Merge branch 'master' into Basilisk-releasev2018.11.04

43 files changed, 24693 insertions, 0 deletions

diff --git a/third_party/aom/av1/common/x86/av1_convolve_horiz_rs_sse4.c b/third_party/aom/av1/common/x86/av1_convolve_horiz_rs_sse4.c
new file mode 100644
index 000000000..8aa14696f
--- /dev/null
+++ b/third_party/aom/av1/common/x86/av1_convolve_horiz_rs_sse4.c
@@ -0,0 +1,228 @@
+/*
+ * 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/av1_rtcd.h"
+
+#include "av1/common/convolve.h"
+#include "av1/common/resize.h"
+#include "aom_dsp/x86/synonyms.h"
+
+// Note: If the crop width is not a multiple of 4, then, unlike the C version,
+// this function will overwrite some of the padding on the right hand side of
+// the frame. This padding appears to be trashed anyway, so this should not
+// affect the running of the decoder.
+void av1_convolve_horiz_rs_sse4_1(const uint8_t *src, int src_stride,
+                                  uint8_t *dst, int dst_stride, int w, int h,
+                                  const int16_t *x_filters, int x0_qn,
+                                  int x_step_qn) {
+  assert(UPSCALE_NORMATIVE_TAPS == 8);
+
+  src -= UPSCALE_NORMATIVE_TAPS / 2 - 1;
+
+  const __m128i round_add = _mm_set1_epi32((1 << FILTER_BITS) >> 1);
+  const __m128i zero = _mm_setzero_si128();
+
+  const uint8_t *src_y;
+  uint8_t *dst_y;
+  int x_qn = x0_qn;
+  for (int x = 0; x < w; x += 4, x_qn += 4 * x_step_qn) {
+    const int x_filter_idx0 =
+        ((x_qn + 0 * x_step_qn) & RS_SCALE_SUBPEL_MASK) >> RS_SCALE_EXTRA_BITS;
+    const int x_filter_idx1 =
+        ((x_qn + 1 * x_step_qn) & RS_SCALE_SUBPEL_MASK) >> RS_SCALE_EXTRA_BITS;
+    const int x_filter_idx2 =
+        ((x_qn + 2 * x_step_qn) & RS_SCALE_SUBPEL_MASK) >> RS_SCALE_EXTRA_BITS;
+    const int x_filter_idx3 =
+        ((x_qn + 3 * x_step_qn) & RS_SCALE_SUBPEL_MASK) >> RS_SCALE_EXTRA_BITS;
+
+    assert(x_filter_idx0 <= RS_SUBPEL_MASK);
+    assert(x_filter_idx1 <= RS_SUBPEL_MASK);
+    assert(x_filter_idx2 <= RS_SUBPEL_MASK);
+    assert(x_filter_idx3 <= RS_SUBPEL_MASK);
+
+    const int16_t *const x_filter0 =
+        &x_filters[x_filter_idx0 * UPSCALE_NORMATIVE_TAPS];
+    const int16_t *const x_filter1 =
+        &x_filters[x_filter_idx1 * UPSCALE_NORMATIVE_TAPS];
+    const int16_t *const x_filter2 =
+        &x_filters[x_filter_idx2 * UPSCALE_NORMATIVE_TAPS];
+    const int16_t *const x_filter3 =
+        &x_filters[x_filter_idx3 * UPSCALE_NORMATIVE_TAPS];
+
+    const __m128i fil0_16 = xx_loadu_128(x_filter0);
+    const __m128i fil1_16 = xx_loadu_128(x_filter1);
+    const __m128i fil2_16 = xx_loadu_128(x_filter2);
+    const __m128i fil3_16 = xx_loadu_128(x_filter3);
+
+    src_y = src;
+    dst_y = dst;
+    for (int y = 0; y < h; y++, src_y += src_stride, dst_y += dst_stride) {
+      const uint8_t *const src_x0 =
+          &src_y[(x_qn + 0 * x_step_qn) >> RS_SCALE_SUBPEL_BITS];
+      const uint8_t *const src_x1 =
+          &src_y[(x_qn + 1 * x_step_qn) >> RS_SCALE_SUBPEL_BITS];
+      const uint8_t *const src_x2 =
+          &src_y[(x_qn + 2 * x_step_qn) >> RS_SCALE_SUBPEL_BITS];
+      const uint8_t *const src_x3 =
+          &src_y[(x_qn + 3 * x_step_qn) >> RS_SCALE_SUBPEL_BITS];
+
+      // Load up the source data. This is 8-bit input data, so each load
+      // gets 8 pixels.
+      const __m128i src0_8 = xx_loadl_64(src_x0);
+      const __m128i src1_8 = xx_loadl_64(src_x1);
+      const __m128i src2_8 = xx_loadl_64(src_x2);
+      const __m128i src3_8 = xx_loadl_64(src_x3);
+
+      // Now zero-extend up to 16-bit precision, i.e.
+      // [ 00 00 00 00 hg fe dc ba ] -> [ 0h 0g 0f 0e 0d 0c 0b 0a ]
+      const __m128i src0_16 = _mm_cvtepu8_epi16(src0_8);
+      const __m128i src1_16 = _mm_cvtepu8_epi16(src1_8);
+      const __m128i src2_16 = _mm_cvtepu8_epi16(src2_8);
+      const __m128i src3_16 = _mm_cvtepu8_epi16(src3_8);
+
+      // Multiply by filter coefficients (results in a 32-bit value),
+      // and add adjacent pairs, i.e.
+      // ([ s7 s6 s5 s4 s3 s2 s1 s0], [ f7 f6 f5 f4 f3 f2 f1 f0 ])
+      // -> [ {s7*f7+s6*f6} {s5*f5+s4*f4} {s3*f3+s2*f2} {s1*f1+s0*f0} ]
+      const __m128i conv0_32 = _mm_madd_epi16(src0_16, fil0_16);
+      const __m128i conv1_32 = _mm_madd_epi16(src1_16, fil1_16);
+      const __m128i conv2_32 = _mm_madd_epi16(src2_16, fil2_16);
+      const __m128i conv3_32 = _mm_madd_epi16(src3_16, fil3_16);
+
+      // Reduce horizontally and add, i.e.
+      // ([ D C B A ], [ S R Q P ]) -> [ S+R Q+P D+C B+A ]
+      const __m128i conv01_32 = _mm_hadd_epi32(conv0_32, conv1_32);
+      const __m128i conv23_32 = _mm_hadd_epi32(conv2_32, conv3_32);
+
+      const __m128i conv0123_32 = _mm_hadd_epi32(conv01_32, conv23_32);
+
+      // Divide down by (1 << FILTER_BITS), rounding to nearest.
+      const __m128i shifted_32 =
+          _mm_srai_epi32(_mm_add_epi32(conv0123_32, round_add), FILTER_BITS);
+
+      // Pack 32-bit values into 16-bit values, i.e.
+      // ([ D C B A ], [ 0 0 0 0 ]) -> [ 0 0 0 0 D C B A ]
+      const __m128i shifted_16 = _mm_packus_epi32(shifted_32, zero);
+
+      // Pack 16-bit values into 8-bit values, i.e.
+      // ([ 0 0 0 0 D C B A ], [ 0 0 0 0 0 0 0 0 ])
+      // -> [ 0 0 0 0 0 0 DC BA ]
+      const __m128i shifted_8 = _mm_packus_epi16(shifted_16, zero);
+
+      // Write to the output
+      xx_storel_32(&dst_y[x], shifted_8);
+    }
+  }
+}
+
+// Note: If the crop width is not a multiple of 4, then, unlike the C version,
+// this function will overwrite some of the padding on the right hand side of
+// the frame. This padding appears to be trashed anyway, so this should not
+// affect the running of the decoder.
+void av1_highbd_convolve_horiz_rs_sse4_1(const uint16_t *src, int src_stride,
+                                         uint16_t *dst, int dst_stride, int w,
+                                         int h, const int16_t *x_filters,
+                                         int x0_qn, int x_step_qn, int bd) {
+  assert(UPSCALE_NORMATIVE_TAPS == 8);
+  assert(bd == 8 || bd == 10 || bd == 12);
+
+  src -= UPSCALE_NORMATIVE_TAPS / 2 - 1;
+
+  const __m128i round_add = _mm_set1_epi32((1 << FILTER_BITS) >> 1);
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i clip_maximum = _mm_set1_epi16((1 << bd) - 1);
+
+  const uint16_t *src_y;
+  uint16_t *dst_y;
+  int x_qn = x0_qn;
+  for (int x = 0; x < w; x += 4, x_qn += 4 * x_step_qn) {
+    const int x_filter_idx0 =
+        ((x_qn + 0 * x_step_qn) & RS_SCALE_SUBPEL_MASK) >> RS_SCALE_EXTRA_BITS;
+    const int x_filter_idx1 =
+        ((x_qn + 1 * x_step_qn) & RS_SCALE_SUBPEL_MASK) >> RS_SCALE_EXTRA_BITS;
+    const int x_filter_idx2 =
+        ((x_qn + 2 * x_step_qn) & RS_SCALE_SUBPEL_MASK) >> RS_SCALE_EXTRA_BITS;
+    const int x_filter_idx3 =
+        ((x_qn + 3 * x_step_qn) & RS_SCALE_SUBPEL_MASK) >> RS_SCALE_EXTRA_BITS;
+
+    assert(x_filter_idx0 <= RS_SUBPEL_MASK);
+    assert(x_filter_idx1 <= RS_SUBPEL_MASK);
+    assert(x_filter_idx2 <= RS_SUBPEL_MASK);
+    assert(x_filter_idx3 <= RS_SUBPEL_MASK);
+
+    const int16_t *const x_filter0 =
+        &x_filters[x_filter_idx0 * UPSCALE_NORMATIVE_TAPS];
+    const int16_t *const x_filter1 =
+        &x_filters[x_filter_idx1 * UPSCALE_NORMATIVE_TAPS];
+    const int16_t *const x_filter2 =
+        &x_filters[x_filter_idx2 * UPSCALE_NORMATIVE_TAPS];
+    const int16_t *const x_filter3 =
+        &x_filters[x_filter_idx3 * UPSCALE_NORMATIVE_TAPS];
+
+    const __m128i fil0_16 = xx_loadu_128(x_filter0);
+    const __m128i fil1_16 = xx_loadu_128(x_filter1);
+    const __m128i fil2_16 = xx_loadu_128(x_filter2);
+    const __m128i fil3_16 = xx_loadu_128(x_filter3);
+
+    src_y = src;
+    dst_y = dst;
+    for (int y = 0; y < h; y++, src_y += src_stride, dst_y += dst_stride) {
+      const uint16_t *const src_x0 =
+          &src_y[(x_qn + 0 * x_step_qn) >> RS_SCALE_SUBPEL_BITS];
+      const uint16_t *const src_x1 =
+          &src_y[(x_qn + 1 * x_step_qn) >> RS_SCALE_SUBPEL_BITS];
+      const uint16_t *const src_x2 =
+          &src_y[(x_qn + 2 * x_step_qn) >> RS_SCALE_SUBPEL_BITS];
+      const uint16_t *const src_x3 =
+          &src_y[(x_qn + 3 * x_step_qn) >> RS_SCALE_SUBPEL_BITS];
+
+      // Load up the source data. This is 16-bit input data, so each load
+      // gets 8 pixels.
+      const __m128i src0_16 = xx_loadu_128(src_x0);
+      const __m128i src1_16 = xx_loadu_128(src_x1);
+      const __m128i src2_16 = xx_loadu_128(src_x2);
+      const __m128i src3_16 = xx_loadu_128(src_x3);
+
+      // Multiply by filter coefficients (results in a 32-bit value),
+      // and add adjacent pairs, i.e.
+      // ([ s7 s6 s5 s4 s3 s2 s1 s0], [ f7 f6 f5 f4 f3 f2 f1 f0 ])
+      // -> [ {s7*f7+s6*f6} {s5*f5+s4*f4} {s3*f3+s2*f2} {s1*f1+s0*f0} ]
+      const __m128i conv0_32 = _mm_madd_epi16(src0_16, fil0_16);
+      const __m128i conv1_32 = _mm_madd_epi16(src1_16, fil1_16);
+      const __m128i conv2_32 = _mm_madd_epi16(src2_16, fil2_16);
+      const __m128i conv3_32 = _mm_madd_epi16(src3_16, fil3_16);
+
+      // Reduce horizontally and add, i.e.
+      // ([ D C B A ], [ S R Q P ]) -> [ S+R Q+P D+C B+A ]
+      const __m128i conv01_32 = _mm_hadd_epi32(conv0_32, conv1_32);
+      const __m128i conv23_32 = _mm_hadd_epi32(conv2_32, conv3_32);
+
+      const __m128i conv0123_32 = _mm_hadd_epi32(conv01_32, conv23_32);
+
+      // Divide down by (1 << FILTER_BITS), rounding to nearest.
+      const __m128i shifted_32 =
+          _mm_srai_epi32(_mm_add_epi32(conv0123_32, round_add), FILTER_BITS);
+
+      // Pack 32-bit values into 16-bit values, i.e.
+      // ([ D C B A ], [ 0 0 0 0 ]) -> [ 0 0 0 0 D C B A ]
+      const __m128i shifted_16 = _mm_packus_epi32(shifted_32, zero);
+
+      // Clip the values at (1 << bd) - 1
+      const __m128i clipped_16 = _mm_min_epi16(shifted_16, clip_maximum);
+
+      // Write to the output
+      xx_storel_64(&dst_y[x], clipped_16);
+    }
+  }
+}
diff --git a/third_party/aom/av1/common/x86/av1_convolve_scale_sse4.c b/third_party/aom/av1/common/x86/av1_convolve_scale_sse4.c
new file mode 100644
index 000000000..d9fb53785
--- /dev/null
+++ b/third_party/aom/av1/common/x86/av1_convolve_scale_sse4.c
@@ -0,0 +1,499 @@
+/*
+ * 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 <assert.h>
+#include <smmintrin.h>
+
+#include "config/aom_dsp_rtcd.h"
+
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/aom_filter.h"
+#include "av1/common/convolve.h"
+
+// A specialised version of hfilter, the horizontal filter for
+// av1_convolve_2d_scale_sse4_1. This version only supports 8 tap filters.
+static void hfilter8(const uint8_t *src, int src_stride, int16_t *dst, int w,
+                     int h, int subpel_x_qn, int x_step_qn,
+                     const InterpFilterParams *filter_params, unsigned round) {
+  const int bd = 8;
+  const int ntaps = 8;
+
+  src -= ntaps / 2 - 1;
+
+  int32_t round_add32 = (1 << round) / 2 + (1 << (bd + FILTER_BITS - 1));
+  const __m128i round_add = _mm_set1_epi32(round_add32);
+  const __m128i round_shift = _mm_cvtsi32_si128(round);
+
+  int x_qn = subpel_x_qn;
+  for (int x = 0; x < w; ++x, x_qn += x_step_qn) {
+    const uint8_t *const src_col = src + (x_qn >> SCALE_SUBPEL_BITS);
+    const int filter_idx = (x_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS;
+    assert(filter_idx < SUBPEL_SHIFTS);
+    const int16_t *filter =
+        av1_get_interp_filter_subpel_kernel(filter_params, filter_idx);
+
+    // Load the filter coefficients
+    const __m128i coefflo = _mm_loadu_si128((__m128i *)filter);
+    const __m128i zero = _mm_castps_si128(_mm_setzero_ps());
+
+    int y;
+    for (y = 0; y <= h - 4; y += 4) {
+      const uint8_t *const src0 = src_col + y * src_stride;
+      const uint8_t *const src1 = src0 + 1 * src_stride;
+      const uint8_t *const src2 = src0 + 2 * src_stride;
+      const uint8_t *const src3 = src0 + 3 * src_stride;
+
+      // Load up source data. This is 8-bit input data; each load is just
+      // loading the lower half of the register and gets 8 pixels
+      const __m128i data08 = _mm_loadl_epi64((__m128i *)src0);
+      const __m128i data18 = _mm_loadl_epi64((__m128i *)src1);
+      const __m128i data28 = _mm_loadl_epi64((__m128i *)src2);
+      const __m128i data38 = _mm_loadl_epi64((__m128i *)src3);
+
+      // Now zero-extend up to 16-bit precision by interleaving with
+      // zeros. Drop the upper half of each register (which just had zeros)
+      const __m128i data0lo = _mm_unpacklo_epi8(data08, zero);
+      const __m128i data1lo = _mm_unpacklo_epi8(data18, zero);
+      const __m128i data2lo = _mm_unpacklo_epi8(data28, zero);
+      const __m128i data3lo = _mm_unpacklo_epi8(data38, zero);
+
+      // Multiply by coefficients
+      const __m128i conv0lo = _mm_madd_epi16(data0lo, coefflo);
+      const __m128i conv1lo = _mm_madd_epi16(data1lo, coefflo);
+      const __m128i conv2lo = _mm_madd_epi16(data2lo, coefflo);
+      const __m128i conv3lo = _mm_madd_epi16(data3lo, coefflo);
+
+      // Reduce horizontally and add
+      const __m128i conv01lo = _mm_hadd_epi32(conv0lo, conv1lo);
+      const __m128i conv23lo = _mm_hadd_epi32(conv2lo, conv3lo);
+      const __m128i conv = _mm_hadd_epi32(conv01lo, conv23lo);
+
+      // Divide down by (1 << round), rounding to nearest.
+      __m128i shifted =
+          _mm_sra_epi32(_mm_add_epi32(conv, round_add), round_shift);
+
+      shifted = _mm_packus_epi32(shifted, shifted);
+      // Write transposed to the output
+      _mm_storel_epi64((__m128i *)(dst + y + x * h), shifted);
+    }
+    for (; y < h; ++y) {
+      const uint8_t *const src_row = src_col + y * src_stride;
+
+      int32_t sum = (1 << (bd + FILTER_BITS - 1));
+      for (int k = 0; k < ntaps; ++k) {
+        sum += filter[k] * src_row[k];
+      }
+
+      dst[y + x * h] = ROUND_POWER_OF_TWO(sum, round);
+    }
+  }
+}
+
+static __m128i convolve_16_8(const int16_t *src, __m128i coeff) {
+  __m128i data = _mm_loadu_si128((__m128i *)src);
+  return _mm_madd_epi16(data, coeff);
+}
+
+// A specialised version of vfilter, the vertical filter for
+// av1_convolve_2d_scale_sse4_1. This version only supports 8 tap filters.
+static void vfilter8(const int16_t *src, int src_stride, uint8_t *dst,
+                     int dst_stride, int w, int h, int subpel_y_qn,
+                     int y_step_qn, const InterpFilterParams *filter_params,
+                     const ConvolveParams *conv_params, int bd) {
+  const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0;
+  const int ntaps = 8;
+
+  const __m128i round_shift = _mm_cvtsi32_si128(conv_params->round_1);
+
+  const int32_t sub32 = ((1 << (offset_bits - conv_params->round_1)) +
+                         (1 << (offset_bits - conv_params->round_1 - 1)));
+  const __m128i sub = _mm_set1_epi16(sub32);
+
+  CONV_BUF_TYPE *dst16 = conv_params->dst;
+  const int dst16_stride = conv_params->dst_stride;
+  const int bits =
+      FILTER_BITS * 2 - conv_params->round_0 - conv_params->round_1;
+  const __m128i bits_shift = _mm_cvtsi32_si128(bits);
+  const __m128i bits_const = _mm_set1_epi16(((1 << bits) >> 1));
+  const __m128i round_shift_add =
+      _mm_set1_epi32(((1 << conv_params->round_1) >> 1));
+  const __m128i res_add_const = _mm_set1_epi32(1 << offset_bits);
+
+  const int w0 = conv_params->fwd_offset;
+  const int w1 = conv_params->bck_offset;
+  const __m128i wt0 = _mm_set1_epi16(w0);
+  const __m128i wt1 = _mm_set1_epi16(w1);
+  const __m128i wt = _mm_unpacklo_epi16(wt0, wt1);
+
+  int y_qn = subpel_y_qn;
+  for (int y = 0; y < h; ++y, y_qn += y_step_qn) {
+    const int16_t *src_y = src + (y_qn >> SCALE_SUBPEL_BITS);
+    const int filter_idx = (y_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS;
+    assert(filter_idx < SUBPEL_SHIFTS);
+    const int16_t *filter =
+        av1_get_interp_filter_subpel_kernel(filter_params, filter_idx);
+
+    const __m128i coeff0716 = _mm_loadu_si128((__m128i *)filter);
+    int x;
+    for (x = 0; x <= w - 4; x += 4) {
+      const int16_t *const src0 = src_y + x * src_stride;
+      const int16_t *const src1 = src0 + 1 * src_stride;
+      const int16_t *const src2 = src0 + 2 * src_stride;
+      const int16_t *const src3 = src0 + 3 * src_stride;
+
+      // Load the source data for the three rows, adding the three registers of
+      // convolved products to one as we go (conv0..conv3) to avoid the
+      // register pressure getting too high.
+      const __m128i conv0 = convolve_16_8(src0, coeff0716);
+      const __m128i conv1 = convolve_16_8(src1, coeff0716);
+      const __m128i conv2 = convolve_16_8(src2, coeff0716);
+      const __m128i conv3 = convolve_16_8(src3, coeff0716);
+
+      // Now reduce horizontally to get one lane for each result
+      const __m128i conv01 = _mm_hadd_epi32(conv0, conv1);
+      const __m128i conv23 = _mm_hadd_epi32(conv2, conv3);
+      __m128i conv = _mm_hadd_epi32(conv01, conv23);
+
+      conv = _mm_add_epi32(conv, res_add_const);
+      // Divide down by (1 << round_1), rounding to nearest and subtract sub32.
+      __m128i shifted =
+          _mm_sra_epi32(_mm_add_epi32(conv, round_shift_add), round_shift);
+
+      uint8_t *dst_x = dst + y * dst_stride + x;
+      CONV_BUF_TYPE *dst_16_x = dst16 + y * dst16_stride + x;
+      __m128i result;
+      __m128i shifted_16 = _mm_packus_epi32(shifted, shifted);
+
+      if (conv_params->is_compound) {
+        if (conv_params->do_average) {
+          const __m128i p_16 = _mm_loadl_epi64((__m128i *)dst_16_x);
+          if (conv_params->use_jnt_comp_avg) {
+            const __m128i p_16_lo = _mm_unpacklo_epi16(p_16, shifted_16);
+            const __m128i wt_res_lo = _mm_madd_epi16(p_16_lo, wt);
+            const __m128i shifted_32 =
+                _mm_srai_epi32(wt_res_lo, DIST_PRECISION_BITS);
+            shifted_16 = _mm_packus_epi32(shifted_32, shifted_32);
+          } else {
+            shifted_16 = _mm_srai_epi16(_mm_add_epi16(p_16, shifted_16), 1);
+          }
+          const __m128i subbed = _mm_sub_epi16(shifted_16, sub);
+          result = _mm_sra_epi16(_mm_add_epi16(subbed, bits_const), bits_shift);
+          const __m128i result_8 = _mm_packus_epi16(result, result);
+          *(uint32_t *)dst_x = _mm_cvtsi128_si32(result_8);
+        } else {
+          _mm_storel_epi64((__m128i *)dst_16_x, shifted_16);
+        }
+      } else {
+        const __m128i subbed = _mm_sub_epi16(shifted_16, sub);
+        result = _mm_sra_epi16(_mm_add_epi16(subbed, bits_const), bits_shift);
+        const __m128i result_8 = _mm_packus_epi16(result, result);
+        *(uint32_t *)dst_x = _mm_cvtsi128_si32(result_8);
+      }
+    }
+    for (; x < w; ++x) {
+      const int16_t *src_x = src_y + x * src_stride;
+      int32_t sum = 1 << offset_bits;
+      for (int k = 0; k < ntaps; ++k) sum += filter[k] * src_x[k];
+      CONV_BUF_TYPE res = ROUND_POWER_OF_TWO(sum, conv_params->round_1);
+
+      if (conv_params->is_compound) {
+        if (conv_params->do_average) {
+          int32_t tmp = dst16[y * dst16_stride + x];
+          if (conv_params->use_jnt_comp_avg) {
+            tmp = tmp * conv_params->fwd_offset + res * conv_params->bck_offset;
+            tmp = tmp >> DIST_PRECISION_BITS;
+          } else {
+            tmp += res;
+            tmp = tmp >> 1;
+          }
+          /* Subtract round offset and convolve round */
+          tmp = tmp - sub32;
+          dst[y * dst_stride + x] = clip_pixel(ROUND_POWER_OF_TWO(tmp, bits));
+        } else {
+          dst16[y * dst16_stride + x] = res;
+        }
+      } else {
+        /* Subtract round offset and convolve round */
+        int32_t tmp = res - ((1 << (offset_bits - conv_params->round_1)) +
+                             (1 << (offset_bits - conv_params->round_1 - 1)));
+        dst[y * dst_stride + x] = clip_pixel(ROUND_POWER_OF_TWO(tmp, bits));
+      }
+    }
+  }
+}
+void av1_convolve_2d_scale_sse4_1(const uint8_t *src, int src_stride,
+                                  uint8_t *dst8, int dst8_stride, int w, int h,
+                                  const InterpFilterParams *filter_params_x,
+                                  const InterpFilterParams *filter_params_y,
+                                  const int subpel_x_qn, const int x_step_qn,
+                                  const int subpel_y_qn, const int y_step_qn,
+                                  ConvolveParams *conv_params) {
+  // TODO(yaowu): remove unnecessary initializations
+  int16_t tmp[(2 * MAX_SB_SIZE + MAX_FILTER_TAP) * MAX_SB_SIZE] = { 0 };
+  int im_h = (((h - 1) * y_step_qn + subpel_y_qn) >> SCALE_SUBPEL_BITS) +
+             filter_params_y->taps;
+
+  const int xtaps = filter_params_x->taps;
+  const int ytaps = filter_params_y->taps;
+  const int fo_vert = ytaps / 2 - 1;
+  assert((xtaps == 8) && (ytaps == 8));
+  (void)xtaps;
+
+  // horizontal filter
+  hfilter8(src - fo_vert * src_stride, src_stride, tmp, w, im_h, subpel_x_qn,
+           x_step_qn, filter_params_x, conv_params->round_0);
+
+  // vertical filter (input is transposed)
+  vfilter8(tmp, im_h, dst8, dst8_stride, w, h, subpel_y_qn, y_step_qn,
+           filter_params_y, conv_params, 8);
+}
+
+// A specialised version of hfilter, the horizontal filter for
+// av1_highbd_convolve_2d_scale_sse4_1. This version only supports 8 tap
+// filters.
+static void highbd_hfilter8(const uint16_t *src, int src_stride, int16_t *dst,
+                            int w, int h, int subpel_x_qn, int x_step_qn,
+                            const InterpFilterParams *filter_params,
+                            unsigned round, int bd) {
+  const int ntaps = 8;
+
+  src -= ntaps / 2 - 1;
+
+  int32_t round_add32 = (1 << round) / 2 + (1 << (bd + FILTER_BITS - 1));
+  const __m128i round_add = _mm_set1_epi32(round_add32);
+  const __m128i round_shift = _mm_cvtsi32_si128(round);
+
+  int x_qn = subpel_x_qn;
+  for (int x = 0; x < w; ++x, x_qn += x_step_qn) {
+    const uint16_t *const src_col = src + (x_qn >> SCALE_SUBPEL_BITS);
+    const int filter_idx = (x_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS;
+    assert(filter_idx < SUBPEL_SHIFTS);
+    const int16_t *filter =
+        av1_get_interp_filter_subpel_kernel(filter_params, filter_idx);
+
+    // Load the filter coefficients
+    const __m128i coefflo = _mm_loadu_si128((__m128i *)filter);
+
+    int y;
+    for (y = 0; y <= h - 4; y += 4) {
+      const uint16_t *const src0 = src_col + y * src_stride;
+      const uint16_t *const src1 = src0 + 1 * src_stride;
+      const uint16_t *const src2 = src0 + 2 * src_stride;
+      const uint16_t *const src3 = src0 + 3 * src_stride;
+
+      // Load up source data. This is 16-bit input data, so each load gets the 8
+      // pixels we need.
+      const __m128i data0lo = _mm_loadu_si128((__m128i *)src0);
+      const __m128i data1lo = _mm_loadu_si128((__m128i *)src1);
+      const __m128i data2lo = _mm_loadu_si128((__m128i *)src2);
+      const __m128i data3lo = _mm_loadu_si128((__m128i *)src3);
+
+      // Multiply by coefficients
+      const __m128i conv0lo = _mm_madd_epi16(data0lo, coefflo);
+      const __m128i conv1lo = _mm_madd_epi16(data1lo, coefflo);
+      const __m128i conv2lo = _mm_madd_epi16(data2lo, coefflo);
+      const __m128i conv3lo = _mm_madd_epi16(data3lo, coefflo);
+
+      // Reduce horizontally and add
+      const __m128i conv01lo = _mm_hadd_epi32(conv0lo, conv1lo);
+      const __m128i conv23lo = _mm_hadd_epi32(conv2lo, conv3lo);
+      const __m128i conv = _mm_hadd_epi32(conv01lo, conv23lo);
+
+      // Divide down by (1 << round), rounding to nearest.
+      __m128i shifted =
+          _mm_sra_epi32(_mm_add_epi32(conv, round_add), round_shift);
+
+      shifted = _mm_packus_epi32(shifted, shifted);
+      // Write transposed to the output
+      _mm_storel_epi64((__m128i *)(dst + y + x * h), shifted);
+    }
+    for (; y < h; ++y) {
+      const uint16_t *const src_row = src_col + y * src_stride;
+
+      int32_t sum = (1 << (bd + FILTER_BITS - 1));
+      for (int k = 0; k < ntaps; ++k) {
+        sum += filter[k] * src_row[k];
+      }
+
+      dst[y + x * h] = ROUND_POWER_OF_TWO(sum, round);
+    }
+  }
+}
+// A specialised version of vfilter, the vertical filter for
+// av1_highbd_convolve_2d_scale_sse4_1. This version only supports 8 tap
+// filters.
+static void highbd_vfilter8(const int16_t *src, int src_stride, uint16_t *dst,
+                            int dst_stride, int w, int h, int subpel_y_qn,
+                            int y_step_qn,
+                            const InterpFilterParams *filter_params,
+                            const ConvolveParams *conv_params, int bd) {
+  const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0;
+  const int ntaps = 8;
+
+  const __m128i round_shift = _mm_cvtsi32_si128(conv_params->round_1);
+
+  const int32_t sub32 = ((1 << (offset_bits - conv_params->round_1)) +
+                         (1 << (offset_bits - conv_params->round_1 - 1)));
+  const __m128i sub = _mm_set1_epi32(sub32);
+
+  CONV_BUF_TYPE *dst16 = conv_params->dst;
+  const int dst16_stride = conv_params->dst_stride;
+  const __m128i clip_pixel_ =
+      _mm_set1_epi16(bd == 10 ? 1023 : (bd == 12 ? 4095 : 255));
+  const int bits =
+      FILTER_BITS * 2 - conv_params->round_0 - conv_params->round_1;
+  const __m128i bits_shift = _mm_cvtsi32_si128(bits);
+  const __m128i bits_const = _mm_set1_epi32(((1 << bits) >> 1));
+  const __m128i round_shift_add =
+      _mm_set1_epi32(((1 << conv_params->round_1) >> 1));
+  const __m128i res_add_const = _mm_set1_epi32(1 << offset_bits);
+  const int round_bits =
+      2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+  __m128i round_bits_shift = _mm_cvtsi32_si128(round_bits);
+  __m128i round_bits_const = _mm_set1_epi32(((1 << round_bits) >> 1));
+
+  const int w0 = conv_params->fwd_offset;
+  const int w1 = conv_params->bck_offset;
+  const __m128i wt0 = _mm_set1_epi32(w0);
+  const __m128i wt1 = _mm_set1_epi32(w1);
+
+  int y_qn = subpel_y_qn;
+  for (int y = 0; y < h; ++y, y_qn += y_step_qn) {
+    const int16_t *src_y = src + (y_qn >> SCALE_SUBPEL_BITS);
+    const int filter_idx = (y_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS;
+    assert(filter_idx < SUBPEL_SHIFTS);
+    const int16_t *filter =
+        av1_get_interp_filter_subpel_kernel(filter_params, filter_idx);
+
+    const __m128i coeff0716 = _mm_loadu_si128((__m128i *)filter);
+    int x;
+    for (x = 0; x <= w - 4; x += 4) {
+      const int16_t *const src0 = src_y + x * src_stride;
+      const int16_t *const src1 = src0 + 1 * src_stride;
+      const int16_t *const src2 = src0 + 2 * src_stride;
+      const int16_t *const src3 = src0 + 3 * src_stride;
+
+      // Load the source data for the three rows, adding the three registers of
+      // convolved products to one as we go (conv0..conv3) to avoid the
+      // register pressure getting too high.
+      const __m128i conv0 = convolve_16_8(src0, coeff0716);
+      const __m128i conv1 = convolve_16_8(src1, coeff0716);
+      const __m128i conv2 = convolve_16_8(src2, coeff0716);
+      const __m128i conv3 = convolve_16_8(src3, coeff0716);
+
+      // Now reduce horizontally to get one lane for each result
+      const __m128i conv01 = _mm_hadd_epi32(conv0, conv1);
+      const __m128i conv23 = _mm_hadd_epi32(conv2, conv3);
+      __m128i conv = _mm_hadd_epi32(conv01, conv23);
+      conv = _mm_add_epi32(conv, res_add_const);
+
+      // Divide down by (1 << round_1), rounding to nearest and subtract sub32.
+      __m128i shifted =
+          _mm_sra_epi32(_mm_add_epi32(conv, round_shift_add), round_shift);
+
+      uint16_t *dst_x = dst + y * dst_stride + x;
+      CONV_BUF_TYPE *dst_16_x = dst16 + y * dst16_stride + x;
+
+      __m128i result;
+      if (conv_params->is_compound) {
+        if (conv_params->do_average) {
+          __m128i p_32 =
+              _mm_cvtepu16_epi32(_mm_loadl_epi64((__m128i *)dst_16_x));
+
+          if (conv_params->use_jnt_comp_avg) {
+            shifted = _mm_add_epi32(_mm_mullo_epi32(p_32, wt0),
+                                    _mm_mullo_epi32(shifted, wt1));
+            shifted = _mm_srai_epi32(shifted, DIST_PRECISION_BITS);
+          } else {
+            shifted = _mm_srai_epi32(_mm_add_epi32(p_32, shifted), 1);
+          }
+          __m128i res32 = _mm_sub_epi32(shifted, sub);
+          res32 = _mm_sra_epi32(_mm_add_epi32(res32, round_bits_const),
+                                round_bits_shift);
+
+          __m128i res16 = _mm_packus_epi32(res32, res32);
+          res16 = _mm_min_epi16(res16, clip_pixel_);
+          _mm_storel_epi64((__m128i *)dst_x, res16);
+        } else {
+          __m128i shifted_16 = _mm_packus_epi32(shifted, shifted);
+          _mm_storel_epi64((__m128i *)dst_16_x, shifted_16);
+        }
+      } else {
+        const __m128i subbed = _mm_sub_epi32(shifted, sub);
+        result = _mm_sra_epi16(_mm_add_epi32(subbed, bits_const), bits_shift);
+        result = _mm_packus_epi32(result, result);
+        result = _mm_min_epi16(result, clip_pixel_);
+        _mm_storel_epi64((__m128i *)dst_x, result);
+      }
+    }
+
+    for (; x < w; ++x) {
+      const int16_t *src_x = src_y + x * src_stride;
+      int32_t sum = 1 << offset_bits;
+      for (int k = 0; k < ntaps; ++k) sum += filter[k] * src_x[k];
+      CONV_BUF_TYPE res = ROUND_POWER_OF_TWO(sum, conv_params->round_1);
+      if (conv_params->is_compound) {
+        if (conv_params->do_average) {
+          int32_t tmp = dst16[y * dst16_stride + x];
+          if (conv_params->use_jnt_comp_avg) {
+            tmp = tmp * conv_params->fwd_offset + res * conv_params->bck_offset;
+            tmp = tmp >> DIST_PRECISION_BITS;
+          } else {
+            tmp += res;
+            tmp = tmp >> 1;
+          }
+          /* Subtract round offset and convolve round */
+          tmp = tmp - ((1 << (offset_bits - conv_params->round_1)) +
+                       (1 << (offset_bits - conv_params->round_1 - 1)));
+          dst[y * dst_stride + x] =
+              clip_pixel_highbd(ROUND_POWER_OF_TWO(tmp, bits), bd);
+        } else {
+          dst16[y * dst16_stride + x] = res;
+        }
+      } else {
+        /* Subtract round offset and convolve round */
+        int32_t tmp = res - ((1 << (offset_bits - conv_params->round_1)) +
+                             (1 << (offset_bits - conv_params->round_1 - 1)));
+        dst[y * dst_stride + x] =
+            clip_pixel_highbd(ROUND_POWER_OF_TWO(tmp, bits), bd);
+      }
+    }
+  }
+}
+
+void av1_highbd_convolve_2d_scale_sse4_1(
+    const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w,
+    int h, const InterpFilterParams *filter_params_x,
+    const InterpFilterParams *filter_params_y, const int subpel_x_qn,
+    const int x_step_qn, const int subpel_y_qn, const int y_step_qn,
+    ConvolveParams *conv_params, int bd) {
+  // TODO(yaowu): Move this out of stack
+  DECLARE_ALIGNED(16, int16_t,
+                  tmp[(2 * MAX_SB_SIZE + MAX_FILTER_TAP) * MAX_SB_SIZE]);
+  int im_h = (((h - 1) * y_step_qn + subpel_y_qn) >> SCALE_SUBPEL_BITS) +
+             filter_params_y->taps;
+  const int xtaps = filter_params_x->taps;
+  const int ytaps = filter_params_y->taps;
+  const int fo_vert = ytaps / 2 - 1;
+
+  memset(tmp, 0, sizeof(tmp));
+  assert((xtaps == 8) && (ytaps == 8));
+  (void)xtaps;
+
+  // horizontal filter
+  highbd_hfilter8(src - fo_vert * src_stride, src_stride, tmp, w, im_h,
+                  subpel_x_qn, x_step_qn, filter_params_x, conv_params->round_0,
+                  bd);
+
+  // vertical filter (input is transposed)
+  highbd_vfilter8(tmp, im_h, dst, dst_stride, w, h, subpel_y_qn, y_step_qn,
+                  filter_params_y, conv_params, bd);
+}
diff --git a/third_party/aom/av1/common/x86/av1_highbd_convolve_sse4.c b/third_party/aom/av1/common/x86/av1_highbd_convolve_sse4.c
new file mode 100644
index 000000000..212d3bd72
--- /dev/null
+++ b/third_party/aom/av1/common/x86/av1_highbd_convolve_sse4.c
@@ -0,0 +1,205 @@
+/*
+ * 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 <smmintrin.h>
+
+#include "config/av1_rtcd.h"
+
+#include "av1/common/filter.h"
+
+typedef void (*TransposeSave)(int width, int pixelsNum, uint32_t *src,
+                              int src_stride, uint16_t *dst, int dst_stride,
+                              int bd);
+
+// pixelsNum 0: write all 4 pixels
+//           1/2/3: residual pixels 1/2/3
+static void writePixel(__m128i *u, int width, int pixelsNum, uint16_t *dst,
+                       int dst_stride) {
+  if (2 == width) {
+    if (0 == pixelsNum) {
+      *(int *)dst = _mm_cvtsi128_si32(u[0]);
+      *(int *)(dst + dst_stride) = _mm_cvtsi128_si32(u[1]);
+      *(int *)(dst + 2 * dst_stride) = _mm_cvtsi128_si32(u[2]);
+      *(int *)(dst + 3 * dst_stride) = _mm_cvtsi128_si32(u[3]);
+    } else if (1 == pixelsNum) {
+      *(int *)dst = _mm_cvtsi128_si32(u[0]);
+    } else if (2 == pixelsNum) {
+      *(int *)dst = _mm_cvtsi128_si32(u[0]);
+      *(int *)(dst + dst_stride) = _mm_cvtsi128_si32(u[1]);
+    } else if (3 == pixelsNum) {
+      *(int *)dst = _mm_cvtsi128_si32(u[0]);
+      *(int *)(dst + dst_stride) = _mm_cvtsi128_si32(u[1]);
+      *(int *)(dst + 2 * dst_stride) = _mm_cvtsi128_si32(u[2]);
+    }
+  } else {
+    if (0 == pixelsNum) {
+      _mm_storel_epi64((__m128i *)dst, u[0]);
+      _mm_storel_epi64((__m128i *)(dst + dst_stride), u[1]);
+      _mm_storel_epi64((__m128i *)(dst + 2 * dst_stride), u[2]);
+      _mm_storel_epi64((__m128i *)(dst + 3 * dst_stride), u[3]);
+    } else if (1 == pixelsNum) {
+      _mm_storel_epi64((__m128i *)dst, u[0]);
+    } else if (2 == pixelsNum) {
+      _mm_storel_epi64((__m128i *)dst, u[0]);
+      _mm_storel_epi64((__m128i *)(dst + dst_stride), u[1]);
+    } else if (3 == pixelsNum) {
+      _mm_storel_epi64((__m128i *)dst, u[0]);
+      _mm_storel_epi64((__m128i *)(dst + dst_stride), u[1]);
+      _mm_storel_epi64((__m128i *)(dst + 2 * dst_stride), u[2]);
+    }
+  }
+}
+
+// 16-bit pixels clip with bd (10/12)
+static void highbd_clip(__m128i *p, int numVecs, int bd) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i one = _mm_set1_epi16(1);
+  const __m128i max = _mm_sub_epi16(_mm_slli_epi16(one, bd), one);
+  __m128i clamped, mask;
+  int i;
+
+  for (i = 0; i < numVecs; i++) {
+    mask = _mm_cmpgt_epi16(p[i], max);
+    clamped = _mm_andnot_si128(mask, p[i]);
+    mask = _mm_and_si128(mask, max);
+    clamped = _mm_or_si128(mask, clamped);
+    mask = _mm_cmpgt_epi16(clamped, zero);
+    p[i] = _mm_and_si128(clamped, mask);
+  }
+}
+
+static void transClipPixel(uint32_t *src, int src_stride, __m128i *u, int bd) {
+  __m128i v0, v1;
+  __m128i rnd = _mm_set1_epi32(1 << (FILTER_BITS - 1));
+
+  u[0] = _mm_loadu_si128((__m128i const *)src);
+  u[1] = _mm_loadu_si128((__m128i const *)(src + src_stride));
+  u[2] = _mm_loadu_si128((__m128i const *)(src + 2 * src_stride));
+  u[3] = _mm_loadu_si128((__m128i const *)(src + 3 * src_stride));
+
+  u[0] = _mm_add_epi32(u[0], rnd);
+  u[1] = _mm_add_epi32(u[1], rnd);
+  u[2] = _mm_add_epi32(u[2], rnd);
+  u[3] = _mm_add_epi32(u[3], rnd);
+
+  u[0] = _mm_srai_epi32(u[0], FILTER_BITS);
+  u[1] = _mm_srai_epi32(u[1], FILTER_BITS);
+  u[2] = _mm_srai_epi32(u[2], FILTER_BITS);
+  u[3] = _mm_srai_epi32(u[3], FILTER_BITS);
+
+  u[0] = _mm_packus_epi32(u[0], u[1]);
+  u[1] = _mm_packus_epi32(u[2], u[3]);
+
+  highbd_clip(u, 2, bd);
+
+  v0 = _mm_unpacklo_epi16(u[0], u[1]);
+  v1 = _mm_unpackhi_epi16(u[0], u[1]);
+
+  u[0] = _mm_unpacklo_epi16(v0, v1);
+  u[2] = _mm_unpackhi_epi16(v0, v1);
+
+  u[1] = _mm_srli_si128(u[0], 8);
+  u[3] = _mm_srli_si128(u[2], 8);
+}
+
+// pixelsNum = 0     : all 4 rows of pixels will be saved.
+// pixelsNum = 1/2/3 : residual 1/2/4 rows of pixels will be saved.
+void trans_save_4x4(int width, int pixelsNum, uint32_t *src, int src_stride,
+                    uint16_t *dst, int dst_stride, int bd) {
+  __m128i u[4];
+  transClipPixel(src, src_stride, u, bd);
+  writePixel(u, width, pixelsNum, dst, dst_stride);
+}
+
+void trans_accum_save_4x4(int width, int pixelsNum, uint32_t *src,
+                          int src_stride, uint16_t *dst, int dst_stride,
+                          int bd) {
+  __m128i u[4], v[4];
+  const __m128i ones = _mm_set1_epi16(1);
+
+  transClipPixel(src, src_stride, u, bd);
+
+  v[0] = _mm_loadl_epi64((__m128i const *)dst);
+  v[1] = _mm_loadl_epi64((__m128i const *)(dst + dst_stride));
+  v[2] = _mm_loadl_epi64((__m128i const *)(dst + 2 * dst_stride));
+  v[3] = _mm_loadl_epi64((__m128i const *)(dst + 3 * dst_stride));
+
+  u[0] = _mm_add_epi16(u[0], v[0]);
+  u[1] = _mm_add_epi16(u[1], v[1]);
+  u[2] = _mm_add_epi16(u[2], v[2]);
+  u[3] = _mm_add_epi16(u[3], v[3]);
+
+  u[0] = _mm_add_epi16(u[0], ones);
+  u[1] = _mm_add_epi16(u[1], ones);
+  u[2] = _mm_add_epi16(u[2], ones);
+  u[3] = _mm_add_epi16(u[3], ones);
+
+  u[0] = _mm_srai_epi16(u[0], 1);
+  u[1] = _mm_srai_epi16(u[1], 1);
+  u[2] = _mm_srai_epi16(u[2], 1);
+  u[3] = _mm_srai_epi16(u[3], 1);
+
+  writePixel(u, width, pixelsNum, dst, dst_stride);
+}
+
+// Vertical convolutional filter
+
+typedef void (*WritePixels)(__m128i *u, int bd, uint16_t *dst);
+
+static void highbdRndingPacks(__m128i *u) {
+  __m128i rnd = _mm_set1_epi32(1 << (FILTER_BITS - 1));
+  u[0] = _mm_add_epi32(u[0], rnd);
+  u[0] = _mm_srai_epi32(u[0], FILTER_BITS);
+  u[0] = _mm_packus_epi32(u[0], u[0]);
+}
+
+static void write2pixelsOnly(__m128i *u, int bd, uint16_t *dst) {
+  highbdRndingPacks(u);
+  highbd_clip(u, 1, bd);
+  *(uint32_t *)dst = _mm_cvtsi128_si32(u[0]);
+}
+
+static void write2pixelsAccum(__m128i *u, int bd, uint16_t *dst) {
+  __m128i v = _mm_loadl_epi64((__m128i const *)dst);
+  const __m128i ones = _mm_set1_epi16(1);
+
+  highbdRndingPacks(u);
+  highbd_clip(u, 1, bd);
+
+  v = _mm_add_epi16(v, u[0]);
+  v = _mm_add_epi16(v, ones);
+  v = _mm_srai_epi16(v, 1);
+  *(uint32_t *)dst = _mm_cvtsi128_si32(v);
+}
+
+WritePixels write2pixelsTab[2] = { write2pixelsOnly, write2pixelsAccum };
+
+static void write4pixelsOnly(__m128i *u, int bd, uint16_t *dst) {
+  highbdRndingPacks(u);
+  highbd_clip(u, 1, bd);
+  _mm_storel_epi64((__m128i *)dst, u[0]);
+}
+
+static void write4pixelsAccum(__m128i *u, int bd, uint16_t *dst) {
+  __m128i v = _mm_loadl_epi64((__m128i const *)dst);
+  const __m128i ones = _mm_set1_epi16(1);
+
+  highbdRndingPacks(u);
+  highbd_clip(u, 1, bd);
+
+  v = _mm_add_epi16(v, u[0]);
+  v = _mm_add_epi16(v, ones);
+  v = _mm_srai_epi16(v, 1);
+  _mm_storel_epi64((__m128i *)dst, v);
+}
+
+WritePixels write4pixelsTab[2] = { write4pixelsOnly, write4pixelsAccum };
diff --git a/third_party/aom/av1/common/x86/av1_inv_txfm_avx2.c b/third_party/aom/av1/common/x86/av1_inv_txfm_avx2.c
new file mode 100644
index 000000000..5db2ccf6c
--- /dev/null
+++ b/third_party/aom/av1/common/x86/av1_inv_txfm_avx2.c
@@ -0,0 +1,1945 @@
+/*
+ * 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 "config/aom_config.h"
+
+#include "config/av1_rtcd.h"
+
+#include "av1/common/av1_inv_txfm1d_cfg.h"
+#include "av1/common/x86/av1_txfm_sse2.h"
+#include "av1/common/x86/av1_inv_txfm_avx2.h"
+#include "av1/common/x86/av1_inv_txfm_ssse3.h"
+
+// TODO(venkatsanampudi@ittiam.com): move this to header file
+
+// Sqrt2, Sqrt2^2, Sqrt2^3, Sqrt2^4, Sqrt2^5
+static int32_t NewSqrt2list[TX_SIZES] = { 5793, 2 * 4096, 2 * 5793, 4 * 4096,
+                                          4 * 5793 };
+
+static INLINE void idct16_stage5_avx2(__m256i *x1, const int32_t *cospi,
+                                      const __m256i _r, int8_t cos_bit) {
+  const __m256i cospi_m32_p32 = pair_set_w16_epi16(-cospi[32], cospi[32]);
+  const __m256i cospi_p32_p32 = pair_set_w16_epi16(cospi[32], cospi[32]);
+  btf_16_adds_subs_avx2(&x1[0], &x1[3]);
+  btf_16_adds_subs_avx2(&x1[1], &x1[2]);
+  btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x1[5], &x1[6], _r, cos_bit);
+
+  btf_16_adds_subs_avx2(&x1[8], &x1[11]);
+  btf_16_adds_subs_avx2(&x1[9], &x1[10]);
+  btf_16_adds_subs_avx2(&x1[15], &x1[12]);
+  btf_16_adds_subs_avx2(&x1[14], &x1[13]);
+}
+
+static INLINE void idct16_stage6_avx2(__m256i *x, const int32_t *cospi,
+                                      const __m256i _r, int8_t cos_bit) {
+  const __m256i cospi_m32_p32 = pair_set_w16_epi16(-cospi[32], cospi[32]);
+  const __m256i cospi_p32_p32 = pair_set_w16_epi16(cospi[32], cospi[32]);
+  btf_16_adds_subs_avx2(&x[0], &x[7]);
+  btf_16_adds_subs_avx2(&x[1], &x[6]);
+  btf_16_adds_subs_avx2(&x[2], &x[5]);
+  btf_16_adds_subs_avx2(&x[3], &x[4]);
+  btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x[10], &x[13], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x[11], &x[12], _r, cos_bit);
+}
+
+static INLINE void idct16_stage7_avx2(__m256i *output, __m256i *x1) {
+  btf_16_adds_subs_out_avx2(&output[0], &output[15], x1[0], x1[15]);
+  btf_16_adds_subs_out_avx2(&output[1], &output[14], x1[1], x1[14]);
+  btf_16_adds_subs_out_avx2(&output[2], &output[13], x1[2], x1[13]);
+  btf_16_adds_subs_out_avx2(&output[3], &output[12], x1[3], x1[12]);
+  btf_16_adds_subs_out_avx2(&output[4], &output[11], x1[4], x1[11]);
+  btf_16_adds_subs_out_avx2(&output[5], &output[10], x1[5], x1[10]);
+  btf_16_adds_subs_out_avx2(&output[6], &output[9], x1[6], x1[9]);
+  btf_16_adds_subs_out_avx2(&output[7], &output[8], x1[7], x1[8]);
+}
+
+static void idct16_new_avx2(const __m256i *input, __m256i *output,
+                            int8_t cos_bit) {
+  (void)(cos_bit);
+  const int32_t *cospi = cospi_arr(INV_COS_BIT);
+  const __m256i _r = _mm256_set1_epi32(1 << (INV_COS_BIT - 1));
+
+  __m256i cospi_p60_m04 = pair_set_w16_epi16(cospi[60], -cospi[4]);
+  __m256i cospi_p04_p60 = pair_set_w16_epi16(cospi[4], cospi[60]);
+  __m256i cospi_p28_m36 = pair_set_w16_epi16(cospi[28], -cospi[36]);
+  __m256i cospi_p36_p28 = pair_set_w16_epi16(cospi[36], cospi[28]);
+  __m256i cospi_p44_m20 = pair_set_w16_epi16(cospi[44], -cospi[20]);
+  __m256i cospi_p20_p44 = pair_set_w16_epi16(cospi[20], cospi[44]);
+  __m256i cospi_p12_m52 = pair_set_w16_epi16(cospi[12], -cospi[52]);
+  __m256i cospi_p52_p12 = pair_set_w16_epi16(cospi[52], cospi[12]);
+  __m256i cospi_p56_m08 = pair_set_w16_epi16(cospi[56], -cospi[8]);
+  __m256i cospi_p08_p56 = pair_set_w16_epi16(cospi[8], cospi[56]);
+  __m256i cospi_p24_m40 = pair_set_w16_epi16(cospi[24], -cospi[40]);
+  __m256i cospi_p40_p24 = pair_set_w16_epi16(cospi[40], cospi[24]);
+  __m256i cospi_p32_p32 = pair_set_w16_epi16(cospi[32], cospi[32]);
+  __m256i cospi_p32_m32 = pair_set_w16_epi16(cospi[32], -cospi[32]);
+  __m256i cospi_p48_m16 = pair_set_w16_epi16(cospi[48], -cospi[16]);
+  __m256i cospi_p16_p48 = pair_set_w16_epi16(cospi[16], cospi[48]);
+  __m256i cospi_m16_p48 = pair_set_w16_epi16(-cospi[16], cospi[48]);
+  __m256i cospi_p48_p16 = pair_set_w16_epi16(cospi[48], cospi[16]);
+  __m256i cospi_m48_m16 = pair_set_w16_epi16(-cospi[48], -cospi[16]);
+
+  // stage 1
+  __m256i x1[16];
+  x1[0] = input[0];
+  x1[1] = input[8];
+  x1[2] = input[4];
+  x1[3] = input[12];
+  x1[4] = input[2];
+  x1[5] = input[10];
+  x1[6] = input[6];
+  x1[7] = input[14];
+  x1[8] = input[1];
+  x1[9] = input[9];
+  x1[10] = input[5];
+  x1[11] = input[13];
+  x1[12] = input[3];
+  x1[13] = input[11];
+  x1[14] = input[7];
+  x1[15] = input[15];
+
+  // stage 2
+  btf_16_w16_avx2(cospi_p60_m04, cospi_p04_p60, &x1[8], &x1[15], _r, cos_bit);
+  btf_16_w16_avx2(cospi_p28_m36, cospi_p36_p28, &x1[9], &x1[14], _r, cos_bit);
+  btf_16_w16_avx2(cospi_p44_m20, cospi_p20_p44, &x1[10], &x1[13], _r, cos_bit);
+  btf_16_w16_avx2(cospi_p12_m52, cospi_p52_p12, &x1[11], &x1[12], _r, cos_bit);
+
+  // stage 3
+  btf_16_w16_avx2(cospi_p56_m08, cospi_p08_p56, &x1[4], &x1[7], _r, cos_bit);
+  btf_16_w16_avx2(cospi_p24_m40, cospi_p40_p24, &x1[5], &x1[6], _r, cos_bit);
+  btf_16_adds_subs_avx2(&x1[8], &x1[9]);
+  btf_16_adds_subs_avx2(&x1[11], &x1[10]);
+  btf_16_adds_subs_avx2(&x1[12], &x1[13]);
+  btf_16_adds_subs_avx2(&x1[15], &x1[14]);
+
+  // stage 4
+  btf_16_w16_avx2(cospi_p32_p32, cospi_p32_m32, &x1[0], &x1[1], _r, cos_bit);
+  btf_16_w16_avx2(cospi_p48_m16, cospi_p16_p48, &x1[2], &x1[3], _r, cos_bit);
+  btf_16_adds_subs_avx2(&x1[4], &x1[5]);
+  btf_16_adds_subs_avx2(&x1[7], &x1[6]);
+  btf_16_w16_avx2(cospi_m16_p48, cospi_p48_p16, &x1[9], &x1[14], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m48_m16, cospi_m16_p48, &x1[10], &x1[13], _r, cos_bit);
+
+  idct16_stage5_avx2(x1, cospi, _r, cos_bit);
+  idct16_stage6_avx2(x1, cospi, _r, cos_bit);
+  idct16_stage7_avx2(output, x1);
+}
+
+static void idct16_low8_new_avx2(const __m256i *input, __m256i *output,
+                                 int8_t cos_bit) {
+  (void)(cos_bit);
+  const int32_t *cospi = cospi_arr(INV_COS_BIT);
+  const __m256i _r = _mm256_set1_epi32(1 << (INV_COS_BIT - 1));
+
+  const __m256i cospi_m16_p48 = pair_set_w16_epi16(-cospi[16], cospi[48]);
+  const __m256i cospi_p48_p16 = pair_set_w16_epi16(cospi[48], cospi[16]);
+  const __m256i cospi_m48_m16 = pair_set_w16_epi16(-cospi[48], -cospi[16]);
+
+  // stage 1
+  __m256i x1[16];
+  x1[0] = input[0];
+  x1[2] = input[4];
+  x1[4] = input[2];
+  x1[6] = input[6];
+  x1[8] = input[1];
+  x1[10] = input[5];
+  x1[12] = input[3];
+  x1[14] = input[7];
+
+  // stage 2
+  btf_16_w16_0_avx2(cospi[60], cospi[4], x1[8], x1[8], x1[15]);
+  btf_16_w16_0_avx2(-cospi[36], cospi[28], x1[14], x1[9], x1[14]);
+  btf_16_w16_0_avx2(cospi[44], cospi[20], x1[10], x1[10], x1[13]);
+  btf_16_w16_0_avx2(-cospi[52], cospi[12], x1[12], x1[11], x1[12]);
+
+  // stage 3
+  btf_16_w16_0_avx2(cospi[56], cospi[8], x1[4], x1[4], x1[7]);
+  btf_16_w16_0_avx2(-cospi[40], cospi[24], x1[6], x1[5], x1[6]);
+  btf_16_adds_subs_avx2(&x1[8], &x1[9]);
+  btf_16_adds_subs_avx2(&x1[11], &x1[10]);
+  btf_16_adds_subs_avx2(&x1[12], &x1[13]);
+  btf_16_adds_subs_avx2(&x1[15], &x1[14]);
+
+  // stage 4
+  btf_16_w16_0_avx2(cospi[32], cospi[32], x1[0], x1[0], x1[1]);
+  btf_16_w16_0_avx2(cospi[48], cospi[16], x1[2], x1[2], x1[3]);
+  btf_16_adds_subs_avx2(&x1[4], &x1[5]);
+  btf_16_adds_subs_avx2(&x1[7], &x1[6]);
+  btf_16_w16_avx2(cospi_m16_p48, cospi_p48_p16, &x1[9], &x1[14], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m48_m16, cospi_m16_p48, &x1[10], &x1[13], _r, cos_bit);
+
+  idct16_stage5_avx2(x1, cospi, _r, cos_bit);
+  idct16_stage6_avx2(x1, cospi, _r, cos_bit);
+  idct16_stage7_avx2(output, x1);
+}
+
+static void idct16_low1_new_avx2(const __m256i *input, __m256i *output,
+                                 int8_t cos_bit) {
+  (void)(cos_bit);
+  const int32_t *cospi = cospi_arr(INV_COS_BIT);
+
+  // stage 1
+  __m256i x1[2];
+  x1[0] = input[0];
+
+  // stage 2
+  // stage 3
+  // stage 4
+  btf_16_w16_0_avx2(cospi[32], cospi[32], x1[0], x1[0], x1[1]);
+
+  // stage 5
+  // stage 6
+  output[0] = x1[0];
+  output[1] = x1[1];
+  output[2] = x1[1];
+  output[3] = x1[0];
+  output[4] = x1[0];
+  output[5] = x1[1];
+  output[6] = x1[1];
+  output[7] = x1[0];
+  output[8] = x1[0];
+  output[9] = x1[1];
+  output[10] = x1[1];
+  output[11] = x1[0];
+  output[12] = x1[0];
+  output[13] = x1[1];
+  output[14] = x1[1];
+  output[15] = x1[0];
+}
+
+static INLINE void iadst16_stage3_avx2(__m256i *x) {
+  btf_16_adds_subs_avx2(&x[0], &x[8]);
+  btf_16_adds_subs_avx2(&x[1], &x[9]);
+  btf_16_adds_subs_avx2(&x[2], &x[10]);
+  btf_16_adds_subs_avx2(&x[3], &x[11]);
+  btf_16_adds_subs_avx2(&x[4], &x[12]);
+  btf_16_adds_subs_avx2(&x[5], &x[13]);
+  btf_16_adds_subs_avx2(&x[6], &x[14]);
+  btf_16_adds_subs_avx2(&x[7], &x[15]);
+}
+
+static INLINE void iadst16_stage4_avx2(__m256i *x, const int32_t *cospi,
+                                       const __m256i _r, int8_t cos_bit) {
+  const __m256i cospi_p08_p56 = pair_set_w16_epi16(cospi[8], cospi[56]);
+  const __m256i cospi_p56_m08 = pair_set_w16_epi16(cospi[56], -cospi[8]);
+  const __m256i cospi_p40_p24 = pair_set_w16_epi16(cospi[40], cospi[24]);
+  const __m256i cospi_p24_m40 = pair_set_w16_epi16(cospi[24], -cospi[40]);
+  const __m256i cospi_m56_p08 = pair_set_w16_epi16(-cospi[56], cospi[8]);
+  const __m256i cospi_m24_p40 = pair_set_w16_epi16(-cospi[24], cospi[40]);
+  btf_16_w16_avx2(cospi_p08_p56, cospi_p56_m08, &x[8], &x[9], _r, cos_bit);
+  btf_16_w16_avx2(cospi_p40_p24, cospi_p24_m40, &x[10], &x[11], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m56_p08, cospi_p08_p56, &x[12], &x[13], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m24_p40, cospi_p40_p24, &x[14], &x[15], _r, cos_bit);
+}
+
+static INLINE void iadst16_stage5_avx2(__m256i *x) {
+  btf_16_adds_subs_avx2(&x[0], &x[4]);
+  btf_16_adds_subs_avx2(&x[1], &x[5]);
+  btf_16_adds_subs_avx2(&x[2], &x[6]);
+  btf_16_adds_subs_avx2(&x[3], &x[7]);
+  btf_16_adds_subs_avx2(&x[8], &x[12]);
+  btf_16_adds_subs_avx2(&x[9], &x[13]);
+  btf_16_adds_subs_avx2(&x[10], &x[14]);
+  btf_16_adds_subs_avx2(&x[11], &x[15]);
+}
+
+static INLINE void iadst16_stage6_avx2(__m256i *x, const int32_t *cospi,
+                                       const __m256i _r, int8_t cos_bit) {
+  const __m256i cospi_p16_p48 = pair_set_w16_epi16(cospi[16], cospi[48]);
+  const __m256i cospi_p48_m16 = pair_set_w16_epi16(cospi[48], -cospi[16]);
+  const __m256i cospi_m48_p16 = pair_set_w16_epi16(-cospi[48], cospi[16]);
+  btf_16_w16_avx2(cospi_p16_p48, cospi_p48_m16, &x[4], &x[5], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m48_p16, cospi_p16_p48, &x[6], &x[7], _r, cos_bit);
+  btf_16_w16_avx2(cospi_p16_p48, cospi_p48_m16, &x[12], &x[13], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m48_p16, cospi_p16_p48, &x[14], &x[15], _r, cos_bit);
+}
+
+static INLINE void iadst16_stage7_avx2(__m256i *x) {
+  btf_16_adds_subs_avx2(&x[0], &x[2]);
+  btf_16_adds_subs_avx2(&x[1], &x[3]);
+  btf_16_adds_subs_avx2(&x[4], &x[6]);
+  btf_16_adds_subs_avx2(&x[5], &x[7]);
+  btf_16_adds_subs_avx2(&x[8], &x[10]);
+  btf_16_adds_subs_avx2(&x[9], &x[11]);
+  btf_16_adds_subs_avx2(&x[12], &x[14]);
+  btf_16_adds_subs_avx2(&x[13], &x[15]);
+}
+
+static INLINE void iadst16_stage8_avx2(__m256i *x1, const int32_t *cospi,
+                                       const __m256i _r, int8_t cos_bit) {
+  const __m256i cospi_p32_p32 = pair_set_w16_epi16(cospi[32], cospi[32]);
+  const __m256i cospi_p32_m32 = pair_set_w16_epi16(cospi[32], -cospi[32]);
+  btf_16_w16_avx2(cospi_p32_p32, cospi_p32_m32, &x1[2], &x1[3], _r, cos_bit);
+  btf_16_w16_avx2(cospi_p32_p32, cospi_p32_m32, &x1[6], &x1[7], _r, cos_bit);
+  btf_16_w16_avx2(cospi_p32_p32, cospi_p32_m32, &x1[10], &x1[11], _r, cos_bit);
+  btf_16_w16_avx2(cospi_p32_p32, cospi_p32_m32, &x1[14], &x1[15], _r, cos_bit);
+}
+
+static INLINE void iadst16_stage9_avx2(__m256i *output, __m256i *x1) {
+  const __m256i __zero = _mm256_setzero_si256();
+  output[0] = x1[0];
+  output[1] = _mm256_subs_epi16(__zero, x1[8]);
+  output[2] = x1[12];
+  output[3] = _mm256_subs_epi16(__zero, x1[4]);
+  output[4] = x1[6];
+  output[5] = _mm256_subs_epi16(__zero, x1[14]);
+  output[6] = x1[10];
+  output[7] = _mm256_subs_epi16(__zero, x1[2]);
+  output[8] = x1[3];
+  output[9] = _mm256_subs_epi16(__zero, x1[11]);
+  output[10] = x1[15];
+  output[11] = _mm256_subs_epi16(__zero, x1[7]);
+  output[12] = x1[5];
+  output[13] = _mm256_subs_epi16(__zero, x1[13]);
+  output[14] = x1[9];
+  output[15] = _mm256_subs_epi16(__zero, x1[1]);
+}
+
+static void iadst16_new_avx2(const __m256i *input, __m256i *output,
+                             int8_t cos_bit) {
+  (void)(cos_bit);
+  const int32_t *cospi = cospi_arr(INV_COS_BIT);
+
+  const __m256i _r = _mm256_set1_epi32(1 << (INV_COS_BIT - 1));
+
+  __m256i cospi_p02_p62 = pair_set_w16_epi16(cospi[2], cospi[62]);
+  __m256i cospi_p62_m02 = pair_set_w16_epi16(cospi[62], -cospi[2]);
+  __m256i cospi_p10_p54 = pair_set_w16_epi16(cospi[10], cospi[54]);
+  __m256i cospi_p54_m10 = pair_set_w16_epi16(cospi[54], -cospi[10]);
+  __m256i cospi_p18_p46 = pair_set_w16_epi16(cospi[18], cospi[46]);
+  __m256i cospi_p46_m18 = pair_set_w16_epi16(cospi[46], -cospi[18]);
+  __m256i cospi_p26_p38 = pair_set_w16_epi16(cospi[26], cospi[38]);
+  __m256i cospi_p38_m26 = pair_set_w16_epi16(cospi[38], -cospi[26]);
+  __m256i cospi_p34_p30 = pair_set_w16_epi16(cospi[34], cospi[30]);
+  __m256i cospi_p30_m34 = pair_set_w16_epi16(cospi[30], -cospi[34]);
+  __m256i cospi_p42_p22 = pair_set_w16_epi16(cospi[42], cospi[22]);
+  __m256i cospi_p22_m42 = pair_set_w16_epi16(cospi[22], -cospi[42]);
+  __m256i cospi_p50_p14 = pair_set_w16_epi16(cospi[50], cospi[14]);
+  __m256i cospi_p14_m50 = pair_set_w16_epi16(cospi[14], -cospi[50]);
+  __m256i cospi_p58_p06 = pair_set_w16_epi16(cospi[58], cospi[6]);
+  __m256i cospi_p06_m58 = pair_set_w16_epi16(cospi[6], -cospi[58]);
+
+  // stage 1
+  __m256i x1[16];
+  x1[0] = input[15];
+  x1[1] = input[0];
+  x1[2] = input[13];
+  x1[3] = input[2];
+  x1[4] = input[11];
+  x1[5] = input[4];
+  x1[6] = input[9];
+  x1[7] = input[6];
+  x1[8] = input[7];
+  x1[9] = input[8];
+  x1[10] = input[5];
+  x1[11] = input[10];
+  x1[12] = input[3];
+  x1[13] = input[12];
+  x1[14] = input[1];
+  x1[15] = input[14];
+
+  // stage 2
+  btf_16_w16_avx2(cospi_p02_p62, cospi_p62_m02, &x1[0], &x1[1], _r, cos_bit);
+  btf_16_w16_avx2(cospi_p10_p54, cospi_p54_m10, &x1[2], &x1[3], _r, cos_bit);
+  btf_16_w16_avx2(cospi_p18_p46, cospi_p46_m18, &x1[4], &x1[5], _r, cos_bit);
+  btf_16_w16_avx2(cospi_p26_p38, cospi_p38_m26, &x1[6], &x1[7], _r, cos_bit);
+  btf_16_w16_avx2(cospi_p34_p30, cospi_p30_m34, &x1[8], &x1[9], _r, cos_bit);
+  btf_16_w16_avx2(cospi_p42_p22, cospi_p22_m42, &x1[10], &x1[11], _r, cos_bit);
+  btf_16_w16_avx2(cospi_p50_p14, cospi_p14_m50, &x1[12], &x1[13], _r, cos_bit);
+  btf_16_w16_avx2(cospi_p58_p06, cospi_p06_m58, &x1[14], &x1[15], _r, cos_bit);
+
+  iadst16_stage3_avx2(x1);
+  iadst16_stage4_avx2(x1, cospi, _r, cos_bit);
+  iadst16_stage5_avx2(x1);
+  iadst16_stage6_avx2(x1, cospi, _r, cos_bit);
+  iadst16_stage7_avx2(x1);
+  iadst16_stage8_avx2(x1, cospi, _r, cos_bit);
+  iadst16_stage9_avx2(output, x1);
+}
+
+static void iadst16_low8_new_avx2(const __m256i *input, __m256i *output,
+                                  int8_t cos_bit) {
+  (void)(cos_bit);
+  const int32_t *cospi = cospi_arr(INV_COS_BIT);
+  const __m256i _r = _mm256_set1_epi32(1 << (INV_COS_BIT - 1));
+
+  // stage 1
+  __m256i x1[16];
+  x1[1] = input[0];
+  x1[3] = input[2];
+  x1[5] = input[4];
+  x1[7] = input[6];
+  x1[8] = input[7];
+  x1[10] = input[5];
+  x1[12] = input[3];
+  x1[14] = input[1];
+
+  // stage 2
+  btf_16_w16_0_avx2(cospi[62], -cospi[2], x1[1], x1[0], x1[1]);
+  btf_16_w16_0_avx2(cospi[54], -cospi[10], x1[3], x1[2], x1[3]);
+  btf_16_w16_0_avx2(cospi[46], -cospi[18], x1[5], x1[4], x1[5]);
+  btf_16_w16_0_avx2(cospi[38], -cospi[26], x1[7], x1[6], x1[7]);
+  btf_16_w16_0_avx2(cospi[34], cospi[30], x1[8], x1[8], x1[9]);
+  btf_16_w16_0_avx2(cospi[42], cospi[22], x1[10], x1[10], x1[11]);
+  btf_16_w16_0_avx2(cospi[50], cospi[14], x1[12], x1[12], x1[13]);
+  btf_16_w16_0_avx2(cospi[58], cospi[06], x1[14], x1[14], x1[15]);
+
+  iadst16_stage3_avx2(x1);
+  iadst16_stage4_avx2(x1, cospi, _r, cos_bit);
+  iadst16_stage5_avx2(x1);
+  iadst16_stage6_avx2(x1, cospi, _r, cos_bit);
+  iadst16_stage7_avx2(x1);
+  iadst16_stage8_avx2(x1, cospi, _r, cos_bit);
+  iadst16_stage9_avx2(output, x1);
+}
+
+static void iadst16_low1_new_avx2(const __m256i *input, __m256i *output,
+                                  int8_t cos_bit) {
+  (void)(cos_bit);
+  const int32_t *cospi = cospi_arr(INV_COS_BIT);
+  const __m256i _r = _mm256_set1_epi32(1 << (INV_COS_BIT - 1));
+
+  const __m256i cospi_p08_p56 = pair_set_w16_epi16(cospi[8], cospi[56]);
+  const __m256i cospi_p56_m08 = pair_set_w16_epi16(cospi[56], -cospi[8]);
+  const __m256i cospi_p16_p48 = pair_set_w16_epi16(cospi[16], cospi[48]);
+  const __m256i cospi_p48_m16 = pair_set_w16_epi16(cospi[48], -cospi[16]);
+
+  // stage 1
+  __m256i x1[16];
+  x1[1] = input[0];
+
+  // stage 2
+  btf_16_w16_0_avx2(cospi[62], -cospi[2], x1[1], x1[0], x1[1]);
+
+  // stage 3
+  x1[8] = x1[0];
+  x1[9] = x1[1];
+
+  // stage 4
+  btf_16_w16_avx2(cospi_p08_p56, cospi_p56_m08, &x1[8], &x1[9], _r, cos_bit);
+
+  // stage 5
+  x1[4] = x1[0];
+  x1[5] = x1[1];
+
+  x1[12] = x1[8];
+  x1[13] = x1[9];
+
+  // stage 6
+  btf_16_w16_avx2(cospi_p16_p48, cospi_p48_m16, &x1[4], &x1[5], _r, cos_bit);
+  btf_16_w16_avx2(cospi_p16_p48, cospi_p48_m16, &x1[12], &x1[13], _r, cos_bit);
+
+  // stage 7
+  x1[2] = x1[0];
+  x1[3] = x1[1];
+  x1[6] = x1[4];
+  x1[7] = x1[5];
+  x1[10] = x1[8];
+  x1[11] = x1[9];
+  x1[14] = x1[12];
+  x1[15] = x1[13];
+
+  iadst16_stage8_avx2(x1, cospi, _r, cos_bit);
+  iadst16_stage9_avx2(output, x1);
+}
+
+static INLINE void idct32_high16_stage3_avx2(__m256i *x) {
+  btf_16_adds_subs_avx2(&x[16], &x[17]);
+  btf_16_adds_subs_avx2(&x[19], &x[18]);
+  btf_16_adds_subs_avx2(&x[20], &x[21]);
+  btf_16_adds_subs_avx2(&x[23], &x[22]);
+  btf_16_adds_subs_avx2(&x[24], &x[25]);
+  btf_16_adds_subs_avx2(&x[27], &x[26]);
+  btf_16_adds_subs_avx2(&x[28], &x[29]);
+  btf_16_adds_subs_avx2(&x[31], &x[30]);
+}
+
+static INLINE void idct32_high16_stage4_avx2(__m256i *x, const int32_t *cospi,
+                                             const __m256i _r, int8_t cos_bit) {
+  const __m256i cospi_m08_p56 = pair_set_w16_epi16(-cospi[8], cospi[56]);
+  const __m256i cospi_p56_p08 = pair_set_w16_epi16(cospi[56], cospi[8]);
+  const __m256i cospi_m56_m08 = pair_set_w16_epi16(-cospi[56], -cospi[8]);
+  const __m256i cospi_m40_p24 = pair_set_w16_epi16(-cospi[40], cospi[24]);
+  const __m256i cospi_p24_p40 = pair_set_w16_epi16(cospi[24], cospi[40]);
+  const __m256i cospi_m24_m40 = pair_set_w16_epi16(-cospi[24], -cospi[40]);
+  btf_16_w16_avx2(cospi_m08_p56, cospi_p56_p08, &x[17], &x[30], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m56_m08, cospi_m08_p56, &x[18], &x[29], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m40_p24, cospi_p24_p40, &x[21], &x[26], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m24_m40, cospi_m40_p24, &x[22], &x[25], _r, cos_bit);
+}
+
+static INLINE void idct32_high24_stage5_avx2(__m256i *x, const int32_t *cospi,
+                                             const __m256i _r, int8_t cos_bit) {
+  const __m256i cospi_m16_p48 = pair_set_w16_epi16(-cospi[16], cospi[48]);
+  const __m256i cospi_p48_p16 = pair_set_w16_epi16(cospi[48], cospi[16]);
+  const __m256i cospi_m48_m16 = pair_set_w16_epi16(-cospi[48], -cospi[16]);
+  btf_16_w16_avx2(cospi_m16_p48, cospi_p48_p16, &x[9], &x[14], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m48_m16, cospi_m16_p48, &x[10], &x[13], _r, cos_bit);
+  btf_16_adds_subs_avx2(&x[16], &x[19]);
+  btf_16_adds_subs_avx2(&x[17], &x[18]);
+  btf_16_adds_subs_avx2(&x[23], &x[20]);
+  btf_16_adds_subs_avx2(&x[22], &x[21]);
+  btf_16_adds_subs_avx2(&x[24], &x[27]);
+  btf_16_adds_subs_avx2(&x[25], &x[26]);
+  btf_16_adds_subs_avx2(&x[31], &x[28]);
+  btf_16_adds_subs_avx2(&x[30], &x[29]);
+}
+
+static INLINE void idct32_high28_stage6_avx2(__m256i *x, const int32_t *cospi,
+                                             const __m256i _r, int8_t cos_bit) {
+  const __m256i cospi_m32_p32 = pair_set_w16_epi16(-cospi[32], cospi[32]);
+  const __m256i cospi_p32_p32 = pair_set_w16_epi16(cospi[32], cospi[32]);
+  const __m256i cospi_m16_p48 = pair_set_w16_epi16(-cospi[16], cospi[48]);
+  const __m256i cospi_p48_p16 = pair_set_w16_epi16(cospi[48], cospi[16]);
+  const __m256i cospi_m48_m16 = pair_set_w16_epi16(-cospi[48], -cospi[16]);
+  btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x[5], &x[6], _r, cos_bit);
+  btf_16_adds_subs_avx2(&x[8], &x[11]);
+  btf_16_adds_subs_avx2(&x[9], &x[10]);
+  btf_16_adds_subs_avx2(&x[15], &x[12]);
+  btf_16_adds_subs_avx2(&x[14], &x[13]);
+  btf_16_w16_avx2(cospi_m16_p48, cospi_p48_p16, &x[18], &x[29], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m16_p48, cospi_p48_p16, &x[19], &x[28], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m48_m16, cospi_m16_p48, &x[20], &x[27], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m48_m16, cospi_m16_p48, &x[21], &x[26], _r, cos_bit);
+}
+
+static INLINE void idct32_stage7_avx2(__m256i *x, const int32_t *cospi,
+                                      const __m256i _r, int8_t cos_bit) {
+  const __m256i cospi_m32_p32 = pair_set_w16_epi16(-cospi[32], cospi[32]);
+  const __m256i cospi_p32_p32 = pair_set_w16_epi16(cospi[32], cospi[32]);
+  btf_16_adds_subs_avx2(&x[0], &x[7]);
+  btf_16_adds_subs_avx2(&x[1], &x[6]);
+  btf_16_adds_subs_avx2(&x[2], &x[5]);
+  btf_16_adds_subs_avx2(&x[3], &x[4]);
+  btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x[10], &x[13], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x[11], &x[12], _r, cos_bit);
+  btf_16_adds_subs_avx2(&x[16], &x[23]);
+  btf_16_adds_subs_avx2(&x[17], &x[22]);
+  btf_16_adds_subs_avx2(&x[18], &x[21]);
+  btf_16_adds_subs_avx2(&x[19], &x[20]);
+  btf_16_adds_subs_avx2(&x[31], &x[24]);
+  btf_16_adds_subs_avx2(&x[30], &x[25]);
+  btf_16_adds_subs_avx2(&x[29], &x[26]);
+  btf_16_adds_subs_avx2(&x[28], &x[27]);
+}
+
+static INLINE void idct32_stage8_avx2(__m256i *x, const int32_t *cospi,
+                                      const __m256i _r, int8_t cos_bit) {
+  const __m256i cospi_m32_p32 = pair_set_w16_epi16(-cospi[32], cospi[32]);
+  const __m256i cospi_p32_p32 = pair_set_w16_epi16(cospi[32], cospi[32]);
+  btf_16_adds_subs_avx2(&x[0], &x[15]);
+  btf_16_adds_subs_avx2(&x[1], &x[14]);
+  btf_16_adds_subs_avx2(&x[2], &x[13]);
+  btf_16_adds_subs_avx2(&x[3], &x[12]);
+  btf_16_adds_subs_avx2(&x[4], &x[11]);
+  btf_16_adds_subs_avx2(&x[5], &x[10]);
+  btf_16_adds_subs_avx2(&x[6], &x[9]);
+  btf_16_adds_subs_avx2(&x[7], &x[8]);
+  btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x[20], &x[27], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x[21], &x[26], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x[22], &x[25], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x[23], &x[24], _r, cos_bit);
+}
+
+static INLINE void idct32_stage9_avx2(__m256i *output, __m256i *x) {
+  btf_16_adds_subs_out_avx2(&output[0], &output[31], x[0], x[31]);
+  btf_16_adds_subs_out_avx2(&output[1], &output[30], x[1], x[30]);
+  btf_16_adds_subs_out_avx2(&output[2], &output[29], x[2], x[29]);
+  btf_16_adds_subs_out_avx2(&output[3], &output[28], x[3], x[28]);
+  btf_16_adds_subs_out_avx2(&output[4], &output[27], x[4], x[27]);
+  btf_16_adds_subs_out_avx2(&output[5], &output[26], x[5], x[26]);
+  btf_16_adds_subs_out_avx2(&output[6], &output[25], x[6], x[25]);
+  btf_16_adds_subs_out_avx2(&output[7], &output[24], x[7], x[24]);
+  btf_16_adds_subs_out_avx2(&output[8], &output[23], x[8], x[23]);
+  btf_16_adds_subs_out_avx2(&output[9], &output[22], x[9], x[22]);
+  btf_16_adds_subs_out_avx2(&output[10], &output[21], x[10], x[21]);
+  btf_16_adds_subs_out_avx2(&output[11], &output[20], x[11], x[20]);
+  btf_16_adds_subs_out_avx2(&output[12], &output[19], x[12], x[19]);
+  btf_16_adds_subs_out_avx2(&output[13], &output[18], x[13], x[18]);
+  btf_16_adds_subs_out_avx2(&output[14], &output[17], x[14], x[17]);
+  btf_16_adds_subs_out_avx2(&output[15], &output[16], x[15], x[16]);
+}
+
+static void idct32_low1_new_avx2(const __m256i *input, __m256i *output,
+                                 int8_t cos_bit) {
+  (void)cos_bit;
+  const int32_t *cospi = cospi_arr(INV_COS_BIT);
+
+  // stage 1
+  __m256i x[2];
+  x[0] = input[0];
+
+  // stage 2
+  // stage 3
+  // stage 4
+  // stage 5
+  btf_16_w16_0_avx2(cospi[32], cospi[32], x[0], x[0], x[1]);
+
+  // stage 6
+  // stage 7
+  // stage 8
+  // stage 9
+  output[0] = x[0];
+  output[31] = x[0];
+  output[1] = x[1];
+  output[30] = x[1];
+  output[2] = x[1];
+  output[29] = x[1];
+  output[3] = x[0];
+  output[28] = x[0];
+  output[4] = x[0];
+  output[27] = x[0];
+  output[5] = x[1];
+  output[26] = x[1];
+  output[6] = x[1];
+  output[25] = x[1];
+  output[7] = x[0];
+  output[24] = x[0];
+  output[8] = x[0];
+  output[23] = x[0];
+  output[9] = x[1];
+  output[22] = x[1];
+  output[10] = x[1];
+  output[21] = x[1];
+  output[11] = x[0];
+  output[20] = x[0];
+  output[12] = x[0];
+  output[19] = x[0];
+  output[13] = x[1];
+  output[18] = x[1];
+  output[14] = x[1];
+  output[17] = x[1];
+  output[15] = x[0];
+  output[16] = x[0];
+}
+
+static void idct32_low8_new_avx2(const __m256i *input, __m256i *output,
+                                 int8_t cos_bit) {
+  (void)cos_bit;
+  const int32_t *cospi = cospi_arr(INV_COS_BIT);
+  const __m256i _r = _mm256_set1_epi32(1 << (INV_COS_BIT - 1));
+
+  // stage 1
+  __m256i x[32];
+  x[0] = input[0];
+  x[4] = input[4];
+  x[8] = input[2];
+  x[12] = input[6];
+  x[16] = input[1];
+  x[20] = input[5];
+  x[24] = input[3];
+  x[28] = input[7];
+
+  // stage 2
+  btf_16_w16_0_avx2(cospi[62], cospi[2], x[16], x[16], x[31]);
+  btf_16_w16_0_avx2(-cospi[50], cospi[14], x[28], x[19], x[28]);
+  btf_16_w16_0_avx2(cospi[54], cospi[10], x[20], x[20], x[27]);
+  btf_16_w16_0_avx2(-cospi[58], cospi[6], x[24], x[23], x[24]);
+
+  // stage 3
+  btf_16_w16_0_avx2(cospi[60], cospi[4], x[8], x[8], x[15]);
+  btf_16_w16_0_avx2(-cospi[52], cospi[12], x[12], x[11], x[12]);
+  x[17] = x[16];
+  x[18] = x[19];
+  x[21] = x[20];
+  x[22] = x[23];
+  x[25] = x[24];
+  x[26] = x[27];
+  x[29] = x[28];
+  x[30] = x[31];
+
+  // stage 4
+  btf_16_w16_0_avx2(cospi[56], cospi[8], x[4], x[4], x[7]);
+  x[9] = x[8];
+  x[10] = x[11];
+  x[13] = x[12];
+  x[14] = x[15];
+  idct32_high16_stage4_avx2(x, cospi, _r, cos_bit);
+
+  // stage 5
+  btf_16_w16_0_avx2(cospi[32], cospi[32], x[0], x[0], x[1]);
+  x[5] = x[4];
+  x[6] = x[7];
+  idct32_high24_stage5_avx2(x, cospi, _r, cos_bit);
+  // stage 6
+  x[3] = x[0];
+  x[2] = x[1];
+  idct32_high28_stage6_avx2(x, cospi, _r, cos_bit);
+
+  idct32_stage7_avx2(x, cospi, _r, cos_bit);
+  idct32_stage8_avx2(x, cospi, _r, cos_bit);
+  idct32_stage9_avx2(output, x);
+}
+
+static void idct32_low16_new_avx2(const __m256i *input, __m256i *output,
+                                  int8_t cos_bit) {
+  (void)cos_bit;
+  const int32_t *cospi = cospi_arr(INV_COS_BIT);
+  const __m256i _r = _mm256_set1_epi32(1 << (INV_COS_BIT - 1));
+
+  // stage 1
+  __m256i x[32];
+  x[0] = input[0];
+  x[2] = input[8];
+  x[4] = input[4];
+  x[6] = input[12];
+  x[8] = input[2];
+  x[10] = input[10];
+  x[12] = input[6];
+  x[14] = input[14];
+  x[16] = input[1];
+  x[18] = input[9];
+  x[20] = input[5];
+  x[22] = input[13];
+  x[24] = input[3];
+  x[26] = input[11];
+  x[28] = input[7];
+  x[30] = input[15];
+
+  // stage 2
+  btf_16_w16_0_avx2(cospi[62], cospi[2], x[16], x[16], x[31]);
+  btf_16_w16_0_avx2(-cospi[34], cospi[30], x[30], x[17], x[30]);
+  btf_16_w16_0_avx2(cospi[46], cospi[18], x[18], x[18], x[29]);
+  btf_16_w16_0_avx2(-cospi[50], cospi[14], x[28], x[19], x[28]);
+  btf_16_w16_0_avx2(cospi[54], cospi[10], x[20], x[20], x[27]);
+  btf_16_w16_0_avx2(-cospi[42], cospi[22], x[26], x[21], x[26]);
+  btf_16_w16_0_avx2(cospi[38], cospi[26], x[22], x[22], x[25]);
+  btf_16_w16_0_avx2(-cospi[58], cospi[6], x[24], x[23], x[24]);
+
+  // stage 3
+  btf_16_w16_0_avx2(cospi[60], cospi[4], x[8], x[8], x[15]);
+  btf_16_w16_0_avx2(-cospi[36], cospi[28], x[14], x[9], x[14]);
+  btf_16_w16_0_avx2(cospi[44], cospi[20], x[10], x[10], x[13]);
+  btf_16_w16_0_avx2(-cospi[52], cospi[12], x[12], x[11], x[12]);
+  idct32_high16_stage3_avx2(x);
+
+  // stage 4
+  btf_16_w16_0_avx2(cospi[56], cospi[8], x[4], x[4], x[7]);
+  btf_16_w16_0_avx2(-cospi[40], cospi[24], x[6], x[5], x[6]);
+  btf_16_adds_subs_avx2(&x[8], &x[9]);
+  btf_16_adds_subs_avx2(&x[11], &x[10]);
+  btf_16_adds_subs_avx2(&x[12], &x[13]);
+  btf_16_adds_subs_avx2(&x[15], &x[14]);
+  idct32_high16_stage4_avx2(x, cospi, _r, cos_bit);
+
+  // stage 5
+  btf_16_w16_0_avx2(cospi[32], cospi[32], x[0], x[0], x[1]);
+  btf_16_w16_0_avx2(cospi[48], cospi[16], x[2], x[2], x[3]);
+  btf_16_adds_subs_avx2(&x[4], &x[5]);
+  btf_16_adds_subs_avx2(&x[7], &x[6]);
+  idct32_high24_stage5_avx2(x, cospi, _r, cos_bit);
+
+  btf_16_adds_subs_avx2(&x[0], &x[3]);
+  btf_16_adds_subs_avx2(&x[1], &x[2]);
+  idct32_high28_stage6_avx2(x, cospi, _r, cos_bit);
+
+  idct32_stage7_avx2(x, cospi, _r, cos_bit);
+  idct32_stage8_avx2(x, cospi, _r, cos_bit);
+  idct32_stage9_avx2(output, x);
+}
+
+static void idct32_new_avx2(const __m256i *input, __m256i *output,
+                            int8_t cos_bit) {
+  (void)(cos_bit);
+  const int32_t *cospi = cospi_arr(INV_COS_BIT);
+  const __m256i _r = _mm256_set1_epi32(1 << (INV_COS_BIT - 1));
+
+  __m256i cospi_p62_m02 = pair_set_w16_epi16(cospi[62], -cospi[2]);
+  __m256i cospi_p02_p62 = pair_set_w16_epi16(cospi[2], cospi[62]);
+  __m256i cospi_p30_m34 = pair_set_w16_epi16(cospi[30], -cospi[34]);
+  __m256i cospi_p34_p30 = pair_set_w16_epi16(cospi[34], cospi[30]);
+  __m256i cospi_p46_m18 = pair_set_w16_epi16(cospi[46], -cospi[18]);
+  __m256i cospi_p18_p46 = pair_set_w16_epi16(cospi[18], cospi[46]);
+  __m256i cospi_p14_m50 = pair_set_w16_epi16(cospi[14], -cospi[50]);
+  __m256i cospi_p50_p14 = pair_set_w16_epi16(cospi[50], cospi[14]);
+  __m256i cospi_p54_m10 = pair_set_w16_epi16(cospi[54], -cospi[10]);
+  __m256i cospi_p10_p54 = pair_set_w16_epi16(cospi[10], cospi[54]);
+  __m256i cospi_p22_m42 = pair_set_w16_epi16(cospi[22], -cospi[42]);
+  __m256i cospi_p42_p22 = pair_set_w16_epi16(cospi[42], cospi[22]);
+  __m256i cospi_p38_m26 = pair_set_w16_epi16(cospi[38], -cospi[26]);
+  __m256i cospi_p26_p38 = pair_set_w16_epi16(cospi[26], cospi[38]);
+  __m256i cospi_p06_m58 = pair_set_w16_epi16(cospi[6], -cospi[58]);
+  __m256i cospi_p58_p06 = pair_set_w16_epi16(cospi[58], cospi[6]);
+  __m256i cospi_p60_m04 = pair_set_w16_epi16(cospi[60], -cospi[4]);
+  __m256i cospi_p04_p60 = pair_set_w16_epi16(cospi[4], cospi[60]);
+  __m256i cospi_p28_m36 = pair_set_w16_epi16(cospi[28], -cospi[36]);
+  __m256i cospi_p36_p28 = pair_set_w16_epi16(cospi[36], cospi[28]);
+  __m256i cospi_p44_m20 = pair_set_w16_epi16(cospi[44], -cospi[20]);
+  __m256i cospi_p20_p44 = pair_set_w16_epi16(cospi[20], cospi[44]);
+  __m256i cospi_p12_m52 = pair_set_w16_epi16(cospi[12], -cospi[52]);
+  __m256i cospi_p52_p12 = pair_set_w16_epi16(cospi[52], cospi[12]);
+  __m256i cospi_p56_m08 = pair_set_w16_epi16(cospi[56], -cospi[8]);
+  __m256i cospi_p08_p56 = pair_set_w16_epi16(cospi[8], cospi[56]);
+  __m256i cospi_p24_m40 = pair_set_w16_epi16(cospi[24], -cospi[40]);
+  __m256i cospi_p40_p24 = pair_set_w16_epi16(cospi[40], cospi[24]);
+  __m256i cospi_p32_p32 = pair_set_w16_epi16(cospi[32], cospi[32]);
+  __m256i cospi_p32_m32 = pair_set_w16_epi16(cospi[32], -cospi[32]);
+  __m256i cospi_p48_m16 = pair_set_w16_epi16(cospi[48], -cospi[16]);
+  __m256i cospi_p16_p48 = pair_set_w16_epi16(cospi[16], cospi[48]);
+
+  // stage 1
+  __m256i x1[32];
+  x1[0] = input[0];
+  x1[1] = input[16];
+  x1[2] = input[8];
+  x1[3] = input[24];
+  x1[4] = input[4];
+  x1[5] = input[20];
+  x1[6] = input[12];
+  x1[7] = input[28];
+  x1[8] = input[2];
+  x1[9] = input[18];
+  x1[10] = input[10];
+  x1[11] = input[26];
+  x1[12] = input[6];
+  x1[13] = input[22];
+  x1[14] = input[14];
+  x1[15] = input[30];
+  x1[16] = input[1];
+  x1[17] = input[17];
+  x1[18] = input[9];
+  x1[19] = input[25];
+  x1[20] = input[5];
+  x1[21] = input[21];
+  x1[22] = input[13];
+  x1[23] = input[29];
+  x1[24] = input[3];
+  x1[25] = input[19];
+  x1[26] = input[11];
+  x1[27] = input[27];
+  x1[28] = input[7];
