        src_reg = _mm256_loadu_si256((__m256i const *)(src));                 \
+        AVG_NEXT_SRC(src_reg, 1)                                              \
+        for (i = 0; i < height; i++) {                                        \
+          /* save current source average */                                   \
+          src_avg = src_reg;                                                  \
+          src += src_stride;                                                  \
+          LOAD_SRC_DST                                                        \
+          AVG_NEXT_SRC(src_reg, 1)                                            \
+          /* average between previous average to current average */           \
+          src_avg = _mm256_avg_epu8(src_avg, src_reg);                        \
+          sec_reg = _mm256_loadu_si256((__m256i const *)(sec));               \
+          src_avg = _mm256_avg_epu8(src_avg, sec_reg);                        \
+          sec += sec_stride;                                                  \
+          /* expand each byte to 2 bytes */                                   \
+          MERGE_WITH_SRC(src_avg, zero_reg)                                   \
+          CALC_SUM_SSE_INSIDE_LOOP                                            \
+          dst += dst_stride;                                                  \
+        }                                                                     \
+        /* x_offset = 4  and y_offset = bilin interpolation */                \
+      } else {                                                                \
+        __m256i filter, pw8, src_next_reg, src_avg;                           \
+        y_offset <<= 5;                                                       \
+        filter = _mm256_load_si256(                                           \
+            (__m256i const *)(bilinear_filters_avx2 + y_offset));             \
+        pw8 = _mm256_set1_epi16(8);                                           \
+        /* load source and another source starting from the next */           \
+        /* following byte */                                                  \
+        src_reg = _mm256_loadu_si256((__m256i const *)(src));                 \
+        AVG_NEXT_SRC(src_reg, 1)                                              \
+        for (i = 0; i < height; i++) {                                        \
+          /* save current source average */                                   \
+          src_avg = src_reg;                                                  \
+          src += src_stride;                                                  \
+          LOAD_SRC_DST                                                        \
+          AVG_NEXT_SRC(src_reg, 1)                                            \
+          MERGE_WITH_SRC(src_avg, src_reg)                                    \
+          FILTER_SRC(filter)                                                  \
+          src_avg = _mm256_packus_epi16(exp_src_lo, exp_src_hi);              \
+          sec_reg = _mm256_loadu_si256((__m256i const *)(sec));               \
+          src_avg = _mm256_avg_epu8(src_avg, sec_reg);                        \
+          /* expand each byte to 2 bytes */                                   \
+          MERGE_WITH_SRC(src_avg, zero_reg)                                   \
+          sec += sec_stride;                                                  \
+          CALC_SUM_SSE_INSIDE_LOOP                                            \
+          dst += dst_stride;                                                  \
+        }                                                                     \
+      }                                                                       \
+      /* x_offset = bilin interpolation and y_offset = 0 */                   \
+    } else {                                                                  \
+      if (y_offset == 0) {                                                    \
+        __m256i filter, pw8, src_next_reg;                                    \
+        x_offset <<= 5;                                                       \
+        filter = _mm256_load_si256(                                           \
+            (__m256i const *)(bilinear_filters_avx2 + x_offset));             \
+        pw8 = _mm256_set1_epi16(8);                                           \
+        for (i = 0; i < height; i++) {                                        \
+          LOAD_SRC_DST                                                        \
+          MERGE_NEXT_SRC(src_reg, 1)                                          \
+          FILTER_SRC(filter)                                                  \
+          src_reg = _mm256_packus_epi16(exp_src_lo, exp_src_hi);              \
+          sec_reg = _mm256_loadu_si256((__m256i const *)(sec));               \
+          src_reg = _mm256_avg_epu8(src_reg, sec_reg);                        \
+          MERGE_WITH_SRC(src_reg, zero_reg)                                   \
+          sec += sec_stride;                                                  \
+          CALC_SUM_SSE_INSIDE_LOOP                                            \
+          src += src_stride;                                                  \
+          dst += dst_stride;                                                  \
+        }                                                                     \
+        /* x_offset = bilin interpolation and y_offset = 4 */                 \
+      } else if (y_offset == 4) {                                             \
+        __m256i filter, pw8, src_next_reg, src_pack;                          \
+        x_offset <<= 5;                                                       \
+        filter = _mm256_load_si256(                                           \
+            (__m256i const *)(bilinear_filters_avx2 + x_offset));             \
+        pw8 = _mm256_set1_epi16(8);                                           \
+        src_reg = _mm256_loadu_si256((__m256i const *)(src));                 \
+        MERGE_NEXT_SRC(src_reg, 1)                                            \
+        FILTER_SRC(filter)                                                    \
+        /* convert each 16 bit to 8 bit to each low and high lane source */   \
+        src_pack = _mm256_packus_epi16(exp_src_lo, exp_src_hi);               \
+        for (i = 0; i < height; i++) {                                        \
+          src += src_stride;                                                  \
+          LOAD_SRC_DST                                                        \
+          MERGE_NEXT_SRC(src_reg, 1)                                          \
+          FILTER_SRC(filter)                                                  \
+          src_reg = _mm256_packus_epi16(exp_src_lo, exp_src_hi);              \
+          /* average between previous pack to the current */                  \
+          src_pack = _mm256_avg_epu8(src_pack, src_reg);                      \
+          sec_reg = _mm256_loadu_si256((__m256i const *)(sec));               \
+          src_pack = _mm256_avg_epu8(src_pack, sec_reg);                      \
+          sec += sec_stride;                                                  \
+          MERGE_WITH_SRC(src_pack, zero_reg)                                  \
+          src_pack = src_reg;                                                 \
+          CALC_SUM_SSE_INSIDE_LOOP                                            \