+  x1[29] = input[23];
+  x1[30] = input[15];
+  x1[31] = input[31];
+
+  // stage 2
+  btf_16_w16_avx2(cospi_p62_m02, cospi_p02_p62, &x1[16], &x1[31], _r, cos_bit);
+  btf_16_w16_avx2(cospi_p30_m34, cospi_p34_p30, &x1[17], &x1[30], _r, cos_bit);
+  btf_16_w16_avx2(cospi_p46_m18, cospi_p18_p46, &x1[18], &x1[29], _r, cos_bit);
+  btf_16_w16_avx2(cospi_p14_m50, cospi_p50_p14, &x1[19], &x1[28], _r, cos_bit);
+  btf_16_w16_avx2(cospi_p54_m10, cospi_p10_p54, &x1[20], &x1[27], _r, cos_bit);
+  btf_16_w16_avx2(cospi_p22_m42, cospi_p42_p22, &x1[21], &x1[26], _r, cos_bit);
+  btf_16_w16_avx2(cospi_p38_m26, cospi_p26_p38, &x1[22], &x1[25], _r, cos_bit);
+  btf_16_w16_avx2(cospi_p06_m58, cospi_p58_p06, &x1[23], &x1[24], _r, cos_bit);
+
+  // stage 3
+  btf_16_w16_avx2(cospi_p60_m04, cospi_p04_p60, &x1[8], &x1[15], _r, cos_bit);
+  btf_16_w16_avx2(cospi_p28_m36, cospi_p36_p28, &x1[9], &x1[14], _r, cos_bit);
+  btf_16_w16_avx2(cospi_p44_m20, cospi_p20_p44, &x1[10], &x1[13], _r, cos_bit);
+  btf_16_w16_avx2(cospi_p12_m52, cospi_p52_p12, &x1[11], &x1[12], _r, cos_bit);
+  idct32_high16_stage3_avx2(x1);
+
+  // stage 4
+  btf_16_w16_avx2(cospi_p56_m08, cospi_p08_p56, &x1[4], &x1[7], _r, cos_bit);
+  btf_16_w16_avx2(cospi_p24_m40, cospi_p40_p24, &x1[5], &x1[6], _r, cos_bit);
+  btf_16_adds_subs_avx2(&x1[8], &x1[9]);
+  btf_16_adds_subs_avx2(&x1[11], &x1[10]);
+  btf_16_adds_subs_avx2(&x1[12], &x1[13]);
+  btf_16_adds_subs_avx2(&x1[15], &x1[14]);
+  idct32_high16_stage4_avx2(x1, cospi, _r, cos_bit);
+
+  // stage 5
+  btf_16_w16_avx2(cospi_p32_p32, cospi_p32_m32, &x1[0], &x1[1], _r, cos_bit);
+  btf_16_w16_avx2(cospi_p48_m16, cospi_p16_p48, &x1[2], &x1[3], _r, cos_bit);
+  btf_16_adds_subs_avx2(&x1[4], &x1[5]);
+  btf_16_adds_subs_avx2(&x1[7], &x1[6]);
+  idct32_high24_stage5_avx2(x1, cospi, _r, cos_bit);
+
+  // stage 6
+  btf_16_adds_subs_avx2(&x1[0], &x1[3]);
+  btf_16_adds_subs_avx2(&x1[1], &x1[2]);
+  idct32_high28_stage6_avx2(x1, cospi, _r, cos_bit);
+
+  idct32_stage7_avx2(x1, cospi, _r, cos_bit);
+  idct32_stage8_avx2(x1, cospi, _r, cos_bit);
+  idct32_stage9_avx2(output, x1);
+}
+
+static INLINE void idct64_stage4_high32_avx2(__m256i *x, const int32_t *cospi,
+                                             const __m256i _r, int8_t cos_bit) {
+  (void)cos_bit;
+  const __m256i cospi_m04_p60 = pair_set_w16_epi16(-cospi[4], cospi[60]);
+  const __m256i cospi_p60_p04 = pair_set_w16_epi16(cospi[60], cospi[4]);
+  const __m256i cospi_m60_m04 = pair_set_w16_epi16(-cospi[60], -cospi[4]);
+  const __m256i cospi_m36_p28 = pair_set_w16_epi16(-cospi[36], cospi[28]);
+  const __m256i cospi_p28_p36 = pair_set_w16_epi16(cospi[28], cospi[36]);
+  const __m256i cospi_m28_m36 = pair_set_w16_epi16(-cospi[28], -cospi[36]);
+  const __m256i cospi_m20_p44 = pair_set_w16_epi16(-cospi[20], cospi[44]);
+  const __m256i cospi_p44_p20 = pair_set_w16_epi16(cospi[44], cospi[20]);
+  const __m256i cospi_m44_m20 = pair_set_w16_epi16(-cospi[44], -cospi[20]);
+  const __m256i cospi_m52_p12 = pair_set_w16_epi16(-cospi[52], cospi[12]);
+  const __m256i cospi_p12_p52 = pair_set_w16_epi16(cospi[12], cospi[52]);
+  const __m256i cospi_m12_m52 = pair_set_w16_epi16(-cospi[12], -cospi[52]);
+  btf_16_w16_avx2(cospi_m04_p60, cospi_p60_p04, &x[33], &x[62], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m60_m04, cospi_m04_p60, &x[34], &x[61], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m36_p28, cospi_p28_p36, &x[37], &x[58], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m28_m36, cospi_m36_p28, &x[38], &x[57], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m20_p44, cospi_p44_p20, &x[41], &x[54], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m44_m20, cospi_m20_p44, &x[42], &x[53], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m52_p12, cospi_p12_p52, &x[45], &x[50], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m12_m52, cospi_m52_p12, &x[46], &x[49], _r, cos_bit);
+}
+
+static INLINE void idct64_stage5_high48_avx2(__m256i *x, const int32_t *cospi,
+                                             const __m256i _r, int8_t cos_bit) {
+  (void)cos_bit;
+  const __m256i cospi_m08_p56 = pair_set_w16_epi16(-cospi[8], cospi[56]);
+  const __m256i cospi_p56_p08 = pair_set_w16_epi16(cospi[56], cospi[8]);
+  const __m256i cospi_m56_m08 = pair_set_w16_epi16(-cospi[56], -cospi[8]);
+  const __m256i cospi_m40_p24 = pair_set_w16_epi16(-cospi[40], cospi[24]);
+  const __m256i cospi_p24_p40 = pair_set_w16_epi16(cospi[24], cospi[40]);
+  const __m256i cospi_m24_m40 = pair_set_w16_epi16(-cospi[24], -cospi[40]);
+  btf_16_w16_avx2(cospi_m08_p56, cospi_p56_p08, &x[17], &x[30], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m56_m08, cospi_m08_p56, &x[18], &x[29], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m40_p24, cospi_p24_p40, &x[21], &x[26], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m24_m40, cospi_m40_p24, &x[22], &x[25], _r, cos_bit);
+  btf_16_adds_subs_avx2(&x[32], &x[35]);
+  btf_16_adds_subs_avx2(&x[33], &x[34]);
+  btf_16_adds_subs_avx2(&x[39], &x[36]);
+  btf_16_adds_subs_avx2(&x[38], &x[37]);
+  btf_16_adds_subs_avx2(&x[40], &x[43]);
+  btf_16_adds_subs_avx2(&x[41], &x[42]);
+  btf_16_adds_subs_avx2(&x[47], &x[44]);
+  btf_16_adds_subs_avx2(&x[46], &x[45]);
+  btf_16_adds_subs_avx2(&x[48], &x[51]);
+  btf_16_adds_subs_avx2(&x[49], &x[50]);
+  btf_16_adds_subs_avx2(&x[55], &x[52]);
+  btf_16_adds_subs_avx2(&x[54], &x[53]);
+  btf_16_adds_subs_avx2(&x[56], &x[59]);
+  btf_16_adds_subs_avx2(&x[57], &x[58]);
+  btf_16_adds_subs_avx2(&x[63], &x[60]);
+  btf_16_adds_subs_avx2(&x[62], &x[61]);
+}
+
+static INLINE void idct64_stage6_high32_avx2(__m256i *x, const int32_t *cospi,
+                                             const __m256i _r, int8_t cos_bit) {
+  (void)cos_bit;
+  const __m256i cospi_m08_p56 = pair_set_w16_epi16(-cospi[8], cospi[56]);
+  const __m256i cospi_p56_p08 = pair_set_w16_epi16(cospi[56], cospi[8]);
+  const __m256i cospi_m56_m08 = pair_set_w16_epi16(-cospi[56], -cospi[8]);
+  const __m256i cospi_m40_p24 = pair_set_w16_epi16(-cospi[40], cospi[24]);
+  const __m256i cospi_p24_p40 = pair_set_w16_epi16(cospi[24], cospi[40]);
+  const __m256i cospi_m24_m40 = pair_set_w16_epi16(-cospi[24], -cospi[40]);
+  btf_16_w16_avx2(cospi_m08_p56, cospi_p56_p08, &x[34], &x[61], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m08_p56, cospi_p56_p08, &x[35], &x[60], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m56_m08, cospi_m08_p56, &x[36], &x[59], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m56_m08, cospi_m08_p56, &x[37], &x[58], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m40_p24, cospi_p24_p40, &x[42], &x[53], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m40_p24, cospi_p24_p40, &x[43], &x[52], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m24_m40, cospi_m40_p24, &x[44], &x[51], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m24_m40, cospi_m40_p24, &x[45], &x[50], _r, cos_bit);
+}
+
+static INLINE void idct64_stage6_high48_avx2(__m256i *x, const int32_t *cospi,
+                                             const __m256i _r, int8_t cos_bit) {
+  btf_16_adds_subs_avx2(&x[16], &x[19]);
+  btf_16_adds_subs_avx2(&x[17], &x[18]);
+  btf_16_adds_subs_avx2(&x[23], &x[20]);
+  btf_16_adds_subs_avx2(&x[22], &x[21]);
+  btf_16_adds_subs_avx2(&x[24], &x[27]);
+  btf_16_adds_subs_avx2(&x[25], &x[26]);
+  btf_16_adds_subs_avx2(&x[31], &x[28]);
+  btf_16_adds_subs_avx2(&x[30], &x[29]);
+  idct64_stage6_high32_avx2(x, cospi, _r, cos_bit);
+}
+
+static INLINE void idct64_stage7_high48_avx2(__m256i *x, const int32_t *cospi,
+                                             const __m256i _r, int8_t cos_bit) {
+  (void)cos_bit;
+  const __m256i cospi_m16_p48 = pair_set_w16_epi16(-cospi[16], cospi[48]);
+  const __m256i cospi_p48_p16 = pair_set_w16_epi16(cospi[48], cospi[16]);
+  const __m256i cospi_m48_m16 = pair_set_w16_epi16(-cospi[48], -cospi[16]);
+  btf_16_w16_avx2(cospi_m16_p48, cospi_p48_p16, &x[18], &x[29], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m16_p48, cospi_p48_p16, &x[19], &x[28], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m48_m16, cospi_m16_p48, &x[20], &x[27], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m48_m16, cospi_m16_p48, &x[21], &x[26], _r, cos_bit);
+  btf_16_adds_subs_avx2(&x[32], &x[39]);
+  btf_16_adds_subs_avx2(&x[33], &x[38]);
+  btf_16_adds_subs_avx2(&x[34], &x[37]);
+  btf_16_adds_subs_avx2(&x[35], &x[36]);
+  btf_16_adds_subs_avx2(&x[47], &x[40]);
+  btf_16_adds_subs_avx2(&x[46], &x[41]);
+  btf_16_adds_subs_avx2(&x[45], &x[42]);
+  btf_16_adds_subs_avx2(&x[44], &x[43]);
+  btf_16_adds_subs_avx2(&x[48], &x[55]);
+  btf_16_adds_subs_avx2(&x[49], &x[54]);
+  btf_16_adds_subs_avx2(&x[50], &x[53]);
+  btf_16_adds_subs_avx2(&x[51], &x[52]);
+  btf_16_adds_subs_avx2(&x[63], &x[56]);
+  btf_16_adds_subs_avx2(&x[62], &x[57]);
+  btf_16_adds_subs_avx2(&x[61], &x[58]);
+  btf_16_adds_subs_avx2(&x[60], &x[59]);
+}
+
+static INLINE void idct64_stage8_high48_avx2(__m256i *x, const int32_t *cospi,
+                                             const __m256i _r, int8_t cos_bit) {
+  (void)cos_bit;
+  const __m256i cospi_m16_p48 = pair_set_w16_epi16(-cospi[16], cospi[48]);
+  const __m256i cospi_p48_p16 = pair_set_w16_epi16(cospi[48], cospi[16]);
+  const __m256i cospi_m48_m16 = pair_set_w16_epi16(-cospi[48], -cospi[16]);
+  btf_16_adds_subs_avx2(&x[16], &x[23]);
+  btf_16_adds_subs_avx2(&x[17], &x[22]);
+  btf_16_adds_subs_avx2(&x[18], &x[21]);
+  btf_16_adds_subs_avx2(&x[19], &x[20]);
+  btf_16_adds_subs_avx2(&x[31], &x[24]);
+  btf_16_adds_subs_avx2(&x[30], &x[25]);
+  btf_16_adds_subs_avx2(&x[29], &x[26]);
+  btf_16_adds_subs_avx2(&x[28], &x[27]);
+  btf_16_w16_avx2(cospi_m16_p48, cospi_p48_p16, &x[36], &x[59], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m16_p48, cospi_p48_p16, &x[37], &x[58], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m16_p48, cospi_p48_p16, &x[38], &x[57], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m16_p48, cospi_p48_p16, &x[39], &x[56], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m48_m16, cospi_m16_p48, &x[40], &x[55], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m48_m16, cospi_m16_p48, &x[41], &x[54], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m48_m16, cospi_m16_p48, &x[42], &x[53], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m48_m16, cospi_m16_p48, &x[43], &x[52], _r, cos_bit);
+}
+
+static INLINE void idct64_stage9_avx2(__m256i *x, const int32_t *cospi,
+                                      const __m256i _r, int8_t cos_bit) {
+  (void)cos_bit;
+  const __m256i cospi_m32_p32 = pair_set_w16_epi16(-cospi[32], cospi[32]);
+  const __m256i cospi_p32_p32 = pair_set_w16_epi16(cospi[32], cospi[32]);
+  btf_16_adds_subs_avx2(&x[0], &x[15]);
+  btf_16_adds_subs_avx2(&x[1], &x[14]);
+  btf_16_adds_subs_avx2(&x[2], &x[13]);
+  btf_16_adds_subs_avx2(&x[3], &x[12]);
+  btf_16_adds_subs_avx2(&x[4], &x[11]);
+  btf_16_adds_subs_avx2(&x[5], &x[10]);
+  btf_16_adds_subs_avx2(&x[6], &x[9]);
+  btf_16_adds_subs_avx2(&x[7], &x[8]);
+  btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x[20], &x[27], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x[21], &x[26], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x[22], &x[25], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x[23], &x[24], _r, cos_bit);
+  btf_16_adds_subs_avx2(&x[32], &x[47]);
+  btf_16_adds_subs_avx2(&x[33], &x[46]);
+  btf_16_adds_subs_avx2(&x[34], &x[45]);
+  btf_16_adds_subs_avx2(&x[35], &x[44]);
+  btf_16_adds_subs_avx2(&x[36], &x[43]);
+  btf_16_adds_subs_avx2(&x[37], &x[42]);
+  btf_16_adds_subs_avx2(&x[38], &x[41]);
+  btf_16_adds_subs_avx2(&x[39], &x[40]);
+  btf_16_adds_subs_avx2(&x[63], &x[48]);
+  btf_16_adds_subs_avx2(&x[62], &x[49]);
+  btf_16_adds_subs_avx2(&x[61], &x[50]);
+  btf_16_adds_subs_avx2(&x[60], &x[51]);
+  btf_16_adds_subs_avx2(&x[59], &x[52]);
+  btf_16_adds_subs_avx2(&x[58], &x[53]);
+  btf_16_adds_subs_avx2(&x[57], &x[54]);
+  btf_16_adds_subs_avx2(&x[56], &x[55]);
+}
+
+static INLINE void idct64_stage10_avx2(__m256i *x, const int32_t *cospi,
+                                       const __m256i _r, int8_t cos_bit) {
+  (void)cos_bit;
+  const __m256i cospi_m32_p32 = pair_set_w16_epi16(-cospi[32], cospi[32]);
+  const __m256i cospi_p32_p32 = pair_set_w16_epi16(cospi[32], cospi[32]);
+  btf_16_adds_subs_avx2(&x[0], &x[31]);
+  btf_16_adds_subs_avx2(&x[1], &x[30]);
+  btf_16_adds_subs_avx2(&x[2], &x[29]);
+  btf_16_adds_subs_avx2(&x[3], &x[28]);
+  btf_16_adds_subs_avx2(&x[4], &x[27]);
+  btf_16_adds_subs_avx2(&x[5], &x[26]);
+  btf_16_adds_subs_avx2(&x[6], &x[25]);
+  btf_16_adds_subs_avx2(&x[7], &x[24]);
+  btf_16_adds_subs_avx2(&x[8], &x[23]);
+  btf_16_adds_subs_avx2(&x[9], &x[22]);
+  btf_16_adds_subs_avx2(&x[10], &x[21]);
+  btf_16_adds_subs_avx2(&x[11], &x[20]);
+  btf_16_adds_subs_avx2(&x[12], &x[19]);
+  btf_16_adds_subs_avx2(&x[13], &x[18]);
+  btf_16_adds_subs_avx2(&x[14], &x[17]);
+  btf_16_adds_subs_avx2(&x[15], &x[16]);
+  btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x[40], &x[55], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x[41], &x[54], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x[42], &x[53], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x[43], &x[52], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x[44], &x[51], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x[45], &x[50], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x[46], &x[49], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x[47], &x[48], _r, cos_bit);
+}
+
+static INLINE void idct64_stage11_avx2(__m256i *output, __m256i *x) {
+  btf_16_adds_subs_out_avx2(&output[0], &output[63], x[0], x[63]);
+  btf_16_adds_subs_out_avx2(&output[1], &output[62], x[1], x[62]);
+  btf_16_adds_subs_out_avx2(&output[2], &output[61], x[2], x[61]);
+  btf_16_adds_subs_out_avx2(&output[3], &output[60], x[3], x[60]);
+  btf_16_adds_subs_out_avx2(&output[4], &output[59], x[4], x[59]);
+  btf_16_adds_subs_out_avx2(&output[5], &output[58], x[5], x[58]);
+  btf_16_adds_subs_out_avx2(&output[6], &output[57], x[6], x[57]);
+  btf_16_adds_subs_out_avx2(&output[7], &output[56], x[7], x[56]);
+  btf_16_adds_subs_out_avx2(&output[8], &output[55], x[8], x[55]);
+  btf_16_adds_subs_out_avx2(&output[9], &output[54], x[9], x[54]);
+  btf_16_adds_subs_out_avx2(&output[10], &output[53], x[10], x[53]);
+  btf_16_adds_subs_out_avx2(&output[11], &output[52], x[11], x[52]);
+  btf_16_adds_subs_out_avx2(&output[12], &output[51], x[12], x[51]);
+  btf_16_adds_subs_out_avx2(&output[13], &output[50], x[13], x[50]);
+  btf_16_adds_subs_out_avx2(&output[14], &output[49], x[14], x[49]);
+  btf_16_adds_subs_out_avx2(&output[15], &output[48], x[15], x[48]);
+  btf_16_adds_subs_out_avx2(&output[16], &output[47], x[16], x[47]);
+  btf_16_adds_subs_out_avx2(&output[17], &output[46], x[17], x[46]);
+  btf_16_adds_subs_out_avx2(&output[18], &output[45], x[18], x[45]);
+  btf_16_adds_subs_out_avx2(&output[19], &output[44], x[19], x[44]);
+  btf_16_adds_subs_out_avx2(&output[20], &output[43], x[20], x[43]);
+  btf_16_adds_subs_out_avx2(&output[21], &output[42], x[21], x[42]);
+  btf_16_adds_subs_out_avx2(&output[22], &output[41], x[22], x[41]);
+  btf_16_adds_subs_out_avx2(&output[23], &output[40], x[23], x[40]);
+  btf_16_adds_subs_out_avx2(&output[24], &output[39], x[24], x[39]);
+  btf_16_adds_subs_out_avx2(&output[25], &output[38], x[25], x[38]);
+  btf_16_adds_subs_out_avx2(&output[26], &output[37], x[26], x[37]);
+  btf_16_adds_subs_out_avx2(&output[27], &output[36], x[27], x[36]);
+  btf_16_adds_subs_out_avx2(&output[28], &output[35], x[28], x[35]);
+  btf_16_adds_subs_out_avx2(&output[29], &output[34], x[29], x[34]);
+  btf_16_adds_subs_out_avx2(&output[30], &output[33], x[30], x[33]);
+  btf_16_adds_subs_out_avx2(&output[31], &output[32], x[31], x[32]);
+}
+
+static void idct64_low1_new_avx2(const __m256i *input, __m256i *output,
+                                 int8_t cos_bit) {
+  (void)cos_bit;
+  const int32_t *cospi = cospi_arr(INV_COS_BIT);
+
+  // stage 1
+  __m256i x[32];
+  x[0] = input[0];
+
+  // stage 2
+  // stage 3
+  // stage 4
+  // stage 5
+  // stage 6
+  btf_16_w16_0_avx2(cospi[32], cospi[32], x[0], x[0], x[1]);
+
+  // stage 7
+  // stage 8
+  // stage 9
+  // stage 10
+  // stage 11
+  output[0] = x[0];
+  output[63] = x[0];
+  output[1] = x[1];
+  output[62] = x[1];
+  output[2] = x[1];
+  output[61] = x[1];
+  output[3] = x[0];
+  output[60] = x[0];
+  output[4] = x[0];
+  output[59] = x[0];
+  output[5] = x[1];
+  output[58] = x[1];
+  output[6] = x[1];
+  output[57] = x[1];
+  output[7] = x[0];
+  output[56] = x[0];
+  output[8] = x[0];
+  output[55] = x[0];
+  output[9] = x[1];
+  output[54] = x[1];
+  output[10] = x[1];
+  output[53] = x[1];
+  output[11] = x[0];
+  output[52] = x[0];
+  output[12] = x[0];
+  output[51] = x[0];
+  output[13] = x[1];
+  output[50] = x[1];
+  output[14] = x[1];
+  output[49] = x[1];
+  output[15] = x[0];
+  output[48] = x[0];
+  output[16] = x[0];
+  output[47] = x[0];
+  output[17] = x[1];
+  output[46] = x[1];
+  output[18] = x[1];
+  output[45] = x[1];
+  output[19] = x[0];
+  output[44] = x[0];
+  output[20] = x[0];
+  output[43] = x[0];
+  output[21] = x[1];
+  output[42] = x[1];
+  output[22] = x[1];
+  output[41] = x[1];
+  output[23] = x[0];
+  output[40] = x[0];
+  output[24] = x[0];
+  output[39] = x[0];
+  output[25] = x[1];
+  output[38] = x[1];
+  output[26] = x[1];
+  output[37] = x[1];
+  output[27] = x[0];
+  output[36] = x[0];
+  output[28] = x[0];
+  output[35] = x[0];
+  output[29] = x[1];
+  output[34] = x[1];
+  output[30] = x[1];
+  output[33] = x[1];
+  output[31] = x[0];
+  output[32] = x[0];
+}
+
+static void idct64_low8_new_avx2(const __m256i *input, __m256i *output,
+                                 int8_t cos_bit) {
+  (void)cos_bit;
+  const int32_t *cospi = cospi_arr(INV_COS_BIT);
+  const __m256i _r = _mm256_set1_epi32(1 << (INV_COS_BIT - 1));
+  const __m256i cospi_m04_p60 = pair_set_w16_epi16(-cospi[4], cospi[60]);
+  const __m256i cospi_p60_p04 = pair_set_w16_epi16(cospi[60], cospi[4]);
+  const __m256i cospi_m36_p28 = pair_set_w16_epi16(-cospi[36], cospi[28]);
+  const __m256i cospi_m28_m36 = pair_set_w16_epi16(-cospi[28], -cospi[36]);
+  const __m256i cospi_m20_p44 = pair_set_w16_epi16(-cospi[20], cospi[44]);
+  const __m256i cospi_p44_p20 = pair_set_w16_epi16(cospi[44], cospi[20]);
+  const __m256i cospi_m52_p12 = pair_set_w16_epi16(-cospi[52], cospi[12]);
+  const __m256i cospi_m12_m52 = pair_set_w16_epi16(-cospi[12], -cospi[52]);
+  const __m256i cospi_m08_p56 = pair_set_w16_epi16(-cospi[8], cospi[56]);
+  const __m256i cospi_p56_p08 = pair_set_w16_epi16(cospi[56], cospi[8]);
+  const __m256i cospi_m40_p24 = pair_set_w16_epi16(-cospi[40], cospi[24]);
+  const __m256i cospi_m24_m40 = pair_set_w16_epi16(-cospi[24], -cospi[40]);
+  const __m256i cospi_p32_p32 = pair_set_w16_epi16(cospi[32], cospi[32]);
+  const __m256i cospi_m16_p48 = pair_set_w16_epi16(-cospi[16], cospi[48]);
+  const __m256i cospi_p48_p16 = pair_set_w16_epi16(cospi[48], cospi[16]);
+  const __m256i cospi_m32_p32 = pair_set_w16_epi16(-cospi[32], cospi[32]);
+
+  // stage 1
+  __m256i x[64];
+  x[0] = input[0];
+  x[8] = input[4];
+  x[16] = input[2];
+  x[24] = input[6];
+  x[32] = input[1];
+  x[40] = input[5];
+  x[48] = input[3];
+  x[56] = input[7];
+
+  // stage 2
+  btf_16_w16_0_avx2(cospi[63], cospi[1], x[32], x[32], x[63]);
+  btf_16_w16_0_avx2(-cospi[57], cospi[7], x[56], x[39], x[56]);
+  btf_16_w16_0_avx2(cospi[59], cospi[5], x[40], x[40], x[55]);
+  btf_16_w16_0_avx2(-cospi[61], cospi[3], x[48], x[47], x[48]);
+
+  // stage 3
+  btf_16_w16_0_avx2(cospi[62], cospi[2], x[16], x[16], x[31]);
+  btf_16_w16_0_avx2(-cospi[58], cospi[6], x[24], x[23], x[24]);
+  x[33] = x[32];
+  x[38] = x[39];
+  x[41] = x[40];
+  x[46] = x[47];
+  x[49] = x[48];
+  x[54] = x[55];
+  x[57] = x[56];
+  x[62] = x[63];
+
+  // stage 4
+  btf_16_w16_0_avx2(cospi[60], cospi[4], x[8], x[8], x[15]);
+  x[17] = x[16];
+  x[22] = x[23];
+  x[25] = x[24];
+  x[30] = x[31];
+  btf_16_w16_avx2(cospi_m04_p60, cospi_p60_p04, &x[33], &x[62], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m28_m36, cospi_m36_p28, &x[38], &x[57], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m20_p44, cospi_p44_p20, &x[41], &x[54], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m12_m52, cospi_m52_p12, &x[46], &x[49], _r, cos_bit);
+
+  // stage 5
+  x[9] = x[8];
+  x[14] = x[15];
+  btf_16_w16_avx2(cospi_m08_p56, cospi_p56_p08, &x[17], &x[30], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m24_m40, cospi_m40_p24, &x[22], &x[25], _r, cos_bit);
+  x[35] = x[32];
+  x[34] = x[33];
+  x[36] = x[39];
+  x[37] = x[38];
+  x[43] = x[40];
+  x[42] = x[41];
+  x[44] = x[47];
+  x[45] = x[46];
+  x[51] = x[48];
+  x[50] = x[49];
+  x[52] = x[55];
+  x[53] = x[54];
+  x[59] = x[56];
+  x[58] = x[57];
+  x[60] = x[63];
+  x[61] = x[62];
+
+  // stage 6
+  btf_16_w16_0_avx2(cospi[32], cospi[32], x[0], x[0], x[1]);
+  btf_16_w16_avx2(cospi_m16_p48, cospi_p48_p16, &x[9], &x[14], _r, cos_bit);
+  x[19] = x[16];
+  x[18] = x[17];
+  x[20] = x[23];
+  x[21] = x[22];
+  x[27] = x[24];
+  x[26] = x[25];
+  x[28] = x[31];
+  x[29] = x[30];
+  idct64_stage6_high32_avx2(x, cospi, _r, cos_bit);
+
+  // stage 7
+  x[3] = x[0];
+  x[2] = x[1];
+  x[11] = x[8];
+  x[10] = x[9];
+  x[12] = x[15];
+  x[13] = x[14];
+  idct64_stage7_high48_avx2(x, cospi, _r, cos_bit);
+
+  // stage 8
+  x[7] = x[0];
+  x[6] = x[1];
+  x[5] = x[2];
+  x[4] = x[3];
+  x[9] = x[9];
+  btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x[10], &x[13], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x[11], &x[12], _r, cos_bit);
+  idct64_stage8_high48_avx2(x, cospi, _r, cos_bit);
+
+  idct64_stage9_avx2(x, cospi, _r, cos_bit);
+  idct64_stage10_avx2(x, cospi, _r, cos_bit);
+  idct64_stage11_avx2(output, x);
+}
+
+static void idct64_low16_new_avx2(const __m256i *input, __m256i *output,
+                                  int8_t cos_bit) {
+  (void)cos_bit;
+  const int32_t *cospi = cospi_arr(INV_COS_BIT);
+  const __m256i _r = _mm256_set1_epi32(1 << (INV_COS_BIT - 1));
+
+  const __m256i cospi_p32_p32 = pair_set_w16_epi16(cospi[32], cospi[32]);
+  const __m256i cospi_m16_p48 = pair_set_w16_epi16(-cospi[16], cospi[48]);
+  const __m256i cospi_p48_p16 = pair_set_w16_epi16(cospi[48], cospi[16]);
+  const __m256i cospi_m48_m16 = pair_set_w16_epi16(-cospi[48], -cospi[16]);
+  const __m256i cospi_m32_p32 = pair_set_w16_epi16(-cospi[32], cospi[32]);
+
+  // stage 1
+  __m256i x[64];
+  x[0] = input[0];
+  x[4] = input[8];
+  x[8] = input[4];
+  x[12] = input[12];
+  x[16] = input[2];
+  x[20] = input[10];
+  x[24] = input[6];
+  x[28] = input[14];
+  x[32] = input[1];
+  x[36] = input[9];
+  x[40] = input[5];
+  x[44] = input[13];
+  x[48] = input[3];
+  x[52] = input[11];
+  x[56] = input[7];
+  x[60] = input[15];
+
+  // stage 2
+  btf_16_w16_0_avx2(cospi[63], cospi[1], x[32], x[32], x[63]);
+  btf_16_w16_0_avx2(-cospi[49], cospi[15], x[60], x[35], x[60]);
+  btf_16_w16_0_avx2(cospi[55], cospi[9], x[36], x[36], x[59]);
+  btf_16_w16_0_avx2(-cospi[57], cospi[7], x[56], x[39], x[56]);
+  btf_16_w16_0_avx2(cospi[59], cospi[5], x[40], x[40], x[55]);
+  btf_16_w16_0_avx2(-cospi[53], cospi[11], x[52], x[43], x[52]);
+  btf_16_w16_0_avx2(cospi[51], cospi[13], x[44], x[44], x[51]);
+  btf_16_w16_0_avx2(-cospi[61], cospi[3], x[48], x[47], x[48]);
+
+  // stage 3
+  btf_16_w16_0_avx2(cospi[62], cospi[2], x[16], x[16], x[31]);
+  btf_16_w16_0_avx2(-cospi[50], cospi[14], x[28], x[19], x[28]);
+  btf_16_w16_0_avx2(cospi[54], cospi[10], x[20], x[20], x[27]);
+  btf_16_w16_0_avx2(-cospi[58], cospi[6], x[24], x[23], x[24]);
+  x[33] = x[32];
+  x[34] = x[35];
+  x[37] = x[36];
+  x[38] = x[39];
+  x[41] = x[40];
+  x[42] = x[43];
+  x[45] = x[44];
+  x[46] = x[47];
+  x[49] = x[48];
+  x[50] = x[51];
+  x[53] = x[52];
+  x[54] = x[55];
+  x[57] = x[56];
+  x[58] = x[59];
+  x[61] = x[60];
+  x[62] = x[63];
+
+  // stage 4
+  btf_16_w16_0_avx2(cospi[60], cospi[4], x[8], x[8], x[15]);
+  btf_16_w16_0_avx2(-cospi[52], cospi[12], x[12], x[11], x[12]);
+  x[17] = x[16];
+  x[18] = x[19];
+  x[21] = x[20];
+  x[22] = x[23];
+  x[25] = x[24];
+  x[26] = x[27];
+  x[29] = x[28];
+  x[30] = x[31];
+  idct64_stage4_high32_avx2(x, cospi, _r, cos_bit);
+
+  // stage 5
+  btf_16_w16_0_avx2(cospi[56], cospi[8], x[4], x[4], x[7]);
+  x[9] = x[8];
+  x[10] = x[11];
+  x[13] = x[12];
+  x[14] = x[15];
+  idct64_stage5_high48_avx2(x, cospi, _r, cos_bit);
+
+  // stage 6
+  btf_16_w16_0_avx2(cospi[32], cospi[32], x[0], x[0], x[1]);
+  x[5] = x[4];
+  x[6] = x[7];
+  btf_16_w16_avx2(cospi_m16_p48, cospi_p48_p16, &x[9], &x[14], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m48_m16, cospi_m16_p48, &x[10], &x[13], _r, cos_bit);
+  idct64_stage6_high48_avx2(x, cospi, _r, cos_bit);
+
+  // stage 7
+  x[3] = x[0];
+  x[2] = x[1];
+  btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x[5], &x[6], _r, cos_bit);
+  btf_16_adds_subs_avx2(&x[8], &x[11]);
+  btf_16_adds_subs_avx2(&x[9], &x[10]);
+  btf_16_adds_subs_avx2(&x[15], &x[12]);
+  btf_16_adds_subs_avx2(&x[14], &x[13]);
+  idct64_stage7_high48_avx2(x, cospi, _r, cos_bit);
+
+  // stage 8
+  btf_16_adds_subs_avx2(&x[0], &x[7]);
+  btf_16_adds_subs_avx2(&x[1], &x[6]);
+  btf_16_adds_subs_avx2(&x[2], &x[5]);
+  btf_16_adds_subs_avx2(&x[3], &x[4]);
+  btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x[10], &x[13], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x[11], &x[12], _r, cos_bit);
+  idct64_stage8_high48_avx2(x, cospi, _r, cos_bit);
+
+  idct64_stage9_avx2(x, cospi, _r, cos_bit);
+  idct64_stage10_avx2(x, cospi, _r, cos_bit);
+  idct64_stage11_avx2(output, x);
+}
+
+static void idct64_low32_new_avx2(const __m256i *input, __m256i *output,
+                                  int8_t cos_bit) {
+  (void)cos_bit;
+  const int32_t *cospi = cospi_arr(INV_COS_BIT);
+  const __m256i _r = _mm256_set1_epi32(1 << (INV_COS_BIT - 1));
+
+  const __m256i cospi_p32_p32 = pair_set_w16_epi16(cospi[32], cospi[32]);
+  const __m256i cospi_m16_p48 = pair_set_w16_epi16(-cospi[16], cospi[48]);
+  const __m256i cospi_p48_p16 = pair_set_w16_epi16(cospi[48], cospi[16]);
+  const __m256i cospi_m48_m16 = pair_set_w16_epi16(-cospi[48], -cospi[16]);
+  const __m256i cospi_m32_p32 = pair_set_w16_epi16(-cospi[32], cospi[32]);
+
+  // stage 1
+  __m256i x[64];
+  x[0] = input[0];
+  x[2] = input[16];
+  x[4] = input[8];
+  x[6] = input[24];
+  x[8] = input[4];
+  x[10] = input[20];
+  x[12] = input[12];
+  x[14] = input[28];
+  x[16] = input[2];
+  x[18] = input[18];
+  x[20] = input[10];
+  x[22] = input[26];
+  x[24] = input[6];
+  x[26] = input[22];
+  x[28] = input[14];
+  x[30] = input[30];
+  x[32] = input[1];
+  x[34] = input[17];
+  x[36] = input[9];
+  x[38] = input[25];
+  x[40] = input[5];
+  x[42] = input[21];
+  x[44] = input[13];
+  x[46] = input[29];
+  x[48] = input[3];
+  x[50] = input[19];
+  x[52] = input[11];
+  x[54] = input[27];
+  x[56] = input[7];
+  x[58] = input[23];
+  x[60] = input[15];
+  x[62] = input[31];
+
+  // stage 2
+  btf_16_w16_0_avx2(cospi[63], cospi[1], x[32], x[32], x[63]);
+  btf_16_w16_0_avx2(-cospi[33], cospi[31], x[62], x[33], x[62]);
+  btf_16_w16_0_avx2(cospi[47], cospi[17], x[34], x[34], x[61]);
+  btf_16_w16_0_avx2(-cospi[49], cospi[15], x[60], x[35], x[60]);
+  btf_16_w16_0_avx2(cospi[55], cospi[9], x[36], x[36], x[59]);
+  btf_16_w16_0_avx2(-cospi[41], cospi[23], x[58], x[37], x[58]);
+  btf_16_w16_0_avx2(cospi[39], cospi[25], x[38], x[38], x[57]);
+  btf_16_w16_0_avx2(-cospi[57], cospi[7], x[56], x[39], x[56]);
+  btf_16_w16_0_avx2(cospi[59], cospi[5], x[40], x[40], x[55]);
+  btf_16_w16_0_avx2(-cospi[37], cospi[27], x[54], x[41], x[54]);
+  btf_16_w16_0_avx2(cospi[43], cospi[21], x[42], x[42], x[53]);
+  btf_16_w16_0_avx2(-cospi[53], cospi[11], x[52], x[43], x[52]);
+  btf_16_w16_0_avx2(cospi[51], cospi[13], x[44], x[44], x[51]);
+  btf_16_w16_0_avx2(-cospi[45], cospi[19], x[50], x[45], x[50]);
+  btf_16_w16_0_avx2(cospi[35], cospi[29], x[46], x[46], x[49]);
+  btf_16_w16_0_avx2(-cospi[61], cospi[3], x[48], x[47], x[48]);
+
+  // stage 3
+  btf_16_w16_0_avx2(cospi[62], cospi[2], x[16], x[16], x[31]);
+  btf_16_w16_0_avx2(-cospi[34], cospi[30], x[30], x[17], x[30]);
+  btf_16_w16_0_avx2(cospi[46], cospi[18], x[18], x[18], x[29]);
+  btf_16_w16_0_avx2(-cospi[50], cospi[14], x[28], x[19], x[28]);
+  btf_16_w16_0_avx2(cospi[54], cospi[10], x[20], x[20], x[27]);
+  btf_16_w16_0_avx2(-cospi[42], cospi[22], x[26], x[21], x[26]);
+  btf_16_w16_0_avx2(cospi[38], cospi[26], x[22], x[22], x[25]);
+  btf_16_w16_0_avx2(-cospi[58], cospi[6], x[24], x[23], x[24]);
+  btf_16_adds_subs_avx2(&x[32], &x[33]);
+  btf_16_adds_subs_avx2(&x[35], &x[34]);
+  btf_16_adds_subs_avx2(&x[36], &x[37]);
+  btf_16_adds_subs_avx2(&x[39], &x[38]);
+  btf_16_adds_subs_avx2(&x[40], &x[41]);
+  btf_16_adds_subs_avx2(&x[43], &x[42]);
+  btf_16_adds_subs_avx2(&x[44], &x[45]);
+  btf_16_adds_subs_avx2(&x[47], &x[46]);
+  btf_16_adds_subs_avx2(&x[48], &x[49]);
+  btf_16_adds_subs_avx2(&x[51], &x[50]);
+  btf_16_adds_subs_avx2(&x[52], &x[53]);
+  btf_16_adds_subs_avx2(&x[55], &x[54]);
+  btf_16_adds_subs_avx2(&x[56], &x[57]);
+  btf_16_adds_subs_avx2(&x[59], &x[58]);
+  btf_16_adds_subs_avx2(&x[60], &x[61]);
+  btf_16_adds_subs_avx2(&x[63], &x[62]);
+
+  // stage 4
+  btf_16_w16_0_avx2(cospi[60], cospi[4], x[8], x[8], x[15]);
+  btf_16_w16_0_avx2(-cospi[36], cospi[28], x[14], x[9], x[14]);
+  btf_16_w16_0_avx2(cospi[44], cospi[20], x[10], x[10], x[13]);
+  btf_16_w16_0_avx2(-cospi[52], cospi[12], x[12], x[11], x[12]);
+  btf_16_adds_subs_avx2(&x[16], &x[17]);
+  btf_16_adds_subs_avx2(&x[19], &x[18]);
+  btf_16_adds_subs_avx2(&x[20], &x[21]);
+  btf_16_adds_subs_avx2(&x[23], &x[22]);
+  btf_16_adds_subs_avx2(&x[24], &x[25]);
+  btf_16_adds_subs_avx2(&x[27], &x[26]);
+  btf_16_adds_subs_avx2(&x[28], &x[29]);
+  btf_16_adds_subs_avx2(&x[31], &x[30]);
+  idct64_stage4_high32_avx2(x, cospi, _r, cos_bit);
+
+  // stage 5
+  btf_16_w16_0_avx2(cospi[56], cospi[8], x[4], x[4], x[7]);
+  btf_16_w16_0_avx2(-cospi[40], cospi[24], x[6], x[5], x[6]);
+  btf_16_adds_subs_avx2(&x[8], &x[9]);
+  btf_16_adds_subs_avx2(&x[11], &x[10]);
+  btf_16_adds_subs_avx2(&x[12], &x[13]);
+  btf_16_adds_subs_avx2(&x[15], &x[14]);
+  idct64_stage5_high48_avx2(x, cospi, _r, cos_bit);
+
+  // stage 6
+  btf_16_w16_0_avx2(cospi[32], cospi[32], x[0], x[0], x[1]);
+  btf_16_w16_0_avx2(cospi[48], cospi[16], x[2], x[2], x[3]);
+  btf_16_adds_subs_avx2(&x[4], &x[5]);
+  btf_16_adds_subs_avx2(&x[7], &x[6]);
+  btf_16_w16_avx2(cospi_m16_p48, cospi_p48_p16, &x[9], &x[14], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m48_m16, cospi_m16_p48, &x[10], &x[13], _r, cos_bit);
+  idct64_stage6_high48_avx2(x, cospi, _r, cos_bit);
+
+  // stage 7
+  btf_16_adds_subs_avx2(&x[0], &x[3]);
+  btf_16_adds_subs_avx2(&x[1], &x[2]);
+  btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x[5], &x[6], _r, cos_bit);
+  btf_16_adds_subs_avx2(&x[8], &x[11]);
+  btf_16_adds_subs_avx2(&x[9], &x[10]);
+  btf_16_adds_subs_avx2(&x[15], &x[12]);
+  btf_16_adds_subs_avx2(&x[14], &x[13]);
+  idct64_stage7_high48_avx2(x, cospi, _r, cos_bit);
+
+  // stage 8
+  btf_16_adds_subs_avx2(&x[0], &x[7]);
+  btf_16_adds_subs_avx2(&x[1], &x[6]);
+  btf_16_adds_subs_avx2(&x[2], &x[5]);
+  btf_16_adds_subs_avx2(&x[3], &x[4]);
+  btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x[10], &x[13], _r, cos_bit);
+  btf_16_w16_avx2(cospi_m32_p32, cospi_p32_p32, &x[11], &x[12], _r, cos_bit);
+  idct64_stage8_high48_avx2(x, cospi, _r, cos_bit);
+
+  // stage 9~11
+  idct64_stage9_avx2(x, cospi, _r, cos_bit);
+  idct64_stage10_avx2(x, cospi, _r, cos_bit);
+  idct64_stage11_avx2(output, x);
+}
+
+// 1D functions process 16 pixels at one time.
+static const transform_1d_avx2
+    lowbd_txfm_all_1d_zeros_w16_arr[TX_SIZES][ITX_TYPES_1D][4] = {
+      {
+          { NULL, NULL, NULL, NULL },
+          { NULL, NULL, NULL, NULL },
+          { NULL, NULL, NULL, NULL },
+      },
+      { { NULL, NULL, NULL, NULL },
+        { NULL, NULL, NULL, NULL },
+        { NULL, NULL, NULL, NULL } },
+      {
+          { idct16_low1_new_avx2, idct16_low8_new_avx2, idct16_new_avx2, NULL },
+          { iadst16_low1_new_avx2, iadst16_low8_new_avx2, iadst16_new_avx2,
+            NULL },
+          { NULL, NULL, NULL, NULL },
+      },
+      { { idct32_low1_new_avx2, idct32_low8_new_avx2, idct32_low16_new_avx2,
+          idct32_new_avx2 },
+        { NULL, NULL, NULL, NULL },
+        { NULL, NULL, NULL, NULL } },
+      { { idct64_low1_new_avx2, idct64_low8_new_avx2, idct64_low16_new_avx2,
+          idct64_low32_new_avx2 },
+        { NULL, NULL, NULL, NULL },
+        { NULL, NULL, NULL, NULL } }
+    };
+
+// only process w >= 16 h >= 16
+static INLINE void lowbd_inv_txfm2d_add_no_identity_avx2(
+    const int32_t *input, uint8_t *output, int stride, TX_TYPE tx_type,
+    TX_SIZE tx_size, int eob) {
+  __m256i buf1[64 * 16];
+  int eobx, eoby;
+  get_eobx_eoby_scan_default(&eobx, &eoby, tx_size, eob);
+  const int8_t *shift = inv_txfm_shift_ls[tx_size];
+  const int txw_idx = get_txw_idx(tx_size);
+  const int txh_idx = get_txh_idx(tx_size);
+  const int cos_bit_col = inv_cos_bit_col[txw_idx][txh_idx];
+  const int cos_bit_row = inv_cos_bit_row[txw_idx][txh_idx];
+  const int txfm_size_col = tx_size_wide[tx_size];
+  const int txfm_size_row = tx_size_high[tx_size];
+  const int buf_size_w_div16 = txfm_size_col >> 4;
+  const int buf_size_nonzero_w_div16 = (eobx + 16) >> 4;
+  const int buf_size_nonzero_h_div16 = (eoby + 16) >> 4;
+  const int input_stride = AOMMIN(32, txfm_size_col);
+  const int rect_type = get_rect_tx_log_ratio(txfm_size_col, txfm_size_row);
+
+  const int fun_idx_x = lowbd_txfm_all_1d_zeros_idx[eobx];
+  const int fun_idx_y = lowbd_txfm_all_1d_zeros_idx[eoby];
+  const transform_1d_avx2 row_txfm =
+      lowbd_txfm_all_1d_zeros_w16_arr[txw_idx][hitx_1d_tab[tx_type]][fun_idx_x];
+  const transform_1d_avx2 col_txfm =
+      lowbd_txfm_all_1d_zeros_w16_arr[txh_idx][vitx_1d_tab[tx_type]][fun_idx_y];
+
+  assert(col_txfm != NULL);
+  assert(row_txfm != NULL);
+  int ud_flip, lr_flip;
+  get_flip_cfg(tx_type, &ud_flip, &lr_flip);
+  for (int i = 0; i < buf_size_nonzero_h_div16; i++) {
+    __m256i buf0[64];
+    const int32_t *input_row = input + (i << 4) * input_stride;
+    for (int j = 0; j < buf_size_nonzero_w_div16; ++j) {
+      __m256i *buf0_cur = buf0 + j * 16;
+      const int32_t *input_cur = input_row + j * 16;
+      load_buffer_32bit_to_16bit_w16_avx2(input_cur, input_stride, buf0_cur,
+                                          16);
+      transpose_16bit_16x16_avx2(buf0_cur, buf0_cur);
+    }
+    if (rect_type == 1 || rect_type == -1) {
+      round_shift_avx2(buf0, buf0, input_stride);  // rect special code
+    }
+    row_txfm(buf0, buf0, cos_bit_row);
+    round_shift_16bit_w16_avx2(buf0, txfm_size_col, shift[0]);
+
+    __m256i *buf1_cur = buf1 + (i << 4);
+    if (lr_flip) {
+      for (int j = 0; j < buf_size_w_div16; ++j) {
+        __m256i temp[16];
+        flip_buf_avx2(buf0 + 16 * j, temp, 16);
+        int offset = txfm_size_row * (buf_size_w_div16 - 1 - j);
+        transpose_16bit_16x16_avx2(temp, buf1_cur + offset);
+      }
+    } else {
+      for (int j = 0; j < buf_size_w_div16; ++j) {
+        transpose_16bit_16x16_avx2(buf0 + 16 * j, buf1_cur + txfm_size_row * j);
+      }
+    }
+  }
+  for (int i = 0; i < buf_size_w_div16; i++) {
+    __m256i *buf1_cur = buf1 + i * txfm_size_row;
+    col_txfm(buf1_cur, buf1_cur, cos_bit_col);
+    round_shift_16bit_w16_avx2(buf1_cur, txfm_size_row, shift[1]);
+  }
+  for (int i = 0; i < buf_size_w_div16; i++) {
+    lowbd_write_buffer_16xn_avx2(buf1 + i * txfm_size_row, output + 16 * i,
+                                 stride, ud_flip, txfm_size_row);
+  }
+}
+
+static INLINE void iidentity_row_16xn_avx2(__m256i *out, const int32_t *input,
+                                           int stride, int shift, int height,
+                                           int txw_idx, int rect_type) {
+  const int32_t *input_row = input;
+  const __m256i scale = _mm256_set1_epi16(NewSqrt2list[txw_idx]);
+  const __m256i _r = _mm256_set1_epi16((1 << (NewSqrt2Bits - 1)) +
+                                       (1 << (NewSqrt2Bits - shift - 1)));
+  const __m256i one = _mm256_set1_epi16(1);
+  const __m256i scale__r = _mm256_unpacklo_epi16(scale, _r);
+  if (rect_type != 1 && rect_type != -1) {
+    for (int i = 0; i < height; ++i) {
+      const __m256i src = load_32bit_to_16bit_w16_avx2(input_row);
+      input_row += stride;
+      __m256i lo = _mm256_unpacklo_epi16(src, one);
+      __m256i hi = _mm256_unpackhi_epi16(src, one);
+      lo = _mm256_madd_epi16(lo, scale__r);
+      hi = _mm256_madd_epi16(hi, scale__r);
+      lo = _mm256_srai_epi32(lo, NewSqrt2Bits - shift);
+      hi = _mm256_srai_epi32(hi, NewSqrt2Bits - shift);
+      out[i] = _mm256_packs_epi32(lo, hi);
+    }
+  } else {
+    const __m256i rect_scale =
+        _mm256_set1_epi16(NewInvSqrt2 << (15 - NewSqrt2Bits));
+    for (int i = 0; i < height; ++i) {
+      __m256i src = load_32bit_to_16bit_w16_avx2(input_row);
+      src = _mm256_mulhrs_epi16(src, rect_scale);
+      input_row += stride;
+      __m256i lo = _mm256_unpacklo_epi16(src, one);
+      __m256i hi = _mm256_unpackhi_epi16(src, one);
+      lo = _mm256_madd_epi16(lo, scale__r);
+      hi = _mm256_madd_epi16(hi, scale__r);
+      lo = _mm256_srai_epi32(lo, NewSqrt2Bits - shift);
+      hi = _mm256_srai_epi32(hi, NewSqrt2Bits - shift);
+      out[i] = _mm256_packs_epi32(lo, hi);
+    }
+  }
+}
+
+static INLINE void iidentity_col_16xn_avx2(uint8_t *output, int stride,
+                                           __m256i *buf, int shift, int height,
+                                           int txh_idx) {
+  const __m256i scale = _mm256_set1_epi16(NewSqrt2list[txh_idx]);
+  const __m256i scale__r = _mm256_set1_epi16(1 << (NewSqrt2Bits - 1));
+  const __m256i shift__r = _mm256_set1_epi32(1 << (-shift - 1));
+  const __m256i one = _mm256_set1_epi16(1);
+  const __m256i scale_coeff = _mm256_unpacklo_epi16(scale, scale__r);
+  for (int h = 0; h < height; ++h) {
+    __m256i lo = _mm256_unpacklo_epi16(buf[h], one);
+    __m256i hi = _mm256_unpackhi_epi16(buf[h], one);
+    lo = _mm256_madd_epi16(lo, scale_coeff);
+    hi = _mm256_madd_epi16(hi, scale_coeff);
+    lo = _mm256_srai_epi32(lo, NewSqrt2Bits);
+    hi = _mm256_srai_epi32(hi, NewSqrt2Bits);
+    lo = _mm256_add_epi32(lo, shift__r);
+    hi = _mm256_add_epi32(hi, shift__r);
+    lo = _mm256_srai_epi32(lo, -shift);
+    hi = _mm256_srai_epi32(hi, -shift);
+    const __m256i x = _mm256_packs_epi32(lo, hi);
+    write_recon_w16_avx2(x, output);
+    output += stride;
+  }
+}
+
+static INLINE void lowbd_inv_txfm2d_add_idtx_avx2(const int32_t *input,
+                                                  uint8_t *output, int stride,
+                                                  TX_SIZE tx_size,
+                                                  int32_t eob) {
+  (void)eob;
+  const int8_t *shift = inv_txfm_shift_ls[tx_size];
+  const int txw_idx = get_txw_idx(tx_size);
+  const int txh_idx = get_txh_idx(tx_size);
+  const int txfm_size_col = tx_size_wide[tx_size];
+  const int txfm_size_row = tx_size_high[tx_size];
+  const int input_stride = AOMMIN(32, txfm_size_col);
+  const int row_max = AOMMIN(32, txfm_size_row);
+  const int rect_type = get_rect_tx_log_ratio(txfm_size_col, txfm_size_row);
+  __m256i buf[32];
+  for (int i = 0; i < input_stride; i += 16) {
+    iidentity_row_16xn_avx2(buf, input + i, input_stride, shift[0], row_max,
+                            txw_idx, rect_type);
+    iidentity_col_16xn_avx2(output + i, stride, buf, shift[1], row_max,
+                            txh_idx);
+  }
+}
+
+static INLINE void lowbd_inv_txfm2d_add_h_identity_avx2(
+    const int32_t *input, uint8_t *output, int stride, TX_TYPE tx_type,
+    TX_SIZE tx_size, int eob) {
+  int eobx, eoby;
+  get_eobx_eoby_scan_h_identity(&eobx, &eoby, tx_size, eob);
+  const int8_t *shift = inv_txfm_shift_ls[tx_size];
+  const int txw_idx = get_txw_idx(tx_size);
+  const int txh_idx = get_txh_idx(tx_size);
+  const int cos_bit_col = inv_cos_bit_col[txw_idx][txh_idx];
+  const int txfm_size_col = tx_size_wide[tx_size];
+  const int txfm_size_row = tx_size_high[tx_size];
+  const int txfm_size_col_notzero = AOMMIN(32, txfm_size_col);
+  const int input_stride = txfm_size_col_notzero;
+  const int buf_size_w_div16 = (eobx + 16) >> 4;
+  const int rect_type = get_rect_tx_log_ratio(txfm_size_col, txfm_size_row);
+
+  const int fun_idx_y = lowbd_txfm_all_1d_zeros_idx[eoby];
+  const transform_1d_avx2 col_txfm =
+      lowbd_txfm_all_1d_zeros_w16_arr[txh_idx][vitx_1d_tab[tx_type]][fun_idx_y];
+
+  assert(col_txfm != NULL);
+
+  int ud_flip, lr_flip;
+  get_flip_cfg(tx_type, &ud_flip, &lr_flip);
+  for (int i = 0; i < buf_size_w_div16; i++) {
+    __m256i buf0[64];
+    iidentity_row_16xn_avx2(buf0, input + (i << 4), input_stride, shift[0],
+                            eoby + 1, txw_idx, rect_type);
+    col_txfm(buf0, buf0, cos_bit_col);
+    __m256i mshift = _mm256_set1_epi16(1 << (15 + shift[1]));
+    int k = ud_flip ? (txfm_size_row - 1) : 0;
+    const int step = ud_flip ? -1 : 1;
+    for (int j = 0; j < txfm_size_row; ++j, k += step) {
+      __m256i res = _mm256_mulhrs_epi16(buf0[k], mshift);
+      write_recon_w16_avx2(res, output + (i << 4) + j * stride);
+    }
+  }
+}
+
+static INLINE void lowbd_inv_txfm2d_add_v_identity_avx2(
+    const int32_t *input, uint8_t *output, int stride, TX_TYPE tx_type,
+    TX_SIZE tx_size, int eob) {
+  __m256i buf1[64];
+  int eobx, eoby;
+  get_eobx_eoby_scan_v_identity(&eobx, &eoby, tx_size, eob);
+  const int8_t *shift = inv_txfm_shift_ls[tx_size];
+  const int txw_idx = get_txw_idx(tx_size);
+  const int txh_idx = get_txh_idx(tx_size);
+  const int cos_bit_row = inv_cos_bit_row[txw_idx][txh_idx];
+  const int txfm_size_col = tx_size_wide[tx_size];
+  const int txfm_size_row = tx_size_high[tx_size];
+  const int buf_size_w_div16 = txfm_size_col >> 4;
+  const int buf_size_h_div16 = (eoby + 16) >> 4;
+  const int input_stride = AOMMIN(32, txfm_size_col);
+  const int rect_type = get_rect_tx_log_ratio(txfm_size_col, txfm_size_row);
+
+  const int fun_idx_x = lowbd_txfm_all_1d_zeros_idx[eobx];
+  const transform_1d_avx2 row_txfm =
+      lowbd_txfm_all_1d_zeros_w16_arr[txw_idx][hitx_1d_tab[tx_type]][fun_idx_x];
+
+  assert(row_txfm != NULL);
+
+  int ud_flip, lr_flip;
+  get_flip_cfg(tx_type, &ud_flip, &lr_flip);
+  for (int i = 0; i < buf_size_h_div16; i++) {
+    __m256i buf0[64];
+    const int32_t *input_row = input + i * input_stride * 16;
+    for (int j = 0; j < AOMMIN(4, buf_size_w_div16); ++j) {
+      __m256i *buf0_cur = buf0 + j * 16;
+      load_buffer_32bit_to_16bit_w16_avx2(input_row + j * 16, input_stride,
+                                          buf0_cur, 16);
+      transpose_16bit_16x16_avx2(buf0_cur, buf0_cur);
+    }
+    if (rect_type == 1 || rect_type == -1) {
+      round_shift_avx2(buf0, buf0, input_stride);  // rect special code
+    }
+    row_txfm(buf0, buf0, cos_bit_row);
+    round_shift_16bit_w16_avx2(buf0, txfm_size_col, shift[0]);
+    __m256i *_buf1 = buf1;
+    if (lr_flip) {
+      for (int j = 0; j < buf_size_w_div16; ++j) {
+        __m256i temp[16];
+        flip_buf_avx2(buf0 + 16 * j, temp, 16);
+        transpose_16bit_16x16_avx2(temp,
+                                   _buf1 + 16 * (buf_size_w_div16 - 1 - j));
+      }
+    } else {
+      for (int j = 0; j < buf_size_w_div16; ++j) {
+        transpose_16bit_16x16_avx2(buf0 + 16 * j, _buf1 + 16 * j);
+      }
+    }
+    for (int j = 0; j < buf_size_w_div16; ++j) {
+      iidentity_col_16xn_avx2(output + i * 16 * stride + j * 16, stride,
+                              buf1 + j * 16, shift[1], 16, txh_idx);
+    }
+  }
+}
+
+// for 32x32,32x64,64x32,64x64,16x32,32x16,64x16,16x64
+static INLINE void lowbd_inv_txfm2d_add_universe_avx2(
+    const int32_t *input, uint8_t *output, int stride, TX_TYPE tx_type,
+    TX_SIZE tx_size, int eob) {
+  (void)eob;
+  switch (tx_type) {
+    case DCT_DCT:
+    case ADST_DCT:   // ADST in vertical, DCT in horizontal
+    case DCT_ADST:   // DCT  in vertical, ADST in horizontal
+    case ADST_ADST:  // ADST in both directions
+    case FLIPADST_DCT:
+    case DCT_FLIPADST:
+    case FLIPADST_FLIPADST:
+    case ADST_FLIPADST:
+    case FLIPADST_ADST:
+      lowbd_inv_txfm2d_add_no_identity_avx2(input, output, stride, tx_type,
+                                            tx_size, eob);
+      break;
+    case IDTX:
+      lowbd_inv_txfm2d_add_idtx_avx2(input, output, stride, tx_size, eob);
+      break;
+    case V_DCT:
+    case V_ADST:
+    case V_FLIPADST:
+      lowbd_inv_txfm2d_add_h_identity_avx2(input, output, stride, tx_type,
+                                           tx_size, eob);
+      break;
+    case H_DCT:
+    case H_ADST:
+    case H_FLIPADST:
+      lowbd_inv_txfm2d_add_v_identity_avx2(input, output, stride, tx_type,
+                                           tx_size, eob);
+      break;
+    default:
+      av1_lowbd_inv_txfm2d_add_ssse3(input, output, stride, tx_type, tx_size,
+                                     eob);
+      break;
+  }
+}
+
+void av1_lowbd_inv_txfm2d_add_avx2(const int32_t *input, uint8_t *output,
+                                   int stride, TX_TYPE tx_type, TX_SIZE tx_size,
+                                   int eob) {
+  switch (tx_size) {
+    case TX_4X4:
+    case TX_8X8:
+    case TX_4X8:
+    case TX_8X4:
+    case TX_8X16:
+    case TX_16X8:
+    case TX_4X16:
+    case TX_16X4:
+    case TX_8X32:
+    case TX_32X8:
+      av1_lowbd_inv_txfm2d_add_ssse3(input, output, stride, tx_type, tx_size,
+                                     eob);
+      break;
+    case TX_16X16:
+    case TX_32X32:
+    case TX_64X64:
+    case TX_16X32:
+    case TX_32X16:
+    case TX_32X64:
+    case TX_64X32:
+    case TX_16X64:
+    case TX_64X16:
+    default:
+      lowbd_inv_txfm2d_add_universe_avx2(input, output, stride, tx_type,
+                                         tx_size, eob);
+      break;
+  }
+}
+
+void av1_inv_txfm_add_avx2(const tran_low_t *dqcoeff, uint8_t *dst, int stride,
+                           const TxfmParam *txfm_param) {
+  const TX_TYPE tx_type = txfm_param->tx_type;
+  if (!txfm_param->lossless) {
+    av1_lowbd_inv_txfm2d_add_avx2(dqcoeff, dst, stride, tx_type,
+                                  txfm_param->tx_size, txfm_param->eob);
+  } else {
+    av1_inv_txfm_add_c(dqcoeff, dst, stride, txfm_param);
+  }
+}
diff --git a/third_party/aom/av1/common/x86/av1_inv_txfm_avx2.h b/third_party/aom/av1/common/x86/av1_inv_txfm_avx2.h
new file mode 100644
index 000000000..f74cbaeaa
--- /dev/null
+++ b/third_party/aom/av1/common/x86/av1_inv_txfm_avx2.h
@@ -0,0 +1,71 @@
+/*
+ * 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_AV1_COMMON_X86_AV1_INV_TXFM_AVX2_H_
+#define AOM_AV1_COMMON_X86_AV1_INV_TXFM_AVX2_H_
+
+#include <immintrin.h>
+
+#include "config/aom_config.h"
+#include "config/av1_rtcd.h"
+
+#include "aom/aom_integer.h"
+#include "aom_dsp/x86/transpose_sse2.h"
+#include "aom_dsp/x86/txfm_common_sse2.h"
+#include "aom_dsp/x86/txfm_common_avx2.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// half input is zero
+#define btf_16_w16_0_avx2(w0, w1, in, out0, out1)  \
+  {                                                \
+    const __m256i _w0 = _mm256_set1_epi16(w0 * 8); \
+    const __m256i _w1 = _mm256_set1_epi16(w1 * 8); \
+    const __m256i _in = in;                        \
+    out0 = _mm256_mulhrs_epi16(_in, _w0);          \
+    out1 = _mm256_mulhrs_epi16(_in, _w1);          \
+  }
+
+static INLINE void round_shift_avx2(const __m256i *input, __m256i *output,
+                                    int size) {
+  const __m256i scale = _mm256_set1_epi16(NewInvSqrt2 * 8);
+  for (int i = 0; i < size; ++i) {
+    output[i] = _mm256_mulhrs_epi16(input[i], scale);
+  }
+}
+
+static INLINE void write_recon_w16_avx2(__m256i res, uint8_t *output) {
+  __m128i pred = _mm_loadu_si128((__m128i const *)(output));
+  __m256i u = _mm256_adds_epi16(_mm256_cvtepu8_epi16(pred), res);
+  __m128i y = _mm256_castsi256_si128(
+      _mm256_permute4x64_epi64(_mm256_packus_epi16(u, u), 168));
+  _mm_storeu_si128((__m128i *)(output), y);
+}
+
+static INLINE void lowbd_write_buffer_16xn_avx2(__m256i *in, uint8_t *output,
+                                                int stride, int flipud,
+                                                int height) {
+  int j = flipud ? (height - 1) : 0;
+  const int step = flipud ? -1 : 1;
+  for (int i = 0; i < height; ++i, j += step) {
+    write_recon_w16_avx2(in[j], output + i * stride);
+  }
+}
+
+void av1_lowbd_inv_txfm2d_add_avx2(const int32_t *input, uint8_t *output,
+                                   int stride, TX_TYPE tx_type, TX_SIZE tx_size,
+                                   int eob);
+#ifdef __cplusplus
+}
+#endif
+
+#endif  // AOM_AV1_COMMON_X86_AV1_INV_TXFM_AVX2_H_
diff --git a/third_party/aom/av1/common/x86/av1_inv_txfm_ssse3.c b/third_party/aom/av1/common/x86/av1_inv_txfm_ssse3.c
new file mode 100644
index 000000000..995bc3da4
--- /dev/null
+++ b/third_party/aom/av1/common/x86/av1_inv_txfm_ssse3.c
@@ -0,0 +1,2923 @@
+/*
+ * 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 "config/aom_config.h"
+#include "config/av1_rtcd.h"
+
+#include "av1/common/av1_inv_txfm1d_cfg.h"
+#include "av1/common/x86/av1_inv_txfm_ssse3.h"
+#include "av1/common/x86/av1_txfm_sse2.h"
+
+// TODO(venkatsanampudi@ittiam.com): move this to header file
+
+// Sqrt2, Sqrt2^2, Sqrt2^3, Sqrt2^4, Sqrt2^5
+static int32_t NewSqrt2list[TX_SIZES] = { 5793, 2 * 4096, 2 * 5793, 4 * 4096,
+                                          4 * 5793 };
+
+// TODO(binpengsmail@gmail.com): replace some for loop with do {} while
+
+static void idct4_new_sse2(const __m128i *input, __m128i *output,
+                           int8_t cos_bit) {
+  (void)cos_bit;
+  const int32_t *cospi = cospi_arr(INV_COS_BIT);
+  const __m128i __rounding = _mm_set1_epi32(1 << (INV_COS_BIT - 1));
+
+  const __m128i cospi_p32_p32 = pair_set_epi16(cospi[32], cospi[32]);
+  const __m128i cospi_p32_m32 = pair_set_epi16(cospi[32], -cospi[32]);
+  const __m128i cospi_p48_m16 = pair_set_epi16(cospi[48], -cospi[16]);
+  const __m128i cospi_p16_p48 = pair_set_epi16(cospi[16], cospi[48]);
+
+  // stage 1
+  __m128i x[4];
+  x[0] = input[0];
+  x[1] = input[2];
+  x[2] = input[1];
+  x[3] = input[3];
+
+  // stage 2
+  btf_16_sse2(cospi_p32_p32, cospi_p32_m32, x[0], x[1], x[0], x[1]);
+  btf_16_sse2(cospi_p48_m16, cospi_p16_p48, x[2], x[3], x[2], x[3]);
+
+  // stage 3
+  btf_16_adds_subs_out_sse2(output[0], output[3], x[0], x[3]);
+  btf_16_adds_subs_out_sse2(output[1], output[2], x[1], x[2]);
+}
+
+void idct4_w4_new_sse2(const __m128i *input, __m128i *output, int8_t cos_bit) {
+  (void)cos_bit;
+  const int32_t *cospi = cospi_arr(INV_COS_BIT);
+  const __m128i __rounding = _mm_set1_epi32(1 << (INV_COS_BIT - 1));
+
+  const __m128i cospi_p32_p32 = pair_set_epi16(cospi[32], cospi[32]);
+  const __m128i cospi_p32_m32 = pair_set_epi16(cospi[32], -cospi[32]);
+  const __m128i cospi_p48_m16 = pair_set_epi16(cospi[48], -cospi[16]);
+  const __m128i cospi_p16_p48 = pair_set_epi16(cospi[16], cospi[48]);
+
+  // stage 1
+  __m128i x[4];
+  x[0] = input[0];
+  x[1] = input[2];
+  x[2] = input[1];
+  x[3] = input[3];
+
+  // stage 2
+  btf_16_4p_sse2(cospi_p32_p32, cospi_p32_m32, x[0], x[1], x[0], x[1]);
+  btf_16_4p_sse2(cospi_p48_m16, cospi_p16_p48, x[2], x[3], x[2], x[3]);
+
+  // stage 3
+  btf_16_adds_subs_out_sse2(output[0], output[3], x[0], x[3]);
+  btf_16_adds_subs_out_sse2(output[1], output[2], x[1], x[2]);
+}
+
+void idct8_low1_new_ssse3(const __m128i *input, __m128i *output,
+                          int8_t cos_bit) {
+  (void)cos_bit;
+  const int32_t *cospi = cospi_arr(INV_COS_BIT);
+
+  // stage 1
+  __m128i x[2];
+  x[0] = input[0];
+
+  // stage 2
+  // stage 3
+  btf_16_ssse3(cospi[32], cospi[32], x[0], x[0], x[1]);
+
+  // stage 4
+  // stage 5
+  output[0] = x[0];
+  output[7] = x[0];
+  output[1] = x[1];
+  output[6] = x[1];
+  output[2] = x[1];
+  output[5] = x[1];
+  output[3] = x[0];
+  output[4] = x[0];
+}
+
+void idct8_new_sse2(const __m128i *input, __m128i *output, int8_t cos_bit) {
+  (void)cos_bit;
+  const int32_t *cospi = cospi_arr(INV_COS_BIT);
+  const __m128i __rounding = _mm_set1_epi32(1 << (INV_COS_BIT - 1));
+
+  const __m128i cospi_p56_m08 = pair_set_epi16(cospi[56], -cospi[8]);
+  const __m128i cospi_p08_p56 = pair_set_epi16(cospi[8], cospi[56]);
+  const __m128i cospi_p24_m40 = pair_set_epi16(cospi[24], -cospi[40]);
+  const __m128i cospi_p40_p24 = pair_set_epi16(cospi[40], cospi[24]);
+  const __m128i cospi_p32_p32 = pair_set_epi16(cospi[32], cospi[32]);
+  const __m128i cospi_p32_m32 = pair_set_epi16(cospi[32], -cospi[32]);
+  const __m128i cospi_p48_m16 = pair_set_epi16(cospi[48], -cospi[16]);
+  const __m128i cospi_p16_p48 = pair_set_epi16(cospi[16], cospi[48]);
+  const __m128i cospi_m32_p32 = pair_set_epi16(-cospi[32], cospi[32]);
+
+  // stage 1
+  __m128i x[8];
+  x[0] = input[0];
+  x[1] = input[4];
+  x[2] = input[2];
+  x[3] = input[6];
+  x[4] = input[1];
+  x[5] = input[5];
+  x[6] = input[3];
+  x[7] = input[7];
+
+  // stage 2
+  btf_16_sse2(cospi_p56_m08, cospi_p08_p56, x[4], x[7], x[4], x[7]);
+  btf_16_sse2(cospi_p24_m40, cospi_p40_p24, x[5], x[6], x[5], x[6]);
+
+  // stage 3
+  btf_16_sse2(cospi_p32_p32, cospi_p32_m32, x[0], x[1], x[0], x[1]);
+  btf_16_sse2(cospi_p48_m16, cospi_p16_p48, x[2], x[3], x[2], x[3]);
+  btf_16_adds_subs_sse2(x[4], x[5]);
+  btf_16_subs_adds_sse2(x[7], x[6]);
+
+  // stage 4
+  btf_16_adds_subs_sse2(x[0], x[3]);
+  btf_16_adds_subs_sse2(x[1], x[2]);
+  btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x[5], x[6], x[5], x[6]);
+
+  // stage 5
+  btf_16_adds_subs_out_sse2(output[0], output[7], x[0], x[7]);
+  btf_16_adds_subs_out_sse2(output[1], output[6], x[1], x[6]);
+  btf_16_adds_subs_out_sse2(output[2], output[5], x[2], x[5]);
+  btf_16_adds_subs_out_sse2(output[3], output[4], x[3], x[4]);
+}
+
+void idct8_w4_new_sse2(const __m128i *input, __m128i *output, int8_t cos_bit) {
+  (void)cos_bit;
+  const int32_t *cospi = cospi_arr(INV_COS_BIT);
+  const __m128i __rounding = _mm_set1_epi32(1 << (INV_COS_BIT - 1));
+
+  const __m128i cospi_p56_m08 = pair_set_epi16(cospi[56], -cospi[8]);
+  const __m128i cospi_p08_p56 = pair_set_epi16(cospi[8], cospi[56]);
+  const __m128i cospi_p24_m40 = pair_set_epi16(cospi[24], -cospi[40]);
+  const __m128i cospi_p40_p24 = pair_set_epi16(cospi[40], cospi[24]);
+  const __m128i cospi_p32_p32 = pair_set_epi16(cospi[32], cospi[32]);
+  const __m128i cospi_p32_m32 = pair_set_epi16(cospi[32], -cospi[32]);
+  const __m128i cospi_p48_m16 = pair_set_epi16(cospi[48], -cospi[16]);
+  const __m128i cospi_p16_p48 = pair_set_epi16(cospi[16], cospi[48]);
+  const __m128i cospi_m32_p32 = pair_set_epi16(-cospi[32], cospi[32]);
+
+  // stage 1
+  __m128i x[8];
+  x[0] = input[0];
+  x[1] = input[4];
+  x[2] = input[2];
+  x[3] = input[6];
+  x[4] = input[1];
+  x[5] = input[5];
+  x[6] = input[3];
+  x[7] = input[7];
+
+  // stage 2
+  btf_16_4p_sse2(cospi_p56_m08, cospi_p08_p56, x[4], x[7], x[4], x[7]);
+  btf_16_4p_sse2(cospi_p24_m40, cospi_p40_p24, x[5], x[6], x[5], x[6]);
+
+  // stage 3
+  btf_16_4p_sse2(cospi_p32_p32, cospi_p32_m32, x[0], x[1], x[0], x[1]);
+  btf_16_4p_sse2(cospi_p48_m16, cospi_p16_p48, x[2], x[3], x[2], x[3]);
+  btf_16_adds_subs_sse2(x[4], x[5]);
+  btf_16_subs_adds_sse2(x[7], x[6]);
+
+  // stage 4
+  btf_16_adds_subs_sse2(x[0], x[3]);
+  btf_16_adds_subs_sse2(x[1], x[2]);
+  btf_16_4p_sse2(cospi_m32_p32, cospi_p32_p32, x[5], x[6], x[5], x[6]);
+
+  // stage 5
+  btf_16_adds_subs_out_sse2(output[0], output[7], x[0], x[7]);
+  btf_16_adds_subs_out_sse2(output[1], output[6], x[1], x[6]);
+  btf_16_adds_subs_out_sse2(output[2], output[5], x[2], x[5]);
+  btf_16_adds_subs_out_sse2(output[3], output[4], x[3], x[4]);
+}
+
+static INLINE void idct16_stage5_sse2(__m128i *x, const int32_t *cospi,
+                                      const __m128i __rounding,
+                                      int8_t cos_bit) {
+  const __m128i cospi_m32_p32 = pair_set_epi16(-cospi[32], cospi[32]);
+  const __m128i cospi_p32_p32 = pair_set_epi16(cospi[32], cospi[32]);
+  btf_16_adds_subs_sse2(x[0], x[3]);
+  btf_16_adds_subs_sse2(x[1], x[2]);
+  btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x[5], x[6], x[5], x[6]);
+  btf_16_adds_subs_sse2(x[8], x[11]);
+  btf_16_adds_subs_sse2(x[9], x[10]);
+  btf_16_subs_adds_sse2(x[15], x[12]);
+  btf_16_subs_adds_sse2(x[14], x[13]);
+}
+
+static INLINE void idct16_stage6_sse2(__m128i *x, const int32_t *cospi,
+                                      const __m128i __rounding,
+                                      int8_t cos_bit) {
+  const __m128i cospi_m32_p32 = pair_set_epi16(-cospi[32], cospi[32]);
+  const __m128i cospi_p32_p32 = pair_set_epi16(cospi[32], cospi[32]);
+  btf_16_adds_subs_sse2(x[0], x[7]);
+  btf_16_adds_subs_sse2(x[1], x[6]);
+  btf_16_adds_subs_sse2(x[2], x[5]);
+  btf_16_adds_subs_sse2(x[3], x[4]);
+  btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x[10], x[13], x[10], x[13]);
+  btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x[11], x[12], x[11], x[12]);
+}
+
+static INLINE void idct16_stage7_sse2(__m128i *output, __m128i *x) {
+  btf_16_adds_subs_out_sse2(output[0], output[15], x[0], x[15]);
+  btf_16_adds_subs_out_sse2(output[1], output[14], x[1], x[14]);
+  btf_16_adds_subs_out_sse2(output[2], output[13], x[2], x[13]);
+  btf_16_adds_subs_out_sse2(output[3], output[12], x[3], x[12]);
+  btf_16_adds_subs_out_sse2(output[4], output[11], x[4], x[11]);
+  btf_16_adds_subs_out_sse2(output[5], output[10], x[5], x[10]);
+  btf_16_adds_subs_out_sse2(output[6], output[9], x[6], x[9]);
+  btf_16_adds_subs_out_sse2(output[7], output[8], x[7], x[8]);
+}
+
+static void idct16_low1_new_ssse3(const __m128i *input, __m128i *output,
+                                  int8_t cos_bit) {
+  (void)cos_bit;
+  const int32_t *cospi = cospi_arr(INV_COS_BIT);
+
+  // stage 1
+  __m128i x[2];
+  x[0] = input[0];
+
+  // stage 2
+  // stage 3
+  // stage 4
+  btf_16_ssse3(cospi[32], cospi[32], x[0], x[0], x[1]);
+
+  // stage 5
+  // stage 6
+  // stage 7
+  output[0] = x[0];
+  output[15] = x[0];
+  output[1] = x[1];
+  output[14] = x[1];
+  output[2] = x[1];
+  output[13] = x[1];
+  output[3] = x[0];
+  output[12] = x[0];
+  output[4] = x[0];
+  output[11] = x[0];
+  output[5] = x[1];
+  output[10] = x[1];
+  output[6] = x[1];
+  output[9] = x[1];
+  output[7] = x[0];
+  output[8] = x[0];
+}
+
+static void idct16_low8_new_ssse3(const __m128i *input, __m128i *output,
+                                  int8_t cos_bit) {
+  (void)cos_bit;
+  const int32_t *cospi = cospi_arr(INV_COS_BIT);
+  const __m128i __rounding = _mm_set1_epi32(1 << (INV_COS_BIT - 1));
+  const __m128i cospi_m16_p48 = pair_set_epi16(-cospi[16], cospi[48]);
+  const __m128i cospi_p48_p16 = pair_set_epi16(cospi[48], cospi[16]);
+  const __m128i cospi_m48_m16 = pair_set_epi16(-cospi[48], -cospi[16]);
+
+  // stage 1
+  __m128i x[16];
+  x[0] = input[0];
+  x[2] = input[4];
+  x[4] = input[2];
+  x[6] = input[6];
+  x[8] = input[1];
+  x[10] = input[5];
+  x[12] = input[3];
+  x[14] = input[7];
+
+  // stage 2
+  btf_16_ssse3(cospi[60], cospi[4], x[8], x[8], x[15]);
+  btf_16_ssse3(-cospi[36], cospi[28], x[14], x[9], x[14]);
+  btf_16_ssse3(cospi[44], cospi[20], x[10], x[10], x[13]);
+  btf_16_ssse3(-cospi[52], cospi[12], x[12], x[11], x[12]);
+
+  // stage 3
+  btf_16_ssse3(cospi[56], cospi[8], x[4], x[4], x[7]);
+  btf_16_ssse3(-cospi[40], cospi[24], x[6], x[5], x[6]);
+  btf_16_adds_subs_sse2(x[8], x[9]);
+  btf_16_subs_adds_sse2(x[11], x[10]);
+  btf_16_adds_subs_sse2(x[12], x[13]);
+  btf_16_subs_adds_sse2(x[15], x[14]);
+
+  // stage 4
+  btf_16_ssse3(cospi[32], cospi[32], x[0], x[0], x[1]);
+  btf_16_ssse3(cospi[48], cospi[16], x[2], x[2], x[3]);
+  btf_16_adds_subs_sse2(x[4], x[5]);
+  btf_16_subs_adds_sse2(x[7], x[6]);
+  btf_16_sse2(cospi_m16_p48, cospi_p48_p16, x[9], x[14], x[9], x[14]);
+  btf_16_sse2(cospi_m48_m16, cospi_m16_p48, x[10], x[13], x[10], x[13]);
+
+  idct16_stage5_sse2(x, cospi, __rounding, cos_bit);
+  idct16_stage6_sse2(x, cospi, __rounding, cos_bit);
+  idct16_stage7_sse2(output, x);
+}
+
+void idct16_new_sse2(const __m128i *input, __m128i *output, int8_t cos_bit) {
+  (void)cos_bit;
+  const int32_t *cospi = cospi_arr(INV_COS_BIT);
+  const __m128i __rounding = _mm_set1_epi32(1 << (INV_COS_BIT - 1));
+
+  const __m128i cospi_p60_m04 = pair_set_epi16(cospi[60], -cospi[4]);
+  const __m128i cospi_p04_p60 = pair_set_epi16(cospi[4], cospi[60]);
+  const __m128i cospi_p28_m36 = pair_set_epi16(cospi[28], -cospi[36]);
+  const __m128i cospi_p36_p28 = pair_set_epi16(cospi[36], cospi[28]);
+  const __m128i cospi_p44_m20 = pair_set_epi16(cospi[44], -cospi[20]);
+  const __m128i cospi_p20_p44 = pair_set_epi16(cospi[20], cospi[44]);
+  const __m128i cospi_p12_m52 = pair_set_epi16(cospi[12], -cospi[52]);
+  const __m128i cospi_p52_p12 = pair_set_epi16(cospi[52], cospi[12]);
+  const __m128i cospi_p56_m08 = pair_set_epi16(cospi[56], -cospi[8]);
+  const __m128i cospi_p08_p56 = pair_set_epi16(cospi[8], cospi[56]);
+  const __m128i cospi_p24_m40 = pair_set_epi16(cospi[24], -cospi[40]);
+  const __m128i cospi_p40_p24 = pair_set_epi16(cospi[40], cospi[24]);
+  const __m128i cospi_p32_p32 = pair_set_epi16(cospi[32], cospi[32]);
+  const __m128i cospi_p32_m32 = pair_set_epi16(cospi[32], -cospi[32]);
+  const __m128i cospi_p48_m16 = pair_set_epi16(cospi[48], -cospi[16]);
+  const __m128i cospi_p16_p48 = pair_set_epi16(cospi[16], cospi[48]);
+  const __m128i cospi_m16_p48 = pair_set_epi16(-cospi[16], cospi[48]);
+  const __m128i cospi_p48_p16 = pair_set_epi16(cospi[48], cospi[16]);
+  const __m128i cospi_m48_m16 = pair_set_epi16(-cospi[48], -cospi[16]);
+
+  // stage 1
+  __m128i x[16];
+  x[0] = input[0];
+  x[1] = input[8];
+  x[2] = input[4];
+  x[3] = input[12];
+  x[4] = input[2];
+  x[5] = input[10];
+  x[6] = input[6];
+  x[7] = input[14];
+  x[8] = input[1];
+  x[9] = input[9];
+  x[10] = input[5];
+  x[11] = input[13];
+  x[12] = input[3];
+  x[13] = input[11];
+  x[14] = input[7];
+  x[15] = input[15];
+
+  // stage 2
+  btf_16_sse2(cospi_p60_m04, cospi_p04_p60, x[8], x[15], x[8], x[15]);
+  btf_16_sse2(cospi_p28_m36, cospi_p36_p28, x[9], x[14], x[9], x[14]);
+  btf_16_sse2(cospi_p44_m20, cospi_p20_p44, x[10], x[13], x[10], x[13]);
+  btf_16_sse2(cospi_p12_m52, cospi_p52_p12, x[11], x[12], x[11], x[12]);
+
+  // stage 3
+  btf_16_sse2(cospi_p56_m08, cospi_p08_p56, x[4], x[7], x[4], x[7]);
+  btf_16_sse2(cospi_p24_m40, cospi_p40_p24, x[5], x[6], x[5], x[6]);
+  btf_16_adds_subs_sse2(x[8], x[9]);
+  btf_16_subs_adds_sse2(x[11], x[10]);
+  btf_16_adds_subs_sse2(x[12], x[13]);
+  btf_16_subs_adds_sse2(x[15], x[14]);
+
+  // stage 4
+  btf_16_sse2(cospi_p32_p32, cospi_p32_m32, x[0], x[1], x[0], x[1]);
+  btf_16_sse2(cospi_p48_m16, cospi_p16_p48, x[2], x[3], x[2], x[3]);
+  btf_16_adds_subs_sse2(x[4], x[5]);
+  btf_16_subs_adds_sse2(x[7], x[6]);
+  btf_16_sse2(cospi_m16_p48, cospi_p48_p16, x[9], x[14], x[9], x[14]);
+  btf_16_sse2(cospi_m48_m16, cospi_m16_p48, x[10], x[13], x[10], x[13]);
+
+  // stage 5~7
+  idct16_stage5_sse2(x, cospi, __rounding, cos_bit);
+  idct16_stage6_sse2(x, cospi, __rounding, cos_bit);
+  idct16_stage7_sse2(output, x);
+}
+
+void idct16_w4_new_sse2(const __m128i *input, __m128i *output, int8_t cos_bit) {
+  (void)cos_bit;
+  const int32_t *cospi = cospi_arr(INV_COS_BIT);
+  const __m128i __rounding = _mm_set1_epi32(1 << (INV_COS_BIT - 1));
+
+  const __m128i cospi_p60_m04 = pair_set_epi16(cospi[60], -cospi[4]);
+  const __m128i cospi_p04_p60 = pair_set_epi16(cospi[4], cospi[60]);
+  const __m128i cospi_p28_m36 = pair_set_epi16(cospi[28], -cospi[36]);
+  const __m128i cospi_p36_p28 = pair_set_epi16(cospi[36], cospi[28]);
+  const __m128i cospi_p44_m20 = pair_set_epi16(cospi[44], -cospi[20]);
+  const __m128i cospi_p20_p44 = pair_set_epi16(cospi[20], cospi[44]);
+  const __m128i cospi_p12_m52 = pair_set_epi16(cospi[12], -cospi[52]);
+  const __m128i cospi_p52_p12 = pair_set_epi16(cospi[52], cospi[12]);
+  const __m128i cospi_p56_m08 = pair_set_epi16(cospi[56], -cospi[8]);
+  const __m128i cospi_p08_p56 = pair_set_epi16(cospi[8], cospi[56]);
+  const __m128i cospi_p24_m40 = pair_set_epi16(cospi[24], -cospi[40]);
+  const __m128i cospi_p40_p24 = pair_set_epi16(cospi[40], cospi[24]);
+  const __m128i cospi_p32_p32 = pair_set_epi16(cospi[32], cospi[32]);
+  const __m128i cospi_p32_m32 = pair_set_epi16(cospi[32], -cospi[32]);
+  const __m128i cospi_p48_m16 = pair_set_epi16(cospi[48], -cospi[16]);
+  const __m128i cospi_p16_p48 = pair_set_epi16(cospi[16], cospi[48]);
+  const __m128i cospi_m16_p48 = pair_set_epi16(-cospi[16], cospi[48]);
+  const __m128i cospi_p48_p16 = pair_set_epi16(cospi[48], cospi[16]);
+  const __m128i cospi_m48_m16 = pair_set_epi16(-cospi[48], -cospi[16]);
+  const __m128i cospi_m32_p32 = pair_set_epi16(-cospi[32], cospi[32]);
+
+  // stage 1
+  __m128i x[16];
+  x[0] = input[0];
+  x[1] = input[8];
+  x[2] = input[4];
+  x[3] = input[12];
+  x[4] = input[2];
+  x[5] = input[10];
+  x[6] = input[6];
+  x[7] = input[14];
+  x[8] = input[1];
+  x[9] = input[9];
+  x[10] = input[5];
+  x[11] = input[13];
+  x[12] = input[3];
+  x[13] = input[11];
+  x[14] = input[7];
+  x[15] = input[15];
+
+  // stage 2
+  btf_16_4p_sse2(cospi_p60_m04, cospi_p04_p60, x[8], x[15], x[8], x[15]);
+  btf_16_4p_sse2(cospi_p28_m36, cospi_p36_p28, x[9], x[14], x[9], x[14]);
+  btf_16_4p_sse2(cospi_p44_m20, cospi_p20_p44, x[10], x[13], x[10], x[13]);
+  btf_16_4p_sse2(cospi_p12_m52, cospi_p52_p12, x[11], x[12], x[11], x[12]);
+
+  // stage 3
+  btf_16_4p_sse2(cospi_p56_m08, cospi_p08_p56, x[4], x[7], x[4], x[7]);
+  btf_16_4p_sse2(cospi_p24_m40, cospi_p40_p24, x[5], x[6], x[5], x[6]);
+  btf_16_adds_subs_sse2(x[8], x[9]);
+  btf_16_subs_adds_sse2(x[11], x[10]);
+  btf_16_adds_subs_sse2(x[12], x[13]);
+  btf_16_subs_adds_sse2(x[15], x[14]);
+
+  // stage 4
+  btf_16_4p_sse2(cospi_p32_p32, cospi_p32_m32, x[0], x[1], x[0], x[1]);
+  btf_16_4p_sse2(cospi_p48_m16, cospi_p16_p48, x[2], x[3], x[2], x[3]);
+  btf_16_adds_subs_sse2(x[4], x[5]);
+  btf_16_subs_adds_sse2(x[7], x[6]);
+  btf_16_4p_sse2(cospi_m16_p48, cospi_p48_p16, x[9], x[14], x[9], x[14]);
+  btf_16_4p_sse2(cospi_m48_m16, cospi_m16_p48, x[10], x[13], x[10], x[13]);
+
+  // stage 5
+  btf_16_adds_subs_sse2(x[0], x[3]);
+  btf_16_adds_subs_sse2(x[1], x[2]);
+  btf_16_4p_sse2(cospi_m32_p32, cospi_p32_p32, x[5], x[6], x[5], x[6]);
+  btf_16_adds_subs_sse2(x[8], x[11]);
+  btf_16_adds_subs_sse2(x[9], x[10]);
+  btf_16_subs_adds_sse2(x[15], x[12]);
+  btf_16_subs_adds_sse2(x[14], x[13]);
+
+  // stage 6
+  btf_16_adds_subs_sse2(x[0], x[7]);
+  btf_16_adds_subs_sse2(x[1], x[6]);
+  btf_16_adds_subs_sse2(x[2], x[5]);
+  btf_16_adds_subs_sse2(x[3], x[4]);
+  btf_16_4p_sse2(cospi_m32_p32, cospi_p32_p32, x[10], x[13], x[10], x[13]);
+  btf_16_4p_sse2(cospi_m32_p32, cospi_p32_p32, x[11], x[12], x[11], x[12]);
+
+  // stage 7
+  idct16_stage7_sse2(output, x);
+}
+
+static INLINE void idct32_high16_stage3_sse2(__m128i *x) {
+  btf_16_adds_subs_sse2(x[16], x[17]);
+  btf_16_subs_adds_sse2(x[19], x[18]);
+  btf_16_adds_subs_sse2(x[20], x[21]);
+  btf_16_subs_adds_sse2(x[23], x[22]);
+  btf_16_adds_subs_sse2(x[24], x[25]);
+  btf_16_subs_adds_sse2(x[27], x[26]);
+  btf_16_adds_subs_sse2(x[28], x[29]);
+  btf_16_subs_adds_sse2(x[31], x[30]);
+}
+
+static INLINE void idct32_high16_stage4_sse2(__m128i *x, const int32_t *cospi,
+                                             const __m128i __rounding,
+                                             int8_t cos_bit) {
+  const __m128i cospi_m08_p56 = pair_set_epi16(-cospi[8], cospi[56]);
+  const __m128i cospi_p56_p08 = pair_set_epi16(cospi[56], cospi[8]);
+  const __m128i cospi_m56_m08 = pair_set_epi16(-cospi[56], -cospi[8]);
+  const __m128i cospi_m40_p24 = pair_set_epi16(-cospi[40], cospi[24]);
+  const __m128i cospi_p24_p40 = pair_set_epi16(cospi[24], cospi[40]);
+  const __m128i cospi_m24_m40 = pair_set_epi16(-cospi[24], -cospi[40]);
+  btf_16_sse2(cospi_m08_p56, cospi_p56_p08, x[17], x[30], x[17], x[30]);
+  btf_16_sse2(cospi_m56_m08, cospi_m08_p56, x[18], x[29], x[18], x[29]);
+  btf_16_sse2(cospi_m40_p24, cospi_p24_p40, x[21], x[26], x[21], x[26]);
+  btf_16_sse2(cospi_m24_m40, cospi_m40_p24, x[22], x[25], x[22], x[25]);
+}
+
+static INLINE void idct32_high24_stage5_sse2(__m128i *x, const int32_t *cospi,
+                                             const __m128i __rounding,
+                                             int8_t cos_bit) {
+  const __m128i cospi_m16_p48 = pair_set_epi16(-cospi[16], cospi[48]);
+  const __m128i cospi_p48_p16 = pair_set_epi16(cospi[48], cospi[16]);
+  const __m128i cospi_m48_m16 = pair_set_epi16(-cospi[48], -cospi[16]);
+  btf_16_sse2(cospi_m16_p48, cospi_p48_p16, x[9], x[14], x[9], x[14]);
+  btf_16_sse2(cospi_m48_m16, cospi_m16_p48, x[10], x[13], x[10], x[13]);
+  btf_16_adds_subs_sse2(x[16], x[19]);
+  btf_16_adds_subs_sse2(x[17], x[18]);
+  btf_16_subs_adds_sse2(x[23], x[20]);
+  btf_16_subs_adds_sse2(x[22], x[21]);
+  btf_16_adds_subs_sse2(x[24], x[27]);
+  btf_16_adds_subs_sse2(x[25], x[26]);
+  btf_16_subs_adds_sse2(x[31], x[28]);
+  btf_16_subs_adds_sse2(x[30], x[29]);
+}
+
+static INLINE void idct32_high28_stage6_sse2(__m128i *x, const int32_t *cospi,
+                                             const __m128i __rounding,
+                                             int8_t cos_bit) {
+  const __m128i cospi_m32_p32 = pair_set_epi16(-cospi[32], cospi[32]);
+  const __m128i cospi_p32_p32 = pair_set_epi16(cospi[32], cospi[32]);
+  const __m128i cospi_m16_p48 = pair_set_epi16(-cospi[16], cospi[48]);
+  const __m128i cospi_p48_p16 = pair_set_epi16(cospi[48], cospi[16]);
+  const __m128i cospi_m48_m16 = pair_set_epi16(-cospi[48], -cospi[16]);
+  btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x[5], x[6], x[5], x[6]);
+  btf_16_adds_subs_sse2(x[8], x[11]);
+  btf_16_adds_subs_sse2(x[9], x[10]);
+  btf_16_subs_adds_sse2(x[15], x[12]);
+  btf_16_subs_adds_sse2(x[14], x[13]);
+  btf_16_sse2(cospi_m16_p48, cospi_p48_p16, x[18], x[29], x[18], x[29]);
+  btf_16_sse2(cospi_m16_p48, cospi_p48_p16, x[19], x[28], x[19], x[28]);
+  btf_16_sse2(cospi_m48_m16, cospi_m16_p48, x[20], x[27], x[20], x[27]);
+  btf_16_sse2(cospi_m48_m16, cospi_m16_p48, x[21], x[26], x[21], x[26]);
+}
+
+static INLINE void idct32_stage7_sse2(__m128i *x, const int32_t *cospi,
+                                      const __m128i __rounding,
+                                      int8_t cos_bit) {
+  const __m128i cospi_m32_p32 = pair_set_epi16(-cospi[32], cospi[32]);
+  const __m128i cospi_p32_p32 = pair_set_epi16(cospi[32], cospi[32]);
+  btf_16_adds_subs_sse2(x[0], x[7]);
+  btf_16_adds_subs_sse2(x[1], x[6]);
+  btf_16_adds_subs_sse2(x[2], x[5]);
+  btf_16_adds_subs_sse2(x[3], x[4]);
+  btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x[10], x[13], x[10], x[13]);
+  btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x[11], x[12], x[11], x[12]);
+  btf_16_adds_subs_sse2(x[16], x[23]);
+  btf_16_adds_subs_sse2(x[17], x[22]);
+  btf_16_adds_subs_sse2(x[18], x[21]);
+  btf_16_adds_subs_sse2(x[19], x[20]);
+  btf_16_subs_adds_sse2(x[31], x[24]);
+  btf_16_subs_adds_sse2(x[30], x[25]);
+  btf_16_subs_adds_sse2(x[29], x[26]);
+  btf_16_subs_adds_sse2(x[28], x[27]);
+}
+
+static INLINE void idct32_stage8_sse2(__m128i *x, const int32_t *cospi,
+                                      const __m128i __rounding,
+                                      int8_t cos_bit) {
+  const __m128i cospi_m32_p32 = pair_set_epi16(-cospi[32], cospi[32]);
+  const __m128i cospi_p32_p32 = pair_set_epi16(cospi[32], cospi[32]);
+  btf_16_adds_subs_sse2(x[0], x[15]);
+  btf_16_adds_subs_sse2(x[1], x[14]);
+  btf_16_adds_subs_sse2(x[2], x[13]);
+  btf_16_adds_subs_sse2(x[3], x[12]);
+  btf_16_adds_subs_sse2(x[4], x[11]);
+  btf_16_adds_subs_sse2(x[5], x[10]);
+  btf_16_adds_subs_sse2(x[6], x[9]);
+  btf_16_adds_subs_sse2(x[7], x[8]);
+  btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x[20], x[27], x[20], x[27]);
+  btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x[21], x[26], x[21], x[26]);
+  btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x[22], x[25], x[22], x[25]);
+  btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x[23], x[24], x[23], x[24]);
+}
+
+static INLINE void idct32_stage9_sse2(__m128i *output, __m128i *x) {
+  btf_16_adds_subs_out_sse2(output[0], output[31], x[0], x[31]);
+  btf_16_adds_subs_out_sse2(output[1], output[30], x[1], x[30]);
+  btf_16_adds_subs_out_sse2(output[2], output[29], x[2], x[29]);
+  btf_16_adds_subs_out_sse2(output[3], output[28], x[3], x[28]);
+  btf_16_adds_subs_out_sse2(output[4], output[27], x[4], x[27]);
+  btf_16_adds_subs_out_sse2(output[5], output[26], x[5], x[26]);
+  btf_16_adds_subs_out_sse2(output[6], output[25], x[6], x[25]);
+  btf_16_adds_subs_out_sse2(output[7], output[24], x[7], x[24]);
+  btf_16_adds_subs_out_sse2(output[8], output[23], x[8], x[23]);
+  btf_16_adds_subs_out_sse2(output[9], output[22], x[9], x[22]);
+  btf_16_adds_subs_out_sse2(output[10], output[21], x[10], x[21]);
+  btf_16_adds_subs_out_sse2(output[11], output[20], x[11], x[20]);
+  btf_16_adds_subs_out_sse2(output[12], output[19], x[12], x[19]);
+  btf_16_adds_subs_out_sse2(output[13], output[18], x[13], x[18]);
+  btf_16_adds_subs_out_sse2(output[14], output[17], x[14], x[17]);
+  btf_16_adds_subs_out_sse2(output[15], output[16], x[15], x[16]);
+}
+
+static void idct32_low1_new_ssse3(const __m128i *input, __m128i *output,
+                                  int8_t cos_bit) {
+  (void)cos_bit;
+  const int32_t *cospi = cospi_arr(INV_COS_BIT);
+
+  // stage 1
+  __m128i x[2];
+  x[0] = input[0];
+
+  // stage 2
+  // stage 3
+  // stage 4
+  // stage 5
+  btf_16_ssse3(cospi[32], cospi[32], x[0], x[0], x[1]);
+
+  // stage 6
+  // stage 7
+  // stage 8
+  // stage 9
+  output[0] = x[0];
+  output[31] = x[0];
+  output[1] = x[1];
+  output[30] = x[1];
+  output[2] = x[1];
+  output[29] = x[1];
+  output[3] = x[0];
+  output[28] = x[0];
+  output[4] = x[0];
+  output[27] = x[0];
+  output[5] = x[1];
+  output[26] = x[1];
+  output[6] = x[1];
+  output[25] = x[1];
+  output[7] = x[0];
+  output[24] = x[0];
+  output[8] = x[0];
+  output[23] = x[0];
+  output[9] = x[1];
+  output[22] = x[1];
+  output[10] = x[1];
+  output[21] = x[1];
+  output[11] = x[0];
+  output[20] = x[0];
+  output[12] = x[0];
+  output[19] = x[0];
+  output[13] = x[1];
+  output[18] = x[1];
+  output[14] = x[1];
+  output[17] = x[1];
+  output[15] = x[0];
+  output[16] = x[0];
+}
+
+static void idct32_low8_new_ssse3(const __m128i *input, __m128i *output,
+                                  int8_t cos_bit) {
+  (void)cos_bit;
+  const int32_t *cospi = cospi_arr(INV_COS_BIT);
+  const __m128i __rounding = _mm_set1_epi32(1 << (INV_COS_BIT - 1));
+
+  // stage 1
+  __m128i x[32];
+  x[0] = input[0];
+  x[4] = input[4];
+  x[8] = input[2];
+  x[12] = input[6];
+  x[16] = input[1];
+  x[20] = input[5];
+  x[24] = input[3];
+  x[28] = input[7];
+
+  // stage 2
+  btf_16_ssse3(cospi[62], cospi[2], x[16], x[16], x[31]);
+  btf_16_ssse3(-cospi[50], cospi[14], x[28], x[19], x[28]);
+  btf_16_ssse3(cospi[54], cospi[10], x[20], x[20], x[27]);
+  btf_16_ssse3(-cospi[58], cospi[6], x[24], x[23], x[24]);
+
+  // stage 3
+  btf_16_ssse3(cospi[60], cospi[4], x[8], x[8], x[15]);
+  btf_16_ssse3(-cospi[52], cospi[12], x[12], x[11], x[12]);
+  x[17] = x[16];
+  x[18] = x[19];
+  x[21] = x[20];
+  x[22] = x[23];
+  x[25] = x[24];
+  x[26] = x[27];
+  x[29] = x[28];
+  x[30] = x[31];
+
+  // stage 4
+  btf_16_ssse3(cospi[56], cospi[8], x[4], x[4], x[7]);
+  x[9] = x[8];
+  x[10] = x[11];
+  x[13] = x[12];
+  x[14] = x[15];
+  idct32_high16_stage4_sse2(x, cospi, __rounding, cos_bit);
+
+  // stage 5
+  btf_16_ssse3(cospi[32], cospi[32], x[0], x[0], x[1]);
+  x[5] = x[4];
+  x[6] = x[7];
+  idct32_high24_stage5_sse2(x, cospi, __rounding, cos_bit);
+  // stage 6
+  x[3] = x[0];
+  x[2] = x[1];
+  idct32_high28_stage6_sse2(x, cospi, __rounding, cos_bit);
+
+  idct32_stage7_sse2(x, cospi, __rounding, cos_bit);
+  idct32_stage8_sse2(x, cospi, __rounding, cos_bit);
+  idct32_stage9_sse2(output, x);
+}
+
+static void idct32_low16_new_ssse3(const __m128i *input, __m128i *output,
+                                   int8_t cos_bit) {
+  (void)cos_bit;
+  const int32_t *cospi = cospi_arr(INV_COS_BIT);
+  const __m128i __rounding = _mm_set1_epi32(1 << (INV_COS_BIT - 1));
+
+  // stage 1
+  __m128i x[32];
+  x[0] = input[0];
+  x[2] = input[8];
+  x[4] = input[4];
+  x[6] = input[12];
+  x[8] = input[2];
+  x[10] = input[10];
+  x[12] = input[6];
+  x[14] = input[14];
+  x[16] = input[1];
+  x[18] = input[9];
+  x[20] = input[5];
+  x[22] = input[13];
+  x[24] = input[3];
+  x[26] = input[11];
+  x[28] = input[7];
+  x[30] = input[15];
+
+  // stage 2
+  btf_16_ssse3(cospi[62], cospi[2], x[16], x[16], x[31]);
+  btf_16_ssse3(-cospi[34], cospi[30], x[30], x[17], x[30]);
+  btf_16_ssse3(cospi[46], cospi[18], x[18], x[18], x[29]);
+  btf_16_ssse3(-cospi[50], cospi[14], x[28], x[19], x[28]);
+  btf_16_ssse3(cospi[54], cospi[10], x[20], x[20], x[27]);
+  btf_16_ssse3(-cospi[42], cospi[22], x[26], x[21], x[26]);
+  btf_16_ssse3(cospi[38], cospi[26], x[22], x[22], x[25]);
+  btf_16_ssse3(-cospi[58], cospi[6], x[24], x[23], x[24]);
+
+  // stage 3
+  btf_16_ssse3(cospi[60], cospi[4], x[8], x[8], x[15]);
+  btf_16_ssse3(-cospi[36], cospi[28], x[14], x[9], x[14]);
+  btf_16_ssse3(cospi[44], cospi[20], x[10], x[10], x[13]);
+  btf_16_ssse3(-cospi[52], cospi[12], x[12], x[11], x[12]);
+  idct32_high16_stage3_sse2(x);
+
+  // stage 4
+  btf_16_ssse3(cospi[56], cospi[8], x[4], x[4], x[7]);
+  btf_16_ssse3(-cospi[40], cospi[24], x[6], x[5], x[6]);
+  btf_16_adds_subs_sse2(x[8], x[9]);
+  btf_16_subs_adds_sse2(x[11], x[10]);
+  btf_16_adds_subs_sse2(x[12], x[13]);
+  btf_16_subs_adds_sse2(x[15], x[14]);
+  idct32_high16_stage4_sse2(x, cospi, __rounding, cos_bit);
+
+  // stage 5
+  btf_16_ssse3(cospi[32], cospi[32], x[0], x[0], x[1]);
+  btf_16_ssse3(cospi[48], cospi[16], x[2], x[2], x[3]);
+  btf_16_adds_subs_sse2(x[4], x[5]);
+  btf_16_subs_adds_sse2(x[7], x[6]);
+  idct32_high24_stage5_sse2(x, cospi, __rounding, cos_bit);
+
+  btf_16_adds_subs_sse2(x[0], x[3]);
+  btf_16_adds_subs_sse2(x[1], x[2]);
+  idct32_high28_stage6_sse2(x, cospi, __rounding, cos_bit);
+
+  idct32_stage7_sse2(x, cospi, __rounding, cos_bit);
+  idct32_stage8_sse2(x, cospi, __rounding, cos_bit);
+  idct32_stage9_sse2(output, x);
+}
+
+static void idct32_new_sse2(const __m128i *input, __m128i *output,
+                            int8_t cos_bit) {
+  (void)cos_bit;
+  const int32_t *cospi = cospi_arr(INV_COS_BIT);
+  const __m128i __rounding = _mm_set1_epi32(1 << (INV_COS_BIT - 1));
+
+  const __m128i cospi_p62_m02 = pair_set_epi16(cospi[62], -cospi[2]);
+  const __m128i cospi_p02_p62 = pair_set_epi16(cospi[2], cospi[62]);
+  const __m128i cospi_p30_m34 = pair_set_epi16(cospi[30], -cospi[34]);
+  const __m128i cospi_p34_p30 = pair_set_epi16(cospi[34], cospi[30]);
+  const __m128i cospi_p46_m18 = pair_set_epi16(cospi[46], -cospi[18]);
+  const __m128i cospi_p18_p46 = pair_set_epi16(cospi[18], cospi[46]);
+  const __m128i cospi_p14_m50 = pair_set_epi16(cospi[14], -cospi[50]);
+  const __m128i cospi_p50_p14 = pair_set_epi16(cospi[50], cospi[14]);
+  const __m128i cospi_p54_m10 = pair_set_epi16(cospi[54], -cospi[10]);
+  const __m128i cospi_p10_p54 = pair_set_epi16(cospi[10], cospi[54]);
+  const __m128i cospi_p22_m42 = pair_set_epi16(cospi[22], -cospi[42]);
+  const __m128i cospi_p42_p22 = pair_set_epi16(cospi[42], cospi[22]);
+  const __m128i cospi_p38_m26 = pair_set_epi16(cospi[38], -cospi[26]);
+  const __m128i cospi_p26_p38 = pair_set_epi16(cospi[26], cospi[38]);
+  const __m128i cospi_p06_m58 = pair_set_epi16(cospi[6], -cospi[58]);
+  const __m128i cospi_p58_p06 = pair_set_epi16(cospi[58], cospi[6]);
+  const __m128i cospi_p60_m04 = pair_set_epi16(cospi[60], -cospi[4]);
+  const __m128i cospi_p04_p60 = pair_set_epi16(cospi[4], cospi[60]);
+  const __m128i cospi_p28_m36 = pair_set_epi16(cospi[28], -cospi[36]);
+  const __m128i cospi_p36_p28 = pair_set_epi16(cospi[36], cospi[28]);
+  const __m128i cospi_p44_m20 = pair_set_epi16(cospi[44], -cospi[20]);
+  const __m128i cospi_p20_p44 = pair_set_epi16(cospi[20], cospi[44]);
+  const __m128i cospi_p12_m52 = pair_set_epi16(cospi[12], -cospi[52]);
+  const __m128i cospi_p52_p12 = pair_set_epi16(cospi[52], cospi[12]);
+  const __m128i cospi_p56_m08 = pair_set_epi16(cospi[56], -cospi[8]);
+  const __m128i cospi_p08_p56 = pair_set_epi16(cospi[8], cospi[56]);
+  const __m128i cospi_p24_m40 = pair_set_epi16(cospi[24], -cospi[40]);
+  const __m128i cospi_p40_p24 = pair_set_epi16(cospi[40], cospi[24]);
+  const __m128i cospi_p32_p32 = pair_set_epi16(cospi[32], cospi[32]);
+  const __m128i cospi_p32_m32 = pair_set_epi16(cospi[32], -cospi[32]);
+  const __m128i cospi_p48_m16 = pair_set_epi16(cospi[48], -cospi[16]);
+  const __m128i cospi_p16_p48 = pair_set_epi16(cospi[16], cospi[48]);
+
+  // stage 1
+  __m128i x[32];
+  x[0] = input[0];
+  x[1] = input[16];
+  x[2] = input[8];
+  x[3] = input[24];
+  x[4] = input[4];
+  x[5] = input[20];
+  x[6] = input[12];
+  x[7] = input[28];
+  x[8] = input[2];
+  x[9] = input[18];
+  x[10] = input[10];
+  x[11] = input[26];
+  x[12] = input[6];
+  x[13] = input[22];
+  x[14] = input[14];
+  x[15] = input[30];
+  x[16] = input[1];
+  x[17] = input[17];
+  x[18] = input[9];
+  x[19] = input[25];
+  x[20] = input[5];
+  x[21] = input[21];
+  x[22] = input[13];
+  x[23] = input[29];
+  x[24] = input[3];
+  x[25] = input[19];
+  x[26] = input[11];
+  x[27] = input[27];
+  x[28] = input[7];
+  x[29] = input[23];
+  x[30] = input[15];
+  x[31] = input[31];
+
+  // stage 2
+  btf_16_sse2(cospi_p62_m02, cospi_p02_p62, x[16], x[31], x[16], x[31]);
+  btf_16_sse2(cospi_p30_m34, cospi_p34_p30, x[17], x[30], x[17], x[30]);
+  btf_16_sse2(cospi_p46_m18, cospi_p18_p46, x[18], x[29], x[18], x[29]);
+  btf_16_sse2(cospi_p14_m50, cospi_p50_p14, x[19], x[28], x[19], x[28]);
+  btf_16_sse2(cospi_p54_m10, cospi_p10_p54, x[20], x[27], x[20], x[27]);
+  btf_16_sse2(cospi_p22_m42, cospi_p42_p22, x[21], x[26], x[21], x[26]);
+  btf_16_sse2(cospi_p38_m26, cospi_p26_p38, x[22], x[25], x[22], x[25]);
+  btf_16_sse2(cospi_p06_m58, cospi_p58_p06, x[23], x[24], x[23], x[24]);
+
+  // stage 3
+  btf_16_sse2(cospi_p60_m04, cospi_p04_p60, x[8], x[15], x[8], x[15]);
+  btf_16_sse2(cospi_p28_m36, cospi_p36_p28, x[9], x[14], x[9], x[14]);
+  btf_16_sse2(cospi_p44_m20, cospi_p20_p44, x[10], x[13], x[10], x[13]);
+  btf_16_sse2(cospi_p12_m52, cospi_p52_p12, x[11], x[12], x[11], x[12]);
+  idct32_high16_stage3_sse2(x);
+
+  // stage 4
+  btf_16_sse2(cospi_p56_m08, cospi_p08_p56, x[4], x[7], x[4], x[7]);
+  btf_16_sse2(cospi_p24_m40, cospi_p40_p24, x[5], x[6], x[5], x[6]);
+  btf_16_adds_subs_sse2(x[8], x[9]);
+  btf_16_subs_adds_sse2(x[11], x[10]);
+  btf_16_adds_subs_sse2(x[12], x[13]);
+  btf_16_subs_adds_sse2(x[15], x[14]);
+  idct32_high16_stage4_sse2(x, cospi, __rounding, cos_bit);
+
+  // stage 5
+  btf_16_sse2(cospi_p32_p32, cospi_p32_m32, x[0], x[1], x[0], x[1]);
+  btf_16_sse2(cospi_p48_m16, cospi_p16_p48, x[2], x[3], x[2], x[3]);
+  btf_16_adds_subs_sse2(x[4], x[5]);
+  btf_16_adds_subs_sse2(x[7], x[6]);
+  idct32_high24_stage5_sse2(x, cospi, __rounding, cos_bit);
+
+  // stage 6
+  btf_16_adds_subs_sse2(x[0], x[3]);
+  btf_16_adds_subs_sse2(x[1], x[2]);
+  idct32_high28_stage6_sse2(x, cospi, __rounding, cos_bit);
+
+  // stage 7~8
+  idct32_stage7_sse2(x, cospi, __rounding, cos_bit);
+  idct32_stage8_sse2(x, cospi, __rounding, cos_bit);
+  idct32_stage9_sse2(output, x);
+}
+
+static INLINE void idct64_stage4_high32_sse2(__m128i *x, const int32_t *cospi,
+                                             const __m128i __rounding,
+                                             int8_t cos_bit) {
+  const __m128i cospi_m04_p60 = pair_set_epi16(-cospi[4], cospi[60]);
+  const __m128i cospi_p60_p04 = pair_set_epi16(cospi[60], cospi[4]);
+  const __m128i cospi_m60_m04 = pair_set_epi16(-cospi[60], -cospi[4]);
+  const __m128i cospi_m36_p28 = pair_set_epi16(-cospi[36], cospi[28]);
+  const __m128i cospi_p28_p36 = pair_set_epi16(cospi[28], cospi[36]);
+  const __m128i cospi_m28_m36 = pair_set_epi16(-cospi[28], -cospi[36]);
+  const __m128i cospi_m20_p44 = pair_set_epi16(-cospi[20], cospi[44]);
+  const __m128i cospi_p44_p20 = pair_set_epi16(cospi[44], cospi[20]);
+  const __m128i cospi_m44_m20 = pair_set_epi16(-cospi[44], -cospi[20]);
+  const __m128i cospi_m52_p12 = pair_set_epi16(-cospi[52], cospi[12]);
+  const __m128i cospi_p12_p52 = pair_set_epi16(cospi[12], cospi[52]);
+  const __m128i cospi_m12_m52 = pair_set_epi16(-cospi[12], -cospi[52]);
+  btf_16_sse2(cospi_m04_p60, cospi_p60_p04, x[33], x[62], x[33], x[62]);
+  btf_16_sse2(cospi_m60_m04, cospi_m04_p60, x[34], x[61], x[34], x[61]);
+  btf_16_sse2(cospi_m36_p28, cospi_p28_p36, x[37], x[58], x[37], x[58]);
+  btf_16_sse2(cospi_m28_m36, cospi_m36_p28, x[38], x[57], x[38], x[57]);
+  btf_16_sse2(cospi_m20_p44, cospi_p44_p20, x[41], x[54], x[41], x[54]);
+  btf_16_sse2(cospi_m44_m20, cospi_m20_p44, x[42], x[53], x[42], x[53]);
+  btf_16_sse2(cospi_m52_p12, cospi_p12_p52, x[45], x[50], x[45], x[50]);
+  btf_16_sse2(cospi_m12_m52, cospi_m52_p12, x[46], x[49], x[46], x[49]);
+}
+
+static INLINE void idct64_stage5_high48_sse2(__m128i *x, const int32_t *cospi,
+                                             const __m128i __rounding,
+                                             int8_t cos_bit) {
+  const __m128i cospi_m08_p56 = pair_set_epi16(-cospi[8], cospi[56]);
+  const __m128i cospi_p56_p08 = pair_set_epi16(cospi[56], cospi[8]);
+  const __m128i cospi_m56_m08 = pair_set_epi16(-cospi[56], -cospi[8]);
+  const __m128i cospi_m40_p24 = pair_set_epi16(-cospi[40], cospi[24]);
+  const __m128i cospi_p24_p40 = pair_set_epi16(cospi[24], cospi[40]);
+  const __m128i cospi_m24_m40 = pair_set_epi16(-cospi[24], -cospi[40]);
+  btf_16_sse2(cospi_m08_p56, cospi_p56_p08, x[17], x[30], x[17], x[30]);
+  btf_16_sse2(cospi_m56_m08, cospi_m08_p56, x[18], x[29], x[18], x[29]);
+  btf_16_sse2(cospi_m40_p24, cospi_p24_p40, x[21], x[26], x[21], x[26]);
+  btf_16_sse2(cospi_m24_m40, cospi_m40_p24, x[22], x[25], x[22], x[25]);
+  btf_16_adds_subs_sse2(x[32], x[35]);
+  btf_16_adds_subs_sse2(x[33], x[34]);
+  btf_16_subs_adds_sse2(x[39], x[36]);
+  btf_16_subs_adds_sse2(x[38], x[37]);
+  btf_16_adds_subs_sse2(x[40], x[43]);
+  btf_16_adds_subs_sse2(x[41], x[42]);
+  btf_16_subs_adds_sse2(x[47], x[44]);
+  btf_16_subs_adds_sse2(x[46], x[45]);
+  btf_16_adds_subs_sse2(x[48], x[51]);
+  btf_16_adds_subs_sse2(x[49], x[50]);
+  btf_16_subs_adds_sse2(x[55], x[52]);
+  btf_16_subs_adds_sse2(x[54], x[53]);
+  btf_16_adds_subs_sse2(x[56], x[59]);
+  btf_16_adds_subs_sse2(x[57], x[58]);
+  btf_16_subs_adds_sse2(x[63], x[60]);
+  btf_16_subs_adds_sse2(x[62], x[61]);
+}
+
+static INLINE void idct64_stage6_high32_sse2(__m128i *x, const int32_t *cospi,
+                                             const __m128i __rounding,
+                                             int8_t cos_bit) {
+  const __m128i cospi_m08_p56 = pair_set_epi16(-cospi[8], cospi[56]);
+  const __m128i cospi_p56_p08 = pair_set_epi16(cospi[56], cospi[8]);
+  const __m128i cospi_m56_m08 = pair_set_epi16(-cospi[56], -cospi[8]);
+  const __m128i cospi_m40_p24 = pair_set_epi16(-cospi[40], cospi[24]);
+  const __m128i cospi_p24_p40 = pair_set_epi16(cospi[24], cospi[40]);
+  const __m128i cospi_m24_m40 = pair_set_epi16(-cospi[24], -cospi[40]);
+  btf_16_sse2(cospi_m08_p56, cospi_p56_p08, x[34], x[61], x[34], x[61]);
+  btf_16_sse2(cospi_m08_p56, cospi_p56_p08, x[35], x[60], x[35], x[60]);
+  btf_16_sse2(cospi_m56_m08, cospi_m08_p56, x[36], x[59], x[36], x[59]);
+  btf_16_sse2(cospi_m56_m08, cospi_m08_p56, x[37], x[58], x[37], x[58]);
+  btf_16_sse2(cospi_m40_p24, cospi_p24_p40, x[42], x[53], x[42], x[53]);
+  btf_16_sse2(cospi_m40_p24, cospi_p24_p40, x[43], x[52], x[43], x[52]);
+  btf_16_sse2(cospi_m24_m40, cospi_m40_p24, x[44], x[51], x[44], x[51]);
+  btf_16_sse2(cospi_m24_m40, cospi_m40_p24, x[45], x[50], x[45], x[50]);
+}
+
+static INLINE void idct64_stage6_high48_sse2(__m128i *x, const int32_t *cospi,
+                                             const __m128i __rounding,
+                                             int8_t cos_bit) {
+  btf_16_adds_subs_sse2(x[16], x[19]);
+  btf_16_adds_subs_sse2(x[17], x[18]);
+  btf_16_subs_adds_sse2(x[23], x[20]);
+  btf_16_subs_adds_sse2(x[22], x[21]);
+  btf_16_adds_subs_sse2(x[24], x[27]);
+  btf_16_adds_subs_sse2(x[25], x[26]);
+  btf_16_subs_adds_sse2(x[31], x[28]);
+  btf_16_subs_adds_sse2(x[30], x[29]);
+  idct64_stage6_high32_sse2(x, cospi, __rounding, cos_bit);
+}
+
+static INLINE void idct64_stage7_high48_sse2(__m128i *x, const int32_t *cospi,
+                                             const __m128i __rounding,
+                                             int8_t cos_bit) {
+  const __m128i cospi_m16_p48 = pair_set_epi16(-cospi[16], cospi[48]);
+  const __m128i cospi_p48_p16 = pair_set_epi16(cospi[48], cospi[16]);
+  const __m128i cospi_m48_m16 = pair_set_epi16(-cospi[48], -cospi[16]);
+  btf_16_sse2(cospi_m16_p48, cospi_p48_p16, x[18], x[29], x[18], x[29]);
+  btf_16_sse2(cospi_m16_p48, cospi_p48_p16, x[19], x[28], x[19], x[28]);
+  btf_16_sse2(cospi_m48_m16, cospi_m16_p48, x[20], x[27], x[20], x[27]);
+  btf_16_sse2(cospi_m48_m16, cospi_m16_p48, x[21], x[26], x[21], x[26]);
+  btf_16_adds_subs_sse2(x[32], x[39]);
+  btf_16_adds_subs_sse2(x[33], x[38]);
+  btf_16_adds_subs_sse2(x[34], x[37]);
+  btf_16_adds_subs_sse2(x[35], x[36]);
+  btf_16_subs_adds_sse2(x[47], x[40]);
+  btf_16_subs_adds_sse2(x[46], x[41]);
+  btf_16_subs_adds_sse2(x[45], x[42]);
+  btf_16_subs_adds_sse2(x[44], x[43]);
+  btf_16_adds_subs_sse2(x[48], x[55]);
+  btf_16_adds_subs_sse2(x[49], x[54]);
+  btf_16_adds_subs_sse2(x[50], x[53]);
+  btf_16_adds_subs_sse2(x[51], x[52]);
+  btf_16_subs_adds_sse2(x[63], x[56]);
+  btf_16_subs_adds_sse2(x[62], x[57]);
+  btf_16_subs_adds_sse2(x[61], x[58]);
+  btf_16_subs_adds_sse2(x[60], x[59]);
+}
+
+static INLINE void idct64_stage8_high48_sse2(__m128i *x, const int32_t *cospi,
+                                             const __m128i __rounding,
+                                             int8_t cos_bit) {
+  const __m128i cospi_m16_p48 = pair_set_epi16(-cospi[16], cospi[48]);
+  const __m128i cospi_p48_p16 = pair_set_epi16(cospi[48], cospi[16]);
+  const __m128i cospi_m48_m16 = pair_set_epi16(-cospi[48], -cospi[16]);
+  btf_16_adds_subs_sse2(x[16], x[23]);
+  btf_16_adds_subs_sse2(x[17], x[22]);
+  btf_16_adds_subs_sse2(x[18], x[21]);
+  btf_16_adds_subs_sse2(x[19], x[20]);
+  btf_16_subs_adds_sse2(x[31], x[24]);
+  btf_16_subs_adds_sse2(x[30], x[25]);
+  btf_16_subs_adds_sse2(x[29], x[26]);
+  btf_16_subs_adds_sse2(x[28], x[27]);
+  btf_16_sse2(cospi_m16_p48, cospi_p48_p16, x[36], x[59], x[36], x[59]);
+  btf_16_sse2(cospi_m16_p48, cospi_p48_p16, x[37], x[58], x[37], x[58]);
+  btf_16_sse2(cospi_m16_p48, cospi_p48_p16, x[38], x[57], x[38], x[57]);
+  btf_16_sse2(cospi_m16_p48, cospi_p48_p16, x[39], x[56], x[39], x[56]);
+  btf_16_sse2(cospi_m48_m16, cospi_m16_p48, x[40], x[55], x[40], x[55]);
+  btf_16_sse2(cospi_m48_m16, cospi_m16_p48, x[41], x[54], x[41], x[54]);
+  btf_16_sse2(cospi_m48_m16, cospi_m16_p48, x[42], x[53], x[42], x[53]);
+  btf_16_sse2(cospi_m48_m16, cospi_m16_p48, x[43], x[52], x[43], x[52]);
+}
+
+static INLINE void idct64_stage9_sse2(__m128i *x, const int32_t *cospi,
+                                      const __m128i __rounding,
+                                      int8_t cos_bit) {
+  const __m128i cospi_m32_p32 = pair_set_epi16(-cospi[32], cospi[32]);
+  const __m128i cospi_p32_p32 = pair_set_epi16(cospi[32], cospi[32]);
+  btf_16_adds_subs_sse2(x[0], x[15]);
+  btf_16_adds_subs_sse2(x[1], x[14]);
+  btf_16_adds_subs_sse2(x[2], x[13]);
+  btf_16_adds_subs_sse2(x[3], x[12]);
+  btf_16_adds_subs_sse2(x[4], x[11]);
+  btf_16_adds_subs_sse2(x[5], x[10]);
+  btf_16_adds_subs_sse2(x[6], x[9]);
+  btf_16_adds_subs_sse2(x[7], x[8]);
+  btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x[20], x[27], x[20], x[27]);
+  btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x[21], x[26], x[21], x[26]);
+  btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x[22], x[25], x[22], x[25]);
+  btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x[23], x[24], x[23], x[24]);
+  btf_16_adds_subs_sse2(x[32], x[47]);
+  btf_16_adds_subs_sse2(x[33], x[46]);
+  btf_16_adds_subs_sse2(x[34], x[45]);
+  btf_16_adds_subs_sse2(x[35], x[44]);
+  btf_16_adds_subs_sse2(x[36], x[43]);
+  btf_16_adds_subs_sse2(x[37], x[42]);
+  btf_16_adds_subs_sse2(x[38], x[41]);
+  btf_16_adds_subs_sse2(x[39], x[40]);
+  btf_16_subs_adds_sse2(x[63], x[48]);
+  btf_16_subs_adds_sse2(x[62], x[49]);
+  btf_16_subs_adds_sse2(x[61], x[50]);
+  btf_16_subs_adds_sse2(x[60], x[51]);
+  btf_16_subs_adds_sse2(x[59], x[52]);
+  btf_16_subs_adds_sse2(x[58], x[53]);
+  btf_16_subs_adds_sse2(x[57], x[54]);
+  btf_16_subs_adds_sse2(x[56], x[55]);
+}
+
+static INLINE void idct64_stage10_sse2(__m128i *x, const int32_t *cospi,
+                                       const __m128i __rounding,
+                                       int8_t cos_bit) {
+  const __m128i cospi_m32_p32 = pair_set_epi16(-cospi[32], cospi[32]);
+  const __m128i cospi_p32_p32 = pair_set_epi16(cospi[32], cospi[32]);
+  btf_16_adds_subs_sse2(x[0], x[31]);
+  btf_16_adds_subs_sse2(x[1], x[30]);
+  btf_16_adds_subs_sse2(x[2], x[29]);
+  btf_16_adds_subs_sse2(x[3], x[28]);
+  btf_16_adds_subs_sse2(x[4], x[27]);
+  btf_16_adds_subs_sse2(x[5], x[26]);
+  btf_16_adds_subs_sse2(x[6], x[25]);
+  btf_16_adds_subs_sse2(x[7], x[24]);
+  btf_16_adds_subs_sse2(x[8], x[23]);
+  btf_16_adds_subs_sse2(x[9], x[22]);
+  btf_16_adds_subs_sse2(x[10], x[21]);
+  btf_16_adds_subs_sse2(x[11], x[20]);
+  btf_16_adds_subs_sse2(x[12], x[19]);
+  btf_16_adds_subs_sse2(x[13], x[18]);
+  btf_16_adds_subs_sse2(x[14], x[17]);
+  btf_16_adds_subs_sse2(x[15], x[16]);
+  btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x[40], x[55], x[40], x[55]);
+  btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x[41], x[54], x[41], x[54]);
+  btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x[42], x[53], x[42], x[53]);
+  btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x[43], x[52], x[43], x[52]);
+  btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x[44], x[51], x[44], x[51]);
+  btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x[45], x[50], x[45], x[50]);
+  btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x[46], x[49], x[46], x[49]);
+  btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x[47], x[48], x[47], x[48]);
+}
+
+static INLINE void idct64_stage11_sse2(__m128i *output, __m128i *x) {
+  btf_16_adds_subs_out_sse2(output[0], output[63], x[0], x[63]);
+  btf_16_adds_subs_out_sse2(output[1], output[62], x[1], x[62]);
+  btf_16_adds_subs_out_sse2(output[2], output[61], x[2], x[61]);
+  btf_16_adds_subs_out_sse2(output[3], output[60], x[3], x[60]);
+  btf_16_adds_subs_out_sse2(output[4], output[59], x[4], x[59]);
+  btf_16_adds_subs_out_sse2(output[5], output[58], x[5], x[58]);
+  btf_16_adds_subs_out_sse2(output[6], output[57], x[6], x[57]);
+  btf_16_adds_subs_out_sse2(output[7], output[56], x[7], x[56]);
+  btf_16_adds_subs_out_sse2(output[8], output[55], x[8], x[55]);
+  btf_16_adds_subs_out_sse2(output[9], output[54], x[9], x[54]);
+  btf_16_adds_subs_out_sse2(output[10], output[53], x[10], x[53]);
+  btf_16_adds_subs_out_sse2(output[11], output[52], x[11], x[52]);
+  btf_16_adds_subs_out_sse2(output[12], output[51], x[12], x[51]);
+  btf_16_adds_subs_out_sse2(output[13], output[50], x[13], x[50]);
+  btf_16_adds_subs_out_sse2(output[14], output[49], x[14], x[49]);
+  btf_16_adds_subs_out_sse2(output[15], output[48], x[15], x[48]);
+  btf_16_adds_subs_out_sse2(output[16], output[47], x[16], x[47]);
+  btf_16_adds_subs_out_sse2(output[17], output[46], x[17], x[46]);
+  btf_16_adds_subs_out_sse2(output[18], output[45], x[18], x[45]);
+  btf_16_adds_subs_out_sse2(output[19], output[44], x[19], x[44]);
+  btf_16_adds_subs_out_sse2(output[20], output[43], x[20], x[43]);
+  btf_16_adds_subs_out_sse2(output[21], output[42], x[21], x[42]);
+  btf_16_adds_subs_out_sse2(output[22], output[41], x[22], x[41]);
+  btf_16_adds_subs_out_sse2(output[23], output[40], x[23], x[40]);
+  btf_16_adds_subs_out_sse2(output[24], output[39], x[24], x[39]);
+  btf_16_adds_subs_out_sse2(output[25], output[38], x[25], x[38]);
+  btf_16_adds_subs_out_sse2(output[26], output[37], x[26], x[37]);
+  btf_16_adds_subs_out_sse2(output[27], output[36], x[27], x[36]);
+  btf_16_adds_subs_out_sse2(output[28], output[35], x[28], x[35]);
+  btf_16_adds_subs_out_sse2(output[29], output[34], x[29], x[34]);
+  btf_16_adds_subs_out_sse2(output[30], output[33], x[30], x[33]);
+  btf_16_adds_subs_out_sse2(output[31], output[32], x[31], x[32]);
+}
+
+static void idct64_low1_new_ssse3(const __m128i *input, __m128i *output,
+                                  int8_t cos_bit) {
+  (void)cos_bit;
+  const int32_t *cospi = cospi_arr(INV_COS_BIT);
+
+  // stage 1
+  __m128i x[32];
+  x[0] = input[0];
+
+  // stage 2
+  // stage 3
+  // stage 4
+  // stage 5
+  // stage 6
+  btf_16_ssse3(cospi[32], cospi[32], x[0], x[0], x[1]);
+
+  // stage 7
+  // stage 8
+  // stage 9
+  // stage 10
+  // stage 11
+  output[0] = x[0];
+  output[63] = x[0];
+  output[1] = x[1];
+  output[62] = x[1];
+  output[2] = x[1];
+  output[61] = x[1];
+  output[3] = x[0];
+  output[60] = x[0];
+  output[4] = x[0];
+  output[59] = x[0];
+  output[5] = x[1];
+  output[58] = x[1];
+  output[6] = x[1];
+  output[57] = x[1];
+  output[7] = x[0];
+  output[56] = x[0];
+  output[8] = x[0];
+  output[55] = x[0];
+  output[9] = x[1];
+  output[54] = x[1];
+  output[10] = x[1];
+  output[53] = x[1];
+  output[11] = x[0];
+  output[52] = x[0];
+  output[12] = x[0];
+  output[51] = x[0];
+  output[13] = x[1];
+  output[50] = x[1];
+  output[14] = x[1];
+  output[49] = x[1];
+  output[15] = x[0];
+  output[48] = x[0];
+  output[16] = x[0];
+  output[47] = x[0];
+  output[17] = x[1];
+  output[46] = x[1];
+  output[18] = x[1];
+  output[45] = x[1];
+  output[19] = x[0];
+  output[44] = x[0];
+  output[20] = x[0];
+  output[43] = x[0];
+  output[21] = x[1];
+  output[42] = x[1];
+  output[22] = x[1];
+  output[41] = x[1];
+  output[23] = x[0];
+  output[40] = x[0];
+  output[24] = x[0];
+  output[39] = x[0];
+  output[25] = x[1];
+  output[38] = x[1];
+  output[26] = x[1];
+  output[37] = x[1];
+  output[27] = x[0];
+  output[36] = x[0];
+  output[28] = x[0];
+  output[35] = x[0];
+  output[29] = x[1];
+  output[34] = x[1];
+  output[30] = x[1];
+  output[33] = x[1];
+  output[31] = x[0];
+  output[32] = x[0];
+}
+
+static void idct64_low8_new_ssse3(const __m128i *input, __m128i *output,
+                                  int8_t cos_bit) {
+  (void)cos_bit;
+  const int32_t *cospi = cospi_arr(INV_COS_BIT);
+  const __m128i __rounding = _mm_set1_epi32(1 << (INV_COS_BIT - 1));
+  const __m128i cospi_m04_p60 = pair_set_epi16(-cospi[4], cospi[60]);
+  const __m128i cospi_p60_p04 = pair_set_epi16(cospi[60], cospi[4]);
+  const __m128i cospi_m36_p28 = pair_set_epi16(-cospi[36], cospi[28]);
+  const __m128i cospi_m28_m36 = pair_set_epi16(-cospi[28], -cospi[36]);
+  const __m128i cospi_m20_p44 = pair_set_epi16(-cospi[20], cospi[44]);
+  const __m128i cospi_p44_p20 = pair_set_epi16(cospi[44], cospi[20]);
+  const __m128i cospi_m52_p12 = pair_set_epi16(-cospi[52], cospi[12]);
+  const __m128i cospi_m12_m52 = pair_set_epi16(-cospi[12], -cospi[52]);
+  const __m128i cospi_m08_p56 = pair_set_epi16(-cospi[8], cospi[56]);
+  const __m128i cospi_p56_p08 = pair_set_epi16(cospi[56], cospi[8]);
+  const __m128i cospi_m40_p24 = pair_set_epi16(-cospi[40], cospi[24]);
+  const __m128i cospi_m24_m40 = pair_set_epi16(-cospi[24], -cospi[40]);
+  const __m128i cospi_p32_p32 = pair_set_epi16(cospi[32], cospi[32]);
+  const __m128i cospi_m16_p48 = pair_set_epi16(-cospi[16], cospi[48]);
+  const __m128i cospi_p48_p16 = pair_set_epi16(cospi[48], cospi[16]);
+  const __m128i cospi_m32_p32 = pair_set_epi16(-cospi[32], cospi[32]);
+
+  // stage 1
+  __m128i x[64];
+  x[0] = input[0];
+  x[8] = input[4];
+  x[16] = input[2];
+  x[24] = input[6];
+  x[32] = input[1];
+  x[40] = input[5];
+  x[48] = input[3];
+  x[56] = input[7];
+
+  // stage 2
+  btf_16_ssse3(cospi[63], cospi[1], x[32], x[32], x[63]);
+  btf_16_ssse3(-cospi[57], cospi[7], x[56], x[39], x[56]);
+  btf_16_ssse3(cospi[59], cospi[5], x[40], x[40], x[55]);
+  btf_16_ssse3(-cospi[61], cospi[3], x[48], x[47], x[48]);
+
+  // stage 3
+  btf_16_ssse3(cospi[62], cospi[2], x[16], x[16], x[31]);
+  btf_16_ssse3(-cospi[58], cospi[6], x[24], x[23], x[24]);
+  x[33] = x[32];
+  x[38] = x[39];
+  x[41] = x[40];
+  x[46] = x[47];
+  x[49] = x[48];
+  x[54] = x[55];
+  x[57] = x[56];
+  x[62] = x[63];
+
+  // stage 4
+  btf_16_ssse3(cospi[60], cospi[4], x[8], x[8], x[15]);
+  x[17] = x[16];
+  x[22] = x[23];
+  x[25] = x[24];
+  x[30] = x[31];
+  btf_16_sse2(cospi_m04_p60, cospi_p60_p04, x[33], x[62], x[33], x[62]);
+  btf_16_sse2(cospi_m28_m36, cospi_m36_p28, x[38], x[57], x[38], x[57]);
+  btf_16_sse2(cospi_m20_p44, cospi_p44_p20, x[41], x[54], x[41], x[54]);
+  btf_16_sse2(cospi_m12_m52, cospi_m52_p12, x[46], x[49], x[46], x[49]);
+
+  // stage 5
+  x[9] = x[8];
+  x[14] = x[15];
+  btf_16_sse2(cospi_m08_p56, cospi_p56_p08, x[17], x[30], x[17], x[30]);
+  btf_16_sse2(cospi_m24_m40, cospi_m40_p24, x[22], x[25], x[22], x[25]);
+  x[35] = x[32];
+  x[34] = x[33];
+  x[36] = x[39];
+  x[37] = x[38];
+  x[43] = x[40];
+  x[42] = x[41];
+  x[44] = x[47];
+  x[45] = x[46];
+  x[51] = x[48];
+  x[50] = x[49];
+  x[52] = x[55];
+  x[53] = x[54];
+  x[59] = x[56];
+  x[58] = x[57];
+  x[60] = x[63];
+  x[61] = x[62];
+
+  // stage 6
+  btf_16_ssse3(cospi[32], cospi[32], x[0], x[0], x[1]);
+  btf_16_sse2(cospi_m16_p48, cospi_p48_p16, x[9], x[14], x[9], x[14]);
+  x[19] = x[16];
+  x[18] = x[17];
+  x[20] = x[23];
+  x[21] = x[22];
+  x[27] = x[24];
+  x[26] = x[25];
+  x[28] = x[31];
+  x[29] = x[30];
+  idct64_stage6_high32_sse2(x, cospi, __rounding, cos_bit);
+
+  // stage 7
+  x[3] = x[0];
+  x[2] = x[1];
+  x[11] = x[8];
+  x[10] = x[9];
+  x[12] = x[15];
+  x[13] = x[14];
+  idct64_stage7_high48_sse2(x, cospi, __rounding, cos_bit);
+
+  // stage 8
+  x[7] = x[0];
+  x[6] = x[1];
+  x[5] = x[2];
+  x[4] = x[3];
+  x[9] = x[9];
+  btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x[10], x[13], x[10], x[13]);
+  btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x[11], x[12], x[11], x[12]);
+  idct64_stage8_high48_sse2(x, cospi, __rounding, cos_bit);
+
+  idct64_stage9_sse2(x, cospi, __rounding, cos_bit);
+  idct64_stage10_sse2(x, cospi, __rounding, cos_bit);
+  idct64_stage11_sse2(output, x);
+}
+
+static void idct64_low16_new_ssse3(const __m128i *input, __m128i *output,
+                                   int8_t cos_bit) {
+  (void)cos_bit;
+  const int32_t *cospi = cospi_arr(INV_COS_BIT);
+  const __m128i __rounding = _mm_set1_epi32(1 << (INV_COS_BIT - 1));
+
+  const __m128i cospi_p32_p32 = pair_set_epi16(cospi[32], cospi[32]);
+  const __m128i cospi_m16_p48 = pair_set_epi16(-cospi[16], cospi[48]);
+  const __m128i cospi_p48_p16 = pair_set_epi16(cospi[48], cospi[16]);
+  const __m128i cospi_m48_m16 = pair_set_epi16(-cospi[48], -cospi[16]);
+  const __m128i cospi_m32_p32 = pair_set_epi16(-cospi[32], cospi[32]);
+
+  // stage 1
+  __m128i x[64];
+  x[0] = input[0];
+  x[4] = input[8];
+  x[8] = input[4];
+  x[12] = input[12];
+  x[16] = input[2];
+  x[20] = input[10];
+  x[24] = input[6];
+  x[28] = input[14];
+  x[32] = input[1];
+  x[36] = input[9];
+  x[40] = input[5];
+  x[44] = input[13];
+  x[48] = input[3];
+  x[52] = input[11];
+  x[56] = input[7];
+  x[60] = input[15];
+
+  // stage 2
+  btf_16_ssse3(cospi[63], cospi[1], x[32], x[32], x[63]);
+  btf_16_ssse3(-cospi[49], cospi[15], x[60], x[35], x[60]);
+  btf_16_ssse3(cospi[55], cospi[9], x[36], x[36], x[59]);
+  btf_16_ssse3(-cospi[57], cospi[7], x[56], x[39], x[56]);
+  btf_16_ssse3(cospi[59], cospi[5], x[40], x[40], x[55]);
+  btf_16_ssse3(-cospi[53], cospi[11], x[52], x[43], x[52]);
+  btf_16_ssse3(cospi[51], cospi[13], x[44], x[44], x[51]);
+  btf_16_ssse3(-cospi[61], cospi[3], x[48], x[47], x[48]);
+
+  // stage 3
+  btf_16_ssse3(cospi[62], cospi[2], x[16], x[16], x[31]);
+  btf_16_ssse3(-cospi[50], cospi[14], x[28], x[19], x[28]);
+  btf_16_ssse3(cospi[54], cospi[10], x[20], x[20], x[27]);
+  btf_16_ssse3(-cospi[58], cospi[6], x[24], x[23], x[24]);
+  x[33] = x[32];
+  x[34] = x[35];
+  x[37] = x[36];
+  x[38] = x[39];
+  x[41] = x[40];
+  x[42] = x[43];
+  x[45] = x[44];
+  x[46] = x[47];
+  x[49] = x[48];
+  x[50] = x[51];
+  x[53] = x[52];
+  x[54] = x[55];
+  x[57] = x[56];
+  x[58] = x[59];
+  x[61] = x[60];
+  x[62] = x[63];
+
+  // stage 4
+  btf_16_ssse3(cospi[60], cospi[4], x[8], x[8], x[15]);
+  btf_16_ssse3(-cospi[52], cospi[12], x[12], x[11], x[12]);
+  x[17] = x[16];
+  x[18] = x[19];
+  x[21] = x[20];
+  x[22] = x[23];
+  x[25] = x[24];
+  x[26] = x[27];
+  x[29] = x[28];
+  x[30] = x[31];
+  idct64_stage4_high32_sse2(x, cospi, __rounding, cos_bit);
+
+  // stage 5
+  btf_16_ssse3(cospi[56], cospi[8], x[4], x[4], x[7]);
+  x[9] = x[8];
+  x[10] = x[11];
+  x[13] = x[12];
+  x[14] = x[15];
+  idct64_stage5_high48_sse2(x, cospi, __rounding, cos_bit);
+
+  // stage 6
+  btf_16_ssse3(cospi[32], cospi[32], x[0], x[0], x[1]);
+  x[5] = x[4];
+  x[6] = x[7];
+  btf_16_sse2(cospi_m16_p48, cospi_p48_p16, x[9], x[14], x[9], x[14]);
+  btf_16_sse2(cospi_m48_m16, cospi_m16_p48, x[10], x[13], x[10], x[13]);
+  idct64_stage6_high48_sse2(x, cospi, __rounding, cos_bit);
+
+  // stage 7
+  x[3] = x[0];
+  x[2] = x[1];
+  btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x[5], x[6], x[5], x[6]);
+  btf_16_adds_subs_sse2(x[8], x[11]);
+  btf_16_adds_subs_sse2(x[9], x[10]);
+  btf_16_subs_adds_sse2(x[15], x[12]);
+  btf_16_subs_adds_sse2(x[14], x[13]);
+  idct64_stage7_high48_sse2(x, cospi, __rounding, cos_bit);
+
+  // stage 8
+  btf_16_adds_subs_sse2(x[0], x[7]);
+  btf_16_adds_subs_sse2(x[1], x[6]);
+  btf_16_adds_subs_sse2(x[2], x[5]);
+  btf_16_adds_subs_sse2(x[3], x[4]);
+  btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x[10], x[13], x[10], x[13]);
+  btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x[11], x[12], x[11], x[12]);
+  idct64_stage8_high48_sse2(x, cospi, __rounding, cos_bit);
+
+  idct64_stage9_sse2(x, cospi, __rounding, cos_bit);
+  idct64_stage10_sse2(x, cospi, __rounding, cos_bit);
+  idct64_stage11_sse2(output, x);
+}
+
+static void idct64_low32_new_ssse3(const __m128i *input, __m128i *output,
+                                   int8_t cos_bit) {
+  (void)cos_bit;
+  const int32_t *cospi = cospi_arr(INV_COS_BIT);
+  const __m128i __rounding = _mm_set1_epi32(1 << (INV_COS_BIT - 1));
+
+  const __m128i cospi_p32_p32 = pair_set_epi16(cospi[32], cospi[32]);
+  const __m128i cospi_m16_p48 = pair_set_epi16(-cospi[16], cospi[48]);
+  const __m128i cospi_p48_p16 = pair_set_epi16(cospi[48], cospi[16]);
+  const __m128i cospi_m48_m16 = pair_set_epi16(-cospi[48], -cospi[16]);
+  const __m128i cospi_m32_p32 = pair_set_epi16(-cospi[32], cospi[32]);
+
+  // stage 1
+  __m128i x[64];
+  x[0] = input[0];
+  x[2] = input[16];
+  x[4] = input[8];
+  x[6] = input[24];
+  x[8] = input[4];
+  x[10] = input[20];
+  x[12] = input[12];
+  x[14] = input[28];
+  x[16] = input[2];
+  x[18] = input[18];
+  x[20] = input[10];
+  x[22] = input[26];
+  x[24] = input[6];
+  x[26] = input[22];
+  x[28] = input[14];
+  x[30] = input[30];
+  x[32] = input[1];
+  x[34] = input[17];
+  x[36] = input[9];
+  x[38] = input[25];
+  x[40] = input[5];
+  x[42] = input[21];
+  x[44] = input[13];
+  x[46] = input[29];
+  x[48] = input[3];
+  x[50] = input[19];
+  x[52] = input[11];
+  x[54] = input[27];
+  x[56] = input[7];
+  x[58] = input[23];
+  x[60] = input[15];
+  x[62] = input[31];
+
+  // stage 2
+  btf_16_ssse3(cospi[63], cospi[1], x[32], x[32], x[63]);
+  btf_16_ssse3(-cospi[33], cospi[31], x[62], x[33], x[62]);
+  btf_16_ssse3(cospi[47], cospi[17], x[34], x[34], x[61]);
+  btf_16_ssse3(-cospi[49], cospi[15], x[60], x[35], x[60]);
+  btf_16_ssse3(cospi[55], cospi[9], x[36], x[36], x[59]);
+  btf_16_ssse3(-cospi[41], cospi[23], x[58], x[37], x[58]);
+  btf_16_ssse3(cospi[39], cospi[25], x[38], x[38], x[57]);
+  btf_16_ssse3(-cospi[57], cospi[7], x[56], x[39], x[56]);
+  btf_16_ssse3(cospi[59], cospi[5], x[40], x[40], x[55]);
+  btf_16_ssse3(-cospi[37], cospi[27], x[54], x[41], x[54]);
+  btf_16_ssse3(cospi[43], cospi[21], x[42], x[42], x[53]);
+  btf_16_ssse3(-cospi[53], cospi[11], x[52], x[43], x[52]);
+  btf_16_ssse3(cospi[51], cospi[13], x[44], x[44], x[51]);
+  btf_16_ssse3(-cospi[45], cospi[19], x[50], x[45], x[50]);
+  btf_16_ssse3(cospi[35], cospi[29], x[46], x[46], x[49]);
+  btf_16_ssse3(-cospi[61], cospi[3], x[48], x[47], x[48]);
+
+  // stage 3
+  btf_16_ssse3(cospi[62], cospi[2], x[16], x[16], x[31]);
+  btf_16_ssse3(-cospi[34], cospi[30], x[30], x[17], x[30]);
+  btf_16_ssse3(cospi[46], cospi[18], x[18], x[18], x[29]);
+  btf_16_ssse3(-cospi[50], cospi[14], x[28], x[19], x[28]);
+  btf_16_ssse3(cospi[54], cospi[10], x[20], x[20], x[27]);
+  btf_16_ssse3(-cospi[42], cospi[22], x[26], x[21], x[26]);
+  btf_16_ssse3(cospi[38], cospi[26], x[22], x[22], x[25]);
+  btf_16_ssse3(-cospi[58], cospi[6], x[24], x[23], x[24]);
+  btf_16_adds_subs_sse2(x[32], x[33]);
+  btf_16_subs_adds_sse2(x[35], x[34]);
+  btf_16_adds_subs_sse2(x[36], x[37]);
+  btf_16_subs_adds_sse2(x[39], x[38]);
+  btf_16_adds_subs_sse2(x[40], x[41]);
+  btf_16_subs_adds_sse2(x[43], x[42]);
+  btf_16_adds_subs_sse2(x[44], x[45]);
+  btf_16_subs_adds_sse2(x[47], x[46]);
+  btf_16_adds_subs_sse2(x[48], x[49]);
+  btf_16_subs_adds_sse2(x[51], x[50]);
+  btf_16_adds_subs_sse2(x[52], x[53]);
+  btf_16_subs_adds_sse2(x[55], x[54]);
+  btf_16_adds_subs_sse2(x[56], x[57]);
+  btf_16_subs_adds_sse2(x[59], x[58]);
+  btf_16_adds_subs_sse2(x[60], x[61]);
+  btf_16_subs_adds_sse2(x[63], x[62]);
+
+  // stage 4
+  btf_16_ssse3(cospi[60], cospi[4], x[8], x[8], x[15]);
+  btf_16_ssse3(-cospi[36], cospi[28], x[14], x[9], x[14]);
+  btf_16_ssse3(cospi[44], cospi[20], x[10], x[10], x[13]);
+  btf_16_ssse3(-cospi[52], cospi[12], x[12], x[11], x[12]);
+  btf_16_adds_subs_sse2(x[16], x[17]);
+  btf_16_subs_adds_sse2(x[19], x[18]);
+  btf_16_adds_subs_sse2(x[20], x[21]);
+  btf_16_subs_adds_sse2(x[23], x[22]);
+  btf_16_adds_subs_sse2(x[24], x[25]);
+  btf_16_subs_adds_sse2(x[27], x[26]);
+  btf_16_adds_subs_sse2(x[28], x[29]);
+  btf_16_subs_adds_sse2(x[31], x[30]);
+  idct64_stage4_high32_sse2(x, cospi, __rounding, cos_bit);
+
+  // stage 5
+  btf_16_ssse3(cospi[56], cospi[8], x[4], x[4], x[7]);
+  btf_16_ssse3(-cospi[40], cospi[24], x[6], x[5], x[6]);
+  btf_16_adds_subs_sse2(x[8], x[9]);
+  btf_16_subs_adds_sse2(x[11], x[10]);
+  btf_16_adds_subs_sse2(x[12], x[13]);
+  btf_16_subs_adds_sse2(x[15], x[14]);
+  idct64_stage5_high48_sse2(x, cospi, __rounding, cos_bit);
+
+  // stage 6
+  btf_16_ssse3(cospi[32], cospi[32], x[0], x[0], x[1]);
+  btf_16_ssse3(cospi[48], cospi[16], x[2], x[2], x[3]);
+  btf_16_adds_subs_sse2(x[4], x[5]);
+  btf_16_subs_adds_sse2(x[7], x[6]);
+  btf_16_sse2(cospi_m16_p48, cospi_p48_p16, x[9], x[14], x[9], x[14]);
+  btf_16_sse2(cospi_m48_m16, cospi_m16_p48, x[10], x[13], x[10], x[13]);
+  idct64_stage6_high48_sse2(x, cospi, __rounding, cos_bit);
+
+  // stage 7
+  btf_16_adds_subs_sse2(x[0], x[3]);
+  btf_16_adds_subs_sse2(x[1], x[2]);
+  btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x[5], x[6], x[5], x[6]);
+  btf_16_adds_subs_sse2(x[8], x[11]);
+  btf_16_adds_subs_sse2(x[9], x[10]);
+  btf_16_subs_adds_sse2(x[15], x[12]);
+  btf_16_subs_adds_sse2(x[14], x[13]);
+  idct64_stage7_high48_sse2(x, cospi, __rounding, cos_bit);
+
+  // stage 8
+  btf_16_adds_subs_sse2(x[0], x[7]);
+  btf_16_adds_subs_sse2(x[1], x[6]);
+  btf_16_adds_subs_sse2(x[2], x[5]);
+  btf_16_adds_subs_sse2(x[3], x[4]);
+  btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x[10], x[13], x[10], x[13]);
+  btf_16_sse2(cospi_m32_p32, cospi_p32_p32, x[11], x[12], x[11], x[12]);
+  idct64_stage8_high48_sse2(x, cospi, __rounding, cos_bit);
+
+  // stage 9~11
+  idct64_stage9_sse2(x, cospi, __rounding, cos_bit);
+  idct64_stage10_sse2(x, cospi, __rounding, cos_bit);
+  idct64_stage11_sse2(output, x);
+}
+
+void iadst4_new_sse2(const __m128i *input, __m128i *output, int8_t cos_bit) {
+  (void)cos_bit;
+  const int32_t *sinpi = sinpi_arr(INV_COS_BIT);
+  const __m128i sinpi_p01_p04 = pair_set_epi16(sinpi[1], sinpi[4]);
+  const __m128i sinpi_p02_m01 = pair_set_epi16(sinpi[2], -sinpi[1]);
+  const __m128i sinpi_p03_p02 = pair_set_epi16(sinpi[3], sinpi[2]);
+  const __m128i sinpi_p03_m04 = pair_set_epi16(sinpi[3], -sinpi[4]);
+  const __m128i sinpi_p03_m03 = pair_set_epi16(sinpi[3], -sinpi[3]);
+  const __m128i sinpi_0_p03 = pair_set_epi16(0, sinpi[3]);
+  const __m128i sinpi_p04_p02 = pair_set_epi16(sinpi[4], sinpi[2]);
+  const __m128i sinpi_m03_m01 = pair_set_epi16(-sinpi[3], -sinpi[1]);
+  __m128i x0[4];
+  x0[0] = input[0];
+  x0[1] = input[1];
+  x0[2] = input[2];
+  x0[3] = input[3];
+
+  __m128i u[4];
+  u[0] = _mm_unpacklo_epi16(x0[0], x0[2]);
+  u[1] = _mm_unpackhi_epi16(x0[0], x0[2]);
+  u[2] = _mm_unpacklo_epi16(x0[1], x0[3]);
+  u[3] = _mm_unpackhi_epi16(x0[1], x0[3]);
+
+  __m128i x1[16];
+  x1[0] = _mm_madd_epi16(u[0], sinpi_p01_p04);  // x0*sin1 + x2*sin4
+  x1[1] = _mm_madd_epi16(u[1], sinpi_p01_p04);
+  x1[2] = _mm_madd_epi16(u[0], sinpi_p02_m01);  // x0*sin2 - x2*sin1
+  x1[3] = _mm_madd_epi16(u[1], sinpi_p02_m01);
+  x1[4] = _mm_madd_epi16(u[2], sinpi_p03_p02);  // x1*sin3 + x3*sin2
+  x1[5] = _mm_madd_epi16(u[3], sinpi_p03_p02);
+  x1[6] = _mm_madd_epi16(u[2], sinpi_p03_m04);  // x1*sin3 - x3*sin4
+  x1[7] = _mm_madd_epi16(u[3], sinpi_p03_m04);
+  x1[8] = _mm_madd_epi16(u[0], sinpi_p03_m03);  // x0*sin3 - x2*sin3
+  x1[9] = _mm_madd_epi16(u[1], sinpi_p03_m03);
+  x1[10] = _mm_madd_epi16(u[2], sinpi_0_p03);  // x2*sin3
+  x1[11] = _mm_madd_epi16(u[3], sinpi_0_p03);
+  x1[12] = _mm_madd_epi16(u[0], sinpi_p04_p02);  // x0*sin4 + x2*sin2
+  x1[13] = _mm_madd_epi16(u[1], sinpi_p04_p02);
+  x1[14] = _mm_madd_epi16(u[2], sinpi_m03_m01);  // -x1*sin3 - x3*sin1
+  x1[15] = _mm_madd_epi16(u[3], sinpi_m03_m01);
+
+  __m128i x2[8];
+  x2[0] = _mm_add_epi32(x1[0], x1[4]);  // x0*sin1 +x2*sin4 +x1*sin3 +x3*sin2
+  x2[1] = _mm_add_epi32(x1[1], x1[5]);
+  x2[2] = _mm_add_epi32(x1[2], x1[6]);  // x0*sin2 -x2*sin1 +x1*sin3 -x3*sin4
+  x2[3] = _mm_add_epi32(x1[3], x1[7]);
+  x2[4] = _mm_add_epi32(x1[8], x1[10]);  // x0*sin3 -x2*sin3 +x3*sin3
+  x2[5] = _mm_add_epi32(x1[9], x1[11]);
+  x2[6] = _mm_add_epi32(x1[12], x1[14]);  // x0*sin1 +x2*sin4 +x0*sin2 -x2*sin1
+  x2[7] = _mm_add_epi32(x1[13], x1[15]);
+
+  const __m128i rounding = _mm_set1_epi32(1 << (INV_COS_BIT - 1));
+  for (int i = 0; i < 4; ++i) {
+    __m128i out0 = _mm_add_epi32(x2[2 * i], rounding);
+    __m128i out1 = _mm_add_epi32(x2[2 * i + 1], rounding);
+    out0 = _mm_srai_epi32(out0, INV_COS_BIT);
+    out1 = _mm_srai_epi32(out1, INV_COS_BIT);
+    output[i] = _mm_packs_epi32(out0, out1);
+  }
+}
+
+// TODO(binpengsmail@gmail.com):
+// To explore the reuse of VP9 versions of corresponding SSE2 functions and
+// evaluate whether there is a possibility for further speedup.
+void iadst4_w4_new_sse2(const __m128i *input, __m128i *output, int8_t cos_bit) {
+  (void)cos_bit;
+  const int32_t *sinpi = sinpi_arr(INV_COS_BIT);
+  const __m128i sinpi_p01_p04 = pair_set_epi16(sinpi[1], sinpi[4]);
+  const __m128i sinpi_p02_m01 = pair_set_epi16(sinpi[2], -sinpi[1]);
+  const __m128i sinpi_p03_p02 = pair_set_epi16(sinpi[3], sinpi[2]);
+  const __m128i sinpi_p03_m04 = pair_set_epi16(sinpi[3], -sinpi[4]);
+  const __m128i sinpi_p03_m03 = pair_set_epi16(sinpi[3], -sinpi[3]);
+  const __m128i sinpi_0_p03 = pair_set_epi16(0, sinpi[3]);
+  const __m128i sinpi_p04_p02 = pair_set_epi16(sinpi[4], sinpi[2]);
+  const __m128i sinpi_m03_m01 = pair_set_epi16(-sinpi[3], -sinpi[1]);
+  __m128i x0[4];
+  x0[0] = input[0];
+  x0[1] = input[1];
+  x0[2] = input[2];
+  x0[3] = input[3];
+
+  __m128i u[2];
+  u[0] = _mm_unpacklo_epi16(x0[0], x0[2]);
+  u[1] = _mm_unpacklo_epi16(x0[1], x0[3]);
+
+  __m128i x1[8];
+  x1[0] = _mm_madd_epi16(u[0], sinpi_p01_p04);  // x0*sin1 + x2*sin4
+  x1[1] = _mm_madd_epi16(u[0], sinpi_p02_m01);  // x0*sin2 - x2*sin1
+  x1[2] = _mm_madd_epi16(u[1], sinpi_p03_p02);  // x1*sin3 + x3*sin2
+  x1[3] = _mm_madd_epi16(u[1], sinpi_p03_m04);  // x1*sin3 - x3*sin4
+  x1[4] = _mm_madd_epi16(u[0], sinpi_p03_m03);  // x0*sin3 - x2*sin3
+  x1[5] = _mm_madd_epi16(u[1], sinpi_0_p03);    // x2*sin3
+  x1[6] = _mm_madd_epi16(u[0], sinpi_p04_p02);  // x0*sin4 + x2*sin2
+  x1[7] = _mm_madd_epi16(u[1], sinpi_m03_m01);  // -x1*sin3 - x3*sin1
+
+  __m128i x2[4];
+  x2[0] = _mm_add_epi32(x1[0], x1[2]);  // x0*sin1 + x2*sin4 + x1*sin3 + x3*sin2
+  x2[1] = _mm_add_epi32(x1[1], x1[3]);  // x0*sin2 - x2*sin1 + x1*sin3 - x3*sin4
+  x2[2] = _mm_add_epi32(x1[4], x1[5]);  // x0*sin3 - x2*sin3 + x3*sin3
+  x2[3] = _mm_add_epi32(x1[6], x1[7]);  // x0*sin4 + x2*sin2 - x1*sin3 - x3*sin1
+
+  const __m128i rounding = _mm_set1_epi32(1 << (INV_COS_BIT - 1));
+  for (int i = 0; i < 4; ++i) {
+    __m128i out0 = _mm_add_epi32(x2[i], rounding);
+    out0 = _mm_srai_epi32(out0, INV_COS_BIT);
+    output[i] = _mm_packs_epi32(out0, out0);
+  }
+}
+
+static void iadst8_low1_new_ssse3(const __m128i *input, __m128i *output,
+                                  int8_t cos_bit) {
+  (void)cos_bit;
+  const int32_t *cospi = cospi_arr(INV_COS_BIT);
+  const __m128i __zero = _mm_setzero_si128();
+  const __m128i __rounding = _mm_set1_epi32(1 << (INV_COS_BIT - 1));
+
+  const __m128i cospi_p16_p48 = pair_set_epi16(cospi[16], cospi[48]);
+  const __m128i cospi_p48_m16 = pair_set_epi16(cospi[48], -cospi[16]);
+  const __m128i cospi_p32_p32 = pair_set_epi16(cospi[32], cospi[32]);
+  const __m128i cospi_p32_m32 = pair_set_epi16(cospi[32], -cospi[32]);
+
+  // stage 1
+  __m128i x[8];
+  x[1] = input[0];
+
+  // stage 2
+  btf_16_ssse3(cospi[60], -cospi[4], x[1], x[0], x[1]);
+
+  // stage 3
+  x[4] = x[0];
+  x[5] = x[1];
+
+  // stage 4
+  btf_16_sse2(cospi_p16_p48, cospi_p48_m16, x[4], x[5], x[4], x[5]);
+
+  // stage 5
+  x[2] = x[0];
+  x[3] = x[1];
+  x[6] = x[4];
+  x[7] = x[5];
+
+  // stage 6
+  btf_16_sse2(cospi_p32_p32, cospi_p32_m32, x[2], x[3], x[2], x[3]);
+  btf_16_sse2(cospi_p32_p32, cospi_p32_m32, x[6], x[7], x[6], x[7]);
+
+  // stage 7
+  output[0] = x[0];
+  output[1] = _mm_subs_epi16(__zero, x[4]);
+  output[2] = x[6];
+  output[3] = _mm_subs_epi16(__zero, x[2]);
+  output[4] = x[3];
+  output[5] = _mm_subs_epi16(__zero, x[7]);
+  output[6] = x[5];
+  output[7] = _mm_subs_epi16(__zero, x[1]);
+}
+
+void iadst8_new_sse2(const __m128i *input, __m128i *output, int8_t cos_bit) {
+  (void)cos_bit;
+  const int32_t *cospi = cospi_arr(INV_COS_BIT);
+  const __m128i __zero = _mm_setzero_si128();
+  const __m128i __rounding = _mm_set1_epi32(1 << (INV_COS_BIT - 1));
+
+  const __m128i cospi_p04_p60 = pair_set_epi16(cospi[4], cospi[60]);
+  const __m128i cospi_p60_m04 = pair_set_epi16(cospi[60], -cospi[4]);
+  const __m128i cospi_p20_p44 = pair_set_epi16(cospi[20], cospi[44]);
+  const __m128i cospi_p44_m20 = pair_set_epi16(cospi[44], -cospi[20]);
+  const __m128i cospi_p36_p28 = pair_set_epi16(cospi[36], cospi[28]);
+  const __m128i cospi_p28_m36 = pair_set_epi16(cospi[28], -cospi[36]);
+  const __m128i cospi_p52_p12 = pair_set_epi16(cospi[52], cospi[12]);
+  const __m128i cospi_p12_m52 = pair_set_epi16(cospi[12], -cospi[52]);
+  const __m128i cospi_p16_p48 = pair_set_epi16(cospi[16], cospi[48]);
+  const __m128i cospi_p48_m16 = pair_set_epi16(cospi[48], -cospi[16]);
+  const __m128i cospi_m48_p16 = pair_set_epi16(-cospi[48], cospi[16]);
+  const __m128i cospi_p32_p32 = pair_set_epi16(cospi[32], cospi[32]);
+  const __m128i cospi_p32_m32 = pair_set_epi16(cospi[32], -cospi[32]);
+
+  // stage 1
+  __m128i x[8];
+  x[0] = input[7];
+  x[1] = input[0];
+  x[2] = input[5];
+  x[3] = input[2];
+  x[4] = input[3];
+  x[5] = input[4];
+  x[6] = input[1];
+  x[7] = input[6];
+
+  // stage 2
+  btf_16_sse2(cospi_p04_p60, cospi_p60_m04, x[0], x[1], x[0], x[1]);
+  btf_16_sse2(cospi_p20_p44, cospi_p44_m20, x[2], x[3], x[2], x[3]);
+  btf_16_sse2(cospi_p36_p28, cospi_p28_m36, x[4], x[5], x[4], x[5]);
+  btf_16_sse2(cospi_p52_p12, cospi_p12_m52, x[6], x[7], x[6], x[7]);
+
+  // stage 3
+  btf_16_adds_subs_sse2(x[0], x[4]);
+  btf_16_adds_subs_sse2(x[1], x[5]);
+  btf_16_adds_subs_sse2(x[2], x[6]);
+  btf_16_adds_subs_sse2(x[3], x[7]);
+
+  // stage 4
+  btf_16_sse2(cospi_p16_p48, cospi_p48_m16, x[4], x[5], x[4], x[5]);
+  btf_16_sse2(cospi_m48_p16, cospi_p16_p48, x[6], x[7], x[6], x[7]);
+
+  // stage 5
+  btf_16_adds_subs_sse2(x[0], x[2]);
+  btf_16_adds_subs_sse2(x[1], x[3]);
+  btf_16_adds_subs_sse2(x[4], x[6]);
+  btf_16_adds_subs_sse2(x[5], x[7]);
+
+  // stage 6
+  btf_16_sse2(cospi_p32_p32, cospi_p32_m32, x[2], x[3], x[2], x[3]);
+  btf_16_sse2(cospi_p32_p32, cospi_p32_m32, x[6], x[7], x[6], x[7]);
+
+  // stage 7
+  output[0] = x[0];
+  output[1] = _mm_subs_epi16(__zero, x[4]);
+  output[2] = x[6];
+  output[3] = _mm_subs_epi16(__zero, x[2]);
+  output[4] = x[3];
+  output[5] = _mm_subs_epi16(__zero, x[7]);
+  output[6] = x[5];
+  output[7] = _mm_subs_epi16(__zero, x[1]);
+}
+
+void iadst8_w4_new_sse2(const __m128i *input, __m128i *output, int8_t cos_bit) {
+  (void)cos_bit;
+  const int32_t *cospi = cospi_arr(INV_COS_BIT);
+  const __m128i __zero = _mm_setzero_si128();
+  const __m128i __rounding = _mm_set1_epi32(1 << (INV_COS_BIT - 1));
+
+  const __m128i cospi_p04_p60 = pair_set_epi16(cospi[4], cospi[60]);
+  const __m128i cospi_p60_m04 = pair_set_epi16(cospi[60], -cospi[4]);
+  const __m128i cospi_p20_p44 = pair_set_epi16(cospi[20], cospi[44]);
+  const __m128i cospi_p44_m20 = pair_set_epi16(cospi[44], -cospi[20]);
+  const __m128i cospi_p36_p28 = pair_set_epi16(cospi[36], cospi[28]);
+  const __m128i cospi_p28_m36 = pair_set_epi16(cospi[28], -cospi[36]);
+  const __m128i cospi_p52_p12 = pair_set_epi16(cospi[52], cospi[12]);
+  const __m128i cospi_p12_m52 = pair_set_epi16(cospi[12], -cospi[52]);
+  const __m128i cospi_p16_p48 = pair_set_epi16(cospi[16], cospi[48]);
+  const __m128i cospi_p48_m16 = pair_set_epi16(cospi[48], -cospi[16]);
+  const __m128i cospi_m48_p16 = pair_set_epi16(-cospi[48], cospi[16]);
+  const __m128i cospi_p32_p32 = pair_set_epi16(cospi[32], cospi[32]);
+  const __m128i cospi_p32_m32 = pair_set_epi16(cospi[32], -cospi[32]);
+
+  // stage 1
+  __m128i x[8];
+  x[0] = input[7];
+  x[1] = input[0];
+  x[2] = input[5];
+  x[3] = input[2];
+  x[4] = input[3];
+  x[5] = input[4];
+  x[6] = input[1];
+  x[7] = input[6];
+
+  // stage 2
+  btf_16_4p_sse2(cospi_p04_p60, cospi_p60_m04, x[0], x[1], x[0], x[1]);
+  btf_16_4p_sse2(cospi_p20_p44, cospi_p44_m20, x[2], x[3], x[2], x[3]);
+  btf_16_4p_sse2(cospi_p36_p28, cospi_p28_m36, x[4], x[5], x[4], x[5]);
+  btf_16_4p_sse2(cospi_p52_p12, cospi_p12_m52, x[6], x[7], x[6], x[7]);
+
+  // stage 3
+  btf_16_adds_subs_sse2(x[0], x[4]);
+  btf_16_adds_subs_sse2(x[1], x[5]);
+  btf_16_adds_subs_sse2(x[2], x[6]);
+  btf_16_adds_subs_sse2(x[3], x[7]);
+
+  // stage 4
+  btf_16_4p_sse2(cospi_p16_p48, cospi_p48_m16, x[4], x[5], x[4], x[5]);
+  btf_16_4p_sse2(cospi_m48_p16, cospi_p16_p48, x[6], x[7], x[6], x[7]);
+
+  // stage 5
+  btf_16_adds_subs_sse2(x[0], x[2]);
+  btf_16_adds_subs_sse2(x[1], x[3]);
+  btf_16_adds_subs_sse2(x[4], x[6]);
+  btf_16_adds_subs_sse2(x[5], x[7]);
+
+  // stage 6
+  btf_16_4p_sse2(cospi_p32_p32, cospi_p32_m32, x[2], x[3], x[2], x[3]);
+  btf_16_4p_sse2(cospi_p32_p32, cospi_p32_m32, x[6], x[7], x[6], x[7]);
+
+  // stage 7
+  output[0] = x[0];
+  output[1] = _mm_subs_epi16(__zero, x[4]);
+  output[2] = x[6];
+  output[3] = _mm_subs_epi16(__zero, x[2]);
+  output[4] = x[3];
+  output[5] = _mm_subs_epi16(__zero, x[7]);
+  output[6] = x[5];
+  output[7] = _mm_subs_epi16(__zero, x[1]);
+}
+
+static INLINE void iadst16_stage3_ssse3(__m128i *x) {
+  btf_16_adds_subs_sse2(x[0], x[8]);
+  btf_16_adds_subs_sse2(x[1], x[9]);
+  btf_16_adds_subs_sse2(x[2], x[10]);
+  btf_16_adds_subs_sse2(x[3], x[11]);
+  btf_16_adds_subs_sse2(x[4], x[12]);
+  btf_16_adds_subs_sse2(x[5], x[13]);
+  btf_16_adds_subs_sse2(x[6], x[14]);
+  btf_16_adds_subs_sse2(x[7], x[15]);
+}
+
+static INLINE void iadst16_stage4_ssse3(__m128i *x, const int32_t *cospi,
+                                        const __m128i __rounding,
+                                        int8_t cos_bit) {
+  const __m128i cospi_p08_p56 = pair_set_epi16(cospi[8], cospi[56]);
+  const __m128i cospi_p56_m08 = pair_set_epi16(cospi[56], -cospi[8]);
+  const __m128i cospi_p40_p24 = pair_set_epi16(cospi[40], cospi[24]);
+  const __m128i cospi_p24_m40 = pair_set_epi16(cospi[24], -cospi[40]);
+  const __m128i cospi_m56_p08 = pair_set_epi16(-cospi[56], cospi[8]);
+  const __m128i cospi_m24_p40 = pair_set_epi16(-cospi[24], cospi[40]);
+  btf_16_sse2(cospi_p08_p56, cospi_p56_m08, x[8], x[9], x[8], x[9]);
+  btf_16_sse2(cospi_p40_p24, cospi_p24_m40, x[10], x[11], x[10], x[11]);
+  btf_16_sse2(cospi_m56_p08, cospi_p08_p56, x[12], x[13], x[12], x[13]);
+  btf_16_sse2(cospi_m24_p40, cospi_p40_p24, x[14], x[15], x[14], x[15]);
+}
+
+static INLINE void iadst16_stage5_ssse3(__m128i *x) {
+  btf_16_adds_subs_sse2(x[0], x[4]);
+  btf_16_adds_subs_sse2(x[1], x[5]);
+  btf_16_adds_subs_sse2(x[2], x[6]);
+  btf_16_adds_subs_sse2(x[3], x[7]);
+  btf_16_adds_subs_sse2(x[8], x[12]);
+  btf_16_adds_subs_sse2(x[9], x[13]);
+  btf_16_adds_subs_sse2(x[10], x[14]);
+  btf_16_adds_subs_sse2(x[11], x[15]);
+}
+
+static INLINE void iadst16_stage6_ssse3(__m128i *x, const int32_t *cospi,
+                                        const __m128i __rounding,
+                                        int8_t cos_bit) {
+  const __m128i cospi_p16_p48 = pair_set_epi16(cospi[16], cospi[48]);
+  const __m128i cospi_p48_m16 = pair_set_epi16(cospi[48], -cospi[16]);
+  const __m128i cospi_m48_p16 = pair_set_epi16(-cospi[48], cospi[16]);
+  btf_16_sse2(cospi_p16_p48, cospi_p48_m16, x[4], x[5], x[4], x[5]);
+  btf_16_sse2(cospi_m48_p16, cospi_p16_p48, x[6], x[7], x[6], x[7]);
+  btf_16_sse2(cospi_p16_p48, cospi_p48_m16, x[12], x[13], x[12], x[13]);
+  btf_16_sse2(cospi_m48_p16, cospi_p16_p48, x[14], x[15], x[14], x[15]);
+}
+
+static INLINE void iadst16_stage7_ssse3(__m128i *x) {
+  btf_16_adds_subs_sse2(x[0], x[2]);
+  btf_16_adds_subs_sse2(x[1], x[3]);
+  btf_16_adds_subs_sse2(x[4], x[6]);
+  btf_16_adds_subs_sse2(x[5], x[7]);
+  btf_16_adds_subs_sse2(x[8], x[10]);
+  btf_16_adds_subs_sse2(x[9], x[11]);
+  btf_16_adds_subs_sse2(x[12], x[14]);
+  btf_16_adds_subs_sse2(x[13], x[15]);
+}
+
+static INLINE void iadst16_stage8_ssse3(__m128i *x, const int32_t *cospi,
+                                        const __m128i __rounding,
+                                        int8_t cos_bit) {
+  const __m128i cospi_p32_p32 = pair_set_epi16(cospi[32], cospi[32]);
+  const __m128i cospi_p32_m32 = pair_set_epi16(cospi[32], -cospi[32]);
+  btf_16_sse2(cospi_p32_p32, cospi_p32_m32, x[2], x[3], x[2], x[3]);
+  btf_16_sse2(cospi_p32_p32, cospi_p32_m32, x[6], x[7], x[6], x[7]);
+  btf_16_sse2(cospi_p32_p32, cospi_p32_m32, x[10], x[11], x[10], x[11]);
+  btf_16_sse2(cospi_p32_p32, cospi_p32_m32, x[14], x[15], x[14], x[15]);
+}
+
+static INLINE void iadst16_stage9_ssse3(__m128i *output, __m128i *x) {
+  const __m128i __zero = _mm_setzero_si128();
+  output[0] = x[0];
+  output[1] = _mm_subs_epi16(__zero, x[8]);
+  output[2] = x[12];
+  output[3] = _mm_subs_epi16(__zero, x[4]);
+  output[4] = x[6];
+  output[5] = _mm_subs_epi16(__zero, x[14]);
+  output[6] = x[10];
+  output[7] = _mm_subs_epi16(__zero, x[2]);
+  output[8] = x[3];
+  output[9] = _mm_subs_epi16(__zero, x[11]);
+  output[10] = x[15];
+  output[11] = _mm_subs_epi16(__zero, x[7]);
+  output[12] = x[5];
+  output[13] = _mm_subs_epi16(__zero, x[13]);
+  output[14] = x[9];
+  output[15] = _mm_subs_epi16(__zero, x[1]);
+}
+
+static void iadst16_low1_new_ssse3(const __m128i *input, __m128i *output,
+                                   int8_t cos_bit) {
+  (void)cos_bit;
+  const int32_t *cospi = cospi_arr(INV_COS_BIT);
+  const __m128i __rounding = _mm_set1_epi32(1 << (INV_COS_BIT - 1));
+
+  const __m128i cospi_p08_p56 = pair_set_epi16(cospi[8], cospi[56]);
+  const __m128i cospi_p56_m08 = pair_set_epi16(cospi[56], -cospi[8]);
+  const __m128i cospi_p16_p48 = pair_set_epi16(cospi[16], cospi[48]);
+  const __m128i cospi_p48_m16 = pair_set_epi16(cospi[48], -cospi[16]);
+
+  // stage 1
+  __m128i x[16];
+  x[1] = input[0];
+
+  // stage 2
+  btf_16_ssse3(cospi[62], -cospi[2], x[1], x[0], x[1]);
+
+  // stage 3
+  x[8] = x[0];
+  x[9] = x[1];
+
+  // stage 4
+  btf_16_sse2(cospi_p08_p56, cospi_p56_m08, x[8], x[9], x[8], x[9]);
+
+  // stage 5
+  x[4] = x[0];
+  x[5] = x[1];
+  x[12] = x[8];
+  x[13] = x[9];
+
+  // stage 6
+  btf_16_sse2(cospi_p16_p48, cospi_p48_m16, x[4], x[5], x[4], x[5]);
+  btf_16_sse2(cospi_p16_p48, cospi_p48_m16, x[12], x[13], x[12], x[13]);
+
+  // stage 7
+  x[2] = x[0];
+  x[3] = x[1];
+  x[6] = x[4];
+  x[7] = x[5];
+  x[10] = x[8];
+  x[11] = x[9];
+  x[14] = x[12];
+  x[15] = x[13];
+
+  iadst16_stage8_ssse3(x, cospi, __rounding, cos_bit);
+  iadst16_stage9_ssse3(output, x);
+}
+
+static void iadst16_low8_new_ssse3(const __m128i *input, __m128i *output,
+                                   int8_t cos_bit) {
+  (void)cos_bit;
+  const int32_t *cospi = cospi_arr(INV_COS_BIT);
+  const __m128i __rounding = _mm_set1_epi32(1 << (INV_COS_BIT - 1));
+
+  // stage 1
+  __m128i x[16];
+  x[1] = input[0];
+  x[3] = input[2];
+  x[5] = input[4];
+  x[7] = input[6];
+  x[8] = input[7];
+  x[10] = input[5];
+  x[12] = input[3];
+  x[14] = input[1];
+
+  // stage 2
+  btf_16_ssse3(cospi[62], -cospi[2], x[1], x[0], x[1]);
+  btf_16_ssse3(cospi[54], -cospi[10], x[3], x[2], x[3]);
+  btf_16_ssse3(cospi[46], -cospi[18], x[5], x[4], x[5]);
+  btf_16_ssse3(cospi[38], -cospi[26], x[7], x[6], x[7]);
+  btf_16_ssse3(cospi[34], cospi[30], x[8], x[8], x[9]);
+  btf_16_ssse3(cospi[42], cospi[22], x[10], x[10], x[11]);
+  btf_16_ssse3(cospi[50], cospi[14], x[12], x[12], x[13]);
+  btf_16_ssse3(cospi[58], cospi[6], x[14], x[14], x[15]);
+
+  // stage 3
+  iadst16_stage3_ssse3(x);
+  iadst16_stage4_ssse3(x, cospi, __rounding, cos_bit);
+  iadst16_stage5_ssse3(x);
+  iadst16_stage6_ssse3(x, cospi, __rounding, cos_bit);
+  iadst16_stage7_ssse3(x);
+  iadst16_stage8_ssse3(x, cospi, __rounding, cos_bit);
+  iadst16_stage9_ssse3(output, x);
+}
+void iadst16_new_sse2(const __m128i *input, __m128i *output, int8_t cos_bit) {
+  (void)cos_bit;
+  const int32_t *cospi = cospi_arr(INV_COS_BIT);
+  const __m128i __rounding = _mm_set1_epi32(1 << (INV_COS_BIT - 1));
+  const __m128i cospi_p02_p62 = pair_set_epi16(cospi[2], cospi[62]);
+  const __m128i cospi_p62_m02 = pair_set_epi16(cospi[62], -cospi[2]);
+  const __m128i cospi_p10_p54 = pair_set_epi16(cospi[10], cospi[54]);
+  const __m128i cospi_p54_m10 = pair_set_epi16(cospi[54], -cospi[10]);
+  const __m128i cospi_p18_p46 = pair_set_epi16(cospi[18], cospi[46]);
+  const __m128i cospi_p46_m18 = pair_set_epi16(cospi[46], -cospi[18]);
+  const __m128i cospi_p26_p38 = pair_set_epi16(cospi[26], cospi[38]);
+  const __m128i cospi_p38_m26 = pair_set_epi16(cospi[38], -cospi[26]);
+  const __m128i cospi_p34_p30 = pair_set_epi16(cospi[34], cospi[30]);
+  const __m128i cospi_p30_m34 = pair_set_epi16(cospi[30], -cospi[34]);
+  const __m128i cospi_p42_p22 = pair_set_epi16(cospi[42], cospi[22]);
+  const __m128i cospi_p22_m42 = pair_set_epi16(cospi[22], -cospi[42]);
+  const __m128i cospi_p50_p14 = pair_set_epi16(cospi[50], cospi[14]);
+  const __m128i cospi_p14_m50 = pair_set_epi16(cospi[14], -cospi[50]);
+  const __m128i cospi_p58_p06 = pair_set_epi16(cospi[58], cospi[6]);
+  const __m128i cospi_p06_m58 = pair_set_epi16(cospi[6], -cospi[58]);
+
+  // stage 1
+  __m128i x[16];
+  x[0] = input[15];
+  x[1] = input[0];
+  x[2] = input[13];
+  x[3] = input[2];
+  x[4] = input[11];
+  x[5] = input[4];
+  x[6] = input[9];
+  x[7] = input[6];
+  x[8] = input[7];
+  x[9] = input[8];
+  x[10] = input[5];
+  x[11] = input[10];
+  x[12] = input[3];
+  x[13] = input[12];
+  x[14] = input[1];
+  x[15] = input[14];
+
+  // stage 2
+  btf_16_sse2(cospi_p02_p62, cospi_p62_m02, x[0], x[1], x[0], x[1]);
+  btf_16_sse2(cospi_p10_p54, cospi_p54_m10, x[2], x[3], x[2], x[3]);
+  btf_16_sse2(cospi_p18_p46, cospi_p46_m18, x[4], x[5], x[4], x[5]);
+  btf_16_sse2(cospi_p26_p38, cospi_p38_m26, x[6], x[7], x[6], x[7]);
+  btf_16_sse2(cospi_p34_p30, cospi_p30_m34, x[8], x[9], x[8], x[9]);
+  btf_16_sse2(cospi_p42_p22, cospi_p22_m42, x[10], x[11], x[10], x[11]);
+  btf_16_sse2(cospi_p50_p14, cospi_p14_m50, x[12], x[13], x[12], x[13]);
+  btf_16_sse2(cospi_p58_p06, cospi_p06_m58, x[14], x[15], x[14], x[15]);
+
+  // stage 3~9
+  iadst16_stage3_ssse3(x);
+  iadst16_stage4_ssse3(x, cospi, __rounding, cos_bit);
+  iadst16_stage5_ssse3(x);
+  iadst16_stage6_ssse3(x, cospi, __rounding, cos_bit);
+  iadst16_stage7_ssse3(x);
+  iadst16_stage8_ssse3(x, cospi, __rounding, cos_bit);
+  iadst16_stage9_ssse3(output, x);
+}
+
+void iadst16_w4_new_sse2(const __m128i *input, __m128i *output,
+                         int8_t cos_bit) {
+  (void)cos_bit;
+  const int32_t *cospi = cospi_arr(INV_COS_BIT);
+  const __m128i __rounding = _mm_set1_epi32(1 << (INV_COS_BIT - 1));
+
+  const __m128i cospi_p02_p62 = pair_set_epi16(cospi[2], cospi[62]);
+  const __m128i cospi_p62_m02 = pair_set_epi16(cospi[62], -cospi[2]);
+  const __m128i cospi_p10_p54 = pair_set_epi16(cospi[10], cospi[54]);
+  const __m128i cospi_p54_m10 = pair_set_epi16(cospi[54], -cospi[10]);
+  const __m128i cospi_p18_p46 = pair_set_epi16(cospi[18], cospi[46]);
+  const __m128i cospi_p46_m18 = pair_set_epi16(cospi[46], -cospi[18]);
+  const __m128i cospi_p26_p38 = pair_set_epi16(cospi[26], cospi[38]);
+  const __m128i cospi_p38_m26 = pair_set_epi16(cospi[38], -cospi[26]);
+  const __m128i cospi_p34_p30 = pair_set_epi16(cospi[34], cospi[30]);
+  const __m128i cospi_p30_m34 = pair_set_epi16(cospi[30], -cospi[34]);
+  const __m128i cospi_p42_p22 = pair_set_epi16(cospi[42], cospi[22]);
+  const __m128i cospi_p22_m42 = pair_set_epi16(cospi[22], -cospi[42]);
+  const __m128i cospi_p50_p14 = pair_set_epi16(cospi[50], cospi[14]);
+  const __m128i cospi_p14_m50 = pair_set_epi16(cospi[14], -cospi[50]);
+  const __m128i cospi_p58_p06 = pair_set_epi16(cospi[58], cospi[6]);
+  const __m128i cospi_p06_m58 = pair_set_epi16(cospi[6], -cospi[58]);
+  const __m128i cospi_p08_p56 = pair_set_epi16(cospi[8], cospi[56]);
+  const __m128i cospi_p56_m08 = pair_set_epi16(cospi[56], -cospi[8]);
+  const __m128i cospi_p40_p24 = pair_set_epi16(cospi[40], cospi[24]);
+  const __m128i cospi_p24_m40 = pair_set_epi16(cospi[24], -cospi[40]);
+  const __m128i cospi_m56_p08 = pair_set_epi16(-cospi[56], cospi[8]);
+  const __m128i cospi_m24_p40 = pair_set_epi16(-cospi[24], cospi[40]);
+  const __m128i cospi_p16_p48 = pair_set_epi16(cospi[16], cospi[48]);
+  const __m128i cospi_p48_m16 = pair_set_epi16(cospi[48], -cospi[16]);
+  const __m128i cospi_m48_p16 = pair_set_epi16(-cospi[48], cospi[16]);
+  const __m128i cospi_p32_p32 = pair_set_epi16(cospi[32], cospi[32]);
+  const __m128i cospi_p32_m32 = pair_set_epi16(cospi[32], -cospi[32]);
+
+  // stage 1
+  __m128i x[16];
+  x[0] = input[15];
+  x[1] = input[0];
+  x[2] = input[13];
+  x[3] = input[2];
+  x[4] = input[11];
+  x[5] = input[4];
+  x[6] = input[9];
+  x[7] = input[6];
+  x[8] = input[7];
+  x[9] = input[8];
+  x[10] = input[5];
+  x[11] = input[10];
+  x[12] = input[3];
+  x[13] = input[12];
+  x[14] = input[1];
+  x[15] = input[14];
+
+  // stage 2
+  btf_16_4p_sse2(cospi_p02_p62, cospi_p62_m02, x[0], x[1], x[0], x[1]);
+  btf_16_4p_sse2(cospi_p10_p54, cospi_p54_m10, x[2], x[3], x[2], x[3]);
+  btf_16_4p_sse2(cospi_p18_p46, cospi_p46_m18, x[4], x[5], x[4], x[5]);
+  btf_16_4p_sse2(cospi_p26_p38, cospi_p38_m26, x[6], x[7], x[6], x[7]);
+  btf_16_4p_sse2(cospi_p34_p30, cospi_p30_m34, x[8], x[9], x[8], x[9]);
+  btf_16_4p_sse2(cospi_p42_p22, cospi_p22_m42, x[10], x[11], x[10], x[11]);
+  btf_16_4p_sse2(cospi_p50_p14, cospi_p14_m50, x[12], x[13], x[12], x[13]);
+  btf_16_4p_sse2(cospi_p58_p06, cospi_p06_m58, x[14], x[15], x[14], x[15]);
+
+  // stage 3
+  iadst16_stage3_ssse3(x);
+
+  // stage 4
+  btf_16_4p_sse2(cospi_p08_p56, cospi_p56_m08, x[8], x[9], x[8], x[9]);
+  btf_16_4p_sse2(cospi_p40_p24, cospi_p24_m40, x[10], x[11], x[10], x[11]);
+  btf_16_4p_sse2(cospi_m56_p08, cospi_p08_p56, x[12], x[13], x[12], x[13]);
+  btf_16_4p_sse2(cospi_m24_p40, cospi_p40_p24, x[14], x[15], x[14], x[15]);
+
+  // stage 5
+  iadst16_stage5_ssse3(x);
+
+  // stage 6
+  btf_16_4p_sse2(cospi_p16_p48, cospi_p48_m16, x[4], x[5], x[4], x[5]);
+  btf_16_4p_sse2(cospi_m48_p16, cospi_p16_p48, x[6], x[7], x[6], x[7]);
+  btf_16_4p_sse2(cospi_p16_p48, cospi_p48_m16, x[12], x[13], x[12], x[13]);
+  btf_16_4p_sse2(cospi_m48_p16, cospi_p16_p48, x[14], x[15], x[14], x[15]);
+
+  // stage 7
+  iadst16_stage7_ssse3(x);
+
+  // stage 8
+  btf_16_4p_sse2(cospi_p32_p32, cospi_p32_m32, x[2], x[3], x[2], x[3]);
+  btf_16_4p_sse2(cospi_p32_p32, cospi_p32_m32, x[6], x[7], x[6], x[7]);
+  btf_16_4p_sse2(cospi_p32_p32, cospi_p32_m32, x[10], x[11], x[10], x[11]);
+  btf_16_4p_sse2(cospi_p32_p32, cospi_p32_m32, x[14], x[15], x[14], x[15]);
+
+  // stage 9
+  iadst16_stage9_ssse3(output, x);
+}
+
+static void iidentity4_new_ssse3(const __m128i *input, __m128i *output,
+                                 int8_t cos_bit) {
+  (void)cos_bit;
+  const int16_t scale_fractional = (NewSqrt2 - (1 << NewSqrt2Bits));
+  const __m128i scale = _mm_set1_epi16(scale_fractional << (15 - NewSqrt2Bits));
+  for (int i = 0; i < 4; ++i) {
+    __m128i x = _mm_mulhrs_epi16(input[i], scale);
+    output[i] = _mm_adds_epi16(x, input[i]);
+  }
+}
+
+static void iidentity8_new_sse2(const __m128i *input, __m128i *output,
+                                int8_t cos_bit) {
+  (void)cos_bit;
+  for (int i = 0; i < 8; ++i) {
+    output[i] = _mm_adds_epi16(input[i], input[i]);
+  }
+}
+
+static void iidentity16_new_ssse3(const __m128i *input, __m128i *output,
+                                  int8_t cos_bit) {
+  (void)cos_bit;
+  const int16_t scale_fractional = 2 * (NewSqrt2 - (1 << NewSqrt2Bits));
+  const __m128i scale = _mm_set1_epi16(scale_fractional << (15 - NewSqrt2Bits));
+  for (int i = 0; i < 16; ++i) {
+    __m128i x = _mm_mulhrs_epi16(input[i], scale);
+    __m128i srcx2 = _mm_adds_epi16(input[i], input[i]);
+    output[i] = _mm_adds_epi16(x, srcx2);
+  }
+}
+
+static INLINE __m128i lowbd_get_recon_8x8_sse2(const __m128i pred,
+                                               __m128i res) {
+  const __m128i zero = _mm_setzero_si128();
+  __m128i x0 = _mm_adds_epi16(res, _mm_unpacklo_epi8(pred, zero));
+  return _mm_packus_epi16(x0, x0);
+}
+
+static INLINE void lowbd_write_buffer_4xn_sse2(__m128i *in, uint8_t *output,
+                                               int stride, int flipud,
+                                               const int height) {
+  int j = flipud ? (height - 1) : 0;
+  const int step = flipud ? -1 : 1;
+  const __m128i zero = _mm_setzero_si128();
+  for (int i = 0; i < height; ++i, j += step) {
+    const __m128i v = _mm_cvtsi32_si128(*((uint32_t *)(output + i * stride)));
+    __m128i u = _mm_adds_epi16(in[j], _mm_unpacklo_epi8(v, zero));
+    u = _mm_packus_epi16(u, zero);
+    *((uint32_t *)(output + i * stride)) = _mm_cvtsi128_si32(u);
+  }
+}
+
+static INLINE void lowbd_write_buffer_8xn_sse2(__m128i *in, uint8_t *output,
+                                               int stride, int flipud,
+                                               const int height) {
+  int j = flipud ? (height - 1) : 0;
+  const int step = flipud ? -1 : 1;
+  for (int i = 0; i < height; ++i, j += step) {
+    const __m128i v = _mm_loadl_epi64((__m128i const *)(output + i * stride));
+    const __m128i u = lowbd_get_recon_8x8_sse2(v, in[j]);
+    _mm_storel_epi64((__m128i *)(output + i * stride), u);
+  }
+}
+
+// 1D functions process process 8 pixels at one time.
+static const transform_1d_ssse3
+    lowbd_txfm_all_1d_w8_arr[TX_SIZES][ITX_TYPES_1D] = {
+      { idct4_new_sse2, iadst4_new_sse2, iidentity4_new_ssse3 },
+      { idct8_new_sse2, iadst8_new_sse2, iidentity8_new_sse2 },
+      { idct16_new_sse2, iadst16_new_sse2, iidentity16_new_ssse3 },
+      { idct32_new_sse2, NULL, NULL },
+      { idct64_low32_new_ssse3, NULL, NULL },
+    };
+
+// functions for blocks with eob at DC and within
+// topleft 8x8, 16x16, 32x32 corner
+static const transform_1d_ssse3
+    lowbd_txfm_all_1d_zeros_w8_arr[TX_SIZES][ITX_TYPES_1D][4] = {
+      {
+          { idct4_new_sse2, idct4_new_sse2, NULL, NULL },
+          { iadst4_new_sse2, iadst4_new_sse2, NULL, NULL },
+          { iidentity4_new_ssse3, iidentity4_new_ssse3, NULL, NULL },
+      },
+      { { idct8_low1_new_ssse3, idct8_new_sse2, NULL, NULL },
+        { iadst8_low1_new_ssse3, iadst8_new_sse2, NULL, NULL },
+        { iidentity8_new_sse2, iidentity8_new_sse2, NULL, NULL } },
+      {
+          { idct16_low1_new_ssse3, idct16_low8_new_ssse3, idct16_new_sse2,
+            NULL },
+          { iadst16_low1_new_ssse3, iadst16_low8_new_ssse3, iadst16_new_sse2,
+            NULL },
+          { NULL, NULL, NULL, NULL },
+      },
+      { { idct32_low1_new_ssse3, idct32_low8_new_ssse3, idct32_low16_new_ssse3,
+          idct32_new_sse2 },
+        { NULL, NULL, NULL, NULL },
+        { NULL, NULL, NULL, NULL } },
+      { { idct64_low1_new_ssse3, idct64_low8_new_ssse3, idct64_low16_new_ssse3,
+          idct64_low32_new_ssse3 },
+        { NULL, NULL, NULL, NULL },
+        { NULL, NULL, NULL, NULL } }
+    };
+
+// 1D functions process process 4 pixels at one time.