+          dst += dst_stride;                                                  \
+        }                                                                     \
+        /* x_offset = bilin interpolation and y_offset = bilin interpolation  \
+         */                                                                   \
+      } else {                                                                \
+        __m256i xfilter, yfilter, pw8, src_next_reg, src_pack;                \
+        x_offset <<= 5;                                                       \
+        xfilter = _mm256_load_si256(                                          \
+            (__m256i const *)(bilinear_filters_avx2 + x_offset));             \
+        y_offset <<= 5;                                                       \
+        yfilter = _mm256_load_si256(                                          \
+            (__m256i const *)(bilinear_filters_avx2 + y_offset));             \
+        pw8 = _mm256_set1_epi16(8);                                           \
+        /* load source and another source starting from the next */           \
+        /* following byte */                                                  \
+        src_reg = _mm256_loadu_si256((__m256i const *)(src));                 \
+        MERGE_NEXT_SRC(src_reg, 1)                                            \
+                                                                              \
+        FILTER_SRC(xfilter)                                                   \
+        /* convert each 16 bit to 8 bit to each low and high lane source */   \
+        src_pack = _mm256_packus_epi16(exp_src_lo, exp_src_hi);               \
+        for (i = 0; i < height; i++) {                                        \
+          src += src_stride;                                                  \
+          LOAD_SRC_DST                                                        \
+          MERGE_NEXT_SRC(src_reg, 1)                                          \
+          FILTER_SRC(xfilter)                                                 \
+          src_reg = _mm256_packus_epi16(exp_src_lo, exp_src_hi);              \
+          /* merge previous pack to current pack source */                    \
+          MERGE_WITH_SRC(src_pack, src_reg)                                   \
+          /* filter the source */                                             \
+          FILTER_SRC(yfilter)                                                 \
+          src_pack = _mm256_packus_epi16(exp_src_lo, exp_src_hi);             \
+          sec_reg = _mm256_loadu_si256((__m256i const *)(sec));               \
+          src_pack = _mm256_avg_epu8(src_pack, sec_reg);                      \
+          MERGE_WITH_SRC(src_pack, zero_reg)                                  \
+          src_pack = src_reg;                                                 \
+          sec += sec_stride;                                                  \
+          CALC_SUM_SSE_INSIDE_LOOP                                            \
+          dst += dst_stride;                                                  \
+        }                                                                     \
+      }                                                                       \
+    }                                                                         \
+    CALC_SUM_AND_SSE                                                          \
+    _mm256_zeroupper();                                                       \
+    return sum;                                                               \
+  }                                                                           \
+  unsigned int aom_sub_pixel_avg_variance32x##height##_avx2(                  \
+      const uint8_t *src, int src_stride, int x_offset, int y_offset,         \
+      const uint8_t *dst, int dst_stride, unsigned int *sse,                  \
+      const uint8_t *sec_ptr) {                                               \
+    const int sum = aom_sub_pixel_avg_variance32x##height##_imp_avx2(         \
+        src, src_stride, x_offset, y_offset, dst, dst_stride, sec_ptr, 32,    \
+        sse);                                                                 \
+    return *sse - (unsigned int)(((int64_t)sum * sum) >> (5 + log2height));   \
+  }
+
+MAKE_SUB_PIXEL_AVG_VAR_32XH(64, 6)
+MAKE_SUB_PIXEL_AVG_VAR_32XH(32, 5)
+MAKE_SUB_PIXEL_AVG_VAR_32XH(16, 4)
+
+#define AOM_SUB_PIXEL_AVG_VAR_AVX2(w, h, wf, hf, wlog2, hlog2)            \
+  unsigned int aom_sub_pixel_avg_variance##w##x##h##_avx2(                \
+      const uint8_t *src, int src_stride, int x_offset, int y_offset,     \
+      const uint8_t *dst, int dst_stride, unsigned int *sse_ptr,          \
+      const uint8_t *sec) {                                               \
+    unsigned int sse = 0;                                                 \
+    int se = 0;                                                           \
+    for (int i = 0; i < (w / wf); ++i) {                                  \
+      const uint8_t *src_ptr = src;                                       \
+      const uint8_t *dst_ptr = dst;                                       \
+      const uint8_t *sec_ptr = sec;                                       \
+      for (int j = 0; j < (h / hf); ++j) {                                \
+        unsigned int sse2;                                                \
+        const int se2 = aom_sub_pixel_avg_variance##wf##x##hf##_imp_avx2( \
+            src_ptr, src_stride, x_offset, y_offset, dst_ptr, dst_stride, \
+            sec_ptr, w, &sse2);                                           \
+        dst_ptr += hf * dst_stride;                                       \
+        src_ptr += hf * src_stride;                                       \
+        sec_ptr += hf * w;                                                \
+        se += se2;                                                        \
+        sse += sse2;                                                      \
+      }                                                                   \
+      src += wf;                                                          \
+      dst += wf;                                                          \
+      sec += wf;                                                          \
+    }                                                                     \
+    *sse_ptr = sse;                                                       \
+    return sse - (unsigned int)(((int64_t)se * se) >> (wlog2 + hlog2));   \
+  }
+
+// Note: hf = AOMMIN(h, 64) to avoid overflow in helper by capping height.
+AOM_SUB_PIXEL_AVG_VAR_AVX2(128, 128, 32, 64, 7, 7)
+AOM_SUB_PIXEL_AVG_VAR_AVX2(128, 64, 32, 64, 7, 6)
+AOM_SUB_PIXEL_AVG_VAR_AVX2(64, 128, 32, 64, 6, 7)
+AOM_SUB_PIXEL_AVG_VAR_AVX2(64, 64, 32, 64, 6, 6)
+AOM_SUB_PIXEL_AVG_VAR_AVX2(64, 32, 32, 32, 6, 5)
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 0000000000..699002195b
--- /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 int8_t f0 = (int8_t)(filter[0] >> 1);
+  const int8_t f1 = (int8_t)(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
new file mode 100644
index 0000000000..faec9cf73d
--- /dev/null
+++ b/third_party/aom/aom_dsp/x86/variance_sse2.c
@@ -0,0 +1,802 @@
+/*
+ * Copyright (c) 2016, Alliance for Open Media. All rights reserved
+ *
+ * This source code is subject to the terms of the BSD 2 Clause License and
+ * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