+// used in 4x4, 4x8, 4x16, 8x4, 16x4
+static const transform_1d_ssse3
+    lowbd_txfm_all_1d_w4_arr[TX_SIZES][ITX_TYPES_1D] = {
+      { idct4_w4_new_sse2, iadst4_w4_new_sse2, iidentity4_new_ssse3 },
+      { idct8_w4_new_sse2, iadst8_w4_new_sse2, iidentity8_new_sse2 },
+      { idct16_w4_new_sse2, iadst16_w4_new_sse2, iidentity16_new_ssse3 },
+      { NULL, NULL, NULL },
+      { NULL, NULL, NULL },
+    };
+
+static INLINE void iidentity_row_8xn_ssse3(__m128i *out, const int32_t *input,
+                                           int stride, int shift, int height,
+                                           int txw_idx, int rect_type) {
+  const int32_t *input_row = input;
+  const __m128i scale = _mm_set1_epi16(NewSqrt2list[txw_idx]);
+  const __m128i rounding = _mm_set1_epi16((1 << (NewSqrt2Bits - 1)) +
+                                          (1 << (NewSqrt2Bits - shift - 1)));
+  const __m128i one = _mm_set1_epi16(1);
+  const __m128i scale_rounding = _mm_unpacklo_epi16(scale, rounding);
+  if (rect_type != 1 && rect_type != -1) {
+    for (int i = 0; i < height; ++i) {
+      const __m128i src = load_32bit_to_16bit(input_row);
+      input_row += stride;
+      __m128i lo = _mm_unpacklo_epi16(src, one);
+      __m128i hi = _mm_unpackhi_epi16(src, one);
+      lo = _mm_madd_epi16(lo, scale_rounding);
+      hi = _mm_madd_epi16(hi, scale_rounding);
+      lo = _mm_srai_epi32(lo, NewSqrt2Bits - shift);
+      hi = _mm_srai_epi32(hi, NewSqrt2Bits - shift);
+      out[i] = _mm_packs_epi32(lo, hi);
+    }
+  } else {
+    const __m128i rect_scale =
+        _mm_set1_epi16(NewInvSqrt2 << (15 - NewSqrt2Bits));
+    for (int i = 0; i < height; ++i) {
+      __m128i src = load_32bit_to_16bit(input_row);
+      src = _mm_mulhrs_epi16(src, rect_scale);
+      input_row += stride;
+      __m128i lo = _mm_unpacklo_epi16(src, one);
+      __m128i hi = _mm_unpackhi_epi16(src, one);
+      lo = _mm_madd_epi16(lo, scale_rounding);
+      hi = _mm_madd_epi16(hi, scale_rounding);
+      lo = _mm_srai_epi32(lo, NewSqrt2Bits - shift);
+      hi = _mm_srai_epi32(hi, NewSqrt2Bits - shift);
+      out[i] = _mm_packs_epi32(lo, hi);
+    }
+  }
+}
+
+static INLINE void iidentity_col_8xn_ssse3(uint8_t *output, int stride,
+                                           __m128i *buf, int shift, int height,
+                                           int txh_idx) {
+  const __m128i scale = _mm_set1_epi16(NewSqrt2list[txh_idx]);
+  const __m128i scale_rounding = _mm_set1_epi16(1 << (NewSqrt2Bits - 1));
+  const __m128i shift_rounding = _mm_set1_epi32(1 << (-shift - 1));
+  const __m128i one = _mm_set1_epi16(1);
+  const __m128i scale_coeff = _mm_unpacklo_epi16(scale, scale_rounding);
+  const __m128i zero = _mm_setzero_si128();
+  for (int h = 0; h < height; ++h) {
+    __m128i lo = _mm_unpacklo_epi16(buf[h], one);
+    __m128i hi = _mm_unpackhi_epi16(buf[h], one);
+    lo = _mm_madd_epi16(lo, scale_coeff);
+    hi = _mm_madd_epi16(hi, scale_coeff);
+    lo = _mm_srai_epi32(lo, NewSqrt2Bits);
+    hi = _mm_srai_epi32(hi, NewSqrt2Bits);
+    lo = _mm_add_epi32(lo, shift_rounding);
+    hi = _mm_add_epi32(hi, shift_rounding);
+    lo = _mm_srai_epi32(lo, -shift);
+    hi = _mm_srai_epi32(hi, -shift);
+    __m128i x = _mm_packs_epi32(lo, hi);
+
+    const __m128i pred = _mm_loadl_epi64((__m128i const *)(output));
+    x = _mm_adds_epi16(x, _mm_unpacklo_epi8(pred, zero));
+    const __m128i u = _mm_packus_epi16(x, x);
+    _mm_storel_epi64((__m128i *)(output), u);
+    output += stride;
+  }
+}
+
+static INLINE void lowbd_inv_txfm2d_add_idtx_ssse3(const int32_t *input,
+                                                   uint8_t *output, int stride,
+                                                   TX_SIZE tx_size) {
+  const int8_t *shift = inv_txfm_shift_ls[tx_size];
+  const int txw_idx = get_txw_idx(tx_size);
+  const int txh_idx = get_txh_idx(tx_size);
+  const int txfm_size_col = tx_size_wide[tx_size];
+  const int txfm_size_row = tx_size_high[tx_size];
+  const int input_stride = AOMMIN(32, txfm_size_col);
+  const int row_max = AOMMIN(32, txfm_size_row);
+  const int rect_type = get_rect_tx_log_ratio(txfm_size_col, txfm_size_row);
+  __m128i buf[32];
+
+  for (int i = 0; i < (input_stride >> 3); ++i) {
+    iidentity_row_8xn_ssse3(buf, input + 8 * i, input_stride, shift[0], row_max,
+                            txw_idx, rect_type);
+    iidentity_col_8xn_ssse3(output + 8 * i, stride, buf, shift[1], row_max,
+                            txh_idx);
+  }
+}
+
+void lowbd_inv_txfm2d_add_4x4_ssse3(const int32_t *input, uint8_t *output,
+                                    int stride, TX_TYPE tx_type,
+                                    TX_SIZE tx_size_, int eob) {
+  (void)tx_size_;
+  (void)eob;
+  __m128i buf[4];
+  const TX_SIZE tx_size = TX_4X4;
+  const int8_t *shift = inv_txfm_shift_ls[tx_size];
+  const int txw_idx = get_txw_idx(tx_size);
+  const int txh_idx = get_txh_idx(tx_size);
+  const int cos_bit_row = inv_cos_bit_row[txw_idx][txh_idx];
+  const int cos_bit_col = inv_cos_bit_col[txw_idx][txh_idx];
+  const int txfm_size_col = tx_size_wide[tx_size];
+  const int txfm_size_row = tx_size_high[tx_size];
+
+  const transform_1d_ssse3 row_txfm =
+      lowbd_txfm_all_1d_w4_arr[txw_idx][hitx_1d_tab[tx_type]];
+  const transform_1d_ssse3 col_txfm =
+      lowbd_txfm_all_1d_w4_arr[txh_idx][vitx_1d_tab[tx_type]];
+
+  int ud_flip, lr_flip;
+  get_flip_cfg(tx_type, &ud_flip, &lr_flip);
+  load_buffer_32bit_to_16bit_w4(input, txfm_size_col, buf, txfm_size_row);
+  transpose_16bit_4x4(buf, buf);
+  row_txfm(buf, buf, cos_bit_row);
+  if (lr_flip) {
+    __m128i temp[4];
+    flip_buf_sse2(buf, temp, txfm_size_col);
+    transpose_16bit_4x4(temp, buf);
+  } else {
+    transpose_16bit_4x4(buf, buf);
+  }
+  col_txfm(buf, buf, cos_bit_col);
+  round_shift_16bit_ssse3(buf, txfm_size_row, shift[1]);
+  lowbd_write_buffer_4xn_sse2(buf, output, stride, ud_flip, txfm_size_row);
+}
+
+static INLINE __m128i lowbd_get_recon_16x16_sse2(const __m128i pred,
+                                                 __m128i res0, __m128i res1) {
+  const __m128i zero = _mm_setzero_si128();
+  __m128i x0 = _mm_unpacklo_epi8(pred, zero);
+  __m128i x1 = _mm_unpackhi_epi8(pred, zero);
+  x0 = _mm_adds_epi16(res0, x0);
+  x1 = _mm_adds_epi16(res1, x1);
+  return _mm_packus_epi16(x0, x1);
+}
+
+static INLINE void lowbd_write_buffer_16xn_sse2(__m128i *in, uint8_t *output,
+                                                int stride, int flipud,
+                                                int height) {
+  int j = flipud ? (height - 1) : 0;
+  const int step = flipud ? -1 : 1;
+  for (int i = 0; i < height; ++i, j += step) {
+    __m128i v = _mm_loadu_si128((__m128i const *)(output + i * stride));
+    __m128i u = lowbd_get_recon_16x16_sse2(v, in[j], in[j + height]);
+    _mm_storeu_si128((__m128i *)(output + i * stride), u);
+  }
+}
+
+static INLINE void round_shift_ssse3(const __m128i *input, __m128i *output,
+                                     int size) {
+  const __m128i scale = _mm_set1_epi16(NewInvSqrt2 * 8);
+  for (int i = 0; i < size; ++i) {
+    output[i] = _mm_mulhrs_epi16(input[i], scale);
+  }
+}
+
+static INLINE void lowbd_inv_txfm2d_add_no_identity_ssse3(
+    const int32_t *input, uint8_t *output, int stride, TX_TYPE tx_type,
+    TX_SIZE tx_size, int eob) {
+  __m128i buf1[64 * 8];
+  int eobx, eoby;
+  get_eobx_eoby_scan_default(&eobx, &eoby, tx_size, eob);
+  const int8_t *shift = inv_txfm_shift_ls[tx_size];
+  const int txw_idx = get_txw_idx(tx_size);
+  const int txh_idx = get_txh_idx(tx_size);
+  const int cos_bit_col = inv_cos_bit_col[txw_idx][txh_idx];
+  const int cos_bit_row = inv_cos_bit_row[txw_idx][txh_idx];
+  const int txfm_size_col = tx_size_wide[tx_size];
+  const int txfm_size_row = tx_size_high[tx_size];
+  const int buf_size_w_div8 = txfm_size_col >> 3;
+  const int buf_size_nonzero_w_div8 = (eobx + 8) >> 3;
+  const int buf_size_nonzero_h_div8 = (eoby + 8) >> 3;
+  const int input_stride = AOMMIN(32, txfm_size_col);
+  const int rect_type = get_rect_tx_log_ratio(txfm_size_col, txfm_size_row);
+
+  const int fun_idx_x = lowbd_txfm_all_1d_zeros_idx[eobx];
+  const int fun_idx_y = lowbd_txfm_all_1d_zeros_idx[eoby];
+  const transform_1d_ssse3 row_txfm =
+      lowbd_txfm_all_1d_zeros_w8_arr[txw_idx][hitx_1d_tab[tx_type]][fun_idx_x];
+  const transform_1d_ssse3 col_txfm =
+      lowbd_txfm_all_1d_zeros_w8_arr[txh_idx][vitx_1d_tab[tx_type]][fun_idx_y];
+
+  assert(col_txfm != NULL);
+  assert(row_txfm != NULL);
+  int ud_flip, lr_flip;
+  get_flip_cfg(tx_type, &ud_flip, &lr_flip);
+  for (int i = 0; i < buf_size_nonzero_h_div8; i++) {
+    __m128i buf0[64];
+    const int32_t *input_row = input + i * input_stride * 8;
+    for (int j = 0; j < buf_size_nonzero_w_div8; ++j) {
+      __m128i *buf0_cur = buf0 + j * 8;
+      load_buffer_32bit_to_16bit(input_row + j * 8, input_stride, buf0_cur, 8);
+      transpose_16bit_8x8(buf0_cur, buf0_cur);
+    }
+    if (rect_type == 1 || rect_type == -1) {
+      round_shift_ssse3(buf0, buf0, input_stride);  // rect special code
+    }
+    row_txfm(buf0, buf0, cos_bit_row);
+    round_shift_16bit_ssse3(buf0, txfm_size_col, shift[0]);
+    __m128i *_buf1 = buf1 + i * 8;
+    if (lr_flip) {
+      for (int j = 0; j < buf_size_w_div8; ++j) {
+        __m128i temp[8];
+        flip_buf_sse2(buf0 + 8 * j, temp, 8);
+        transpose_16bit_8x8(temp,
+                            _buf1 + txfm_size_row * (buf_size_w_div8 - 1 - j));
+      }
+    } else {
+      for (int j = 0; j < buf_size_w_div8; ++j) {
+        transpose_16bit_8x8(buf0 + 8 * j, _buf1 + txfm_size_row * j);
+      }
+    }
+  }
+  for (int i = 0; i < buf_size_w_div8; i++) {
+    col_txfm(buf1 + i * txfm_size_row, buf1 + i * txfm_size_row, cos_bit_col);
+    round_shift_16bit_ssse3(buf1 + i * txfm_size_row, txfm_size_row, shift[1]);
+  }
+
+  if (txfm_size_col >= 16) {
+    for (int i = 0; i < (txfm_size_col >> 4); i++) {
+      lowbd_write_buffer_16xn_sse2(buf1 + i * txfm_size_row * 2,
+                                   output + 16 * i, stride, ud_flip,
+                                   txfm_size_row);
+    }
+  } else if (txfm_size_col == 8) {
+    lowbd_write_buffer_8xn_sse2(buf1, output, stride, ud_flip, txfm_size_row);
+  }
+}
+
+static INLINE void lowbd_inv_txfm2d_add_h_identity_ssse3(
+    const int32_t *input, uint8_t *output, int stride, TX_TYPE tx_type,
+    TX_SIZE tx_size, int eob) {
+  const int8_t *shift = inv_txfm_shift_ls[tx_size];
+  int eobx, eoby;
+  get_eobx_eoby_scan_h_identity(&eobx, &eoby, tx_size, eob);
+  const int txw_idx = get_txw_idx(tx_size);
+  const int txh_idx = get_txh_idx(tx_size);
+  const int cos_bit_col = inv_cos_bit_col[txw_idx][txh_idx];
+  const int txfm_size_col = tx_size_wide[tx_size];
+  const int txfm_size_row = tx_size_high[tx_size];
+  const int buf_size_w_div8 = (eobx + 8) >> 3;
+  const int input_stride = AOMMIN(32, txfm_size_col);
+  const int rect_type = get_rect_tx_log_ratio(txfm_size_col, txfm_size_row);
+
+  const int fun_idx = lowbd_txfm_all_1d_zeros_idx[eoby];
+  assert(fun_idx < 5);
+  const transform_1d_ssse3 col_txfm =
+      lowbd_txfm_all_1d_zeros_w8_arr[txh_idx][vitx_1d_tab[tx_type]][fun_idx];
+
+  assert(col_txfm != NULL);
+
+  int ud_flip, lr_flip;
+  get_flip_cfg(tx_type, &ud_flip, &lr_flip);
+  for (int i = 0; i < buf_size_w_div8; i++) {
+    __m128i buf0[64];
+    iidentity_row_8xn_ssse3(buf0, input + 8 * i, input_stride, shift[0],
+                            eoby + 1, txw_idx, rect_type);
+    col_txfm(buf0, buf0, cos_bit_col);
+    __m128i mshift = _mm_set1_epi16(1 << (15 + shift[1]));
+    int k = ud_flip ? (txfm_size_row - 1) : 0;
+    const int step = ud_flip ? -1 : 1;
+    uint8_t *out = output + 8 * i;
+    for (int j = 0; j < txfm_size_row; ++j, k += step) {
+      const __m128i v = _mm_loadl_epi64((__m128i const *)(out));
+      __m128i res = _mm_mulhrs_epi16(buf0[k], mshift);
+      const __m128i u = lowbd_get_recon_8x8_sse2(v, res);
+      _mm_storel_epi64((__m128i *)(out), u);
+      out += stride;
+    }
+  }
+}
+
+static INLINE void lowbd_inv_txfm2d_add_v_identity_ssse3(
+    const int32_t *input, uint8_t *output, int stride, TX_TYPE tx_type,
+    TX_SIZE tx_size, int eob) {
+  __m128i buf1[64];
+  int eobx, eoby;
+  get_eobx_eoby_scan_v_identity(&eobx, &eoby, tx_size, eob);
+  const int8_t *shift = inv_txfm_shift_ls[tx_size];
+  const int txw_idx = get_txw_idx(tx_size);
+  const int txh_idx = get_txh_idx(tx_size);
+  const int cos_bit_row = inv_cos_bit_row[txw_idx][txh_idx];
+  const int txfm_size_col = tx_size_wide[tx_size];
+  const int txfm_size_row = tx_size_high[tx_size];
+  const int buf_size_w_div8 = txfm_size_col >> 3;
+  const int buf_size_h_div8 = (eoby + 8) >> 3;
+  const int input_stride = AOMMIN(32, txfm_size_col);
+  const int rect_type = get_rect_tx_log_ratio(txfm_size_col, txfm_size_row);
+
+  const int fun_idx = lowbd_txfm_all_1d_zeros_idx[eobx];
+  const transform_1d_ssse3 row_txfm =
+      lowbd_txfm_all_1d_zeros_w8_arr[txw_idx][hitx_1d_tab[tx_type]][fun_idx];
+
+  assert(row_txfm != NULL);
+  int ud_flip, lr_flip;
+  get_flip_cfg(tx_type, &ud_flip, &lr_flip);
+  for (int i = 0; i < buf_size_h_div8; i++) {
+    __m128i buf0[64];
+    const int32_t *input_row = input + i * input_stride * 8;
+    for (int j = 0; j < AOMMIN(4, buf_size_w_div8); ++j) {
+      __m128i *buf0_cur = buf0 + j * 8;
+      load_buffer_32bit_to_16bit(input_row + j * 8, input_stride, buf0_cur, 8);
+      transpose_16bit_8x8(buf0_cur, buf0_cur);
+    }
+    if (rect_type == 1 || rect_type == -1) {
+      round_shift_ssse3(buf0, buf0, input_stride);  // rect special code
+    }
+    row_txfm(buf0, buf0, cos_bit_row);
+    round_shift_16bit_ssse3(buf0, txfm_size_col, shift[0]);
+    __m128i *_buf1 = buf1;
+    if (lr_flip) {
+      for (int j = 0; j < buf_size_w_div8; ++j) {
+        __m128i temp[8];
+        flip_buf_sse2(buf0 + 8 * j, temp, 8);
+        transpose_16bit_8x8(temp, _buf1 + 8 * (buf_size_w_div8 - 1 - j));
+      }
+    } else {
+      for (int j = 0; j < buf_size_w_div8; ++j) {
+        transpose_16bit_8x8(buf0 + 8 * j, _buf1 + 8 * j);
+      }
+    }
+
+    for (int j = 0; j < buf_size_w_div8; ++j) {
+      iidentity_col_8xn_ssse3(output + i * 8 * stride + j * 8, stride,
+                              buf1 + j * 8, shift[1], 8, txh_idx);
+    }
+  }
+}
+
+// for 32x32,32x64,64x32,64x64,32x8,8x32,16x32,32x16,64x16,16x64
+static INLINE void lowbd_inv_txfm2d_add_universe_ssse3(
+    const int32_t *input, uint8_t *output, int stride, TX_TYPE tx_type,
+    TX_SIZE tx_size, int eob) {
+  switch (tx_type) {
+    case DCT_DCT:
+      lowbd_inv_txfm2d_add_no_identity_ssse3(input, output, stride, tx_type,
+                                             tx_size, eob);
+      break;
+    case IDTX:
+      lowbd_inv_txfm2d_add_idtx_ssse3(input, output, stride, tx_size);
+      break;
+    case V_DCT:
+    case V_ADST:
+    case V_FLIPADST:
+      lowbd_inv_txfm2d_add_h_identity_ssse3(input, output, stride, tx_type,
+                                            tx_size, eob);
+      break;
+    case H_DCT:
+    case H_ADST:
+    case H_FLIPADST:
+      lowbd_inv_txfm2d_add_v_identity_ssse3(input, output, stride, tx_type,
+                                            tx_size, eob);
+      break;
+    default:
+      lowbd_inv_txfm2d_add_no_identity_ssse3(input, output, stride, tx_type,
+                                             tx_size, eob);
+      break;
+  }
+}
+
+void lowbd_inv_txfm2d_add_4x8_ssse3(const int32_t *input, uint8_t *output,
+                                    int stride, TX_TYPE tx_type,
+                                    TX_SIZE tx_size_, int eob) {
+  (void)tx_size_;
+  (void)eob;
+  __m128i buf[8];
+  const TX_SIZE tx_size = TX_4X8;
+  const int8_t *shift = inv_txfm_shift_ls[tx_size];
+  const int txw_idx = get_txw_idx(tx_size);
+  const int txh_idx = get_txh_idx(tx_size);
+  const int cos_bit_row = inv_cos_bit_row[txw_idx][txh_idx];
+  const int cos_bit_col = inv_cos_bit_col[txw_idx][txh_idx];
+  const int txfm_size_col = tx_size_wide[tx_size];
+  const int txfm_size_row = tx_size_high[tx_size];
+
+  const transform_1d_ssse3 row_txfm =
+      lowbd_txfm_all_1d_w8_arr[txw_idx][hitx_1d_tab[tx_type]];
+  const transform_1d_ssse3 col_txfm =
+      lowbd_txfm_all_1d_w4_arr[txh_idx][vitx_1d_tab[tx_type]];
+
+  int ud_flip, lr_flip;
+  get_flip_cfg(tx_type, &ud_flip, &lr_flip);
+  load_buffer_32bit_to_16bit_w4(input, txfm_size_col, buf, txfm_size_row);
+  transpose_16bit_4x8(buf, buf);
+  round_shift_ssse3(buf, buf, txfm_size_col);  // rect special code
+  row_txfm(buf, buf, cos_bit_row);
+  // round_shift_16bit_ssse3(buf, txfm_size_col, shift[0]);// shift[0] is 0
+  if (lr_flip) {
+    __m128i temp[4];
+    flip_buf_sse2(buf, temp, txfm_size_col);
+    transpose_16bit_8x4(temp, buf);
+  } else {
+    transpose_16bit_8x4(buf, buf);
+  }
+  col_txfm(buf, buf, cos_bit_col);
+  round_shift_16bit_ssse3(buf, txfm_size_row, shift[1]);
+  lowbd_write_buffer_4xn_sse2(buf, output, stride, ud_flip, txfm_size_row);
+}
+
+void lowbd_inv_txfm2d_add_8x4_ssse3(const int32_t *input, uint8_t *output,
+                                    int stride, TX_TYPE tx_type,
+                                    TX_SIZE tx_size_, int eob) {
+  (void)tx_size_;
+  (void)eob;
+  __m128i buf[8];
+  const TX_SIZE tx_size = TX_8X4;
+  const int8_t *shift = inv_txfm_shift_ls[tx_size];
+  const int txw_idx = get_txw_idx(tx_size);
+  const int txh_idx = get_txh_idx(tx_size);
+  const int cos_bit_row = inv_cos_bit_row[txw_idx][txh_idx];
+  const int cos_bit_col = inv_cos_bit_col[txw_idx][txh_idx];
+  const int txfm_size_col = tx_size_wide[tx_size];
+  const int txfm_size_row = tx_size_high[tx_size];
+
+  const transform_1d_ssse3 row_txfm =
+      lowbd_txfm_all_1d_w4_arr[txw_idx][hitx_1d_tab[tx_type]];
+  const transform_1d_ssse3 col_txfm =
+      lowbd_txfm_all_1d_w8_arr[txh_idx][vitx_1d_tab[tx_type]];
+
+  int ud_flip, lr_flip;
+  get_flip_cfg(tx_type, &ud_flip, &lr_flip);
+  load_buffer_32bit_to_16bit(input, txfm_size_col, buf, txfm_size_row);
+  transpose_16bit_8x4(buf, buf);
+  round_shift_ssse3(buf, buf, txfm_size_col);  // rect special code
+  row_txfm(buf, buf, cos_bit_row);
+  // round_shift_16bit_ssse3(buf, txfm_size_col, shift[0]); // shift[0] is 0
+  if (lr_flip) {
+    __m128i temp[8];
+    flip_buf_sse2(buf, temp, txfm_size_col);
+    transpose_16bit_4x8(temp, buf);
+  } else {
+    transpose_16bit_4x8(buf, buf);
+  }
+  col_txfm(buf, buf, cos_bit_col);
+  round_shift_16bit_ssse3(buf, txfm_size_row, shift[1]);
+  lowbd_write_buffer_8xn_sse2(buf, output, stride, ud_flip, txfm_size_row);
+}
+
+void lowbd_inv_txfm2d_add_4x16_ssse3(const int32_t *input, uint8_t *output,
+                                     int stride, TX_TYPE tx_type,
+                                     TX_SIZE tx_size_, int eob) {
+  (void)tx_size_;
+  (void)eob;
+  __m128i buf[16];
+  const TX_SIZE tx_size = TX_4X16;
+  const int8_t *shift = inv_txfm_shift_ls[tx_size];
+  const int txw_idx = get_txw_idx(tx_size);
+  const int txh_idx = get_txh_idx(tx_size);
+  const int cos_bit_row = inv_cos_bit_row[txw_idx][txh_idx];
+  const int cos_bit_col = inv_cos_bit_col[txw_idx][txh_idx];
+  const int txfm_size_col = tx_size_wide[tx_size];
+  const int txfm_size_row = tx_size_high[tx_size];
+
+  const transform_1d_ssse3 row_txfm =
+      lowbd_txfm_all_1d_w8_arr[txw_idx][hitx_1d_tab[tx_type]];
+  const transform_1d_ssse3 col_txfm =
+      lowbd_txfm_all_1d_w4_arr[txh_idx][vitx_1d_tab[tx_type]];
+
+  int ud_flip, lr_flip;
+  get_flip_cfg(tx_type, &ud_flip, &lr_flip);
+
+  const int row_one_loop = 8;
+  for (int i = 0; i < 2; ++i) {
+    const int32_t *input_cur = input + i * txfm_size_col * row_one_loop;
+    __m128i *buf_cur = buf + i * row_one_loop;
+    load_buffer_32bit_to_16bit_w4(input_cur, txfm_size_col, buf_cur,
+                                  row_one_loop);
+    transpose_16bit_4x8(buf_cur, buf_cur);
+    row_txfm(buf_cur, buf_cur, cos_bit_row);
+    round_shift_16bit_ssse3(buf_cur, row_one_loop, shift[0]);
+    if (lr_flip) {
+      __m128i temp[8];
+      flip_buf_sse2(buf_cur, temp, txfm_size_col);
+      transpose_16bit_8x4(temp, buf_cur);
+    } else {
+      transpose_16bit_8x4(buf_cur, buf_cur);
+    }
+  }
+  col_txfm(buf, buf, cos_bit_col);
+  round_shift_16bit_ssse3(buf, txfm_size_row, shift[1]);
+  lowbd_write_buffer_4xn_sse2(buf, output, stride, ud_flip, txfm_size_row);
+}
+
+void lowbd_inv_txfm2d_add_16x4_ssse3(const int32_t *input, uint8_t *output,
+                                     int stride, TX_TYPE tx_type,
+                                     TX_SIZE tx_size_, int eob) {
+  (void)tx_size_;
+  (void)eob;
+  __m128i buf[16];
+  const TX_SIZE tx_size = TX_16X4;
+  const int8_t *shift = inv_txfm_shift_ls[tx_size];
+  const int txw_idx = get_txw_idx(tx_size);
+  const int txh_idx = get_txh_idx(tx_size);
+  const int cos_bit_row = inv_cos_bit_row[txw_idx][txh_idx];
+  const int cos_bit_col = inv_cos_bit_col[txw_idx][txh_idx];
+  const int txfm_size_col = tx_size_wide[tx_size];
+  const int txfm_size_row = tx_size_high[tx_size];
+  const int buf_size_w_div8 = txfm_size_col >> 3;
+
+  const transform_1d_ssse3 row_txfm =
+      lowbd_txfm_all_1d_w4_arr[txw_idx][hitx_1d_tab[tx_type]];
+  const transform_1d_ssse3 col_txfm =
+      lowbd_txfm_all_1d_w8_arr[txh_idx][vitx_1d_tab[tx_type]];
+
+  int ud_flip, lr_flip;
+  get_flip_cfg(tx_type, &ud_flip, &lr_flip);
+  const int row_one_loop = 8;
+  for (int i = 0; i < buf_size_w_div8; ++i) {
+    const int32_t *input_cur = input + i * row_one_loop;
+    __m128i *buf_cur = buf + i * row_one_loop;
+    load_buffer_32bit_to_16bit(input_cur, txfm_size_col, buf_cur,
+                               txfm_size_row);
+    transpose_16bit_8x4(buf_cur, buf_cur);
+  }
+  row_txfm(buf, buf, cos_bit_row);
+  round_shift_16bit_ssse3(buf, txfm_size_col, shift[0]);
+  if (lr_flip) {
+    __m128i temp[16];
+    flip_buf_sse2(buf, temp, 16);
+    transpose_16bit_4x8(temp, buf);
+    transpose_16bit_4x8(temp + 8, buf + 8);
+  } else {
+    transpose_16bit_4x8(buf, buf);
+    transpose_16bit_4x8(buf + row_one_loop, buf + row_one_loop);
+  }
+  for (int i = 0; i < buf_size_w_div8; i++) {
+    col_txfm(buf + i * row_one_loop, buf + i * row_one_loop, cos_bit_col);
+    round_shift_16bit_ssse3(buf + i * row_one_loop, txfm_size_row, shift[1]);
+  }
+  lowbd_write_buffer_8xn_sse2(buf, output, stride, ud_flip, 4);
+  lowbd_write_buffer_8xn_sse2(buf + 8, output + 8, stride, ud_flip, 4);
+}
+
+void av1_lowbd_inv_txfm2d_add_ssse3(const int32_t *input, uint8_t *output,
+                                    int stride, TX_TYPE tx_type,
+                                    TX_SIZE tx_size, int eob) {
+  switch (tx_size) {
+    case TX_4X4:
+      lowbd_inv_txfm2d_add_4x4_ssse3(input, output, stride, tx_type, tx_size,
+                                     eob);
+      break;
+    case TX_4X8:
+      lowbd_inv_txfm2d_add_4x8_ssse3(input, output, stride, tx_type, tx_size,
+                                     eob);
+      break;
+    case TX_8X4:
+      lowbd_inv_txfm2d_add_8x4_ssse3(input, output, stride, tx_type, tx_size,
+                                     eob);
+      break;
+    case TX_4X16:
+      lowbd_inv_txfm2d_add_4x16_ssse3(input, output, stride, tx_type, tx_size,
+                                      eob);
+      break;
+    case TX_16X4:
+      lowbd_inv_txfm2d_add_16x4_ssse3(input, output, stride, tx_type, tx_size,
+                                      eob);
+      break;
+    default:
+      lowbd_inv_txfm2d_add_universe_ssse3(input, output, stride, tx_type,
+                                          tx_size, eob);
+      break;
+  }
+}
+void av1_inv_txfm_add_ssse3(const tran_low_t *dqcoeff, uint8_t *dst, int stride,
+                            const TxfmParam *txfm_param) {
+  const TX_TYPE tx_type = txfm_param->tx_type;
+  if (!txfm_param->lossless) {
+    av1_lowbd_inv_txfm2d_add_ssse3(dqcoeff, dst, stride, tx_type,
+                                   txfm_param->tx_size, txfm_param->eob);
+  } else {
+    av1_inv_txfm_add_c(dqcoeff, dst, stride, txfm_param);
+  }
+}
diff --git a/third_party/aom/av1/common/x86/av1_inv_txfm_ssse3.h b/third_party/aom/av1/common/x86/av1_inv_txfm_ssse3.h
new file mode 100644
index 000000000..66bd339d1
--- /dev/null
+++ b/third_party/aom/av1/common/x86/av1_inv_txfm_ssse3.h
@@ -0,0 +1,232 @@
+/*
+ * 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_AV1_COMMON_X86_AV1_INV_TXFM_SSSE3_H_
+#define AOM_AV1_COMMON_X86_AV1_INV_TXFM_SSSE3_H_
+
+#include <emmintrin.h>  // SSE2
+#include <tmmintrin.h>  // SSSE3
+
+#include "config/aom_config.h"
+#include "config/av1_rtcd.h"
+
+#include "aom/aom_integer.h"
+#include "aom_dsp/x86/transpose_sse2.h"
+#include "aom_dsp/x86/txfm_common_sse2.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define btf_16_ssse3(w0, w1, in, out0, out1)    \
+  do {                                          \
+    const __m128i _w0 = _mm_set1_epi16(w0 * 8); \
+    const __m128i _w1 = _mm_set1_epi16(w1 * 8); \
+    const __m128i _in = in;                     \
+    out0 = _mm_mulhrs_epi16(_in, _w0);          \
+    out1 = _mm_mulhrs_epi16(_in, _w1);          \
+  } while (0)
+
+#define btf_16_adds_subs_sse2(in0, in1) \
+  do {                                  \
+    const __m128i _in0 = in0;           \
+    const __m128i _in1 = in1;           \
+    in0 = _mm_adds_epi16(_in0, _in1);   \
+    in1 = _mm_subs_epi16(_in0, _in1);   \
+  } while (0)
+
+#define btf_16_subs_adds_sse2(in0, in1) \
+  do {                                  \
+    const __m128i _in0 = in0;           \
+    const __m128i _in1 = in1;           \
+    in1 = _mm_subs_epi16(_in0, _in1);   \
+    in0 = _mm_adds_epi16(_in0, _in1);   \
+  } while (0)
+
+#define btf_16_adds_subs_out_sse2(out0, out1, in0, in1) \
+  do {                                                  \
+    const __m128i _in0 = in0;                           \
+    const __m128i _in1 = in1;                           \
+    out0 = _mm_adds_epi16(_in0, _in1);                  \
+    out1 = _mm_subs_epi16(_in0, _in1);                  \
+  } while (0)
+
+static INLINE void round_shift_16bit_ssse3(__m128i *in, int size, int bit) {
+  if (bit < 0) {
+    const __m128i scale = _mm_set1_epi16(1 << (15 + bit));
+    for (int i = 0; i < size; ++i) {
+      in[i] = _mm_mulhrs_epi16(in[i], scale);
+    }
+  } else if (bit > 0) {
+    for (int i = 0; i < size; ++i) {
+      in[i] = _mm_slli_epi16(in[i], bit);
+    }
+  }
+}
+
+// 1D itx types
+typedef enum ATTRIBUTE_PACKED {
+  IDCT_1D,
+  IADST_1D,
+  IFLIPADST_1D = IADST_1D,
+  IIDENTITY_1D,
+  ITX_TYPES_1D,
+} ITX_TYPE_1D;
+
+static const ITX_TYPE_1D vitx_1d_tab[TX_TYPES] = {
+  IDCT_1D,      IADST_1D,     IDCT_1D,      IADST_1D,
+  IFLIPADST_1D, IDCT_1D,      IFLIPADST_1D, IADST_1D,
+  IFLIPADST_1D, IIDENTITY_1D, IDCT_1D,      IIDENTITY_1D,
+  IADST_1D,     IIDENTITY_1D, IFLIPADST_1D, IIDENTITY_1D,
+};
+
+static const ITX_TYPE_1D hitx_1d_tab[TX_TYPES] = {
+  IDCT_1D,      IDCT_1D,      IADST_1D,     IADST_1D,
+  IDCT_1D,      IFLIPADST_1D, IFLIPADST_1D, IFLIPADST_1D,
+  IADST_1D,     IIDENTITY_1D, IIDENTITY_1D, IDCT_1D,
+  IIDENTITY_1D, IADST_1D,     IIDENTITY_1D, IFLIPADST_1D,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_eob_to_eobxy_8x8_default[8]) = {
+  0x0707, 0x0707, 0x0707, 0x0707, 0x0707, 0x0707, 0x0707, 0x0707,
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                av1_eob_to_eobxy_16x16_default[16]) = {
+  0x0707, 0x0707, 0x0f0f, 0x0f0f, 0x0f0f, 0x0f0f, 0x0f0f, 0x0f0f,
+  0x0f0f, 0x0f0f, 0x0f0f, 0x0f0f, 0x0f0f, 0x0f0f, 0x0f0f, 0x0f0f,
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                av1_eob_to_eobxy_32x32_default[32]) = {
+  0x0707, 0x0f0f, 0x0f0f, 0x0f0f, 0x1f1f, 0x1f1f, 0x1f1f, 0x1f1f,
+  0x1f1f, 0x1f1f, 0x1f1f, 0x1f1f, 0x1f1f, 0x1f1f, 0x1f1f, 0x1f1f,
+  0x1f1f, 0x1f1f, 0x1f1f, 0x1f1f, 0x1f1f, 0x1f1f, 0x1f1f, 0x1f1f,
+  0x1f1f, 0x1f1f, 0x1f1f, 0x1f1f, 0x1f1f, 0x1f1f, 0x1f1f, 0x1f1f,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_eob_to_eobxy_8x16_default[16]) = {
+  0x0707, 0x0707, 0x0707, 0x0707, 0x0707, 0x0f07, 0x0f07, 0x0f07,
+  0x0f07, 0x0f07, 0x0f07, 0x0f07, 0x0f07, 0x0f07, 0x0f07, 0x0f07,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_eob_to_eobxy_16x8_default[8]) = {
+  0x0707, 0x0707, 0x070f, 0x070f, 0x070f, 0x070f, 0x070f, 0x070f,
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                av1_eob_to_eobxy_16x32_default[32]) = {
+  0x0707, 0x0707, 0x0f0f, 0x0f0f, 0x0f0f, 0x0f0f, 0x0f0f, 0x0f0f,
+  0x0f0f, 0x1f0f, 0x1f0f, 0x1f0f, 0x1f0f, 0x1f0f, 0x1f0f, 0x1f0f,
+  0x1f0f, 0x1f0f, 0x1f0f, 0x1f0f, 0x1f0f, 0x1f0f, 0x1f0f, 0x1f0f,
+  0x1f0f, 0x1f0f, 0x1f0f, 0x1f0f, 0x1f0f, 0x1f0f, 0x1f0f, 0x1f0f,
+};
+
+DECLARE_ALIGNED(16, static const int16_t,
+                av1_eob_to_eobxy_32x16_default[16]) = {
+  0x0707, 0x0f0f, 0x0f0f, 0x0f0f, 0x0f1f, 0x0f1f, 0x0f1f, 0x0f1f,
+  0x0f1f, 0x0f1f, 0x0f1f, 0x0f1f, 0x0f1f, 0x0f1f, 0x0f1f, 0x0f1f,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_eob_to_eobxy_8x32_default[32]) = {
+  0x0707, 0x0707, 0x0707, 0x0707, 0x0707, 0x0f07, 0x0f07, 0x0f07,
+  0x0f07, 0x0f07, 0x0f07, 0x0f07, 0x0f07, 0x1f07, 0x1f07, 0x1f07,
+  0x1f07, 0x1f07, 0x1f07, 0x1f07, 0x1f07, 0x1f07, 0x1f07, 0x1f07,
+  0x1f07, 0x1f07, 0x1f07, 0x1f07, 0x1f07, 0x1f07, 0x1f07, 0x1f07,
+};
+
+DECLARE_ALIGNED(16, static const int16_t, av1_eob_to_eobxy_32x8_default[8]) = {
+  0x0707, 0x070f, 0x070f, 0x071f, 0x071f, 0x071f, 0x071f, 0x071f,
+};
+
+DECLARE_ALIGNED(16, static const int16_t *,
+                av1_eob_to_eobxy_default[TX_SIZES_ALL]) = {
+  NULL,
+  av1_eob_to_eobxy_8x8_default,
+  av1_eob_to_eobxy_16x16_default,
+  av1_eob_to_eobxy_32x32_default,
+  av1_eob_to_eobxy_32x32_default,
+  NULL,
+  NULL,
+  av1_eob_to_eobxy_8x16_default,
+  av1_eob_to_eobxy_16x8_default,
+  av1_eob_to_eobxy_16x32_default,
+  av1_eob_to_eobxy_32x16_default,
+  av1_eob_to_eobxy_32x32_default,
+  av1_eob_to_eobxy_32x32_default,
+  NULL,
+  NULL,
+  av1_eob_to_eobxy_8x32_default,
+  av1_eob_to_eobxy_32x8_default,
+  av1_eob_to_eobxy_16x32_default,
+  av1_eob_to_eobxy_32x16_default,
+};
+
+static const int lowbd_txfm_all_1d_zeros_idx[32] = {
+  0, 1, 1, 1, 1, 1, 1, 1, 2, 2, 2, 2, 2, 2, 2, 2,
+  3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3,
+};
+
+// Transform block width in log2 for eob (size of 64 map to 32)
+static const int tx_size_wide_log2_eob[TX_SIZES_ALL] = {
+  2, 3, 4, 5, 5, 2, 3, 3, 4, 4, 5, 5, 5, 2, 4, 3, 5, 4, 5,
+};
+
+static INLINE void get_eobx_eoby_scan_default(int *eobx, int *eoby,
+                                              TX_SIZE tx_size, int eob) {
+  if (eob == 1) {
+    *eobx = 0;
+    *eoby = 0;
+    return;
+  }
+
+  const int tx_w_log2 = tx_size_wide_log2_eob[tx_size];
+  const int eob_row = (eob - 1) >> tx_w_log2;
+  const int eobxy = av1_eob_to_eobxy_default[tx_size][eob_row];
+  *eobx = eobxy & 0xFF;
+  *eoby = eobxy >> 8;
+}
+
+static int eob_fill[32] = {
+  0,  7,  7,  7,  7,  7,  7,  7,  15, 15, 15, 15, 15, 15, 15, 15,
+  31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31, 31,
+};
+
+static INLINE void get_eobx_eoby_scan_h_identity(int *eobx, int *eoby,
+                                                 TX_SIZE tx_size, int eob) {
+  eob -= 1;
+  const int txfm_size_col = tx_size_wide[tx_size];
+  const int eobx_max = AOMMIN(32, txfm_size_col) - 1;
+  *eobx = (eob >= eobx_max) ? eobx_max : eob_fill[eob];
+  const int temp_eoby = eob / (eobx_max + 1);
+  assert(temp_eoby < 32);
+  *eoby = eob_fill[temp_eoby];
+}
+
+static INLINE void get_eobx_eoby_scan_v_identity(int *eobx, int *eoby,
+                                                 TX_SIZE tx_size, int eob) {
+  eob -= 1;
+  const int txfm_size_row = tx_size_high[tx_size];
+  const int eoby_max = AOMMIN(32, txfm_size_row) - 1;
+  *eobx = eob / (eoby_max + 1);
+  *eoby = (eob >= eoby_max) ? eoby_max : eob_fill[eob];
+}
+
+typedef void (*transform_1d_ssse3)(const __m128i *input, __m128i *output,
+                                   int8_t cos_bit);
+
+void av1_lowbd_inv_txfm2d_add_ssse3(const int32_t *input, uint8_t *output,
+                                    int stride, TX_TYPE tx_type,
+                                    TX_SIZE tx_size, int eob);
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // AOM_AV1_COMMON_X86_AV1_INV_TXFM_SSSE3_H_
diff --git a/third_party/aom/av1/common/x86/av1_txfm_sse2.h b/third_party/aom/av1/common/x86/av1_txfm_sse2.h
new file mode 100644
index 000000000..77aeb6eb1
--- /dev/null
+++ b/third_party/aom/av1/common/x86/av1_txfm_sse2.h
@@ -0,0 +1,317 @@
+/*
+ * 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_AV1_COMMON_X86_AV1_TXFM_SSE2_H_
+#define AOM_AV1_COMMON_X86_AV1_TXFM_SSE2_H_
+
+#include <emmintrin.h>  // SSE2
+
+#include "config/aom_config.h"
+#include "config/av1_rtcd.h"
+
+#include "aom/aom_integer.h"
+#include "aom_dsp/x86/transpose_sse2.h"
+#include "aom_dsp/x86/txfm_common_sse2.h"
+#include "av1/common/av1_txfm.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+static INLINE void btf_16_w4_sse2(
+    const __m128i *const w0, const __m128i *const w1, const __m128i __rounding,
+    const int8_t cos_bit, const __m128i *const in0, const __m128i *const in1,
+    __m128i *const out0, __m128i *const out1) {
+  const __m128i t0 = _mm_unpacklo_epi16(*in0, *in1);
+  const __m128i u0 = _mm_madd_epi16(t0, *w0);
+  const __m128i v0 = _mm_madd_epi16(t0, *w1);
+  const __m128i a0 = _mm_add_epi32(u0, __rounding);
+  const __m128i b0 = _mm_add_epi32(v0, __rounding);
+  const __m128i c0 = _mm_srai_epi32(a0, cos_bit);
+  const __m128i d0 = _mm_srai_epi32(b0, cos_bit);
+
+  *out0 = _mm_packs_epi32(c0, c0);
+  *out1 = _mm_packs_epi32(d0, c0);
+}
+
+#define btf_16_4p_sse2(w0, w1, in0, in1, out0, out1) \
+  {                                                  \
+    __m128i t0 = _mm_unpacklo_epi16(in0, in1);       \
+    __m128i u0 = _mm_madd_epi16(t0, w0);             \
+    __m128i v0 = _mm_madd_epi16(t0, w1);             \
+                                                     \
+    __m128i a0 = _mm_add_epi32(u0, __rounding);      \
+    __m128i b0 = _mm_add_epi32(v0, __rounding);      \
+                                                     \
+    __m128i c0 = _mm_srai_epi32(a0, cos_bit);        \
+    __m128i d0 = _mm_srai_epi32(b0, cos_bit);        \
+                                                     \
+    out0 = _mm_packs_epi32(c0, c0);                  \
+    out1 = _mm_packs_epi32(d0, d0);                  \
+  }
+
+#define btf_16_sse2(w0, w1, in0, in1, out0, out1) \
+  {                                               \
+    __m128i t0 = _mm_unpacklo_epi16(in0, in1);    \
+    __m128i t1 = _mm_unpackhi_epi16(in0, in1);    \
+    __m128i u0 = _mm_madd_epi16(t0, w0);          \
+    __m128i u1 = _mm_madd_epi16(t1, w0);          \
+    __m128i v0 = _mm_madd_epi16(t0, w1);          \
+    __m128i v1 = _mm_madd_epi16(t1, w1);          \
+                                                  \
+    __m128i a0 = _mm_add_epi32(u0, __rounding);   \
+    __m128i a1 = _mm_add_epi32(u1, __rounding);   \
+    __m128i b0 = _mm_add_epi32(v0, __rounding);   \
+    __m128i b1 = _mm_add_epi32(v1, __rounding);   \
+                                                  \
+    __m128i c0 = _mm_srai_epi32(a0, cos_bit);     \
+    __m128i c1 = _mm_srai_epi32(a1, cos_bit);     \
+    __m128i d0 = _mm_srai_epi32(b0, cos_bit);     \
+    __m128i d1 = _mm_srai_epi32(b1, cos_bit);     \
+                                                  \
+    out0 = _mm_packs_epi32(c0, c1);               \
+    out1 = _mm_packs_epi32(d0, d1);               \
+  }
+
+static INLINE __m128i load_16bit_to_16bit(const int16_t *a) {
+  return _mm_load_si128((const __m128i *)a);
+}
+
+static INLINE __m128i load_32bit_to_16bit(const int32_t *a) {
+  const __m128i a_low = _mm_load_si128((const __m128i *)a);
+  return _mm_packs_epi32(a_low, *(const __m128i *)(a + 4));
+}
+
+static INLINE __m128i load_32bit_to_16bit_w4(const int32_t *a) {
+  const __m128i a_low = _mm_load_si128((const __m128i *)a);
+  return _mm_packs_epi32(a_low, a_low);
+}
+
+// Store 4 16 bit values. Sign extend the values.
+static INLINE void store_16bit_to_32bit_w4(const __m128i a, int32_t *const b) {
+  const __m128i a_lo = _mm_unpacklo_epi16(a, a);
+  const __m128i a_1 = _mm_srai_epi32(a_lo, 16);
+  _mm_store_si128((__m128i *)b, a_1);
+}
+
+// Store 8 16 bit values. Sign extend the values.
+static INLINE void store_16bit_to_32bit(__m128i a, int32_t *b) {
+  const __m128i a_lo = _mm_unpacklo_epi16(a, a);
+  const __m128i a_hi = _mm_unpackhi_epi16(a, a);
+  const __m128i a_1 = _mm_srai_epi32(a_lo, 16);
+  const __m128i a_2 = _mm_srai_epi32(a_hi, 16);
+  _mm_store_si128((__m128i *)b, a_1);
+  _mm_store_si128((__m128i *)(b + 4), a_2);
+}
+
+static INLINE __m128i scale_round_sse2(const __m128i a, const int scale) {
+  const __m128i scale_rounding = pair_set_epi16(scale, 1 << (NewSqrt2Bits - 1));
+  const __m128i b = _mm_madd_epi16(a, scale_rounding);
+  return _mm_srai_epi32(b, NewSqrt2Bits);
+}
+
+static INLINE void store_rect_16bit_to_32bit_w4(const __m128i a,
+                                                int32_t *const b) {
+  const __m128i one = _mm_set1_epi16(1);
+  const __m128i a_lo = _mm_unpacklo_epi16(a, one);
+  const __m128i b_lo = scale_round_sse2(a_lo, NewSqrt2);
+  _mm_store_si128((__m128i *)b, b_lo);
+}
+
+static INLINE void store_rect_16bit_to_32bit(const __m128i a,
+                                             int32_t *const b) {
+  const __m128i one = _mm_set1_epi16(1);
+  const __m128i a_lo = _mm_unpacklo_epi16(a, one);
+  const __m128i a_hi = _mm_unpackhi_epi16(a, one);
+  const __m128i b_lo = scale_round_sse2(a_lo, NewSqrt2);
+  const __m128i b_hi = scale_round_sse2(a_hi, NewSqrt2);
+  _mm_store_si128((__m128i *)b, b_lo);
+  _mm_store_si128((__m128i *)(b + 4), b_hi);
+}
+
+static INLINE void load_buffer_16bit_to_16bit_w4(const int16_t *const in,
+                                                 const int stride,
+                                                 __m128i *const out,
+                                                 const int out_size) {
+  for (int i = 0; i < out_size; ++i) {
+    out[i] = _mm_loadl_epi64((const __m128i *)(in + i * stride));
+  }
+}
+
+static INLINE void load_buffer_16bit_to_16bit_w4_flip(const int16_t *const in,
+                                                      const int stride,
+                                                      __m128i *const out,
+                                                      const int out_size) {
+  for (int i = 0; i < out_size; ++i) {
+    out[out_size - i - 1] = _mm_loadl_epi64((const __m128i *)(in + i * stride));
+  }
+}
+
+static INLINE void load_buffer_16bit_to_16bit(const int16_t *in, int stride,
+                                              __m128i *out, int out_size) {
+  for (int i = 0; i < out_size; ++i) {
+    out[i] = load_16bit_to_16bit(in + i * stride);
+  }
+}
+
+static INLINE void load_buffer_16bit_to_16bit_flip(const int16_t *in,
+                                                   int stride, __m128i *out,
+                                                   int out_size) {
+  for (int i = 0; i < out_size; ++i) {
+    out[out_size - i - 1] = load_16bit_to_16bit(in + i * stride);
+  }
+}
+
+static INLINE void load_buffer_32bit_to_16bit(const int32_t *in, int stride,
+                                              __m128i *out, int out_size) {
+  for (int i = 0; i < out_size; ++i) {
+    out[i] = load_32bit_to_16bit(in + i * stride);
+  }
+}
+
+static INLINE void load_buffer_32bit_to_16bit_w4(const int32_t *in, int stride,
+                                                 __m128i *out, int out_size) {
+  for (int i = 0; i < out_size; ++i) {
+    out[i] = load_32bit_to_16bit_w4(in + i * stride);
+  }
+}
+
+static INLINE void load_buffer_32bit_to_16bit_flip(const int32_t *in,
+                                                   int stride, __m128i *out,
+                                                   int out_size) {
+  for (int i = 0; i < out_size; ++i) {
+    out[out_size - i - 1] = load_32bit_to_16bit(in + i * stride);
+  }
+}
+
+static INLINE void store_buffer_16bit_to_32bit_w4(const __m128i *const in,
+                                                  int32_t *const out,
+                                                  const int stride,
+                                                  const int out_size) {
+  for (int i = 0; i < out_size; ++i) {
+    store_16bit_to_32bit_w4(in[i], out + i * stride);
+  }
+}
+
+static INLINE void store_buffer_16bit_to_32bit_w8(const __m128i *const in,
+                                                  int32_t *const out,
+                                                  const int stride,
+                                                  const int out_size) {
+  for (int i = 0; i < out_size; ++i) {
+    store_16bit_to_32bit(in[i], out + i * stride);
+  }
+}
+
+static INLINE void store_rect_buffer_16bit_to_32bit_w4(const __m128i *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_w4(in[i], out + i * stride);
+  }
+}
+
+static INLINE void store_rect_buffer_16bit_to_32bit_w8(const __m128i *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(in[i], out + i * stride);
+  }
+}
+
+static INLINE void store_buffer_16bit_to_16bit_8x8(const __m128i *in,
+                                                   uint16_t *out,
+                                                   const int stride) {
+  for (int i = 0; i < 8; ++i) {
+    _mm_store_si128((__m128i *)(out + i * stride), in[i]);
+  }
+}
+
+static INLINE void round_shift_16bit(__m128i *in, int size, int bit) {
+  if (bit < 0) {
+    bit = -bit;
+    __m128i rounding = _mm_set1_epi16(1 << (bit - 1));
+    for (int i = 0; i < size; ++i) {
+      in[i] = _mm_adds_epi16(in[i], rounding);
+      in[i] = _mm_srai_epi16(in[i], bit);
+    }
+  } else if (bit > 0) {
+    for (int i = 0; i < size; ++i) {
+      in[i] = _mm_slli_epi16(in[i], bit);
+    }
+  }
+}
+
+static INLINE void flip_buf_sse2(__m128i *in, __m128i *out, int size) {
+  for (int i = 0; i < size; ++i) {
+    out[size - i - 1] = in[i];
+  }
+}
+
+void av1_lowbd_fwd_txfm2d_4x4_sse2(const int16_t *input, int32_t *output,
+                                   int stride, TX_TYPE tx_type, int bd);
+
+void av1_lowbd_fwd_txfm2d_4x8_sse2(const int16_t *input, int32_t *output,
+                                   int stride, TX_TYPE tx_type, int bd);
+
+void av1_lowbd_fwd_txfm2d_4x16_sse2(const int16_t *input, int32_t *output,
+                                    int stride, TX_TYPE tx_type, int bd);
+
+void av1_lowbd_fwd_txfm2d_8x4_sse2(const int16_t *input, int32_t *output,
+                                   int stride, TX_TYPE tx_type, int bd);
+
+void av1_lowbd_fwd_txfm2d_8x8_sse2(const int16_t *input, int32_t *output,
+                                   int stride, TX_TYPE tx_type, int bd);
+
+void av1_lowbd_fwd_txfm2d_8x16_sse2(const int16_t *input, int32_t *output,
+                                    int stride, TX_TYPE tx_type, int bd);
+
+void av1_lowbd_fwd_txfm2d_8x32_sse2(const int16_t *input, int32_t *output,
+                                    int stride, TX_TYPE tx_type, int bd);
+
+void av1_lowbd_fwd_txfm2d_16x4_sse2(const int16_t *input, int32_t *output,
+                                    int stride, TX_TYPE tx_type, int bd);
+
+void av1_lowbd_fwd_txfm2d_16x8_sse2(const int16_t *input, int32_t *output,
+                                    int stride, TX_TYPE tx_type, int bd);
+
+void av1_lowbd_fwd_txfm2d_16x16_sse2(const int16_t *input, int32_t *output,
+                                     int stride, TX_TYPE tx_type, int bd);
+
+void av1_lowbd_fwd_txfm2d_16x32_sse2(const int16_t *input, int32_t *output,
+                                     int stride, TX_TYPE tx_type, int bd);
+
+void av1_lowbd_fwd_txfm2d_32x8_sse2(const int16_t *input, int32_t *output,
+                                    int stride, TX_TYPE tx_type, int bd);
+
+void av1_lowbd_fwd_txfm2d_32x16_sse2(const int16_t *input, int32_t *output,
+                                     int stride, TX_TYPE tx_type, int bd);
+
+void av1_lowbd_fwd_txfm2d_32x32_sse2(const int16_t *input, int32_t *output,
+                                     int stride, TX_TYPE tx_type, int bd);
+
+void av1_lowbd_fwd_txfm2d_16x64_sse2(const int16_t *input, int32_t *output,
+                                     int stride, TX_TYPE tx_type, int bd);
+
+void av1_lowbd_fwd_txfm2d_64x16_sse2(const int16_t *input, int32_t *output,
+                                     int stride, TX_TYPE tx_type, int bd);
+
+typedef void (*transform_1d_sse2)(const __m128i *input, __m128i *output,
+                                  int8_t cos_bit);
+
+typedef struct {
+  transform_1d_sse2 col, row;  // vertical and horizontal
+} transform_2d_sse2;
+
+#ifdef __cplusplus
+}
+#endif  // __cplusplus
+#endif  // AOM_AV1_COMMON_X86_AV1_TXFM_SSE2_H_
diff --git a/third_party/aom/av1/common/x86/av1_txfm_sse4.c b/third_party/aom/av1/common/x86/av1_txfm_sse4.c
new file mode 100644
index 000000000..90b9879cc
--- /dev/null
+++ b/third_party/aom/av1/common/x86/av1_txfm_sse4.c
@@ -0,0 +1,21 @@
+/*
+ * 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 "config/aom_dsp_rtcd.h"
+
+#include "av1/common/av1_txfm.h"
+#include "av1/common/x86/av1_txfm_sse4.h"
+
+void av1_round_shift_array_sse4_1(int32_t *arr, int size, int bit) {
+  __m128i *const vec = (__m128i *)arr;
+  const int vec_size = size >> 2;
+  av1_round_shift_array_32_sse4_1(vec, vec, vec_size, bit);
+}
diff --git a/third_party/aom/av1/common/x86/av1_txfm_sse4.h b/third_party/aom/av1/common/x86/av1_txfm_sse4.h
new file mode 100644
index 000000000..6cad821b1
--- /dev/null
+++ b/third_party/aom/av1/common/x86/av1_txfm_sse4.h
@@ -0,0 +1,72 @@
+/*
+ * 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_AV1_COMMON_X86_AV1_TXFM_SSE4_H_
+#define AOM_AV1_COMMON_X86_AV1_TXFM_SSE4_H_
+
+#include <smmintrin.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+static INLINE __m128i av1_round_shift_32_sse4_1(__m128i vec, int bit) {
+  __m128i tmp, round;
+  round = _mm_set1_epi32(1 << (bit - 1));
+  tmp = _mm_add_epi32(vec, round);
+  return _mm_srai_epi32(tmp, bit);
+}
+
+static INLINE void av1_round_shift_array_32_sse4_1(__m128i *input,
+                                                   __m128i *output,
+                                                   const int size,
+                                                   const int bit) {
+  if (bit > 0) {
+    int i;
+    for (i = 0; i < size; i++) {
+      output[i] = av1_round_shift_32_sse4_1(input[i], bit);
+    }
+  } else {
+    int i;
+    for (i = 0; i < size; i++) {
+      output[i] = _mm_slli_epi32(input[i], -bit);
+    }
+  }
+}
+
+static INLINE void av1_round_shift_rect_array_32_sse4_1(__m128i *input,
+                                                        __m128i *output,
+                                                        const int size,
+                                                        const int bit,
+                                                        const int val) {
+  const __m128i sqrt2 = _mm_set1_epi32(val);
+  if (bit > 0) {
+    int i;
+    for (i = 0; i < size; i++) {
+      const __m128i r0 = av1_round_shift_32_sse4_1(input[i], bit);
+      const __m128i r1 = _mm_mullo_epi32(sqrt2, r0);
+      output[i] = av1_round_shift_32_sse4_1(r1, NewSqrt2Bits);
+    }
+  } else {
+    int i;
+    for (i = 0; i < size; i++) {
+      const __m128i r0 = _mm_slli_epi32(input[i], -bit);
+      const __m128i r1 = _mm_mullo_epi32(sqrt2, r0);
+      output[i] = av1_round_shift_32_sse4_1(r1, NewSqrt2Bits);
+    }
+  }
+}
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif  // AOM_AV1_COMMON_X86_AV1_TXFM_SSE4_H_
diff --git a/third_party/aom/av1/common/x86/cfl_avx2.c b/third_party/aom/av1/common/x86/cfl_avx2.c
new file mode 100644
index 000000000..a8bfdcce6
--- /dev/null
+++ b/third_party/aom/av1/common/x86/cfl_avx2.c
@@ -0,0 +1,491 @@
+/*
+ * 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 "av1/common/cfl.h"
+
+#include "av1/common/x86/cfl_simd.h"
+
+#define CFL_GET_SUBSAMPLE_FUNCTION_AVX2(sub, bd)                           \
+  CFL_SUBSAMPLE(avx2, sub, bd, 32, 32)                                     \
+  CFL_SUBSAMPLE(avx2, sub, bd, 32, 16)                                     \
+  CFL_SUBSAMPLE(avx2, sub, bd, 32, 8)                                      \
+  cfl_subsample_##bd##_fn cfl_get_luma_subsampling_##sub##_##bd##_avx2(    \
+      TX_SIZE tx_size) {                                                   \
+    static const cfl_subsample_##bd##_fn subfn_##sub[TX_SIZES_ALL] = {     \
+      subsample_##bd##_##sub##_4x4_ssse3,   /* 4x4 */                      \
+      subsample_##bd##_##sub##_8x8_ssse3,   /* 8x8 */                      \
+      subsample_##bd##_##sub##_16x16_ssse3, /* 16x16 */                    \
+      subsample_##bd##_##sub##_32x32_avx2,  /* 32x32 */                    \
+      cfl_subsample_##bd##_null,            /* 64x64 (invalid CFL size) */ \
+      subsample_##bd##_##sub##_4x8_ssse3,   /* 4x8 */                      \
+      subsample_##bd##_##sub##_8x4_ssse3,   /* 8x4 */                      \
+      subsample_##bd##_##sub##_8x16_ssse3,  /* 8x16 */                     \
+      subsample_##bd##_##sub##_16x8_ssse3,  /* 16x8 */                     \
+      subsample_##bd##_##sub##_16x32_ssse3, /* 16x32 */                    \
+      subsample_##bd##_##sub##_32x16_avx2,  /* 32x16 */                    \
+      cfl_subsample_##bd##_null,            /* 32x64 (invalid CFL size) */ \
+      cfl_subsample_##bd##_null,            /* 64x32 (invalid CFL size) */ \
+      subsample_##bd##_##sub##_4x16_ssse3,  /* 4x16  */                    \
+      subsample_##bd##_##sub##_16x4_ssse3,  /* 16x4  */                    \
+      subsample_##bd##_##sub##_8x32_ssse3,  /* 8x32  */                    \
+      subsample_##bd##_##sub##_32x8_avx2,   /* 32x8  */                    \
+      cfl_subsample_##bd##_null,            /* 16x64 (invalid CFL size) */ \
+      cfl_subsample_##bd##_null,            /* 64x16 (invalid CFL size) */ \
+    };                                                                     \
+    return subfn_##sub[tx_size];                                           \
+  }
+
+/**
+ * Adds 4 pixels (in a 2x2 grid) and multiplies them by 2. Resulting in a more
+ * precise version of a box filter 4:2:0 pixel subsampling in Q3.
+ *
+ * The CfL prediction buffer is always of size CFL_BUF_SQUARE. However, the
+ * active area is specified using width and height.
+ *
+ * Note: We don't need to worry about going over the active area, as long as we
+ * stay inside the CfL prediction buffer.
+ *
+ * Note: For 4:2:0 luma subsampling, the width will never be greater than 16.
+ */
+static void cfl_luma_subsampling_420_lbd_avx2(const uint8_t *input,
+                                              int input_stride,
+                                              uint16_t *pred_buf_q3, int width,
+                                              int height) {
+  (void)width;                               // Forever 32
+  const __m256i twos = _mm256_set1_epi8(2);  // Thirty two twos
+  const int luma_stride = input_stride << 1;
+  __m256i *row = (__m256i *)pred_buf_q3;
+  const __m256i *row_end = row + (height >> 1) * CFL_BUF_LINE_I256;
+  do {
+    __m256i top = _mm256_loadu_si256((__m256i *)input);
+    __m256i bot = _mm256_loadu_si256((__m256i *)(input + input_stride));
+
+    __m256i top_16x16 = _mm256_maddubs_epi16(top, twos);
+    __m256i bot_16x16 = _mm256_maddubs_epi16(bot, twos);
+    __m256i sum_16x16 = _mm256_add_epi16(top_16x16, bot_16x16);
+
+    _mm256_storeu_si256(row, sum_16x16);
+
+    input += luma_stride;
+  } while ((row += CFL_BUF_LINE_I256) < row_end);
+}
+
+CFL_GET_SUBSAMPLE_FUNCTION_AVX2(420, lbd)
+
+/**
+ * Adds 2 pixels (in a 2x1 grid) and multiplies them by 4. Resulting in a more
+ * precise version of a box filter 4:2:2 pixel subsampling in Q3.
+ *
+ * The CfL prediction buffer is always of size CFL_BUF_SQUARE. However, the
+ * active area is specified using width and height.
+ *
+ * Note: We don't need to worry about going over the active area, as long as we
+ * stay inside the CfL prediction buffer.
+ */
+static void cfl_luma_subsampling_422_lbd_avx2(const uint8_t *input,
+                                              int input_stride,
+                                              uint16_t *pred_buf_q3, int width,
+                                              int height) {
+  (void)width;                                // Forever 32
+  const __m256i fours = _mm256_set1_epi8(4);  // Thirty two fours
+  __m256i *row = (__m256i *)pred_buf_q3;
+  const __m256i *row_end = row + height * CFL_BUF_LINE_I256;
+  do {
+    __m256i top = _mm256_loadu_si256((__m256i *)input);
+    __m256i top_16x16 = _mm256_maddubs_epi16(top, fours);
+    _mm256_storeu_si256(row, top_16x16);
+    input += input_stride;
+  } while ((row += CFL_BUF_LINE_I256) < row_end);
+}
+
+CFL_GET_SUBSAMPLE_FUNCTION_AVX2(422, lbd)
+
+/**
+ * Multiplies the pixels by 8 (scaling in Q3). The AVX2 subsampling is only
+ * performed on block of width 32.
+ *
+ * The CfL prediction buffer is always of size CFL_BUF_SQUARE. However, the
+ * active area is specified using width and height.
+ *
+ * Note: We don't need to worry about going over the active area, as long as we
+ * stay inside the CfL prediction buffer.
+ */
+static void cfl_luma_subsampling_444_lbd_avx2(const uint8_t *input,
+                                              int input_stride,
+                                              uint16_t *pred_buf_q3, int width,
+                                              int height) {
+  (void)width;  // Forever 32
+  __m256i *row = (__m256i *)pred_buf_q3;
+  const __m256i *row_end = row + height * CFL_BUF_LINE_I256;
+  const __m256i zeros = _mm256_setzero_si256();
+  do {
+    __m256i top = _mm256_loadu_si256((__m256i *)input);
+    top = _mm256_permute4x64_epi64(top, _MM_SHUFFLE(3, 1, 2, 0));
+
+    __m256i row_lo = _mm256_unpacklo_epi8(top, zeros);
+    row_lo = _mm256_slli_epi16(row_lo, 3);
+    __m256i row_hi = _mm256_unpackhi_epi8(top, zeros);
+    row_hi = _mm256_slli_epi16(row_hi, 3);
+
+    _mm256_storeu_si256(row, row_lo);
+    _mm256_storeu_si256(row + 1, row_hi);
+
+    input += input_stride;
+  } while ((row += CFL_BUF_LINE_I256) < row_end);
+}
+
+CFL_GET_SUBSAMPLE_FUNCTION_AVX2(444, lbd)
+
+/**
+ * Adds 4 pixels (in a 2x2 grid) and multiplies them by 2. Resulting in a more
+ * precise version of a box filter 4:2:0 pixel subsampling in Q3.
+ *
+ * The CfL prediction buffer is always of size CFL_BUF_SQUARE. However, the
+ * active area is specified using width and height.
+ *
+ * Note: We don't need to worry about going over the active area, as long as we
+ * stay inside the CfL prediction buffer.
+ *
+ * Note: For 4:2:0 luma subsampling, the width will never be greater than 16.
+ */
+static void cfl_luma_subsampling_420_hbd_avx2(const uint16_t *input,
+                                              int input_stride,
+                                              uint16_t *pred_buf_q3, int width,
+                                              int height) {
+  (void)width;  // Forever 32
+  const int luma_stride = input_stride << 1;
+  __m256i *row = (__m256i *)pred_buf_q3;
+  const __m256i *row_end = row + (height >> 1) * CFL_BUF_LINE_I256;
+  do {
+    __m256i top = _mm256_loadu_si256((__m256i *)input);
+    __m256i bot = _mm256_loadu_si256((__m256i *)(input + input_stride));
+    __m256i sum = _mm256_add_epi16(top, bot);
+
+    __m256i top_1 = _mm256_loadu_si256((__m256i *)(input + 16));
+    __m256i bot_1 = _mm256_loadu_si256((__m256i *)(input + 16 + input_stride));
+    __m256i sum_1 = _mm256_add_epi16(top_1, bot_1);
+
+    __m256i hsum = _mm256_hadd_epi16(sum, sum_1);
+    hsum = _mm256_permute4x64_epi64(hsum, _MM_SHUFFLE(3, 1, 2, 0));
+    hsum = _mm256_add_epi16(hsum, hsum);
+
+    _mm256_storeu_si256(row, hsum);
+
+    input += luma_stride;
+  } while ((row += CFL_BUF_LINE_I256) < row_end);
+}
+
+CFL_GET_SUBSAMPLE_FUNCTION_AVX2(420, hbd)
+
+/**
+ * Adds 2 pixels (in a 2x1 grid) and multiplies them by 4. Resulting in a more
+ * precise version of a box filter 4:2:2 pixel subsampling in Q3.
+ *
+ * The CfL prediction buffer is always of size CFL_BUF_SQUARE. However, the
+ * active area is specified using width and height.
+ *
+ * Note: We don't need to worry about going over the active area, as long as we
+ * stay inside the CfL prediction buffer.
+ *
+ */
+static void cfl_luma_subsampling_422_hbd_avx2(const uint16_t *input,
+                                              int input_stride,
+                                              uint16_t *pred_buf_q3, int width,
+                                              int height) {
+  (void)width;  // Forever 32
+  __m256i *row = (__m256i *)pred_buf_q3;
+  const __m256i *row_end = row + height * CFL_BUF_LINE_I256;
+  do {
+    __m256i top = _mm256_loadu_si256((__m256i *)input);
+    __m256i top_1 = _mm256_loadu_si256((__m256i *)(input + 16));
+    __m256i hsum = _mm256_hadd_epi16(top, top_1);
+    hsum = _mm256_permute4x64_epi64(hsum, _MM_SHUFFLE(3, 1, 2, 0));
+    hsum = _mm256_slli_epi16(hsum, 2);
+
+    _mm256_storeu_si256(row, hsum);
+
+    input += input_stride;
+  } while ((row += CFL_BUF_LINE_I256) < row_end);
+}
+
+CFL_GET_SUBSAMPLE_FUNCTION_AVX2(422, hbd)
+
+static void cfl_luma_subsampling_444_hbd_avx2(const uint16_t *input,
+                                              int input_stride,
+                                              uint16_t *pred_buf_q3, int width,
+                                              int height) {
+  (void)width;  // Forever 32
+  __m256i *row = (__m256i *)pred_buf_q3;
+  const __m256i *row_end = row + height * CFL_BUF_LINE_I256;
+  do {
+    __m256i top = _mm256_loadu_si256((__m256i *)input);
+    __m256i top_1 = _mm256_loadu_si256((__m256i *)(input + 16));
+    _mm256_storeu_si256(row, _mm256_slli_epi16(top, 3));
+    _mm256_storeu_si256(row + 1, _mm256_slli_epi16(top_1, 3));
+    input += input_stride;
+  } while ((row += CFL_BUF_LINE_I256) < row_end);
+}
+
+CFL_GET_SUBSAMPLE_FUNCTION_AVX2(444, hbd)
+
+static INLINE __m256i predict_unclipped(const __m256i *input, __m256i alpha_q12,
+                                        __m256i alpha_sign, __m256i dc_q0) {
+  __m256i ac_q3 = _mm256_loadu_si256(input);
+  __m256i ac_sign = _mm256_sign_epi16(alpha_sign, ac_q3);
+  __m256i scaled_luma_q0 =
+      _mm256_mulhrs_epi16(_mm256_abs_epi16(ac_q3), alpha_q12);
+  scaled_luma_q0 = _mm256_sign_epi16(scaled_luma_q0, ac_sign);
+  return _mm256_add_epi16(scaled_luma_q0, dc_q0);
+}
+
+static INLINE void cfl_predict_lbd_avx2(const int16_t *pred_buf_q3,
+                                        uint8_t *dst, int dst_stride,
+                                        int alpha_q3, int width, int height) {
+  (void)width;
+  const __m256i alpha_sign = _mm256_set1_epi16(alpha_q3);
+  const __m256i alpha_q12 = _mm256_slli_epi16(_mm256_abs_epi16(alpha_sign), 9);
+  const __m256i dc_q0 = _mm256_set1_epi16(*dst);
+  __m256i *row = (__m256i *)pred_buf_q3;
+  const __m256i *row_end = row + height * CFL_BUF_LINE_I256;
+
+  do {
+    __m256i res = predict_unclipped(row, alpha_q12, alpha_sign, dc_q0);
+    __m256i next = predict_unclipped(row + 1, alpha_q12, alpha_sign, dc_q0);
+    res = _mm256_packus_epi16(res, next);
+    res = _mm256_permute4x64_epi64(res, _MM_SHUFFLE(3, 1, 2, 0));
+    _mm256_storeu_si256((__m256i *)dst, res);
+    dst += dst_stride;
+  } while ((row += CFL_BUF_LINE_I256) < row_end);
+}
+
+CFL_PREDICT_X(avx2, 32, 8, lbd);
+CFL_PREDICT_X(avx2, 32, 16, lbd);
+CFL_PREDICT_X(avx2, 32, 32, lbd);
+
+cfl_predict_lbd_fn get_predict_lbd_fn_avx2(TX_SIZE tx_size) {
+  static const cfl_predict_lbd_fn pred[TX_SIZES_ALL] = {
+    predict_lbd_4x4_ssse3,   /* 4x4 */
+    predict_lbd_8x8_ssse3,   /* 8x8 */
+    predict_lbd_16x16_ssse3, /* 16x16 */
+    predict_lbd_32x32_avx2,  /* 32x32 */
+    cfl_predict_lbd_null,    /* 64x64 (invalid CFL size) */
+    predict_lbd_4x8_ssse3,   /* 4x8 */
+    predict_lbd_8x4_ssse3,   /* 8x4 */
+    predict_lbd_8x16_ssse3,  /* 8x16 */
+    predict_lbd_16x8_ssse3,  /* 16x8 */
+    predict_lbd_16x32_ssse3, /* 16x32 */
+    predict_lbd_32x16_avx2,  /* 32x16 */
+    cfl_predict_lbd_null,    /* 32x64 (invalid CFL size) */
+    cfl_predict_lbd_null,    /* 64x32 (invalid CFL size) */
+    predict_lbd_4x16_ssse3,  /* 4x16  */
+    predict_lbd_16x4_ssse3,  /* 16x4  */
+    predict_lbd_8x32_ssse3,  /* 8x32  */
+    predict_lbd_32x8_avx2,   /* 32x8  */
+    cfl_predict_lbd_null,    /* 16x64 (invalid CFL size) */
+    cfl_predict_lbd_null,    /* 64x16 (invalid CFL size) */
+  };
+  // Modulo TX_SIZES_ALL to ensure that an attacker won't be able to index the
+  // function pointer array out of bounds.
+  return pred[tx_size % TX_SIZES_ALL];
+}
+
+static __m256i highbd_max_epi16(int bd) {
+  const __m256i neg_one = _mm256_set1_epi16(-1);
+  // (1 << bd) - 1 => -(-1 << bd) -1 => -1 - (-1 << bd) => -1 ^ (-1 << bd)
+  return _mm256_xor_si256(_mm256_slli_epi16(neg_one, bd), neg_one);
+}
+
+static __m256i highbd_clamp_epi16(__m256i u, __m256i zero, __m256i max) {
+  return _mm256_max_epi16(_mm256_min_epi16(u, max), zero);
+}
+
+static INLINE void cfl_predict_hbd_avx2(const int16_t *pred_buf_q3,
+                                        uint16_t *dst, int dst_stride,
+                                        int alpha_q3, int bd, int width,
+                                        int height) {
+  // Use SSSE3 version for smaller widths
+  assert(width == 16 || width == 32);
+  const __m256i alpha_sign = _mm256_set1_epi16(alpha_q3);
+  const __m256i alpha_q12 = _mm256_slli_epi16(_mm256_abs_epi16(alpha_sign), 9);
+  const __m256i dc_q0 = _mm256_loadu_si256((__m256i *)dst);
+  const __m256i max = highbd_max_epi16(bd);
+
+  __m256i *row = (__m256i *)pred_buf_q3;
+  const __m256i *row_end = row + height * CFL_BUF_LINE_I256;
+  do {
+    const __m256i res = predict_unclipped(row, alpha_q12, alpha_sign, dc_q0);
+    _mm256_storeu_si256((__m256i *)dst,
+                        highbd_clamp_epi16(res, _mm256_setzero_si256(), max));
+    if (width == 32) {
+      const __m256i res_1 =
+          predict_unclipped(row + 1, alpha_q12, alpha_sign, dc_q0);
+      _mm256_storeu_si256(
+          (__m256i *)(dst + 16),
+          highbd_clamp_epi16(res_1, _mm256_setzero_si256(), max));
+    }
+    dst += dst_stride;
+  } while ((row += CFL_BUF_LINE_I256) < row_end);
+}
+
+CFL_PREDICT_X(avx2, 16, 4, hbd)
+CFL_PREDICT_X(avx2, 16, 8, hbd)
+CFL_PREDICT_X(avx2, 16, 16, hbd)
+CFL_PREDICT_X(avx2, 16, 32, hbd)
+CFL_PREDICT_X(avx2, 32, 8, hbd)
+CFL_PREDICT_X(avx2, 32, 16, hbd)
+CFL_PREDICT_X(avx2, 32, 32, hbd)
+
+cfl_predict_hbd_fn get_predict_hbd_fn_avx2(TX_SIZE tx_size) {
+  static const cfl_predict_hbd_fn pred[TX_SIZES_ALL] = {
+    predict_hbd_4x4_ssse3,  /* 4x4 */
+    predict_hbd_8x8_ssse3,  /* 8x8 */
+    predict_hbd_16x16_avx2, /* 16x16 */
+    predict_hbd_32x32_avx2, /* 32x32 */
+    cfl_predict_hbd_null,   /* 64x64 (invalid CFL size) */
+    predict_hbd_4x8_ssse3,  /* 4x8 */
+    predict_hbd_8x4_ssse3,  /* 8x4 */
+    predict_hbd_8x16_ssse3, /* 8x16 */
+    predict_hbd_16x8_avx2,  /* 16x8 */
+    predict_hbd_16x32_avx2, /* 16x32 */
+    predict_hbd_32x16_avx2, /* 32x16 */
+    cfl_predict_hbd_null,   /* 32x64 (invalid CFL size) */
+    cfl_predict_hbd_null,   /* 64x32 (invalid CFL size) */
+    predict_hbd_4x16_ssse3, /* 4x16  */
+    predict_hbd_16x4_avx2,  /* 16x4  */
+    predict_hbd_8x32_ssse3, /* 8x32  */
+    predict_hbd_32x8_avx2,  /* 32x8  */
+    cfl_predict_hbd_null,   /* 16x64 (invalid CFL size) */
+    cfl_predict_hbd_null,   /* 64x16 (invalid CFL size) */
+  };
+  // Modulo TX_SIZES_ALL to ensure that an attacker won't be able to index the
+  // function pointer array out of bounds.
+  return pred[tx_size % TX_SIZES_ALL];
+}
+
+// Returns a vector where all the (32-bits) elements are the sum of all the
+// lanes in a.
+static INLINE __m256i fill_sum_epi32(__m256i a) {
+  // Given that a == [A, B, C, D, E, F, G, H]
+  a = _mm256_hadd_epi32(a, a);
+  // Given that A' == A + B, C' == C + D, E' == E + F, G' == G + H
+  // a == [A', C', A', C', E', G', E', G']
+  a = _mm256_permute4x64_epi64(a, _MM_SHUFFLE(3, 1, 2, 0));
+  // a == [A', C', E', G', A', C', E', G']
+  a = _mm256_hadd_epi32(a, a);
+  // Given that A'' == A' + C' and E'' == E' + G'
+  // a == [A'', E'', A'', E'', A'', E'', A'', E'']
+  return _mm256_hadd_epi32(a, a);
+  // Given that A''' == A'' + E''
+  // a == [A''', A''', A''', A''', A''', A''', A''', A''']
+}
+
+static INLINE __m256i _mm256_addl_epi16(__m256i a) {
+  return _mm256_add_epi32(_mm256_unpacklo_epi16(a, _mm256_setzero_si256()),
+                          _mm256_unpackhi_epi16(a, _mm256_setzero_si256()));
+}
+
+static INLINE void subtract_average_avx2(const uint16_t *src_ptr,
+                                         int16_t *dst_ptr, int width,
+                                         int height, int round_offset,
+                                         int num_pel_log2) {
+  // Use SSE2 version for smaller widths
+  assert(width == 16 || width == 32);
+
+  const __m256i *src = (__m256i *)src_ptr;
+  const __m256i *const end = src + height * CFL_BUF_LINE_I256;
+  // To maximize usage of the AVX2 registers, we sum two rows per loop
+  // iteration
+  const int step = 2 * CFL_BUF_LINE_I256;
+
+  __m256i sum = _mm256_setzero_si256();
+  // For width 32, we use a second sum accumulator to reduce accumulator
+  // dependencies in the loop.
+  __m256i sum2;
+  if (width == 32) sum2 = _mm256_setzero_si256();
+
+  do {
+    // Add top row to the bottom row
+    __m256i l0 = _mm256_add_epi16(_mm256_loadu_si256(src),
+                                  _mm256_loadu_si256(src + CFL_BUF_LINE_I256));
+    sum = _mm256_add_epi32(sum, _mm256_addl_epi16(l0));
+    if (width == 32) { /* Don't worry, this if it gets optimized out. */
+      // Add the second part of the top row to the second part of the bottom row
+      __m256i l1 =
+          _mm256_add_epi16(_mm256_loadu_si256(src + 1),
+                           _mm256_loadu_si256(src + 1 + CFL_BUF_LINE_I256));
+      sum2 = _mm256_add_epi32(sum2, _mm256_addl_epi16(l1));
+    }
+    src += step;
+  } while (src < end);
+  // Combine both sum accumulators
+  if (width == 32) sum = _mm256_add_epi32(sum, sum2);
+
+  __m256i fill = fill_sum_epi32(sum);
+
+  __m256i avg_epi16 = _mm256_srli_epi32(
+      _mm256_add_epi32(fill, _mm256_set1_epi32(round_offset)), num_pel_log2);
+  avg_epi16 = _mm256_packs_epi32(avg_epi16, avg_epi16);
+
+  // Store and subtract loop
+  src = (__m256i *)src_ptr;
+  __m256i *dst = (__m256i *)dst_ptr;
+  do {
+    _mm256_storeu_si256(dst,
+                        _mm256_sub_epi16(_mm256_loadu_si256(src), avg_epi16));
+    if (width == 32) {
+      _mm256_storeu_si256(
+          dst + 1, _mm256_sub_epi16(_mm256_loadu_si256(src + 1), avg_epi16));
+    }
+    src += CFL_BUF_LINE_I256;
+    dst += CFL_BUF_LINE_I256;
+  } while (src < end);
+}
+
+// Declare wrappers for AVX2 sizes
+CFL_SUB_AVG_X(avx2, 16, 4, 32, 6)
+CFL_SUB_AVG_X(avx2, 16, 8, 64, 7)
+CFL_SUB_AVG_X(avx2, 16, 16, 128, 8)
+CFL_SUB_AVG_X(avx2, 16, 32, 256, 9)
+CFL_SUB_AVG_X(avx2, 32, 8, 128, 8)
+CFL_SUB_AVG_X(avx2, 32, 16, 256, 9)
+CFL_SUB_AVG_X(avx2, 32, 32, 512, 10)
+
+// Based on the observation that for small blocks AVX2 does not outperform
+// SSE2, we call the SSE2 code for block widths 4 and 8.
+cfl_subtract_average_fn get_subtract_average_fn_avx2(TX_SIZE tx_size) {
+  static const cfl_subtract_average_fn sub_avg[TX_SIZES_ALL] = {
+    subtract_average_4x4_sse2,   /* 4x4 */
+    subtract_average_8x8_sse2,   /* 8x8 */
+    subtract_average_16x16_avx2, /* 16x16 */
+    subtract_average_32x32_avx2, /* 32x32 */
+    cfl_subtract_average_null,   /* 64x64 (invalid CFL size) */
+    subtract_average_4x8_sse2,   /* 4x8 */
+    subtract_average_8x4_sse2,   /* 8x4 */
+    subtract_average_8x16_sse2,  /* 8x16 */
+    subtract_average_16x8_avx2,  /* 16x8 */
+    subtract_average_16x32_avx2, /* 16x32 */
+    subtract_average_32x16_avx2, /* 32x16 */
+    cfl_subtract_average_null,   /* 32x64 (invalid CFL size) */
+    cfl_subtract_average_null,   /* 64x32 (invalid CFL size) */
+    subtract_average_4x16_sse2,  /* 4x16 */
+    subtract_average_16x4_avx2,  /* 16x4 */
+    subtract_average_8x32_sse2,  /* 8x32 */
+    subtract_average_32x8_avx2,  /* 32x8 */
+    cfl_subtract_average_null,   /* 16x64 (invalid CFL size) */
+    cfl_subtract_average_null,   /* 64x16 (invalid CFL size) */
+  };
+  // Modulo TX_SIZES_ALL to ensure that an attacker won't be able to
+  // index the function pointer array out of bounds.
+  return sub_avg[tx_size % TX_SIZES_ALL];
+}
diff --git a/third_party/aom/av1/common/x86/cfl_simd.h b/third_party/aom/av1/common/x86/cfl_simd.h
new file mode 100644
index 000000000..3b342cd4e
--- /dev/null
+++ b/third_party/aom/av1/common/x86/cfl_simd.h
@@ -0,0 +1,243 @@
+/*
+ * 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_AV1_COMMON_X86_CFL_SIMD_H_
+#define AOM_AV1_COMMON_X86_CFL_SIMD_H_
+
+#include "av1/common/blockd.h"
+
+// SSSE3 version is optimal for with == 4, we reuse them in AVX2
+void subsample_lbd_420_4x4_ssse3(const uint8_t *input, int input_stride,
+                                 uint16_t *output_q3);
+void subsample_lbd_420_4x8_ssse3(const uint8_t *input, int input_stride,
+                                 uint16_t *output_q3);
+void subsample_lbd_420_4x16_ssse3(const uint8_t *input, int input_stride,
+                                  uint16_t *output_q3);
+
+// SSSE3 version is optimal for with == 8, we reuse it in AVX2
+void subsample_lbd_420_8x4_ssse3(const uint8_t *input, int input_stride,
+                                 uint16_t *output_q3);
+void subsample_lbd_420_8x8_ssse3(const uint8_t *input, int input_stride,
+                                 uint16_t *output_q3);
+void subsample_lbd_420_8x16_ssse3(const uint8_t *input, int input_stride,
+                                  uint16_t *output_q3);
+void subsample_lbd_420_8x32_ssse3(const uint8_t *input, int input_stride,
+                                  uint16_t *output_q3);
+
+// SSSE3 version is optimal for with == 16, we reuse it in AVX2
+void subsample_lbd_420_16x4_ssse3(const uint8_t *input, int input_stride,
+                                  uint16_t *output_q3);
+void subsample_lbd_420_16x8_ssse3(const uint8_t *input, int input_stride,
+                                  uint16_t *output_q3);
+void subsample_lbd_420_16x16_ssse3(const uint8_t *input, int input_stride,
+                                   uint16_t *output_q3);
+void subsample_lbd_420_16x32_ssse3(const uint8_t *input, int input_stride,
+                                   uint16_t *output_q3);
+
+// SSSE3 version is optimal for with == 4, we reuse them in AVX2
+void subsample_lbd_422_4x4_ssse3(const uint8_t *input, int input_stride,
+                                 uint16_t *output_q3);
+void subsample_lbd_422_4x8_ssse3(const uint8_t *input, int input_stride,
+                                 uint16_t *output_q3);
+void subsample_lbd_422_4x16_ssse3(const uint8_t *input, int input_stride,
+                                  uint16_t *output_q3);
+
+// SSSE3 version is optimal for with == 8, we reuse it in AVX2
+void subsample_lbd_422_8x4_ssse3(const uint8_t *input, int input_stride,
+                                 uint16_t *output_q3);
+void subsample_lbd_422_8x8_ssse3(const uint8_t *input, int input_stride,
+                                 uint16_t *output_q3);
+void subsample_lbd_422_8x16_ssse3(const uint8_t *input, int input_stride,
+                                  uint16_t *output_q3);
+void subsample_lbd_422_8x32_ssse3(const uint8_t *input, int input_stride,
+                                  uint16_t *output_q3);
+
+// SSSE3 version is optimal for with == 16, we reuse it in AVX2
+void subsample_lbd_422_16x4_ssse3(const uint8_t *input, int input_stride,
+                                  uint16_t *output_q3);
+void subsample_lbd_422_16x8_ssse3(const uint8_t *input, int input_stride,
+                                  uint16_t *output_q3);
+void subsample_lbd_422_16x16_ssse3(const uint8_t *input, int input_stride,
+                                   uint16_t *output_q3);
+void subsample_lbd_422_16x32_ssse3(const uint8_t *input, int input_stride,
+                                   uint16_t *output_q3);
+
+// SSSE3 version is optimal for with == 4, we reuse them in AVX2
+void subsample_lbd_444_4x4_ssse3(const uint8_t *input, int input_stride,
+                                 uint16_t *output_q3);
+void subsample_lbd_444_4x8_ssse3(const uint8_t *input, int input_stride,
+                                 uint16_t *output_q3);
+void subsample_lbd_444_4x16_ssse3(const uint8_t *input, int input_stride,
+                                  uint16_t *output_q3);
+
+// SSSE3 version is optimal for with == 8, we reuse it in AVX2
+void subsample_lbd_444_8x4_ssse3(const uint8_t *input, int input_stride,
+                                 uint16_t *output_q3);
+void subsample_lbd_444_8x8_ssse3(const uint8_t *input, int input_stride,
+                                 uint16_t *output_q3);
+void subsample_lbd_444_8x16_ssse3(const uint8_t *input, int input_stride,
+                                  uint16_t *output_q3);
+void subsample_lbd_444_8x32_ssse3(const uint8_t *input, int input_stride,
+                                  uint16_t *output_q3);
+
+// SSSE3 version is optimal for with == 16, we reuse it in AVX2
+void subsample_lbd_444_16x4_ssse3(const uint8_t *input, int input_stride,
+                                  uint16_t *output_q3);
+void subsample_lbd_444_16x8_ssse3(const uint8_t *input, int input_stride,
+                                  uint16_t *output_q3);
+void subsample_lbd_444_16x16_ssse3(const uint8_t *input, int input_stride,
+                                   uint16_t *output_q3);
+void subsample_lbd_444_16x32_ssse3(const uint8_t *input, int input_stride,
+                                   uint16_t *output_q3);
+
+void subsample_hbd_420_4x4_ssse3(const uint16_t *input, int input_stride,
+                                 uint16_t *output_q3);
+void subsample_hbd_420_4x8_ssse3(const uint16_t *input, int input_stride,
+                                 uint16_t *output_q3);
+void subsample_hbd_420_4x16_ssse3(const uint16_t *input, int input_stride,
+                                  uint16_t *output_q3);
+
+// SSSE3 version is optimal for with == 8, we reuse it in AVX2
+void subsample_hbd_420_8x4_ssse3(const uint16_t *input, int input_stride,
+                                 uint16_t *output_q3);
+void subsample_hbd_420_8x8_ssse3(const uint16_t *input, int input_stride,
+                                 uint16_t *output_q3);
+void subsample_hbd_420_8x16_ssse3(const uint16_t *input, int input_stride,
+                                  uint16_t *output_q3);
+void subsample_hbd_420_8x32_ssse3(const uint16_t *input, int input_stride,
+                                  uint16_t *output_q3);
+
+// SSSE3 version is faster for with == 16, we reuse it in AVX2
+void subsample_hbd_420_16x4_ssse3(const uint16_t *input, int input_stride,
+                                  uint16_t *output_q3);
+void subsample_hbd_420_16x8_ssse3(const uint16_t *input, int input_stride,
+                                  uint16_t *output_q3);
+void subsample_hbd_420_16x16_ssse3(const uint16_t *input, int input_stride,
+                                   uint16_t *output_q3);
+void subsample_hbd_420_16x32_ssse3(const uint16_t *input, int input_stride,
+                                   uint16_t *output_q3);
+
+void subsample_hbd_422_4x4_ssse3(const uint16_t *input, int input_stride,
+                                 uint16_t *output_q3);
+void subsample_hbd_422_4x8_ssse3(const uint16_t *input, int input_stride,
+                                 uint16_t *output_q3);
+void subsample_hbd_422_4x16_ssse3(const uint16_t *input, int input_stride,
+                                  uint16_t *output_q3);
+
+// SSSE3 version is optimal for with == 8, we reuse it in AVX2
+void subsample_hbd_422_8x4_ssse3(const uint16_t *input, int input_stride,
+                                 uint16_t *output_q3);
+void subsample_hbd_422_8x8_ssse3(const uint16_t *input, int input_stride,
+                                 uint16_t *output_q3);
+void subsample_hbd_422_8x16_ssse3(const uint16_t *input, int input_stride,
+                                  uint16_t *output_q3);
+void subsample_hbd_422_8x32_ssse3(const uint16_t *input, int input_stride,
+                                  uint16_t *output_q3);
+
+// SSSE3 version is faster for with == 16, we reuse it in AVX2
+void subsample_hbd_422_16x4_ssse3(const uint16_t *input, int input_stride,
+                                  uint16_t *output_q3);
+void subsample_hbd_422_16x8_ssse3(const uint16_t *input, int input_stride,
+                                  uint16_t *output_q3);
+void subsample_hbd_422_16x16_ssse3(const uint16_t *input, int input_stride,
+                                   uint16_t *output_q3);
+void subsample_hbd_422_16x32_ssse3(const uint16_t *input, int input_stride,
+                                   uint16_t *output_q3);
+
+void subsample_hbd_444_4x4_ssse3(const uint16_t *input, int input_stride,
+                                 uint16_t *output_q3);
+void subsample_hbd_444_4x8_ssse3(const uint16_t *input, int input_stride,
+                                 uint16_t *output_q3);
+void subsample_hbd_444_4x16_ssse3(const uint16_t *input, int input_stride,
+                                  uint16_t *output_q3);
+
+// SSSE3 version is optimal for with == 8, we reuse it in AVX2
+void subsample_hbd_444_8x4_ssse3(const uint16_t *input, int input_stride,
+                                 uint16_t *output_q3);
+void subsample_hbd_444_8x8_ssse3(const uint16_t *input, int input_stride,
+                                 uint16_t *output_q3);
+void subsample_hbd_444_8x16_ssse3(const uint16_t *input, int input_stride,
+                                  uint16_t *output_q3);
+void subsample_hbd_444_8x32_ssse3(const uint16_t *input, int input_stride,
+                                  uint16_t *output_q3);
+
+// SSSE3 version is faster for with == 16, we reuse it in AVX2
+void subsample_hbd_444_16x4_ssse3(const uint16_t *input, int input_stride,
+                                  uint16_t *output_q3);
+void subsample_hbd_444_16x8_ssse3(const uint16_t *input, int input_stride,
+                                  uint16_t *output_q3);
+void subsample_hbd_444_16x16_ssse3(const uint16_t *input, int input_stride,
+                                   uint16_t *output_q3);
+void subsample_hbd_444_16x32_ssse3(const uint16_t *input, int input_stride,
+                                   uint16_t *output_q3);
+
+// SSE2 version is optimal for with == 4, we reuse them in AVX2
+void subtract_average_4x4_sse2(const uint16_t *src, int16_t *dst);
+void subtract_average_4x8_sse2(const uint16_t *src, int16_t *dst);
+void subtract_average_4x16_sse2(const uint16_t *src, int16_t *dst);
+
+// SSE2 version is optimal for with == 8, we reuse them in AVX2
+void subtract_average_8x4_sse2(const uint16_t *src, int16_t *dst);
+void subtract_average_8x8_sse2(const uint16_t *src, int16_t *dst);
+void subtract_average_8x16_sse2(const uint16_t *src, int16_t *dst);
+void subtract_average_8x32_sse2(const uint16_t *src, int16_t *dst);
+
+void predict_lbd_4x4_ssse3(const int16_t *pred_buf_q3, uint8_t *dst,
+                           int dst_stride, int alpha_q3);
+void predict_lbd_4x8_ssse3(const int16_t *pred_buf_q3, uint8_t *dst,
+                           int dst_stride, int alpha_q3);
+void predict_lbd_4x16_ssse3(const int16_t *pred_buf_q3, uint8_t *dst,
+                            int dst_stride, int alpha_q3);
+
+void predict_lbd_8x4_ssse3(const int16_t *pred_buf_q3, uint8_t *dst,
+                           int dst_stride, int alpha_q3);
+void predict_lbd_8x8_ssse3(const int16_t *pred_buf_q3, uint8_t *dst,
+                           int dst_stride, int alpha_q3);
+void predict_lbd_8x16_ssse3(const int16_t *pred_buf_q3, uint8_t *dst,
+                            int dst_stride, int alpha_q3);
+void predict_lbd_8x32_ssse3(const int16_t *pred_buf_q3, uint8_t *dst,
+                            int dst_stride, int alpha_q3);
+
+void predict_lbd_16x4_ssse3(const int16_t *pred_buf_q3, uint8_t *dst,
+                            int dst_stride, int alpha_q3);
+void predict_lbd_16x8_ssse3(const int16_t *pred_buf_q3, uint8_t *dst,
+                            int dst_stride, int alpha_q3);
+void predict_lbd_16x16_ssse3(const int16_t *pred_buf_q3, uint8_t *dst,
+                             int dst_stride, int alpha_q3);
+void predict_lbd_16x32_ssse3(const int16_t *pred_buf_q3, uint8_t *dst,
+                             int dst_stride, int alpha_q3);
+
+void predict_hbd_4x4_ssse3(const int16_t *pred_buf_q3, uint16_t *dst,
+                           int dst_stride, int alpha_q3, int bd);
+void predict_hbd_4x8_ssse3(const int16_t *pred_buf_q3, uint16_t *dst,
+                           int dst_stride, int alpha_q3, int bd);
+void predict_hbd_4x16_ssse3(const int16_t *pred_buf_q3, uint16_t *dst,
+                            int dst_stride, int alpha_q3, int bd);
+
+void predict_hbd_8x4_ssse3(const int16_t *pred_buf_q3, uint16_t *dst,
+                           int dst_stride, int alpha_q3, int bd);
+void predict_hbd_8x8_ssse3(const int16_t *pred_buf_q3, uint16_t *dst,
+                           int dst_stride, int alpha_q3, int bd);
+void predict_hbd_8x16_ssse3(const int16_t *pred_buf_q3, uint16_t *dst,
+                            int dst_stride, int alpha_q3, int bd);
+void predict_hbd_8x32_ssse3(const int16_t *pred_buf_q3, uint16_t *dst,
+                            int dst_stride, int alpha_q3, int bd);
+
+void predict_hbd_16x4_ssse3(const int16_t *pred_buf_q3, uint16_t *dst,
+                            int dst_stride, int alpha_q3, int bd);
+void predict_hbd_16x8_ssse3(const int16_t *pred_buf_q3, uint16_t *dst,
+                            int dst_stride, int alpha_q3, int bd);
+void predict_hbd_16x16_ssse3(const int16_t *pred_buf_q3, uint16_t *dst,
+                             int dst_stride, int alpha_q3, int bd);
+void predict_hbd_16x32_ssse3(const int16_t *pred_buf_q3, uint16_t *dst,
+                             int dst_stride, int alpha_q3, int bd);
+
+#endif  // AOM_AV1_COMMON_X86_CFL_SIMD_H_
diff --git a/third_party/aom/av1/common/x86/cfl_sse2.c b/third_party/aom/av1/common/x86/cfl_sse2.c
new file mode 100644
index 000000000..4783fe098
--- /dev/null
+++ b/third_party/aom/av1/common/x86/cfl_sse2.c
@@ -0,0 +1,89 @@
+/*
+ * 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 "av1/common/cfl.h"
+#include "config/av1_rtcd.h"
+
+static INLINE __m128i fill_sum_epi32(__m128i l0) {
+  l0 = _mm_add_epi32(l0, _mm_shuffle_epi32(l0, _MM_SHUFFLE(1, 0, 3, 2)));
+  return _mm_add_epi32(l0, _mm_shuffle_epi32(l0, _MM_SHUFFLE(2, 3, 0, 1)));
+}
+
+static INLINE void subtract_average_sse2(const uint16_t *src_ptr,
+                                         int16_t *dst_ptr, int width,
+                                         int height, int round_offset,
+                                         int num_pel_log2) {
+  const __m128i zeros = _mm_setzero_si128();
+  const __m128i round_offset_epi32 = _mm_set1_epi32(round_offset);
+  const __m128i *src = (__m128i *)src_ptr;
+  const __m128i *const end = src + height * CFL_BUF_LINE_I128;
+  const int step = CFL_BUF_LINE_I128 * (1 + (width == 8) + 3 * (width == 4));
+
+  __m128i sum = zeros;
+  do {
+    __m128i l0;
+    if (width == 4) {
+      l0 = _mm_add_epi16(_mm_loadl_epi64(src),
+                         _mm_loadl_epi64(src + CFL_BUF_LINE_I128));
+      __m128i l1 = _mm_add_epi16(_mm_loadl_epi64(src + 2 * CFL_BUF_LINE_I128),
+                                 _mm_loadl_epi64(src + 3 * CFL_BUF_LINE_I128));
+      sum = _mm_add_epi32(sum, _mm_add_epi32(_mm_unpacklo_epi16(l0, zeros),
+                                             _mm_unpacklo_epi16(l1, zeros)));
+    } else {
+      if (width == 8) {
+        l0 = _mm_add_epi16(_mm_loadu_si128(src),
+                           _mm_loadu_si128(src + CFL_BUF_LINE_I128));
+      } else {
+        l0 = _mm_add_epi16(_mm_loadu_si128(src), _mm_loadu_si128(src + 1));
+      }
+      sum = _mm_add_epi32(sum, _mm_add_epi32(_mm_unpacklo_epi16(l0, zeros),
+                                             _mm_unpackhi_epi16(l0, zeros)));
+      if (width == 32) {
+        l0 = _mm_add_epi16(_mm_loadu_si128(src + 2), _mm_loadu_si128(src + 3));
+        sum = _mm_add_epi32(sum, _mm_add_epi32(_mm_unpacklo_epi16(l0, zeros),
+                                               _mm_unpackhi_epi16(l0, zeros)));
+      }
+    }
+    src += step;
+  } while (src < end);
+
+  sum = fill_sum_epi32(sum);
+
+  __m128i avg_epi16 =
+      _mm_srli_epi32(_mm_add_epi32(sum, round_offset_epi32), num_pel_log2);
+  avg_epi16 = _mm_packs_epi32(avg_epi16, avg_epi16);
+
+  src = (__m128i *)src_ptr;
+  __m128i *dst = (__m128i *)dst_ptr;
+  do {
+    if (width == 4) {
+      _mm_storel_epi64(dst, _mm_sub_epi16(_mm_loadl_epi64(src), avg_epi16));
+    } else {
+      _mm_storeu_si128(dst, _mm_sub_epi16(_mm_loadu_si128(src), avg_epi16));
+      if (width > 8) {
+        _mm_storeu_si128(dst + 1,
+                         _mm_sub_epi16(_mm_loadu_si128(src + 1), avg_epi16));
+        if (width == 32) {
+          _mm_storeu_si128(dst + 2,
+                           _mm_sub_epi16(_mm_loadu_si128(src + 2), avg_epi16));
+          _mm_storeu_si128(dst + 3,
+                           _mm_sub_epi16(_mm_loadu_si128(src + 3), avg_epi16));
+        }
+      }
+    }
+    src += CFL_BUF_LINE_I128;
+    dst += CFL_BUF_LINE_I128;
+  } while (src < end);
+}
+
+CFL_SUB_AVG_FN(sse2)
diff --git a/third_party/aom/av1/common/x86/cfl_ssse3.c b/third_party/aom/av1/common/x86/cfl_ssse3.c
new file mode 100644
index 000000000..bbf007295
--- /dev/null
+++ b/third_party/aom/av1/common/x86/cfl_ssse3.c
@@ -0,0 +1,393 @@
+/*
+ * 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/av1_rtcd.h"
+
+#include "av1/common/cfl.h"
+
+#include "av1/common/x86/cfl_simd.h"
+
+// Load 32-bit integer from memory into the first element of dst.