+ * was not distributed with this source code in the LICENSE file, you can
+ * obtain it at www.aomedia.org/license/software. If the Alliance for Open
+ * Media Patent License 1.0 was not distributed with this source code in the
+ * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
+ */
+
+#include <assert.h>
+#include <emmintrin.h>  // SSE2
+
+#include "config/aom_config.h"
+#include "config/aom_dsp_rtcd.h"
+
+#include "aom_dsp/blend.h"
+#include "aom_dsp/x86/mem_sse2.h"
+#include "aom_dsp/x86/synonyms.h"
+#include "aom_ports/mem.h"
+
+unsigned int aom_get_mb_ss_sse2(const int16_t *src) {
+  __m128i vsum = _mm_setzero_si128();
+  int i;
+
+  for (i = 0; i < 32; ++i) {
+    const __m128i v = xx_loadu_128(src);
+    vsum = _mm_add_epi32(vsum, _mm_madd_epi16(v, v));
+    src += 8;
+  }
+
+  vsum = _mm_add_epi32(vsum, _mm_srli_si128(vsum, 8));
+  vsum = _mm_add_epi32(vsum, _mm_srli_si128(vsum, 4));
+  return (unsigned int)_mm_cvtsi128_si32(vsum);
+}
+
+static INLINE __m128i load4x2_sse2(const uint8_t *const p, const int stride) {
+  const __m128i p0 = _mm_cvtsi32_si128(loadu_int32(p + 0 * stride));
+  const __m128i p1 = _mm_cvtsi32_si128(loadu_int32(p + 1 * stride));
+  return _mm_unpacklo_epi8(_mm_unpacklo_epi32(p0, p1), _mm_setzero_si128());
+}
+
+static INLINE __m128i load8_8to16_sse2(const uint8_t *const p) {
+  const __m128i p0 = _mm_loadl_epi64((const __m128i *)p);
+  return _mm_unpacklo_epi8(p0, _mm_setzero_si128());
+}
+
+static INLINE void load16_8to16_sse2(const uint8_t *const p, __m128i *out) {
+  const __m128i p0 = _mm_loadu_si128((const __m128i *)p);
+  out[0] = _mm_unpacklo_epi8(p0, _mm_setzero_si128());  // lower 8 values
+  out[1] = _mm_unpackhi_epi8(p0, _mm_setzero_si128());  // upper 8 values
+}
+
+// Accumulate 4 32bit numbers in val to 1 32bit number
+static INLINE unsigned int add32x4_sse2(__m128i val) {
+  val = _mm_add_epi32(val, _mm_srli_si128(val, 8));
+  val = _mm_add_epi32(val, _mm_srli_si128(val, 4));
+  return (unsigned int)_mm_cvtsi128_si32(val);
+}
+
+// Accumulate 8 16bit in sum to 4 32bit number
+static INLINE __m128i sum_to_32bit_sse2(const __m128i sum) {
+  const __m128i sum_lo = _mm_srai_epi32(_mm_unpacklo_epi16(sum, sum), 16);
+  const __m128i sum_hi = _mm_srai_epi32(_mm_unpackhi_epi16(sum, sum), 16);
+  return _mm_add_epi32(sum_lo, sum_hi);
+}
+
+static INLINE void variance_kernel_sse2(const __m128i src, const __m128i ref,
+                                        __m128i *const sse,
+                                        __m128i *const sum) {
+  const __m128i diff = _mm_sub_epi16(src, ref);
+  *sse = _mm_add_epi32(*sse, _mm_madd_epi16(diff, diff));
+  *sum = _mm_add_epi16(*sum, diff);
+}
+
+// Can handle 128 pixels' diff sum (such as 8x16 or 16x8)
+// Slightly faster than variance_final_256_pel_sse2()
+// diff sum of 128 pixels can still fit in 16bit integer
+static INLINE void variance_final_128_pel_sse2(__m128i vsse, __m128i vsum,
+                                               unsigned int *const sse,
+                                               int *const sum) {
+  *sse = add32x4_sse2(vsse);
+
+  vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 8));
+  vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 4));
+  vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 2));
+  *sum = (int16_t)_mm_extract_epi16(vsum, 0);
+}
+
+// Can handle 256 pixels' diff sum (such as 16x16)
+static INLINE void variance_final_256_pel_sse2(__m128i vsse, __m128i vsum,
+                                               unsigned int *const sse,
+                                               int *const sum) {
+  *sse = add32x4_sse2(vsse);
+
+  vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 8));
+  vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 4));
+  *sum = (int16_t)_mm_extract_epi16(vsum, 0);
+  *sum += (int16_t)_mm_extract_epi16(vsum, 1);
+}
+
+// Can handle 512 pixels' diff sum (such as 16x32 or 32x16)
+static INLINE void variance_final_512_pel_sse2(__m128i vsse, __m128i vsum,
+                                               unsigned int *const sse,
+                                               int *const sum) {
+  *sse = add32x4_sse2(vsse);
+
+  vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 8));
+  vsum = _mm_unpacklo_epi16(vsum, vsum);
+  vsum = _mm_srai_epi32(vsum, 16);
+  *sum = (int)add32x4_sse2(vsum);
+}
+
+// Can handle 1024 pixels' diff sum (such as 32x32)
+static INLINE void variance_final_1024_pel_sse2(__m128i vsse, __m128i vsum,
+                                                unsigned int *const sse,
+                                                int *const sum) {
+  *sse = add32x4_sse2(vsse);
+
+  vsum = sum_to_32bit_sse2(vsum);
+  *sum = (int)add32x4_sse2(vsum);
+}
+
+static INLINE void variance4_sse2(const uint8_t *src, const int src_stride,
+                                  const uint8_t *ref, const int ref_stride,
+                                  const int h, __m128i *const sse,
+                                  __m128i *const sum) {
+  assert(h <= 256);  // May overflow for larger height.
+  *sum = _mm_setzero_si128();
+
+  for (int i = 0; i < h; i += 2) {
+    const __m128i s = load4x2_sse2(src, src_stride);
+    const __m128i r = load4x2_sse2(ref, ref_stride);
+
+    variance_kernel_sse2(s, r, sse, sum);
+    src += 2 * src_stride;
+    ref += 2 * ref_stride;
+  }
+}
+
+static INLINE void variance8_sse2(const uint8_t *src, const int src_stride,
+                                  const uint8_t *ref, const int ref_stride,
+                                  const int h, __m128i *const sse,
+                                  __m128i *const sum) {
+  assert(h <= 128);  // May overflow for larger height.
+  *sum = _mm_setzero_si128();
+  *sse = _mm_setzero_si128();
+  for (int i = 0; i < h; i++) {
+    const __m128i s = load8_8to16_sse2(src);
+    const __m128i r = load8_8to16_sse2(ref);
+
+    variance_kernel_sse2(s, r, sse, sum);
+    src += src_stride;
+    ref += ref_stride;
+  }
+}
+
+static INLINE void variance16_kernel_sse2(const uint8_t *const src,
+                                          const uint8_t *const ref,
+                                          __m128i *const sse,
+                                          __m128i *const sum) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i s = _mm_loadu_si128((const __m128i *)src);
+  const __m128i r = _mm_loadu_si128((const __m128i *)ref);
+  const __m128i src0 = _mm_unpacklo_epi8(s, zero);
+  const __m128i ref0 = _mm_unpacklo_epi8(r, zero);
+  const __m128i src1 = _mm_unpackhi_epi8(s, zero);
+  const __m128i ref1 = _mm_unpackhi_epi8(r, zero);
+
+  variance_kernel_sse2(src0, ref0, sse, sum);
+  variance_kernel_sse2(src1, ref1, sse, sum);
+}
+
+static INLINE void variance16_sse2(const uint8_t *src, const int src_stride,
+                                   const uint8_t *ref, const int ref_stride,
+                                   const int h, __m128i *const sse,
+                                   __m128i *const sum) {
+  assert(h <= 64);  // May overflow for larger height.
+  *sum = _mm_setzero_si128();
+
+  for (int i = 0; i < h; ++i) {
+    variance16_kernel_sse2(src, ref, sse, sum);
+    src += src_stride;
+    ref += ref_stride;
+  }
+}
+
+static INLINE void variance32_sse2(const uint8_t *src, const int src_stride,
+                                   const uint8_t *ref, const int ref_stride,
+                                   const int h, __m128i *const sse,
+                                   __m128i *const sum) {
+  assert(h <= 32);  // May overflow for larger height.
+  // Don't initialize sse here since it's an accumulation.
+  *sum = _mm_setzero_si128();
+
+  for (int i = 0; i < h; ++i) {
+    variance16_kernel_sse2(src + 0, ref + 0, sse, sum);
+    variance16_kernel_sse2(src + 16, ref + 16, sse, sum);
+    src += src_stride;
+    ref += ref_stride;
+  }
+}
+
+static INLINE void variance64_sse2(const uint8_t *src, const int src_stride,
+                                   const uint8_t *ref, const int ref_stride,
+                                   const int h, __m128i *const sse,
+                                   __m128i *const sum) {
+  assert(h <= 16);  // May overflow for larger height.
+  *sum = _mm_setzero_si128();
+
+  for (int i = 0; i < h; ++i) {
+    variance16_kernel_sse2(src + 0, ref + 0, sse, sum);
+    variance16_kernel_sse2(src + 16, ref + 16, sse, sum);
+    variance16_kernel_sse2(src + 32, ref + 32, sse, sum);
+    variance16_kernel_sse2(src + 48, ref + 48, sse, sum);
+    src += src_stride;
+    ref += ref_stride;
+  }
+}
+
+static INLINE void variance128_sse2(const uint8_t *src, const int src_stride,
+                                    const uint8_t *ref, const int ref_stride,
+                                    const int h, __m128i *const sse,
+                                    __m128i *const sum) {
+  assert(h <= 8);  // May overflow for larger height.
+  *sum = _mm_setzero_si128();
+
+  for (int i = 0; i < h; ++i) {
+    for (int j = 0; j < 4; ++j) {
+      const int offset0 = j << 5;
+      const int offset1 = offset0 + 16;
+      variance16_kernel_sse2(src + offset0, ref + offset0, sse, sum);
+      variance16_kernel_sse2(src + offset1, ref + offset1, sse, sum);
+    }
+    src += src_stride;
+    ref += ref_stride;
+  }
+}
+
+void aom_get_var_sse_sum_8x8_quad_sse2(const uint8_t *src_ptr, int src_stride,
+                                       const uint8_t *ref_ptr, int ref_stride,
+                                       uint32_t *sse8x8, int *sum8x8,
+                                       unsigned int *tot_sse, int *tot_sum,
+                                       uint32_t *var8x8) {
+  // Loop over 4 8x8 blocks. Process one 8x32 block.
+  for (int k = 0; k < 4; k++) {
+    const uint8_t *src = src_ptr;
+    const uint8_t *ref = ref_ptr;
+    __m128i vsum = _mm_setzero_si128();
+    __m128i vsse = _mm_setzero_si128();
+    for (int i = 0; i < 8; i++) {
+      const __m128i s = load8_8to16_sse2(src + (k * 8));
+      const __m128i r = load8_8to16_sse2(ref + (k * 8));
+      const __m128i diff = _mm_sub_epi16(s, r);
+      vsse = _mm_add_epi32(vsse, _mm_madd_epi16(diff, diff));
+      vsum = _mm_add_epi16(vsum, diff);
+
+      src += src_stride;
+      ref += ref_stride;
+    }
+    variance_final_128_pel_sse2(vsse, vsum, &sse8x8[k], &sum8x8[k]);
+  }
+
+  // Calculate variance at 8x8 level and total sse, sum of 8x32 block.
+  *tot_sse += sse8x8[0] + sse8x8[1] + sse8x8[2] + sse8x8[3];
+  *tot_sum += sum8x8[0] + sum8x8[1] + sum8x8[2] + sum8x8[3];
+  for (int i = 0; i < 4; i++)
+    var8x8[i] = sse8x8[i] - (uint32_t)(((int64_t)sum8x8[i] * sum8x8[i]) >> 6);
+}
+
+void aom_get_var_sse_sum_16x16_dual_sse2(const uint8_t *src_ptr, int src_stride,
+                                         const uint8_t *ref_ptr, int ref_stride,
+                                         uint32_t *sse16x16,
+                                         unsigned int *tot_sse, int *tot_sum,
+                                         uint32_t *var16x16) {
+  int sum16x16[2] = { 0 };
+  // Loop over 2 16x16 blocks. Process one 16x32 block.
+  for (int k = 0; k < 2; k++) {
+    const uint8_t *src = src_ptr;
+    const uint8_t *ref = ref_ptr;
+    __m128i vsum = _mm_setzero_si128();
+    __m128i vsse = _mm_setzero_si128();
+    for (int i = 0; i < 16; i++) {
+      __m128i s[2];
+      __m128i r[2];
+      load16_8to16_sse2(src + (k * 16), s);
+      load16_8to16_sse2(ref + (k * 16), r);
+      const __m128i diff0 = _mm_sub_epi16(s[0], r[0]);
+      const __m128i diff1 = _mm_sub_epi16(s[1], r[1]);
+      vsse = _mm_add_epi32(vsse, _mm_madd_epi16(diff0, diff0));
+      vsse = _mm_add_epi32(vsse, _mm_madd_epi16(diff1, diff1));
+      vsum = _mm_add_epi16(vsum, _mm_add_epi16(diff0, diff1));
+      src += src_stride;
+      ref += ref_stride;
+    }
+    variance_final_256_pel_sse2(vsse, vsum, &sse16x16[k], &sum16x16[k]);
+  }
+
+  // Calculate variance at 16x16 level and total sse, sum of 16x32 block.