+static INLINE __m128i _mm_loadh_epi32(__m128i const *mem_addr) {
+  return _mm_cvtsi32_si128(*((int *)mem_addr));
+}
+
+// Store 32-bit integer from the first element of a into memory.
+static INLINE void _mm_storeh_epi32(__m128i const *mem_addr, __m128i a) {
+  *((int *)mem_addr) = _mm_cvtsi128_si32(a);
+}
+
+/**
+ * Adds 4 pixels (in a 2x2 grid) and multiplies them by 2. Resulting in a more
+ * precise version of a box filter 4:2:0 pixel subsampling in Q3.
+ *
+ * The CfL prediction buffer is always of size CFL_BUF_SQUARE. However, the
+ * active area is specified using width and height.
+ *
+ * Note: We don't need to worry about going over the active area, as long as we
+ * stay inside the CfL prediction buffer.
+ */
+static INLINE void cfl_luma_subsampling_420_lbd_ssse3(const uint8_t *input,
+                                                      int input_stride,
+                                                      uint16_t *pred_buf_q3,
+                                                      int width, int height) {
+  const __m128i twos = _mm_set1_epi8(2);
+  __m128i *pred_buf_m128i = (__m128i *)pred_buf_q3;
+  const __m128i *end = pred_buf_m128i + (height >> 1) * CFL_BUF_LINE_I128;
+  const int luma_stride = input_stride << 1;
+  do {
+    if (width == 4) {
+      __m128i top = _mm_loadh_epi32((__m128i *)input);
+      top = _mm_maddubs_epi16(top, twos);
+      __m128i bot = _mm_loadh_epi32((__m128i *)(input + input_stride));
+      bot = _mm_maddubs_epi16(bot, twos);
+      const __m128i sum = _mm_add_epi16(top, bot);
+      _mm_storeh_epi32(pred_buf_m128i, sum);
+    } else if (width == 8) {
+      __m128i top = _mm_loadl_epi64((__m128i *)input);
+      top = _mm_maddubs_epi16(top, twos);
+      __m128i bot = _mm_loadl_epi64((__m128i *)(input + input_stride));
+      bot = _mm_maddubs_epi16(bot, twos);
+      const __m128i sum = _mm_add_epi16(top, bot);
+      _mm_storel_epi64(pred_buf_m128i, sum);
+    } else {
+      __m128i top = _mm_loadu_si128((__m128i *)input);
+      top = _mm_maddubs_epi16(top, twos);
+      __m128i bot = _mm_loadu_si128((__m128i *)(input + input_stride));
+      bot = _mm_maddubs_epi16(bot, twos);
+      const __m128i sum = _mm_add_epi16(top, bot);
+      _mm_storeu_si128(pred_buf_m128i, sum);
+      if (width == 32) {
+        __m128i top_1 = _mm_loadu_si128(((__m128i *)input) + 1);
+        __m128i bot_1 =
+            _mm_loadu_si128(((__m128i *)(input + input_stride)) + 1);
+        top_1 = _mm_maddubs_epi16(top_1, twos);
+        bot_1 = _mm_maddubs_epi16(bot_1, twos);
+        __m128i sum_1 = _mm_add_epi16(top_1, bot_1);
+        _mm_storeu_si128(pred_buf_m128i + 1, sum_1);
+      }
+    }
+    input += luma_stride;
+    pred_buf_m128i += CFL_BUF_LINE_I128;
+  } while (pred_buf_m128i < end);
+}
+
+/**
+ * Adds 2 pixels (in a 2x1 grid) and multiplies them by 4. Resulting in a more
+ * precise version of a box filter 4:2:2 pixel subsampling in Q3.
+ *
+ * The CfL prediction buffer is always of size CFL_BUF_SQUARE. However, the
+ * active area is specified using width and height.
+ *
+ * Note: We don't need to worry about going over the active area, as long as we
+ * stay inside the CfL prediction buffer.
+ */
+static INLINE void cfl_luma_subsampling_422_lbd_ssse3(const uint8_t *input,
+                                                      int input_stride,
+                                                      uint16_t *pred_buf_q3,
+                                                      int width, int height) {
+  const __m128i fours = _mm_set1_epi8(4);
+  __m128i *pred_buf_m128i = (__m128i *)pred_buf_q3;
+  const __m128i *end = pred_buf_m128i + height * CFL_BUF_LINE_I128;
+  do {
+    if (width == 4) {
+      __m128i top = _mm_loadh_epi32((__m128i *)input);
+      top = _mm_maddubs_epi16(top, fours);
+      _mm_storeh_epi32(pred_buf_m128i, top);
+    } else if (width == 8) {
+      __m128i top = _mm_loadl_epi64((__m128i *)input);
+      top = _mm_maddubs_epi16(top, fours);
+      _mm_storel_epi64(pred_buf_m128i, top);
+    } else {
+      __m128i top = _mm_loadu_si128((__m128i *)input);
+      top = _mm_maddubs_epi16(top, fours);
+      _mm_storeu_si128(pred_buf_m128i, top);
+      if (width == 32) {
+        __m128i top_1 = _mm_loadu_si128(((__m128i *)input) + 1);
+        top_1 = _mm_maddubs_epi16(top_1, fours);
+        _mm_storeu_si128(pred_buf_m128i + 1, top_1);
+      }
+    }
+    input += input_stride;
+    pred_buf_m128i += CFL_BUF_LINE_I128;
+  } while (pred_buf_m128i < end);
+}
+
+/**
+ * Multiplies the pixels by 8 (scaling in Q3).
+ *
+ * The CfL prediction buffer is always of size CFL_BUF_SQUARE. However, the
+ * active area is specified using width and height.
+ *
+ * Note: We don't need to worry about going over the active area, as long as we
+ * stay inside the CfL prediction buffer.
+ */
+static INLINE void cfl_luma_subsampling_444_lbd_ssse3(const uint8_t *input,
+                                                      int input_stride,
+                                                      uint16_t *pred_buf_q3,
+                                                      int width, int height) {
+  const __m128i zeros = _mm_setzero_si128();
+  const int luma_stride = input_stride;
+  __m128i *pred_buf_m128i = (__m128i *)pred_buf_q3;
+  const __m128i *end = pred_buf_m128i + height * CFL_BUF_LINE_I128;
+  do {
+    if (width == 4) {
+      __m128i row = _mm_loadh_epi32((__m128i *)input);
+      row = _mm_unpacklo_epi8(row, zeros);
+      _mm_storel_epi64(pred_buf_m128i, _mm_slli_epi16(row, 3));
+    } else if (width == 8) {
+      __m128i row = _mm_loadl_epi64((__m128i *)input);
+      row = _mm_unpacklo_epi8(row, zeros);
+      _mm_storeu_si128(pred_buf_m128i, _mm_slli_epi16(row, 3));
+    } else {
+      __m128i row = _mm_loadu_si128((__m128i *)input);
+      const __m128i row_lo = _mm_unpacklo_epi8(row, zeros);
+      const __m128i row_hi = _mm_unpackhi_epi8(row, zeros);
+      _mm_storeu_si128(pred_buf_m128i, _mm_slli_epi16(row_lo, 3));
+      _mm_storeu_si128(pred_buf_m128i + 1, _mm_slli_epi16(row_hi, 3));
+      if (width == 32) {
+        __m128i row_1 = _mm_loadu_si128(((__m128i *)input) + 1);
+        const __m128i row_1_lo = _mm_unpacklo_epi8(row_1, zeros);
+        const __m128i row_1_hi = _mm_unpackhi_epi8(row_1, zeros);
+        _mm_storeu_si128(pred_buf_m128i + 2, _mm_slli_epi16(row_1_lo, 3));
+        _mm_storeu_si128(pred_buf_m128i + 3, _mm_slli_epi16(row_1_hi, 3));
+      }
+    }
+    input += luma_stride;
+    pred_buf_m128i += CFL_BUF_LINE_I128;
+  } while (pred_buf_m128i < end);
+}
+
+/**
+ * Adds 4 pixels (in a 2x2 grid) and multiplies them by 2. Resulting in a more
+ * precise version of a box filter 4:2:0 pixel subsampling in Q3.
+ *
+ * The CfL prediction buffer is always of size CFL_BUF_SQUARE. However, the
+ * active area is specified using width and height.
+ *
+ * Note: We don't need to worry about going over the active area, as long as we
+ * stay inside the CfL prediction buffer.
+ */
+static INLINE void cfl_luma_subsampling_420_hbd_ssse3(const uint16_t *input,
+                                                      int input_stride,
+                                                      uint16_t *pred_buf_q3,
+                                                      int width, int height) {
+  const uint16_t *end = pred_buf_q3 + (height >> 1) * CFL_BUF_LINE;
+  const int luma_stride = input_stride << 1;
+  do {
+    if (width == 4) {
+      const __m128i top = _mm_loadl_epi64((__m128i *)input);
+      const __m128i bot = _mm_loadl_epi64((__m128i *)(input + input_stride));
+      __m128i sum = _mm_add_epi16(top, bot);
+      sum = _mm_hadd_epi16(sum, sum);
+      *((int *)pred_buf_q3) = _mm_cvtsi128_si32(_mm_add_epi16(sum, sum));
+    } else {
+      const __m128i top = _mm_loadu_si128((__m128i *)input);
+      const __m128i bot = _mm_loadu_si128((__m128i *)(input + input_stride));
+      __m128i sum = _mm_add_epi16(top, bot);
+      if (width == 8) {
+        sum = _mm_hadd_epi16(sum, sum);
+        _mm_storel_epi64((__m128i *)pred_buf_q3, _mm_add_epi16(sum, sum));
+      } else {
+        const __m128i top_1 = _mm_loadu_si128(((__m128i *)input) + 1);
+        const __m128i bot_1 =
+            _mm_loadu_si128(((__m128i *)(input + input_stride)) + 1);
+        sum = _mm_hadd_epi16(sum, _mm_add_epi16(top_1, bot_1));
+        _mm_storeu_si128((__m128i *)pred_buf_q3, _mm_add_epi16(sum, sum));
+        if (width == 32) {
+          const __m128i top_2 = _mm_loadu_si128(((__m128i *)input) + 2);
+          const __m128i bot_2 =
+              _mm_loadu_si128(((__m128i *)(input + input_stride)) + 2);
+          const __m128i top_3 = _mm_loadu_si128(((__m128i *)input) + 3);
+          const __m128i bot_3 =
+              _mm_loadu_si128(((__m128i *)(input + input_stride)) + 3);
+          const __m128i sum_2 = _mm_add_epi16(top_2, bot_2);
+          const __m128i sum_3 = _mm_add_epi16(top_3, bot_3);
+          __m128i next_sum = _mm_hadd_epi16(sum_2, sum_3);
+          _mm_storeu_si128(((__m128i *)pred_buf_q3) + 1,
+                           _mm_add_epi16(next_sum, next_sum));
+        }
+      }
+    }
+    input += luma_stride;
+  } while ((pred_buf_q3 += CFL_BUF_LINE) < end);
+}
+
+/**
+ * Adds 2 pixels (in a 2x1 grid) and multiplies them by 4. Resulting in a more
+ * precise version of a box filter 4:2:2 pixel subsampling in Q3.
+ *
+ * The CfL prediction buffer is always of size CFL_BUF_SQUARE. However, the
+ * active area is specified using width and height.
+ *
+ * Note: We don't need to worry about going over the active area, as long as we
+ * stay inside the CfL prediction buffer.
+ */
+static INLINE void cfl_luma_subsampling_422_hbd_ssse3(const uint16_t *input,
+                                                      int input_stride,
+                                                      uint16_t *pred_buf_q3,
+                                                      int width, int height) {
+  __m128i *pred_buf_m128i = (__m128i *)pred_buf_q3;
+  const __m128i *end = pred_buf_m128i + height * CFL_BUF_LINE_I128;
+  do {
+    if (width == 4) {
+      const __m128i top = _mm_loadl_epi64((__m128i *)input);
+      const __m128i sum = _mm_slli_epi16(_mm_hadd_epi16(top, top), 2);
+      _mm_storeh_epi32(pred_buf_m128i, sum);
+    } else {
+      const __m128i top = _mm_loadu_si128((__m128i *)input);
+      if (width == 8) {
+        const __m128i sum = _mm_slli_epi16(_mm_hadd_epi16(top, top), 2);
+        _mm_storel_epi64(pred_buf_m128i, sum);
+      } else {
+        const __m128i top_1 = _mm_loadu_si128(((__m128i *)input) + 1);
+        const __m128i sum = _mm_slli_epi16(_mm_hadd_epi16(top, top_1), 2);
+        _mm_storeu_si128(pred_buf_m128i, sum);
+        if (width == 32) {
+          const __m128i top_2 = _mm_loadu_si128(((__m128i *)input) + 2);
+          const __m128i top_3 = _mm_loadu_si128(((__m128i *)input) + 3);
+          const __m128i sum_1 = _mm_slli_epi16(_mm_hadd_epi16(top_2, top_3), 2);
+          _mm_storeu_si128(pred_buf_m128i + 1, sum_1);
+        }
+      }
+    }
+    pred_buf_m128i += CFL_BUF_LINE_I128;
+    input += input_stride;
+  } while (pred_buf_m128i < end);
+}
+
+static INLINE void cfl_luma_subsampling_444_hbd_ssse3(const uint16_t *input,
+                                                      int input_stride,
+                                                      uint16_t *pred_buf_q3,
+                                                      int width, int height) {
+  const uint16_t *end = pred_buf_q3 + height * CFL_BUF_LINE;
+  do {
+    if (width == 4) {
+      const __m128i row = _mm_slli_epi16(_mm_loadl_epi64((__m128i *)input), 3);
+      _mm_storel_epi64((__m128i *)pred_buf_q3, row);
+    } else {
+      const __m128i row = _mm_slli_epi16(_mm_loadu_si128((__m128i *)input), 3);
+      _mm_storeu_si128((__m128i *)pred_buf_q3, row);
+      if (width >= 16) {
+        __m128i row_1 = _mm_loadu_si128(((__m128i *)input) + 1);
+        row_1 = _mm_slli_epi16(row_1, 3);
+        _mm_storeu_si128(((__m128i *)pred_buf_q3) + 1, row_1);
+        if (width == 32) {
+          __m128i row_2 = _mm_loadu_si128(((__m128i *)input) + 2);
+          row_2 = _mm_slli_epi16(row_2, 3);
+          _mm_storeu_si128(((__m128i *)pred_buf_q3) + 2, row_2);
+          __m128i row_3 = _mm_loadu_si128(((__m128i *)input) + 3);
+          row_3 = _mm_slli_epi16(row_3, 3);
+          _mm_storeu_si128(((__m128i *)pred_buf_q3) + 3, row_3);
+        }
+      }
+    }
+    input += input_stride;
+    pred_buf_q3 += CFL_BUF_LINE;
+  } while (pred_buf_q3 < end);
+}
+
+CFL_GET_SUBSAMPLE_FUNCTION(ssse3)
+
+static INLINE __m128i predict_unclipped(const __m128i *input, __m128i alpha_q12,
+                                        __m128i alpha_sign, __m128i dc_q0) {
+  __m128i ac_q3 = _mm_loadu_si128(input);
+  __m128i ac_sign = _mm_sign_epi16(alpha_sign, ac_q3);
+  __m128i scaled_luma_q0 = _mm_mulhrs_epi16(_mm_abs_epi16(ac_q3), alpha_q12);
+  scaled_luma_q0 = _mm_sign_epi16(scaled_luma_q0, ac_sign);
+  return _mm_add_epi16(scaled_luma_q0, dc_q0);
+}
+
+static INLINE void cfl_predict_lbd_ssse3(const int16_t *pred_buf_q3,
+                                         uint8_t *dst, int dst_stride,
+                                         int alpha_q3, int width, int height) {
+  const __m128i alpha_sign = _mm_set1_epi16(alpha_q3);
+  const __m128i alpha_q12 = _mm_slli_epi16(_mm_abs_epi16(alpha_sign), 9);
+  const __m128i dc_q0 = _mm_set1_epi16(*dst);
+  __m128i *row = (__m128i *)pred_buf_q3;
+  const __m128i *row_end = row + height * CFL_BUF_LINE_I128;
+  do {
+    __m128i res = predict_unclipped(row, alpha_q12, alpha_sign, dc_q0);
+    if (width < 16) {
+      res = _mm_packus_epi16(res, res);
+      if (width == 4)
+        _mm_storeh_epi32((__m128i *)dst, res);
+      else
+        _mm_storel_epi64((__m128i *)dst, res);
+    } else {
+      __m128i next = predict_unclipped(row + 1, alpha_q12, alpha_sign, dc_q0);
+      res = _mm_packus_epi16(res, next);
+      _mm_storeu_si128((__m128i *)dst, res);
+      if (width == 32) {
+        res = predict_unclipped(row + 2, alpha_q12, alpha_sign, dc_q0);
+        next = predict_unclipped(row + 3, alpha_q12, alpha_sign, dc_q0);
+        res = _mm_packus_epi16(res, next);
+        _mm_storeu_si128((__m128i *)(dst + 16), res);
+      }
+    }
+    dst += dst_stride;
+  } while ((row += CFL_BUF_LINE_I128) < row_end);
+}
+
+CFL_PREDICT_FN(ssse3, lbd)
+
+static INLINE __m128i highbd_max_epi16(int bd) {
+  const __m128i neg_one = _mm_set1_epi16(-1);
+  // (1 << bd) - 1 => -(-1 << bd) -1 => -1 - (-1 << bd) => -1 ^ (-1 << bd)
+  return _mm_xor_si128(_mm_slli_epi16(neg_one, bd), neg_one);
+}
+
+static INLINE __m128i highbd_clamp_epi16(__m128i u, __m128i zero, __m128i max) {
+  return _mm_max_epi16(_mm_min_epi16(u, max), zero);
+}
+
+static INLINE void cfl_predict_hbd_ssse3(const int16_t *pred_buf_q3,
+                                         uint16_t *dst, int dst_stride,
+                                         int alpha_q3, int bd, int width,
+                                         int height) {
+  const __m128i alpha_sign = _mm_set1_epi16(alpha_q3);
+  const __m128i alpha_q12 = _mm_slli_epi16(_mm_abs_epi16(alpha_sign), 9);
+  const __m128i dc_q0 = _mm_set1_epi16(*dst);
+  const __m128i max = highbd_max_epi16(bd);
+  const __m128i zeros = _mm_setzero_si128();
+  __m128i *row = (__m128i *)pred_buf_q3;
+  const __m128i *row_end = row + height * CFL_BUF_LINE_I128;
+  do {
+    __m128i res = predict_unclipped(row, alpha_q12, alpha_sign, dc_q0);
+    res = highbd_clamp_epi16(res, zeros, max);
+    if (width == 4) {
+      _mm_storel_epi64((__m128i *)dst, res);
+    } else {
+      _mm_storeu_si128((__m128i *)dst, res);
+    }
+    if (width >= 16) {
+      const __m128i res_1 =
+          predict_unclipped(row + 1, alpha_q12, alpha_sign, dc_q0);
+      _mm_storeu_si128(((__m128i *)dst) + 1,
+                       highbd_clamp_epi16(res_1, zeros, max));
+    }
+    if (width == 32) {
+      const __m128i res_2 =
+          predict_unclipped(row + 2, alpha_q12, alpha_sign, dc_q0);
+      _mm_storeu_si128((__m128i *)(dst + 16),
+                       highbd_clamp_epi16(res_2, zeros, max));
+      const __m128i res_3 =
+          predict_unclipped(row + 3, alpha_q12, alpha_sign, dc_q0);
+      _mm_storeu_si128((__m128i *)(dst + 24),
+                       highbd_clamp_epi16(res_3, zeros, max));
+    }
+    dst += dst_stride;
+  } while ((row += CFL_BUF_LINE_I128) < row_end);
+}
+
+CFL_PREDICT_FN(ssse3, hbd)
diff --git a/third_party/aom/av1/common/x86/convolve_2d_avx2.c b/third_party/aom/av1/common/x86/convolve_2d_avx2.c
new file mode 100644
index 000000000..0acafd044
--- /dev/null
+++ b/third_party/aom/av1/common/x86/convolve_2d_avx2.c
@@ -0,0 +1,283 @@
+/*
+ * 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/convolve_avx2.h"
+#include "aom_dsp/x86/convolve_common_intrin.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/aom_filter.h"
+#include "aom_dsp/x86/synonyms.h"
+#include "av1/common/convolve.h"
+
+void av1_convolve_2d_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst,
+                             int dst_stride, int w, int h,
+                             const InterpFilterParams *filter_params_x,
+                             const InterpFilterParams *filter_params_y,
+                             const int subpel_x_q4, const int subpel_y_q4,
+                             ConvolveParams *conv_params) {
+  const int bd = 8;
+
+  DECLARE_ALIGNED(32, int16_t, im_block[(MAX_SB_SIZE + MAX_FILTER_TAP) * 8]);
+  int im_h = h + filter_params_y->taps - 1;
+  int im_stride = 8;
+  int i, j;
+  const int fo_vert = filter_params_y->taps / 2 - 1;
+  const int fo_horiz = filter_params_x->taps / 2 - 1;
+  const uint8_t *const src_ptr = src - fo_vert * src_stride - fo_horiz;
+
+  const int bits =
+      FILTER_BITS * 2 - conv_params->round_0 - conv_params->round_1;
+  const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0;
+
+  __m256i filt[4], coeffs_h[4], coeffs_v[4];
+
+  assert(conv_params->round_0 > 0);
+
+  filt[0] = _mm256_load_si256((__m256i const *)filt_global_avx2);
+  filt[1] = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32));
+  filt[2] = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32 * 2));
+  filt[3] = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32 * 3));
+
+  prepare_coeffs_lowbd(filter_params_x, subpel_x_q4, coeffs_h);
+  prepare_coeffs(filter_params_y, subpel_y_q4, coeffs_v);
+
+  const __m256i round_const_h = _mm256_set1_epi16(
+      ((1 << (conv_params->round_0 - 1)) >> 1) + (1 << (bd + FILTER_BITS - 2)));
+  const __m128i round_shift_h = _mm_cvtsi32_si128(conv_params->round_0 - 1);
+
+  const __m256i sum_round_v = _mm256_set1_epi32(
+      (1 << offset_bits) + ((1 << conv_params->round_1) >> 1));
+  const __m128i sum_shift_v = _mm_cvtsi32_si128(conv_params->round_1);
+
+  const __m256i round_const_v = _mm256_set1_epi32(
+      ((1 << bits) >> 1) - (1 << (offset_bits - conv_params->round_1)) -
+      ((1 << (offset_bits - conv_params->round_1)) >> 1));
+  const __m128i round_shift_v = _mm_cvtsi32_si128(bits);
+
+  for (j = 0; j < w; j += 8) {
+    for (i = 0; i < im_h; i += 2) {
+      __m256i data = _mm256_castsi128_si256(
+          _mm_loadu_si128((__m128i *)&src_ptr[(i * src_stride) + j]));
+
+      // Load the next line
+      if (i + 1 < im_h)
+        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);
+    }
+
+    /* Vertical filter */
+    {
+      __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);
+
+        // Combine V round and 2F-H-V round into a single rounding
+        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);
+
+        /* rounding code */
+        // 16 bit conversion
+        const __m256i res_16bit = _mm256_packs_epi32(res_a_round, res_b_round);
+        // 8 bit conversion and saturation to uint8
+        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);
+
+        // Store values into the destination buffer
+        __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 = _mm_cvtsi128_si32(res_0);
+          *(uint16_t *)p_1 = _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];
+      }
+    }
+  }
+}
+
+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 av1_convolve_2d_copy_sr_avx2(const uint8_t *src, int src_stride,
+                                  uint8_t *dst, int dst_stride, int w, int h,
+                                  const InterpFilterParams *filter_params_x,
+                                  const InterpFilterParams *filter_params_y,
+                                  const int subpel_x_q4, const int subpel_y_q4,
+                                  ConvolveParams *conv_params) {
+  (void)filter_params_x;
+  (void)filter_params_y;
+  (void)subpel_x_q4;
+  (void)subpel_y_q4;
+  (void)conv_params;
+
+  if (w >= 16) {
+    assert(!((intptr_t)dst % 16));
+    assert(!(dst_stride % 16));
+  }
+
+  if (w == 2) {
+    do {
+      memcpy(dst, src, 2 * sizeof(*src));
+      src += src_stride;
+      dst += dst_stride;
+      memcpy(dst, src, 2 * sizeof(*src));
+      src += src_stride;
+      dst += dst_stride;
+      h -= 2;
+    } while (h);
+  } else if (w == 4) {
+    do {
+      memcpy(dst, src, 4 * sizeof(*src));
+      src += src_stride;
+      dst += dst_stride;
+      memcpy(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);
+  }
+}
diff --git a/third_party/aom/av1/common/x86/convolve_2d_sse2.c b/third_party/aom/av1/common/x86/convolve_2d_sse2.c
new file mode 100644
index 000000000..b1a62a4f6
--- /dev/null
+++ b/third_party/aom/av1/common/x86/convolve_2d_sse2.c
@@ -0,0 +1,472 @@
+/*
+ * 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 "config/av1_rtcd.h"
+
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/aom_filter.h"
+#include "aom_dsp/x86/convolve_sse2.h"
+#include "av1/common/convolve.h"
+
+void av1_convolve_2d_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst,
+                             int dst_stride, int w, int h,
+                             const InterpFilterParams *filter_params_x,
+                             const InterpFilterParams *filter_params_y,
+                             const int subpel_x_q4, const int subpel_y_q4,
+                             ConvolveParams *conv_params) {
+  const int bd = 8;
+
+  DECLARE_ALIGNED(16, int16_t,
+                  im_block[(MAX_SB_SIZE + MAX_FILTER_TAP - 1) * MAX_SB_SIZE]);
+  int im_h = h + filter_params_y->taps - 1;
+  int im_stride = MAX_SB_SIZE;
+  int i, j;
+  const int fo_vert = filter_params_y->taps / 2 - 1;
+  const int fo_horiz = filter_params_x->taps / 2 - 1;
+  const uint8_t *const src_ptr = src - fo_vert * src_stride - fo_horiz;
+
+  const __m128i zero = _mm_setzero_si128();
+  const int bits =
+      FILTER_BITS * 2 - conv_params->round_0 - conv_params->round_1;
+  const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0;
+
+  assert(conv_params->round_0 > 0);
+
+  /* Horizontal filter */
+  {
+    const int16_t *x_filter = av1_get_interp_filter_subpel_kernel(
+        filter_params_x, subpel_x_q4 & SUBPEL_MASK);
+    const __m128i coeffs_x = _mm_loadu_si128((__m128i *)x_filter);
+
+    // coeffs 0 1 0 1 2 3 2 3
+    const __m128i tmp_0 = _mm_unpacklo_epi32(coeffs_x, coeffs_x);
+    // coeffs 4 5 4 5 6 7 6 7
+    const __m128i tmp_1 = _mm_unpackhi_epi32(coeffs_x, coeffs_x);
+
+    // coeffs 0 1 0 1 0 1 0 1
+    const __m128i coeff_01 = _mm_unpacklo_epi64(tmp_0, tmp_0);
+    // coeffs 2 3 2 3 2 3 2 3
+    const __m128i coeff_23 = _mm_unpackhi_epi64(tmp_0, tmp_0);
+    // coeffs 4 5 4 5 4 5 4 5
+    const __m128i coeff_45 = _mm_unpacklo_epi64(tmp_1, tmp_1);
+    // coeffs 6 7 6 7 6 7 6 7
+    const __m128i coeff_67 = _mm_unpackhi_epi64(tmp_1, tmp_1);
+
+    const __m128i round_const = _mm_set1_epi32(
+        (1 << (bd + FILTER_BITS - 1)) + ((1 << conv_params->round_0) >> 1));
+    const __m128i round_shift = _mm_cvtsi32_si128(conv_params->round_0);
+
+    for (i = 0; i < im_h; ++i) {
+      for (j = 0; j < w; j += 8) {
+        const __m128i data =
+            _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + j]);
+
+        // Filter even-index pixels
+        const __m128i src_0 = _mm_unpacklo_epi8(data, zero);
+        const __m128i res_0 = _mm_madd_epi16(src_0, coeff_01);
+        const __m128i src_2 = _mm_unpacklo_epi8(_mm_srli_si128(data, 2), zero);
+        const __m128i res_2 = _mm_madd_epi16(src_2, coeff_23);
+        const __m128i src_4 = _mm_unpacklo_epi8(_mm_srli_si128(data, 4), zero);
+        const __m128i res_4 = _mm_madd_epi16(src_4, coeff_45);
+        const __m128i src_6 = _mm_unpacklo_epi8(_mm_srli_si128(data, 6), zero);
+        const __m128i res_6 = _mm_madd_epi16(src_6, coeff_67);
+
+        __m128i res_even = _mm_add_epi32(_mm_add_epi32(res_0, res_4),
+                                         _mm_add_epi32(res_2, res_6));
+        res_even =
+            _mm_sra_epi32(_mm_add_epi32(res_even, round_const), round_shift);
+
+        // Filter odd-index pixels
+        const __m128i src_1 = _mm_unpacklo_epi8(_mm_srli_si128(data, 1), zero);
+        const __m128i res_1 = _mm_madd_epi16(src_1, coeff_01);
+        const __m128i src_3 = _mm_unpacklo_epi8(_mm_srli_si128(data, 3), zero);
+        const __m128i res_3 = _mm_madd_epi16(src_3, coeff_23);
+        const __m128i src_5 = _mm_unpacklo_epi8(_mm_srli_si128(data, 5), zero);
+        const __m128i res_5 = _mm_madd_epi16(src_5, coeff_45);
+        const __m128i src_7 = _mm_unpacklo_epi8(_mm_srli_si128(data, 7), zero);
+        const __m128i res_7 = _mm_madd_epi16(src_7, coeff_67);
+
+        __m128i res_odd = _mm_add_epi32(_mm_add_epi32(res_1, res_5),
+                                        _mm_add_epi32(res_3, res_7));
+        res_odd =
+            _mm_sra_epi32(_mm_add_epi32(res_odd, round_const), round_shift);
+
+        // Pack in the column order 0, 2, 4, 6, 1, 3, 5, 7
+        __m128i res = _mm_packs_epi32(res_even, res_odd);
+        _mm_storeu_si128((__m128i *)&im_block[i * im_stride + j], res);
+      }
+    }
+  }
+
+  /* Vertical filter */
+  {
+    const int16_t *y_filter = av1_get_interp_filter_subpel_kernel(
+        filter_params_y, subpel_y_q4 & SUBPEL_MASK);
+    const __m128i coeffs_y = _mm_loadu_si128((__m128i *)y_filter);
+
+    // coeffs 0 1 0 1 2 3 2 3
+    const __m128i tmp_0 = _mm_unpacklo_epi32(coeffs_y, coeffs_y);
+    // coeffs 4 5 4 5 6 7 6 7
+    const __m128i tmp_1 = _mm_unpackhi_epi32(coeffs_y, coeffs_y);
+
+    // coeffs 0 1 0 1 0 1 0 1
+    const __m128i coeff_01 = _mm_unpacklo_epi64(tmp_0, tmp_0);
+    // coeffs 2 3 2 3 2 3 2 3
+    const __m128i coeff_23 = _mm_unpackhi_epi64(tmp_0, tmp_0);
+    // coeffs 4 5 4 5 4 5 4 5
+    const __m128i coeff_45 = _mm_unpacklo_epi64(tmp_1, tmp_1);
+    // coeffs 6 7 6 7 6 7 6 7
+    const __m128i coeff_67 = _mm_unpackhi_epi64(tmp_1, tmp_1);
+
+    const __m128i sum_round =
+        _mm_set1_epi32((1 << offset_bits) + ((1 << conv_params->round_1) >> 1));
+    const __m128i sum_shift = _mm_cvtsi32_si128(conv_params->round_1);
+
+    const __m128i round_const = _mm_set1_epi32(
+        ((1 << bits) >> 1) - (1 << (offset_bits - conv_params->round_1)) -
+        ((1 << (offset_bits - conv_params->round_1)) >> 1));
+    const __m128i round_shift = _mm_cvtsi32_si128(bits);
+
+    for (i = 0; i < h; ++i) {
+      for (j = 0; j < w; j += 8) {
+        // Filter even-index pixels
+        const int16_t *data = &im_block[i * im_stride + j];
+        const __m128i src_0 =
+            _mm_unpacklo_epi16(*(__m128i *)(data + 0 * im_stride),
+                               *(__m128i *)(data + 1 * im_stride));
+        const __m128i src_2 =
+            _mm_unpacklo_epi16(*(__m128i *)(data + 2 * im_stride),
+                               *(__m128i *)(data + 3 * im_stride));
+        const __m128i src_4 =
+            _mm_unpacklo_epi16(*(__m128i *)(data + 4 * im_stride),
+                               *(__m128i *)(data + 5 * im_stride));
+        const __m128i src_6 =
+            _mm_unpacklo_epi16(*(__m128i *)(data + 6 * im_stride),
+                               *(__m128i *)(data + 7 * im_stride));
+
+        const __m128i res_0 = _mm_madd_epi16(src_0, coeff_01);
+        const __m128i res_2 = _mm_madd_epi16(src_2, coeff_23);
+        const __m128i res_4 = _mm_madd_epi16(src_4, coeff_45);
+        const __m128i res_6 = _mm_madd_epi16(src_6, coeff_67);
+
+        const __m128i res_even = _mm_add_epi32(_mm_add_epi32(res_0, res_2),
+                                               _mm_add_epi32(res_4, res_6));
+
+        // Filter odd-index pixels
+        const __m128i src_1 =
+            _mm_unpackhi_epi16(*(__m128i *)(data + 0 * im_stride),
+                               *(__m128i *)(data + 1 * im_stride));
+        const __m128i src_3 =
+            _mm_unpackhi_epi16(*(__m128i *)(data + 2 * im_stride),
+                               *(__m128i *)(data + 3 * im_stride));
+        const __m128i src_5 =
+            _mm_unpackhi_epi16(*(__m128i *)(data + 4 * im_stride),
+                               *(__m128i *)(data + 5 * im_stride));
+        const __m128i src_7 =
+            _mm_unpackhi_epi16(*(__m128i *)(data + 6 * im_stride),
+                               *(__m128i *)(data + 7 * im_stride));
+
+        const __m128i res_1 = _mm_madd_epi16(src_1, coeff_01);
+        const __m128i res_3 = _mm_madd_epi16(src_3, coeff_23);
+        const __m128i res_5 = _mm_madd_epi16(src_5, coeff_45);
+        const __m128i res_7 = _mm_madd_epi16(src_7, coeff_67);
+
+        const __m128i res_odd = _mm_add_epi32(_mm_add_epi32(res_1, res_3),
+                                              _mm_add_epi32(res_5, res_7));
+
+        // Rearrange pixels back into the order 0 ... 7
+        const __m128i res_lo = _mm_unpacklo_epi32(res_even, res_odd);
+        const __m128i res_hi = _mm_unpackhi_epi32(res_even, res_odd);
+
+        __m128i res_lo_round =
+            _mm_sra_epi32(_mm_add_epi32(res_lo, sum_round), sum_shift);
+        __m128i res_hi_round =
+            _mm_sra_epi32(_mm_add_epi32(res_hi, sum_round), sum_shift);
+
+        res_lo_round = _mm_sra_epi32(_mm_add_epi32(res_lo_round, round_const),
+                                     round_shift);
+        res_hi_round = _mm_sra_epi32(_mm_add_epi32(res_hi_round, round_const),
+                                     round_shift);
+
+        const __m128i res16 = _mm_packs_epi32(res_lo_round, res_hi_round);
+        const __m128i res = _mm_packus_epi16(res16, res16);
+
+        // Accumulate values into the destination buffer
+        __m128i *const p = (__m128i *)&dst[i * dst_stride + j];
+
+        if (w == 2) {
+          *(uint16_t *)p = _mm_cvtsi128_si32(res);
+        } else if (w == 4) {
+          *(uint32_t *)p = _mm_cvtsi128_si32(res);
+        } else {
+          _mm_storel_epi64(p, res);
+        }
+      }
+    }
+  }
+}
+
+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 av1_convolve_2d_copy_sr_sse2(const uint8_t *src, int src_stride,
+                                  uint8_t *dst, int dst_stride, int w, int h,
+                                  const InterpFilterParams *filter_params_x,
+                                  const InterpFilterParams *filter_params_y,
+                                  const int subpel_x_q4, const int subpel_y_q4,
+                                  ConvolveParams *conv_params) {
+  (void)filter_params_x;
+  (void)filter_params_y;
+  (void)subpel_x_q4;
+  (void)subpel_y_q4;
+  (void)conv_params;
+
+  if (w >= 16) {
+    assert(!((intptr_t)dst % 16));
+    assert(!(dst_stride % 16));
+  }
+
+  if (w == 2) {
+    do {
+      memcpy(dst, src, 2 * sizeof(*src));
+      src += src_stride;
+      dst += dst_stride;
+      memcpy(dst, src, 2 * sizeof(*src));
+      src += src_stride;
+      dst += dst_stride;
+      h -= 2;
+    } while (h);
+  } else if (w == 4) {
+    do {
+      memcpy(dst, src, 4 * sizeof(*src));
+      src += src_stride;
+      dst += dst_stride;
+      memcpy(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);
+  }
+}
+
+void av1_jnt_convolve_2d_copy_sse2(const uint8_t *src, int src_stride,
+                                   uint8_t *dst0, int dst_stride0, int w, int h,
+                                   const InterpFilterParams *filter_params_x,
+                                   const InterpFilterParams *filter_params_y,
+                                   const int subpel_x_q4, const int subpel_y_q4,
+                                   ConvolveParams *conv_params) {
+  const int bd = 8;
+  CONV_BUF_TYPE *dst = conv_params->dst;
+  int dst_stride = conv_params->dst_stride;
+  (void)filter_params_x;
+  (void)filter_params_y;
+  (void)subpel_x_q4;
+  (void)subpel_y_q4;
+
+  const int bits =
+      FILTER_BITS * 2 - conv_params->round_1 - conv_params->round_0;
+  const int do_average = conv_params->do_average;
+  const int use_jnt_comp_avg = conv_params->use_jnt_comp_avg;
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i left_shift = _mm_cvtsi32_si128(bits);
+  int i, j;
+
+  const int w0 = conv_params->fwd_offset;
+  const int w1 = conv_params->bck_offset;
+  const __m128i wt0 = _mm_set1_epi16(w0);
+  const __m128i wt1 = _mm_set1_epi16(w1);
+  const __m128i wt = _mm_unpacklo_epi16(wt0, wt1);
+
+  const int offset_0 =
+      bd + 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+  const int offset = (1 << offset_0) + (1 << (offset_0 - 1));
+  const __m128i offset_const = _mm_set1_epi16(offset);
+  const int rounding_shift =
+      2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+  const __m128i rounding_const = _mm_set1_epi16((1 << rounding_shift) >> 1);
+
+  assert((w % 4) == 0);
+
+  if (!(w % 16)) {
+    for (i = 0; i < h; ++i) {
+      for (j = 0; j < w; j += 16) {
+        const __m128i d8 = _mm_loadu_si128((__m128i *)&src[j]);
+
+        const __m128i d16_lo = _mm_unpacklo_epi8(d8, zero);
+        const __m128i d16_hi = _mm_unpackhi_epi8(d8, zero);
+
+        const __m128i res_lo = _mm_sll_epi16(d16_lo, left_shift);
+        const __m128i res_unsigned_lo = _mm_add_epi16(res_lo, offset_const);
+
+        const __m128i res_hi = _mm_sll_epi16(d16_hi, left_shift);
+        const __m128i res_unsigned_hi = _mm_add_epi16(res_hi, offset_const);
+
+        if (do_average) {
+          const __m128i data_ref_0_lo = _mm_loadu_si128((__m128i *)(&dst[j]));
+          const __m128i data_ref_0_hi =
+              _mm_loadu_si128((__m128i *)(&dst[j + 8]));
+
+          const __m128i comp_avg_res_lo =
+              comp_avg(&data_ref_0_lo, &res_unsigned_lo, &wt, use_jnt_comp_avg);
+
+          const __m128i round_result_lo = convolve_rounding(
+              &comp_avg_res_lo, &offset_const, &rounding_const, rounding_shift);
+
+          const __m128i comp_avg_res_hi =
+              comp_avg(&data_ref_0_hi, &res_unsigned_hi, &wt, use_jnt_comp_avg);
+
+          const __m128i round_result_hi = convolve_rounding(
+              &comp_avg_res_hi, &offset_const, &rounding_const, rounding_shift);
+
+          const __m128i res_8 =
+              _mm_packus_epi16(round_result_lo, round_result_hi);
+
+          _mm_store_si128((__m128i *)(&dst0[j]), res_8);
+        } else {
+          _mm_store_si128((__m128i *)(&dst[j]), res_unsigned_lo);
+          _mm_store_si128((__m128i *)(&dst[j + 8]), res_unsigned_hi);
+        }
+      }
+      src += src_stride;
+      dst += dst_stride;
+      dst0 += dst_stride0;
+    }
+  } else {
+    for (i = 0; i < h; ++i) {
+      for (j = 0; j < w; j += 8) {
+        const __m128i d8 = _mm_loadl_epi64((__m128i *)&src[j]);
+        const __m128i d16_0 = _mm_unpacklo_epi8(d8, zero);
+
+        const __m128i res = _mm_sll_epi16(d16_0, left_shift);
+        const __m128i res_unsigned = _mm_add_epi16(res, offset_const);
+
+        if (do_average) {
+          const __m128i data_ref_0 = _mm_loadu_si128((__m128i *)(&dst[j]));
+
+          const __m128i comp_avg_res =
+              comp_avg(&data_ref_0, &res_unsigned, &wt, use_jnt_comp_avg);
+
+          const __m128i round_result = convolve_rounding(
+              &comp_avg_res, &offset_const, &rounding_const, rounding_shift);
+
+          const __m128i res_8 = _mm_packus_epi16(round_result, round_result);
+
+          if (w > 4)
+            _mm_storel_epi64((__m128i *)(&dst0[j]), res_8);
+          else
+            *(uint32_t *)(&dst0[j]) = _mm_cvtsi128_si32(res_8);
+        } else {
+          _mm_store_si128((__m128i *)(&dst[j]), res_unsigned);
+        }
+      }
+      src += src_stride;
+      dst += dst_stride;
+      dst0 += dst_stride0;
+    }
+  }
+}
diff --git a/third_party/aom/av1/common/x86/convolve_avx2.c b/third_party/aom/av1/common/x86/convolve_avx2.c
new file mode 100644
index 000000000..0e91ea947
--- /dev/null
+++ b/third_party/aom/av1/common/x86/convolve_avx2.c
@@ -0,0 +1,277 @@
+/*
+ * 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/aom_dsp_common.h"
+#include "aom_dsp/x86/convolve_avx2.h"
+#include "aom_dsp/x86/synonyms.h"
+
+void av1_convolve_y_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst,
+                            int dst_stride, int w, int h,
+                            const InterpFilterParams *filter_params_x,
+                            const InterpFilterParams *filter_params_y,
+                            const int subpel_x_q4, const int subpel_y_q4,
+                            ConvolveParams *conv_params) {
+  int i, j;
+  const int fo_vert = filter_params_y->taps / 2 - 1;
+  const uint8_t *const src_ptr = src - fo_vert * src_stride;
+
+  // right shift is F-1 because we are already dividing
+  // filter co-efficients by 2
+  const int right_shift_bits = (FILTER_BITS - 1);
+  const __m128i right_shift = _mm_cvtsi32_si128(right_shift_bits);
+  const __m256i right_shift_const =
+      _mm256_set1_epi16((1 << right_shift_bits) >> 1);
+  __m256i coeffs[4], s[8];
+
+  assert(conv_params->round_0 <= FILTER_BITS);
+  assert(((conv_params->round_0 + conv_params->round_1) <= (FILTER_BITS + 1)) ||
+         ((conv_params->round_0 + conv_params->round_1) == (2 * FILTER_BITS)));
+
+  prepare_coeffs_lowbd(filter_params_y, subpel_y_q4, coeffs);
+
+  (void)filter_params_x;
+  (void)subpel_x_q4;
+  (void)conv_params;
+
+  for (j = 0; j < w; j += 16) {
+    const uint8_t *data = &src_ptr[j];
+    __m256i src6;
+
+    // Load lines a and b. Line a to lower 128, line b to upper 128
+    const __m256i src_01a = _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);
+
+    const __m256i src_12a = _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);
+
+    const __m256i src_23a = _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);
+
+    const __m256i src_34a = _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);
+
+    const __m256i src_45a = _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)));
+    const __m256i src_56a = _mm256_permute2x128_si256(
+        _mm256_castsi128_si256(
+            _mm_loadu_si128((__m128i *)(data + 5 * src_stride))),
+        src6, 0x20);
+
+    s[0] = _mm256_unpacklo_epi8(src_01a, src_12a);
+    s[1] = _mm256_unpacklo_epi8(src_23a, src_34a);
+    s[2] = _mm256_unpacklo_epi8(src_45a, src_56a);
+
+    s[4] = _mm256_unpackhi_epi8(src_01a, src_12a);
+    s[5] = _mm256_unpackhi_epi8(src_23a, src_34a);
+    s[6] = _mm256_unpackhi_epi8(src_45a, src_56a);
+
+    for (i = 0; i < h; i += 2) {
+      data = &src_ptr[i * src_stride + j];
+      const __m256i src_67a = _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 src_78a = _mm256_permute2x128_si256(
+          _mm256_castsi128_si256(
+              _mm_loadu_si128((__m128i *)(data + 7 * src_stride))),
+          src6, 0x20);
+
+      s[3] = _mm256_unpacklo_epi8(src_67a, src_78a);
+      s[7] = _mm256_unpackhi_epi8(src_67a, src_78a);
+
+      const __m256i res_lo = convolve_lowbd(s, coeffs);
+
+      /* rounding code */
+      // shift by F - 1
+      const __m256i res_16b_lo = _mm256_sra_epi16(
+          _mm256_add_epi16(res_lo, right_shift_const), right_shift);
+      // 8 bit conversion and saturation to uint8
+      __m256i res_8b_lo = _mm256_packus_epi16(res_16b_lo, res_16b_lo);
+
+      if (w - j > 8) {
+        const __m256i res_hi = convolve_lowbd(s + 4, coeffs);
+
+        /* rounding code */
+        // shift by F - 1
+        const __m256i res_16b_hi = _mm256_sra_epi16(
+            _mm256_add_epi16(res_hi, right_shift_const), right_shift);
+        // 8 bit conversion and saturation to uint8
+        __m256i res_8b_hi = _mm256_packus_epi16(res_16b_hi, res_16b_hi);
+
+        __m256i res_a = _mm256_unpacklo_epi64(res_8b_lo, res_8b_hi);
+
+        const __m128i res_0 = _mm256_castsi256_si128(res_a);
+        const __m128i res_1 = _mm256_extracti128_si256(res_a, 1);
+
+        _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j], res_0);
+        _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j + dst_stride],
+                         res_1);
+      } else {
+        const __m128i res_0 = _mm256_castsi256_si128(res_8b_lo);
+        const __m128i res_1 = _mm256_extracti128_si256(res_8b_lo, 1);
+        if (w - j > 4) {
+          _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j], res_0);
+          _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j + dst_stride],
+                           res_1);
+        } else if (w - j > 2) {
+          xx_storel_32(&dst[i * dst_stride + j], res_0);
+          xx_storel_32(&dst[i * dst_stride + j + dst_stride], res_1);
+        } else {
+          __m128i *const p_0 = (__m128i *)&dst[i * dst_stride + j];
+          __m128i *const p_1 = (__m128i *)&dst[i * dst_stride + j + dst_stride];
+          *(uint16_t *)p_0 = _mm_cvtsi128_si32(res_0);
+          *(uint16_t *)p_1 = _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];
+    }
+  }
+}
+
+void av1_convolve_x_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst,
+                            int dst_stride, int w, int h,
+                            const InterpFilterParams *filter_params_x,
+                            const InterpFilterParams *filter_params_y,
+                            const int subpel_x_q4, const int subpel_y_q4,
+                            ConvolveParams *conv_params) {
+  int i, j;
+  const int fo_horiz = filter_params_x->taps / 2 - 1;
+  const uint8_t *const src_ptr = src - fo_horiz;
+  const int bits = FILTER_BITS - conv_params->round_0;
+
+  __m256i filt[4], coeffs[4];
+
+  filt[0] = _mm256_load_si256((__m256i const *)filt_global_avx2);
+  filt[1] = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32));
+  filt[2] = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32 * 2));
+  filt[3] = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32 * 3));
+
+  prepare_coeffs_lowbd(filter_params_x, subpel_x_q4, coeffs);
+
+  const __m256i round_0_const =
+      _mm256_set1_epi16((1 << (conv_params->round_0 - 1)) >> 1);
+  const __m128i round_0_shift = _mm_cvtsi32_si128(conv_params->round_0 - 1);
+  const __m256i round_const = _mm256_set1_epi16((1 << bits) >> 1);
+  const __m128i round_shift = _mm_cvtsi32_si128(bits);
+
+  (void)filter_params_y;
+  (void)subpel_y_q4;
+
+  assert(bits >= 0);
+  assert((FILTER_BITS - conv_params->round_1) >= 0 ||
+         ((conv_params->round_0 + conv_params->round_1) == 2 * FILTER_BITS));
+  assert(conv_params->round_0 > 0);
+
+  if (w <= 8) {
+    for (i = 0; i < h; i += 2) {
+      const __m256i data = _mm256_permute2x128_si256(
+          _mm256_castsi128_si256(
+              _mm_loadu_si128((__m128i *)(&src_ptr[i * src_stride]))),
+          _mm256_castsi128_si256(_mm_loadu_si128(
+              (__m128i *)(&src_ptr[i * src_stride + src_stride]))),
+          0x20);
+
+      __m256i res_16b = convolve_lowbd_x(data, coeffs, filt);
+
+      res_16b = _mm256_sra_epi16(_mm256_add_epi16(res_16b, round_0_const),
+                                 round_0_shift);
+
+      res_16b =
+          _mm256_sra_epi16(_mm256_add_epi16(res_16b, round_const), round_shift);
+
+      /* rounding code */
+      // 8 bit conversion and saturation to uint8
+      __m256i res_8b = _mm256_packus_epi16(res_16b, res_16b);
+
+      const __m128i res_0 = _mm256_castsi256_si128(res_8b);
+      const __m128i res_1 = _mm256_extracti128_si256(res_8b, 1);
+      if (w > 4) {
+        _mm_storel_epi64((__m128i *)&dst[i * dst_stride], res_0);
+        _mm_storel_epi64((__m128i *)&dst[i * dst_stride + dst_stride], res_1);
+      } else if (w > 2) {
+        xx_storel_32(&dst[i * dst_stride], res_0);
+        xx_storel_32(&dst[i * dst_stride + dst_stride], res_1);
+      } else {
+        __m128i *const p_0 = (__m128i *)&dst[i * dst_stride];
+        __m128i *const p_1 = (__m128i *)&dst[i * dst_stride + dst_stride];
+        *(uint16_t *)p_0 = _mm_cvtsi128_si32(res_0);
+        *(uint16_t *)p_1 = _mm_cvtsi128_si32(res_1);
+      }
+    }
+  } else {
+    for (i = 0; i < h; ++i) {
+      for (j = 0; j < w; j += 16) {
+        // 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 8 9 10 11 12 13 14 15 16 17 18
+        // 19 20 21 22 23
+        const __m256i data = _mm256_inserti128_si256(
+            _mm256_loadu_si256((__m256i *)&src_ptr[(i * src_stride) + j]),
+            _mm_loadu_si128((__m128i *)&src_ptr[(i * src_stride) + (j + 8)]),
+            1);
+
+        __m256i res_16b = convolve_lowbd_x(data, coeffs, filt);
+
+        res_16b = _mm256_sra_epi16(_mm256_add_epi16(res_16b, round_0_const),
+                                   round_0_shift);
+
+        res_16b = _mm256_sra_epi16(_mm256_add_epi16(res_16b, round_const),
+                                   round_shift);
+
+        /* rounding code */
+        // 8 bit conversion and saturation to uint8
+        __m256i res_8b = _mm256_packus_epi16(res_16b, res_16b);
+
+        // Store values into the destination buffer
+        // 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
+        res_8b = _mm256_permute4x64_epi64(res_8b, 216);
+        __m128i res = _mm256_castsi256_si128(res_8b);
+        _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j], res);
+      }
+    }
+  }
+}
diff --git a/third_party/aom/av1/common/x86/convolve_sse2.c b/third_party/aom/av1/common/x86/convolve_sse2.c
new file mode 100644
index 000000000..5016642de
--- /dev/null
+++ b/third_party/aom/av1/common/x86/convolve_sse2.c
@@ -0,0 +1,338 @@
+/*
+ * 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/av1_rtcd.h"
+
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/aom_filter.h"
+#include "aom_dsp/x86/convolve_common_intrin.h"
+#include "av1/common/convolve.h"
+
+static INLINE void prepare_coeffs(const InterpFilterParams *const filter_params,
+                                  const int subpel_q4,
+                                  __m128i *const coeffs /* [4] */) {
+  const int16_t *const y_filter = av1_get_interp_filter_subpel_kernel(
+      filter_params, subpel_q4 & SUBPEL_MASK);
+  const __m128i coeffs_y = _mm_loadu_si128((__m128i *)y_filter);
+  // coeffs 0 1 0 1 2 3 2 3
+  const __m128i tmp_0 = _mm_unpacklo_epi32(coeffs_y, coeffs_y);
+  // coeffs 4 5 4 5 6 7 6 7
+  const __m128i tmp_1 = _mm_unpackhi_epi32(coeffs_y, coeffs_y);
+
+  coeffs[0] = _mm_unpacklo_epi64(tmp_0, tmp_0);  // coeffs 0 1 0 1 0 1 0 1
+  coeffs[1] = _mm_unpackhi_epi64(tmp_0, tmp_0);  // coeffs 2 3 2 3 2 3 2 3
+  coeffs[2] = _mm_unpacklo_epi64(tmp_1, tmp_1);  // coeffs 4 5 4 5 4 5 4 5
+  coeffs[3] = _mm_unpackhi_epi64(tmp_1, tmp_1);  // coeffs 6 7 6 7 6 7 6 7
+}
+
+static INLINE __m128i convolve(const __m128i *const s,
+                               const __m128i *const 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 d = _mm_add_epi32(_mm_add_epi32(d0, d1), _mm_add_epi32(d2, d3));
+  return d;
+}
+
+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);
+}
+
+void av1_convolve_y_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst,
+                            int dst_stride, int w, int h,
+                            const InterpFilterParams *filter_params_x,
+                            const InterpFilterParams *filter_params_y,
+                            const int subpel_x_q4, const int subpel_y_q4,
+                            ConvolveParams *conv_params) {
+  const int fo_vert = filter_params_y->taps / 2 - 1;
+  const uint8_t *src_ptr = src - fo_vert * src_stride;
+  const __m128i round_const = _mm_set1_epi32((1 << FILTER_BITS) >> 1);
+  const __m128i round_shift = _mm_cvtsi32_si128(FILTER_BITS);
+  __m128i coeffs[4];
+
+  (void)filter_params_x;
+  (void)subpel_x_q4;
+  (void)conv_params;
+
+  assert(conv_params->round_0 <= FILTER_BITS);
+  assert(((conv_params->round_0 + conv_params->round_1) <= (FILTER_BITS + 1)) ||
+         ((conv_params->round_0 + conv_params->round_1) == (2 * FILTER_BITS)));
+
+  prepare_coeffs(filter_params_y, subpel_y_q4, coeffs);
+
+  if (w <= 4) {
+    __m128i s[8], src6, res, res_round, res16;
+    uint32_t res_int;
+    src6 = _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 6 * src_stride));
+    s[0] = _mm_unpacklo_epi8(
+        _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 0 * src_stride)),
+        _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 1 * src_stride)));
+    s[1] = _mm_unpacklo_epi8(
+        _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 1 * src_stride)),
+        _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 2 * src_stride)));
+    s[2] = _mm_unpacklo_epi8(
+        _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 2 * src_stride)),
+        _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 3 * src_stride)));
+    s[3] = _mm_unpacklo_epi8(
+        _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 3 * src_stride)),
+        _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 4 * src_stride)));
+    s[4] = _mm_unpacklo_epi8(
+        _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 4 * src_stride)),
+        _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 5 * src_stride)));
+    s[5] = _mm_unpacklo_epi8(
+        _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 5 * src_stride)), src6);
+
+    do {
+      s[6] = _mm_unpacklo_epi8(
+          src6, _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 7 * src_stride)));
+      src6 = _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 8 * src_stride));
+      s[7] = _mm_unpacklo_epi8(
+          _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 7 * src_stride)), src6);
+
+      res = convolve_lo_y(s + 0, coeffs);
+      res_round = _mm_sra_epi32(_mm_add_epi32(res, round_const), round_shift);
+      res16 = _mm_packs_epi32(res_round, res_round);
+      res_int = _mm_cvtsi128_si32(_mm_packus_epi16(res16, res16));
+
+      if (w == 2)
+        *(uint16_t *)dst = res_int;
+      else
+        *(uint32_t *)dst = res_int;
+
+      src_ptr += src_stride;
+      dst += dst_stride;
+
+      res = convolve_lo_y(s + 1, coeffs);
+      res_round = _mm_sra_epi32(_mm_add_epi32(res, round_const), round_shift);
+      res16 = _mm_packs_epi32(res_round, res_round);
+      res_int = _mm_cvtsi128_si32(_mm_packus_epi16(res16, res16));
+
+      if (w == 2)
+        *(uint16_t *)dst = res_int;
+      else
+        *(uint32_t *)dst = res_int;
+
+      src_ptr += src_stride;
+      dst += dst_stride;
+
+      s[0] = s[2];
+      s[1] = s[3];
+      s[2] = s[4];
+      s[3] = s[5];
+      s[4] = s[6];
+      s[5] = s[7];
+      h -= 2;
+    } while (h);
+  } else {
+    assert(!(w % 8));
+    int j = 0;
+    do {
+      __m128i s[8], src6, res_lo, res_hi;
+      __m128i res_lo_round, res_hi_round, res16, res;
+      const uint8_t *data = &src_ptr[j];
+
+      src6 = _mm_loadl_epi64((__m128i *)(data + 6 * src_stride));
+      s[0] = _mm_unpacklo_epi8(
+          _mm_loadl_epi64((__m128i *)(data + 0 * src_stride)),
+          _mm_loadl_epi64((__m128i *)(data + 1 * src_stride)));
+      s[1] = _mm_unpacklo_epi8(
+          _mm_loadl_epi64((__m128i *)(data + 1 * src_stride)),
+          _mm_loadl_epi64((__m128i *)(data + 2 * src_stride)));
+      s[2] = _mm_unpacklo_epi8(
+          _mm_loadl_epi64((__m128i *)(data + 2 * src_stride)),
+          _mm_loadl_epi64((__m128i *)(data + 3 * src_stride)));
+      s[3] = _mm_unpacklo_epi8(
+          _mm_loadl_epi64((__m128i *)(data + 3 * src_stride)),
+          _mm_loadl_epi64((__m128i *)(data + 4 * src_stride)));
+      s[4] = _mm_unpacklo_epi8(
+          _mm_loadl_epi64((__m128i *)(data + 4 * src_stride)),
+          _mm_loadl_epi64((__m128i *)(data + 5 * src_stride)));
+      s[5] = _mm_unpacklo_epi8(
+          _mm_loadl_epi64((__m128i *)(data + 5 * src_stride)), src6);
+
+      int i = 0;
+      do {
+        data = &src_ptr[i * src_stride + j];
+        s[6] = _mm_unpacklo_epi8(
+            src6, _mm_loadl_epi64((__m128i *)(data + 7 * src_stride)));
+        src6 = _mm_loadl_epi64((__m128i *)(data + 8 * src_stride));
+        s[7] = _mm_unpacklo_epi8(
+            _mm_loadl_epi64((__m128i *)(data + 7 * src_stride)), src6);
+
+        res_lo = convolve_lo_y(s, coeffs);  // Filter low index pixels
+        res_hi = convolve_hi_y(s, coeffs);  // Filter high index pixels
+
+        res_lo_round =
+            _mm_sra_epi32(_mm_add_epi32(res_lo, round_const), round_shift);
+        res_hi_round =
+            _mm_sra_epi32(_mm_add_epi32(res_hi, round_const), round_shift);
+
+        res16 = _mm_packs_epi32(res_lo_round, res_hi_round);
+        res = _mm_packus_epi16(res16, res16);
+
+        _mm_storel_epi64((__m128i *)(dst + i * dst_stride + j), res);
+        i++;
+
+        res_lo = convolve_lo_y(s + 1, coeffs);  // Filter low index pixels
+        res_hi = convolve_hi_y(s + 1, coeffs);  // Filter high index pixels
+
+        res_lo_round =
+            _mm_sra_epi32(_mm_add_epi32(res_lo, round_const), round_shift);
+        res_hi_round =
+            _mm_sra_epi32(_mm_add_epi32(res_hi, round_const), round_shift);
+
+        res16 = _mm_packs_epi32(res_lo_round, res_hi_round);
+        res = _mm_packus_epi16(res16, res16);
+
+        _mm_storel_epi64((__m128i *)(dst + i * dst_stride + j), res);
+        i++;
+
+        s[0] = s[2];
+        s[1] = s[3];
+        s[2] = s[4];
+        s[3] = s[5];
+        s[4] = s[6];
+        s[5] = s[7];
+      } while (i < h);
+      j += 8;
+    } while (j < w);
+  }
+}
+
+void av1_convolve_x_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst,
+                            int dst_stride, int w, int h,
+                            const InterpFilterParams *filter_params_x,
+                            const InterpFilterParams *filter_params_y,
+                            const int subpel_x_q4, const int subpel_y_q4,
+                            ConvolveParams *conv_params) {
+  const int fo_horiz = filter_params_x->taps / 2 - 1;
+  const uint8_t *src_ptr = src - fo_horiz;
+  const int bits = FILTER_BITS - conv_params->round_0;
+  const __m128i round_0_const =
+      _mm_set1_epi32((1 << conv_params->round_0) >> 1);
+  const __m128i round_const = _mm_set1_epi32((1 << bits) >> 1);
+  const __m128i round_0_shift = _mm_cvtsi32_si128(conv_params->round_0);
+  const __m128i round_shift = _mm_cvtsi32_si128(bits);
+  __m128i coeffs[4];
+
+  (void)filter_params_y;
+  (void)subpel_y_q4;
+
+  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_q4, coeffs);
+
+  if (w <= 4) {
+    do {
+      const __m128i data = _mm_loadu_si128((__m128i *)src_ptr);
+      __m128i s[4];
+
+      s[0] = _mm_unpacklo_epi8(data, _mm_srli_si128(data, 1));
+      s[1] =
+          _mm_unpacklo_epi8(_mm_srli_si128(data, 2), _mm_srli_si128(data, 3));
+      s[2] =
+          _mm_unpacklo_epi8(_mm_srli_si128(data, 4), _mm_srli_si128(data, 5));
+      s[3] =
+          _mm_unpacklo_epi8(_mm_srli_si128(data, 6), _mm_srli_si128(data, 7));
+      const __m128i res_lo = convolve_lo_x(s, coeffs);
+      __m128i res_lo_round =
+          _mm_sra_epi32(_mm_add_epi32(res_lo, round_0_const), round_0_shift);
+      res_lo_round =
+          _mm_sra_epi32(_mm_add_epi32(res_lo_round, round_const), round_shift);
+
+      const __m128i res16 = _mm_packs_epi32(res_lo_round, res_lo_round);
+      const __m128i res = _mm_packus_epi16(res16, res16);
+
+      uint32_t r = _mm_cvtsi128_si32(res);
+      if (w == 2)
+        *(uint16_t *)dst = r;
+      else
+        *(uint32_t *)dst = r;
+
+      src_ptr += src_stride;
+      dst += dst_stride;
+    } while (--h);
+  } else {
+    assert(!(w % 8));
+    int i = 0;
+    do {
+      int j = 0;
+      do {
+        const __m128i data =
+            _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + j]);
+        __m128i s[4];
+
+        // Filter even-index pixels
+        s[0] = data;
+        s[1] = _mm_srli_si128(data, 2);
+        s[2] = _mm_srli_si128(data, 4);
+        s[3] = _mm_srli_si128(data, 6);
+        const __m128i res_even = convolve_lo_x(s, coeffs);
+
+        // Filter odd-index pixels
+        s[0] = _mm_srli_si128(data, 1);
+        s[1] = _mm_srli_si128(data, 3);
+        s[2] = _mm_srli_si128(data, 5);
+        s[3] = _mm_srli_si128(data, 7);
+        const __m128i res_odd = convolve_lo_x(s, coeffs);
+
+        // Rearrange pixels back into the order 0 ... 7
+        const __m128i res_lo = _mm_unpacklo_epi32(res_even, res_odd);
+        const __m128i res_hi = _mm_unpackhi_epi32(res_even, res_odd);
+        __m128i res_lo_round =
+            _mm_sra_epi32(_mm_add_epi32(res_lo, round_0_const), round_0_shift);
+        res_lo_round = _mm_sra_epi32(_mm_add_epi32(res_lo_round, round_const),
+                                     round_shift);
+        __m128i res_hi_round =
+            _mm_sra_epi32(_mm_add_epi32(res_hi, round_0_const), round_0_shift);
+        res_hi_round = _mm_sra_epi32(_mm_add_epi32(res_hi_round, round_const),
+                                     round_shift);
+
+        const __m128i res16 = _mm_packs_epi32(res_lo_round, res_hi_round);
+        const __m128i res = _mm_packus_epi16(res16, res16);
+
+        _mm_storel_epi64((__m128i *)(dst + i * dst_stride + j), res);
+        j += 8;
+      } while (j < w);
+    } while (++i < h);
+  }
+}
diff --git a/third_party/aom/av1/common/x86/filterintra_sse4.c b/third_party/aom/av1/common/x86/filterintra_sse4.c
new file mode 100644
index 000000000..c11edc1d4
--- /dev/null
+++ b/third_party/aom/av1/common/x86/filterintra_sse4.c
@@ -0,0 +1,75 @@
+/*
+ * 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 "config/av1_rtcd.h"
+
+#include "aom_dsp/x86/synonyms.h"
+#include "av1/common/enums.h"
+#include "av1/common/reconintra.h"
+
+void av1_filter_intra_predictor_sse4_1(uint8_t *dst, ptrdiff_t stride,
+                                       TX_SIZE tx_size, const uint8_t *above,
+                                       const uint8_t *left, int mode) {
+  int r, c;
+  uint8_t buffer[33][33];
+  const int bw = tx_size_wide[tx_size];
+  const int bh = tx_size_high[tx_size];
+
+  assert(bw <= 32 && bh <= 32);
+
+  // The initialization is just for silencing Jenkins static analysis warnings
+  for (r = 0; r < bh + 1; ++r)
+    memset(buffer[r], 0, (bw + 1) * sizeof(buffer[0][0]));
+
+  for (r = 0; r < bh; ++r) buffer[r + 1][0] = left[r];
+  memcpy(buffer[0], &above[-1], (bw + 1) * sizeof(uint8_t));
+
+  const __m128i f1f0 = xx_load_128(av1_filter_intra_taps[mode][0]);
+  const __m128i f3f2 = xx_load_128(av1_filter_intra_taps[mode][2]);
+  const __m128i f5f4 = xx_load_128(av1_filter_intra_taps[mode][4]);
+  const __m128i f7f6 = xx_load_128(av1_filter_intra_taps[mode][6]);
+  const __m128i filter_intra_scale_bits =
+      _mm_set1_epi16(1 << (15 - FILTER_INTRA_SCALE_BITS));
+
+  for (r = 1; r < bh + 1; r += 2) {
+    for (c = 1; c < bw + 1; c += 4) {
+      DECLARE_ALIGNED(16, uint8_t, p[8]);
+      memcpy(p, &buffer[r - 1][c - 1], 5 * sizeof(uint8_t));
+      p[5] = buffer[r][c - 1];
+      p[6] = buffer[r + 1][c - 1];
+      p[7] = 0;
+      const __m128i p_b = xx_loadl_64(p);
+      const __m128i in = _mm_unpacklo_epi64(p_b, p_b);
+      const __m128i out_01 = _mm_maddubs_epi16(in, f1f0);
+      const __m128i out_23 = _mm_maddubs_epi16(in, f3f2);
+      const __m128i out_45 = _mm_maddubs_epi16(in, f5f4);
+      const __m128i out_67 = _mm_maddubs_epi16(in, f7f6);
+      const __m128i out_0123 = _mm_hadd_epi16(out_01, out_23);
+      const __m128i out_4567 = _mm_hadd_epi16(out_45, out_67);
+      const __m128i out_01234567 = _mm_hadd_epi16(out_0123, out_4567);
+      // Rounding
+      const __m128i round_w =
+          _mm_mulhrs_epi16(out_01234567, filter_intra_scale_bits);
+      const __m128i out_r = _mm_packus_epi16(round_w, round_w);
+      const __m128i out_r1 = _mm_srli_si128(out_r, 4);
+      // Storing
+      xx_storel_32(&buffer[r][c], out_r);
+      xx_storel_32(&buffer[r + 1][c], out_r1);
+    }
+  }
+
+  for (r = 0; r < bh; ++r) {
+    memcpy(dst, &buffer[r + 1][1], bw * sizeof(uint8_t));
+    dst += stride;
+  }
+}
diff --git a/third_party/aom/av1/common/x86/highbd_convolve_2d_avx2.c b/third_party/aom/av1/common/x86/highbd_convolve_2d_avx2.c
new file mode 100644
index 000000000..ae68f0bbb
--- /dev/null
+++ b/third_party/aom/av1/common/x86/highbd_convolve_2d_avx2.c
@@ -0,0 +1,326 @@
+/*
+ * 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 <assert.h>
+
+#include "config/aom_dsp_rtcd.h"
+
+#include "aom_dsp/x86/convolve_avx2.h"
+#include "aom_dsp/x86/synonyms.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/aom_filter.h"
+#include "av1/common/convolve.h"
+
+void av1_highbd_convolve_2d_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 InterpFilterParams *filter_params_y,
+                                    const int subpel_x_q4,
+                                    const int subpel_y_q4,
+                                    ConvolveParams *conv_params, int bd) {
+  DECLARE_ALIGNED(32, int16_t, im_block[(MAX_SB_SIZE + MAX_FILTER_TAP) * 8]);
+  int im_h = h + filter_params_y->taps - 1;
+  int im_stride = 8;
+  int i, j;
+  const int fo_vert = filter_params_y->taps / 2 - 1;
+  const int fo_horiz = filter_params_x->taps / 2 - 1;
+  const uint16_t *const src_ptr = src - fo_vert * src_stride - 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[8], coeffs_y[4], coeffs_x[4];
+
+  const __m256i round_const_x = _mm256_set1_epi32(
+      ((1 << conv_params->round_0) >> 1) + (1 << (bd + FILTER_BITS - 1)));
+  const __m128i round_shift_x = _mm_cvtsi32_si128(conv_params->round_0);
+
+  const __m256i round_const_y = _mm256_set1_epi32(
+      ((1 << conv_params->round_1) >> 1) -
+      (1 << (bd + 2 * FILTER_BITS - conv_params->round_0 - 1)));
+  const __m128i round_shift_y = _mm_cvtsi32_si128(conv_params->round_1);
+
+  const int bits =
+      FILTER_BITS * 2 - conv_params->round_0 - conv_params->round_1;
+  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_x, subpel_x_q4, coeffs_x);
+  prepare_coeffs(filter_params_y, subpel_y_q4, coeffs_y);
+
+  for (j = 0; j < w; j += 8) {
+    /* Horizontal filter */
+    {
+      for (i = 0; i < im_h; i += 2) {
+        const __m256i row0 =
+            _mm256_loadu_si256((__m256i *)&src_ptr[i * src_stride + j]);
+        __m256i row1 = _mm256_set1_epi16(0);
+        if (i + 1 < im_h)
+          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);
+
+        __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);
+
+        _mm256_store_si256((__m256i *)&im_block[i * im_stride], res);
+      }
+    }
+
+    /* Vertical filter */
+    {
+      __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);
+        __m256i res_a_round = _mm256_sra_epi32(
+            _mm256_add_epi32(res_a, round_const_y), round_shift_y);
+
+        res_a_round = _mm256_sra_epi32(
+            _mm256_add_epi32(res_a_round, 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_y), round_shift_y);
+          res_b_round =
+              _mm256_sra_epi32(_mm256_add_epi32(res_b_round, 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((__m128i *)&dst[i * dst_stride + j],
+                       _mm256_castsi256_si128(res_a_round));
+          xx_storel_32((__m128i *)&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];
+      }
+    }
+  }
+}
+
+static INLINE void 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 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 av1_highbd_convolve_2d_copy_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 InterpFilterParams *filter_params_y, const int subpel_x_q4,
+    const int subpel_y_q4, ConvolveParams *conv_params, int bd) {
+  (void)filter_params_x;
+  (void)filter_params_y;
+  (void)subpel_x_q4;
+  (void)subpel_y_q4;
+  (void)conv_params;
+  (void)bd;
+
+  if (w >= 16) {
+    assert(!((intptr_t)dst % 16));
+    assert(!(dst_stride % 16));
+  }
+
+  if (w == 2) {
+    do {
+      memcpy(dst, src, 2 * sizeof(*src));
+      src += src_stride;
+      dst += dst_stride;
+      memcpy(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 {
+      copy_64(src, dst);
+      src += src_stride;
+      dst += dst_stride;
+      copy_64(src, dst);
+      src += src_stride;
+      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);
+  }
+}
diff --git a/third_party/aom/av1/common/x86/highbd_convolve_2d_sse2.c b/third_party/aom/av1/common/x86/highbd_convolve_2d_sse2.c
new file mode 100644
index 000000000..15f8872c1
--- /dev/null
+++ b/third_party/aom/av1/common/x86/highbd_convolve_2d_sse2.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 <emmintrin.h>
+#include <assert.h>
+
+#include "config/aom_dsp_rtcd.h"
+
+#include "aom_dsp/aom_filter.h"
+
+static INLINE void 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 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 av1_highbd_convolve_2d_copy_sr_sse2(
+    const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w,
+    int h, const InterpFilterParams *filter_params_x,
+    const InterpFilterParams *filter_params_y, const int subpel_x_q4,
+    const int subpel_y_q4, ConvolveParams *conv_params, int bd) {
+  (void)filter_params_x;
+  (void)filter_params_y;
+  (void)subpel_x_q4;
+  (void)subpel_y_q4;
+  (void)conv_params;
+  (void)bd;
+  if (w >= 16) {
+    assert(!((intptr_t)dst % 16));
+    assert(!(dst_stride % 16));
+  }
+
+  if (w == 2) {
+    do {
+      __m128i s = _mm_loadl_epi64((__m128i *)src);
+      *(uint32_t *)dst = _mm_cvtsi128_si32(s);
+      src += src_stride;
+      dst += dst_stride;
+      s = _mm_loadl_epi64((__m128i *)src);
+      *(uint32_t *)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 {
+      copy_64(src, dst);
+      src += src_stride;
+      dst += dst_stride;
+      copy_64(src, dst);
+      src += src_stride;
+      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);
+  }
+}
diff --git a/third_party/aom/av1/common/x86/highbd_convolve_2d_sse4.c b/third_party/aom/av1/common/x86/highbd_convolve_2d_sse4.c
new file mode 100644
index 000000000..3f8dafb4b
--- /dev/null
+++ b/third_party/aom/av1/common/x86/highbd_convolve_2d_sse4.c
@@ -0,0 +1,420 @@
+/*
+ * 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 <smmintrin.h>
+#include <assert.h>
+
+#include "config/aom_dsp_rtcd.h"
+
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/aom_filter.h"
+#include "aom_dsp/x86/convolve_sse2.h"
+#include "aom_dsp/x86/convolve_sse4_1.h"
+#include "av1/common/convolve.h"
+
+void av1_highbd_jnt_convolve_2d_copy_sse4_1(
+    const uint16_t *src, int src_stride, uint16_t *dst0, int dst_stride0, int w,
+    int h, const InterpFilterParams *filter_params_x,
+    const InterpFilterParams *filter_params_y, const int subpel_x_q4,
+    const int subpel_y_q4, ConvolveParams *conv_params, int bd) {
+  CONV_BUF_TYPE *dst = conv_params->dst;
+  int dst_stride = conv_params->dst_stride;
+  (void)filter_params_x;
+  (void)filter_params_y;
+  (void)subpel_x_q4;
+  (void)subpel_y_q4;
+
+  const int bits =
+      FILTER_BITS * 2 - conv_params->round_1 - conv_params->round_0;
+  const __m128i left_shift = _mm_cvtsi32_si128(bits);
+  const int do_average = conv_params->do_average;
+  const int use_jnt_comp_avg = conv_params->use_jnt_comp_avg;
+  const int w0 = conv_params->fwd_offset;
+  const int w1 = conv_params->bck_offset;
+  const __m128i wt0 = _mm_set1_epi32(w0);
+  const __m128i wt1 = _mm_set1_epi32(w1);
+  const __m128i zero = _mm_setzero_si128();
+  int i, j;
+
+  const int offset_0 =
+      bd + 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+  const int offset = (1 << offset_0) + (1 << (offset_0 - 1));
+  const __m128i offset_const = _mm_set1_epi32(offset);
+  const __m128i offset_const_16b = _mm_set1_epi16(offset);
+  const int rounding_shift =
+      2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+  const __m128i rounding_const = _mm_set1_epi32((1 << rounding_shift) >> 1);
+  const __m128i clip_pixel_to_bd =
+      _mm_set1_epi16(bd == 10 ? 1023 : (bd == 12 ? 4095 : 255));
+
+  assert(bits <= 4);
+
+  if (!(w % 8)) {
+    for (i = 0; i < h; i += 1) {
+      for (j = 0; j < w; j += 8) {
+        const __m128i src_16bit =
+            _mm_loadu_si128((__m128i *)(&src[i * src_stride + j]));
+        const __m128i res = _mm_sll_epi16(src_16bit, left_shift);
+        if (do_average) {
+          const __m128i data_0 =
+              _mm_loadu_si128((__m128i *)(&dst[i * dst_stride + j]));
+
+          const __m128i data_ref_0_lo = _mm_unpacklo_epi16(data_0, zero);
+          const __m128i data_ref_0_hi = _mm_unpackhi_epi16(data_0, zero);
+
+          const __m128i res_32b_lo = _mm_unpacklo_epi16(res, zero);
+          const __m128i res_unsigned_lo =
+              _mm_add_epi32(res_32b_lo, offset_const);
+
+          const __m128i comp_avg_res_lo = highbd_comp_avg_sse4_1(
+              &data_ref_0_lo, &res_unsigned_lo, &wt0, &wt1, use_jnt_comp_avg);
+
+          const __m128i res_32b_hi = _mm_unpackhi_epi16(res, zero);
+          const __m128i res_unsigned_hi =
+              _mm_add_epi32(res_32b_hi, offset_const);
+
+          const __m128i comp_avg_res_hi = highbd_comp_avg_sse4_1(
+              &data_ref_0_hi, &res_unsigned_hi, &wt0, &wt1, use_jnt_comp_avg);
+
+          const __m128i round_result_lo = highbd_convolve_rounding_sse2(
+              &comp_avg_res_lo, &offset_const, &rounding_const, rounding_shift);
+          const __m128i round_result_hi = highbd_convolve_rounding_sse2(
+              &comp_avg_res_hi, &offset_const, &rounding_const, rounding_shift);
+
+          const __m128i res_16b =
+              _mm_packus_epi32(round_result_lo, round_result_hi);
+          const __m128i res_clip = _mm_min_epi16(res_16b, clip_pixel_to_bd);
+
+          _mm_store_si128((__m128i *)(&dst0[i * dst_stride0 + j]), res_clip);
+        } else {
+          const __m128i res_unsigned_16b =
+              _mm_adds_epu16(res, offset_const_16b);
+
+          _mm_store_si128((__m128i *)(&dst[i * dst_stride + j]),
+                          res_unsigned_16b);
+        }
+      }
+    }
+  } else if (!(w % 4)) {
+    for (i = 0; i < h; i += 2) {
+      for (j = 0; j < w; j += 4) {
+        const __m128i src_row_0 =
+            _mm_loadl_epi64((__m128i *)(&src[i * src_stride + j]));
+        const __m128i src_row_1 =
+            _mm_loadl_epi64((__m128i *)(&src[i * src_stride + j + src_stride]));
+        const __m128i src_10 = _mm_unpacklo_epi64(src_row_0, src_row_1);
+
+        const __m128i res = _mm_sll_epi16(src_10, left_shift);
+
+        if (do_average) {
+          const __m128i data_0 =
+              _mm_loadl_epi64((__m128i *)(&dst[i * dst_stride + j]));
+          const __m128i data_1 = _mm_loadl_epi64(
+              (__m128i *)(&dst[i * dst_stride + j + dst_stride]));
+
+          const __m128i data_ref_0 = _mm_unpacklo_epi16(data_0, zero);
+          const __m128i data_ref_1 = _mm_unpacklo_epi16(data_1, zero);
+
+          const __m128i res_32b = _mm_unpacklo_epi16(res, zero);
+          const __m128i res_unsigned_lo = _mm_add_epi32(res_32b, offset_const);
+
+          const __m128i res_32b_hi = _mm_unpackhi_epi16(res, zero);
+          const __m128i res_unsigned_hi =
+              _mm_add_epi32(res_32b_hi, offset_const);
+
+          const __m128i comp_avg_res_lo = highbd_comp_avg_sse4_1(
+              &data_ref_0, &res_unsigned_lo, &wt0, &wt1, use_jnt_comp_avg);
+          const __m128i comp_avg_res_hi = highbd_comp_avg_sse4_1(
+              &data_ref_1, &res_unsigned_hi, &wt0, &wt1, use_jnt_comp_avg);
+
+          const __m128i round_result_lo = highbd_convolve_rounding_sse2(
+              &comp_avg_res_lo, &offset_const, &rounding_const, rounding_shift);
+          const __m128i round_result_hi = highbd_convolve_rounding_sse2(
+              &comp_avg_res_hi, &offset_const, &rounding_const, rounding_shift);
+
+          const __m128i res_16b =
+              _mm_packus_epi32(round_result_lo, round_result_hi);
+          const __m128i res_clip = _mm_min_epi16(res_16b, clip_pixel_to_bd);
+
+          const __m128i res_1 = _mm_srli_si128(res_clip, 8);
+
+          _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_clip);
+          _mm_storel_epi64(
+              (__m128i *)(&dst0[i * dst_stride0 + j + dst_stride0]), res_1);
+        } else {
+          const __m128i res_unsigned_16b =
+              _mm_adds_epu16(res, offset_const_16b);
+
+          const __m128i res_1 = _mm_srli_si128(res_unsigned_16b, 8);
+
+          _mm_storel_epi64((__m128i *)(&dst[i * dst_stride + j]),
+                           res_unsigned_16b);
+          _mm_storel_epi64((__m128i *)(&dst[i * dst_stride + j + dst_stride]),
+                           res_1);
+        }
+      }
+    }
+  }
+}
+
+void av1_highbd_jnt_convolve_2d_sse4_1(
+    const uint16_t *src, int src_stride, uint16_t *dst0, int dst_stride0, int w,
+    int h, const InterpFilterParams *filter_params_x,
+    const InterpFilterParams *filter_params_y, const int subpel_x_q4,
+    const int subpel_y_q4, ConvolveParams *conv_params, int bd) {
+  DECLARE_ALIGNED(16, int16_t,
+                  im_block[(MAX_SB_SIZE + MAX_FILTER_TAP - 1) * MAX_SB_SIZE]);
+  CONV_BUF_TYPE *dst = conv_params->dst;
+  int dst_stride = conv_params->dst_stride;
+  int im_h = h + filter_params_y->taps - 1;
+  int im_stride = MAX_SB_SIZE;
+  int i, j;
+  const int do_average = conv_params->do_average;
+  const int use_jnt_comp_avg = conv_params->use_jnt_comp_avg;
+  const int fo_vert = filter_params_y->taps / 2 - 1;
+  const int fo_horiz = filter_params_x->taps / 2 - 1;
+  const uint16_t *const src_ptr = src - fo_vert * src_stride - fo_horiz;
+
+  const int w0 = conv_params->fwd_offset;
+  const int w1 = conv_params->bck_offset;
+  const __m128i wt0 = _mm_set1_epi32(w0);
+  const __m128i wt1 = _mm_set1_epi32(w1);
+
+  const int offset_0 =
+      bd + 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+  const int offset = (1 << offset_0) + (1 << (offset_0 - 1));
+  const __m128i offset_const = _mm_set1_epi32(offset);
+  const int rounding_shift =
+      2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+  const __m128i rounding_const = _mm_set1_epi32((1 << rounding_shift) >> 1);
+  const __m128i clip_pixel_to_bd =
+      _mm_set1_epi16(bd == 10 ? 1023 : (bd == 12 ? 4095 : 255));
+
+  // 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);
+
+  /* Horizontal filter */
+  {
+    const int16_t *x_filter = av1_get_interp_filter_subpel_kernel(
+        filter_params_x, subpel_x_q4 & SUBPEL_MASK);
+    const __m128i coeffs_x = _mm_loadu_si128((__m128i *)x_filter);
+
+    // coeffs 0 1 0 1 2 3 2 3
+    const __m128i tmp_0 = _mm_unpacklo_epi32(coeffs_x, coeffs_x);
+    // coeffs 4 5 4 5 6 7 6 7
+    const __m128i tmp_1 = _mm_unpackhi_epi32(coeffs_x, coeffs_x);
+
+    // coeffs 0 1 0 1 0 1 0 1
+    const __m128i coeff_01 = _mm_unpacklo_epi64(tmp_0, tmp_0);
+    // coeffs 2 3 2 3 2 3 2 3
+    const __m128i coeff_23 = _mm_unpackhi_epi64(tmp_0, tmp_0);
+    // coeffs 4 5 4 5 4 5 4 5
+    const __m128i coeff_45 = _mm_unpacklo_epi64(tmp_1, tmp_1);
+    // coeffs 6 7 6 7 6 7 6 7
+    const __m128i coeff_67 = _mm_unpackhi_epi64(tmp_1, tmp_1);
+
+    const __m128i round_const = _mm_set1_epi32(
+        ((1 << conv_params->round_0) >> 1) + (1 << (bd + FILTER_BITS - 1)));
+    const __m128i round_shift = _mm_cvtsi32_si128(conv_params->round_0);
+
+    for (i = 0; i < im_h; ++i) {
+      for (j = 0; j < w; j += 8) {
+        const __m128i data =
+            _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + j]);
+        const __m128i data2 =
+            _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + j + 8]);
+
+        // Filter even-index pixels
+        const __m128i res_0 = _mm_madd_epi16(data, coeff_01);
+        const __m128i res_2 =
+            _mm_madd_epi16(_mm_alignr_epi8(data2, data, 4), coeff_23);
+        const __m128i res_4 =
+            _mm_madd_epi16(_mm_alignr_epi8(data2, data, 8), coeff_45);
+        const __m128i res_6 =
+            _mm_madd_epi16(_mm_alignr_epi8(data2, data, 12), coeff_67);
+
+        __m128i res_even = _mm_add_epi32(_mm_add_epi32(res_0, res_4),
+                                         _mm_add_epi32(res_2, res_6));
+        res_even =
+            _mm_sra_epi32(_mm_add_epi32(res_even, round_const), round_shift);
+
+        // Filter odd-index pixels
+        const __m128i res_1 =
+            _mm_madd_epi16(_mm_alignr_epi8(data2, data, 2), coeff_01);
+        const __m128i res_3 =
+            _mm_madd_epi16(_mm_alignr_epi8(data2, data, 6), coeff_23);
+        const __m128i res_5 =
+            _mm_madd_epi16(_mm_alignr_epi8(data2, data, 10), coeff_45);
+        const __m128i res_7 =
+            _mm_madd_epi16(_mm_alignr_epi8(data2, data, 14), coeff_67);
+
+        __m128i res_odd = _mm_add_epi32(_mm_add_epi32(res_1, res_5),
+                                        _mm_add_epi32(res_3, res_7));
+        res_odd =
+            _mm_sra_epi32(_mm_add_epi32(res_odd, round_const), round_shift);
+
+        // Pack in the column order 0, 2, 4, 6, 1, 3, 5, 7
+        __m128i res = _mm_packs_epi32(res_even, res_odd);
+        _mm_storeu_si128((__m128i *)&im_block[i * im_stride + j], res);
+      }
+    }
+  }
+
+  /* Vertical filter */
+  {
+    const int16_t *y_filter = av1_get_interp_filter_subpel_kernel(
+        filter_params_y, subpel_y_q4 & SUBPEL_MASK);
+    const __m128i coeffs_y = _mm_loadu_si128((__m128i *)y_filter);
+
+    // coeffs 0 1 0 1 2 3 2 3
+    const __m128i tmp_0 = _mm_unpacklo_epi32(coeffs_y, coeffs_y);
+    // coeffs 4 5 4 5 6 7 6 7
+    const __m128i tmp_1 = _mm_unpackhi_epi32(coeffs_y, coeffs_y);
+
+    // coeffs 0 1 0 1 0 1 0 1
+    const __m128i coeff_01 = _mm_unpacklo_epi64(tmp_0, tmp_0);
+    // coeffs 2 3 2 3 2 3 2 3
+    const __m128i coeff_23 = _mm_unpackhi_epi64(tmp_0, tmp_0);
+    // coeffs 4 5 4 5 4 5 4 5
+    const __m128i coeff_45 = _mm_unpacklo_epi64(tmp_1, tmp_1);
+    // coeffs 6 7 6 7 6 7 6 7
+    const __m128i coeff_67 = _mm_unpackhi_epi64(tmp_1, tmp_1);
+
+    const __m128i round_const = _mm_set1_epi32(
+        ((1 << conv_params->round_1) >> 1) -
+        (1 << (bd + 2 * FILTER_BITS - conv_params->round_0 - 1)));
+    const __m128i round_shift = _mm_cvtsi32_si128(conv_params->round_1);
+
+    for (i = 0; i < h; ++i) {
+      for (j = 0; j < w; j += 8) {
+        // Filter even-index pixels
+        const int16_t *data = &im_block[i * im_stride + j];
+        const __m128i src_0 =
+            _mm_unpacklo_epi16(*(__m128i *)(data + 0 * im_stride),
+                               *(__m128i *)(data + 1 * im_stride));
+        const __m128i src_2 =
+            _mm_unpacklo_epi16(*(__m128i *)(data + 2 * im_stride),
+                               *(__m128i *)(data + 3 * im_stride));
+        const __m128i src_4 =
+            _mm_unpacklo_epi16(*(__m128i *)(data + 4 * im_stride),
+                               *(__m128i *)(data + 5 * im_stride));
+        const __m128i src_6 =
+            _mm_unpacklo_epi16(*(__m128i *)(data + 6 * im_stride),
+                               *(__m128i *)(data + 7 * im_stride));
+
+        const __m128i res_0 = _mm_madd_epi16(src_0, coeff_01);
+        const __m128i res_2 = _mm_madd_epi16(src_2, coeff_23);
+        const __m128i res_4 = _mm_madd_epi16(src_4, coeff_45);
+        const __m128i res_6 = _mm_madd_epi16(src_6, coeff_67);
+
+        const __m128i res_even = _mm_add_epi32(_mm_add_epi32(res_0, res_2),
+                                               _mm_add_epi32(res_4, res_6));
+
+        // Filter odd-index pixels
+        const __m128i src_1 =
+            _mm_unpackhi_epi16(*(__m128i *)(data + 0 * im_stride),
+                               *(__m128i *)(data + 1 * im_stride));
+        const __m128i src_3 =
+            _mm_unpackhi_epi16(*(__m128i *)(data + 2 * im_stride),
+                               *(__m128i *)(data + 3 * im_stride));
+        const __m128i src_5 =
+            _mm_unpackhi_epi16(*(__m128i *)(data + 4 * im_stride),
+                               *(__m128i *)(data + 5 * im_stride));
+        const __m128i src_7 =
+            _mm_unpackhi_epi16(*(__m128i *)(data + 6 * im_stride),
+                               *(__m128i *)(data + 7 * im_stride));
+
+        const __m128i res_1 = _mm_madd_epi16(src_1, coeff_01);
+        const __m128i res_3 = _mm_madd_epi16(src_3, coeff_23);
+        const __m128i res_5 = _mm_madd_epi16(src_5, coeff_45);
+        const __m128i res_7 = _mm_madd_epi16(src_7, coeff_67);
+
+        const __m128i res_odd = _mm_add_epi32(_mm_add_epi32(res_1, res_3),
+                                              _mm_add_epi32(res_5, res_7));
+
+        // Rearrange pixels back into the order 0 ... 7
+        const __m128i res_lo = _mm_unpacklo_epi32(res_even, res_odd);
+        const __m128i res_hi = _mm_unpackhi_epi32(res_even, res_odd);
+
+        const __m128i res_lo_round =
+            _mm_sra_epi32(_mm_add_epi32(res_lo, round_const), round_shift);
+
+        const __m128i res_unsigned_lo =
+            _mm_add_epi32(res_lo_round, offset_const);
+
+        if (w < 8) {
+          if (do_average) {
+            const __m128i data_0 =
+                _mm_loadl_epi64((__m128i *)(&dst[i * dst_stride + j]));
+
+            const __m128i data_ref_0 = _mm_cvtepu16_epi32(data_0);
+
+            const __m128i comp_avg_res = highbd_comp_avg_sse4_1(
+                &data_ref_0, &res_unsigned_lo, &wt0, &wt1, use_jnt_comp_avg);
+
+            const __m128i round_result = highbd_convolve_rounding_sse2(
+                &comp_avg_res, &offset_const, &rounding_const, rounding_shift);
+
+            const __m128i res_16b =
+                _mm_packus_epi32(round_result, round_result);
+            const __m128i res_clip = _mm_min_epi16(res_16b, clip_pixel_to_bd);
+
+            _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_clip);
+          } else {
+            const __m128i res_16b =
+                _mm_packus_epi32(res_unsigned_lo, res_unsigned_lo);
+            _mm_storel_epi64((__m128i *)(&dst[i * dst_stride + j]), res_16b);
+          }
+        } else {
+          const __m128i res_hi_round =
+              _mm_sra_epi32(_mm_add_epi32(res_hi, round_const), round_shift);
+
+          const __m128i res_unsigned_hi =
+              _mm_add_epi32(res_hi_round, offset_const);
+
+          if (do_average) {
+            const __m128i data_lo =
+                _mm_loadl_epi64((__m128i *)(&dst[i * dst_stride + j]));
+            const __m128i data_hi =
+                _mm_loadl_epi64((__m128i *)(&dst[i * dst_stride + j + 4]));
+
+            const __m128i data_ref_0_lo = _mm_cvtepu16_epi32(data_lo);
+            const __m128i data_ref_0_hi = _mm_cvtepu16_epi32(data_hi);
+
+            const __m128i comp_avg_res_lo = highbd_comp_avg_sse4_1(
+                &data_ref_0_lo, &res_unsigned_lo, &wt0, &wt1, use_jnt_comp_avg);
+            const __m128i comp_avg_res_hi = highbd_comp_avg_sse4_1(
+                &data_ref_0_hi, &res_unsigned_hi, &wt0, &wt1, use_jnt_comp_avg);
+
+            const __m128i round_result_lo =
+                highbd_convolve_rounding_sse2(&comp_avg_res_lo, &offset_const,
+                                              &rounding_const, rounding_shift);
+            const __m128i round_result_hi =
+                highbd_convolve_rounding_sse2(&comp_avg_res_hi, &offset_const,
+                                              &rounding_const, rounding_shift);
+
+            const __m128i res_16b =
+                _mm_packus_epi32(round_result_lo, round_result_hi);
+            const __m128i res_clip = _mm_min_epi16(res_16b, clip_pixel_to_bd);
+
+            _mm_store_si128((__m128i *)(&dst0[i * dst_stride0 + j]), res_clip);
+          } else {
+            const __m128i res_16b =
+                _mm_packus_epi32(res_unsigned_lo, res_unsigned_hi);
+            _mm_store_si128((__m128i *)(&dst[i * dst_stride + j]), res_16b);
+          }
+        }
+      }
+    }
+  }
+}
diff --git a/third_party/aom/av1/common/x86/highbd_convolve_2d_ssse3.c b/third_party/aom/av1/common/x86/highbd_convolve_2d_ssse3.c
new file mode 100644
index 000000000..1d029db39
--- /dev/null
+++ b/third_party/aom/av1/common/x86/highbd_convolve_2d_ssse3.c
@@ -0,0 +1,217 @@
+/*
+ * 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 <assert.h>
+
+#include "config/aom_dsp_rtcd.h"
+
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/aom_filter.h"
+#include "aom_dsp/x86/convolve_sse2.h"
+#include "av1/common/convolve.h"
+
+void av1_highbd_convolve_2d_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 InterpFilterParams *filter_params_y, const int subpel_x_q4,
+    const int subpel_y_q4, ConvolveParams *conv_params, int bd) {
+  DECLARE_ALIGNED(32, int16_t, im_block[(MAX_SB_SIZE + MAX_FILTER_TAP) * 8]);
+  int im_h = h + filter_params_y->taps - 1;
+  int im_stride = 8;
+  int i, j;
+  const int fo_vert = filter_params_y->taps / 2 - 1;
+  const int fo_horiz = filter_params_x->taps / 2 - 1;
+  const uint16_t *const src_ptr = src - fo_vert * src_stride - 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);
+  __m128i coeffs_x[4], coeffs_y[4], s[16];
+
+  const __m128i round_const_x = _mm_set1_epi32(
+      ((1 << conv_params->round_0) >> 1) + (1 << (bd + FILTER_BITS - 1)));
+  const __m128i round_shift_x = _mm_cvtsi32_si128(conv_params->round_0);
+
+  const __m128i round_const_y =
+      _mm_set1_epi32(((1 << conv_params->round_1) >> 1) -
+                     (1 << (bd + 2 * FILTER_BITS - conv_params->round_0 - 1)));
+  const __m128i round_shift_y = _mm_cvtsi32_si128(conv_params->round_1);
+
+  const int bits =
+      FILTER_BITS * 2 - conv_params->round_0 - conv_params->round_1;
+  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();
+
+  prepare_coeffs(filter_params_x, subpel_x_q4, coeffs_x);
+  prepare_coeffs(filter_params_y, subpel_y_q4, coeffs_y);
+
+  for (j = 0; j < w; j += 8) {
+    /* Horizontal filter */
+    {
+      for (i = 0; i < im_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);
+
+        __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);
+
+        _mm_store_si128((__m128i *)&im_block[i * im_stride], res);
+      }
+    }
+    /* Vertical filter */
+    {
+      __m128i s0 = _mm_loadu_si128((__m128i *)(im_block + 0 * im_stride));
+      __m128i s1 = _mm_loadu_si128((__m128i *)(im_block + 1 * im_stride));
+      __m128i s2 = _mm_loadu_si128((__m128i *)(im_block + 2 * im_stride));
+      __m128i s3 = _mm_loadu_si128((__m128i *)(im_block + 3 * im_stride));
+      __m128i s4 = _mm_loadu_si128((__m128i *)(im_block + 4 * im_stride));
+      __m128i s5 = _mm_loadu_si128((__m128i *)(im_block + 5 * im_stride));
+      __m128i s6 = _mm_loadu_si128((__m128i *)(im_block + 6 * im_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) {
+        const int16_t *data = &im_block[i * im_stride];
+
+        __m128i s7 = _mm_loadu_si128((__m128i *)(data + 7 * im_stride));
+        __m128i s8 = _mm_loadu_si128((__m128i *)(data + 8 * im_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_y), round_shift_y);
+        res_a_round0 = _mm_sra_epi32(
+            _mm_add_epi32(res_a_round0, 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_y), round_shift_y);
+        res_a_round1 = _mm_sra_epi32(
+            _mm_add_epi32(res_a_round1, 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_y), round_shift_y);
+          res_b_round0 = _mm_sra_epi32(
+              _mm_add_epi32(res_b_round0, 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_y), round_shift_y);
+          res_b_round1 = _mm_sra_epi32(
+              _mm_add_epi32(res_b_round1, 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);
+
+          *((uint32_t *)(&dst[i * dst_stride + j])) =
+              _mm_cvtsi128_si32(res_a_round0);
+
+          *((uint32_t *)(&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;
+      }
+    }
+  }
+}
diff --git a/third_party/aom/av1/common/x86/highbd_inv_txfm_avx2.c b/third_party/aom/av1/common/x86/highbd_inv_txfm_avx2.c
new file mode 100644
index 000000000..ade2af03e
--- /dev/null
+++ b/third_party/aom/av1/common/x86/highbd_inv_txfm_avx2.c
@@ -0,0 +1,1349 @@
+/*
+ * 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/av1_rtcd.h"
+
+#include "av1/common/av1_inv_txfm1d_cfg.h"
+#include "av1/common/idct.h"
+#include "av1/common/x86/av1_inv_txfm_ssse3.h"
+#include "av1/common/x86/highbd_txfm_utility_sse4.h"
+
+// Note:
+//  Total 32x4 registers to represent 32x32 block coefficients.