+  *tot_sse += sse16x16[0] + sse16x16[1];
+  *tot_sum += sum16x16[0] + sum16x16[1];
+  for (int i = 0; i < 2; i++)
+    var16x16[i] =
+        sse16x16[i] - (uint32_t)(((int64_t)sum16x16[i] * sum16x16[i]) >> 8);
+}
+
+#define AOM_VAR_NO_LOOP_SSE2(bw, bh, bits, max_pixels)                        \
+  unsigned int aom_variance##bw##x##bh##_sse2(                                \
+      const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, \
+      unsigned int *sse) {                                                    \
+    __m128i vsse = _mm_setzero_si128();                                       \
+    __m128i vsum;                                                             \
+    int sum = 0;                                                              \
+    variance##bw##_sse2(src, src_stride, ref, ref_stride, bh, &vsse, &vsum);  \
+    variance_final_##max_pixels##_pel_sse2(vsse, vsum, sse, &sum);            \
+    assert(sum <= 255 * bw * bh);                                             \
+    assert(sum >= -255 * bw * bh);                                            \
+    return *sse - (uint32_t)(((int64_t)sum * sum) >> bits);                   \
+  }
+
+AOM_VAR_NO_LOOP_SSE2(4, 4, 4, 128)
+AOM_VAR_NO_LOOP_SSE2(4, 8, 5, 128)
+AOM_VAR_NO_LOOP_SSE2(4, 16, 6, 128)
+
+AOM_VAR_NO_LOOP_SSE2(8, 4, 5, 128)
+AOM_VAR_NO_LOOP_SSE2(8, 8, 6, 128)
+AOM_VAR_NO_LOOP_SSE2(8, 16, 7, 128)
+
+AOM_VAR_NO_LOOP_SSE2(16, 8, 7, 128)
+AOM_VAR_NO_LOOP_SSE2(16, 16, 8, 256)
+AOM_VAR_NO_LOOP_SSE2(16, 32, 9, 512)
+
+AOM_VAR_NO_LOOP_SSE2(32, 8, 8, 256)
+AOM_VAR_NO_LOOP_SSE2(32, 16, 9, 512)
+AOM_VAR_NO_LOOP_SSE2(32, 32, 10, 1024)
+
+#if !CONFIG_REALTIME_ONLY
+AOM_VAR_NO_LOOP_SSE2(16, 4, 6, 128)
+AOM_VAR_NO_LOOP_SSE2(8, 32, 8, 256)
+AOM_VAR_NO_LOOP_SSE2(16, 64, 10, 1024)
+#endif
+
+#define AOM_VAR_LOOP_SSE2(bw, bh, bits, uh)                                   \
+  unsigned int aom_variance##bw##x##bh##_sse2(                                \
+      const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, \
+      unsigned int *sse) {                                                    \
+    __m128i vsse = _mm_setzero_si128();                                       \
+    __m128i vsum = _mm_setzero_si128();                                       \
+    for (int i = 0; i < (bh / uh); ++i) {                                     \
+      __m128i vsum16;                                                         \
+      variance##bw##_sse2(src, src_stride, ref, ref_stride, uh, &vsse,        \
+                          &vsum16);                                           \
+      vsum = _mm_add_epi32(vsum, sum_to_32bit_sse2(vsum16));                  \
+      src += (src_stride * uh);                                               \
+      ref += (ref_stride * uh);                                               \
+    }                                                                         \
+    *sse = add32x4_sse2(vsse);                                                \
+    int sum = (int)add32x4_sse2(vsum);                                        \
+    assert(sum <= 255 * bw * bh);                                             \
+    assert(sum >= -255 * bw * bh);                                            \
+    return *sse - (uint32_t)(((int64_t)sum * sum) >> bits);                   \
+  }
+
+AOM_VAR_LOOP_SSE2(32, 64, 11, 32)  // 32x32 * ( 64/32 )
+
+AOM_VAR_LOOP_SSE2(64, 32, 11, 16)   // 64x16 * ( 32/16 )
+AOM_VAR_LOOP_SSE2(64, 64, 12, 16)   // 64x16 * ( 64/16 )
+AOM_VAR_LOOP_SSE2(64, 128, 13, 16)  // 64x16 * ( 128/16 )
+
+AOM_VAR_LOOP_SSE2(128, 64, 13, 8)   // 128x8 * ( 64/8 )
+AOM_VAR_LOOP_SSE2(128, 128, 14, 8)  // 128x8 * ( 128/8 )
+
+#if !CONFIG_REALTIME_ONLY
+AOM_VAR_NO_LOOP_SSE2(64, 16, 10, 1024)
+#endif
+
+unsigned int aom_mse8x8_sse2(const uint8_t *src, int src_stride,
+                             const uint8_t *ref, int ref_stride,
+                             unsigned int *sse) {
+  aom_variance8x8_sse2(src, src_stride, ref, ref_stride, sse);
+  return *sse;
+}
+
+unsigned int aom_mse8x16_sse2(const uint8_t *src, int src_stride,
+                              const uint8_t *ref, int ref_stride,
+                              unsigned int *sse) {
+  aom_variance8x16_sse2(src, src_stride, ref, ref_stride, sse);
+  return *sse;
+}
+
+unsigned int aom_mse16x8_sse2(const uint8_t *src, int src_stride,
+                              const uint8_t *ref, int ref_stride,
+                              unsigned int *sse) {
+  aom_variance16x8_sse2(src, src_stride, ref, ref_stride, sse);
+  return *sse;
+}
+
+unsigned int aom_mse16x16_sse2(const uint8_t *src, int src_stride,
+                               const uint8_t *ref, int ref_stride,
+                               unsigned int *sse) {
+  aom_variance16x16_sse2(src, src_stride, ref, ref_stride, sse);
+  return *sse;
+}
+
+// The 2 unused parameters are place holders for PIC enabled build.
+// These definitions are for functions defined in subpel_variance.asm
+#define DECL(w, opt)                                                           \
+  int aom_sub_pixel_variance##w##xh_##opt(                                     \
+      const uint8_t *src, ptrdiff_t src_stride, int x_offset, int y_offset,    \
+      const uint8_t *dst, ptrdiff_t dst_stride, int height, unsigned int *sse, \
+      void *unused0, void *unused)
+#define DECLS(opt) \
+  DECL(4, opt);    \
+  DECL(8, opt);    \
+  DECL(16, opt)
+
+DECLS(sse2);
+DECLS(ssse3);
+#undef DECLS
+#undef DECL
+
+#define FN(w, h, wf, wlog2, hlog2, opt, cast_prod, cast)                      \
+  unsigned int aom_sub_pixel_variance##w##x##h##_##opt(                       \
+      const uint8_t *src, int src_stride, int x_offset, int y_offset,         \
+      const uint8_t *dst, int dst_stride, unsigned int *sse_ptr) {            \
+    /*Avoid overflow in helper by capping height.*/                           \
+    const int hf = AOMMIN(h, 64);                                             \
+    unsigned int sse = 0;                                                     \
+    int se = 0;                                                               \
+    for (int i = 0; i < (w / wf); ++i) {                                      \
+      const uint8_t *src_ptr = src;                                           \
+      const uint8_t *dst_ptr = dst;                                           \
+      for (int j = 0; j < (h / hf); ++j) {                                    \
+        unsigned int sse2;                                                    \
+        const int se2 = aom_sub_pixel_variance##wf##xh_##opt(                 \
+            src_ptr, src_stride, x_offset, y_offset, dst_ptr, dst_stride, hf, \
+            &sse2, NULL, NULL);                                               \
+        dst_ptr += hf * dst_stride;                                           \
+        src_ptr += hf * src_stride;                                           \
+        se += se2;                                                            \
+        sse += sse2;                                                          \
+      }                                                                       \
+      src += wf;                                                              \
+      dst += wf;                                                              \
+    }                                                                         \
+    *sse_ptr = sse;                                                           \
+    return sse - (unsigned int)(cast_prod(cast se * se) >> (wlog2 + hlog2));  \
+  }
+
+#if !CONFIG_REALTIME_ONLY
+#define FNS(opt)                                    \
+  FN(128, 128, 16, 7, 7, opt, (int64_t), (int64_t)) \
+  FN(128, 64, 16, 7, 6, opt, (int64_t), (int64_t))  \
+  FN(64, 128, 16, 6, 7, opt, (int64_t), (int64_t))  \
+  FN(64, 64, 16, 6, 6, opt, (int64_t), (int64_t))   \
+  FN(64, 32, 16, 6, 5, opt, (int64_t), (int64_t))   \
+  FN(32, 64, 16, 5, 6, opt, (int64_t), (int64_t))   \
+  FN(32, 32, 16, 5, 5, opt, (int64_t), (int64_t))   \
+  FN(32, 16, 16, 5, 4, opt, (int64_t), (int64_t))   \
+  FN(16, 32, 16, 4, 5, opt, (int64_t), (int64_t))   \
+  FN(16, 16, 16, 4, 4, opt, (uint32_t), (int64_t))  \
+  FN(16, 8, 16, 4, 3, opt, (int32_t), (int32_t))    \
+  FN(8, 16, 8, 3, 4, opt, (int32_t), (int32_t))     \
+  FN(8, 8, 8, 3, 3, opt, (int32_t), (int32_t))      \
+  FN(8, 4, 8, 3, 2, opt, (int32_t), (int32_t))      \
+  FN(4, 8, 4, 2, 3, opt, (int32_t), (int32_t))      \
+  FN(4, 4, 4, 2, 2, opt, (int32_t), (int32_t))      \
+  FN(4, 16, 4, 2, 4, opt, (int32_t), (int32_t))     \
+  FN(16, 4, 16, 4, 2, opt, (int32_t), (int32_t))    \
+  FN(8, 32, 8, 3, 5, opt, (uint32_t), (int64_t))    \
+  FN(32, 8, 16, 5, 3, opt, (uint32_t), (int64_t))   \
+  FN(16, 64, 16, 4, 6, opt, (int64_t), (int64_t))   \
+  FN(64, 16, 16, 6, 4, opt, (int64_t), (int64_t))
+#else
+#define FNS(opt)                                    \
+  FN(128, 128, 16, 7, 7, opt, (int64_t), (int64_t)) \
+  FN(128, 64, 16, 7, 6, opt, (int64_t), (int64_t))  \
+  FN(64, 128, 16, 6, 7, opt, (int64_t), (int64_t))  \
+  FN(64, 64, 16, 6, 6, opt, (int64_t), (int64_t))   \
+  FN(64, 32, 16, 6, 5, opt, (int64_t), (int64_t))   \
+  FN(32, 64, 16, 5, 6, opt, (int64_t), (int64_t))   \
+  FN(32, 32, 16, 5, 5, opt, (int64_t), (int64_t))   \
+  FN(32, 16, 16, 5, 4, opt, (int64_t), (int64_t))   \
+  FN(16, 32, 16, 4, 5, opt, (int64_t), (int64_t))   \
+  FN(16, 16, 16, 4, 4, opt, (uint32_t), (int64_t))  \
+  FN(16, 8, 16, 4, 3, opt, (int32_t), (int32_t))    \
+  FN(8, 16, 8, 3, 4, opt, (int32_t), (int32_t))     \
+  FN(8, 8, 8, 3, 3, opt, (int32_t), (int32_t))      \
+  FN(8, 4, 8, 3, 2, opt, (int32_t), (int32_t))      \
+  FN(4, 8, 4, 2, 3, opt, (int32_t), (int32_t))      \
+  FN(4, 4, 4, 2, 2, opt, (int32_t), (int32_t))
+#endif
+
+FNS(sse2)
+FNS(ssse3)
+
+#undef FNS
+#undef FN
+
+// The 2 unused parameters are place holders for PIC enabled build.
+#define DECL(w, opt)                                                        \
+  int aom_sub_pixel_avg_variance##w##xh_##opt(                              \
+      const uint8_t *src, ptrdiff_t src_stride, int x_offset, int y_offset, \
+      const uint8_t *dst, ptrdiff_t dst_stride, const uint8_t *sec,         \
+      ptrdiff_t sec_stride, int height, unsigned int *sse, void *unused0,   \
+      void *unused)
+#define DECLS(opt) \
+  DECL(4, opt);    \
+  DECL(8, opt);    \
+  DECL(16, opt)
+
+DECLS(sse2);
+DECLS(ssse3);
+#undef DECL
+#undef DECLS
+
+#define FN(w, h, wf, wlog2, hlog2, opt, cast_prod, cast)                     \
+  unsigned int aom_sub_pixel_avg_variance##w##x##h##_##opt(                  \
+      const uint8_t *src, int src_stride, int x_offset, int y_offset,        \
+      const uint8_t *dst, int dst_stride, unsigned int *sse_ptr,             \
+      const uint8_t *sec) {                                                  \
+    /*Avoid overflow in helper by capping height.*/                          \
+    const int hf = AOMMIN(h, 64);                                            \
+    unsigned int sse = 0;                                                    \
+    int se = 0;                                                              \
+    for (int i = 0; i < (w / wf); ++i) {                                     \
+      const uint8_t *src_ptr = src;                                          \