+//  For high bit depth, each coefficient is 4-byte.
+//  Each __m256i register holds 8 coefficients.
+//  So each "row" we needs 4 register. Totally 32 rows
+//  Register layout:
+//   v0,   v1,   v2,   v3,
+//   v4,   v5,   v6,   v7,
+//   ... ...
+//   v124, v125, v126, v127
+
+static INLINE __m256i highbd_clamp_epi16_avx2(__m256i u, int bd) {
+  const __m256i zero = _mm256_setzero_si256();
+  const __m256i one = _mm256_set1_epi16(1);
+  const __m256i max = _mm256_sub_epi16(_mm256_slli_epi16(one, bd), one);
+  __m256i clamped, mask;
+
+  mask = _mm256_cmpgt_epi16(u, max);
+  clamped = _mm256_andnot_si256(mask, u);
+  mask = _mm256_and_si256(mask, max);
+  clamped = _mm256_or_si256(mask, clamped);
+  mask = _mm256_cmpgt_epi16(clamped, zero);
+  clamped = _mm256_and_si256(clamped, mask);
+
+  return clamped;
+}
+
+static INLINE __m256i highbd_get_recon_16x8_avx2(const __m256i pred,
+                                                 __m256i res0, __m256i res1,
+                                                 const int bd) {
+  __m256i x0 = _mm256_cvtepi16_epi32(_mm256_castsi256_si128(pred));
+  __m256i x1 = _mm256_cvtepi16_epi32(_mm256_extractf128_si256(pred, 1));
+
+  x0 = _mm256_add_epi32(res0, x0);
+  x1 = _mm256_add_epi32(res1, x1);
+  x0 = _mm256_packus_epi32(x0, x1);
+  x0 = _mm256_permute4x64_epi64(x0, 0xd8);
+  x0 = highbd_clamp_epi16_avx2(x0, bd);
+  return x0;
+}
+
+static INLINE void highbd_write_buffer_16xn_avx2(__m256i *in, uint16_t *output,
+                                                 int stride, int flipud,
+                                                 int height, const int bd) {
+  int j = flipud ? (height - 1) : 0;
+  const int step = flipud ? -1 : 1;
+  for (int i = 0; i < height; ++i, j += step) {
+    __m256i v = _mm256_loadu_si256((__m256i const *)(output + i * stride));
+    __m256i u = highbd_get_recon_16x8_avx2(v, in[j], in[j + height], bd);
+
+    _mm256_storeu_si256((__m256i *)(output + i * stride), u);
+  }
+}
+
+static INLINE __m256i av1_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 av1_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] = av1_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 void transpose_8x8_avx2(const __m256i *in, __m256i *out) {
+  __m256i u0, u1, u2, u3, u4, u5, u6, u7;
+  __m256i x0, x1;
+
+  u0 = _mm256_unpacklo_epi32(in[0], in[1]);
+  u1 = _mm256_unpackhi_epi32(in[0], in[1]);
+
+  u2 = _mm256_unpacklo_epi32(in[2], in[3]);
+  u3 = _mm256_unpackhi_epi32(in[2], in[3]);
+
+  u4 = _mm256_unpacklo_epi32(in[4], in[5]);
+  u5 = _mm256_unpackhi_epi32(in[4], in[5]);
+
+  u6 = _mm256_unpacklo_epi32(in[6], in[7]);
+  u7 = _mm256_unpackhi_epi32(in[6], in[7]);
+
+  x0 = _mm256_unpacklo_epi64(u0, u2);
+  x1 = _mm256_unpacklo_epi64(u4, u6);
+  out[0] = _mm256_permute2f128_si256(x0, x1, 0x20);
+  out[4] = _mm256_permute2f128_si256(x0, x1, 0x31);
+
+  x0 = _mm256_unpackhi_epi64(u0, u2);
+  x1 = _mm256_unpackhi_epi64(u4, u6);
+  out[1] = _mm256_permute2f128_si256(x0, x1, 0x20);
+  out[5] = _mm256_permute2f128_si256(x0, x1, 0x31);
+
+  x0 = _mm256_unpacklo_epi64(u1, u3);
+  x1 = _mm256_unpacklo_epi64(u5, u7);
+  out[2] = _mm256_permute2f128_si256(x0, x1, 0x20);
+  out[6] = _mm256_permute2f128_si256(x0, x1, 0x31);
+
+  x0 = _mm256_unpackhi_epi64(u1, u3);
+  x1 = _mm256_unpackhi_epi64(u5, u7);
+  out[3] = _mm256_permute2f128_si256(x0, x1, 0x20);
+  out[7] = _mm256_permute2f128_si256(x0, x1, 0x31);
+}
+
+static void load_buffer_32x32(const int32_t *coeff, __m256i *in,
+                              int input_stiride, int size) {
+  int i;
+  for (i = 0; i < size; ++i) {
+    in[i] = _mm256_loadu_si256((const __m256i *)(coeff + i * input_stiride));
+  }
+}
+
+static INLINE __m256i half_btf_0_avx2(const __m256i *w0, const __m256i *n0,
+                                      const __m256i *rounding, int bit) {
+  __m256i x;
+  x = _mm256_mullo_epi32(*w0, *n0);
+  x = _mm256_add_epi32(x, *rounding);
+  x = _mm256_srai_epi32(x, bit);
+  return x;
+}
+
+static INLINE __m256i half_btf_avx2(const __m256i *w0, const __m256i *n0,
+                                    const __m256i *w1, const __m256i *n1,
+                                    const __m256i *rounding, int bit) {
+  __m256i x, y;
+
+  x = _mm256_mullo_epi32(*w0, *n0);
+  y = _mm256_mullo_epi32(*w1, *n1);
+  x = _mm256_add_epi32(x, y);
+  x = _mm256_add_epi32(x, *rounding);
+  x = _mm256_srai_epi32(x, bit);
+  return x;
+}
+
+static void addsub_avx2(const __m256i in0, const __m256i in1, __m256i *out0,
+                        __m256i *out1, const __m256i *clamp_lo,
+                        const __m256i *clamp_hi) {
+  __m256i a0 = _mm256_add_epi32(in0, in1);
+  __m256i a1 = _mm256_sub_epi32(in0, in1);
+
+  a0 = _mm256_max_epi32(a0, *clamp_lo);
+  a0 = _mm256_min_epi32(a0, *clamp_hi);
+  a1 = _mm256_max_epi32(a1, *clamp_lo);
+  a1 = _mm256_min_epi32(a1, *clamp_hi);
+
+  *out0 = a0;
+  *out1 = a1;
+}
+
+static void addsub_no_clamp_avx2(const __m256i in0, const __m256i in1,
+                                 __m256i *out0, __m256i *out1) {
+  __m256i a0 = _mm256_add_epi32(in0, in1);
+  __m256i a1 = _mm256_sub_epi32(in0, in1);
+
+  *out0 = a0;
+  *out1 = a1;
+}
+
+static void addsub_shift_avx2(const __m256i in0, const __m256i in1,
+                              __m256i *out0, __m256i *out1,
+                              const __m256i *clamp_lo, const __m256i *clamp_hi,
+                              int shift) {
+  __m256i offset = _mm256_set1_epi32((1 << shift) >> 1);
+  __m256i in0_w_offset = _mm256_add_epi32(in0, offset);
+  __m256i a0 = _mm256_add_epi32(in0_w_offset, in1);
+  __m256i a1 = _mm256_sub_epi32(in0_w_offset, in1);
+
+  a0 = _mm256_sra_epi32(a0, _mm_cvtsi32_si128(shift));
+  a1 = _mm256_sra_epi32(a1, _mm_cvtsi32_si128(shift));
+
+  a0 = _mm256_max_epi32(a0, *clamp_lo);
+  a0 = _mm256_min_epi32(a0, *clamp_hi);
+  a1 = _mm256_max_epi32(a1, *clamp_lo);
+  a1 = _mm256_min_epi32(a1, *clamp_hi);
+
+  *out0 = a0;
+  *out1 = a1;
+}
+
+static INLINE void idct32_stage4_avx2(
+    __m256i *bf1, const __m256i *cospim8, const __m256i *cospi56,
+    const __m256i *cospi8, const __m256i *cospim56, const __m256i *cospim40,
+    const __m256i *cospi24, const __m256i *cospi40, const __m256i *cospim24,
+    const __m256i *rounding, int bit) {
+  __m256i temp1, temp2;
+  temp1 = half_btf_avx2(cospim8, &bf1[17], cospi56, &bf1[30], rounding, bit);
+  bf1[30] = half_btf_avx2(cospi56, &bf1[17], cospi8, &bf1[30], rounding, bit);
+  bf1[17] = temp1;
+
+  temp2 = half_btf_avx2(cospim56, &bf1[18], cospim8, &bf1[29], rounding, bit);
+  bf1[29] = half_btf_avx2(cospim8, &bf1[18], cospi56, &bf1[29], rounding, bit);
+  bf1[18] = temp2;
+
+  temp1 = half_btf_avx2(cospim40, &bf1[21], cospi24, &bf1[26], rounding, bit);
+  bf1[26] = half_btf_avx2(cospi24, &bf1[21], cospi40, &bf1[26], rounding, bit);
+  bf1[21] = temp1;
+
+  temp2 = half_btf_avx2(cospim24, &bf1[22], cospim40, &bf1[25], rounding, bit);
+  bf1[25] = half_btf_avx2(cospim40, &bf1[22], cospi24, &bf1[25], rounding, bit);
+  bf1[22] = temp2;
+}
+
+static INLINE void idct32_stage5_avx2(
+    __m256i *bf1, const __m256i *cospim16, const __m256i *cospi48,
+    const __m256i *cospi16, const __m256i *cospim48, const __m256i *clamp_lo,
+    const __m256i *clamp_hi, const __m256i *rounding, int bit) {
+  __m256i temp1, temp2;
+  temp1 = half_btf_avx2(cospim16, &bf1[9], cospi48, &bf1[14], rounding, bit);
+  bf1[14] = half_btf_avx2(cospi48, &bf1[9], cospi16, &bf1[14], rounding, bit);
+  bf1[9] = temp1;
+
+  temp2 = half_btf_avx2(cospim48, &bf1[10], cospim16, &bf1[13], rounding, bit);
+  bf1[13] = half_btf_avx2(cospim16, &bf1[10], cospi48, &bf1[13], rounding, bit);
+  bf1[10] = temp2;
+
+  addsub_avx2(bf1[16], bf1[19], bf1 + 16, bf1 + 19, clamp_lo, clamp_hi);
+  addsub_avx2(bf1[17], bf1[18], bf1 + 17, bf1 + 18, clamp_lo, clamp_hi);
+  addsub_avx2(bf1[23], bf1[20], bf1 + 23, bf1 + 20, clamp_lo, clamp_hi);
+  addsub_avx2(bf1[22], bf1[21], bf1 + 22, bf1 + 21, clamp_lo, clamp_hi);
+  addsub_avx2(bf1[24], bf1[27], bf1 + 24, bf1 + 27, clamp_lo, clamp_hi);
+  addsub_avx2(bf1[25], bf1[26], bf1 + 25, bf1 + 26, clamp_lo, clamp_hi);
+  addsub_avx2(bf1[31], bf1[28], bf1 + 31, bf1 + 28, clamp_lo, clamp_hi);
+  addsub_avx2(bf1[30], bf1[29], bf1 + 30, bf1 + 29, clamp_lo, clamp_hi);
+}
+
+static INLINE void idct32_stage6_avx2(
+    __m256i *bf1, const __m256i *cospim32, const __m256i *cospi32,
+    const __m256i *cospim16, const __m256i *cospi48, const __m256i *cospi16,
+    const __m256i *cospim48, const __m256i *clamp_lo, const __m256i *clamp_hi,
+    const __m256i *rounding, int bit) {
+  __m256i temp1, temp2;
+  temp1 = half_btf_avx2(cospim32, &bf1[5], cospi32, &bf1[6], rounding, bit);
+  bf1[6] = half_btf_avx2(cospi32, &bf1[5], cospi32, &bf1[6], rounding, bit);
+  bf1[5] = temp1;
+
+  addsub_avx2(bf1[8], bf1[11], bf1 + 8, bf1 + 11, clamp_lo, clamp_hi);
+  addsub_avx2(bf1[9], bf1[10], bf1 + 9, bf1 + 10, clamp_lo, clamp_hi);
+  addsub_avx2(bf1[15], bf1[12], bf1 + 15, bf1 + 12, clamp_lo, clamp_hi);
+  addsub_avx2(bf1[14], bf1[13], bf1 + 14, bf1 + 13, clamp_lo, clamp_hi);
+
+  temp1 = half_btf_avx2(cospim16, &bf1[18], cospi48, &bf1[29], rounding, bit);
+  bf1[29] = half_btf_avx2(cospi48, &bf1[18], cospi16, &bf1[29], rounding, bit);
+  bf1[18] = temp1;
+  temp2 = half_btf_avx2(cospim16, &bf1[19], cospi48, &bf1[28], rounding, bit);
+  bf1[28] = half_btf_avx2(cospi48, &bf1[19], cospi16, &bf1[28], rounding, bit);
+  bf1[19] = temp2;
+  temp1 = half_btf_avx2(cospim48, &bf1[20], cospim16, &bf1[27], rounding, bit);
+  bf1[27] = half_btf_avx2(cospim16, &bf1[20], cospi48, &bf1[27], rounding, bit);
+  bf1[20] = temp1;
+  temp2 = half_btf_avx2(cospim48, &bf1[21], cospim16, &bf1[26], rounding, bit);
+  bf1[26] = half_btf_avx2(cospim16, &bf1[21], cospi48, &bf1[26], rounding, bit);
+  bf1[21] = temp2;
+}
+
+static INLINE void idct32_stage7_avx2(__m256i *bf1, const __m256i *cospim32,
+                                      const __m256i *cospi32,
+                                      const __m256i *clamp_lo,
+                                      const __m256i *clamp_hi,
+                                      const __m256i *rounding, int bit) {
+  __m256i temp1, temp2;
+  addsub_avx2(bf1[0], bf1[7], bf1 + 0, bf1 + 7, clamp_lo, clamp_hi);
+  addsub_avx2(bf1[1], bf1[6], bf1 + 1, bf1 + 6, clamp_lo, clamp_hi);
+  addsub_avx2(bf1[2], bf1[5], bf1 + 2, bf1 + 5, clamp_lo, clamp_hi);
+  addsub_avx2(bf1[3], bf1[4], bf1 + 3, bf1 + 4, clamp_lo, clamp_hi);
+
+  temp1 = half_btf_avx2(cospim32, &bf1[10], cospi32, &bf1[13], rounding, bit);
+  bf1[13] = half_btf_avx2(cospi32, &bf1[10], cospi32, &bf1[13], rounding, bit);
+  bf1[10] = temp1;
+  temp2 = half_btf_avx2(cospim32, &bf1[11], cospi32, &bf1[12], rounding, bit);
+  bf1[12] = half_btf_avx2(cospi32, &bf1[11], cospi32, &bf1[12], rounding, bit);
+  bf1[11] = temp2;
+
+  addsub_avx2(bf1[16], bf1[23], bf1 + 16, bf1 + 23, clamp_lo, clamp_hi);
+  addsub_avx2(bf1[17], bf1[22], bf1 + 17, bf1 + 22, clamp_lo, clamp_hi);
+  addsub_avx2(bf1[18], bf1[21], bf1 + 18, bf1 + 21, clamp_lo, clamp_hi);
+  addsub_avx2(bf1[19], bf1[20], bf1 + 19, bf1 + 20, clamp_lo, clamp_hi);
+  addsub_avx2(bf1[31], bf1[24], bf1 + 31, bf1 + 24, clamp_lo, clamp_hi);
+  addsub_avx2(bf1[30], bf1[25], bf1 + 30, bf1 + 25, clamp_lo, clamp_hi);
+  addsub_avx2(bf1[29], bf1[26], bf1 + 29, bf1 + 26, clamp_lo, clamp_hi);
+  addsub_avx2(bf1[28], bf1[27], bf1 + 28, bf1 + 27, clamp_lo, clamp_hi);
+}
+
+static INLINE void idct32_stage8_avx2(__m256i *bf1, const __m256i *cospim32,
+                                      const __m256i *cospi32,
+                                      const __m256i *clamp_lo,
+                                      const __m256i *clamp_hi,
+                                      const __m256i *rounding, int bit) {
+  __m256i temp1, temp2;
+  addsub_avx2(bf1[0], bf1[15], bf1 + 0, bf1 + 15, clamp_lo, clamp_hi);
+  addsub_avx2(bf1[1], bf1[14], bf1 + 1, bf1 + 14, clamp_lo, clamp_hi);
+  addsub_avx2(bf1[2], bf1[13], bf1 + 2, bf1 + 13, clamp_lo, clamp_hi);
+  addsub_avx2(bf1[3], bf1[12], bf1 + 3, bf1 + 12, clamp_lo, clamp_hi);
+  addsub_avx2(bf1[4], bf1[11], bf1 + 4, bf1 + 11, clamp_lo, clamp_hi);
+  addsub_avx2(bf1[5], bf1[10], bf1 + 5, bf1 + 10, clamp_lo, clamp_hi);
+  addsub_avx2(bf1[6], bf1[9], bf1 + 6, bf1 + 9, clamp_lo, clamp_hi);
+  addsub_avx2(bf1[7], bf1[8], bf1 + 7, bf1 + 8, clamp_lo, clamp_hi);
+
+  temp1 = half_btf_avx2(cospim32, &bf1[20], cospi32, &bf1[27], rounding, bit);
+  bf1[27] = half_btf_avx2(cospi32, &bf1[20], cospi32, &bf1[27], rounding, bit);
+  bf1[20] = temp1;
+  temp2 = half_btf_avx2(cospim32, &bf1[21], cospi32, &bf1[26], rounding, bit);
+  bf1[26] = half_btf_avx2(cospi32, &bf1[21], cospi32, &bf1[26], rounding, bit);
+  bf1[21] = temp2;
+  temp1 = half_btf_avx2(cospim32, &bf1[22], cospi32, &bf1[25], rounding, bit);
+  bf1[25] = half_btf_avx2(cospi32, &bf1[22], cospi32, &bf1[25], rounding, bit);
+  bf1[22] = temp1;
+  temp2 = half_btf_avx2(cospim32, &bf1[23], cospi32, &bf1[24], rounding, bit);
+  bf1[24] = half_btf_avx2(cospi32, &bf1[23], cospi32, &bf1[24], rounding, bit);
+  bf1[23] = temp2;
+}
+
+static INLINE void idct32_stage9_avx2(__m256i *bf1, __m256i *out,
+                                      const int do_cols, const int bd,
+                                      const int out_shift,
+                                      const int log_range) {
+  if (do_cols) {
+    addsub_no_clamp_avx2(bf1[0], bf1[31], out + 0, out + 31);
+    addsub_no_clamp_avx2(bf1[1], bf1[30], out + 1, out + 30);
+    addsub_no_clamp_avx2(bf1[2], bf1[29], out + 2, out + 29);
+    addsub_no_clamp_avx2(bf1[3], bf1[28], out + 3, out + 28);
+    addsub_no_clamp_avx2(bf1[4], bf1[27], out + 4, out + 27);
+    addsub_no_clamp_avx2(bf1[5], bf1[26], out + 5, out + 26);
+    addsub_no_clamp_avx2(bf1[6], bf1[25], out + 6, out + 25);
+    addsub_no_clamp_avx2(bf1[7], bf1[24], out + 7, out + 24);
+    addsub_no_clamp_avx2(bf1[8], bf1[23], out + 8, out + 23);
+    addsub_no_clamp_avx2(bf1[9], bf1[22], out + 9, out + 22);
+    addsub_no_clamp_avx2(bf1[10], bf1[21], out + 10, out + 21);
+    addsub_no_clamp_avx2(bf1[11], bf1[20], out + 11, out + 20);
+    addsub_no_clamp_avx2(bf1[12], bf1[19], out + 12, out + 19);
+    addsub_no_clamp_avx2(bf1[13], bf1[18], out + 13, out + 18);
+    addsub_no_clamp_avx2(bf1[14], bf1[17], out + 14, out + 17);
+    addsub_no_clamp_avx2(bf1[15], bf1[16], out + 15, out + 16);
+  } else {
+    const int log_range_out = AOMMAX(16, bd + 6);
+    const __m256i clamp_lo_out = _mm256_set1_epi32(AOMMAX(
+        -(1 << (log_range_out - 1)), -(1 << (log_range - 1 - out_shift))));
+    const __m256i clamp_hi_out = _mm256_set1_epi32(AOMMIN(
+        (1 << (log_range_out - 1)) - 1, (1 << (log_range - 1 - out_shift))));
+
+    addsub_shift_avx2(bf1[0], bf1[31], out + 0, out + 31, &clamp_lo_out,
+                      &clamp_hi_out, out_shift);
+    addsub_shift_avx2(bf1[1], bf1[30], out + 1, out + 30, &clamp_lo_out,
+                      &clamp_hi_out, out_shift);
+    addsub_shift_avx2(bf1[2], bf1[29], out + 2, out + 29, &clamp_lo_out,
+                      &clamp_hi_out, out_shift);
+    addsub_shift_avx2(bf1[3], bf1[28], out + 3, out + 28, &clamp_lo_out,
+                      &clamp_hi_out, out_shift);
+    addsub_shift_avx2(bf1[4], bf1[27], out + 4, out + 27, &clamp_lo_out,
+                      &clamp_hi_out, out_shift);
+    addsub_shift_avx2(bf1[5], bf1[26], out + 5, out + 26, &clamp_lo_out,
+                      &clamp_hi_out, out_shift);
+    addsub_shift_avx2(bf1[6], bf1[25], out + 6, out + 25, &clamp_lo_out,
+                      &clamp_hi_out, out_shift);
+    addsub_shift_avx2(bf1[7], bf1[24], out + 7, out + 24, &clamp_lo_out,
+                      &clamp_hi_out, out_shift);
+    addsub_shift_avx2(bf1[8], bf1[23], out + 8, out + 23, &clamp_lo_out,
+                      &clamp_hi_out, out_shift);
+    addsub_shift_avx2(bf1[9], bf1[22], out + 9, out + 22, &clamp_lo_out,
+                      &clamp_hi_out, out_shift);
+    addsub_shift_avx2(bf1[10], bf1[21], out + 10, out + 21, &clamp_lo_out,
+                      &clamp_hi_out, out_shift);
+    addsub_shift_avx2(bf1[11], bf1[20], out + 11, out + 20, &clamp_lo_out,
+                      &clamp_hi_out, out_shift);
+    addsub_shift_avx2(bf1[12], bf1[19], out + 12, out + 19, &clamp_lo_out,
+                      &clamp_hi_out, out_shift);
+    addsub_shift_avx2(bf1[13], bf1[18], out + 13, out + 18, &clamp_lo_out,
+                      &clamp_hi_out, out_shift);
+    addsub_shift_avx2(bf1[14], bf1[17], out + 14, out + 17, &clamp_lo_out,
+                      &clamp_hi_out, out_shift);
+    addsub_shift_avx2(bf1[15], bf1[16], out + 15, out + 16, &clamp_lo_out,
+                      &clamp_hi_out, out_shift);
+  }
+}
+
+static void idct32_low1_avx2(__m256i *in, __m256i *out, int bit, int do_cols,
+                             int bd, int out_shift) {
+  const int32_t *cospi = cospi_arr(bit);
+  const __m256i cospi32 = _mm256_set1_epi32(cospi[32]);
+  const __m256i rounding = _mm256_set1_epi32(1 << (bit - 1));
+  const int log_range = AOMMAX(16, bd + (do_cols ? 6 : 8));
+  const __m256i clamp_lo = _mm256_set1_epi32(-(1 << (log_range - 1)));
+  const __m256i clamp_hi = _mm256_set1_epi32((1 << (log_range - 1)) - 1);
+  __m256i x;
+  // stage 0
+  // stage 1
+  // stage 2
+  // stage 3
+  // stage 4
+  // stage 5
+  x = _mm256_mullo_epi32(in[0], cospi32);
+  x = _mm256_add_epi32(x, rounding);
+  x = _mm256_srai_epi32(x, bit);
+
+  // stage 6
+  // stage 7
+  // stage 8
+  // stage 9
+  if (do_cols) {
+    x = _mm256_max_epi32(x, clamp_lo);
+    x = _mm256_min_epi32(x, clamp_hi);
+  } else {
+    const int log_range_out = AOMMAX(16, bd + 6);
+    const __m256i clamp_lo_out = _mm256_set1_epi32(AOMMAX(
+        -(1 << (log_range_out - 1)), -(1 << (log_range - 1 - out_shift))));
+    const __m256i clamp_hi_out = _mm256_set1_epi32(AOMMIN(
+        (1 << (log_range_out - 1)) - 1, (1 << (log_range - 1 - out_shift))));
+    __m256i offset = _mm256_set1_epi32((1 << out_shift) >> 1);
+    x = _mm256_add_epi32(offset, x);
+    x = _mm256_sra_epi32(x, _mm_cvtsi32_si128(out_shift));
+    x = _mm256_max_epi32(x, clamp_lo_out);
+    x = _mm256_min_epi32(x, clamp_hi_out);
+  }
+
+  out[0] = x;
+  out[1] = x;
+  out[2] = x;
+  out[3] = x;
+  out[4] = x;
+  out[5] = x;
+  out[6] = x;
+  out[7] = x;
+  out[8] = x;
+  out[9] = x;
+  out[10] = x;
+  out[11] = x;
+  out[12] = x;
+  out[13] = x;
+  out[14] = x;
+  out[15] = x;
+  out[16] = x;
+  out[17] = x;
+  out[18] = x;
+  out[19] = x;
+  out[20] = x;
+  out[21] = x;
+  out[22] = x;
+  out[23] = x;
+  out[24] = x;
+  out[25] = x;
+  out[26] = x;
+  out[27] = x;
+  out[28] = x;
+  out[29] = x;
+  out[30] = x;
+  out[31] = x;
+}
+
+static void idct32_low8_avx2(__m256i *in, __m256i *out, int bit, int do_cols,
+                             int bd, int out_shift) {
+  const int32_t *cospi = cospi_arr(bit);
+  const __m256i cospi62 = _mm256_set1_epi32(cospi[62]);
+  const __m256i cospi14 = _mm256_set1_epi32(cospi[14]);
+  const __m256i cospi54 = _mm256_set1_epi32(cospi[54]);
+  const __m256i cospi6 = _mm256_set1_epi32(cospi[6]);
+  const __m256i cospi10 = _mm256_set1_epi32(cospi[10]);
+  const __m256i cospi2 = _mm256_set1_epi32(cospi[2]);
+  const __m256i cospim58 = _mm256_set1_epi32(-cospi[58]);
+  const __m256i cospim50 = _mm256_set1_epi32(-cospi[50]);
+  const __m256i cospi60 = _mm256_set1_epi32(cospi[60]);
+  const __m256i cospi12 = _mm256_set1_epi32(cospi[12]);
+  const __m256i cospi4 = _mm256_set1_epi32(cospi[4]);
+  const __m256i cospim52 = _mm256_set1_epi32(-cospi[52]);
+  const __m256i cospi56 = _mm256_set1_epi32(cospi[56]);
+  const __m256i cospi24 = _mm256_set1_epi32(cospi[24]);
+  const __m256i cospi40 = _mm256_set1_epi32(cospi[40]);
+  const __m256i cospi8 = _mm256_set1_epi32(cospi[8]);
+  const __m256i cospim40 = _mm256_set1_epi32(-cospi[40]);
+  const __m256i cospim8 = _mm256_set1_epi32(-cospi[8]);
+  const __m256i cospim56 = _mm256_set1_epi32(-cospi[56]);
+  const __m256i cospim24 = _mm256_set1_epi32(-cospi[24]);
+  const __m256i cospi32 = _mm256_set1_epi32(cospi[32]);
+  const __m256i cospim32 = _mm256_set1_epi32(-cospi[32]);
+  const __m256i cospi48 = _mm256_set1_epi32(cospi[48]);
+  const __m256i cospim48 = _mm256_set1_epi32(-cospi[48]);
+  const __m256i cospi16 = _mm256_set1_epi32(cospi[16]);
+  const __m256i cospim16 = _mm256_set1_epi32(-cospi[16]);
+  const __m256i rounding = _mm256_set1_epi32(1 << (bit - 1));
+  const int log_range = AOMMAX(16, bd + (do_cols ? 6 : 8));
+  const __m256i clamp_lo = _mm256_set1_epi32(-(1 << (log_range - 1)));
+  const __m256i clamp_hi = _mm256_set1_epi32((1 << (log_range - 1)) - 1);
+  __m256i bf1[32];
+
+  {
+    // stage 0
+    // stage 1
+    bf1[0] = in[0];
+    bf1[4] = in[4];
+    bf1[8] = in[2];
+    bf1[12] = in[6];
+    bf1[16] = in[1];
+    bf1[20] = in[5];
+    bf1[24] = in[3];
+    bf1[28] = in[7];
+
+    // stage 2
+    bf1[31] = half_btf_0_avx2(&cospi2, &bf1[16], &rounding, bit);
+    bf1[16] = half_btf_0_avx2(&cospi62, &bf1[16], &rounding, bit);
+    bf1[19] = half_btf_0_avx2(&cospim50, &bf1[28], &rounding, bit);
+    bf1[28] = half_btf_0_avx2(&cospi14, &bf1[28], &rounding, bit);
+    bf1[27] = half_btf_0_avx2(&cospi10, &bf1[20], &rounding, bit);
+    bf1[20] = half_btf_0_avx2(&cospi54, &bf1[20], &rounding, bit);
+    bf1[23] = half_btf_0_avx2(&cospim58, &bf1[24], &rounding, bit);
+    bf1[24] = half_btf_0_avx2(&cospi6, &bf1[24], &rounding, bit);
+
+    // stage 3
+    bf1[15] = half_btf_0_avx2(&cospi4, &bf1[8], &rounding, bit);
+    bf1[8] = half_btf_0_avx2(&cospi60, &bf1[8], &rounding, bit);
+
+    bf1[11] = half_btf_0_avx2(&cospim52, &bf1[12], &rounding, bit);
+    bf1[12] = half_btf_0_avx2(&cospi12, &bf1[12], &rounding, bit);
+    bf1[17] = bf1[16];
+    bf1[18] = bf1[19];
+    bf1[21] = bf1[20];
+    bf1[22] = bf1[23];
+    bf1[25] = bf1[24];
+    bf1[26] = bf1[27];
+    bf1[29] = bf1[28];
+    bf1[30] = bf1[31];
+
+    // stage 4
+    bf1[7] = half_btf_0_avx2(&cospi8, &bf1[4], &rounding, bit);
+    bf1[4] = half_btf_0_avx2(&cospi56, &bf1[4], &rounding, bit);
+
+    bf1[9] = bf1[8];
+    bf1[10] = bf1[11];
+    bf1[13] = bf1[12];
+    bf1[14] = bf1[15];
+
+    idct32_stage4_avx2(bf1, &cospim8, &cospi56, &cospi8, &cospim56, &cospim40,
+                       &cospi24, &cospi40, &cospim24, &rounding, bit);
+
+    // stage 5
+    bf1[0] = half_btf_0_avx2(&cospi32, &bf1[0], &rounding, bit);
+    bf1[1] = bf1[0];
+    bf1[5] = bf1[4];
+    bf1[6] = bf1[7];
+
+    idct32_stage5_avx2(bf1, &cospim16, &cospi48, &cospi16, &cospim48, &clamp_lo,
+                       &clamp_hi, &rounding, bit);
+
+    // stage 6
+    bf1[3] = bf1[0];
+    bf1[2] = bf1[1];
+
+    idct32_stage6_avx2(bf1, &cospim32, &cospi32, &cospim16, &cospi48, &cospi16,
+                       &cospim48, &clamp_lo, &clamp_hi, &rounding, bit);
+
+    // stage 7
+    idct32_stage7_avx2(bf1, &cospim32, &cospi32, &clamp_lo, &clamp_hi,
+                       &rounding, bit);
+
+    // stage 8
+    idct32_stage8_avx2(bf1, &cospim32, &cospi32, &clamp_lo, &clamp_hi,
+                       &rounding, bit);
+
+    // stage 9
+    idct32_stage9_avx2(bf1, out, do_cols, bd, out_shift, log_range);
+  }
+}
+
+static void idct32_low16_avx2(__m256i *in, __m256i *out, int bit, int do_cols,
+                              int bd, int out_shift) {
+  const int32_t *cospi = cospi_arr(bit);
+  const __m256i cospi62 = _mm256_set1_epi32(cospi[62]);
+  const __m256i cospi30 = _mm256_set1_epi32(cospi[30]);
+  const __m256i cospi46 = _mm256_set1_epi32(cospi[46]);
+  const __m256i cospi14 = _mm256_set1_epi32(cospi[14]);
+  const __m256i cospi54 = _mm256_set1_epi32(cospi[54]);
+  const __m256i cospi22 = _mm256_set1_epi32(cospi[22]);
+  const __m256i cospi38 = _mm256_set1_epi32(cospi[38]);
+  const __m256i cospi6 = _mm256_set1_epi32(cospi[6]);
+  const __m256i cospi26 = _mm256_set1_epi32(cospi[26]);
+  const __m256i cospi10 = _mm256_set1_epi32(cospi[10]);
+  const __m256i cospi18 = _mm256_set1_epi32(cospi[18]);
+  const __m256i cospi2 = _mm256_set1_epi32(cospi[2]);
+  const __m256i cospim58 = _mm256_set1_epi32(-cospi[58]);
+  const __m256i cospim42 = _mm256_set1_epi32(-cospi[42]);
+  const __m256i cospim50 = _mm256_set1_epi32(-cospi[50]);
+  const __m256i cospim34 = _mm256_set1_epi32(-cospi[34]);
+  const __m256i cospi60 = _mm256_set1_epi32(cospi[60]);
+  const __m256i cospi28 = _mm256_set1_epi32(cospi[28]);
+  const __m256i cospi44 = _mm256_set1_epi32(cospi[44]);
+  const __m256i cospi12 = _mm256_set1_epi32(cospi[12]);
+  const __m256i cospi20 = _mm256_set1_epi32(cospi[20]);
+  const __m256i cospi4 = _mm256_set1_epi32(cospi[4]);
+  const __m256i cospim52 = _mm256_set1_epi32(-cospi[52]);
+  const __m256i cospim36 = _mm256_set1_epi32(-cospi[36]);
+  const __m256i cospi56 = _mm256_set1_epi32(cospi[56]);
+  const __m256i cospi24 = _mm256_set1_epi32(cospi[24]);
+  const __m256i cospi40 = _mm256_set1_epi32(cospi[40]);
+  const __m256i cospi8 = _mm256_set1_epi32(cospi[8]);
+  const __m256i cospim40 = _mm256_set1_epi32(-cospi[40]);
+  const __m256i cospim8 = _mm256_set1_epi32(-cospi[8]);
+  const __m256i cospim56 = _mm256_set1_epi32(-cospi[56]);
+  const __m256i cospim24 = _mm256_set1_epi32(-cospi[24]);
+  const __m256i cospi32 = _mm256_set1_epi32(cospi[32]);
+  const __m256i cospim32 = _mm256_set1_epi32(-cospi[32]);
+  const __m256i cospi48 = _mm256_set1_epi32(cospi[48]);
+  const __m256i cospim48 = _mm256_set1_epi32(-cospi[48]);
+  const __m256i cospi16 = _mm256_set1_epi32(cospi[16]);
+  const __m256i cospim16 = _mm256_set1_epi32(-cospi[16]);
+  const __m256i rounding = _mm256_set1_epi32(1 << (bit - 1));
+  const int log_range = AOMMAX(16, bd + (do_cols ? 6 : 8));
+  const __m256i clamp_lo = _mm256_set1_epi32(-(1 << (log_range - 1)));
+  const __m256i clamp_hi = _mm256_set1_epi32((1 << (log_range - 1)) - 1);
+  __m256i bf1[32];
+
+  {
+    // stage 0
+    // stage 1
+    bf1[0] = in[0];
+    bf1[2] = in[8];
+    bf1[4] = in[4];
+    bf1[6] = in[12];
+    bf1[8] = in[2];
+    bf1[10] = in[10];
+    bf1[12] = in[6];
+    bf1[14] = in[14];
+    bf1[16] = in[1];
+    bf1[18] = in[9];
+    bf1[20] = in[5];
+    bf1[22] = in[13];
+    bf1[24] = in[3];
+    bf1[26] = in[11];
+    bf1[28] = in[7];
+    bf1[30] = in[15];
+
+    // stage 2
+    bf1[31] = half_btf_0_avx2(&cospi2, &bf1[16], &rounding, bit);
+    bf1[16] = half_btf_0_avx2(&cospi62, &bf1[16], &rounding, bit);
+    bf1[17] = half_btf_0_avx2(&cospim34, &bf1[30], &rounding, bit);
+    bf1[30] = half_btf_0_avx2(&cospi30, &bf1[30], &rounding, bit);
+    bf1[29] = half_btf_0_avx2(&cospi18, &bf1[18], &rounding, bit);
+    bf1[18] = half_btf_0_avx2(&cospi46, &bf1[18], &rounding, bit);
+    bf1[19] = half_btf_0_avx2(&cospim50, &bf1[28], &rounding, bit);
+    bf1[28] = half_btf_0_avx2(&cospi14, &bf1[28], &rounding, bit);
+    bf1[27] = half_btf_0_avx2(&cospi10, &bf1[20], &rounding, bit);
+    bf1[20] = half_btf_0_avx2(&cospi54, &bf1[20], &rounding, bit);
+    bf1[21] = half_btf_0_avx2(&cospim42, &bf1[26], &rounding, bit);
+    bf1[26] = half_btf_0_avx2(&cospi22, &bf1[26], &rounding, bit);
+    bf1[25] = half_btf_0_avx2(&cospi26, &bf1[22], &rounding, bit);
+    bf1[22] = half_btf_0_avx2(&cospi38, &bf1[22], &rounding, bit);
+    bf1[23] = half_btf_0_avx2(&cospim58, &bf1[24], &rounding, bit);
+    bf1[24] = half_btf_0_avx2(&cospi6, &bf1[24], &rounding, bit);
+
+    // stage 3
+    bf1[15] = half_btf_0_avx2(&cospi4, &bf1[8], &rounding, bit);
+    bf1[8] = half_btf_0_avx2(&cospi60, &bf1[8], &rounding, bit);
+    bf1[9] = half_btf_0_avx2(&cospim36, &bf1[14], &rounding, bit);
+    bf1[14] = half_btf_0_avx2(&cospi28, &bf1[14], &rounding, bit);
+    bf1[13] = half_btf_0_avx2(&cospi20, &bf1[10], &rounding, bit);
+    bf1[10] = half_btf_0_avx2(&cospi44, &bf1[10], &rounding, bit);
+    bf1[11] = half_btf_0_avx2(&cospim52, &bf1[12], &rounding, bit);
+    bf1[12] = half_btf_0_avx2(&cospi12, &bf1[12], &rounding, bit);
+
+    addsub_avx2(bf1[16], bf1[17], bf1 + 16, bf1 + 17, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf1[19], bf1[18], bf1 + 19, bf1 + 18, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf1[20], bf1[21], bf1 + 20, bf1 + 21, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf1[23], bf1[22], bf1 + 23, bf1 + 22, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf1[24], bf1[25], bf1 + 24, bf1 + 25, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf1[27], bf1[26], bf1 + 27, bf1 + 26, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf1[28], bf1[29], bf1 + 28, bf1 + 29, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf1[31], bf1[30], bf1 + 31, bf1 + 30, &clamp_lo, &clamp_hi);
+
+    // stage 4
+    bf1[7] = half_btf_0_avx2(&cospi8, &bf1[4], &rounding, bit);
+    bf1[4] = half_btf_0_avx2(&cospi56, &bf1[4], &rounding, bit);
+    bf1[5] = half_btf_0_avx2(&cospim40, &bf1[6], &rounding, bit);
+    bf1[6] = half_btf_0_avx2(&cospi24, &bf1[6], &rounding, bit);
+
+    addsub_avx2(bf1[8], bf1[9], bf1 + 8, bf1 + 9, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf1[11], bf1[10], bf1 + 11, bf1 + 10, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf1[12], bf1[13], bf1 + 12, bf1 + 13, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf1[15], bf1[14], bf1 + 15, bf1 + 14, &clamp_lo, &clamp_hi);
+
+    idct32_stage4_avx2(bf1, &cospim8, &cospi56, &cospi8, &cospim56, &cospim40,
+                       &cospi24, &cospi40, &cospim24, &rounding, bit);
+
+    // stage 5
+    bf1[0] = half_btf_0_avx2(&cospi32, &bf1[0], &rounding, bit);
+    bf1[1] = bf1[0];
+    bf1[3] = half_btf_0_avx2(&cospi16, &bf1[2], &rounding, bit);
+    bf1[2] = half_btf_0_avx2(&cospi48, &bf1[2], &rounding, bit);
+
+    addsub_avx2(bf1[4], bf1[5], bf1 + 4, bf1 + 5, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf1[7], bf1[6], bf1 + 7, bf1 + 6, &clamp_lo, &clamp_hi);
+
+    idct32_stage5_avx2(bf1, &cospim16, &cospi48, &cospi16, &cospim48, &clamp_lo,
+                       &clamp_hi, &rounding, bit);
+
+    // stage 6
+    addsub_avx2(bf1[0], bf1[3], bf1 + 0, bf1 + 3, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf1[1], bf1[2], bf1 + 1, bf1 + 2, &clamp_lo, &clamp_hi);
+
+    idct32_stage6_avx2(bf1, &cospim32, &cospi32, &cospim16, &cospi48, &cospi16,
+                       &cospim48, &clamp_lo, &clamp_hi, &rounding, bit);
+
+    // stage 7
+    idct32_stage7_avx2(bf1, &cospim32, &cospi32, &clamp_lo, &clamp_hi,
+                       &rounding, bit);
+
+    // stage 8
+    idct32_stage8_avx2(bf1, &cospim32, &cospi32, &clamp_lo, &clamp_hi,
+                       &rounding, bit);
+
+    // stage 9
+    idct32_stage9_avx2(bf1, out, do_cols, bd, out_shift, log_range);
+  }
+}
+
+static void idct32_avx2(__m256i *in, __m256i *out, int bit, int do_cols, int bd,
+                        int out_shift) {
+  const int32_t *cospi = cospi_arr(bit);
+  const __m256i cospi62 = _mm256_set1_epi32(cospi[62]);
+  const __m256i cospi30 = _mm256_set1_epi32(cospi[30]);
+  const __m256i cospi46 = _mm256_set1_epi32(cospi[46]);
+  const __m256i cospi14 = _mm256_set1_epi32(cospi[14]);
+  const __m256i cospi54 = _mm256_set1_epi32(cospi[54]);
+  const __m256i cospi22 = _mm256_set1_epi32(cospi[22]);
+  const __m256i cospi38 = _mm256_set1_epi32(cospi[38]);
+  const __m256i cospi6 = _mm256_set1_epi32(cospi[6]);
+  const __m256i cospi58 = _mm256_set1_epi32(cospi[58]);
+  const __m256i cospi26 = _mm256_set1_epi32(cospi[26]);
+  const __m256i cospi42 = _mm256_set1_epi32(cospi[42]);
+  const __m256i cospi10 = _mm256_set1_epi32(cospi[10]);
+  const __m256i cospi50 = _mm256_set1_epi32(cospi[50]);
+  const __m256i cospi18 = _mm256_set1_epi32(cospi[18]);
+  const __m256i cospi34 = _mm256_set1_epi32(cospi[34]);
+  const __m256i cospi2 = _mm256_set1_epi32(cospi[2]);
+  const __m256i cospim58 = _mm256_set1_epi32(-cospi[58]);
+  const __m256i cospim26 = _mm256_set1_epi32(-cospi[26]);
+  const __m256i cospim42 = _mm256_set1_epi32(-cospi[42]);
+  const __m256i cospim10 = _mm256_set1_epi32(-cospi[10]);
+  const __m256i cospim50 = _mm256_set1_epi32(-cospi[50]);
+  const __m256i cospim18 = _mm256_set1_epi32(-cospi[18]);
+  const __m256i cospim34 = _mm256_set1_epi32(-cospi[34]);
+  const __m256i cospim2 = _mm256_set1_epi32(-cospi[2]);
+  const __m256i cospi60 = _mm256_set1_epi32(cospi[60]);
+  const __m256i cospi28 = _mm256_set1_epi32(cospi[28]);
+  const __m256i cospi44 = _mm256_set1_epi32(cospi[44]);
+  const __m256i cospi12 = _mm256_set1_epi32(cospi[12]);
+  const __m256i cospi52 = _mm256_set1_epi32(cospi[52]);
+  const __m256i cospi20 = _mm256_set1_epi32(cospi[20]);
+  const __m256i cospi36 = _mm256_set1_epi32(cospi[36]);
+  const __m256i cospi4 = _mm256_set1_epi32(cospi[4]);
+  const __m256i cospim52 = _mm256_set1_epi32(-cospi[52]);
+  const __m256i cospim20 = _mm256_set1_epi32(-cospi[20]);
+  const __m256i cospim36 = _mm256_set1_epi32(-cospi[36]);
+  const __m256i cospim4 = _mm256_set1_epi32(-cospi[4]);
+  const __m256i cospi56 = _mm256_set1_epi32(cospi[56]);
+  const __m256i cospi24 = _mm256_set1_epi32(cospi[24]);
+  const __m256i cospi40 = _mm256_set1_epi32(cospi[40]);
+  const __m256i cospi8 = _mm256_set1_epi32(cospi[8]);
+  const __m256i cospim40 = _mm256_set1_epi32(-cospi[40]);
+  const __m256i cospim8 = _mm256_set1_epi32(-cospi[8]);
+  const __m256i cospim56 = _mm256_set1_epi32(-cospi[56]);
+  const __m256i cospim24 = _mm256_set1_epi32(-cospi[24]);
+  const __m256i cospi32 = _mm256_set1_epi32(cospi[32]);
+  const __m256i cospim32 = _mm256_set1_epi32(-cospi[32]);
+  const __m256i cospi48 = _mm256_set1_epi32(cospi[48]);
+  const __m256i cospim48 = _mm256_set1_epi32(-cospi[48]);
+  const __m256i cospi16 = _mm256_set1_epi32(cospi[16]);
+  const __m256i cospim16 = _mm256_set1_epi32(-cospi[16]);
+  const __m256i rounding = _mm256_set1_epi32(1 << (bit - 1));
+  const int log_range = AOMMAX(16, bd + (do_cols ? 6 : 8));
+  const __m256i clamp_lo = _mm256_set1_epi32(-(1 << (log_range - 1)));
+  const __m256i clamp_hi = _mm256_set1_epi32((1 << (log_range - 1)) - 1);
+  __m256i bf1[32], bf0[32];
+
+  {
+    // stage 0
+    // stage 1
+    bf1[0] = in[0];
+    bf1[1] = in[16];
+    bf1[2] = in[8];
+    bf1[3] = in[24];
+    bf1[4] = in[4];
+    bf1[5] = in[20];
+    bf1[6] = in[12];
+    bf1[7] = in[28];
+    bf1[8] = in[2];
+    bf1[9] = in[18];
+    bf1[10] = in[10];
+    bf1[11] = in[26];
+    bf1[12] = in[6];
+    bf1[13] = in[22];
+    bf1[14] = in[14];
+    bf1[15] = in[30];
+    bf1[16] = in[1];
+    bf1[17] = in[17];
+    bf1[18] = in[9];
+    bf1[19] = in[25];
+    bf1[20] = in[5];
+    bf1[21] = in[21];
+    bf1[22] = in[13];
+    bf1[23] = in[29];
+    bf1[24] = in[3];
+    bf1[25] = in[19];
+    bf1[26] = in[11];
+    bf1[27] = in[27];
+    bf1[28] = in[7];
+    bf1[29] = in[23];
+    bf1[30] = in[15];
+    bf1[31] = in[31];
+
+    // stage 2
+    bf0[0] = bf1[0];
+    bf0[1] = bf1[1];
+    bf0[2] = bf1[2];
+    bf0[3] = bf1[3];
+    bf0[4] = bf1[4];
+    bf0[5] = bf1[5];
+    bf0[6] = bf1[6];
+    bf0[7] = bf1[7];
+    bf0[8] = bf1[8];
+    bf0[9] = bf1[9];
+    bf0[10] = bf1[10];
+    bf0[11] = bf1[11];
+    bf0[12] = bf1[12];
+    bf0[13] = bf1[13];
+    bf0[14] = bf1[14];
+    bf0[15] = bf1[15];
+    bf0[16] =
+        half_btf_avx2(&cospi62, &bf1[16], &cospim2, &bf1[31], &rounding, bit);
+    bf0[17] =
+        half_btf_avx2(&cospi30, &bf1[17], &cospim34, &bf1[30], &rounding, bit);
+    bf0[18] =
+        half_btf_avx2(&cospi46, &bf1[18], &cospim18, &bf1[29], &rounding, bit);
+    bf0[19] =
+        half_btf_avx2(&cospi14, &bf1[19], &cospim50, &bf1[28], &rounding, bit);
+    bf0[20] =
+        half_btf_avx2(&cospi54, &bf1[20], &cospim10, &bf1[27], &rounding, bit);
+    bf0[21] =
+        half_btf_avx2(&cospi22, &bf1[21], &cospim42, &bf1[26], &rounding, bit);
+    bf0[22] =
+        half_btf_avx2(&cospi38, &bf1[22], &cospim26, &bf1[25], &rounding, bit);
+    bf0[23] =
+        half_btf_avx2(&cospi6, &bf1[23], &cospim58, &bf1[24], &rounding, bit);
+    bf0[24] =
+        half_btf_avx2(&cospi58, &bf1[23], &cospi6, &bf1[24], &rounding, bit);
+    bf0[25] =
+        half_btf_avx2(&cospi26, &bf1[22], &cospi38, &bf1[25], &rounding, bit);
+    bf0[26] =
+        half_btf_avx2(&cospi42, &bf1[21], &cospi22, &bf1[26], &rounding, bit);
+    bf0[27] =
+        half_btf_avx2(&cospi10, &bf1[20], &cospi54, &bf1[27], &rounding, bit);
+    bf0[28] =
+        half_btf_avx2(&cospi50, &bf1[19], &cospi14, &bf1[28], &rounding, bit);
+    bf0[29] =
+        half_btf_avx2(&cospi18, &bf1[18], &cospi46, &bf1[29], &rounding, bit);
+    bf0[30] =
+        half_btf_avx2(&cospi34, &bf1[17], &cospi30, &bf1[30], &rounding, bit);
+    bf0[31] =
+        half_btf_avx2(&cospi2, &bf1[16], &cospi62, &bf1[31], &rounding, bit);
+
+    // stage 3
+    bf1[0] = bf0[0];
+    bf1[1] = bf0[1];
+    bf1[2] = bf0[2];
+    bf1[3] = bf0[3];
+    bf1[4] = bf0[4];
+    bf1[5] = bf0[5];
+    bf1[6] = bf0[6];
+    bf1[7] = bf0[7];
+    bf1[8] =
+        half_btf_avx2(&cospi60, &bf0[8], &cospim4, &bf0[15], &rounding, bit);
+    bf1[9] =
+        half_btf_avx2(&cospi28, &bf0[9], &cospim36, &bf0[14], &rounding, bit);
+    bf1[10] =
+        half_btf_avx2(&cospi44, &bf0[10], &cospim20, &bf0[13], &rounding, bit);
+    bf1[11] =
+        half_btf_avx2(&cospi12, &bf0[11], &cospim52, &bf0[12], &rounding, bit);
+    bf1[12] =
+        half_btf_avx2(&cospi52, &bf0[11], &cospi12, &bf0[12], &rounding, bit);
+    bf1[13] =
+        half_btf_avx2(&cospi20, &bf0[10], &cospi44, &bf0[13], &rounding, bit);
+    bf1[14] =
+        half_btf_avx2(&cospi36, &bf0[9], &cospi28, &bf0[14], &rounding, bit);
+    bf1[15] =
+        half_btf_avx2(&cospi4, &bf0[8], &cospi60, &bf0[15], &rounding, bit);
+
+    addsub_avx2(bf0[16], bf0[17], bf1 + 16, bf1 + 17, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf0[19], bf0[18], bf1 + 19, bf1 + 18, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf0[20], bf0[21], bf1 + 20, bf1 + 21, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf0[23], bf0[22], bf1 + 23, bf1 + 22, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf0[24], bf0[25], bf1 + 24, bf1 + 25, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf0[27], bf0[26], bf1 + 27, bf1 + 26, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf0[28], bf0[29], bf1 + 28, bf1 + 29, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf0[31], bf0[30], bf1 + 31, bf1 + 30, &clamp_lo, &clamp_hi);
+
+    // stage 4
+    bf0[0] = bf1[0];
+    bf0[1] = bf1[1];
+    bf0[2] = bf1[2];
+    bf0[3] = bf1[3];
+    bf0[4] =
+        half_btf_avx2(&cospi56, &bf1[4], &cospim8, &bf1[7], &rounding, bit);
+    bf0[5] =
+        half_btf_avx2(&cospi24, &bf1[5], &cospim40, &bf1[6], &rounding, bit);
+    bf0[6] =
+        half_btf_avx2(&cospi40, &bf1[5], &cospi24, &bf1[6], &rounding, bit);
+    bf0[7] = half_btf_avx2(&cospi8, &bf1[4], &cospi56, &bf1[7], &rounding, bit);
+
+    addsub_avx2(bf1[8], bf1[9], bf0 + 8, bf0 + 9, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf1[11], bf1[10], bf0 + 11, bf0 + 10, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf1[12], bf1[13], bf0 + 12, bf0 + 13, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf1[15], bf1[14], bf0 + 15, bf0 + 14, &clamp_lo, &clamp_hi);
+
+    bf0[16] = bf1[16];
+    bf0[17] =
+        half_btf_avx2(&cospim8, &bf1[17], &cospi56, &bf1[30], &rounding, bit);
+    bf0[18] =
+        half_btf_avx2(&cospim56, &bf1[18], &cospim8, &bf1[29], &rounding, bit);
+    bf0[19] = bf1[19];
+    bf0[20] = bf1[20];
+    bf0[21] =
+        half_btf_avx2(&cospim40, &bf1[21], &cospi24, &bf1[26], &rounding, bit);
+    bf0[22] =
+        half_btf_avx2(&cospim24, &bf1[22], &cospim40, &bf1[25], &rounding, bit);
+    bf0[23] = bf1[23];
+    bf0[24] = bf1[24];
+    bf0[25] =
+        half_btf_avx2(&cospim40, &bf1[22], &cospi24, &bf1[25], &rounding, bit);
+    bf0[26] =
+        half_btf_avx2(&cospi24, &bf1[21], &cospi40, &bf1[26], &rounding, bit);
+    bf0[27] = bf1[27];
+    bf0[28] = bf1[28];
+    bf0[29] =
+        half_btf_avx2(&cospim8, &bf1[18], &cospi56, &bf1[29], &rounding, bit);
+    bf0[30] =
+        half_btf_avx2(&cospi56, &bf1[17], &cospi8, &bf1[30], &rounding, bit);
+    bf0[31] = bf1[31];
+
+    // stage 5
+    bf1[0] =
+        half_btf_avx2(&cospi32, &bf0[0], &cospi32, &bf0[1], &rounding, bit);
+    bf1[1] =
+        half_btf_avx2(&cospi32, &bf0[0], &cospim32, &bf0[1], &rounding, bit);
+    bf1[2] =
+        half_btf_avx2(&cospi48, &bf0[2], &cospim16, &bf0[3], &rounding, bit);
+    bf1[3] =
+        half_btf_avx2(&cospi16, &bf0[2], &cospi48, &bf0[3], &rounding, bit);
+    addsub_avx2(bf0[4], bf0[5], bf1 + 4, bf1 + 5, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf0[7], bf0[6], bf1 + 7, bf1 + 6, &clamp_lo, &clamp_hi);
+    bf1[8] = bf0[8];
+    bf1[9] =
+        half_btf_avx2(&cospim16, &bf0[9], &cospi48, &bf0[14], &rounding, bit);
+    bf1[10] =
+        half_btf_avx2(&cospim48, &bf0[10], &cospim16, &bf0[13], &rounding, bit);
+    bf1[11] = bf0[11];
+    bf1[12] = bf0[12];
+    bf1[13] =
+        half_btf_avx2(&cospim16, &bf0[10], &cospi48, &bf0[13], &rounding, bit);
+    bf1[14] =
+        half_btf_avx2(&cospi48, &bf0[9], &cospi16, &bf0[14], &rounding, bit);
+    bf1[15] = bf0[15];
+    addsub_avx2(bf0[16], bf0[19], bf1 + 16, bf1 + 19, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf0[17], bf0[18], bf1 + 17, bf1 + 18, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf0[23], bf0[20], bf1 + 23, bf1 + 20, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf0[22], bf0[21], bf1 + 22, bf1 + 21, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf0[24], bf0[27], bf1 + 24, bf1 + 27, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf0[25], bf0[26], bf1 + 25, bf1 + 26, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf0[31], bf0[28], bf1 + 31, bf1 + 28, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf0[30], bf0[29], bf1 + 30, bf1 + 29, &clamp_lo, &clamp_hi);
+
+    // stage 6
+    addsub_avx2(bf1[0], bf1[3], bf0 + 0, bf0 + 3, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf1[1], bf1[2], bf0 + 1, bf0 + 2, &clamp_lo, &clamp_hi);
+    bf0[4] = bf1[4];
+    bf0[5] =
+        half_btf_avx2(&cospim32, &bf1[5], &cospi32, &bf1[6], &rounding, bit);
+    bf0[6] =
+        half_btf_avx2(&cospi32, &bf1[5], &cospi32, &bf1[6], &rounding, bit);
+    bf0[7] = bf1[7];
+    addsub_avx2(bf1[8], bf1[11], bf0 + 8, bf0 + 11, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf1[9], bf1[10], bf0 + 9, bf0 + 10, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf1[15], bf1[12], bf0 + 15, bf0 + 12, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf1[14], bf1[13], bf0 + 14, bf0 + 13, &clamp_lo, &clamp_hi);
+    bf0[16] = bf1[16];
+    bf0[17] = bf1[17];
+    bf0[18] =
+        half_btf_avx2(&cospim16, &bf1[18], &cospi48, &bf1[29], &rounding, bit);
+    bf0[19] =
+        half_btf_avx2(&cospim16, &bf1[19], &cospi48, &bf1[28], &rounding, bit);
+    bf0[20] =
+        half_btf_avx2(&cospim48, &bf1[20], &cospim16, &bf1[27], &rounding, bit);
+    bf0[21] =
+        half_btf_avx2(&cospim48, &bf1[21], &cospim16, &bf1[26], &rounding, bit);
+    bf0[22] = bf1[22];
+    bf0[23] = bf1[23];
+    bf0[24] = bf1[24];
+    bf0[25] = bf1[25];
+    bf0[26] =
+        half_btf_avx2(&cospim16, &bf1[21], &cospi48, &bf1[26], &rounding, bit);
+    bf0[27] =
+        half_btf_avx2(&cospim16, &bf1[20], &cospi48, &bf1[27], &rounding, bit);
+    bf0[28] =
+        half_btf_avx2(&cospi48, &bf1[19], &cospi16, &bf1[28], &rounding, bit);
+    bf0[29] =
+        half_btf_avx2(&cospi48, &bf1[18], &cospi16, &bf1[29], &rounding, bit);
+    bf0[30] = bf1[30];
+    bf0[31] = bf1[31];
+
+    // stage 7
+    addsub_avx2(bf0[0], bf0[7], bf1 + 0, bf1 + 7, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf0[1], bf0[6], bf1 + 1, bf1 + 6, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf0[2], bf0[5], bf1 + 2, bf1 + 5, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf0[3], bf0[4], bf1 + 3, bf1 + 4, &clamp_lo, &clamp_hi);
+    bf1[8] = bf0[8];
+    bf1[9] = bf0[9];
+    bf1[10] =
+        half_btf_avx2(&cospim32, &bf0[10], &cospi32, &bf0[13], &rounding, bit);
+    bf1[11] =
+        half_btf_avx2(&cospim32, &bf0[11], &cospi32, &bf0[12], &rounding, bit);
+    bf1[12] =
+        half_btf_avx2(&cospi32, &bf0[11], &cospi32, &bf0[12], &rounding, bit);
+    bf1[13] =
+        half_btf_avx2(&cospi32, &bf0[10], &cospi32, &bf0[13], &rounding, bit);
+    bf1[14] = bf0[14];
+    bf1[15] = bf0[15];
+    addsub_avx2(bf0[16], bf0[23], bf1 + 16, bf1 + 23, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf0[17], bf0[22], bf1 + 17, bf1 + 22, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf0[18], bf0[21], bf1 + 18, bf1 + 21, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf0[19], bf0[20], bf1 + 19, bf1 + 20, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf0[31], bf0[24], bf1 + 31, bf1 + 24, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf0[30], bf0[25], bf1 + 30, bf1 + 25, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf0[29], bf0[26], bf1 + 29, bf1 + 26, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf0[28], bf0[27], bf1 + 28, bf1 + 27, &clamp_lo, &clamp_hi);
+
+    // stage 8
+    addsub_avx2(bf1[0], bf1[15], bf0 + 0, bf0 + 15, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf1[1], bf1[14], bf0 + 1, bf0 + 14, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf1[2], bf1[13], bf0 + 2, bf0 + 13, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf1[3], bf1[12], bf0 + 3, bf0 + 12, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf1[4], bf1[11], bf0 + 4, bf0 + 11, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf1[5], bf1[10], bf0 + 5, bf0 + 10, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf1[6], bf1[9], bf0 + 6, bf0 + 9, &clamp_lo, &clamp_hi);
+    addsub_avx2(bf1[7], bf1[8], bf0 + 7, bf0 + 8, &clamp_lo, &clamp_hi);
+    bf0[16] = bf1[16];
+    bf0[17] = bf1[17];
+    bf0[18] = bf1[18];
+    bf0[19] = bf1[19];
+    bf0[20] =
+        half_btf_avx2(&cospim32, &bf1[20], &cospi32, &bf1[27], &rounding, bit);
+    bf0[21] =
+        half_btf_avx2(&cospim32, &bf1[21], &cospi32, &bf1[26], &rounding, bit);
+    bf0[22] =
+        half_btf_avx2(&cospim32, &bf1[22], &cospi32, &bf1[25], &rounding, bit);
+    bf0[23] =
+        half_btf_avx2(&cospim32, &bf1[23], &cospi32, &bf1[24], &rounding, bit);
+    bf0[24] =
+        half_btf_avx2(&cospi32, &bf1[23], &cospi32, &bf1[24], &rounding, bit);
+    bf0[25] =
+        half_btf_avx2(&cospi32, &bf1[22], &cospi32, &bf1[25], &rounding, bit);
+    bf0[26] =
+        half_btf_avx2(&cospi32, &bf1[21], &cospi32, &bf1[26], &rounding, bit);
+    bf0[27] =
+        half_btf_avx2(&cospi32, &bf1[20], &cospi32, &bf1[27], &rounding, bit);
+    bf0[28] = bf1[28];
+    bf0[29] = bf1[29];
+    bf0[30] = bf1[30];
+    bf0[31] = bf1[31];
+
+    // stage 9
+    if (do_cols) {
+      addsub_no_clamp_avx2(bf0[0], bf0[31], out + 0, out + 31);
+      addsub_no_clamp_avx2(bf0[1], bf0[30], out + 1, out + 30);
+      addsub_no_clamp_avx2(bf0[2], bf0[29], out + 2, out + 29);
+      addsub_no_clamp_avx2(bf0[3], bf0[28], out + 3, out + 28);
+      addsub_no_clamp_avx2(bf0[4], bf0[27], out + 4, out + 27);
+      addsub_no_clamp_avx2(bf0[5], bf0[26], out + 5, out + 26);
+      addsub_no_clamp_avx2(bf0[6], bf0[25], out + 6, out + 25);
+      addsub_no_clamp_avx2(bf0[7], bf0[24], out + 7, out + 24);
+      addsub_no_clamp_avx2(bf0[8], bf0[23], out + 8, out + 23);
+      addsub_no_clamp_avx2(bf0[9], bf0[22], out + 9, out + 22);
+      addsub_no_clamp_avx2(bf0[10], bf0[21], out + 10, out + 21);
+      addsub_no_clamp_avx2(bf0[11], bf0[20], out + 11, out + 20);
+      addsub_no_clamp_avx2(bf0[12], bf0[19], out + 12, out + 19);
+      addsub_no_clamp_avx2(bf0[13], bf0[18], out + 13, out + 18);
+      addsub_no_clamp_avx2(bf0[14], bf0[17], out + 14, out + 17);
+      addsub_no_clamp_avx2(bf0[15], bf0[16], out + 15, out + 16);
+    } else {
+      const int log_range_out = AOMMAX(16, bd + 6);
+      const __m256i clamp_lo_out = _mm256_set1_epi32(AOMMAX(
+          -(1 << (log_range_out - 1)), -(1 << (log_range - 1 - out_shift))));
+      const __m256i clamp_hi_out = _mm256_set1_epi32(AOMMIN(
+          (1 << (log_range_out - 1)) - 1, (1 << (log_range - 1 - out_shift))));
+
+      addsub_shift_avx2(bf0[0], bf0[31], out + 0, out + 31, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+      addsub_shift_avx2(bf0[1], bf0[30], out + 1, out + 30, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+      addsub_shift_avx2(bf0[2], bf0[29], out + 2, out + 29, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+      addsub_shift_avx2(bf0[3], bf0[28], out + 3, out + 28, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+      addsub_shift_avx2(bf0[4], bf0[27], out + 4, out + 27, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+      addsub_shift_avx2(bf0[5], bf0[26], out + 5, out + 26, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+      addsub_shift_avx2(bf0[6], bf0[25], out + 6, out + 25, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+      addsub_shift_avx2(bf0[7], bf0[24], out + 7, out + 24, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+      addsub_shift_avx2(bf0[8], bf0[23], out + 8, out + 23, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+      addsub_shift_avx2(bf0[9], bf0[22], out + 9, out + 22, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+      addsub_shift_avx2(bf0[10], bf0[21], out + 10, out + 21, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+      addsub_shift_avx2(bf0[11], bf0[20], out + 11, out + 20, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+      addsub_shift_avx2(bf0[12], bf0[19], out + 12, out + 19, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+      addsub_shift_avx2(bf0[13], bf0[18], out + 13, out + 18, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+      addsub_shift_avx2(bf0[14], bf0[17], out + 14, out + 17, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+      addsub_shift_avx2(bf0[15], bf0[16], out + 15, out + 16, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+    }
+  }
+}
+
+typedef void (*transform_1d_avx2)(__m256i *in, __m256i *out, int bit,
+                                  int do_cols, int bd, int out_shift);
+
+static const transform_1d_avx2
+    highbd_txfm_all_1d_zeros_w8_arr[TX_SIZES][ITX_TYPES_1D][4] = {
+      {
+          { NULL, NULL, NULL, NULL },
+          { NULL, NULL, NULL, NULL },
+          { NULL, NULL, NULL, NULL },
+      },
+      { { NULL, NULL, NULL, NULL },
+        { NULL, NULL, NULL, NULL },
+        { NULL, NULL, NULL, NULL } },
+      {
+          { NULL, NULL, NULL, NULL },
+          { NULL, NULL, NULL, NULL },
+          { NULL, NULL, NULL, NULL },
+      },
+      { { idct32_low1_avx2, idct32_low8_avx2, idct32_low16_avx2, idct32_avx2 },
+        { NULL, NULL, NULL, NULL },
+        { NULL, NULL, NULL, NULL } },
+
+      { { NULL, NULL, NULL, NULL },
+        { NULL, NULL, NULL, NULL },
+        { NULL, NULL, NULL, NULL } }
+    };
+
+static void highbd_inv_txfm2d_add_no_identity_avx2(const int32_t *input,
+                                                   uint16_t *output, int stride,
+                                                   TX_TYPE tx_type,
+                                                   TX_SIZE tx_size, int eob,
+                                                   const int bd) {
+  __m256i buf1[64 * 2];
+  int eobx, eoby;
+  get_eobx_eoby_scan_default(&eobx, &eoby, tx_size, eob);
+  const int8_t *shift = inv_txfm_shift_ls[tx_size];
+  const int txw_idx = get_txw_idx(tx_size);
+  const int txh_idx = get_txh_idx(tx_size);
+  const int txfm_size_col = tx_size_wide[tx_size];
+  const int txfm_size_row = tx_size_high[tx_size];
+  const int buf_size_w_div8 = txfm_size_col >> 3;
+  const int buf_size_nonzero_w_div8 = (eobx + 8) >> 3;
+  const int buf_size_nonzero_h_div8 = (eoby + 8) >> 3;
+  const int input_stride = AOMMIN(32, txfm_size_col);
+
+  const int fun_idx_x = lowbd_txfm_all_1d_zeros_idx[eobx];
+  const int fun_idx_y = lowbd_txfm_all_1d_zeros_idx[eoby];
+  const transform_1d_avx2 row_txfm =
+      highbd_txfm_all_1d_zeros_w8_arr[txw_idx][hitx_1d_tab[tx_type]][fun_idx_x];
+  const transform_1d_avx2 col_txfm =
+      highbd_txfm_all_1d_zeros_w8_arr[txh_idx][vitx_1d_tab[tx_type]][fun_idx_y];
+
+  assert(col_txfm != NULL);
+  assert(row_txfm != NULL);
+  int ud_flip, lr_flip;
+  get_flip_cfg(tx_type, &ud_flip, &lr_flip);
+
+  // 1st stage: column transform
+  for (int i = 0; i < buf_size_nonzero_h_div8; i++) {
+    __m256i buf0[32];
+    const int32_t *input_row = input + i * input_stride * 8;
+    for (int j = 0; j < buf_size_nonzero_w_div8; ++j) {
+      __m256i *buf0_cur = buf0 + j * 8;
+      load_buffer_32x32(input_row + j * 8, buf0_cur, input_stride, 8);
+
+      transpose_8x8_avx2(&buf0_cur[0], &buf0_cur[0]);
+    }
+
+    row_txfm(buf0, buf0, inv_cos_bit_row[txw_idx][txh_idx], 0, bd, -shift[0]);
+
+    __m256i *_buf1 = buf1 + i * 8;
+    for (int j = 0; j < buf_size_w_div8; ++j) {
+      transpose_8x8_avx2(&buf0[j * 8], &_buf1[j * txfm_size_row]);
+    }
+  }
+  // 2nd stage: column transform
+  for (int i = 0; i < buf_size_w_div8; i++) {
+    col_txfm(buf1 + i * txfm_size_row, buf1 + i * txfm_size_row,
+             inv_cos_bit_col[txw_idx][txh_idx], 1, bd, 0);
+
+    av1_round_shift_array_32_avx2(buf1 + i * txfm_size_row,
+                                  buf1 + i * txfm_size_row, txfm_size_row,
+                                  -shift[1]);
+  }
+
+  // write to buffer
+  {
+    for (int i = 0; i < (txfm_size_col >> 4); i++) {
+      highbd_write_buffer_16xn_avx2(buf1 + i * txfm_size_row * 2,
+                                    output + 16 * i, stride, ud_flip,
+                                    txfm_size_row, bd);
+    }
+  }
+}
+
+void av1_highbd_inv_txfm2d_add_universe_avx2(const int32_t *input,
+                                             uint8_t *output, int stride,
+                                             TX_TYPE tx_type, TX_SIZE tx_size,
+                                             int eob, const int bd) {
+  switch (tx_type) {
+    case DCT_DCT:
+      highbd_inv_txfm2d_add_no_identity_avx2(input, CONVERT_TO_SHORTPTR(output),
+                                             stride, tx_type, tx_size, eob, bd);
+      break;
+    default: assert(0); break;
+  }
+}
+
+void av1_highbd_inv_txfm_add_32x32_avx2(const tran_low_t *input, uint8_t *dest,
+                                        int stride,
+                                        const TxfmParam *txfm_param) {
+  const int bd = txfm_param->bd;
+  const TX_TYPE tx_type = txfm_param->tx_type;
+  const int32_t *src = cast_to_int32(input);
+  switch (tx_type) {
+    case DCT_DCT:
+      av1_highbd_inv_txfm2d_add_universe_avx2(input, dest, stride, tx_type,
+                                              txfm_param->tx_size,
+                                              txfm_param->eob, bd);
+      break;
+      // Assembly version doesn't support IDTX, so use C version for it.
+    case IDTX:
+      av1_inv_txfm2d_add_32x32_c(src, CONVERT_TO_SHORTPTR(dest), stride,
+                                 tx_type, bd);
+      break;
+
+    default: assert(0);
+  }
+}
+
+void av1_highbd_inv_txfm_add_avx2(const tran_low_t *input, uint8_t *dest,
+                                  int stride, const TxfmParam *txfm_param) {
+  assert(av1_ext_tx_used[txfm_param->tx_set_type][txfm_param->tx_type]);
+  const TX_SIZE tx_size = txfm_param->tx_size;
+  switch (tx_size) {
+    case TX_32X32:
+      av1_highbd_inv_txfm_add_32x32_avx2(input, dest, stride, txfm_param);
+      break;
+    case TX_16X16:
+      av1_highbd_inv_txfm_add_16x16_sse4_1(input, dest, stride, txfm_param);
+      break;
+    case TX_8X8:
+      av1_highbd_inv_txfm_add_8x8_sse4_1(input, dest, stride, txfm_param);
+      break;
+    case TX_4X8:
+      av1_highbd_inv_txfm_add_4x8(input, dest, stride, txfm_param);
+      break;
+    case TX_8X4:
+      av1_highbd_inv_txfm_add_8x4(input, dest, stride, txfm_param);
+      break;
+    case TX_8X16:
+      av1_highbd_inv_txfm_add_8x16_sse4_1(input, dest, stride, txfm_param);
+      break;
+    case TX_16X8:
+      av1_highbd_inv_txfm_add_16x8_sse4_1(input, dest, stride, txfm_param);
+      break;
+    case TX_16X32:
+      av1_highbd_inv_txfm_add_16x32(input, dest, stride, txfm_param);
+      break;
+    case TX_32X16:
+      av1_highbd_inv_txfm_add_32x16(input, dest, stride, txfm_param);
+      break;
+    case TX_32X64:
+      av1_highbd_inv_txfm_add_32x64(input, dest, stride, txfm_param);
+      break;
+    case TX_64X32:
+      av1_highbd_inv_txfm_add_64x32(input, dest, stride, txfm_param);
+      break;
+    case TX_4X4:
+      av1_highbd_inv_txfm_add_4x4_sse4_1(input, dest, stride, txfm_param);
+      break;
+    case TX_16X4:
+      av1_highbd_inv_txfm_add_16x4(input, dest, stride, txfm_param);
+      break;
+    case TX_4X16:
+      av1_highbd_inv_txfm_add_4x16(input, dest, stride, txfm_param);
+      break;
+    case TX_8X32:
+      av1_highbd_inv_txfm_add_8x32(input, dest, stride, txfm_param);
+      break;
+    case TX_32X8:
+      av1_highbd_inv_txfm_add_32x8(input, dest, stride, txfm_param);
+      break;
+    case TX_64X64:
+    case TX_16X64:
+    case TX_64X16:
+      av1_highbd_inv_txfm2d_add_universe_sse4_1(
+          input, dest, stride, txfm_param->tx_type, txfm_param->tx_size,
+          txfm_param->eob, txfm_param->bd);
+      break;
+    default: assert(0 && "Invalid transform size"); break;
+  }
+}
diff --git a/third_party/aom/av1/common/x86/highbd_inv_txfm_sse4.c b/third_party/aom/av1/common/x86/highbd_inv_txfm_sse4.c
new file mode 100644
index 000000000..e29e0baf5
--- /dev/null
+++ b/third_party/aom/av1/common/x86/highbd_inv_txfm_sse4.c
@@ -0,0 +1,5348 @@
+/*
+ * 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 <smmintrin.h> /* SSE4.1 */
+
+#include "config/aom_config.h"
+#include "config/av1_rtcd.h"
+
+#include "av1/common/av1_inv_txfm1d_cfg.h"
+#include "av1/common/idct.h"
+#include "av1/common/x86/av1_inv_txfm_ssse3.h"
+#include "av1/common/x86/av1_txfm_sse4.h"
+#include "av1/common/x86/highbd_txfm_utility_sse4.h"
+
+static INLINE __m128i highbd_clamp_epi16(__m128i u, int bd) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i one = _mm_set1_epi16(1);
+  const __m128i max = _mm_sub_epi16(_mm_slli_epi16(one, bd), one);
+  __m128i clamped, mask;
+
+  mask = _mm_cmpgt_epi16(u, max);
+  clamped = _mm_andnot_si128(mask, u);
+  mask = _mm_and_si128(mask, max);
+  clamped = _mm_or_si128(mask, clamped);
+  mask = _mm_cmpgt_epi16(clamped, zero);
+  clamped = _mm_and_si128(clamped, mask);
+
+  return clamped;
+}
+
+static INLINE __m128i highbd_get_recon_8x8_sse4_1(const __m128i pred,
+                                                  __m128i res0, __m128i res1,
+                                                  const int bd) {
+  __m128i x0 = _mm_cvtepi16_epi32(pred);
+  __m128i x1 = _mm_cvtepi16_epi32(_mm_srli_si128(pred, 8));
+
+  x0 = _mm_add_epi32(res0, x0);
+  x1 = _mm_add_epi32(res1, x1);
+  x0 = _mm_packus_epi32(x0, x1);
+  x0 = highbd_clamp_epi16(x0, bd);
+  return x0;
+}
+
+static INLINE void highbd_write_buffer_8xn_sse4_1(__m128i *in, uint16_t *output,
+                                                  int stride, int flipud,
+                                                  int height, const int bd) {
+  int j = flipud ? (height - 1) : 0;
+  const int step = flipud ? -1 : 1;
+  for (int i = 0; i < height; ++i, j += step) {
+    __m128i v = _mm_loadu_si128((__m128i const *)(output + i * stride));
+    __m128i u = highbd_get_recon_8x8_sse4_1(v, in[j], in[j + height], bd);
+
+    _mm_storeu_si128((__m128i *)(output + i * stride), u);
+  }
+}
+
+static INLINE void load_buffer_32bit_input(const int32_t *in, int stride,
+                                           __m128i *out, int out_size) {
+  for (int i = 0; i < out_size; ++i) {
+    out[i] = _mm_loadu_si128((const __m128i *)(in + i * stride));
+  }
+}
+
+static INLINE void load_buffer_4x4(const int32_t *coeff, __m128i *in) {
+  in[0] = _mm_load_si128((const __m128i *)(coeff + 0));
+  in[1] = _mm_load_si128((const __m128i *)(coeff + 4));
+  in[2] = _mm_load_si128((const __m128i *)(coeff + 8));
+  in[3] = _mm_load_si128((const __m128i *)(coeff + 12));
+}
+
+static void addsub_sse4_1(const __m128i in0, const __m128i in1, __m128i *out0,
+                          __m128i *out1, const __m128i *clamp_lo,
+                          const __m128i *clamp_hi) {
+  __m128i a0 = _mm_add_epi32(in0, in1);
+  __m128i a1 = _mm_sub_epi32(in0, in1);
+
+  a0 = _mm_max_epi32(a0, *clamp_lo);
+  a0 = _mm_min_epi32(a0, *clamp_hi);
+  a1 = _mm_max_epi32(a1, *clamp_lo);
+  a1 = _mm_min_epi32(a1, *clamp_hi);
+
+  *out0 = a0;
+  *out1 = a1;
+}
+
+static void addsub_no_clamp_sse4_1(const __m128i in0, const __m128i in1,
+                                   __m128i *out0, __m128i *out1) {
+  __m128i a0 = _mm_add_epi32(in0, in1);
+  __m128i a1 = _mm_sub_epi32(in0, in1);
+
+  *out0 = a0;
+  *out1 = a1;
+}
+
+static void addsub_shift_sse4_1(const __m128i in0, const __m128i in1,
+                                __m128i *out0, __m128i *out1,
+                                const __m128i *clamp_lo,
+                                const __m128i *clamp_hi, int shift) {
+  __m128i offset = _mm_set1_epi32((1 << shift) >> 1);
+  __m128i in0_w_offset = _mm_add_epi32(in0, offset);
+  __m128i a0 = _mm_add_epi32(in0_w_offset, in1);
+  __m128i a1 = _mm_sub_epi32(in0_w_offset, in1);
+
+  a0 = _mm_sra_epi32(a0, _mm_cvtsi32_si128(shift));
+  a1 = _mm_sra_epi32(a1, _mm_cvtsi32_si128(shift));
+
+  a0 = _mm_max_epi32(a0, *clamp_lo);
+  a0 = _mm_min_epi32(a0, *clamp_hi);
+  a1 = _mm_max_epi32(a1, *clamp_lo);
+  a1 = _mm_min_epi32(a1, *clamp_hi);
+
+  *out0 = a0;
+  *out1 = a1;
+}
+
+static INLINE void idct32_stage4_sse4_1(
+    __m128i *bf1, const __m128i *cospim8, const __m128i *cospi56,
+    const __m128i *cospi8, const __m128i *cospim56, const __m128i *cospim40,
+    const __m128i *cospi24, const __m128i *cospi40, const __m128i *cospim24,
+    const __m128i *rounding, int bit) {
+  __m128i temp1, temp2;
+  temp1 = half_btf_sse4_1(cospim8, &bf1[17], cospi56, &bf1[30], rounding, bit);
+  bf1[30] = half_btf_sse4_1(cospi56, &bf1[17], cospi8, &bf1[30], rounding, bit);
+  bf1[17] = temp1;
+
+  temp2 = half_btf_sse4_1(cospim56, &bf1[18], cospim8, &bf1[29], rounding, bit);
+  bf1[29] =
+      half_btf_sse4_1(cospim8, &bf1[18], cospi56, &bf1[29], rounding, bit);
+  bf1[18] = temp2;
+
+  temp1 = half_btf_sse4_1(cospim40, &bf1[21], cospi24, &bf1[26], rounding, bit);
+  bf1[26] =
+      half_btf_sse4_1(cospi24, &bf1[21], cospi40, &bf1[26], rounding, bit);
+  bf1[21] = temp1;
+
+  temp2 =
+      half_btf_sse4_1(cospim24, &bf1[22], cospim40, &bf1[25], rounding, bit);
+  bf1[25] =
+      half_btf_sse4_1(cospim40, &bf1[22], cospi24, &bf1[25], rounding, bit);
+  bf1[22] = temp2;
+}
+
+static INLINE void idct32_stage5_sse4_1(
+    __m128i *bf1, const __m128i *cospim16, const __m128i *cospi48,
+    const __m128i *cospi16, const __m128i *cospim48, const __m128i *clamp_lo,
+    const __m128i *clamp_hi, const __m128i *rounding, int bit) {
+  __m128i temp1, temp2;
+  temp1 = half_btf_sse4_1(cospim16, &bf1[9], cospi48, &bf1[14], rounding, bit);
+  bf1[14] = half_btf_sse4_1(cospi48, &bf1[9], cospi16, &bf1[14], rounding, bit);
+  bf1[9] = temp1;
+
+  temp2 =
+      half_btf_sse4_1(cospim48, &bf1[10], cospim16, &bf1[13], rounding, bit);
+  bf1[13] =
+      half_btf_sse4_1(cospim16, &bf1[10], cospi48, &bf1[13], rounding, bit);
+  bf1[10] = temp2;
+
+  addsub_sse4_1(bf1[16], bf1[19], bf1 + 16, bf1 + 19, clamp_lo, clamp_hi);
+  addsub_sse4_1(bf1[17], bf1[18], bf1 + 17, bf1 + 18, clamp_lo, clamp_hi);
+  addsub_sse4_1(bf1[23], bf1[20], bf1 + 23, bf1 + 20, clamp_lo, clamp_hi);
+  addsub_sse4_1(bf1[22], bf1[21], bf1 + 22, bf1 + 21, clamp_lo, clamp_hi);
+  addsub_sse4_1(bf1[24], bf1[27], bf1 + 24, bf1 + 27, clamp_lo, clamp_hi);
+  addsub_sse4_1(bf1[25], bf1[26], bf1 + 25, bf1 + 26, clamp_lo, clamp_hi);
+  addsub_sse4_1(bf1[31], bf1[28], bf1 + 31, bf1 + 28, clamp_lo, clamp_hi);
+  addsub_sse4_1(bf1[30], bf1[29], bf1 + 30, bf1 + 29, clamp_lo, clamp_hi);
+}
+
+static INLINE void idct32_stage6_sse4_1(
+    __m128i *bf1, const __m128i *cospim32, const __m128i *cospi32,
+    const __m128i *cospim16, const __m128i *cospi48, const __m128i *cospi16,
+    const __m128i *cospim48, const __m128i *clamp_lo, const __m128i *clamp_hi,
+    const __m128i *rounding, int bit) {
+  __m128i temp1, temp2;
+  temp1 = half_btf_sse4_1(cospim32, &bf1[5], cospi32, &bf1[6], rounding, bit);
+  bf1[6] = half_btf_sse4_1(cospi32, &bf1[5], cospi32, &bf1[6], rounding, bit);
+  bf1[5] = temp1;
+
+  addsub_sse4_1(bf1[8], bf1[11], bf1 + 8, bf1 + 11, clamp_lo, clamp_hi);
+  addsub_sse4_1(bf1[9], bf1[10], bf1 + 9, bf1 + 10, clamp_lo, clamp_hi);
+  addsub_sse4_1(bf1[15], bf1[12], bf1 + 15, bf1 + 12, clamp_lo, clamp_hi);
+  addsub_sse4_1(bf1[14], bf1[13], bf1 + 14, bf1 + 13, clamp_lo, clamp_hi);
+
+  temp1 = half_btf_sse4_1(cospim16, &bf1[18], cospi48, &bf1[29], rounding, bit);
+  bf1[29] =
+      half_btf_sse4_1(cospi48, &bf1[18], cospi16, &bf1[29], rounding, bit);
+  bf1[18] = temp1;
+  temp2 = half_btf_sse4_1(cospim16, &bf1[19], cospi48, &bf1[28], rounding, bit);
+  bf1[28] =
+      half_btf_sse4_1(cospi48, &bf1[19], cospi16, &bf1[28], rounding, bit);
+  bf1[19] = temp2;
+  temp1 =
+      half_btf_sse4_1(cospim48, &bf1[20], cospim16, &bf1[27], rounding, bit);
+  bf1[27] =
+      half_btf_sse4_1(cospim16, &bf1[20], cospi48, &bf1[27], rounding, bit);
+  bf1[20] = temp1;
+  temp2 =
+      half_btf_sse4_1(cospim48, &bf1[21], cospim16, &bf1[26], rounding, bit);
+  bf1[26] =
+      half_btf_sse4_1(cospim16, &bf1[21], cospi48, &bf1[26], rounding, bit);
+  bf1[21] = temp2;
+}
+
+static INLINE void idct32_stage7_sse4_1(__m128i *bf1, const __m128i *cospim32,
+                                        const __m128i *cospi32,
+                                        const __m128i *clamp_lo,
+                                        const __m128i *clamp_hi,
+                                        const __m128i *rounding, int bit) {
+  __m128i temp1, temp2;
+  addsub_sse4_1(bf1[0], bf1[7], bf1 + 0, bf1 + 7, clamp_lo, clamp_hi);
+  addsub_sse4_1(bf1[1], bf1[6], bf1 + 1, bf1 + 6, clamp_lo, clamp_hi);
+  addsub_sse4_1(bf1[2], bf1[5], bf1 + 2, bf1 + 5, clamp_lo, clamp_hi);
+  addsub_sse4_1(bf1[3], bf1[4], bf1 + 3, bf1 + 4, clamp_lo, clamp_hi);
+
+  temp1 = half_btf_sse4_1(cospim32, &bf1[10], cospi32, &bf1[13], rounding, bit);
+  bf1[13] =
+      half_btf_sse4_1(cospi32, &bf1[10], cospi32, &bf1[13], rounding, bit);
+  bf1[10] = temp1;
+  temp2 = half_btf_sse4_1(cospim32, &bf1[11], cospi32, &bf1[12], rounding, bit);
+  bf1[12] =
+      half_btf_sse4_1(cospi32, &bf1[11], cospi32, &bf1[12], rounding, bit);
+  bf1[11] = temp2;
+
+  addsub_sse4_1(bf1[16], bf1[23], bf1 + 16, bf1 + 23, clamp_lo, clamp_hi);
+  addsub_sse4_1(bf1[17], bf1[22], bf1 + 17, bf1 + 22, clamp_lo, clamp_hi);
+  addsub_sse4_1(bf1[18], bf1[21], bf1 + 18, bf1 + 21, clamp_lo, clamp_hi);
+  addsub_sse4_1(bf1[19], bf1[20], bf1 + 19, bf1 + 20, clamp_lo, clamp_hi);
+  addsub_sse4_1(bf1[31], bf1[24], bf1 + 31, bf1 + 24, clamp_lo, clamp_hi);
+  addsub_sse4_1(bf1[30], bf1[25], bf1 + 30, bf1 + 25, clamp_lo, clamp_hi);
+  addsub_sse4_1(bf1[29], bf1[26], bf1 + 29, bf1 + 26, clamp_lo, clamp_hi);
+  addsub_sse4_1(bf1[28], bf1[27], bf1 + 28, bf1 + 27, clamp_lo, clamp_hi);
+}
+
+static INLINE void idct32_stage8_sse4_1(__m128i *bf1, const __m128i *cospim32,
+                                        const __m128i *cospi32,
+                                        const __m128i *clamp_lo,
+                                        const __m128i *clamp_hi,
+                                        const __m128i *rounding, int bit) {
+  __m128i temp1, temp2;
+  addsub_sse4_1(bf1[0], bf1[15], bf1 + 0, bf1 + 15, clamp_lo, clamp_hi);
+  addsub_sse4_1(bf1[1], bf1[14], bf1 + 1, bf1 + 14, clamp_lo, clamp_hi);
+  addsub_sse4_1(bf1[2], bf1[13], bf1 + 2, bf1 + 13, clamp_lo, clamp_hi);
+  addsub_sse4_1(bf1[3], bf1[12], bf1 + 3, bf1 + 12, clamp_lo, clamp_hi);
+  addsub_sse4_1(bf1[4], bf1[11], bf1 + 4, bf1 + 11, clamp_lo, clamp_hi);
+  addsub_sse4_1(bf1[5], bf1[10], bf1 + 5, bf1 + 10, clamp_lo, clamp_hi);
+  addsub_sse4_1(bf1[6], bf1[9], bf1 + 6, bf1 + 9, clamp_lo, clamp_hi);
+  addsub_sse4_1(bf1[7], bf1[8], bf1 + 7, bf1 + 8, clamp_lo, clamp_hi);
+
+  temp1 = half_btf_sse4_1(cospim32, &bf1[20], cospi32, &bf1[27], rounding, bit);
+  bf1[27] =
+      half_btf_sse4_1(cospi32, &bf1[20], cospi32, &bf1[27], rounding, bit);
+  bf1[20] = temp1;
+  temp2 = half_btf_sse4_1(cospim32, &bf1[21], cospi32, &bf1[26], rounding, bit);
+  bf1[26] =
+      half_btf_sse4_1(cospi32, &bf1[21], cospi32, &bf1[26], rounding, bit);
+  bf1[21] = temp2;
+  temp1 = half_btf_sse4_1(cospim32, &bf1[22], cospi32, &bf1[25], rounding, bit);
+  bf1[25] =
+      half_btf_sse4_1(cospi32, &bf1[22], cospi32, &bf1[25], rounding, bit);
+  bf1[22] = temp1;
+  temp2 = half_btf_sse4_1(cospim32, &bf1[23], cospi32, &bf1[24], rounding, bit);
+  bf1[24] =
+      half_btf_sse4_1(cospi32, &bf1[23], cospi32, &bf1[24], rounding, bit);
+  bf1[23] = temp2;
+}
+
+static INLINE void idct32_stage9_sse4_1(__m128i *bf1, __m128i *out,
+                                        const int do_cols, const int bd,
+                                        const int out_shift,
+                                        const int log_range) {
+  if (do_cols) {
+    addsub_no_clamp_sse4_1(bf1[0], bf1[31], out + 0, out + 31);
+    addsub_no_clamp_sse4_1(bf1[1], bf1[30], out + 1, out + 30);
+    addsub_no_clamp_sse4_1(bf1[2], bf1[29], out + 2, out + 29);
+    addsub_no_clamp_sse4_1(bf1[3], bf1[28], out + 3, out + 28);
+    addsub_no_clamp_sse4_1(bf1[4], bf1[27], out + 4, out + 27);
+    addsub_no_clamp_sse4_1(bf1[5], bf1[26], out + 5, out + 26);
+    addsub_no_clamp_sse4_1(bf1[6], bf1[25], out + 6, out + 25);
+    addsub_no_clamp_sse4_1(bf1[7], bf1[24], out + 7, out + 24);
+    addsub_no_clamp_sse4_1(bf1[8], bf1[23], out + 8, out + 23);
+    addsub_no_clamp_sse4_1(bf1[9], bf1[22], out + 9, out + 22);
+    addsub_no_clamp_sse4_1(bf1[10], bf1[21], out + 10, out + 21);
+    addsub_no_clamp_sse4_1(bf1[11], bf1[20], out + 11, out + 20);
+    addsub_no_clamp_sse4_1(bf1[12], bf1[19], out + 12, out + 19);
+    addsub_no_clamp_sse4_1(bf1[13], bf1[18], out + 13, out + 18);
+    addsub_no_clamp_sse4_1(bf1[14], bf1[17], out + 14, out + 17);
+    addsub_no_clamp_sse4_1(bf1[15], bf1[16], out + 15, out + 16);
+  } else {
+    const int log_range_out = AOMMAX(16, bd + 6);
+    const __m128i clamp_lo_out = _mm_set1_epi32(AOMMAX(
+        -(1 << (log_range_out - 1)), -(1 << (log_range - 1 - out_shift))));
+    const __m128i clamp_hi_out = _mm_set1_epi32(AOMMIN(
+        (1 << (log_range_out - 1)) - 1, (1 << (log_range - 1 - out_shift))));
+
+    addsub_shift_sse4_1(bf1[0], bf1[31], out + 0, out + 31, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+    addsub_shift_sse4_1(bf1[1], bf1[30], out + 1, out + 30, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+    addsub_shift_sse4_1(bf1[2], bf1[29], out + 2, out + 29, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+    addsub_shift_sse4_1(bf1[3], bf1[28], out + 3, out + 28, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+    addsub_shift_sse4_1(bf1[4], bf1[27], out + 4, out + 27, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+    addsub_shift_sse4_1(bf1[5], bf1[26], out + 5, out + 26, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+    addsub_shift_sse4_1(bf1[6], bf1[25], out + 6, out + 25, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+    addsub_shift_sse4_1(bf1[7], bf1[24], out + 7, out + 24, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+    addsub_shift_sse4_1(bf1[8], bf1[23], out + 8, out + 23, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+    addsub_shift_sse4_1(bf1[9], bf1[22], out + 9, out + 22, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+    addsub_shift_sse4_1(bf1[10], bf1[21], out + 10, out + 21, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+    addsub_shift_sse4_1(bf1[11], bf1[20], out + 11, out + 20, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+    addsub_shift_sse4_1(bf1[12], bf1[19], out + 12, out + 19, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+    addsub_shift_sse4_1(bf1[13], bf1[18], out + 13, out + 18, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+    addsub_shift_sse4_1(bf1[14], bf1[17], out + 14, out + 17, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+    addsub_shift_sse4_1(bf1[15], bf1[16], out + 15, out + 16, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+  }
+}
+
+static void neg_shift_sse4_1(const __m128i in0, const __m128i in1,
+                             __m128i *out0, __m128i *out1,
+                             const __m128i *clamp_lo, const __m128i *clamp_hi,
+                             int shift) {
+  __m128i offset = _mm_set1_epi32((1 << shift) >> 1);
+  __m128i a0 = _mm_add_epi32(offset, in0);
+  __m128i a1 = _mm_sub_epi32(offset, in1);
+
+  a0 = _mm_sra_epi32(a0, _mm_cvtsi32_si128(shift));
+  a1 = _mm_sra_epi32(a1, _mm_cvtsi32_si128(shift));
+
+  a0 = _mm_max_epi32(a0, *clamp_lo);
+  a0 = _mm_min_epi32(a0, *clamp_hi);
+  a1 = _mm_max_epi32(a1, *clamp_lo);
+  a1 = _mm_min_epi32(a1, *clamp_hi);
+
+  *out0 = a0;
+  *out1 = a1;
+}
+
+static void idct4x4_sse4_1(__m128i *in, int bit, int do_cols, int bd) {
+  const int32_t *cospi = cospi_arr(bit);
+  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
+  const __m128i cospi48 = _mm_set1_epi32(cospi[48]);
+  const __m128i cospi16 = _mm_set1_epi32(cospi[16]);
+  const __m128i cospim16 = _mm_set1_epi32(-cospi[16]);
+  const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
+
+  __m128i u0, u1, u2, u3;
+  __m128i v0, v1, v2, v3, x, y;
+
+  v0 = _mm_unpacklo_epi32(in[0], in[1]);
+  v1 = _mm_unpackhi_epi32(in[0], in[1]);
+  v2 = _mm_unpacklo_epi32(in[2], in[3]);
+  v3 = _mm_unpackhi_epi32(in[2], in[3]);
+
+  u0 = _mm_unpacklo_epi64(v0, v2);
+  u1 = _mm_unpackhi_epi64(v0, v2);
+  u2 = _mm_unpacklo_epi64(v1, v3);
+  u3 = _mm_unpackhi_epi64(v1, v3);
+
+  x = _mm_mullo_epi32(u0, cospi32);
+  y = _mm_mullo_epi32(u2, cospi32);
+  v0 = _mm_add_epi32(x, y);
+  v0 = _mm_add_epi32(v0, rnding);
+  v0 = _mm_srai_epi32(v0, bit);
+
+  v1 = _mm_sub_epi32(x, y);
+  v1 = _mm_add_epi32(v1, rnding);
+  v1 = _mm_srai_epi32(v1, bit);
+
+  x = _mm_mullo_epi32(u1, cospi48);
+  y = _mm_mullo_epi32(u3, cospim16);
+  v2 = _mm_add_epi32(x, y);
+  v2 = _mm_add_epi32(v2, rnding);
+  v2 = _mm_srai_epi32(v2, bit);
+
+  x = _mm_mullo_epi32(u1, cospi16);
+  y = _mm_mullo_epi32(u3, cospi48);
+  v3 = _mm_add_epi32(x, y);
+  v3 = _mm_add_epi32(v3, rnding);
+  v3 = _mm_srai_epi32(v3, bit);
+
+  if (do_cols) {
+    addsub_no_clamp_sse4_1(v0, v3, in + 0, in + 3);
+    addsub_no_clamp_sse4_1(v1, v2, in + 1, in + 2);
+  } else {
+    const int log_range = AOMMAX(16, bd + 6);
+    const __m128i clamp_lo = _mm_set1_epi32(-(1 << (log_range - 1)));
+    const __m128i clamp_hi = _mm_set1_epi32((1 << (log_range - 1)) - 1);
+    addsub_sse4_1(v0, v3, in + 0, in + 3, &clamp_lo, &clamp_hi);
+    addsub_sse4_1(v1, v2, in + 1, in + 2, &clamp_lo, &clamp_hi);
+  }
+}
+
+static void iadst4x4_sse4_1(__m128i *in, int bit, int do_cols, int bd) {
+  const int32_t *sinpi = sinpi_arr(bit);
+  const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
+  const __m128i sinpi1 = _mm_set1_epi32((int)sinpi[1]);
+  const __m128i sinpi2 = _mm_set1_epi32((int)sinpi[2]);
+  const __m128i sinpi3 = _mm_set1_epi32((int)sinpi[3]);
+  const __m128i sinpi4 = _mm_set1_epi32((int)sinpi[4]);
+  __m128i t;
+  __m128i s0, s1, s2, s3, s4, s5, s6, s7;
+  __m128i x0, x1, x2, x3;
+  __m128i u0, u1, u2, u3;
+  __m128i v0, v1, v2, v3;
+
+  v0 = _mm_unpacklo_epi32(in[0], in[1]);
+  v1 = _mm_unpackhi_epi32(in[0], in[1]);
+  v2 = _mm_unpacklo_epi32(in[2], in[3]);
+  v3 = _mm_unpackhi_epi32(in[2], in[3]);
+
+  x0 = _mm_unpacklo_epi64(v0, v2);
+  x1 = _mm_unpackhi_epi64(v0, v2);
+  x2 = _mm_unpacklo_epi64(v1, v3);
+  x3 = _mm_unpackhi_epi64(v1, v3);
+
+  s0 = _mm_mullo_epi32(x0, sinpi1);
+  s1 = _mm_mullo_epi32(x0, sinpi2);
+  s2 = _mm_mullo_epi32(x1, sinpi3);
+  s3 = _mm_mullo_epi32(x2, sinpi4);
+  s4 = _mm_mullo_epi32(x2, sinpi1);
+  s5 = _mm_mullo_epi32(x3, sinpi2);
+  s6 = _mm_mullo_epi32(x3, sinpi4);
+  t = _mm_sub_epi32(x0, x2);
+  s7 = _mm_add_epi32(t, x3);
+
+  t = _mm_add_epi32(s0, s3);
+  s0 = _mm_add_epi32(t, s5);
+  t = _mm_sub_epi32(s1, s4);
+  s1 = _mm_sub_epi32(t, s6);
+  s3 = s2;
+  s2 = _mm_mullo_epi32(s7, sinpi3);
+
+  u0 = _mm_add_epi32(s0, s3);
+  u1 = _mm_add_epi32(s1, s3);
+  u2 = s2;
+  t = _mm_add_epi32(s0, s1);
+  u3 = _mm_sub_epi32(t, s3);
+
+  u0 = _mm_add_epi32(u0, rnding);
+  u0 = _mm_srai_epi32(u0, bit);
+
+  u1 = _mm_add_epi32(u1, rnding);
+  u1 = _mm_srai_epi32(u1, bit);
+
+  u2 = _mm_add_epi32(u2, rnding);
+  u2 = _mm_srai_epi32(u2, bit);
+
+  u3 = _mm_add_epi32(u3, rnding);
+  u3 = _mm_srai_epi32(u3, bit);
+
+  if (!do_cols) {
+    const int log_range = AOMMAX(16, bd + 6);
+    const __m128i clamp_lo = _mm_set1_epi32(-(1 << (log_range - 1)));
+    const __m128i clamp_hi = _mm_set1_epi32((1 << (log_range - 1)) - 1);
+
+    u0 = _mm_max_epi32(u0, clamp_lo);
+    u0 = _mm_min_epi32(u0, clamp_hi);
+    u1 = _mm_max_epi32(u1, clamp_lo);
+    u1 = _mm_min_epi32(u1, clamp_hi);
+    u2 = _mm_max_epi32(u2, clamp_lo);
+    u2 = _mm_min_epi32(u2, clamp_hi);
+    u3 = _mm_max_epi32(u3, clamp_lo);
+    u3 = _mm_min_epi32(u3, clamp_hi);
+  }
+
+  in[0] = u0;
+  in[1] = u1;
+  in[2] = u2;
+  in[3] = u3;
+}
+
+static INLINE void round_shift_4x4(__m128i *in, int shift) {
+  __m128i rnding = _mm_set1_epi32(1 << (shift - 1));
+
+  in[0] = _mm_add_epi32(in[0], rnding);
+  in[1] = _mm_add_epi32(in[1], rnding);
+  in[2] = _mm_add_epi32(in[2], rnding);
+  in[3] = _mm_add_epi32(in[3], rnding);
+
+  in[0] = _mm_srai_epi32(in[0], shift);
+  in[1] = _mm_srai_epi32(in[1], shift);
+  in[2] = _mm_srai_epi32(in[2], shift);
+  in[3] = _mm_srai_epi32(in[3], shift);
+}
+
+static void write_buffer_4x4(__m128i *in, uint16_t *output, int stride,
+                             int fliplr, int flipud, int shift, int bd) {
+  const __m128i zero = _mm_setzero_si128();
+  __m128i u0, u1, u2, u3;
+  __m128i v0, v1, v2, v3;
+
+  round_shift_4x4(in, shift);
+
+  v0 = _mm_loadl_epi64((__m128i const *)(output + 0 * stride));
+  v1 = _mm_loadl_epi64((__m128i const *)(output + 1 * stride));
+  v2 = _mm_loadl_epi64((__m128i const *)(output + 2 * stride));
+  v3 = _mm_loadl_epi64((__m128i const *)(output + 3 * stride));
+
+  v0 = _mm_unpacklo_epi16(v0, zero);
+  v1 = _mm_unpacklo_epi16(v1, zero);
+  v2 = _mm_unpacklo_epi16(v2, zero);
+  v3 = _mm_unpacklo_epi16(v3, zero);
+
+  if (fliplr) {
+    in[0] = _mm_shuffle_epi32(in[0], 0x1B);
+    in[1] = _mm_shuffle_epi32(in[1], 0x1B);
+    in[2] = _mm_shuffle_epi32(in[2], 0x1B);
+    in[3] = _mm_shuffle_epi32(in[3], 0x1B);
+  }
+
+  if (flipud) {
+    u0 = _mm_add_epi32(in[3], v0);
+    u1 = _mm_add_epi32(in[2], v1);
+    u2 = _mm_add_epi32(in[1], v2);
+    u3 = _mm_add_epi32(in[0], v3);
+  } else {
+    u0 = _mm_add_epi32(in[0], v0);
+    u1 = _mm_add_epi32(in[1], v1);
+    u2 = _mm_add_epi32(in[2], v2);
+    u3 = _mm_add_epi32(in[3], v3);
+  }
+
+  v0 = _mm_packus_epi32(u0, u1);
+  v2 = _mm_packus_epi32(u2, u3);
+
+  u0 = highbd_clamp_epi16(v0, bd);
+  u2 = highbd_clamp_epi16(v2, bd);
+
+  v0 = _mm_unpacklo_epi64(u0, u0);
+  v1 = _mm_unpackhi_epi64(u0, u0);
+  v2 = _mm_unpacklo_epi64(u2, u2);
+  v3 = _mm_unpackhi_epi64(u2, u2);
+
+  _mm_storel_epi64((__m128i *)(output + 0 * stride), v0);
+  _mm_storel_epi64((__m128i *)(output + 1 * stride), v1);
+  _mm_storel_epi64((__m128i *)(output + 2 * stride), v2);
+  _mm_storel_epi64((__m128i *)(output + 3 * stride), v3);
+}
+
+void av1_inv_txfm2d_add_4x4_sse4_1(const int32_t *coeff, uint16_t *output,
+                                   int stride, TX_TYPE tx_type, int bd) {
+  __m128i in[4];
+  const int8_t *shift = inv_txfm_shift_ls[TX_4X4];
+  const int txw_idx = get_txw_idx(TX_4X4);
+  const int txh_idx = get_txh_idx(TX_4X4);
+
+  switch (tx_type) {
+    case DCT_DCT:
+      load_buffer_4x4(coeff, in);
+      idct4x4_sse4_1(in, inv_cos_bit_row[txw_idx][txh_idx], 0, bd);
+      idct4x4_sse4_1(in, inv_cos_bit_col[txw_idx][txh_idx], 1, bd);
+      write_buffer_4x4(in, output, stride, 0, 0, -shift[1], bd);
+      break;
+    case ADST_DCT:
+      load_buffer_4x4(coeff, in);
+      idct4x4_sse4_1(in, inv_cos_bit_row[txw_idx][txh_idx], 0, bd);
+      iadst4x4_sse4_1(in, inv_cos_bit_col[txw_idx][txh_idx], 1, bd);
+      write_buffer_4x4(in, output, stride, 0, 0, -shift[1], bd);
+      break;
+    case DCT_ADST:
+      load_buffer_4x4(coeff, in);
+      iadst4x4_sse4_1(in, inv_cos_bit_row[txw_idx][txh_idx], 0, bd);
+      idct4x4_sse4_1(in, inv_cos_bit_col[txw_idx][txh_idx], 1, bd);
+      write_buffer_4x4(in, output, stride, 0, 0, -shift[1], bd);
+      break;
+    case ADST_ADST:
+      load_buffer_4x4(coeff, in);
+      iadst4x4_sse4_1(in, inv_cos_bit_row[txw_idx][txh_idx], 0, bd);
+      iadst4x4_sse4_1(in, inv_cos_bit_col[txw_idx][txh_idx], 1, bd);
+      write_buffer_4x4(in, output, stride, 0, 0, -shift[1], bd);
+      break;
+    case FLIPADST_DCT:
+      load_buffer_4x4(coeff, in);
+      idct4x4_sse4_1(in, inv_cos_bit_row[txw_idx][txh_idx], 0, bd);
+      iadst4x4_sse4_1(in, inv_cos_bit_col[txw_idx][txh_idx], 1, bd);
+      write_buffer_4x4(in, output, stride, 0, 1, -shift[1], bd);
+      break;
+    case DCT_FLIPADST:
+      load_buffer_4x4(coeff, in);
+      iadst4x4_sse4_1(in, inv_cos_bit_row[txw_idx][txh_idx], 0, bd);
+      idct4x4_sse4_1(in, inv_cos_bit_col[txw_idx][txh_idx], 1, bd);
+      write_buffer_4x4(in, output, stride, 1, 0, -shift[1], bd);
+      break;
+    case FLIPADST_FLIPADST:
+      load_buffer_4x4(coeff, in);
+      iadst4x4_sse4_1(in, inv_cos_bit_row[txw_idx][txh_idx], 0, bd);
+      iadst4x4_sse4_1(in, inv_cos_bit_col[txw_idx][txh_idx], 1, bd);
+      write_buffer_4x4(in, output, stride, 1, 1, -shift[1], bd);
+      break;
+    case ADST_FLIPADST:
+      load_buffer_4x4(coeff, in);
+      iadst4x4_sse4_1(in, inv_cos_bit_row[txw_idx][txh_idx], 0, bd);
+      iadst4x4_sse4_1(in, inv_cos_bit_col[txw_idx][txh_idx], 1, bd);
+      write_buffer_4x4(in, output, stride, 1, 0, -shift[1], bd);
+      break;
+    case FLIPADST_ADST:
+      load_buffer_4x4(coeff, in);
+      iadst4x4_sse4_1(in, inv_cos_bit_row[txw_idx][txh_idx], 0, bd);
+      iadst4x4_sse4_1(in, inv_cos_bit_col[txw_idx][txh_idx], 1, bd);
+      write_buffer_4x4(in, output, stride, 0, 1, -shift[1], bd);
+      break;
+    default: assert(0);
+  }
+}
+
+// 8x8
+static void load_buffer_8x8(const int32_t *coeff, __m128i *in) {
+  in[0] = _mm_load_si128((const __m128i *)(coeff + 0));
+  in[1] = _mm_load_si128((const __m128i *)(coeff + 4));
+  in[2] = _mm_load_si128((const __m128i *)(coeff + 8));
+  in[3] = _mm_load_si128((const __m128i *)(coeff + 12));
+  in[4] = _mm_load_si128((const __m128i *)(coeff + 16));
+  in[5] = _mm_load_si128((const __m128i *)(coeff + 20));
+  in[6] = _mm_load_si128((const __m128i *)(coeff + 24));
+  in[7] = _mm_load_si128((const __m128i *)(coeff + 28));
+  in[8] = _mm_load_si128((const __m128i *)(coeff + 32));
+  in[9] = _mm_load_si128((const __m128i *)(coeff + 36));
+  in[10] = _mm_load_si128((const __m128i *)(coeff + 40));
+  in[11] = _mm_load_si128((const __m128i *)(coeff + 44));
+  in[12] = _mm_load_si128((const __m128i *)(coeff + 48));
+  in[13] = _mm_load_si128((const __m128i *)(coeff + 52));
+  in[14] = _mm_load_si128((const __m128i *)(coeff + 56));
+  in[15] = _mm_load_si128((const __m128i *)(coeff + 60));
+}
+
+static void idct8x8_sse4_1(__m128i *in, __m128i *out, int bit, int do_cols,
+                           int bd, int out_shift) {
+  const int32_t *cospi = cospi_arr(bit);
+  const __m128i cospi56 = _mm_set1_epi32(cospi[56]);
+  const __m128i cospim8 = _mm_set1_epi32(-cospi[8]);
+  const __m128i cospi24 = _mm_set1_epi32(cospi[24]);
+  const __m128i cospim40 = _mm_set1_epi32(-cospi[40]);
+  const __m128i cospi40 = _mm_set1_epi32(cospi[40]);
+  const __m128i cospi8 = _mm_set1_epi32(cospi[8]);
+  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
+  const __m128i cospi48 = _mm_set1_epi32(cospi[48]);
+  const __m128i cospim16 = _mm_set1_epi32(-cospi[16]);
+  const __m128i cospi16 = _mm_set1_epi32(cospi[16]);
+  const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
+  const int log_range = AOMMAX(16, bd + (do_cols ? 6 : 8));
+  const __m128i clamp_lo = _mm_set1_epi32(-(1 << (log_range - 1)));
+  const __m128i clamp_hi = _mm_set1_epi32((1 << (log_range - 1)) - 1);
+  __m128i u0, u1, u2, u3, u4, u5, u6, u7;
+  __m128i v0, v1, v2, v3, v4, v5, v6, v7;
+  __m128i x, y;
+  int col;
+
+  // Note:
+  //  Even column: 0, 2, ..., 14
+  //  Odd column: 1, 3, ..., 15
+  //  one even column plus one odd column constructs one row (8 coeffs)
+  //  total we have 8 rows (8x8).
+  for (col = 0; col < 2; ++col) {
+    // stage 0
+    // stage 1
+    // stage 2
+    u0 = in[0 * 2 + col];
+    u1 = in[4 * 2 + col];
+    u2 = in[2 * 2 + col];
+    u3 = in[6 * 2 + col];
+
+    x = _mm_mullo_epi32(in[1 * 2 + col], cospi56);
+    y = _mm_mullo_epi32(in[7 * 2 + col], cospim8);
+    u4 = _mm_add_epi32(x, y);
+    u4 = _mm_add_epi32(u4, rnding);
+    u4 = _mm_srai_epi32(u4, bit);
+
+    x = _mm_mullo_epi32(in[1 * 2 + col], cospi8);
+    y = _mm_mullo_epi32(in[7 * 2 + col], cospi56);
+    u7 = _mm_add_epi32(x, y);
+    u7 = _mm_add_epi32(u7, rnding);
+    u7 = _mm_srai_epi32(u7, bit);
+
+    x = _mm_mullo_epi32(in[5 * 2 + col], cospi24);
+    y = _mm_mullo_epi32(in[3 * 2 + col], cospim40);
+    u5 = _mm_add_epi32(x, y);
+    u5 = _mm_add_epi32(u5, rnding);
+    u5 = _mm_srai_epi32(u5, bit);
+
+    x = _mm_mullo_epi32(in[5 * 2 + col], cospi40);
+    y = _mm_mullo_epi32(in[3 * 2 + col], cospi24);
+    u6 = _mm_add_epi32(x, y);
+    u6 = _mm_add_epi32(u6, rnding);
+    u6 = _mm_srai_epi32(u6, bit);
+
+    // stage 3
+    x = _mm_mullo_epi32(u0, cospi32);
+    y = _mm_mullo_epi32(u1, cospi32);
+    v0 = _mm_add_epi32(x, y);
+    v0 = _mm_add_epi32(v0, rnding);
+    v0 = _mm_srai_epi32(v0, bit);
+
+    v1 = _mm_sub_epi32(x, y);
+    v1 = _mm_add_epi32(v1, rnding);
+    v1 = _mm_srai_epi32(v1, bit);
+
+    x = _mm_mullo_epi32(u2, cospi48);
+    y = _mm_mullo_epi32(u3, cospim16);
+    v2 = _mm_add_epi32(x, y);
+    v2 = _mm_add_epi32(v2, rnding);
+    v2 = _mm_srai_epi32(v2, bit);
+
+    x = _mm_mullo_epi32(u2, cospi16);
+    y = _mm_mullo_epi32(u3, cospi48);
+    v3 = _mm_add_epi32(x, y);
+    v3 = _mm_add_epi32(v3, rnding);
+    v3 = _mm_srai_epi32(v3, bit);
+
+    addsub_sse4_1(u4, u5, &v4, &v5, &clamp_lo, &clamp_hi);
+    addsub_sse4_1(u7, u6, &v7, &v6, &clamp_lo, &clamp_hi);
+
+    // stage 4
+    addsub_sse4_1(v0, v3, &u0, &u3, &clamp_lo, &clamp_hi);
+    addsub_sse4_1(v1, v2, &u1, &u2, &clamp_lo, &clamp_hi);
+    u4 = v4;
+    u7 = v7;
+
+    x = _mm_mullo_epi32(v5, cospi32);
+    y = _mm_mullo_epi32(v6, cospi32);
+    u6 = _mm_add_epi32(y, x);
+    u6 = _mm_add_epi32(u6, rnding);
+    u6 = _mm_srai_epi32(u6, bit);
+
+    u5 = _mm_sub_epi32(y, x);
+    u5 = _mm_add_epi32(u5, rnding);
+    u5 = _mm_srai_epi32(u5, bit);
+
+    // stage 5
+    if (do_cols) {
+      addsub_no_clamp_sse4_1(u0, u7, out + 0 * 2 + col, out + 7 * 2 + col);
+      addsub_no_clamp_sse4_1(u1, u6, out + 1 * 2 + col, out + 6 * 2 + col);
+      addsub_no_clamp_sse4_1(u2, u5, out + 2 * 2 + col, out + 5 * 2 + col);
+      addsub_no_clamp_sse4_1(u3, u4, out + 3 * 2 + col, out + 4 * 2 + col);
+    } else {
+      const int log_range_out = AOMMAX(16, bd + 6);
+      const __m128i clamp_lo_out = _mm_set1_epi32(AOMMAX(
+          -(1 << (log_range_out - 1)), -(1 << (log_range - 1 - out_shift))));
+      const __m128i clamp_hi_out = _mm_set1_epi32(AOMMIN(
+          (1 << (log_range_out - 1)) - 1, (1 << (log_range - 1 - out_shift))));
+      addsub_shift_sse4_1(u0, u7, out + 0 * 2 + col, out + 7 * 2 + col,
+                          &clamp_lo_out, &clamp_hi_out, out_shift);
+      addsub_shift_sse4_1(u1, u6, out + 1 * 2 + col, out + 6 * 2 + col,
+                          &clamp_lo_out, &clamp_hi_out, out_shift);
+      addsub_shift_sse4_1(u2, u5, out + 2 * 2 + col, out + 5 * 2 + col,
+                          &clamp_lo_out, &clamp_hi_out, out_shift);
+      addsub_shift_sse4_1(u3, u4, out + 3 * 2 + col, out + 4 * 2 + col,
+                          &clamp_lo_out, &clamp_hi_out, out_shift);
+    }
+  }
+}
+
+static void iadst8x8_sse4_1(__m128i *in, __m128i *out, int bit, int do_cols,
+                            int bd, int out_shift) {
+  const int32_t *cospi = cospi_arr(bit);
+  const __m128i cospi4 = _mm_set1_epi32(cospi[4]);
+  const __m128i cospi60 = _mm_set1_epi32(cospi[60]);
+  const __m128i cospi20 = _mm_set1_epi32(cospi[20]);
+  const __m128i cospi44 = _mm_set1_epi32(cospi[44]);
+  const __m128i cospi36 = _mm_set1_epi32(cospi[36]);
+  const __m128i cospi28 = _mm_set1_epi32(cospi[28]);
+  const __m128i cospi52 = _mm_set1_epi32(cospi[52]);
+  const __m128i cospi12 = _mm_set1_epi32(cospi[12]);
+  const __m128i cospi16 = _mm_set1_epi32(cospi[16]);
+  const __m128i cospi48 = _mm_set1_epi32(cospi[48]);
+  const __m128i cospim48 = _mm_set1_epi32(-cospi[48]);
+  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
+  const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
+  const __m128i kZero = _mm_setzero_si128();
+  const int log_range = AOMMAX(16, bd + (do_cols ? 6 : 8));
+  const __m128i clamp_lo = _mm_set1_epi32(-(1 << (log_range - 1)));
+  const __m128i clamp_hi = _mm_set1_epi32((1 << (log_range - 1)) - 1);
+  __m128i u[8], v[8], x;
+
+  // Even 8 points: 0, 2, ..., 14
+  // stage 0
+  // stage 1
+  // stage 2
+  // (1)
+  u[0] = _mm_mullo_epi32(in[14], cospi4);
+  x = _mm_mullo_epi32(in[0], cospi60);
+  u[0] = _mm_add_epi32(u[0], x);
+  u[0] = _mm_add_epi32(u[0], rnding);
+  u[0] = _mm_srai_epi32(u[0], bit);
+
+  u[1] = _mm_mullo_epi32(in[14], cospi60);
+  x = _mm_mullo_epi32(in[0], cospi4);
+  u[1] = _mm_sub_epi32(u[1], x);
+  u[1] = _mm_add_epi32(u[1], rnding);
+  u[1] = _mm_srai_epi32(u[1], bit);
+
+  // (2)
+  u[2] = _mm_mullo_epi32(in[10], cospi20);
+  x = _mm_mullo_epi32(in[4], cospi44);
+  u[2] = _mm_add_epi32(u[2], x);
+  u[2] = _mm_add_epi32(u[2], rnding);
+  u[2] = _mm_srai_epi32(u[2], bit);
+
+  u[3] = _mm_mullo_epi32(in[10], cospi44);
+  x = _mm_mullo_epi32(in[4], cospi20);
+  u[3] = _mm_sub_epi32(u[3], x);
+  u[3] = _mm_add_epi32(u[3], rnding);
+  u[3] = _mm_srai_epi32(u[3], bit);
+
+  // (3)
+  u[4] = _mm_mullo_epi32(in[6], cospi36);
+  x = _mm_mullo_epi32(in[8], cospi28);
+  u[4] = _mm_add_epi32(u[4], x);
+  u[4] = _mm_add_epi32(u[4], rnding);
+  u[4] = _mm_srai_epi32(u[4], bit);
+
+  u[5] = _mm_mullo_epi32(in[6], cospi28);
+  x = _mm_mullo_epi32(in[8], cospi36);
+  u[5] = _mm_sub_epi32(u[5], x);
+  u[5] = _mm_add_epi32(u[5], rnding);
+  u[5] = _mm_srai_epi32(u[5], bit);
+
+  // (4)
+  u[6] = _mm_mullo_epi32(in[2], cospi52);
+  x = _mm_mullo_epi32(in[12], cospi12);
+  u[6] = _mm_add_epi32(u[6], x);
+  u[6] = _mm_add_epi32(u[6], rnding);
+  u[6] = _mm_srai_epi32(u[6], bit);
+
+  u[7] = _mm_mullo_epi32(in[2], cospi12);
+  x = _mm_mullo_epi32(in[12], cospi52);
+  u[7] = _mm_sub_epi32(u[7], x);
+  u[7] = _mm_add_epi32(u[7], rnding);
+  u[7] = _mm_srai_epi32(u[7], bit);
+
+  // stage 3
+  addsub_sse4_1(u[0], u[4], &v[0], &v[4], &clamp_lo, &clamp_hi);
+  addsub_sse4_1(u[1], u[5], &v[1], &v[5], &clamp_lo, &clamp_hi);
+  addsub_sse4_1(u[2], u[6], &v[2], &v[6], &clamp_lo, &clamp_hi);
+  addsub_sse4_1(u[3], u[7], &v[3], &v[7], &clamp_lo, &clamp_hi);
+
+  // stage 4
+  u[0] = v[0];
+  u[1] = v[1];
+  u[2] = v[2];
+  u[3] = v[3];
+
+  u[4] = _mm_mullo_epi32(v[4], cospi16);
+  x = _mm_mullo_epi32(v[5], cospi48);
+  u[4] = _mm_add_epi32(u[4], x);
+  u[4] = _mm_add_epi32(u[4], rnding);
+  u[4] = _mm_srai_epi32(u[4], bit);
+
+  u[5] = _mm_mullo_epi32(v[4], cospi48);
+  x = _mm_mullo_epi32(v[5], cospi16);
+  u[5] = _mm_sub_epi32(u[5], x);
+  u[5] = _mm_add_epi32(u[5], rnding);
+  u[5] = _mm_srai_epi32(u[5], bit);
+
+  u[6] = _mm_mullo_epi32(v[6], cospim48);
+  x = _mm_mullo_epi32(v[7], cospi16);
+  u[6] = _mm_add_epi32(u[6], x);
+  u[6] = _mm_add_epi32(u[6], rnding);
+  u[6] = _mm_srai_epi32(u[6], bit);
+
+  u[7] = _mm_mullo_epi32(v[6], cospi16);
+  x = _mm_mullo_epi32(v[7], cospim48);
+  u[7] = _mm_sub_epi32(u[7], x);
+  u[7] = _mm_add_epi32(u[7], rnding);
+  u[7] = _mm_srai_epi32(u[7], bit);
+
+  // stage 5
+  addsub_sse4_1(u[0], u[2], &v[0], &v[2], &clamp_lo, &clamp_hi);
+  addsub_sse4_1(u[1], u[3], &v[1], &v[3], &clamp_lo, &clamp_hi);
+  addsub_sse4_1(u[4], u[6], &v[4], &v[6], &clamp_lo, &clamp_hi);
+  addsub_sse4_1(u[5], u[7], &v[5], &v[7], &clamp_lo, &clamp_hi);
+
+  // stage 6
+  u[0] = v[0];
+  u[1] = v[1];
+  u[4] = v[4];
+  u[5] = v[5];
+
+  v[0] = _mm_mullo_epi32(v[2], cospi32);
+  x = _mm_mullo_epi32(v[3], cospi32);
+  u[2] = _mm_add_epi32(v[0], x);
+  u[2] = _mm_add_epi32(u[2], rnding);
+  u[2] = _mm_srai_epi32(u[2], bit);
+
+  u[3] = _mm_sub_epi32(v[0], x);
+  u[3] = _mm_add_epi32(u[3], rnding);
+  u[3] = _mm_srai_epi32(u[3], bit);
+
+  v[0] = _mm_mullo_epi32(v[6], cospi32);
+  x = _mm_mullo_epi32(v[7], cospi32);
+  u[6] = _mm_add_epi32(v[0], x);
+  u[6] = _mm_add_epi32(u[6], rnding);
+  u[6] = _mm_srai_epi32(u[6], bit);
+
+  u[7] = _mm_sub_epi32(v[0], x);
+  u[7] = _mm_add_epi32(u[7], rnding);
+  u[7] = _mm_srai_epi32(u[7], bit);
+
+  // stage 7
+  if (do_cols) {
+    out[0] = u[0];
+    out[2] = _mm_sub_epi32(kZero, u[4]);
+    out[4] = u[6];
+    out[6] = _mm_sub_epi32(kZero, u[2]);
+    out[8] = u[3];
+    out[10] = _mm_sub_epi32(kZero, u[7]);
+    out[12] = u[5];
+    out[14] = _mm_sub_epi32(kZero, u[1]);
+  } else {
+    const int log_range_out = AOMMAX(16, bd + 6);
+    const __m128i clamp_lo_out = _mm_set1_epi32(-(1 << (log_range_out - 1)));
+    const __m128i clamp_hi_out = _mm_set1_epi32((1 << (log_range_out - 1)) - 1);
+
+    neg_shift_sse4_1(u[0], u[4], out + 0, out + 2, &clamp_lo_out, &clamp_hi_out,
+                     out_shift);
+    neg_shift_sse4_1(u[6], u[2], out + 4, out + 6, &clamp_lo_out, &clamp_hi_out,
+                     out_shift);
+    neg_shift_sse4_1(u[3], u[7], out + 8, out + 10, &clamp_lo_out,
+                     &clamp_hi_out, out_shift);
+    neg_shift_sse4_1(u[5], u[1], out + 12, out + 14, &clamp_lo_out,
+                     &clamp_hi_out, out_shift);
+  }
+
+  // Odd 8 points: 1, 3, ..., 15
+  // stage 0
+  // stage 1
+  // stage 2
+  // (1)
+  u[0] = _mm_mullo_epi32(in[15], cospi4);
+  x = _mm_mullo_epi32(in[1], cospi60);
+  u[0] = _mm_add_epi32(u[0], x);
+  u[0] = _mm_add_epi32(u[0], rnding);
+  u[0] = _mm_srai_epi32(u[0], bit);
+
+  u[1] = _mm_mullo_epi32(in[15], cospi60);
+  x = _mm_mullo_epi32(in[1], cospi4);
+  u[1] = _mm_sub_epi32(u[1], x);
+  u[1] = _mm_add_epi32(u[1], rnding);
+  u[1] = _mm_srai_epi32(u[1], bit);
+
+  // (2)
+  u[2] = _mm_mullo_epi32(in[11], cospi20);
+  x = _mm_mullo_epi32(in[5], cospi44);
+  u[2] = _mm_add_epi32(u[2], x);
+  u[2] = _mm_add_epi32(u[2], rnding);
+  u[2] = _mm_srai_epi32(u[2], bit);
+
+  u[3] = _mm_mullo_epi32(in[11], cospi44);
+  x = _mm_mullo_epi32(in[5], cospi20);
+  u[3] = _mm_sub_epi32(u[3], x);
+  u[3] = _mm_add_epi32(u[3], rnding);
+  u[3] = _mm_srai_epi32(u[3], bit);
+
+  // (3)
+  u[4] = _mm_mullo_epi32(in[7], cospi36);
+  x = _mm_mullo_epi32(in[9], cospi28);
+  u[4] = _mm_add_epi32(u[4], x);
+  u[4] = _mm_add_epi32(u[4], rnding);
+  u[4] = _mm_srai_epi32(u[4], bit);
+
+  u[5] = _mm_mullo_epi32(in[7], cospi28);
+  x = _mm_mullo_epi32(in[9], cospi36);
+  u[5] = _mm_sub_epi32(u[5], x);
+  u[5] = _mm_add_epi32(u[5], rnding);
+  u[5] = _mm_srai_epi32(u[5], bit);
+
+  // (4)
+  u[6] = _mm_mullo_epi32(in[3], cospi52);
+  x = _mm_mullo_epi32(in[13], cospi12);
+  u[6] = _mm_add_epi32(u[6], x);
+  u[6] = _mm_add_epi32(u[6], rnding);
+  u[6] = _mm_srai_epi32(u[6], bit);
+
+  u[7] = _mm_mullo_epi32(in[3], cospi12);
+  x = _mm_mullo_epi32(in[13], cospi52);
+  u[7] = _mm_sub_epi32(u[7], x);
+  u[7] = _mm_add_epi32(u[7], rnding);
+  u[7] = _mm_srai_epi32(u[7], bit);
+
+  // stage 3
+  addsub_sse4_1(u[0], u[4], &v[0], &v[4], &clamp_lo, &clamp_hi);
+  addsub_sse4_1(u[1], u[5], &v[1], &v[5], &clamp_lo, &clamp_hi);
+  addsub_sse4_1(u[2], u[6], &v[2], &v[6], &clamp_lo, &clamp_hi);
+  addsub_sse4_1(u[3], u[7], &v[3], &v[7], &clamp_lo, &clamp_hi);
+
+  // stage 4
+  u[0] = v[0];
+  u[1] = v[1];
+  u[2] = v[2];
+  u[3] = v[3];
+
+  u[4] = _mm_mullo_epi32(v[4], cospi16);
+  x = _mm_mullo_epi32(v[5], cospi48);
+  u[4] = _mm_add_epi32(u[4], x);
+  u[4] = _mm_add_epi32(u[4], rnding);
+  u[4] = _mm_srai_epi32(u[4], bit);
+
+  u[5] = _mm_mullo_epi32(v[4], cospi48);
+  x = _mm_mullo_epi32(v[5], cospi16);
+  u[5] = _mm_sub_epi32(u[5], x);
+  u[5] = _mm_add_epi32(u[5], rnding);
+  u[5] = _mm_srai_epi32(u[5], bit);
+
+  u[6] = _mm_mullo_epi32(v[6], cospim48);
+  x = _mm_mullo_epi32(v[7], cospi16);
+  u[6] = _mm_add_epi32(u[6], x);
+  u[6] = _mm_add_epi32(u[6], rnding);
+  u[6] = _mm_srai_epi32(u[6], bit);
+
+  u[7] = _mm_mullo_epi32(v[6], cospi16);
+  x = _mm_mullo_epi32(v[7], cospim48);
+  u[7] = _mm_sub_epi32(u[7], x);
+  u[7] = _mm_add_epi32(u[7], rnding);
+  u[7] = _mm_srai_epi32(u[7], bit);
+
+  // stage 5
+  addsub_sse4_1(u[0], u[2], &v[0], &v[2], &clamp_lo, &clamp_hi);
+  addsub_sse4_1(u[1], u[3], &v[1], &v[3], &clamp_lo, &clamp_hi);
+  addsub_sse4_1(u[4], u[6], &v[4], &v[6], &clamp_lo, &clamp_hi);
+  addsub_sse4_1(u[5], u[7], &v[5], &v[7], &clamp_lo, &clamp_hi);
+
+  // stage 6
+  u[0] = v[0];
+  u[1] = v[1];
+  u[4] = v[4];
+  u[5] = v[5];
+
+  v[0] = _mm_mullo_epi32(v[2], cospi32);
+  x = _mm_mullo_epi32(v[3], cospi32);
+  u[2] = _mm_add_epi32(v[0], x);
+  u[2] = _mm_add_epi32(u[2], rnding);
+  u[2] = _mm_srai_epi32(u[2], bit);
+
+  u[3] = _mm_sub_epi32(v[0], x);
+  u[3] = _mm_add_epi32(u[3], rnding);
+  u[3] = _mm_srai_epi32(u[3], bit);
+
+  v[0] = _mm_mullo_epi32(v[6], cospi32);
+  x = _mm_mullo_epi32(v[7], cospi32);
+  u[6] = _mm_add_epi32(v[0], x);
+  u[6] = _mm_add_epi32(u[6], rnding);
+  u[6] = _mm_srai_epi32(u[6], bit);
+
+  u[7] = _mm_sub_epi32(v[0], x);
+  u[7] = _mm_add_epi32(u[7], rnding);
+  u[7] = _mm_srai_epi32(u[7], bit);
+
+  // stage 7
+  if (do_cols) {
+    out[1] = u[0];
+    out[3] = _mm_sub_epi32(kZero, u[4]);
+    out[5] = u[6];
+    out[7] = _mm_sub_epi32(kZero, u[2]);
+    out[9] = u[3];
+    out[11] = _mm_sub_epi32(kZero, u[7]);
+    out[13] = u[5];
+    out[15] = _mm_sub_epi32(kZero, u[1]);
+  } else {
+    const int log_range_out = AOMMAX(16, bd + 6);
+    const __m128i clamp_lo_out = _mm_set1_epi32(-(1 << (log_range_out - 1)));
+    const __m128i clamp_hi_out = _mm_set1_epi32((1 << (log_range_out - 1)) - 1);
+
+    neg_shift_sse4_1(u[0], u[4], out + 1, out + 3, &clamp_lo_out, &clamp_hi_out,
+                     out_shift);
+    neg_shift_sse4_1(u[6], u[2], out + 5, out + 7, &clamp_lo_out, &clamp_hi_out,
+                     out_shift);
+    neg_shift_sse4_1(u[3], u[7], out + 9, out + 11, &clamp_lo_out,
+                     &clamp_hi_out, out_shift);
+    neg_shift_sse4_1(u[5], u[1], out + 13, out + 15, &clamp_lo_out,
+                     &clamp_hi_out, out_shift);
+  }
+}
+
+static void round_shift_8x8(__m128i *in, int shift) {
+  round_shift_4x4(&in[0], shift);
+  round_shift_4x4(&in[4], shift);
+  round_shift_4x4(&in[8], shift);
+  round_shift_4x4(&in[12], shift);
+}
+
+static __m128i get_recon_8x8(const __m128i pred, __m128i res_lo, __m128i res_hi,
+                             int fliplr, int bd) {
+  __m128i x0, x1;
+  const __m128i zero = _mm_setzero_si128();
+
+  x0 = _mm_unpacklo_epi16(pred, zero);
+  x1 = _mm_unpackhi_epi16(pred, zero);
+
+  if (fliplr) {
+    res_lo = _mm_shuffle_epi32(res_lo, 0x1B);
+    res_hi = _mm_shuffle_epi32(res_hi, 0x1B);
+    x0 = _mm_add_epi32(res_hi, x0);
+    x1 = _mm_add_epi32(res_lo, x1);
+
+  } else {
+    x0 = _mm_add_epi32(res_lo, x0);
+    x1 = _mm_add_epi32(res_hi, x1);
+  }
+
+  x0 = _mm_packus_epi32(x0, x1);
+  return highbd_clamp_epi16(x0, bd);
+}
+
+static void write_buffer_8x8(__m128i *in, uint16_t *output, int stride,
+                             int fliplr, int flipud, int shift, int bd) {
+  __m128i u0, u1, u2, u3, u4, u5, u6, u7;
+  __m128i v0, v1, v2, v3, v4, v5, v6, v7;
+
+  round_shift_8x8(in, shift);
+
+  v0 = _mm_load_si128((__m128i const *)(output + 0 * stride));
+  v1 = _mm_load_si128((__m128i const *)(output + 1 * stride));
+  v2 = _mm_load_si128((__m128i const *)(output + 2 * stride));
+  v3 = _mm_load_si128((__m128i const *)(output + 3 * stride));
+  v4 = _mm_load_si128((__m128i const *)(output + 4 * stride));
+  v5 = _mm_load_si128((__m128i const *)(output + 5 * stride));
+  v6 = _mm_load_si128((__m128i const *)(output + 6 * stride));
+  v7 = _mm_load_si128((__m128i const *)(output + 7 * stride));
+
+  if (flipud) {
+    u0 = get_recon_8x8(v0, in[14], in[15], fliplr, bd);
+    u1 = get_recon_8x8(v1, in[12], in[13], fliplr, bd);
+    u2 = get_recon_8x8(v2, in[10], in[11], fliplr, bd);
+    u3 = get_recon_8x8(v3, in[8], in[9], fliplr, bd);
+    u4 = get_recon_8x8(v4, in[6], in[7], fliplr, bd);
+    u5 = get_recon_8x8(v5, in[4], in[5], fliplr, bd);
+    u6 = get_recon_8x8(v6, in[2], in[3], fliplr, bd);
+    u7 = get_recon_8x8(v7, in[0], in[1], fliplr, bd);
+  } else {
+    u0 = get_recon_8x8(v0, in[0], in[1], fliplr, bd);
+    u1 = get_recon_8x8(v1, in[2], in[3], fliplr, bd);
+    u2 = get_recon_8x8(v2, in[4], in[5], fliplr, bd);
+    u3 = get_recon_8x8(v3, in[6], in[7], fliplr, bd);
+    u4 = get_recon_8x8(v4, in[8], in[9], fliplr, bd);
+    u5 = get_recon_8x8(v5, in[10], in[11], fliplr, bd);
+    u6 = get_recon_8x8(v6, in[12], in[13], fliplr, bd);
+    u7 = get_recon_8x8(v7, in[14], in[15], fliplr, bd);
+  }
+
+  _mm_store_si128((__m128i *)(output + 0 * stride), u0);
+  _mm_store_si128((__m128i *)(output + 1 * stride), u1);
+  _mm_store_si128((__m128i *)(output + 2 * stride), u2);
+  _mm_store_si128((__m128i *)(output + 3 * stride), u3);
+  _mm_store_si128((__m128i *)(output + 4 * stride), u4);
+  _mm_store_si128((__m128i *)(output + 5 * stride), u5);
+  _mm_store_si128((__m128i *)(output + 6 * stride), u6);
+  _mm_store_si128((__m128i *)(output + 7 * stride), u7);
+}
+
+void av1_inv_txfm2d_add_8x8_sse4_1(const int32_t *coeff, uint16_t *output,
+                                   int stride, TX_TYPE tx_type, int bd) {
+  __m128i in[16], out[16];
+  const int8_t *shift = inv_txfm_shift_ls[TX_8X8];
+  const int txw_idx = get_txw_idx(TX_8X8);
+  const int txh_idx = get_txh_idx(TX_8X8);
+
+  switch (tx_type) {
+    case DCT_DCT:
+      load_buffer_8x8(coeff, in);
+      transpose_8x8(in, out);
+      idct8x8_sse4_1(out, in, inv_cos_bit_row[txw_idx][txh_idx], 0, bd,
+                     -shift[0]);
+      transpose_8x8(in, out);
+      idct8x8_sse4_1(out, in, inv_cos_bit_col[txw_idx][txh_idx], 1, bd, 0);
+      write_buffer_8x8(in, output, stride, 0, 0, -shift[1], bd);
+      break;
+    case DCT_ADST:
+      load_buffer_8x8(coeff, in);
+      transpose_8x8(in, out);
+      iadst8x8_sse4_1(out, in, inv_cos_bit_row[txw_idx][txh_idx], 0, bd,
+                      -shift[0]);
+      transpose_8x8(in, out);
+      idct8x8_sse4_1(out, in, inv_cos_bit_col[txw_idx][txh_idx], 1, bd, 0);
+      write_buffer_8x8(in, output, stride, 0, 0, -shift[1], bd);
+      break;
+    case ADST_DCT:
+      load_buffer_8x8(coeff, in);
+      transpose_8x8(in, out);
+      idct8x8_sse4_1(out, in, inv_cos_bit_row[txw_idx][txh_idx], 0, bd,
+                     -shift[0]);
+      transpose_8x8(in, out);
+      iadst8x8_sse4_1(out, in, inv_cos_bit_col[txw_idx][txh_idx], 1, bd, 0);
+      write_buffer_8x8(in, output, stride, 0, 0, -shift[1], bd);
+      break;
+    case ADST_ADST:
+      load_buffer_8x8(coeff, in);
+      transpose_8x8(in, out);
+      iadst8x8_sse4_1(out, in, inv_cos_bit_row[txw_idx][txh_idx], 0, bd,
+                      -shift[0]);
+      transpose_8x8(in, out);
+      iadst8x8_sse4_1(out, in, inv_cos_bit_col[txw_idx][txh_idx], 1, bd, 0);
+      write_buffer_8x8(in, output, stride, 0, 0, -shift[1], bd);
+      break;
+    case FLIPADST_DCT:
+      load_buffer_8x8(coeff, in);
+      transpose_8x8(in, out);
+      idct8x8_sse4_1(out, in, inv_cos_bit_row[txw_idx][txh_idx], 0, bd,
+                     -shift[0]);
+      transpose_8x8(in, out);
+      iadst8x8_sse4_1(out, in, inv_cos_bit_col[txw_idx][txh_idx], 1, bd, 0);
+      write_buffer_8x8(in, output, stride, 0, 1, -shift[1], bd);
+      break;
+    case DCT_FLIPADST:
+      load_buffer_8x8(coeff, in);
+      transpose_8x8(in, out);
+      iadst8x8_sse4_1(out, in, inv_cos_bit_row[txw_idx][txh_idx], 0, bd,
+                      -shift[0]);
+      transpose_8x8(in, out);
+      idct8x8_sse4_1(out, in, inv_cos_bit_col[txw_idx][txh_idx], 1, bd, 0);
+      write_buffer_8x8(in, output, stride, 1, 0, -shift[1], bd);
+      break;
+    case ADST_FLIPADST:
+      load_buffer_8x8(coeff, in);
+      transpose_8x8(in, out);
+      iadst8x8_sse4_1(out, in, inv_cos_bit_row[txw_idx][txh_idx], 0, bd,
+                      -shift[0]);
+      transpose_8x8(in, out);
+      iadst8x8_sse4_1(out, in, inv_cos_bit_col[txw_idx][txh_idx], 1, bd, 0);
+      write_buffer_8x8(in, output, stride, 1, 0, -shift[1], bd);
+      break;
+    case FLIPADST_FLIPADST:
+      load_buffer_8x8(coeff, in);
+      transpose_8x8(in, out);
+      iadst8x8_sse4_1(out, in, inv_cos_bit_row[txw_idx][txh_idx], 0, bd,
+                      -shift[0]);
+      transpose_8x8(in, out);
+      iadst8x8_sse4_1(out, in, inv_cos_bit_col[txw_idx][txh_idx], 1, bd, 0);
+      write_buffer_8x8(in, output, stride, 1, 1, -shift[1], bd);
+      break;
+    case FLIPADST_ADST:
+      load_buffer_8x8(coeff, in);
+      transpose_8x8(in, out);
+      iadst8x8_sse4_1(out, in, inv_cos_bit_row[txw_idx][txh_idx], 0, bd,
+                      -shift[0]);
+      transpose_8x8(in, out);
+      iadst8x8_sse4_1(out, in, inv_cos_bit_col[txw_idx][txh_idx], 1, bd, 0);
+      write_buffer_8x8(in, output, stride, 0, 1, -shift[1], bd);
+      break;
+    default: assert(0);
+  }
+}
+
+static void idct8x8_low1_sse4_1(__m128i *in, __m128i *out, int bit, int do_cols,
+                                int bd, int out_shift) {
+  const int32_t *cospi = cospi_arr(bit);
+  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
+  const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
+  const int log_range = AOMMAX(16, bd + (do_cols ? 6 : 8));
+  __m128i x;
+
+  // stage 0
+  // stage 1
+  // stage 2
+  // stage 3
+  x = _mm_mullo_epi32(in[0], cospi32);
+  x = _mm_add_epi32(x, rnding);
+  x = _mm_srai_epi32(x, bit);
+
+  // stage 4
+  // stage 5
+  if (!do_cols) {
+    const int log_range_out = AOMMAX(16, bd + 6);
+    const __m128i clamp_lo_out = _mm_set1_epi32(AOMMAX(
+        -(1 << (log_range_out - 1)), -(1 << (log_range - 1 - out_shift))));
+    const __m128i clamp_hi_out = _mm_set1_epi32(AOMMIN(
+        (1 << (log_range_out - 1)) - 1, (1 << (log_range - 1 - out_shift))));
+
+    __m128i offset = _mm_set1_epi32((1 << out_shift) >> 1);
+    x = _mm_add_epi32(x, offset);
+    x = _mm_sra_epi32(x, _mm_cvtsi32_si128(out_shift));
+    x = _mm_max_epi32(x, clamp_lo_out);
+    x = _mm_min_epi32(x, clamp_hi_out);
+  }
+
+  out[0] = x;
+  out[1] = x;
+  out[2] = x;
+  out[3] = x;
+  out[4] = x;
+  out[5] = x;
+  out[6] = x;
+  out[7] = x;
+}
+
+static void idct8x8_new_sse4_1(__m128i *in, __m128i *out, int bit, int do_cols,
+                               int bd, int out_shift) {
+  const int32_t *cospi = cospi_arr(bit);
+  const __m128i cospi56 = _mm_set1_epi32(cospi[56]);
+  const __m128i cospim8 = _mm_set1_epi32(-cospi[8]);
+  const __m128i cospi24 = _mm_set1_epi32(cospi[24]);
+  const __m128i cospim40 = _mm_set1_epi32(-cospi[40]);
+  const __m128i cospi40 = _mm_set1_epi32(cospi[40]);
+  const __m128i cospi8 = _mm_set1_epi32(cospi[8]);
+  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
+  const __m128i cospi48 = _mm_set1_epi32(cospi[48]);
+  const __m128i cospim16 = _mm_set1_epi32(-cospi[16]);
+  const __m128i cospi16 = _mm_set1_epi32(cospi[16]);
+  const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
+  const int log_range = AOMMAX(16, bd + (do_cols ? 6 : 8));
+  const __m128i clamp_lo = _mm_set1_epi32(-(1 << (log_range - 1)));
+  const __m128i clamp_hi = _mm_set1_epi32((1 << (log_range - 1)) - 1);
+  __m128i u0, u1, u2, u3, u4, u5, u6, u7;
+  __m128i v0, v1, v2, v3, v4, v5, v6, v7;
+  __m128i x, y;
+
+  // stage 0
+  // stage 1
+  // stage 2
+  u0 = in[0];
+  u1 = in[4];
+  u2 = in[2];
+  u3 = in[6];
+
+  x = _mm_mullo_epi32(in[1], cospi56);
+  y = _mm_mullo_epi32(in[7], cospim8);
+  u4 = _mm_add_epi32(x, y);
+  u4 = _mm_add_epi32(u4, rnding);
+  u4 = _mm_srai_epi32(u4, bit);
+
+  x = _mm_mullo_epi32(in[1], cospi8);
+  y = _mm_mullo_epi32(in[7], cospi56);
+  u7 = _mm_add_epi32(x, y);
+  u7 = _mm_add_epi32(u7, rnding);
+  u7 = _mm_srai_epi32(u7, bit);
+
+  x = _mm_mullo_epi32(in[5], cospi24);
+  y = _mm_mullo_epi32(in[3], cospim40);
+  u5 = _mm_add_epi32(x, y);
+  u5 = _mm_add_epi32(u5, rnding);
+  u5 = _mm_srai_epi32(u5, bit);
+
+  x = _mm_mullo_epi32(in[5], cospi40);
+  y = _mm_mullo_epi32(in[3], cospi24);
+  u6 = _mm_add_epi32(x, y);
+  u6 = _mm_add_epi32(u6, rnding);
+  u6 = _mm_srai_epi32(u6, bit);
+
+  // stage 3
+  x = _mm_mullo_epi32(u0, cospi32);
+  y = _mm_mullo_epi32(u1, cospi32);
+  v0 = _mm_add_epi32(x, y);
+  v0 = _mm_add_epi32(v0, rnding);
+  v0 = _mm_srai_epi32(v0, bit);
+
+  v1 = _mm_sub_epi32(x, y);
+  v1 = _mm_add_epi32(v1, rnding);
+  v1 = _mm_srai_epi32(v1, bit);
+
+  x = _mm_mullo_epi32(u2, cospi48);
+  y = _mm_mullo_epi32(u3, cospim16);
+  v2 = _mm_add_epi32(x, y);
+  v2 = _mm_add_epi32(v2, rnding);
+  v2 = _mm_srai_epi32(v2, bit);
+
+  x = _mm_mullo_epi32(u2, cospi16);
+  y = _mm_mullo_epi32(u3, cospi48);
+  v3 = _mm_add_epi32(x, y);
+  v3 = _mm_add_epi32(v3, rnding);
+  v3 = _mm_srai_epi32(v3, bit);
+
+  addsub_sse4_1(u4, u5, &v4, &v5, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(u7, u6, &v7, &v6, &clamp_lo, &clamp_hi);
+
+  // stage 4
+  addsub_sse4_1(v0, v3, &u0, &u3, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(v1, v2, &u1, &u2, &clamp_lo, &clamp_hi);
+  u4 = v4;
+  u7 = v7;
+
+  x = _mm_mullo_epi32(v5, cospi32);
+  y = _mm_mullo_epi32(v6, cospi32);
+  u6 = _mm_add_epi32(y, x);
+  u6 = _mm_add_epi32(u6, rnding);
+  u6 = _mm_srai_epi32(u6, bit);
+
+  u5 = _mm_sub_epi32(y, x);
+  u5 = _mm_add_epi32(u5, rnding);
+  u5 = _mm_srai_epi32(u5, bit);
+
+  // stage 5
+  if (do_cols) {
+    addsub_no_clamp_sse4_1(u0, u7, out + 0, out + 7);
+    addsub_no_clamp_sse4_1(u1, u6, out + 1, out + 6);
+    addsub_no_clamp_sse4_1(u2, u5, out + 2, out + 5);
+    addsub_no_clamp_sse4_1(u3, u4, out + 3, out + 4);
+  } else {
+    const int log_range_out = AOMMAX(16, bd + 6);
+    const __m128i clamp_lo_out = _mm_set1_epi32(AOMMAX(
+        -(1 << (log_range_out - 1)), -(1 << (log_range - 1 - out_shift))));
+    const __m128i clamp_hi_out = _mm_set1_epi32(AOMMIN(
+        (1 << (log_range_out - 1)) - 1, (1 << (log_range - 1 - out_shift))));
+    addsub_shift_sse4_1(u0, u7, out + 0, out + 7, &clamp_lo_out, &clamp_hi_out,
+                        out_shift);
+    addsub_shift_sse4_1(u1, u6, out + 1, out + 6, &clamp_lo_out, &clamp_hi_out,
+                        out_shift);
+    addsub_shift_sse4_1(u2, u5, out + 2, out + 5, &clamp_lo_out, &clamp_hi_out,
+                        out_shift);
+    addsub_shift_sse4_1(u3, u4, out + 3, out + 4, &clamp_lo_out, &clamp_hi_out,
+                        out_shift);
+  }
+}
+
+static void iadst8x8_low1_sse4_1(__m128i *in, __m128i *out, int bit,
+                                 int do_cols, int bd, int out_shift) {
+  const int32_t *cospi = cospi_arr(bit);
+  const __m128i cospi4 = _mm_set1_epi32(cospi[4]);
+  const __m128i cospi60 = _mm_set1_epi32(cospi[60]);
+  const __m128i cospi16 = _mm_set1_epi32(cospi[16]);
+  const __m128i cospi48 = _mm_set1_epi32(cospi[48]);
+  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
+  const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
+  const __m128i kZero = _mm_setzero_si128();
+  __m128i u[8], x;
+
+  // stage 0
+  // stage 1
+  // stage 2
+
+  x = _mm_mullo_epi32(in[0], cospi60);
+  u[0] = _mm_add_epi32(x, rnding);
+  u[0] = _mm_srai_epi32(u[0], bit);
+
+  x = _mm_mullo_epi32(in[0], cospi4);
+  u[1] = _mm_sub_epi32(kZero, x);
+  u[1] = _mm_add_epi32(u[1], rnding);
+  u[1] = _mm_srai_epi32(u[1], bit);
+
+  // stage 3
+  // stage 4
+  __m128i temp1, temp2;
+  temp1 = _mm_mullo_epi32(u[0], cospi16);
+  x = _mm_mullo_epi32(u[1], cospi48);
+  temp1 = _mm_add_epi32(temp1, x);
+  temp1 = _mm_add_epi32(temp1, rnding);
+  temp1 = _mm_srai_epi32(temp1, bit);
+  u[4] = temp1;
+
+  temp2 = _mm_mullo_epi32(u[0], cospi48);
+  x = _mm_mullo_epi32(u[1], cospi16);
+  u[5] = _mm_sub_epi32(temp2, x);
+  u[5] = _mm_add_epi32(u[5], rnding);
+  u[5] = _mm_srai_epi32(u[5], bit);
+
+  // stage 5
+  // stage 6
+  temp1 = _mm_mullo_epi32(u[0], cospi32);
+  x = _mm_mullo_epi32(u[1], cospi32);
+  u[2] = _mm_add_epi32(temp1, x);
+  u[2] = _mm_add_epi32(u[2], rnding);
+  u[2] = _mm_srai_epi32(u[2], bit);
+
+  u[3] = _mm_sub_epi32(temp1, x);
+  u[3] = _mm_add_epi32(u[3], rnding);
+  u[3] = _mm_srai_epi32(u[3], bit);
+
+  temp1 = _mm_mullo_epi32(u[4], cospi32);
+  x = _mm_mullo_epi32(u[5], cospi32);
+  u[6] = _mm_add_epi32(temp1, x);
+  u[6] = _mm_add_epi32(u[6], rnding);
+  u[6] = _mm_srai_epi32(u[6], bit);
+
+  u[7] = _mm_sub_epi32(temp1, x);
+  u[7] = _mm_add_epi32(u[7], rnding);
+  u[7] = _mm_srai_epi32(u[7], bit);
+
+  // stage 7
+  if (do_cols) {
+    out[0] = u[0];
+    out[1] = _mm_sub_epi32(kZero, u[4]);
+    out[2] = u[6];
+    out[3] = _mm_sub_epi32(kZero, u[2]);
+    out[4] = u[3];
+    out[5] = _mm_sub_epi32(kZero, u[7]);
+    out[6] = u[5];
+    out[7] = _mm_sub_epi32(kZero, u[1]);
+  } else {
+    const int log_range_out = AOMMAX(16, bd + 6);
+    const __m128i clamp_lo_out = _mm_set1_epi32(-(1 << (log_range_out - 1)));
+    const __m128i clamp_hi_out = _mm_set1_epi32((1 << (log_range_out - 1)) - 1);
+
+    neg_shift_sse4_1(u[0], u[4], out + 0, out + 1, &clamp_lo_out, &clamp_hi_out,
+                     out_shift);
+    neg_shift_sse4_1(u[6], u[2], out + 2, out + 3, &clamp_lo_out, &clamp_hi_out,
+                     out_shift);
+    neg_shift_sse4_1(u[3], u[7], out + 4, out + 5, &clamp_lo_out, &clamp_hi_out,
+                     out_shift);
+    neg_shift_sse4_1(u[5], u[1], out + 6, out + 7, &clamp_lo_out, &clamp_hi_out,
+                     out_shift);
+  }
+}
+
+static void iadst8x8_new_sse4_1(__m128i *in, __m128i *out, int bit, int do_cols,
+                                int bd, int out_shift) {
+  const int32_t *cospi = cospi_arr(bit);
+  const __m128i cospi4 = _mm_set1_epi32(cospi[4]);
+  const __m128i cospi60 = _mm_set1_epi32(cospi[60]);
+  const __m128i cospi20 = _mm_set1_epi32(cospi[20]);
+  const __m128i cospi44 = _mm_set1_epi32(cospi[44]);
+  const __m128i cospi36 = _mm_set1_epi32(cospi[36]);
+  const __m128i cospi28 = _mm_set1_epi32(cospi[28]);
+  const __m128i cospi52 = _mm_set1_epi32(cospi[52]);
+  const __m128i cospi12 = _mm_set1_epi32(cospi[12]);
+  const __m128i cospi16 = _mm_set1_epi32(cospi[16]);
+  const __m128i cospi48 = _mm_set1_epi32(cospi[48]);
+  const __m128i cospim48 = _mm_set1_epi32(-cospi[48]);
+  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
+  const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
+  const __m128i kZero = _mm_setzero_si128();
+  const int log_range = AOMMAX(16, bd + (do_cols ? 6 : 8));
+  const __m128i clamp_lo = _mm_set1_epi32(-(1 << (log_range - 1)));
+  const __m128i clamp_hi = _mm_set1_epi32((1 << (log_range - 1)) - 1);
+  __m128i u[8], v[8], x;
+
+  // stage 0
+  // stage 1
+  // stage 2
+
+  u[0] = _mm_mullo_epi32(in[7], cospi4);
+  x = _mm_mullo_epi32(in[0], cospi60);
+  u[0] = _mm_add_epi32(u[0], x);
+  u[0] = _mm_add_epi32(u[0], rnding);
+  u[0] = _mm_srai_epi32(u[0], bit);
+
+  u[1] = _mm_mullo_epi32(in[7], cospi60);
+  x = _mm_mullo_epi32(in[0], cospi4);
+  u[1] = _mm_sub_epi32(u[1], x);
+  u[1] = _mm_add_epi32(u[1], rnding);
+  u[1] = _mm_srai_epi32(u[1], bit);
+
+  // (2)
+  u[2] = _mm_mullo_epi32(in[5], cospi20);
+  x = _mm_mullo_epi32(in[2], cospi44);
+  u[2] = _mm_add_epi32(u[2], x);
+  u[2] = _mm_add_epi32(u[2], rnding);
+  u[2] = _mm_srai_epi32(u[2], bit);
+
+  u[3] = _mm_mullo_epi32(in[5], cospi44);
+  x = _mm_mullo_epi32(in[2], cospi20);
+  u[3] = _mm_sub_epi32(u[3], x);
+  u[3] = _mm_add_epi32(u[3], rnding);
+  u[3] = _mm_srai_epi32(u[3], bit);
+
+  // (3)
+  u[4] = _mm_mullo_epi32(in[3], cospi36);
+  x = _mm_mullo_epi32(in[4], cospi28);
+  u[4] = _mm_add_epi32(u[4], x);
+  u[4] = _mm_add_epi32(u[4], rnding);
+  u[4] = _mm_srai_epi32(u[4], bit);
+
+  u[5] = _mm_mullo_epi32(in[3], cospi28);
+  x = _mm_mullo_epi32(in[4], cospi36);
+  u[5] = _mm_sub_epi32(u[5], x);
+  u[5] = _mm_add_epi32(u[5], rnding);
+  u[5] = _mm_srai_epi32(u[5], bit);
+
+  // (4)
+  u[6] = _mm_mullo_epi32(in[1], cospi52);
+  x = _mm_mullo_epi32(in[6], cospi12);
+  u[6] = _mm_add_epi32(u[6], x);
+  u[6] = _mm_add_epi32(u[6], rnding);
+  u[6] = _mm_srai_epi32(u[6], bit);
+
+  u[7] = _mm_mullo_epi32(in[1], cospi12);
+  x = _mm_mullo_epi32(in[6], cospi52);
+  u[7] = _mm_sub_epi32(u[7], x);
+  u[7] = _mm_add_epi32(u[7], rnding);
+  u[7] = _mm_srai_epi32(u[7], bit);
+
+  // stage 3
+  addsub_sse4_1(u[0], u[4], &v[0], &v[4], &clamp_lo, &clamp_hi);
+  addsub_sse4_1(u[1], u[5], &v[1], &v[5], &clamp_lo, &clamp_hi);
+  addsub_sse4_1(u[2], u[6], &v[2], &v[6], &clamp_lo, &clamp_hi);
+  addsub_sse4_1(u[3], u[7], &v[3], &v[7], &clamp_lo, &clamp_hi);
+
+  // stage 4
+  u[0] = v[0];
+  u[1] = v[1];
+  u[2] = v[2];
+  u[3] = v[3];
+
+  u[4] = _mm_mullo_epi32(v[4], cospi16);
+  x = _mm_mullo_epi32(v[5], cospi48);
+  u[4] = _mm_add_epi32(u[4], x);
+  u[4] = _mm_add_epi32(u[4], rnding);
+  u[4] = _mm_srai_epi32(u[4], bit);
+
+  u[5] = _mm_mullo_epi32(v[4], cospi48);
+  x = _mm_mullo_epi32(v[5], cospi16);
+  u[5] = _mm_sub_epi32(u[5], x);
+  u[5] = _mm_add_epi32(u[5], rnding);
+  u[5] = _mm_srai_epi32(u[5], bit);
+
+  u[6] = _mm_mullo_epi32(v[6], cospim48);
+  x = _mm_mullo_epi32(v[7], cospi16);
+  u[6] = _mm_add_epi32(u[6], x);
+  u[6] = _mm_add_epi32(u[6], rnding);
+  u[6] = _mm_srai_epi32(u[6], bit);
+
+  u[7] = _mm_mullo_epi32(v[6], cospi16);
+  x = _mm_mullo_epi32(v[7], cospim48);
+  u[7] = _mm_sub_epi32(u[7], x);
+  u[7] = _mm_add_epi32(u[7], rnding);
+  u[7] = _mm_srai_epi32(u[7], bit);
+
+  // stage 5
+  addsub_sse4_1(u[0], u[2], &v[0], &v[2], &clamp_lo, &clamp_hi);
+  addsub_sse4_1(u[1], u[3], &v[1], &v[3], &clamp_lo, &clamp_hi);
+  addsub_sse4_1(u[4], u[6], &v[4], &v[6], &clamp_lo, &clamp_hi);
+  addsub_sse4_1(u[5], u[7], &v[5], &v[7], &clamp_lo, &clamp_hi);
+
+  // stage 6
+  u[0] = v[0];
+  u[1] = v[1];
+  u[4] = v[4];
+  u[5] = v[5];
+
+  v[0] = _mm_mullo_epi32(v[2], cospi32);
+  x = _mm_mullo_epi32(v[3], cospi32);
+  u[2] = _mm_add_epi32(v[0], x);
+  u[2] = _mm_add_epi32(u[2], rnding);
+  u[2] = _mm_srai_epi32(u[2], bit);
+
+  u[3] = _mm_sub_epi32(v[0], x);
+  u[3] = _mm_add_epi32(u[3], rnding);
+  u[3] = _mm_srai_epi32(u[3], bit);
+
+  v[0] = _mm_mullo_epi32(v[6], cospi32);
+  x = _mm_mullo_epi32(v[7], cospi32);
+  u[6] = _mm_add_epi32(v[0], x);
+  u[6] = _mm_add_epi32(u[6], rnding);
+  u[6] = _mm_srai_epi32(u[6], bit);
+
+  u[7] = _mm_sub_epi32(v[0], x);
+  u[7] = _mm_add_epi32(u[7], rnding);
+  u[7] = _mm_srai_epi32(u[7], bit);
+
+  // stage 7
+  if (do_cols) {
+    out[0] = u[0];
+    out[1] = _mm_sub_epi32(kZero, u[4]);
+    out[2] = u[6];
+    out[3] = _mm_sub_epi32(kZero, u[2]);
+    out[4] = u[3];
+    out[5] = _mm_sub_epi32(kZero, u[7]);
+    out[6] = u[5];
+    out[7] = _mm_sub_epi32(kZero, u[1]);
+  } else {
+    const int log_range_out = AOMMAX(16, bd + 6);
+    const __m128i clamp_lo_out = _mm_set1_epi32(-(1 << (log_range_out - 1)));
+    const __m128i clamp_hi_out = _mm_set1_epi32((1 << (log_range_out - 1)) - 1);
+
+    neg_shift_sse4_1(u[0], u[4], out + 0, out + 1, &clamp_lo_out, &clamp_hi_out,
+                     out_shift);
+    neg_shift_sse4_1(u[6], u[2], out + 2, out + 3, &clamp_lo_out, &clamp_hi_out,
+                     out_shift);
+    neg_shift_sse4_1(u[3], u[7], out + 4, out + 5, &clamp_lo_out, &clamp_hi_out,
+                     out_shift);
+    neg_shift_sse4_1(u[5], u[1], out + 6, out + 7, &clamp_lo_out, &clamp_hi_out,
+                     out_shift);
+  }
+}
+
+static void idct16x16_low1_sse4_1(__m128i *in, __m128i *out, int bit,
+                                  int do_cols, int bd, int out_shift) {
+  const int32_t *cospi = cospi_arr(bit);
+  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
+  const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
+  const int log_range = AOMMAX(16, bd + (do_cols ? 6 : 8));
+  const __m128i clamp_lo = _mm_set1_epi32(-(1 << (log_range - 1)));
+  const __m128i clamp_hi = _mm_set1_epi32((1 << (log_range - 1)) - 1);
+
+  {
+    // stage 0
+    // stage 1
+    // stage 2
+    // stage 3
+    // stage 4
+    in[0] = _mm_mullo_epi32(in[0], cospi32);
+    in[0] = _mm_add_epi32(in[0], rnding);
+    in[0] = _mm_srai_epi32(in[0], bit);
+
+    // stage 5
+    // stage 6
+    // stage 7
+    if (do_cols) {
+      in[0] = _mm_max_epi32(in[0], clamp_lo);
+      in[0] = _mm_min_epi32(in[0], clamp_hi);
+    } else {
+      const int log_range_out = AOMMAX(16, bd + 6);
+      const __m128i clamp_lo_out = _mm_set1_epi32(AOMMAX(
+          -(1 << (log_range_out - 1)), -(1 << (log_range - 1 - out_shift))));
+      const __m128i clamp_hi_out = _mm_set1_epi32(AOMMIN(
+          (1 << (log_range_out - 1)) - 1, (1 << (log_range - 1 - out_shift))));
+      __m128i offset = _mm_set1_epi32((1 << out_shift) >> 1);
+      in[0] = _mm_add_epi32(in[0], offset);
+      in[0] = _mm_sra_epi32(in[0], _mm_cvtsi32_si128(out_shift));
+      in[0] = _mm_max_epi32(in[0], clamp_lo_out);
+      in[0] = _mm_min_epi32(in[0], clamp_hi_out);
+    }
+
+    out[0] = in[0];
+    out[1] = in[0];
+    out[2] = in[0];
+    out[3] = in[0];
+    out[4] = in[0];
+    out[5] = in[0];
+    out[6] = in[0];
+    out[7] = in[0];
+    out[8] = in[0];
+    out[9] = in[0];
+    out[10] = in[0];
+    out[11] = in[0];
+    out[12] = in[0];
+    out[13] = in[0];
+    out[14] = in[0];
+    out[15] = in[0];
+  }
+}
+
+static void idct16x16_low8_sse4_1(__m128i *in, __m128i *out, int bit,
+                                  int do_cols, int bd, int out_shift) {
+  const int32_t *cospi = cospi_arr(bit);
+  const __m128i cospi60 = _mm_set1_epi32(cospi[60]);
+  const __m128i cospi28 = _mm_set1_epi32(cospi[28]);
+  const __m128i cospi44 = _mm_set1_epi32(cospi[44]);
+  const __m128i cospi20 = _mm_set1_epi32(cospi[20]);
+  const __m128i cospi12 = _mm_set1_epi32(cospi[12]);
+  const __m128i cospi4 = _mm_set1_epi32(cospi[4]);
+  const __m128i cospi56 = _mm_set1_epi32(cospi[56]);
+  const __m128i cospi24 = _mm_set1_epi32(cospi[24]);
+  const __m128i cospim40 = _mm_set1_epi32(-cospi[40]);
+  const __m128i cospi8 = _mm_set1_epi32(cospi[8]);
+  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
+  const __m128i cospi48 = _mm_set1_epi32(cospi[48]);
+  const __m128i cospi16 = _mm_set1_epi32(cospi[16]);
+  const __m128i cospim16 = _mm_set1_epi32(-cospi[16]);
+  const __m128i cospim48 = _mm_set1_epi32(-cospi[48]);
+  const __m128i cospim36 = _mm_set1_epi32(-cospi[36]);
+  const __m128i cospim52 = _mm_set1_epi32(-cospi[52]);
+  const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
+  const int log_range = AOMMAX(16, bd + (do_cols ? 6 : 8));
+  const __m128i clamp_lo = _mm_set1_epi32(-(1 << (log_range - 1)));
+  const __m128i clamp_hi = _mm_set1_epi32((1 << (log_range - 1)) - 1);
+  __m128i u[16], x, y;
+
+  {
+    // stage 0
+    // stage 1
+    u[0] = in[0];
+    u[2] = in[4];
+    u[4] = in[2];
+    u[6] = in[6];
+    u[8] = in[1];
+    u[10] = in[5];
+    u[12] = in[3];
+    u[14] = in[7];
+
+    // stage 2
+    u[15] = half_btf_0_sse4_1(&cospi4, &u[8], &rnding, bit);
+    u[8] = half_btf_0_sse4_1(&cospi60, &u[8], &rnding, bit);
+
+    u[9] = half_btf_0_sse4_1(&cospim36, &u[14], &rnding, bit);
+    u[14] = half_btf_0_sse4_1(&cospi28, &u[14], &rnding, bit);
+
+    u[13] = half_btf_0_sse4_1(&cospi20, &u[10], &rnding, bit);
+    u[10] = half_btf_0_sse4_1(&cospi44, &u[10], &rnding, bit);
+
+    u[11] = half_btf_0_sse4_1(&cospim52, &u[12], &rnding, bit);
+    u[12] = half_btf_0_sse4_1(&cospi12, &u[12], &rnding, bit);
+
+    // stage 3
+    u[7] = half_btf_0_sse4_1(&cospi8, &u[4], &rnding, bit);
+    u[4] = half_btf_0_sse4_1(&cospi56, &u[4], &rnding, bit);
+    u[5] = half_btf_0_sse4_1(&cospim40, &u[6], &rnding, bit);
+    u[6] = half_btf_0_sse4_1(&cospi24, &u[6], &rnding, bit);
+
+    addsub_sse4_1(u[8], u[9], &u[8], &u[9], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(u[11], u[10], &u[11], &u[10], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(u[12], u[13], &u[12], &u[13], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(u[15], u[14], &u[15], &u[14], &clamp_lo, &clamp_hi);
+
+    // stage 4
+    x = _mm_mullo_epi32(u[0], cospi32);
+    u[0] = _mm_add_epi32(x, rnding);
+    u[0] = _mm_srai_epi32(u[0], bit);
+    u[1] = u[0];
+
+    u[3] = half_btf_0_sse4_1(&cospi16, &u[2], &rnding, bit);
+    u[2] = half_btf_0_sse4_1(&cospi48, &u[2], &rnding, bit);
+
+    addsub_sse4_1(u[4], u[5], &u[4], &u[5], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(u[7], u[6], &u[7], &u[6], &clamp_lo, &clamp_hi);
+
+    x = half_btf_sse4_1(&cospim16, &u[9], &cospi48, &u[14], &rnding, bit);
+    u[14] = half_btf_sse4_1(&cospi48, &u[9], &cospi16, &u[14], &rnding, bit);
+    u[9] = x;
+    y = half_btf_sse4_1(&cospim48, &u[10], &cospim16, &u[13], &rnding, bit);
+    u[13] = half_btf_sse4_1(&cospim16, &u[10], &cospi48, &u[13], &rnding, bit);
+    u[10] = y;
+
+    // stage 5
+    addsub_sse4_1(u[0], u[3], &u[0], &u[3], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(u[1], u[2], &u[1], &u[2], &clamp_lo, &clamp_hi);
+
+    x = _mm_mullo_epi32(u[5], cospi32);
+    y = _mm_mullo_epi32(u[6], cospi32);
+    u[5] = _mm_sub_epi32(y, x);
+    u[5] = _mm_add_epi32(u[5], rnding);
+    u[5] = _mm_srai_epi32(u[5], bit);
+
+    u[6] = _mm_add_epi32(y, x);
+    u[6] = _mm_add_epi32(u[6], rnding);
+    u[6] = _mm_srai_epi32(u[6], bit);
+
+    addsub_sse4_1(u[8], u[11], &u[8], &u[11], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(u[9], u[10], &u[9], &u[10], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(u[15], u[12], &u[15], &u[12], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(u[14], u[13], &u[14], &u[13], &clamp_lo, &clamp_hi);
+
+    // stage 6
+    addsub_sse4_1(u[0], u[7], &u[0], &u[7], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(u[1], u[6], &u[1], &u[6], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(u[2], u[5], &u[2], &u[5], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(u[3], u[4], &u[3], &u[4], &clamp_lo, &clamp_hi);
+
+    x = _mm_mullo_epi32(u[10], cospi32);
+    y = _mm_mullo_epi32(u[13], cospi32);
+    u[10] = _mm_sub_epi32(y, x);
+    u[10] = _mm_add_epi32(u[10], rnding);
+    u[10] = _mm_srai_epi32(u[10], bit);
+
+    u[13] = _mm_add_epi32(x, y);
+    u[13] = _mm_add_epi32(u[13], rnding);
+    u[13] = _mm_srai_epi32(u[13], bit);
+
+    x = _mm_mullo_epi32(u[11], cospi32);
+    y = _mm_mullo_epi32(u[12], cospi32);
+    u[11] = _mm_sub_epi32(y, x);
+    u[11] = _mm_add_epi32(u[11], rnding);
+    u[11] = _mm_srai_epi32(u[11], bit);
+
+    u[12] = _mm_add_epi32(x, y);
+    u[12] = _mm_add_epi32(u[12], rnding);
+    u[12] = _mm_srai_epi32(u[12], bit);
+    // stage 7
+    if (do_cols) {
+      addsub_no_clamp_sse4_1(u[0], u[15], out + 0, out + 15);
+      addsub_no_clamp_sse4_1(u[1], u[14], out + 1, out + 14);
+      addsub_no_clamp_sse4_1(u[2], u[13], out + 2, out + 13);
+      addsub_no_clamp_sse4_1(u[3], u[12], out + 3, out + 12);
+      addsub_no_clamp_sse4_1(u[4], u[11], out + 4, out + 11);
+      addsub_no_clamp_sse4_1(u[5], u[10], out + 5, out + 10);
+      addsub_no_clamp_sse4_1(u[6], u[9], out + 6, out + 9);
+      addsub_no_clamp_sse4_1(u[7], u[8], out + 7, out + 8);
+    } else {
+      const int log_range_out = AOMMAX(16, bd + 6);
+      const __m128i clamp_lo_out = _mm_set1_epi32(AOMMAX(
+          -(1 << (log_range_out - 1)), -(1 << (log_range - 1 - out_shift))));
+      const __m128i clamp_hi_out = _mm_set1_epi32(AOMMIN(
+          (1 << (log_range_out - 1)) - 1, (1 << (log_range - 1 - out_shift))));
+
+      addsub_shift_sse4_1(u[0], u[15], out + 0, out + 15, &clamp_lo_out,
+                          &clamp_hi_out, out_shift);
+      addsub_shift_sse4_1(u[1], u[14], out + 1, out + 14, &clamp_lo_out,
+                          &clamp_hi_out, out_shift);
+      addsub_shift_sse4_1(u[2], u[13], out + 2, out + 13, &clamp_lo_out,
+                          &clamp_hi_out, out_shift);
+      addsub_shift_sse4_1(u[3], u[12], out + 3, out + 12, &clamp_lo_out,
+                          &clamp_hi_out, out_shift);
+      addsub_shift_sse4_1(u[4], u[11], out + 4, out + 11, &clamp_lo_out,
+                          &clamp_hi_out, out_shift);
+      addsub_shift_sse4_1(u[5], u[10], out + 5, out + 10, &clamp_lo_out,
+                          &clamp_hi_out, out_shift);
+      addsub_shift_sse4_1(u[6], u[9], out + 6, out + 9, &clamp_lo_out,
+                          &clamp_hi_out, out_shift);
+      addsub_shift_sse4_1(u[7], u[8], out + 7, out + 8, &clamp_lo_out,
+                          &clamp_hi_out, out_shift);
+    }
+  }
+}
+
+static void iadst16x16_low1_sse4_1(__m128i *in, __m128i *out, int bit,
+                                   int do_cols, int bd, int out_shift) {
+  const int32_t *cospi = cospi_arr(bit);
+  const __m128i cospi2 = _mm_set1_epi32(cospi[2]);
+  const __m128i cospi62 = _mm_set1_epi32(cospi[62]);
+  const __m128i cospi8 = _mm_set1_epi32(cospi[8]);
+  const __m128i cospi56 = _mm_set1_epi32(cospi[56]);
+  const __m128i cospi48 = _mm_set1_epi32(cospi[48]);
+  const __m128i cospi16 = _mm_set1_epi32(cospi[16]);
+  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
+  const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
+  const __m128i zero = _mm_setzero_si128();
+  __m128i v[16], x, y, temp1, temp2;
+
+  // Calculate the column 0, 1, 2, 3
+  {
+    // stage 0
+    // stage 1
+    // stage 2
+    x = _mm_mullo_epi32(in[0], cospi62);
+    v[0] = _mm_add_epi32(x, rnding);
+    v[0] = _mm_srai_epi32(v[0], bit);
+
+    x = _mm_mullo_epi32(in[0], cospi2);
+    v[1] = _mm_sub_epi32(zero, x);
+    v[1] = _mm_add_epi32(v[1], rnding);
+    v[1] = _mm_srai_epi32(v[1], bit);
+
+    // stage 3
+    v[8] = v[0];
+    v[9] = v[1];
+
+    // stage 4
+    temp1 = _mm_mullo_epi32(v[8], cospi8);
+    x = _mm_mullo_epi32(v[9], cospi56);
+    temp1 = _mm_add_epi32(temp1, x);
+    temp1 = _mm_add_epi32(temp1, rnding);
+    temp1 = _mm_srai_epi32(temp1, bit);
+
+    temp2 = _mm_mullo_epi32(v[8], cospi56);
+    x = _mm_mullo_epi32(v[9], cospi8);
+    temp2 = _mm_sub_epi32(temp2, x);
+    temp2 = _mm_add_epi32(temp2, rnding);
+    temp2 = _mm_srai_epi32(temp2, bit);
+    v[8] = temp1;
+    v[9] = temp2;
+
+    // stage 5
+    v[4] = v[0];
+    v[5] = v[1];
+    v[12] = v[8];
+    v[13] = v[9];
+
+    // stage 6
+    temp1 = _mm_mullo_epi32(v[4], cospi16);
+    x = _mm_mullo_epi32(v[5], cospi48);
+    temp1 = _mm_add_epi32(temp1, x);
+    temp1 = _mm_add_epi32(temp1, rnding);
+    temp1 = _mm_srai_epi32(temp1, bit);
+
+    temp2 = _mm_mullo_epi32(v[4], cospi48);
+    x = _mm_mullo_epi32(v[5], cospi16);
+    temp2 = _mm_sub_epi32(temp2, x);
+    temp2 = _mm_add_epi32(temp2, rnding);
+    temp2 = _mm_srai_epi32(temp2, bit);
+    v[4] = temp1;
+    v[5] = temp2;
+
+    temp1 = _mm_mullo_epi32(v[12], cospi16);
+    x = _mm_mullo_epi32(v[13], cospi48);
+    temp1 = _mm_add_epi32(temp1, x);
+    temp1 = _mm_add_epi32(temp1, rnding);
+    temp1 = _mm_srai_epi32(temp1, bit);
+
+    temp2 = _mm_mullo_epi32(v[12], cospi48);
+    x = _mm_mullo_epi32(v[13], cospi16);
+    temp2 = _mm_sub_epi32(temp2, x);
+    temp2 = _mm_add_epi32(temp2, rnding);
+    temp2 = _mm_srai_epi32(temp2, bit);
+    v[12] = temp1;
+    v[13] = temp2;
+
+    // stage 7
+    v[2] = v[0];
+    v[3] = v[1];
+    v[6] = v[4];
+    v[7] = v[5];
+    v[10] = v[8];
+    v[11] = v[9];
+    v[14] = v[12];
+    v[15] = v[13];
+
+    // stage 8
+    y = _mm_mullo_epi32(v[2], cospi32);
+    x = _mm_mullo_epi32(v[3], cospi32);
+    v[2] = _mm_add_epi32(y, x);
+    v[2] = _mm_add_epi32(v[2], rnding);
+    v[2] = _mm_srai_epi32(v[2], bit);
+
+    v[3] = _mm_sub_epi32(y, x);
+    v[3] = _mm_add_epi32(v[3], rnding);
+    v[3] = _mm_srai_epi32(v[3], bit);
+
+    y = _mm_mullo_epi32(v[6], cospi32);
+    x = _mm_mullo_epi32(v[7], cospi32);
+    v[6] = _mm_add_epi32(y, x);
+    v[6] = _mm_add_epi32(v[6], rnding);
+    v[6] = _mm_srai_epi32(v[6], bit);
+
+    v[7] = _mm_sub_epi32(y, x);
+    v[7] = _mm_add_epi32(v[7], rnding);