+      const uint8_t *dst_ptr = dst;                                          \
+      const uint8_t *sec_ptr = sec;                                          \
+      for (int j = 0; j < (h / hf); ++j) {                                   \
+        unsigned int sse2;                                                   \
+        const int se2 = aom_sub_pixel_avg_variance##wf##xh_##opt(            \
+            src_ptr, src_stride, x_offset, y_offset, dst_ptr, dst_stride,    \
+            sec_ptr, w, hf, &sse2, NULL, NULL);                              \
+        dst_ptr += hf * dst_stride;                                          \
+        src_ptr += hf * src_stride;                                          \
+        sec_ptr += hf * w;                                                   \
+        se += se2;                                                           \
+        sse += sse2;                                                         \
+      }                                                                      \
+      src += wf;                                                             \
+      dst += wf;                                                             \
+      sec += wf;                                                             \
+    }                                                                        \
+    *sse_ptr = sse;                                                          \
+    return sse - (unsigned int)(cast_prod(cast se * se) >> (wlog2 + hlog2)); \
+  }
+
+#if !CONFIG_REALTIME_ONLY
+#define FNS(opt)                                    \
+  FN(128, 128, 16, 7, 7, opt, (int64_t), (int64_t)) \
+  FN(128, 64, 16, 7, 6, opt, (int64_t), (int64_t))  \
+  FN(64, 128, 16, 6, 7, opt, (int64_t), (int64_t))  \
+  FN(64, 64, 16, 6, 6, opt, (int64_t), (int64_t))   \
+  FN(64, 32, 16, 6, 5, opt, (int64_t), (int64_t))   \
+  FN(32, 64, 16, 5, 6, opt, (int64_t), (int64_t))   \
+  FN(32, 32, 16, 5, 5, opt, (int64_t), (int64_t))   \
+  FN(32, 16, 16, 5, 4, opt, (int64_t), (int64_t))   \
+  FN(16, 32, 16, 4, 5, opt, (int64_t), (int64_t))   \
+  FN(16, 16, 16, 4, 4, opt, (uint32_t), (int64_t))  \
+  FN(16, 8, 16, 4, 3, opt, (uint32_t), (int32_t))   \
+  FN(8, 16, 8, 3, 4, opt, (uint32_t), (int32_t))    \
+  FN(8, 8, 8, 3, 3, opt, (uint32_t), (int32_t))     \
+  FN(8, 4, 8, 3, 2, opt, (uint32_t), (int32_t))     \
+  FN(4, 8, 4, 2, 3, opt, (uint32_t), (int32_t))     \
+  FN(4, 4, 4, 2, 2, opt, (uint32_t), (int32_t))     \
+  FN(4, 16, 4, 2, 4, opt, (int32_t), (int32_t))     \
+  FN(16, 4, 16, 4, 2, opt, (int32_t), (int32_t))    \
+  FN(8, 32, 8, 3, 5, opt, (uint32_t), (int64_t))    \
+  FN(32, 8, 16, 5, 3, opt, (uint32_t), (int64_t))   \
+  FN(16, 64, 16, 4, 6, opt, (int64_t), (int64_t))   \
+  FN(64, 16, 16, 6, 4, opt, (int64_t), (int64_t))
+#else
+#define FNS(opt)                                    \
+  FN(128, 128, 16, 7, 7, opt, (int64_t), (int64_t)) \
+  FN(128, 64, 16, 7, 6, opt, (int64_t), (int64_t))  \
+  FN(64, 128, 16, 6, 7, opt, (int64_t), (int64_t))  \
+  FN(64, 64, 16, 6, 6, opt, (int64_t), (int64_t))   \
+  FN(64, 32, 16, 6, 5, opt, (int64_t), (int64_t))   \
+  FN(32, 64, 16, 5, 6, opt, (int64_t), (int64_t))   \
+  FN(32, 32, 16, 5, 5, opt, (int64_t), (int64_t))   \
+  FN(32, 16, 16, 5, 4, opt, (int64_t), (int64_t))   \
+  FN(16, 32, 16, 4, 5, opt, (int64_t), (int64_t))   \
+  FN(16, 16, 16, 4, 4, opt, (uint32_t), (int64_t))  \
+  FN(16, 8, 16, 4, 3, opt, (uint32_t), (int32_t))   \
+  FN(8, 16, 8, 3, 4, opt, (uint32_t), (int32_t))    \
+  FN(8, 8, 8, 3, 3, opt, (uint32_t), (int32_t))     \
+  FN(8, 4, 8, 3, 2, opt, (uint32_t), (int32_t))     \
+  FN(4, 8, 4, 2, 3, opt, (uint32_t), (int32_t))     \
+  FN(4, 4, 4, 2, 2, opt, (uint32_t), (int32_t))
+#endif
+
+FNS(sse2)
+FNS(ssse3)
+
+#undef FNS
+#undef FN
+
+static INLINE __m128i highbd_comp_mask_pred_line_sse2(const __m128i s0,
+                                                      const __m128i s1,
+                                                      const __m128i a) {
+  const __m128i alpha_max = _mm_set1_epi16((1 << AOM_BLEND_A64_ROUND_BITS));
+  const __m128i round_const =
+      _mm_set1_epi32((1 << AOM_BLEND_A64_ROUND_BITS) >> 1);
+  const __m128i a_inv = _mm_sub_epi16(alpha_max, a);
+
+  const __m128i s_lo = _mm_unpacklo_epi16(s0, s1);
+  const __m128i a_lo = _mm_unpacklo_epi16(a, a_inv);
+  const __m128i pred_lo = _mm_madd_epi16(s_lo, a_lo);
+  const __m128i pred_l = _mm_srai_epi32(_mm_add_epi32(pred_lo, round_const),
+                                        AOM_BLEND_A64_ROUND_BITS);
+
+  const __m128i s_hi = _mm_unpackhi_epi16(s0, s1);
+  const __m128i a_hi = _mm_unpackhi_epi16(a, a_inv);
+  const __m128i pred_hi = _mm_madd_epi16(s_hi, a_hi);
+  const __m128i pred_h = _mm_srai_epi32(_mm_add_epi32(pred_hi, round_const),
+                                        AOM_BLEND_A64_ROUND_BITS);
+
+  const __m128i comp = _mm_packs_epi32(pred_l, pred_h);
+
+  return comp;
+}
+
+void aom_highbd_comp_mask_pred_sse2(uint8_t *comp_pred8, 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 *comp_pred = CONVERT_TO_SHORTPTR(comp_pred8);
+  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 __m128i zero = _mm_setzero_si128();
+
+  if (width == 8) {
+    do {
+      const __m128i s0 = _mm_loadu_si128((const __m128i *)(src0));
+      const __m128i s1 = _mm_loadu_si128((const __m128i *)(src1));
+      const __m128i m_8 = _mm_loadl_epi64((const __m128i *)mask);
+      const __m128i m_16 = _mm_unpacklo_epi8(m_8, zero);
+
+      const __m128i comp = highbd_comp_mask_pred_line_sse2(s0, s1, m_16);
+
+      _mm_storeu_si128((__m128i *)comp_pred, comp);
+
+      src0 += stride0;
+      src1 += stride1;
+      mask += mask_stride;
+      comp_pred += width;
+      i += 1;
+    } while (i < height);
+  } else if (width == 16) {
+    do {
+      const __m128i s0 = _mm_loadu_si128((const __m128i *)(src0));
+      const __m128i s2 = _mm_loadu_si128((const __m128i *)(src0 + 8));
+      const __m128i s1 = _mm_loadu_si128((const __m128i *)(src1));
+      const __m128i s3 = _mm_loadu_si128((const __m128i *)(src1 + 8));
+
+      const __m128i m_8 = _mm_loadu_si128((const __m128i *)mask);
+      const __m128i m01_16 = _mm_unpacklo_epi8(m_8, zero);
+      const __m128i m23_16 = _mm_unpackhi_epi8(m_8, zero);
+
+      const __m128i comp = highbd_comp_mask_pred_line_sse2(s0, s1, m01_16);
+      const __m128i comp1 = highbd_comp_mask_pred_line_sse2(s2, s3, m23_16);
+
+      _mm_storeu_si128((__m128i *)comp_pred, comp);
+      _mm_storeu_si128((__m128i *)(comp_pred + 8), comp1);
+
+      src0 += stride0;
+      src1 += stride1;
+      mask += mask_stride;
+      comp_pred += width;
+      i += 1;
+    } while (i < height);
+  } else {
+    do {
+      for (int x = 0; x < width; x += 32) {
+        for (int j = 0; j < 2; j++) {
+          const __m128i s0 =
+              _mm_loadu_si128((const __m128i *)(src0 + x + j * 16));
+          const __m128i s2 =
+              _mm_loadu_si128((const __m128i *)(src0 + x + 8 + j * 16));
+          const __m128i s1 =
+              _mm_loadu_si128((const __m128i *)(src1 + x + j * 16));
+          const __m128i s3 =
+              _mm_loadu_si128((const __m128i *)(src1 + x + 8 + j * 16));
+
+          const __m128i m_8 =
+              _mm_loadu_si128((const __m128i *)(mask + x + j * 16));
+          const __m128i m01_16 = _mm_unpacklo_epi8(m_8, zero);
+          const __m128i m23_16 = _mm_unpackhi_epi8(m_8, zero);
+
+          const __m128i comp = highbd_comp_mask_pred_line_sse2(s0, s1, m01_16);
+          const __m128i comp1 = highbd_comp_mask_pred_line_sse2(s2, s3, m23_16);
+
+          _mm_storeu_si128((__m128i *)(comp_pred + j * 16), comp);
+          _mm_storeu_si128((__m128i *)(comp_pred + 8 + j * 16), comp1);
+        }
+        comp_pred += 32;
+      }
+      src0 += stride0;
+      src1 += stride1;
+      mask += mask_stride;
+      i += 1;
+    } while (i < height);
+  }
+}
+
+uint64_t aom_mse_4xh_16bit_sse2(uint8_t *dst, int dstride, uint16_t *src,
+                                int sstride, int h) {
+  uint64_t sum = 0;
+  __m128i dst0_8x8, dst1_8x8, dst_16x8;
+  __m128i src0_16x4, src1_16x4, src_16x8;
+  __m128i res0_32x4, res0_64x2, res1_64x2;
+  __m128i sub_result_16x8;
+  const __m128i zeros = _mm_setzero_si128();
+  __m128i square_result = _mm_setzero_si128();
+  for (int i = 0; i < h; i += 2) {
+    dst0_8x8 = _mm_cvtsi32_si128(*(int const *)(&dst[(i + 0) * dstride]));
+    dst1_8x8 = _mm_cvtsi32_si128(*(int const *)(&dst[(i + 1) * dstride]));
+    dst_16x8 = _mm_unpacklo_epi8(_mm_unpacklo_epi32(dst0_8x8, dst1_8x8), zeros);
+
+    src0_16x4 = _mm_loadl_epi64((__m128i const *)(&src[(i + 0) * sstride]));
+    src1_16x4 = _mm_loadl_epi64((__m128i const *)(&src[(i + 1) * sstride]));
+    src_16x8 = _mm_unpacklo_epi64(src0_16x4, src1_16x4);
+
+    sub_result_16x8 = _mm_sub_epi16(src_16x8, dst_16x8);
+
+    res0_32x4 = _mm_madd_epi16(sub_result_16x8, sub_result_16x8);
+
+    res0_64x2 = _mm_unpacklo_epi32(res0_32x4, zeros);
+    res1_64x2 = _mm_unpackhi_epi32(res0_32x4, zeros);
+
+    square_result =
+        _mm_add_epi64(square_result, _mm_add_epi64(res0_64x2, res1_64x2));
+  }
+  const __m128i sum_64x1 =
+      _mm_add_epi64(square_result, _mm_srli_si128(square_result, 8));
+  xx_storel_64(&sum, sum_64x1);
+  return sum;
+}
+
+uint64_t aom_mse_8xh_16bit_sse2(uint8_t *dst, int dstride, uint16_t *src,
+                                int sstride, int h) {
+  uint64_t sum = 0;
+  __m128i dst_8x8, dst_16x8;
+  __m128i src_16x8;
+  __m128i res0_32x4, res0_64x2, res1_64x2;
+  __m128i sub_result_16x8;
+  const __m128i zeros = _mm_setzero_si128();
+  __m128i square_result = _mm_setzero_si128();
+
+  for (int i = 0; i < h; i++) {
+    dst_8x8 = _mm_loadl_epi64((__m128i const *)(&dst[(i + 0) * dstride]));
+    dst_16x8 = _mm_unpacklo_epi8(dst_8x8, zeros);
+
+    src_16x8 = _mm_loadu_si128((__m128i *)&src[i * sstride]);
+
+    sub_result_16x8 = _mm_sub_epi16(src_16x8, dst_16x8);
+
+    res0_32x4 = _mm_madd_epi16(sub_result_16x8, sub_result_16x8);
+
+    res0_64x2 = _mm_unpacklo_epi32(res0_32x4, zeros);
+    res1_64x2 = _mm_unpackhi_epi32(res0_32x4, zeros);
+
+    square_result =
+        _mm_add_epi64(square_result, _mm_add_epi64(res0_64x2, res1_64x2));
+  }
+  const __m128i sum_64x1 =
+      _mm_add_epi64(square_result, _mm_srli_si128(square_result, 8));
+  xx_storel_64(&sum, sum_64x1);
+  return sum;
+}
+
+uint64_t aom_mse_wxh_16bit_sse2(uint8_t *dst, int dstride, uint16_t *src,
+                                int sstride, int w, int h) {
+  assert((w == 8 || w == 4) && (h == 8 || h == 4) &&
+         "w=8/4 and h=8/4 must satisfy");
+  switch (w) {
+    case 4: return aom_mse_4xh_16bit_sse2(dst, dstride, src, sstride, h);
+    case 8: return aom_mse_8xh_16bit_sse2(dst, dstride, src, sstride, h);
+    default: assert(0 && "unsupported width"); return -1;
+  }
+}
+
+uint64_t aom_mse_16xh_16bit_sse2(uint8_t *dst, int dstride, uint16_t *src,
+                                 int w, int h) {
+  assert((w == 8 || w == 4) && (h == 8 || h == 4) &&
+         "w=8/4 and h=8/4 must be satisfied");
+  const int num_blks = 16 / w;
+  uint64_t sum = 0;
+  for (int i = 0; i < num_blks; i++) {
+    sum += aom_mse_wxh_16bit_sse2(dst, dstride, src, w, w, h);
+    dst += w;
+    src += (w * h);
+  }
+  return sum;
+}