+    v[7] = _mm_srai_epi32(v[7], bit);
+
+    y = _mm_mullo_epi32(v[10], cospi32);
+    x = _mm_mullo_epi32(v[11], cospi32);
+    v[10] = _mm_add_epi32(y, x);
+    v[10] = _mm_add_epi32(v[10], rnding);
+    v[10] = _mm_srai_epi32(v[10], bit);
+
+    v[11] = _mm_sub_epi32(y, x);
+    v[11] = _mm_add_epi32(v[11], rnding);
+    v[11] = _mm_srai_epi32(v[11], bit);
+
+    y = _mm_mullo_epi32(v[14], cospi32);
+    x = _mm_mullo_epi32(v[15], cospi32);
+    v[14] = _mm_add_epi32(y, x);
+    v[14] = _mm_add_epi32(v[14], rnding);
+    v[14] = _mm_srai_epi32(v[14], bit);
+
+    v[15] = _mm_sub_epi32(y, x);
+    v[15] = _mm_add_epi32(v[15], rnding);
+    v[15] = _mm_srai_epi32(v[15], bit);
+
+    // stage 9
+    if (do_cols) {
+      out[0] = v[0];
+      out[1] = _mm_sub_epi32(_mm_setzero_si128(), v[8]);
+      out[2] = v[12];
+      out[3] = _mm_sub_epi32(_mm_setzero_si128(), v[4]);
+      out[4] = v[6];
+      out[5] = _mm_sub_epi32(_mm_setzero_si128(), v[14]);
+      out[6] = v[10];
+      out[7] = _mm_sub_epi32(_mm_setzero_si128(), v[2]);
+      out[8] = v[3];
+      out[9] = _mm_sub_epi32(_mm_setzero_si128(), v[11]);
+      out[10] = v[15];
+      out[11] = _mm_sub_epi32(_mm_setzero_si128(), v[7]);
+      out[12] = v[5];
+      out[13] = _mm_sub_epi32(_mm_setzero_si128(), v[13]);
+      out[14] = v[9];
+      out[15] = _mm_sub_epi32(_mm_setzero_si128(), v[1]);
+    } else {
+      const int log_range_out = AOMMAX(16, bd + 6);
+      const __m128i clamp_lo_out = _mm_set1_epi32(-(1 << (log_range_out - 1)));
+      const __m128i clamp_hi_out =
+          _mm_set1_epi32((1 << (log_range_out - 1)) - 1);
+
+      neg_shift_sse4_1(v[0], v[8], out + 0, out + 1, &clamp_lo_out,
+                       &clamp_hi_out, out_shift);
+      neg_shift_sse4_1(v[12], v[4], out + 2, out + 3, &clamp_lo_out,
+                       &clamp_hi_out, out_shift);
+      neg_shift_sse4_1(v[6], v[14], out + 4, out + 5, &clamp_lo_out,
+                       &clamp_hi_out, out_shift);
+      neg_shift_sse4_1(v[10], v[2], out + 6, out + 7, &clamp_lo_out,
+                       &clamp_hi_out, out_shift);
+      neg_shift_sse4_1(v[3], v[11], out + 8, out + 9, &clamp_lo_out,
+                       &clamp_hi_out, out_shift);
+      neg_shift_sse4_1(v[15], v[7], out + 10, out + 11, &clamp_lo_out,
+                       &clamp_hi_out, out_shift);
+      neg_shift_sse4_1(v[5], v[13], out + 12, out + 13, &clamp_lo_out,
+                       &clamp_hi_out, out_shift);
+      neg_shift_sse4_1(v[9], v[1], out + 14, out + 15, &clamp_lo_out,
+                       &clamp_hi_out, out_shift);
+    }
+  }
+}
+
+static void iadst16x16_low8_sse4_1(__m128i *in, __m128i *out, int bit,
+                                   int do_cols, int bd, int out_shift) {
+  const int32_t *cospi = cospi_arr(bit);
+  const __m128i cospi2 = _mm_set1_epi32(cospi[2]);
+  const __m128i cospi62 = _mm_set1_epi32(cospi[62]);
+  const __m128i cospi10 = _mm_set1_epi32(cospi[10]);
+  const __m128i cospi54 = _mm_set1_epi32(cospi[54]);
+  const __m128i cospi18 = _mm_set1_epi32(cospi[18]);
+  const __m128i cospi46 = _mm_set1_epi32(cospi[46]);
+  const __m128i cospi26 = _mm_set1_epi32(cospi[26]);
+  const __m128i cospi38 = _mm_set1_epi32(cospi[38]);
+  const __m128i cospi34 = _mm_set1_epi32(cospi[34]);
+  const __m128i cospi30 = _mm_set1_epi32(cospi[30]);
+  const __m128i cospi42 = _mm_set1_epi32(cospi[42]);
+  const __m128i cospi22 = _mm_set1_epi32(cospi[22]);
+  const __m128i cospi50 = _mm_set1_epi32(cospi[50]);
+  const __m128i cospi14 = _mm_set1_epi32(cospi[14]);
+  const __m128i cospi58 = _mm_set1_epi32(cospi[58]);
+  const __m128i cospi6 = _mm_set1_epi32(cospi[6]);
+  const __m128i cospi8 = _mm_set1_epi32(cospi[8]);
+  const __m128i cospi56 = _mm_set1_epi32(cospi[56]);
+  const __m128i cospi40 = _mm_set1_epi32(cospi[40]);
+  const __m128i cospi24 = _mm_set1_epi32(cospi[24]);
+  const __m128i cospim56 = _mm_set1_epi32(-cospi[56]);
+  const __m128i cospim24 = _mm_set1_epi32(-cospi[24]);
+  const __m128i cospi48 = _mm_set1_epi32(cospi[48]);
+  const __m128i cospi16 = _mm_set1_epi32(cospi[16]);
+  const __m128i cospim48 = _mm_set1_epi32(-cospi[48]);
+  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
+  const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
+  const int log_range = AOMMAX(16, bd + (do_cols ? 6 : 8));
+  const __m128i clamp_lo = _mm_set1_epi32(-(1 << (log_range - 1)));
+  const __m128i clamp_hi = _mm_set1_epi32((1 << (log_range - 1)) - 1);
+  __m128i u[16], x, y;
+
+  // Calculate the column 0, 1, 2, 3
+  {
+    // stage 0
+    // stage 1
+    // stage 2
+    __m128i zero = _mm_setzero_si128();
+    x = _mm_mullo_epi32(in[0], cospi62);
+    u[0] = _mm_add_epi32(x, rnding);
+    u[0] = _mm_srai_epi32(u[0], bit);
+
+    x = _mm_mullo_epi32(in[0], cospi2);
+    u[1] = _mm_sub_epi32(zero, x);
+    u[1] = _mm_add_epi32(u[1], rnding);
+    u[1] = _mm_srai_epi32(u[1], bit);
+
+    x = _mm_mullo_epi32(in[2], cospi54);
+    u[2] = _mm_add_epi32(x, rnding);
+    u[2] = _mm_srai_epi32(u[2], bit);
+
+    x = _mm_mullo_epi32(in[2], cospi10);
+    u[3] = _mm_sub_epi32(zero, x);
+    u[3] = _mm_add_epi32(u[3], rnding);
+    u[3] = _mm_srai_epi32(u[3], bit);
+
+    x = _mm_mullo_epi32(in[4], cospi46);
+    u[4] = _mm_add_epi32(x, rnding);
+    u[4] = _mm_srai_epi32(u[4], bit);
+
+    x = _mm_mullo_epi32(in[4], cospi18);
+    u[5] = _mm_sub_epi32(zero, x);
+    u[5] = _mm_add_epi32(u[5], rnding);
+    u[5] = _mm_srai_epi32(u[5], bit);
+
+    x = _mm_mullo_epi32(in[6], cospi38);
+    u[6] = _mm_add_epi32(x, rnding);
+    u[6] = _mm_srai_epi32(u[6], bit);
+
+    x = _mm_mullo_epi32(in[6], cospi26);
+    u[7] = _mm_sub_epi32(zero, x);
+    u[7] = _mm_add_epi32(u[7], rnding);
+    u[7] = _mm_srai_epi32(u[7], bit);
+
+    u[8] = _mm_mullo_epi32(in[7], cospi34);
+    u[8] = _mm_add_epi32(u[8], rnding);
+    u[8] = _mm_srai_epi32(u[8], bit);
+
+    u[9] = _mm_mullo_epi32(in[7], cospi30);
+    u[9] = _mm_add_epi32(u[9], rnding);
+    u[9] = _mm_srai_epi32(u[9], bit);
+
+    u[10] = _mm_mullo_epi32(in[5], cospi42);
+    u[10] = _mm_add_epi32(u[10], rnding);
+    u[10] = _mm_srai_epi32(u[10], bit);
+
+    u[11] = _mm_mullo_epi32(in[5], cospi22);
+    u[11] = _mm_add_epi32(u[11], rnding);
+    u[11] = _mm_srai_epi32(u[11], bit);
+
+    u[12] = _mm_mullo_epi32(in[3], cospi50);
+    u[12] = _mm_add_epi32(u[12], rnding);
+    u[12] = _mm_srai_epi32(u[12], bit);
+
+    u[13] = _mm_mullo_epi32(in[3], cospi14);
+    u[13] = _mm_add_epi32(u[13], rnding);
+    u[13] = _mm_srai_epi32(u[13], bit);
+
+    u[14] = _mm_mullo_epi32(in[1], cospi58);
+    u[14] = _mm_add_epi32(u[14], rnding);
+    u[14] = _mm_srai_epi32(u[14], bit);
+
+    u[15] = _mm_mullo_epi32(in[1], cospi6);
+    u[15] = _mm_add_epi32(u[15], rnding);
+    u[15] = _mm_srai_epi32(u[15], bit);
+
+    // stage 3
+    addsub_sse4_1(u[0], u[8], &u[0], &u[8], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(u[1], u[9], &u[1], &u[9], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(u[2], u[10], &u[2], &u[10], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(u[3], u[11], &u[3], &u[11], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(u[4], u[12], &u[4], &u[12], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(u[5], u[13], &u[5], &u[13], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(u[6], u[14], &u[6], &u[14], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(u[7], u[15], &u[7], &u[15], &clamp_lo, &clamp_hi);
+
+    // stage 4
+    y = _mm_mullo_epi32(u[8], cospi56);
+    x = _mm_mullo_epi32(u[9], cospi56);
+    u[8] = _mm_mullo_epi32(u[8], cospi8);
+    u[8] = _mm_add_epi32(u[8], x);
+    u[8] = _mm_add_epi32(u[8], rnding);
+    u[8] = _mm_srai_epi32(u[8], bit);
+
+    x = _mm_mullo_epi32(u[9], cospi8);
+    u[9] = _mm_sub_epi32(y, x);
+    u[9] = _mm_add_epi32(u[9], rnding);
+    u[9] = _mm_srai_epi32(u[9], bit);
+
+    x = _mm_mullo_epi32(u[11], cospi24);
+    y = _mm_mullo_epi32(u[10], cospi24);
+    u[10] = _mm_mullo_epi32(u[10], cospi40);
+    u[10] = _mm_add_epi32(u[10], x);
+    u[10] = _mm_add_epi32(u[10], rnding);
+    u[10] = _mm_srai_epi32(u[10], bit);
+
+    x = _mm_mullo_epi32(u[11], cospi40);
+    u[11] = _mm_sub_epi32(y, x);
+    u[11] = _mm_add_epi32(u[11], rnding);
+    u[11] = _mm_srai_epi32(u[11], bit);
+
+    x = _mm_mullo_epi32(u[13], cospi8);
+    y = _mm_mullo_epi32(u[12], cospi8);
+    u[12] = _mm_mullo_epi32(u[12], cospim56);
+    u[12] = _mm_add_epi32(u[12], x);
+    u[12] = _mm_add_epi32(u[12], rnding);
+    u[12] = _mm_srai_epi32(u[12], bit);
+
+    x = _mm_mullo_epi32(u[13], cospim56);
+    u[13] = _mm_sub_epi32(y, x);
+    u[13] = _mm_add_epi32(u[13], rnding);
+    u[13] = _mm_srai_epi32(u[13], bit);
+
+    x = _mm_mullo_epi32(u[15], cospi40);
+    y = _mm_mullo_epi32(u[14], cospi40);
+    u[14] = _mm_mullo_epi32(u[14], cospim24);
+    u[14] = _mm_add_epi32(u[14], x);
+    u[14] = _mm_add_epi32(u[14], rnding);
+    u[14] = _mm_srai_epi32(u[14], bit);
+
+    x = _mm_mullo_epi32(u[15], cospim24);
+    u[15] = _mm_sub_epi32(y, x);
+    u[15] = _mm_add_epi32(u[15], rnding);
+    u[15] = _mm_srai_epi32(u[15], bit);
+
+    // stage 5
+    addsub_sse4_1(u[0], u[4], &u[0], &u[4], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(u[1], u[5], &u[1], &u[5], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(u[2], u[6], &u[2], &u[6], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(u[3], u[7], &u[3], &u[7], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(u[8], u[12], &u[8], &u[12], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(u[9], u[13], &u[9], &u[13], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(u[10], u[14], &u[10], &u[14], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(u[11], u[15], &u[11], &u[15], &clamp_lo, &clamp_hi);
+
+    // stage 6
+    x = _mm_mullo_epi32(u[5], cospi48);
+    y = _mm_mullo_epi32(u[4], cospi48);
+    u[4] = _mm_mullo_epi32(u[4], cospi16);
+    u[4] = _mm_add_epi32(u[4], x);
+    u[4] = _mm_add_epi32(u[4], rnding);
+    u[4] = _mm_srai_epi32(u[4], bit);
+
+    x = _mm_mullo_epi32(u[5], cospi16);
+    u[5] = _mm_sub_epi32(y, x);
+    u[5] = _mm_add_epi32(u[5], rnding);
+    u[5] = _mm_srai_epi32(u[5], bit);
+
+    x = _mm_mullo_epi32(u[7], cospi16);
+    y = _mm_mullo_epi32(u[6], cospi16);
+    u[6] = _mm_mullo_epi32(u[6], cospim48);
+    u[6] = _mm_add_epi32(u[6], x);
+    u[6] = _mm_add_epi32(u[6], rnding);
+    u[6] = _mm_srai_epi32(u[6], bit);
+
+    x = _mm_mullo_epi32(u[7], cospim48);
+    u[7] = _mm_sub_epi32(y, x);
+    u[7] = _mm_add_epi32(u[7], rnding);
+    u[7] = _mm_srai_epi32(u[7], bit);
+
+    x = _mm_mullo_epi32(u[13], cospi48);
+    y = _mm_mullo_epi32(u[12], cospi48);
+    u[12] = _mm_mullo_epi32(u[12], cospi16);
+    u[12] = _mm_add_epi32(u[12], x);
+    u[12] = _mm_add_epi32(u[12], rnding);
+    u[12] = _mm_srai_epi32(u[12], bit);
+
+    x = _mm_mullo_epi32(u[13], cospi16);
+    u[13] = _mm_sub_epi32(y, x);
+    u[13] = _mm_add_epi32(u[13], rnding);
+    u[13] = _mm_srai_epi32(u[13], bit);
+
+    x = _mm_mullo_epi32(u[15], cospi16);
+    y = _mm_mullo_epi32(u[14], cospi16);
+    u[14] = _mm_mullo_epi32(u[14], cospim48);
+    u[14] = _mm_add_epi32(u[14], x);
+    u[14] = _mm_add_epi32(u[14], rnding);
+    u[14] = _mm_srai_epi32(u[14], bit);
+
+    x = _mm_mullo_epi32(u[15], cospim48);
+    u[15] = _mm_sub_epi32(y, x);
+    u[15] = _mm_add_epi32(u[15], rnding);
+    u[15] = _mm_srai_epi32(u[15], bit);
+
+    // stage 7
+    addsub_sse4_1(u[0], u[2], &u[0], &u[2], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(u[1], u[3], &u[1], &u[3], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(u[4], u[6], &u[4], &u[6], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(u[5], u[7], &u[5], &u[7], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(u[8], u[10], &u[8], &u[10], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(u[9], u[11], &u[9], &u[11], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(u[12], u[14], &u[12], &u[14], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(u[13], u[15], &u[13], &u[15], &clamp_lo, &clamp_hi);
+
+    // stage 8
+    y = _mm_mullo_epi32(u[2], cospi32);
+    x = _mm_mullo_epi32(u[3], cospi32);
+    u[2] = _mm_add_epi32(y, x);
+    u[2] = _mm_add_epi32(u[2], rnding);
+    u[2] = _mm_srai_epi32(u[2], bit);
+
+    u[3] = _mm_sub_epi32(y, x);
+    u[3] = _mm_add_epi32(u[3], rnding);
+    u[3] = _mm_srai_epi32(u[3], bit);
+    y = _mm_mullo_epi32(u[6], cospi32);
+    x = _mm_mullo_epi32(u[7], cospi32);
+    u[6] = _mm_add_epi32(y, x);
+    u[6] = _mm_add_epi32(u[6], rnding);
+    u[6] = _mm_srai_epi32(u[6], bit);
+
+    u[7] = _mm_sub_epi32(y, x);
+    u[7] = _mm_add_epi32(u[7], rnding);
+    u[7] = _mm_srai_epi32(u[7], bit);
+
+    y = _mm_mullo_epi32(u[10], cospi32);
+    x = _mm_mullo_epi32(u[11], cospi32);
+    u[10] = _mm_add_epi32(y, x);
+    u[10] = _mm_add_epi32(u[10], rnding);
+    u[10] = _mm_srai_epi32(u[10], bit);
+
+    u[11] = _mm_sub_epi32(y, x);
+    u[11] = _mm_add_epi32(u[11], rnding);
+    u[11] = _mm_srai_epi32(u[11], bit);
+
+    y = _mm_mullo_epi32(u[14], cospi32);
+    x = _mm_mullo_epi32(u[15], cospi32);
+    u[14] = _mm_add_epi32(y, x);
+    u[14] = _mm_add_epi32(u[14], rnding);
+    u[14] = _mm_srai_epi32(u[14], bit);
+
+    u[15] = _mm_sub_epi32(y, x);
+    u[15] = _mm_add_epi32(u[15], rnding);
+    u[15] = _mm_srai_epi32(u[15], bit);
+
+    // stage 9
+    if (do_cols) {
+      out[0] = u[0];
+      out[1] = _mm_sub_epi32(_mm_setzero_si128(), u[8]);
+      out[2] = u[12];
+      out[3] = _mm_sub_epi32(_mm_setzero_si128(), u[4]);
+      out[4] = u[6];
+      out[5] = _mm_sub_epi32(_mm_setzero_si128(), u[14]);
+      out[6] = u[10];
+      out[7] = _mm_sub_epi32(_mm_setzero_si128(), u[2]);
+      out[8] = u[3];
+      out[9] = _mm_sub_epi32(_mm_setzero_si128(), u[11]);
+      out[10] = u[15];
+      out[11] = _mm_sub_epi32(_mm_setzero_si128(), u[7]);
+      out[12] = u[5];
+      out[13] = _mm_sub_epi32(_mm_setzero_si128(), u[13]);
+      out[14] = u[9];
+      out[15] = _mm_sub_epi32(_mm_setzero_si128(), u[1]);
+    } else {
+      const int log_range_out = AOMMAX(16, bd + 6);
+      const __m128i clamp_lo_out = _mm_set1_epi32(-(1 << (log_range_out - 1)));
+      const __m128i clamp_hi_out =
+          _mm_set1_epi32((1 << (log_range_out - 1)) - 1);
+
+      neg_shift_sse4_1(u[0], u[8], out + 0, out + 1, &clamp_lo_out,
+                       &clamp_hi_out, out_shift);
+      neg_shift_sse4_1(u[12], u[4], out + 2, out + 3, &clamp_lo_out,
+                       &clamp_hi_out, out_shift);
+      neg_shift_sse4_1(u[6], u[14], out + 4, out + 5, &clamp_lo_out,
+                       &clamp_hi_out, out_shift);
+      neg_shift_sse4_1(u[10], u[2], out + 6, out + 7, &clamp_lo_out,
+                       &clamp_hi_out, out_shift);
+      neg_shift_sse4_1(u[3], u[11], out + 8, out + 9, &clamp_lo_out,
+                       &clamp_hi_out, out_shift);
+      neg_shift_sse4_1(u[15], u[7], out + 10, out + 11, &clamp_lo_out,
+                       &clamp_hi_out, out_shift);
+      neg_shift_sse4_1(u[5], u[13], out + 12, out + 13, &clamp_lo_out,
+                       &clamp_hi_out, out_shift);
+      neg_shift_sse4_1(u[9], u[1], out + 14, out + 15, &clamp_lo_out,
+                       &clamp_hi_out, out_shift);
+    }
+  }
+}
+
+static void idct16x16_sse4_1(__m128i *in, __m128i *out, int bit, int do_cols,
+                             int bd, int out_shift) {
+  const int32_t *cospi = cospi_arr(bit);
+  const __m128i cospi60 = _mm_set1_epi32(cospi[60]);
+  const __m128i cospim4 = _mm_set1_epi32(-cospi[4]);
+  const __m128i cospi28 = _mm_set1_epi32(cospi[28]);
+  const __m128i cospim36 = _mm_set1_epi32(-cospi[36]);
+  const __m128i cospi44 = _mm_set1_epi32(cospi[44]);
+  const __m128i cospi20 = _mm_set1_epi32(cospi[20]);
+  const __m128i cospim20 = _mm_set1_epi32(-cospi[20]);
+  const __m128i cospi12 = _mm_set1_epi32(cospi[12]);
+  const __m128i cospim52 = _mm_set1_epi32(-cospi[52]);
+  const __m128i cospi52 = _mm_set1_epi32(cospi[52]);
+  const __m128i cospi36 = _mm_set1_epi32(cospi[36]);
+  const __m128i cospi4 = _mm_set1_epi32(cospi[4]);
+  const __m128i cospi56 = _mm_set1_epi32(cospi[56]);
+  const __m128i cospim8 = _mm_set1_epi32(-cospi[8]);
+  const __m128i cospi24 = _mm_set1_epi32(cospi[24]);
+  const __m128i cospim40 = _mm_set1_epi32(-cospi[40]);
+  const __m128i cospi40 = _mm_set1_epi32(cospi[40]);
+  const __m128i cospi8 = _mm_set1_epi32(cospi[8]);
+  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
+  const __m128i cospi48 = _mm_set1_epi32(cospi[48]);
+  const __m128i cospi16 = _mm_set1_epi32(cospi[16]);
+  const __m128i cospim16 = _mm_set1_epi32(-cospi[16]);
+  const __m128i cospim48 = _mm_set1_epi32(-cospi[48]);
+  const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
+  const int log_range = AOMMAX(16, bd + (do_cols ? 6 : 8));
+  const __m128i clamp_lo = _mm_set1_epi32(-(1 << (log_range - 1)));
+  const __m128i clamp_hi = _mm_set1_epi32((1 << (log_range - 1)) - 1);
+  __m128i u[16], v[16], x, y;
+
+  {
+    // stage 0
+    // stage 1
+    u[0] = in[0];
+    u[1] = in[8];
+    u[2] = in[4];
+    u[3] = in[12];
+    u[4] = in[2];
+    u[5] = in[10];
+    u[6] = in[6];
+    u[7] = in[14];
+    u[8] = in[1];
+    u[9] = in[9];
+    u[10] = in[5];
+    u[11] = in[13];
+    u[12] = in[3];
+    u[13] = in[11];
+    u[14] = in[7];
+    u[15] = in[15];
+
+    // stage 2
+    v[0] = u[0];
+    v[1] = u[1];
+    v[2] = u[2];
+    v[3] = u[3];
+    v[4] = u[4];
+    v[5] = u[5];
+    v[6] = u[6];
+    v[7] = u[7];
+
+    v[8] = half_btf_sse4_1(&cospi60, &u[8], &cospim4, &u[15], &rnding, bit);
+    v[9] = half_btf_sse4_1(&cospi28, &u[9], &cospim36, &u[14], &rnding, bit);
+    v[10] = half_btf_sse4_1(&cospi44, &u[10], &cospim20, &u[13], &rnding, bit);
+    v[11] = half_btf_sse4_1(&cospi12, &u[11], &cospim52, &u[12], &rnding, bit);
+    v[12] = half_btf_sse4_1(&cospi52, &u[11], &cospi12, &u[12], &rnding, bit);
+    v[13] = half_btf_sse4_1(&cospi20, &u[10], &cospi44, &u[13], &rnding, bit);
+    v[14] = half_btf_sse4_1(&cospi36, &u[9], &cospi28, &u[14], &rnding, bit);
+    v[15] = half_btf_sse4_1(&cospi4, &u[8], &cospi60, &u[15], &rnding, bit);
+
+    // stage 3
+    u[0] = v[0];
+    u[1] = v[1];
+    u[2] = v[2];
+    u[3] = v[3];
+    u[4] = half_btf_sse4_1(&cospi56, &v[4], &cospim8, &v[7], &rnding, bit);
+    u[5] = half_btf_sse4_1(&cospi24, &v[5], &cospim40, &v[6], &rnding, bit);
+    u[6] = half_btf_sse4_1(&cospi40, &v[5], &cospi24, &v[6], &rnding, bit);
+    u[7] = half_btf_sse4_1(&cospi8, &v[4], &cospi56, &v[7], &rnding, bit);
+    addsub_sse4_1(v[8], v[9], &u[8], &u[9], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(v[11], v[10], &u[11], &u[10], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(v[12], v[13], &u[12], &u[13], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(v[15], v[14], &u[15], &u[14], &clamp_lo, &clamp_hi);
+
+    // stage 4
+    x = _mm_mullo_epi32(u[0], cospi32);
+    y = _mm_mullo_epi32(u[1], cospi32);
+    v[0] = _mm_add_epi32(x, y);
+    v[0] = _mm_add_epi32(v[0], rnding);
+    v[0] = _mm_srai_epi32(v[0], bit);
+
+    v[1] = _mm_sub_epi32(x, y);
+    v[1] = _mm_add_epi32(v[1], rnding);
+    v[1] = _mm_srai_epi32(v[1], bit);
+
+    v[2] = half_btf_sse4_1(&cospi48, &u[2], &cospim16, &u[3], &rnding, bit);
+    v[3] = half_btf_sse4_1(&cospi16, &u[2], &cospi48, &u[3], &rnding, bit);
+    addsub_sse4_1(u[4], u[5], &v[4], &v[5], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(u[7], u[6], &v[7], &v[6], &clamp_lo, &clamp_hi);
+    v[8] = u[8];
+    v[9] = half_btf_sse4_1(&cospim16, &u[9], &cospi48, &u[14], &rnding, bit);
+    v[10] = half_btf_sse4_1(&cospim48, &u[10], &cospim16, &u[13], &rnding, bit);
+    v[11] = u[11];
+    v[12] = u[12];
+    v[13] = half_btf_sse4_1(&cospim16, &u[10], &cospi48, &u[13], &rnding, bit);
+    v[14] = half_btf_sse4_1(&cospi48, &u[9], &cospi16, &u[14], &rnding, bit);
+    v[15] = u[15];
+
+    // stage 5
+    addsub_sse4_1(v[0], v[3], &u[0], &u[3], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(v[1], v[2], &u[1], &u[2], &clamp_lo, &clamp_hi);
+    u[4] = v[4];
+
+    x = _mm_mullo_epi32(v[5], cospi32);
+    y = _mm_mullo_epi32(v[6], cospi32);
+    u[5] = _mm_sub_epi32(y, x);
+    u[5] = _mm_add_epi32(u[5], rnding);
+    u[5] = _mm_srai_epi32(u[5], bit);
+
+    u[6] = _mm_add_epi32(y, x);
+    u[6] = _mm_add_epi32(u[6], rnding);
+    u[6] = _mm_srai_epi32(u[6], bit);
+
+    u[7] = v[7];
+    addsub_sse4_1(v[8], v[11], &u[8], &u[11], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(v[9], v[10], &u[9], &u[10], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(v[15], v[12], &u[15], &u[12], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(v[14], v[13], &u[14], &u[13], &clamp_lo, &clamp_hi);
+
+    // stage 6
+    addsub_sse4_1(u[0], u[7], &v[0], &v[7], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(u[1], u[6], &v[1], &v[6], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(u[2], u[5], &v[2], &v[5], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(u[3], u[4], &v[3], &v[4], &clamp_lo, &clamp_hi);
+    v[8] = u[8];
+    v[9] = u[9];
+
+    x = _mm_mullo_epi32(u[10], cospi32);
+    y = _mm_mullo_epi32(u[13], cospi32);
+    v[10] = _mm_sub_epi32(y, x);
+    v[10] = _mm_add_epi32(v[10], rnding);
+    v[10] = _mm_srai_epi32(v[10], bit);
+
+    v[13] = _mm_add_epi32(x, y);
+    v[13] = _mm_add_epi32(v[13], rnding);
+    v[13] = _mm_srai_epi32(v[13], bit);
+
+    x = _mm_mullo_epi32(u[11], cospi32);
+    y = _mm_mullo_epi32(u[12], cospi32);
+    v[11] = _mm_sub_epi32(y, x);
+    v[11] = _mm_add_epi32(v[11], rnding);
+    v[11] = _mm_srai_epi32(v[11], bit);
+
+    v[12] = _mm_add_epi32(x, y);
+    v[12] = _mm_add_epi32(v[12], rnding);
+    v[12] = _mm_srai_epi32(v[12], bit);
+
+    v[14] = u[14];
+    v[15] = u[15];
+
+    // stage 7
+    if (do_cols) {
+      addsub_no_clamp_sse4_1(v[0], v[15], out + 0, out + 15);
+      addsub_no_clamp_sse4_1(v[1], v[14], out + 1, out + 14);
+      addsub_no_clamp_sse4_1(v[2], v[13], out + 2, out + 13);
+      addsub_no_clamp_sse4_1(v[3], v[12], out + 3, out + 12);
+      addsub_no_clamp_sse4_1(v[4], v[11], out + 4, out + 11);
+      addsub_no_clamp_sse4_1(v[5], v[10], out + 5, out + 10);
+      addsub_no_clamp_sse4_1(v[6], v[9], out + 6, out + 9);
+      addsub_no_clamp_sse4_1(v[7], v[8], out + 7, out + 8);
+    } else {
+      const int log_range_out = AOMMAX(16, bd + 6);
+      const __m128i clamp_lo_out = _mm_set1_epi32(AOMMAX(
+          -(1 << (log_range_out - 1)), -(1 << (log_range - 1 - out_shift))));
+      const __m128i clamp_hi_out = _mm_set1_epi32(AOMMIN(
+          (1 << (log_range_out - 1)) - 1, (1 << (log_range - 1 - out_shift))));
+
+      addsub_shift_sse4_1(v[0], v[15], out + 0, out + 15, &clamp_lo_out,
+                          &clamp_hi_out, out_shift);
+      addsub_shift_sse4_1(v[1], v[14], out + 1, out + 14, &clamp_lo_out,
+                          &clamp_hi_out, out_shift);
+      addsub_shift_sse4_1(v[2], v[13], out + 2, out + 13, &clamp_lo_out,
+                          &clamp_hi_out, out_shift);
+      addsub_shift_sse4_1(v[3], v[12], out + 3, out + 12, &clamp_lo_out,
+                          &clamp_hi_out, out_shift);
+      addsub_shift_sse4_1(v[4], v[11], out + 4, out + 11, &clamp_lo_out,
+                          &clamp_hi_out, out_shift);
+      addsub_shift_sse4_1(v[5], v[10], out + 5, out + 10, &clamp_lo_out,
+                          &clamp_hi_out, out_shift);
+      addsub_shift_sse4_1(v[6], v[9], out + 6, out + 9, &clamp_lo_out,
+                          &clamp_hi_out, out_shift);
+      addsub_shift_sse4_1(v[7], v[8], out + 7, out + 8, &clamp_lo_out,
+                          &clamp_hi_out, out_shift);
+    }
+  }
+}
+
+static void iadst16x16_sse4_1(__m128i *in, __m128i *out, int bit, int do_cols,
+                              int bd, int out_shift) {
+  const int32_t *cospi = cospi_arr(bit);
+  const __m128i cospi2 = _mm_set1_epi32(cospi[2]);
+  const __m128i cospi62 = _mm_set1_epi32(cospi[62]);
+  const __m128i cospi10 = _mm_set1_epi32(cospi[10]);
+  const __m128i cospi54 = _mm_set1_epi32(cospi[54]);
+  const __m128i cospi18 = _mm_set1_epi32(cospi[18]);
+  const __m128i cospi46 = _mm_set1_epi32(cospi[46]);
+  const __m128i cospi26 = _mm_set1_epi32(cospi[26]);
+  const __m128i cospi38 = _mm_set1_epi32(cospi[38]);
+  const __m128i cospi34 = _mm_set1_epi32(cospi[34]);
+  const __m128i cospi30 = _mm_set1_epi32(cospi[30]);
+  const __m128i cospi42 = _mm_set1_epi32(cospi[42]);
+  const __m128i cospi22 = _mm_set1_epi32(cospi[22]);
+  const __m128i cospi50 = _mm_set1_epi32(cospi[50]);
+  const __m128i cospi14 = _mm_set1_epi32(cospi[14]);
+  const __m128i cospi58 = _mm_set1_epi32(cospi[58]);
+  const __m128i cospi6 = _mm_set1_epi32(cospi[6]);
+  const __m128i cospi8 = _mm_set1_epi32(cospi[8]);
+  const __m128i cospi56 = _mm_set1_epi32(cospi[56]);
+  const __m128i cospi40 = _mm_set1_epi32(cospi[40]);
+  const __m128i cospi24 = _mm_set1_epi32(cospi[24]);
+  const __m128i cospim56 = _mm_set1_epi32(-cospi[56]);
+  const __m128i cospim24 = _mm_set1_epi32(-cospi[24]);
+  const __m128i cospi48 = _mm_set1_epi32(cospi[48]);
+  const __m128i cospi16 = _mm_set1_epi32(cospi[16]);
+  const __m128i cospim48 = _mm_set1_epi32(-cospi[48]);
+  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
+  const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
+  const int log_range = AOMMAX(16, bd + (do_cols ? 6 : 8));
+  const __m128i clamp_lo = _mm_set1_epi32(-(1 << (log_range - 1)));
+  const __m128i clamp_hi = _mm_set1_epi32((1 << (log_range - 1)) - 1);
+  __m128i u[16], v[16], x, y;
+
+  // Calculate the column 0, 1, 2, 3
+  {
+    // stage 0
+    // stage 1
+    // stage 2
+    v[0] = _mm_mullo_epi32(in[15], cospi2);
+    x = _mm_mullo_epi32(in[0], cospi62);
+    v[0] = _mm_add_epi32(v[0], x);
+    v[0] = _mm_add_epi32(v[0], rnding);
+    v[0] = _mm_srai_epi32(v[0], bit);
+
+    v[1] = _mm_mullo_epi32(in[15], cospi62);
+    x = _mm_mullo_epi32(in[0], cospi2);
+    v[1] = _mm_sub_epi32(v[1], x);
+    v[1] = _mm_add_epi32(v[1], rnding);
+    v[1] = _mm_srai_epi32(v[1], bit);
+
+    v[2] = _mm_mullo_epi32(in[13], cospi10);
+    x = _mm_mullo_epi32(in[2], cospi54);
+    v[2] = _mm_add_epi32(v[2], x);
+    v[2] = _mm_add_epi32(v[2], rnding);
+    v[2] = _mm_srai_epi32(v[2], bit);
+
+    v[3] = _mm_mullo_epi32(in[13], cospi54);
+    x = _mm_mullo_epi32(in[2], cospi10);
+    v[3] = _mm_sub_epi32(v[3], x);
+    v[3] = _mm_add_epi32(v[3], rnding);
+    v[3] = _mm_srai_epi32(v[3], bit);
+
+    v[4] = _mm_mullo_epi32(in[11], cospi18);
+    x = _mm_mullo_epi32(in[4], cospi46);
+    v[4] = _mm_add_epi32(v[4], x);
+    v[4] = _mm_add_epi32(v[4], rnding);
+    v[4] = _mm_srai_epi32(v[4], bit);
+
+    v[5] = _mm_mullo_epi32(in[11], cospi46);
+    x = _mm_mullo_epi32(in[4], cospi18);
+    v[5] = _mm_sub_epi32(v[5], x);
+    v[5] = _mm_add_epi32(v[5], rnding);
+    v[5] = _mm_srai_epi32(v[5], bit);
+
+    v[6] = _mm_mullo_epi32(in[9], cospi26);
+    x = _mm_mullo_epi32(in[6], cospi38);
+    v[6] = _mm_add_epi32(v[6], x);
+    v[6] = _mm_add_epi32(v[6], rnding);
+    v[6] = _mm_srai_epi32(v[6], bit);
+
+    v[7] = _mm_mullo_epi32(in[9], cospi38);
+    x = _mm_mullo_epi32(in[6], cospi26);
+    v[7] = _mm_sub_epi32(v[7], x);
+    v[7] = _mm_add_epi32(v[7], rnding);
+    v[7] = _mm_srai_epi32(v[7], bit);
+
+    v[8] = _mm_mullo_epi32(in[7], cospi34);
+    x = _mm_mullo_epi32(in[8], cospi30);
+    v[8] = _mm_add_epi32(v[8], x);
+    v[8] = _mm_add_epi32(v[8], rnding);
+    v[8] = _mm_srai_epi32(v[8], bit);
+
+    v[9] = _mm_mullo_epi32(in[7], cospi30);
+    x = _mm_mullo_epi32(in[8], cospi34);
+    v[9] = _mm_sub_epi32(v[9], x);
+    v[9] = _mm_add_epi32(v[9], rnding);
+    v[9] = _mm_srai_epi32(v[9], bit);
+
+    v[10] = _mm_mullo_epi32(in[5], cospi42);
+    x = _mm_mullo_epi32(in[10], cospi22);
+    v[10] = _mm_add_epi32(v[10], x);
+    v[10] = _mm_add_epi32(v[10], rnding);
+    v[10] = _mm_srai_epi32(v[10], bit);
+
+    v[11] = _mm_mullo_epi32(in[5], cospi22);
+    x = _mm_mullo_epi32(in[10], cospi42);
+    v[11] = _mm_sub_epi32(v[11], x);
+    v[11] = _mm_add_epi32(v[11], rnding);
+    v[11] = _mm_srai_epi32(v[11], bit);
+
+    v[12] = _mm_mullo_epi32(in[3], cospi50);
+    x = _mm_mullo_epi32(in[12], cospi14);
+    v[12] = _mm_add_epi32(v[12], x);
+    v[12] = _mm_add_epi32(v[12], rnding);
+    v[12] = _mm_srai_epi32(v[12], bit);
+
+    v[13] = _mm_mullo_epi32(in[3], cospi14);
+    x = _mm_mullo_epi32(in[12], cospi50);
+    v[13] = _mm_sub_epi32(v[13], x);
+    v[13] = _mm_add_epi32(v[13], rnding);
+    v[13] = _mm_srai_epi32(v[13], bit);
+
+    v[14] = _mm_mullo_epi32(in[1], cospi58);
+    x = _mm_mullo_epi32(in[14], cospi6);
+    v[14] = _mm_add_epi32(v[14], x);
+    v[14] = _mm_add_epi32(v[14], rnding);
+    v[14] = _mm_srai_epi32(v[14], bit);
+
+    v[15] = _mm_mullo_epi32(in[1], cospi6);
+    x = _mm_mullo_epi32(in[14], cospi58);
+    v[15] = _mm_sub_epi32(v[15], x);
+    v[15] = _mm_add_epi32(v[15], rnding);
+    v[15] = _mm_srai_epi32(v[15], bit);
+
+    // stage 3
+    addsub_sse4_1(v[0], v[8], &u[0], &u[8], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(v[1], v[9], &u[1], &u[9], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(v[2], v[10], &u[2], &u[10], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(v[3], v[11], &u[3], &u[11], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(v[4], v[12], &u[4], &u[12], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(v[5], v[13], &u[5], &u[13], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(v[6], v[14], &u[6], &u[14], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(v[7], v[15], &u[7], &u[15], &clamp_lo, &clamp_hi);
+
+    // stage 4
+    v[0] = u[0];
+    v[1] = u[1];
+    v[2] = u[2];
+    v[3] = u[3];
+    v[4] = u[4];
+    v[5] = u[5];
+    v[6] = u[6];
+    v[7] = u[7];
+
+    v[8] = _mm_mullo_epi32(u[8], cospi8);
+    x = _mm_mullo_epi32(u[9], cospi56);
+    v[8] = _mm_add_epi32(v[8], x);
+    v[8] = _mm_add_epi32(v[8], rnding);
+    v[8] = _mm_srai_epi32(v[8], bit);
+
+    v[9] = _mm_mullo_epi32(u[8], cospi56);
+    x = _mm_mullo_epi32(u[9], cospi8);
+    v[9] = _mm_sub_epi32(v[9], x);
+    v[9] = _mm_add_epi32(v[9], rnding);
+    v[9] = _mm_srai_epi32(v[9], bit);
+
+    v[10] = _mm_mullo_epi32(u[10], cospi40);
+    x = _mm_mullo_epi32(u[11], cospi24);
+    v[10] = _mm_add_epi32(v[10], x);
+    v[10] = _mm_add_epi32(v[10], rnding);
+    v[10] = _mm_srai_epi32(v[10], bit);
+
+    v[11] = _mm_mullo_epi32(u[10], cospi24);
+    x = _mm_mullo_epi32(u[11], cospi40);
+    v[11] = _mm_sub_epi32(v[11], x);
+    v[11] = _mm_add_epi32(v[11], rnding);
+    v[11] = _mm_srai_epi32(v[11], bit);
+
+    v[12] = _mm_mullo_epi32(u[12], cospim56);
+    x = _mm_mullo_epi32(u[13], cospi8);
+    v[12] = _mm_add_epi32(v[12], x);
+    v[12] = _mm_add_epi32(v[12], rnding);
+    v[12] = _mm_srai_epi32(v[12], bit);
+
+    v[13] = _mm_mullo_epi32(u[12], cospi8);
+    x = _mm_mullo_epi32(u[13], cospim56);
+    v[13] = _mm_sub_epi32(v[13], x);
+    v[13] = _mm_add_epi32(v[13], rnding);
+    v[13] = _mm_srai_epi32(v[13], bit);
+
+    v[14] = _mm_mullo_epi32(u[14], cospim24);
+    x = _mm_mullo_epi32(u[15], cospi40);
+    v[14] = _mm_add_epi32(v[14], x);
+    v[14] = _mm_add_epi32(v[14], rnding);
+    v[14] = _mm_srai_epi32(v[14], bit);
+
+    v[15] = _mm_mullo_epi32(u[14], cospi40);
+    x = _mm_mullo_epi32(u[15], cospim24);
+    v[15] = _mm_sub_epi32(v[15], x);
+    v[15] = _mm_add_epi32(v[15], rnding);
+    v[15] = _mm_srai_epi32(v[15], bit);
+
+    // stage 5
+    addsub_sse4_1(v[0], v[4], &u[0], &u[4], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(v[1], v[5], &u[1], &u[5], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(v[2], v[6], &u[2], &u[6], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(v[3], v[7], &u[3], &u[7], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(v[8], v[12], &u[8], &u[12], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(v[9], v[13], &u[9], &u[13], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(v[10], v[14], &u[10], &u[14], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(v[11], v[15], &u[11], &u[15], &clamp_lo, &clamp_hi);
+
+    // stage 6
+    v[0] = u[0];
+    v[1] = u[1];
+    v[2] = u[2];
+    v[3] = u[3];
+
+    v[4] = _mm_mullo_epi32(u[4], cospi16);
+    x = _mm_mullo_epi32(u[5], cospi48);
+    v[4] = _mm_add_epi32(v[4], x);
+    v[4] = _mm_add_epi32(v[4], rnding);
+    v[4] = _mm_srai_epi32(v[4], bit);
+
+    v[5] = _mm_mullo_epi32(u[4], cospi48);
+    x = _mm_mullo_epi32(u[5], cospi16);
+    v[5] = _mm_sub_epi32(v[5], x);
+    v[5] = _mm_add_epi32(v[5], rnding);
+    v[5] = _mm_srai_epi32(v[5], bit);
+
+    v[6] = _mm_mullo_epi32(u[6], cospim48);
+    x = _mm_mullo_epi32(u[7], cospi16);
+    v[6] = _mm_add_epi32(v[6], x);
+    v[6] = _mm_add_epi32(v[6], rnding);
+    v[6] = _mm_srai_epi32(v[6], bit);
+
+    v[7] = _mm_mullo_epi32(u[6], cospi16);
+    x = _mm_mullo_epi32(u[7], cospim48);
+    v[7] = _mm_sub_epi32(v[7], x);
+    v[7] = _mm_add_epi32(v[7], rnding);
+    v[7] = _mm_srai_epi32(v[7], bit);
+
+    v[8] = u[8];
+    v[9] = u[9];
+    v[10] = u[10];
+    v[11] = u[11];
+
+    v[12] = _mm_mullo_epi32(u[12], cospi16);
+    x = _mm_mullo_epi32(u[13], cospi48);
+    v[12] = _mm_add_epi32(v[12], x);
+    v[12] = _mm_add_epi32(v[12], rnding);
+    v[12] = _mm_srai_epi32(v[12], bit);
+
+    v[13] = _mm_mullo_epi32(u[12], cospi48);
+    x = _mm_mullo_epi32(u[13], cospi16);
+    v[13] = _mm_sub_epi32(v[13], x);
+    v[13] = _mm_add_epi32(v[13], rnding);
+    v[13] = _mm_srai_epi32(v[13], bit);
+
+    v[14] = _mm_mullo_epi32(u[14], cospim48);
+    x = _mm_mullo_epi32(u[15], cospi16);
+    v[14] = _mm_add_epi32(v[14], x);
+    v[14] = _mm_add_epi32(v[14], rnding);
+    v[14] = _mm_srai_epi32(v[14], bit);
+
+    v[15] = _mm_mullo_epi32(u[14], cospi16);
+    x = _mm_mullo_epi32(u[15], cospim48);
+    v[15] = _mm_sub_epi32(v[15], x);
+    v[15] = _mm_add_epi32(v[15], rnding);
+    v[15] = _mm_srai_epi32(v[15], bit);
+
+    // stage 7
+    addsub_sse4_1(v[0], v[2], &u[0], &u[2], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(v[1], v[3], &u[1], &u[3], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(v[4], v[6], &u[4], &u[6], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(v[5], v[7], &u[5], &u[7], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(v[8], v[10], &u[8], &u[10], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(v[9], v[11], &u[9], &u[11], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(v[12], v[14], &u[12], &u[14], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(v[13], v[15], &u[13], &u[15], &clamp_lo, &clamp_hi);
+
+    // stage 8
+    v[0] = u[0];
+    v[1] = u[1];
+
+    y = _mm_mullo_epi32(u[2], cospi32);
+    x = _mm_mullo_epi32(u[3], cospi32);
+    v[2] = _mm_add_epi32(y, x);
+    v[2] = _mm_add_epi32(v[2], rnding);
+    v[2] = _mm_srai_epi32(v[2], bit);
+
+    v[3] = _mm_sub_epi32(y, x);
+    v[3] = _mm_add_epi32(v[3], rnding);
+    v[3] = _mm_srai_epi32(v[3], bit);
+
+    v[4] = u[4];
+    v[5] = u[5];
+
+    y = _mm_mullo_epi32(u[6], cospi32);
+    x = _mm_mullo_epi32(u[7], cospi32);
+    v[6] = _mm_add_epi32(y, x);
+    v[6] = _mm_add_epi32(v[6], rnding);
+    v[6] = _mm_srai_epi32(v[6], bit);
+
+    v[7] = _mm_sub_epi32(y, x);
+    v[7] = _mm_add_epi32(v[7], rnding);
+    v[7] = _mm_srai_epi32(v[7], bit);
+
+    v[8] = u[8];
+    v[9] = u[9];
+
+    y = _mm_mullo_epi32(u[10], cospi32);
+    x = _mm_mullo_epi32(u[11], cospi32);
+    v[10] = _mm_add_epi32(y, x);
+    v[10] = _mm_add_epi32(v[10], rnding);
+    v[10] = _mm_srai_epi32(v[10], bit);
+
+    v[11] = _mm_sub_epi32(y, x);
+    v[11] = _mm_add_epi32(v[11], rnding);
+    v[11] = _mm_srai_epi32(v[11], bit);
+
+    v[12] = u[12];
+    v[13] = u[13];
+
+    y = _mm_mullo_epi32(u[14], cospi32);
+    x = _mm_mullo_epi32(u[15], cospi32);
+    v[14] = _mm_add_epi32(y, x);
+    v[14] = _mm_add_epi32(v[14], rnding);
+    v[14] = _mm_srai_epi32(v[14], bit);
+
+    v[15] = _mm_sub_epi32(y, x);
+    v[15] = _mm_add_epi32(v[15], rnding);
+    v[15] = _mm_srai_epi32(v[15], bit);
+
+    // stage 9
+    if (do_cols) {
+      out[0] = v[0];
+      out[1] = _mm_sub_epi32(_mm_setzero_si128(), v[8]);
+      out[2] = v[12];
+      out[3] = _mm_sub_epi32(_mm_setzero_si128(), v[4]);
+      out[4] = v[6];
+      out[5] = _mm_sub_epi32(_mm_setzero_si128(), v[14]);
+      out[6] = v[10];
+      out[7] = _mm_sub_epi32(_mm_setzero_si128(), v[2]);
+      out[8] = v[3];
+      out[9] = _mm_sub_epi32(_mm_setzero_si128(), v[11]);
+      out[10] = v[15];
+      out[11] = _mm_sub_epi32(_mm_setzero_si128(), v[7]);
+      out[12] = v[5];
+      out[13] = _mm_sub_epi32(_mm_setzero_si128(), v[13]);
+      out[14] = v[9];
+      out[15] = _mm_sub_epi32(_mm_setzero_si128(), v[1]);
+    } else {
+      const int log_range_out = AOMMAX(16, bd + 6);
+      const __m128i clamp_lo_out = _mm_set1_epi32(-(1 << (log_range_out - 1)));
+      const __m128i clamp_hi_out =
+          _mm_set1_epi32((1 << (log_range_out - 1)) - 1);
+
+      neg_shift_sse4_1(v[0], v[8], out + 0, out + 1, &clamp_lo_out,
+                       &clamp_hi_out, out_shift);
+      neg_shift_sse4_1(v[12], v[4], out + 2, out + 3, &clamp_lo_out,
+                       &clamp_hi_out, out_shift);
+      neg_shift_sse4_1(v[6], v[14], out + 4, out + 5, &clamp_lo_out,
+                       &clamp_hi_out, out_shift);
+      neg_shift_sse4_1(v[10], v[2], out + 6, out + 7, &clamp_lo_out,
+                       &clamp_hi_out, out_shift);
+      neg_shift_sse4_1(v[3], v[11], out + 8, out + 9, &clamp_lo_out,
+                       &clamp_hi_out, out_shift);
+      neg_shift_sse4_1(v[15], v[7], out + 10, out + 11, &clamp_lo_out,
+                       &clamp_hi_out, out_shift);
+      neg_shift_sse4_1(v[5], v[13], out + 12, out + 13, &clamp_lo_out,
+                       &clamp_hi_out, out_shift);
+      neg_shift_sse4_1(v[9], v[1], out + 14, out + 15, &clamp_lo_out,
+                       &clamp_hi_out, out_shift);
+    }
+  }
+}
+
+static INLINE void idct64_stage8_sse4_1(
+    __m128i *u, const __m128i *cospim32, const __m128i *cospi32,
+    const __m128i *cospim16, const __m128i *cospi48, const __m128i *cospi16,
+    const __m128i *cospim48, const __m128i *clamp_lo, const __m128i *clamp_hi,
+    const __m128i *rnding, int bit) {
+  int i;
+  __m128i temp1, temp2, temp3, temp4;
+  temp1 = half_btf_sse4_1(cospim32, &u[10], cospi32, &u[13], rnding, bit);
+  u[13] = half_btf_sse4_1(cospi32, &u[10], cospi32, &u[13], rnding, bit);
+  u[10] = temp1;
+  temp2 = half_btf_sse4_1(cospim32, &u[11], cospi32, &u[12], rnding, bit);
+  u[12] = half_btf_sse4_1(cospi32, &u[11], cospi32, &u[12], rnding, bit);
+  u[11] = temp2;
+
+  for (i = 16; i < 20; ++i) {
+    addsub_sse4_1(u[i], u[i ^ 7], &u[i], &u[i ^ 7], clamp_lo, clamp_hi);
+    addsub_sse4_1(u[i ^ 15], u[i ^ 8], &u[i ^ 15], &u[i ^ 8], clamp_lo,
+                  clamp_hi);
+  }
+
+  temp1 = half_btf_sse4_1(cospim16, &u[36], cospi48, &u[59], rnding, bit);
+  temp2 = half_btf_sse4_1(cospim16, &u[37], cospi48, &u[58], rnding, bit);
+  temp3 = half_btf_sse4_1(cospim16, &u[38], cospi48, &u[57], rnding, bit);
+  temp4 = half_btf_sse4_1(cospim16, &u[39], cospi48, &u[56], rnding, bit);
+  u[56] = half_btf_sse4_1(cospi48, &u[39], cospi16, &u[56], rnding, bit);
+  u[57] = half_btf_sse4_1(cospi48, &u[38], cospi16, &u[57], rnding, bit);
+  u[58] = half_btf_sse4_1(cospi48, &u[37], cospi16, &u[58], rnding, bit);
+  u[59] = half_btf_sse4_1(cospi48, &u[36], cospi16, &u[59], rnding, bit);
+  u[36] = temp1;
+  u[37] = temp2;
+  u[38] = temp3;
+  u[39] = temp4;
+
+  temp1 = half_btf_sse4_1(cospim48, &u[40], cospim16, &u[55], rnding, bit);
+  temp2 = half_btf_sse4_1(cospim48, &u[41], cospim16, &u[54], rnding, bit);
+  temp3 = half_btf_sse4_1(cospim48, &u[42], cospim16, &u[53], rnding, bit);
+  temp4 = half_btf_sse4_1(cospim48, &u[43], cospim16, &u[52], rnding, bit);
+  u[52] = half_btf_sse4_1(cospim16, &u[43], cospi48, &u[52], rnding, bit);
+  u[53] = half_btf_sse4_1(cospim16, &u[42], cospi48, &u[53], rnding, bit);
+  u[54] = half_btf_sse4_1(cospim16, &u[41], cospi48, &u[54], rnding, bit);
+  u[55] = half_btf_sse4_1(cospim16, &u[40], cospi48, &u[55], rnding, bit);
+  u[40] = temp1;
+  u[41] = temp2;
+  u[42] = temp3;
+  u[43] = temp4;
+}
+
+static INLINE void idct64_stage9_sse4_1(__m128i *u, const __m128i *cospim32,
+                                        const __m128i *cospi32,
+                                        const __m128i *clamp_lo,
+                                        const __m128i *clamp_hi,
+                                        const __m128i *rnding, int bit) {
+  int i;
+  __m128i temp1, temp2, temp3, temp4;
+  for (i = 0; i < 8; ++i) {
+    addsub_sse4_1(u[i], u[15 - i], &u[i], &u[15 - i], clamp_lo, clamp_hi);
+  }
+
+  temp1 = half_btf_sse4_1(cospim32, &u[20], cospi32, &u[27], rnding, bit);
+  temp2 = half_btf_sse4_1(cospim32, &u[21], cospi32, &u[26], rnding, bit);
+  temp3 = half_btf_sse4_1(cospim32, &u[22], cospi32, &u[25], rnding, bit);
+  temp4 = half_btf_sse4_1(cospim32, &u[23], cospi32, &u[24], rnding, bit);
+  u[24] = half_btf_sse4_1(cospi32, &u[23], cospi32, &u[24], rnding, bit);
+  u[25] = half_btf_sse4_1(cospi32, &u[22], cospi32, &u[25], rnding, bit);
+  u[26] = half_btf_sse4_1(cospi32, &u[21], cospi32, &u[26], rnding, bit);
+  u[27] = half_btf_sse4_1(cospi32, &u[20], cospi32, &u[27], rnding, bit);
+  u[20] = temp1;
+  u[21] = temp2;
+  u[22] = temp3;
+  u[23] = temp4;
+  for (i = 32; i < 40; i++) {
+    addsub_sse4_1(u[i], u[i ^ 15], &u[i], &u[i ^ 15], clamp_lo, clamp_hi);
+  }
+
+  for (i = 48; i < 56; i++) {
+    addsub_sse4_1(u[i ^ 15], u[i], &u[i ^ 15], &u[i], clamp_lo, clamp_hi);
+  }
+}
+
+static INLINE void idct64_stage10_sse4_1(__m128i *u, const __m128i *cospim32,
+                                         const __m128i *cospi32,
+                                         const __m128i *clamp_lo,
+                                         const __m128i *clamp_hi,
+                                         const __m128i *rnding, int bit) {
+  __m128i temp1, temp2, temp3, temp4;
+  for (int i = 0; i < 16; i++) {
+    addsub_sse4_1(u[i], u[31 - i], &u[i], &u[31 - i], clamp_lo, clamp_hi);
+  }
+
+  temp1 = half_btf_sse4_1(cospim32, &u[40], cospi32, &u[55], rnding, bit);
+  temp2 = half_btf_sse4_1(cospim32, &u[41], cospi32, &u[54], rnding, bit);
+  temp3 = half_btf_sse4_1(cospim32, &u[42], cospi32, &u[53], rnding, bit);
+  temp4 = half_btf_sse4_1(cospim32, &u[43], cospi32, &u[52], rnding, bit);
+  u[52] = half_btf_sse4_1(cospi32, &u[43], cospi32, &u[52], rnding, bit);
+  u[53] = half_btf_sse4_1(cospi32, &u[42], cospi32, &u[53], rnding, bit);
+  u[54] = half_btf_sse4_1(cospi32, &u[41], cospi32, &u[54], rnding, bit);
+  u[55] = half_btf_sse4_1(cospi32, &u[40], cospi32, &u[55], rnding, bit);
+  u[40] = temp1;
+  u[41] = temp2;
+  u[42] = temp3;
+  u[43] = temp4;
+
+  temp1 = half_btf_sse4_1(cospim32, &u[44], cospi32, &u[51], rnding, bit);
+  temp2 = half_btf_sse4_1(cospim32, &u[45], cospi32, &u[50], rnding, bit);
+  temp3 = half_btf_sse4_1(cospim32, &u[46], cospi32, &u[49], rnding, bit);
+  temp4 = half_btf_sse4_1(cospim32, &u[47], cospi32, &u[48], rnding, bit);
+  u[48] = half_btf_sse4_1(cospi32, &u[47], cospi32, &u[48], rnding, bit);
+  u[49] = half_btf_sse4_1(cospi32, &u[46], cospi32, &u[49], rnding, bit);
+  u[50] = half_btf_sse4_1(cospi32, &u[45], cospi32, &u[50], rnding, bit);
+  u[51] = half_btf_sse4_1(cospi32, &u[44], cospi32, &u[51], rnding, bit);
+  u[44] = temp1;
+  u[45] = temp2;
+  u[46] = temp3;
+  u[47] = temp4;
+}
+
+static INLINE void idct64_stage11_sse4_1(__m128i *u, __m128i *out, int do_cols,
+                                         int bd, int out_shift,
+                                         const int log_range) {
+  if (do_cols) {
+    for (int i = 0; i < 32; i++) {
+      addsub_no_clamp_sse4_1(u[i], u[63 - i], &out[(i)], &out[(63 - i)]);
+    }
+  } else {
+    const int log_range_out = AOMMAX(16, bd + 6);
+    const __m128i clamp_lo_out = _mm_set1_epi32(AOMMAX(
+        -(1 << (log_range_out - 1)), -(1 << (log_range - 1 - out_shift))));
+    const __m128i clamp_hi_out = _mm_set1_epi32(AOMMIN(
+        (1 << (log_range_out - 1)) - 1, (1 << (log_range - 1 - out_shift))));
+
+    for (int i = 0; i < 32; i++) {
+      addsub_shift_sse4_1(u[i], u[63 - i], &out[(i)], &out[(63 - i)],
+                          &clamp_lo_out, &clamp_hi_out, out_shift);
+    }
+  }
+}
+
+static void idct64x64_low1_sse4_1(__m128i *in, __m128i *out, int bit,
+                                  int do_cols, int bd, int out_shift) {
+  const int32_t *cospi = cospi_arr(bit);
+  const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
+  const int log_range = AOMMAX(16, bd + (do_cols ? 6 : 8));
+  const __m128i clamp_lo = _mm_set1_epi32(-(1 << (log_range - 1)));
+  const __m128i clamp_hi = _mm_set1_epi32((1 << (log_range - 1)) - 1);
+
+  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
+
+  {
+    __m128i x;
+
+    // stage 1
+    // stage 2
+    // stage 3
+    // stage 4
+    // stage 5
+    // stage 6
+    x = half_btf_0_sse4_1(&cospi32, &in[0], &rnding, bit);
+
+    // stage 8
+    // stage 9
+    // stage 10
+    // stage 11
+    if (do_cols) {
+      x = _mm_max_epi32(x, clamp_lo);
+      x = _mm_min_epi32(x, clamp_hi);
+    } else {
+      const int log_range_out = AOMMAX(16, bd + 6);
+      const __m128i clamp_lo_out = _mm_set1_epi32(AOMMAX(
+          -(1 << (log_range_out - 1)), -(1 << (log_range - 1 - out_shift))));
+      const __m128i clamp_hi_out = _mm_set1_epi32(AOMMIN(
+          (1 << (log_range_out - 1)) - 1, (1 << (log_range - 1 - out_shift))));
+
+      __m128i offset = _mm_set1_epi32((1 << out_shift) >> 1);
+      x = _mm_add_epi32(x, offset);
+      x = _mm_sra_epi32(x, _mm_cvtsi32_si128(out_shift));
+
+      x = _mm_max_epi32(x, clamp_lo_out);
+      x = _mm_min_epi32(x, clamp_hi_out);
+    }
+
+    out[0] = x;
+    out[63] = x;
+    out[1] = x;
+    out[62] = x;
+    out[2] = x;
+    out[61] = x;
+    out[3] = x;
+    out[60] = x;
+    out[4] = x;
+    out[59] = x;
+    out[5] = x;
+    out[58] = x;
+    out[6] = x;
+    out[57] = x;
+    out[7] = x;
+    out[56] = x;
+    out[8] = x;
+    out[55] = x;
+    out[9] = x;
+    out[54] = x;
+    out[10] = x;
+    out[53] = x;
+    out[11] = x;
+    out[52] = x;
+    out[12] = x;
+    out[51] = x;
+    out[13] = x;
+    out[50] = x;
+    out[14] = x;
+    out[49] = x;
+    out[15] = x;
+    out[48] = x;
+    out[16] = x;
+    out[47] = x;
+    out[17] = x;
+    out[46] = x;
+    out[18] = x;
+    out[45] = x;
+    out[19] = x;
+    out[44] = x;
+    out[20] = x;
+    out[43] = x;
+    out[21] = x;
+    out[42] = x;
+    out[22] = x;
+    out[41] = x;
+    out[23] = x;
+    out[40] = x;
+    out[24] = x;
+    out[39] = x;
+    out[25] = x;
+    out[38] = x;
+    out[26] = x;
+    out[37] = x;
+    out[27] = x;
+    out[36] = x;
+    out[28] = x;
+    out[35] = x;
+    out[29] = x;
+    out[34] = x;
+    out[30] = x;
+    out[33] = x;
+    out[31] = x;
+    out[32] = x;
+  }
+}
+
+static void idct64x64_low8_sse4_1(__m128i *in, __m128i *out, int bit,
+                                  int do_cols, int bd, int out_shift) {
+  int i, j;
+  const int32_t *cospi = cospi_arr(bit);
+  const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
+  const int log_range = AOMMAX(16, bd + (do_cols ? 6 : 8));
+  const __m128i clamp_lo = _mm_set1_epi32(-(1 << (log_range - 1)));
+  const __m128i clamp_hi = _mm_set1_epi32((1 << (log_range - 1)) - 1);
+
+  const __m128i cospi1 = _mm_set1_epi32(cospi[1]);
+  const __m128i cospi2 = _mm_set1_epi32(cospi[2]);
+  const __m128i cospi3 = _mm_set1_epi32(cospi[3]);
+  const __m128i cospi4 = _mm_set1_epi32(cospi[4]);
+  const __m128i cospi6 = _mm_set1_epi32(cospi[6]);
+  const __m128i cospi8 = _mm_set1_epi32(cospi[8]);
+  const __m128i cospi12 = _mm_set1_epi32(cospi[12]);
+  const __m128i cospi16 = _mm_set1_epi32(cospi[16]);
+  const __m128i cospi20 = _mm_set1_epi32(cospi[20]);
+  const __m128i cospi24 = _mm_set1_epi32(cospi[24]);
+  const __m128i cospi28 = _mm_set1_epi32(cospi[28]);
+  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
+  const __m128i cospi40 = _mm_set1_epi32(cospi[40]);
+  const __m128i cospi44 = _mm_set1_epi32(cospi[44]);
+  const __m128i cospi48 = _mm_set1_epi32(cospi[48]);
+  const __m128i cospi56 = _mm_set1_epi32(cospi[56]);
+  const __m128i cospi60 = _mm_set1_epi32(cospi[60]);
+  const __m128i cospim4 = _mm_set1_epi32(-cospi[4]);
+  const __m128i cospim8 = _mm_set1_epi32(-cospi[8]);
+  const __m128i cospim12 = _mm_set1_epi32(-cospi[12]);
+  const __m128i cospim16 = _mm_set1_epi32(-cospi[16]);
+  const __m128i cospim20 = _mm_set1_epi32(-cospi[20]);
+  const __m128i cospim24 = _mm_set1_epi32(-cospi[24]);
+  const __m128i cospim28 = _mm_set1_epi32(-cospi[28]);
+  const __m128i cospim32 = _mm_set1_epi32(-cospi[32]);
+  const __m128i cospim36 = _mm_set1_epi32(-cospi[36]);
+  const __m128i cospim40 = _mm_set1_epi32(-cospi[40]);
+  const __m128i cospim48 = _mm_set1_epi32(-cospi[48]);
+  const __m128i cospim52 = _mm_set1_epi32(-cospi[52]);
+  const __m128i cospim56 = _mm_set1_epi32(-cospi[56]);
+  const __m128i cospi63 = _mm_set1_epi32(cospi[63]);
+  const __m128i cospim57 = _mm_set1_epi32(-cospi[57]);
+  const __m128i cospi7 = _mm_set1_epi32(cospi[7]);
+  const __m128i cospi5 = _mm_set1_epi32(cospi[5]);
+  const __m128i cospi59 = _mm_set1_epi32(cospi[59]);
+  const __m128i cospim61 = _mm_set1_epi32(-cospi[61]);
+  const __m128i cospim58 = _mm_set1_epi32(-cospi[58]);
+  const __m128i cospi62 = _mm_set1_epi32(cospi[62]);
+
+  {
+    __m128i u[64];
+
+    // stage 1
+    u[0] = in[0];
+    u[8] = in[4];
+    u[16] = in[2];
+    u[24] = in[6];
+    u[32] = in[1];
+    u[40] = in[5];
+    u[48] = in[3];
+    u[56] = in[7];
+
+    // stage 2
+    u[63] = half_btf_0_sse4_1(&cospi1, &u[32], &rnding, bit);
+    u[32] = half_btf_0_sse4_1(&cospi63, &u[32], &rnding, bit);
+    u[39] = half_btf_0_sse4_1(&cospim57, &u[56], &rnding, bit);
+    u[56] = half_btf_0_sse4_1(&cospi7, &u[56], &rnding, bit);
+    u[55] = half_btf_0_sse4_1(&cospi5, &u[40], &rnding, bit);
+    u[40] = half_btf_0_sse4_1(&cospi59, &u[40], &rnding, bit);
+    u[47] = half_btf_0_sse4_1(&cospim61, &u[48], &rnding, bit);
+    u[48] = half_btf_0_sse4_1(&cospi3, &u[48], &rnding, bit);
+
+    // stage 3
+    u[31] = half_btf_0_sse4_1(&cospi2, &u[16], &rnding, bit);
+    u[16] = half_btf_0_sse4_1(&cospi62, &u[16], &rnding, bit);
+    u[23] = half_btf_0_sse4_1(&cospim58, &u[24], &rnding, bit);
+    u[24] = half_btf_0_sse4_1(&cospi6, &u[24], &rnding, bit);
+    u[33] = u[32];
+    u[38] = u[39];
+    u[41] = u[40];
+    u[46] = u[47];
+    u[49] = u[48];
+    u[54] = u[55];
+    u[57] = u[56];
+    u[62] = u[63];
+
+    // stage 4
+    __m128i temp1, temp2;
+    u[15] = half_btf_0_sse4_1(&cospi4, &u[8], &rnding, bit);
+    u[8] = half_btf_0_sse4_1(&cospi60, &u[8], &rnding, bit);
+    u[17] = u[16];
+    u[22] = u[23];
+    u[25] = u[24];
+    u[30] = u[31];
+
+    temp1 = half_btf_sse4_1(&cospim4, &u[33], &cospi60, &u[62], &rnding, bit);
+    u[62] = half_btf_sse4_1(&cospi60, &u[33], &cospi4, &u[62], &rnding, bit);
+    u[33] = temp1;
+
+    temp2 = half_btf_sse4_1(&cospim36, &u[38], &cospi28, &u[57], &rnding, bit);
+    u[38] = half_btf_sse4_1(&cospim28, &u[38], &cospim36, &u[57], &rnding, bit);
+    u[57] = temp2;
+
+    temp1 = half_btf_sse4_1(&cospim20, &u[41], &cospi44, &u[54], &rnding, bit);
+    u[54] = half_btf_sse4_1(&cospi44, &u[41], &cospi20, &u[54], &rnding, bit);
+    u[41] = temp1;
+
+    temp2 = half_btf_sse4_1(&cospim12, &u[46], &cospim52, &u[49], &rnding, bit);
+    u[49] = half_btf_sse4_1(&cospim52, &u[46], &cospi12, &u[49], &rnding, bit);
+    u[46] = temp2;
+
+    // stage 5
+    u[9] = u[8];
+    u[14] = u[15];
+
+    temp1 = half_btf_sse4_1(&cospim8, &u[17], &cospi56, &u[30], &rnding, bit);
+    u[30] = half_btf_sse4_1(&cospi56, &u[17], &cospi8, &u[30], &rnding, bit);
+    u[17] = temp1;
+
+    temp2 = half_btf_sse4_1(&cospim24, &u[22], &cospim40, &u[25], &rnding, bit);
+    u[25] = half_btf_sse4_1(&cospim40, &u[22], &cospi24, &u[25], &rnding, bit);
+    u[22] = temp2;
+
+    u[35] = u[32];
+    u[34] = u[33];
+    u[36] = u[39];
+    u[37] = u[38];
+    u[43] = u[40];
+    u[42] = u[41];
+    u[44] = u[47];
+    u[45] = u[46];
+    u[51] = u[48];
+    u[50] = u[49];
+    u[52] = u[55];
+    u[53] = u[54];
+    u[59] = u[56];
+    u[58] = u[57];
+    u[60] = u[63];
+    u[61] = u[62];
+
+    // stage 6
+    temp1 = half_btf_0_sse4_1(&cospi32, &u[0], &rnding, bit);
+    u[1] = half_btf_0_sse4_1(&cospi32, &u[0], &rnding, bit);
+    u[0] = temp1;
+
+    temp2 = half_btf_sse4_1(&cospim16, &u[9], &cospi48, &u[14], &rnding, bit);
+    u[14] = half_btf_sse4_1(&cospi48, &u[9], &cospi16, &u[14], &rnding, bit);
+    u[9] = temp2;
+    u[19] = u[16];
+    u[18] = u[17];
+    u[20] = u[23];
+    u[21] = u[22];
+    u[27] = u[24];
+    u[26] = u[25];
+    u[28] = u[31];
+    u[29] = u[30];
+
+    temp1 = half_btf_sse4_1(&cospim8, &u[34], &cospi56, &u[61], &rnding, bit);
+    u[61] = half_btf_sse4_1(&cospi56, &u[34], &cospi8, &u[61], &rnding, bit);
+    u[34] = temp1;
+    temp2 = half_btf_sse4_1(&cospim8, &u[35], &cospi56, &u[60], &rnding, bit);
+    u[60] = half_btf_sse4_1(&cospi56, &u[35], &cospi8, &u[60], &rnding, bit);
+    u[35] = temp2;
+    temp1 = half_btf_sse4_1(&cospim56, &u[36], &cospim8, &u[59], &rnding, bit);
+    u[59] = half_btf_sse4_1(&cospim8, &u[36], &cospi56, &u[59], &rnding, bit);
+    u[36] = temp1;
+    temp2 = half_btf_sse4_1(&cospim56, &u[37], &cospim8, &u[58], &rnding, bit);
+    u[58] = half_btf_sse4_1(&cospim8, &u[37], &cospi56, &u[58], &rnding, bit);
+    u[37] = temp2;
+    temp1 = half_btf_sse4_1(&cospim40, &u[42], &cospi24, &u[53], &rnding, bit);
+    u[53] = half_btf_sse4_1(&cospi24, &u[42], &cospi40, &u[53], &rnding, bit);
+    u[42] = temp1;
+    temp2 = half_btf_sse4_1(&cospim40, &u[43], &cospi24, &u[52], &rnding, bit);
+    u[52] = half_btf_sse4_1(&cospi24, &u[43], &cospi40, &u[52], &rnding, bit);
+    u[43] = temp2;
+    temp1 = half_btf_sse4_1(&cospim24, &u[44], &cospim40, &u[51], &rnding, bit);
+    u[51] = half_btf_sse4_1(&cospim40, &u[44], &cospi24, &u[51], &rnding, bit);
+    u[44] = temp1;
+    temp2 = half_btf_sse4_1(&cospim24, &u[45], &cospim40, &u[50], &rnding, bit);
+    u[50] = half_btf_sse4_1(&cospim40, &u[45], &cospi24, &u[50], &rnding, bit);
+    u[45] = temp2;
+
+    // stage 7
+    u[3] = u[0];
+    u[2] = u[1];
+    u[11] = u[8];
+    u[10] = u[9];
+    u[12] = u[15];
+    u[13] = u[14];
+
+    temp1 = half_btf_sse4_1(&cospim16, &u[18], &cospi48, &u[29], &rnding, bit);
+    u[29] = half_btf_sse4_1(&cospi48, &u[18], &cospi16, &u[29], &rnding, bit);
+    u[18] = temp1;
+    temp2 = half_btf_sse4_1(&cospim16, &u[19], &cospi48, &u[28], &rnding, bit);
+    u[28] = half_btf_sse4_1(&cospi48, &u[19], &cospi16, &u[28], &rnding, bit);
+    u[19] = temp2;
+    temp1 = half_btf_sse4_1(&cospim48, &u[20], &cospim16, &u[27], &rnding, bit);
+    u[27] = half_btf_sse4_1(&cospim16, &u[20], &cospi48, &u[27], &rnding, bit);
+    u[20] = temp1;
+    temp2 = half_btf_sse4_1(&cospim48, &u[21], &cospim16, &u[26], &rnding, bit);
+    u[26] = half_btf_sse4_1(&cospim16, &u[21], &cospi48, &u[26], &rnding, bit);
+    u[21] = temp2;
+    for (i = 32; i < 64; i += 16) {
+      for (j = i; j < i + 4; j++) {
+        addsub_sse4_1(u[j], u[j ^ 7], &u[j], &u[j ^ 7], &clamp_lo, &clamp_hi);
+        addsub_sse4_1(u[j ^ 15], u[j ^ 8], &u[j ^ 15], &u[j ^ 8], &clamp_lo,
+                      &clamp_hi);
+      }
+    }
+
+    // stage 8
+    u[7] = u[0];
+    u[6] = u[1];
+    u[5] = u[2];
+    u[4] = u[3];
+    u[9] = u[9];
+
+    idct64_stage8_sse4_1(u, &cospim32, &cospi32, &cospim16, &cospi48, &cospi16,
+                         &cospim48, &clamp_lo, &clamp_hi, &rnding, bit);
+
+    // stage 9
+    idct64_stage9_sse4_1(u, &cospim32, &cospi32, &clamp_lo, &clamp_hi, &rnding,
+                         bit);
+
+    // stage 10
+    idct64_stage10_sse4_1(u, &cospim32, &cospi32, &clamp_lo, &clamp_hi, &rnding,
+                          bit);
+
+    // stage 11
+    idct64_stage11_sse4_1(u, out, do_cols, bd, out_shift, log_range);
+  }
+}
+
+static void idct64x64_low16_sse4_1(__m128i *in, __m128i *out, int bit,
+                                   int do_cols, int bd, int out_shift) {
+  int i, j;
+  const int32_t *cospi = cospi_arr(bit);
+  const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
+  const int log_range = AOMMAX(16, bd + (do_cols ? 6 : 8));
+  const __m128i clamp_lo = _mm_set1_epi32(-(1 << (log_range - 1)));
+  const __m128i clamp_hi = _mm_set1_epi32((1 << (log_range - 1)) - 1);
+
+  const __m128i cospi1 = _mm_set1_epi32(cospi[1]);
+  const __m128i cospi2 = _mm_set1_epi32(cospi[2]);
+  const __m128i cospi3 = _mm_set1_epi32(cospi[3]);
+  const __m128i cospi4 = _mm_set1_epi32(cospi[4]);
+  const __m128i cospi5 = _mm_set1_epi32(cospi[5]);
+  const __m128i cospi6 = _mm_set1_epi32(cospi[6]);
+  const __m128i cospi7 = _mm_set1_epi32(cospi[7]);
+  const __m128i cospi8 = _mm_set1_epi32(cospi[8]);
+  const __m128i cospi9 = _mm_set1_epi32(cospi[9]);
+  const __m128i cospi10 = _mm_set1_epi32(cospi[10]);
+  const __m128i cospi11 = _mm_set1_epi32(cospi[11]);
+  const __m128i cospi12 = _mm_set1_epi32(cospi[12]);
+  const __m128i cospi13 = _mm_set1_epi32(cospi[13]);
+  const __m128i cospi14 = _mm_set1_epi32(cospi[14]);
+  const __m128i cospi15 = _mm_set1_epi32(cospi[15]);
+  const __m128i cospi16 = _mm_set1_epi32(cospi[16]);
+  const __m128i cospi20 = _mm_set1_epi32(cospi[20]);
+  const __m128i cospi24 = _mm_set1_epi32(cospi[24]);
+  const __m128i cospi28 = _mm_set1_epi32(cospi[28]);
+  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
+  const __m128i cospi36 = _mm_set1_epi32(cospi[36]);
+  const __m128i cospi40 = _mm_set1_epi32(cospi[40]);
+  const __m128i cospi44 = _mm_set1_epi32(cospi[44]);
+  const __m128i cospi48 = _mm_set1_epi32(cospi[48]);
+  const __m128i cospi51 = _mm_set1_epi32(cospi[51]);
+  const __m128i cospi52 = _mm_set1_epi32(cospi[52]);
+  const __m128i cospi54 = _mm_set1_epi32(cospi[54]);
+  const __m128i cospi55 = _mm_set1_epi32(cospi[55]);
+  const __m128i cospi56 = _mm_set1_epi32(cospi[56]);
+  const __m128i cospi59 = _mm_set1_epi32(cospi[59]);
+  const __m128i cospi60 = _mm_set1_epi32(cospi[60]);
+  const __m128i cospi62 = _mm_set1_epi32(cospi[62]);
+  const __m128i cospi63 = _mm_set1_epi32(cospi[63]);
+
+  const __m128i cospim4 = _mm_set1_epi32(-cospi[4]);
+  const __m128i cospim8 = _mm_set1_epi32(-cospi[8]);
+  const __m128i cospim12 = _mm_set1_epi32(-cospi[12]);
+  const __m128i cospim16 = _mm_set1_epi32(-cospi[16]);
+  const __m128i cospim20 = _mm_set1_epi32(-cospi[20]);
+  const __m128i cospim24 = _mm_set1_epi32(-cospi[24]);
+  const __m128i cospim28 = _mm_set1_epi32(-cospi[28]);
+  const __m128i cospim32 = _mm_set1_epi32(-cospi[32]);
+  const __m128i cospim36 = _mm_set1_epi32(-cospi[36]);
+  const __m128i cospim40 = _mm_set1_epi32(-cospi[40]);
+  const __m128i cospim44 = _mm_set1_epi32(-cospi[44]);
+  const __m128i cospim48 = _mm_set1_epi32(-cospi[48]);
+  const __m128i cospim49 = _mm_set1_epi32(-cospi[49]);
+  const __m128i cospim50 = _mm_set1_epi32(-cospi[50]);
+  const __m128i cospim52 = _mm_set1_epi32(-cospi[52]);
+  const __m128i cospim53 = _mm_set1_epi32(-cospi[53]);
+  const __m128i cospim56 = _mm_set1_epi32(-cospi[56]);
+  const __m128i cospim57 = _mm_set1_epi32(-cospi[57]);
+  const __m128i cospim58 = _mm_set1_epi32(-cospi[58]);
+  const __m128i cospim60 = _mm_set1_epi32(-cospi[60]);
+  const __m128i cospim61 = _mm_set1_epi32(-cospi[61]);
+
+  {
+    __m128i u[64];
+    __m128i tmp1, tmp2, tmp3, tmp4;
+    // stage 1
+    u[0] = in[0];
+    u[32] = in[1];
+    u[36] = in[9];
+    u[40] = in[5];
+    u[44] = in[13];
+    u[48] = in[3];
+    u[52] = in[11];
+    u[56] = in[7];
+    u[60] = in[15];
+    u[16] = in[2];
+    u[20] = in[10];
+    u[24] = in[6];
+    u[28] = in[14];
+    u[4] = in[8];
+    u[8] = in[4];
+    u[12] = in[12];
+
+    // stage 2
+    u[63] = half_btf_0_sse4_1(&cospi1, &u[32], &rnding, bit);
+    u[32] = half_btf_0_sse4_1(&cospi63, &u[32], &rnding, bit);
+    u[35] = half_btf_0_sse4_1(&cospim49, &u[60], &rnding, bit);
+    u[60] = half_btf_0_sse4_1(&cospi15, &u[60], &rnding, bit);
+    u[59] = half_btf_0_sse4_1(&cospi9, &u[36], &rnding, bit);
+    u[36] = half_btf_0_sse4_1(&cospi55, &u[36], &rnding, bit);
+    u[39] = half_btf_0_sse4_1(&cospim57, &u[56], &rnding, bit);
+    u[56] = half_btf_0_sse4_1(&cospi7, &u[56], &rnding, bit);
+    u[55] = half_btf_0_sse4_1(&cospi5, &u[40], &rnding, bit);
+    u[40] = half_btf_0_sse4_1(&cospi59, &u[40], &rnding, bit);
+    u[43] = half_btf_0_sse4_1(&cospim53, &u[52], &rnding, bit);
+    u[52] = half_btf_0_sse4_1(&cospi11, &u[52], &rnding, bit);
+    u[47] = half_btf_0_sse4_1(&cospim61, &u[48], &rnding, bit);
+    u[48] = half_btf_0_sse4_1(&cospi3, &u[48], &rnding, bit);
+    u[51] = half_btf_0_sse4_1(&cospi13, &u[44], &rnding, bit);
+    u[44] = half_btf_0_sse4_1(&cospi51, &u[44], &rnding, bit);
+
+    // stage 3
+    u[31] = half_btf_0_sse4_1(&cospi2, &u[16], &rnding, bit);
+    u[16] = half_btf_0_sse4_1(&cospi62, &u[16], &rnding, bit);
+    u[19] = half_btf_0_sse4_1(&cospim50, &u[28], &rnding, bit);
+    u[28] = half_btf_0_sse4_1(&cospi14, &u[28], &rnding, bit);
+    u[27] = half_btf_0_sse4_1(&cospi10, &u[20], &rnding, bit);
+    u[20] = half_btf_0_sse4_1(&cospi54, &u[20], &rnding, bit);
+    u[23] = half_btf_0_sse4_1(&cospim58, &u[24], &rnding, bit);
+    u[24] = half_btf_0_sse4_1(&cospi6, &u[24], &rnding, bit);
+    u[33] = u[32];
+    u[34] = u[35];
+    u[37] = u[36];
+    u[38] = u[39];
+    u[41] = u[40];
+    u[42] = u[43];
+    u[45] = u[44];
+    u[46] = u[47];
+    u[49] = u[48];
+    u[50] = u[51];
+    u[53] = u[52];
+    u[54] = u[55];
+    u[57] = u[56];
+    u[58] = u[59];
+    u[61] = u[60];
+    u[62] = u[63];
+
+    // stage 4
+    u[15] = half_btf_0_sse4_1(&cospi4, &u[8], &rnding, bit);
+    u[8] = half_btf_0_sse4_1(&cospi60, &u[8], &rnding, bit);
+    u[11] = half_btf_0_sse4_1(&cospim52, &u[12], &rnding, bit);
+    u[12] = half_btf_0_sse4_1(&cospi12, &u[12], &rnding, bit);
+
+    u[17] = u[16];
+    u[18] = u[19];
+    u[21] = u[20];
+    u[22] = u[23];
+    u[25] = u[24];
+    u[26] = u[27];
+    u[29] = u[28];
+    u[30] = u[31];
+
+    tmp1 = half_btf_sse4_1(&cospim4, &u[33], &cospi60, &u[62], &rnding, bit);
+    tmp2 = half_btf_sse4_1(&cospim60, &u[34], &cospim4, &u[61], &rnding, bit);
+    tmp3 = half_btf_sse4_1(&cospim36, &u[37], &cospi28, &u[58], &rnding, bit);
+    tmp4 = half_btf_sse4_1(&cospim28, &u[38], &cospim36, &u[57], &rnding, bit);
+    u[57] = half_btf_sse4_1(&cospim36, &u[38], &cospi28, &u[57], &rnding, bit);
+    u[58] = half_btf_sse4_1(&cospi28, &u[37], &cospi36, &u[58], &rnding, bit);
+    u[61] = half_btf_sse4_1(&cospim4, &u[34], &cospi60, &u[61], &rnding, bit);
+    u[62] = half_btf_sse4_1(&cospi60, &u[33], &cospi4, &u[62], &rnding, bit);
+    u[33] = tmp1;
+    u[34] = tmp2;
+    u[37] = tmp3;
+    u[38] = tmp4;
+
+    tmp1 = half_btf_sse4_1(&cospim20, &u[41], &cospi44, &u[54], &rnding, bit);
+    tmp2 = half_btf_sse4_1(&cospim44, &u[42], &cospim20, &u[53], &rnding, bit);
+    tmp3 = half_btf_sse4_1(&cospim52, &u[45], &cospi12, &u[50], &rnding, bit);
+    tmp4 = half_btf_sse4_1(&cospim12, &u[46], &cospim52, &u[49], &rnding, bit);
+    u[49] = half_btf_sse4_1(&cospim52, &u[46], &cospi12, &u[49], &rnding, bit);
+    u[50] = half_btf_sse4_1(&cospi12, &u[45], &cospi52, &u[50], &rnding, bit);
+    u[53] = half_btf_sse4_1(&cospim20, &u[42], &cospi44, &u[53], &rnding, bit);
+    u[54] = half_btf_sse4_1(&cospi44, &u[41], &cospi20, &u[54], &rnding, bit);
+    u[41] = tmp1;
+    u[42] = tmp2;
+    u[45] = tmp3;
+    u[46] = tmp4;
+
+    // stage 5
+    u[7] = half_btf_0_sse4_1(&cospi8, &u[4], &rnding, bit);
+    u[4] = half_btf_0_sse4_1(&cospi56, &u[4], &rnding, bit);
+
+    u[9] = u[8];
+    u[10] = u[11];
+    u[13] = u[12];
+    u[14] = u[15];
+
+    tmp1 = half_btf_sse4_1(&cospim8, &u[17], &cospi56, &u[30], &rnding, bit);
+    tmp2 = half_btf_sse4_1(&cospim56, &u[18], &cospim8, &u[29], &rnding, bit);
+    tmp3 = half_btf_sse4_1(&cospim40, &u[21], &cospi24, &u[26], &rnding, bit);
+    tmp4 = half_btf_sse4_1(&cospim24, &u[22], &cospim40, &u[25], &rnding, bit);
+    u[25] = half_btf_sse4_1(&cospim40, &u[22], &cospi24, &u[25], &rnding, bit);
+    u[26] = half_btf_sse4_1(&cospi24, &u[21], &cospi40, &u[26], &rnding, bit);
+    u[29] = half_btf_sse4_1(&cospim8, &u[18], &cospi56, &u[29], &rnding, bit);
+    u[30] = half_btf_sse4_1(&cospi56, &u[17], &cospi8, &u[30], &rnding, bit);
+    u[17] = tmp1;
+    u[18] = tmp2;
+    u[21] = tmp3;
+    u[22] = tmp4;
+
+    for (i = 32; i < 64; i += 8) {
+      addsub_sse4_1(u[i + 0], u[i + 3], &u[i + 0], &u[i + 3], &clamp_lo,
+                    &clamp_hi);
+      addsub_sse4_1(u[i + 1], u[i + 2], &u[i + 1], &u[i + 2], &clamp_lo,
+                    &clamp_hi);
+
+      addsub_sse4_1(u[i + 7], u[i + 4], &u[i + 7], &u[i + 4], &clamp_lo,
+                    &clamp_hi);
+      addsub_sse4_1(u[i + 6], u[i + 5], &u[i + 6], &u[i + 5], &clamp_lo,
+                    &clamp_hi);
+    }
+
+    // stage 6
+    tmp1 = half_btf_0_sse4_1(&cospi32, &u[0], &rnding, bit);
+    u[1] = half_btf_0_sse4_1(&cospi32, &u[0], &rnding, bit);
+    u[0] = tmp1;
+    u[5] = u[4];
+    u[6] = u[7];
+
+    tmp1 = half_btf_sse4_1(&cospim16, &u[9], &cospi48, &u[14], &rnding, bit);
+    u[14] = half_btf_sse4_1(&cospi48, &u[9], &cospi16, &u[14], &rnding, bit);
+    u[9] = tmp1;
+    tmp2 = half_btf_sse4_1(&cospim48, &u[10], &cospim16, &u[13], &rnding, bit);
+    u[13] = half_btf_sse4_1(&cospim16, &u[10], &cospi48, &u[13], &rnding, bit);
+    u[10] = tmp2;
+
+    for (i = 16; i < 32; i += 8) {
+      addsub_sse4_1(u[i + 0], u[i + 3], &u[i + 0], &u[i + 3], &clamp_lo,
+                    &clamp_hi);
+      addsub_sse4_1(u[i + 1], u[i + 2], &u[i + 1], &u[i + 2], &clamp_lo,
+                    &clamp_hi);
+
+      addsub_sse4_1(u[i + 7], u[i + 4], &u[i + 7], &u[i + 4], &clamp_lo,
+                    &clamp_hi);
+      addsub_sse4_1(u[i + 6], u[i + 5], &u[i + 6], &u[i + 5], &clamp_lo,
+                    &clamp_hi);
+    }
+
+    tmp1 = half_btf_sse4_1(&cospim8, &u[34], &cospi56, &u[61], &rnding, bit);
+    tmp2 = half_btf_sse4_1(&cospim8, &u[35], &cospi56, &u[60], &rnding, bit);
+    tmp3 = half_btf_sse4_1(&cospim56, &u[36], &cospim8, &u[59], &rnding, bit);
+    tmp4 = half_btf_sse4_1(&cospim56, &u[37], &cospim8, &u[58], &rnding, bit);
+    u[58] = half_btf_sse4_1(&cospim8, &u[37], &cospi56, &u[58], &rnding, bit);
+    u[59] = half_btf_sse4_1(&cospim8, &u[36], &cospi56, &u[59], &rnding, bit);
+    u[60] = half_btf_sse4_1(&cospi56, &u[35], &cospi8, &u[60], &rnding, bit);
+    u[61] = half_btf_sse4_1(&cospi56, &u[34], &cospi8, &u[61], &rnding, bit);
+    u[34] = tmp1;
+    u[35] = tmp2;
+    u[36] = tmp3;
+    u[37] = tmp4;
+
+    tmp1 = half_btf_sse4_1(&cospim40, &u[42], &cospi24, &u[53], &rnding, bit);
+    tmp2 = half_btf_sse4_1(&cospim40, &u[43], &cospi24, &u[52], &rnding, bit);
+    tmp3 = half_btf_sse4_1(&cospim24, &u[44], &cospim40, &u[51], &rnding, bit);
+    tmp4 = half_btf_sse4_1(&cospim24, &u[45], &cospim40, &u[50], &rnding, bit);
+    u[50] = half_btf_sse4_1(&cospim40, &u[45], &cospi24, &u[50], &rnding, bit);
+    u[51] = half_btf_sse4_1(&cospim40, &u[44], &cospi24, &u[51], &rnding, bit);
+    u[52] = half_btf_sse4_1(&cospi24, &u[43], &cospi40, &u[52], &rnding, bit);
+    u[53] = half_btf_sse4_1(&cospi24, &u[42], &cospi40, &u[53], &rnding, bit);
+    u[42] = tmp1;
+    u[43] = tmp2;
+    u[44] = tmp3;
+    u[45] = tmp4;
+
+    // stage 7
+    u[3] = u[0];
+    u[2] = u[1];
+    tmp1 = half_btf_sse4_1(&cospim32, &u[5], &cospi32, &u[6], &rnding, bit);
+    u[6] = half_btf_sse4_1(&cospi32, &u[5], &cospi32, &u[6], &rnding, bit);
+    u[5] = tmp1;
+    addsub_sse4_1(u[8], u[11], &u[8], &u[11], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(u[9], u[10], &u[9], &u[10], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(u[15], u[12], &u[15], &u[12], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(u[14], u[13], &u[14], &u[13], &clamp_lo, &clamp_hi);
+
+    tmp1 = half_btf_sse4_1(&cospim16, &u[18], &cospi48, &u[29], &rnding, bit);
+    tmp2 = half_btf_sse4_1(&cospim16, &u[19], &cospi48, &u[28], &rnding, bit);
+    tmp3 = half_btf_sse4_1(&cospim48, &u[20], &cospim16, &u[27], &rnding, bit);
+    tmp4 = half_btf_sse4_1(&cospim48, &u[21], &cospim16, &u[26], &rnding, bit);
+    u[26] = half_btf_sse4_1(&cospim16, &u[21], &cospi48, &u[26], &rnding, bit);
+    u[27] = half_btf_sse4_1(&cospim16, &u[20], &cospi48, &u[27], &rnding, bit);
+    u[28] = half_btf_sse4_1(&cospi48, &u[19], &cospi16, &u[28], &rnding, bit);
+    u[29] = half_btf_sse4_1(&cospi48, &u[18], &cospi16, &u[29], &rnding, bit);
+    u[18] = tmp1;
+    u[19] = tmp2;
+    u[20] = tmp3;
+    u[21] = tmp4;
+
+    for (i = 32; i < 64; i += 16) {
+      for (j = i; j < i + 4; j++) {
+        addsub_sse4_1(u[j], u[j ^ 7], &u[j], &u[j ^ 7], &clamp_lo, &clamp_hi);
+        addsub_sse4_1(u[j ^ 15], u[j ^ 8], &u[j ^ 15], &u[j ^ 8], &clamp_lo,
+                      &clamp_hi);
+      }
+    }
+
+    // stage 8
+    for (i = 0; i < 4; ++i) {
+      addsub_sse4_1(u[i], u[7 - i], &u[i], &u[7 - i], &clamp_lo, &clamp_hi);
+    }
+
+    idct64_stage8_sse4_1(u, &cospim32, &cospi32, &cospim16, &cospi48, &cospi16,
+                         &cospim48, &clamp_lo, &clamp_hi, &rnding, bit);
+
+    // stage 9
+    idct64_stage9_sse4_1(u, &cospim32, &cospi32, &clamp_lo, &clamp_hi, &rnding,
+                         bit);
+
+    // stage 10
+    idct64_stage10_sse4_1(u, &cospim32, &cospi32, &clamp_lo, &clamp_hi, &rnding,
+                          bit);
+
+    // stage 11
+    idct64_stage11_sse4_1(u, out, do_cols, bd, out_shift, log_range);
+  }
+}
+
+static void idct64x64_sse4_1(__m128i *in, __m128i *out, int bit, int do_cols,
+                             int bd, int out_shift) {
+  int i, j;
+  const int32_t *cospi = cospi_arr(bit);
+  const __m128i rnding = _mm_set1_epi32(1 << (bit - 1));
+  const int log_range = AOMMAX(16, bd + (do_cols ? 6 : 8));
+  const __m128i clamp_lo = _mm_set1_epi32(-(1 << (log_range - 1)));
+  const __m128i clamp_hi = _mm_set1_epi32((1 << (log_range - 1)) - 1);
+
+  const __m128i cospi1 = _mm_set1_epi32(cospi[1]);
+  const __m128i cospi2 = _mm_set1_epi32(cospi[2]);
+  const __m128i cospi3 = _mm_set1_epi32(cospi[3]);
+  const __m128i cospi4 = _mm_set1_epi32(cospi[4]);
+  const __m128i cospi5 = _mm_set1_epi32(cospi[5]);
+  const __m128i cospi6 = _mm_set1_epi32(cospi[6]);
+  const __m128i cospi7 = _mm_set1_epi32(cospi[7]);
+  const __m128i cospi8 = _mm_set1_epi32(cospi[8]);
+  const __m128i cospi9 = _mm_set1_epi32(cospi[9]);
+  const __m128i cospi10 = _mm_set1_epi32(cospi[10]);
+  const __m128i cospi11 = _mm_set1_epi32(cospi[11]);
+  const __m128i cospi12 = _mm_set1_epi32(cospi[12]);
+  const __m128i cospi13 = _mm_set1_epi32(cospi[13]);
+  const __m128i cospi14 = _mm_set1_epi32(cospi[14]);
+  const __m128i cospi15 = _mm_set1_epi32(cospi[15]);
+  const __m128i cospi16 = _mm_set1_epi32(cospi[16]);
+  const __m128i cospi17 = _mm_set1_epi32(cospi[17]);
+  const __m128i cospi18 = _mm_set1_epi32(cospi[18]);
+  const __m128i cospi19 = _mm_set1_epi32(cospi[19]);
+  const __m128i cospi20 = _mm_set1_epi32(cospi[20]);
+  const __m128i cospi21 = _mm_set1_epi32(cospi[21]);
+  const __m128i cospi22 = _mm_set1_epi32(cospi[22]);
+  const __m128i cospi23 = _mm_set1_epi32(cospi[23]);
+  const __m128i cospi24 = _mm_set1_epi32(cospi[24]);
+  const __m128i cospi25 = _mm_set1_epi32(cospi[25]);
+  const __m128i cospi26 = _mm_set1_epi32(cospi[26]);
+  const __m128i cospi27 = _mm_set1_epi32(cospi[27]);
+  const __m128i cospi28 = _mm_set1_epi32(cospi[28]);
+  const __m128i cospi29 = _mm_set1_epi32(cospi[29]);
+  const __m128i cospi30 = _mm_set1_epi32(cospi[30]);
+  const __m128i cospi31 = _mm_set1_epi32(cospi[31]);
+  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
+  const __m128i cospi35 = _mm_set1_epi32(cospi[35]);
+  const __m128i cospi36 = _mm_set1_epi32(cospi[36]);
+  const __m128i cospi38 = _mm_set1_epi32(cospi[38]);
+  const __m128i cospi39 = _mm_set1_epi32(cospi[39]);
+  const __m128i cospi40 = _mm_set1_epi32(cospi[40]);
+  const __m128i cospi43 = _mm_set1_epi32(cospi[43]);
+  const __m128i cospi44 = _mm_set1_epi32(cospi[44]);
+  const __m128i cospi46 = _mm_set1_epi32(cospi[46]);
+  const __m128i cospi47 = _mm_set1_epi32(cospi[47]);
+  const __m128i cospi48 = _mm_set1_epi32(cospi[48]);
+  const __m128i cospi51 = _mm_set1_epi32(cospi[51]);
+  const __m128i cospi52 = _mm_set1_epi32(cospi[52]);
+  const __m128i cospi54 = _mm_set1_epi32(cospi[54]);
+  const __m128i cospi55 = _mm_set1_epi32(cospi[55]);
+  const __m128i cospi56 = _mm_set1_epi32(cospi[56]);
+  const __m128i cospi59 = _mm_set1_epi32(cospi[59]);
+  const __m128i cospi60 = _mm_set1_epi32(cospi[60]);
+  const __m128i cospi62 = _mm_set1_epi32(cospi[62]);
+  const __m128i cospi63 = _mm_set1_epi32(cospi[63]);
+
+  const __m128i cospim4 = _mm_set1_epi32(-cospi[4]);
+  const __m128i cospim8 = _mm_set1_epi32(-cospi[8]);
+  const __m128i cospim12 = _mm_set1_epi32(-cospi[12]);
+  const __m128i cospim16 = _mm_set1_epi32(-cospi[16]);
+  const __m128i cospim20 = _mm_set1_epi32(-cospi[20]);
+  const __m128i cospim24 = _mm_set1_epi32(-cospi[24]);
+  const __m128i cospim28 = _mm_set1_epi32(-cospi[28]);
+  const __m128i cospim32 = _mm_set1_epi32(-cospi[32]);
+  const __m128i cospim33 = _mm_set1_epi32(-cospi[33]);
+  const __m128i cospim34 = _mm_set1_epi32(-cospi[34]);
+  const __m128i cospim36 = _mm_set1_epi32(-cospi[36]);
+  const __m128i cospim37 = _mm_set1_epi32(-cospi[37]);
+  const __m128i cospim40 = _mm_set1_epi32(-cospi[40]);
+  const __m128i cospim41 = _mm_set1_epi32(-cospi[41]);
+  const __m128i cospim42 = _mm_set1_epi32(-cospi[42]);
+  const __m128i cospim44 = _mm_set1_epi32(-cospi[44]);
+  const __m128i cospim45 = _mm_set1_epi32(-cospi[45]);
+  const __m128i cospim48 = _mm_set1_epi32(-cospi[48]);
+  const __m128i cospim49 = _mm_set1_epi32(-cospi[49]);
+  const __m128i cospim50 = _mm_set1_epi32(-cospi[50]);
+  const __m128i cospim52 = _mm_set1_epi32(-cospi[52]);
+  const __m128i cospim53 = _mm_set1_epi32(-cospi[53]);
+  const __m128i cospim56 = _mm_set1_epi32(-cospi[56]);
+  const __m128i cospim57 = _mm_set1_epi32(-cospi[57]);
+  const __m128i cospim58 = _mm_set1_epi32(-cospi[58]);
+  const __m128i cospim60 = _mm_set1_epi32(-cospi[60]);
+  const __m128i cospim61 = _mm_set1_epi32(-cospi[61]);
+
+  {
+    __m128i u[64], v[64];
+
+    // stage 1
+    u[32] = in[1];
+    u[34] = in[17];
+    u[36] = in[9];
+    u[38] = in[25];
+    u[40] = in[5];
+    u[42] = in[21];
+    u[44] = in[13];
+    u[46] = in[29];
+    u[48] = in[3];
+    u[50] = in[19];
+    u[52] = in[11];
+    u[54] = in[27];
+    u[56] = in[7];
+    u[58] = in[23];
+    u[60] = in[15];
+    u[62] = in[31];
+
+    v[16] = in[2];
+    v[18] = in[18];
+    v[20] = in[10];
+    v[22] = in[26];
+    v[24] = in[6];
+    v[26] = in[22];
+    v[28] = in[14];
+    v[30] = in[30];
+
+    u[8] = in[4];
+    u[10] = in[20];
+    u[12] = in[12];
+    u[14] = in[28];
+
+    v[4] = in[8];
+    v[6] = in[24];
+
+    u[0] = in[0];
+    u[2] = in[16];
+
+    // stage 2
+    v[32] = half_btf_0_sse4_1(&cospi63, &u[32], &rnding, bit);
+    v[33] = half_btf_0_sse4_1(&cospim33, &u[62], &rnding, bit);
+    v[34] = half_btf_0_sse4_1(&cospi47, &u[34], &rnding, bit);
+    v[35] = half_btf_0_sse4_1(&cospim49, &u[60], &rnding, bit);
+    v[36] = half_btf_0_sse4_1(&cospi55, &u[36], &rnding, bit);
+    v[37] = half_btf_0_sse4_1(&cospim41, &u[58], &rnding, bit);
+    v[38] = half_btf_0_sse4_1(&cospi39, &u[38], &rnding, bit);
+    v[39] = half_btf_0_sse4_1(&cospim57, &u[56], &rnding, bit);
+    v[40] = half_btf_0_sse4_1(&cospi59, &u[40], &rnding, bit);
+    v[41] = half_btf_0_sse4_1(&cospim37, &u[54], &rnding, bit);
+    v[42] = half_btf_0_sse4_1(&cospi43, &u[42], &rnding, bit);
+    v[43] = half_btf_0_sse4_1(&cospim53, &u[52], &rnding, bit);
+    v[44] = half_btf_0_sse4_1(&cospi51, &u[44], &rnding, bit);
+    v[45] = half_btf_0_sse4_1(&cospim45, &u[50], &rnding, bit);
+    v[46] = half_btf_0_sse4_1(&cospi35, &u[46], &rnding, bit);
+    v[47] = half_btf_0_sse4_1(&cospim61, &u[48], &rnding, bit);
+    v[48] = half_btf_0_sse4_1(&cospi3, &u[48], &rnding, bit);
+    v[49] = half_btf_0_sse4_1(&cospi29, &u[46], &rnding, bit);
+    v[50] = half_btf_0_sse4_1(&cospi19, &u[50], &rnding, bit);
+    v[51] = half_btf_0_sse4_1(&cospi13, &u[44], &rnding, bit);
+    v[52] = half_btf_0_sse4_1(&cospi11, &u[52], &rnding, bit);
+    v[53] = half_btf_0_sse4_1(&cospi21, &u[42], &rnding, bit);
+    v[54] = half_btf_0_sse4_1(&cospi27, &u[54], &rnding, bit);
+    v[55] = half_btf_0_sse4_1(&cospi5, &u[40], &rnding, bit);
+    v[56] = half_btf_0_sse4_1(&cospi7, &u[56], &rnding, bit);
+    v[57] = half_btf_0_sse4_1(&cospi25, &u[38], &rnding, bit);
+    v[58] = half_btf_0_sse4_1(&cospi23, &u[58], &rnding, bit);
+    v[59] = half_btf_0_sse4_1(&cospi9, &u[36], &rnding, bit);
+    v[60] = half_btf_0_sse4_1(&cospi15, &u[60], &rnding, bit);
+    v[61] = half_btf_0_sse4_1(&cospi17, &u[34], &rnding, bit);
+    v[62] = half_btf_0_sse4_1(&cospi31, &u[62], &rnding, bit);
+    v[63] = half_btf_0_sse4_1(&cospi1, &u[32], &rnding, bit);
+
+    // stage 3
+    u[16] = half_btf_0_sse4_1(&cospi62, &v[16], &rnding, bit);
+    u[17] = half_btf_0_sse4_1(&cospim34, &v[30], &rnding, bit);
+    u[18] = half_btf_0_sse4_1(&cospi46, &v[18], &rnding, bit);
+    u[19] = half_btf_0_sse4_1(&cospim50, &v[28], &rnding, bit);
+    u[20] = half_btf_0_sse4_1(&cospi54, &v[20], &rnding, bit);
+    u[21] = half_btf_0_sse4_1(&cospim42, &v[26], &rnding, bit);
+    u[22] = half_btf_0_sse4_1(&cospi38, &v[22], &rnding, bit);
+    u[23] = half_btf_0_sse4_1(&cospim58, &v[24], &rnding, bit);
+    u[24] = half_btf_0_sse4_1(&cospi6, &v[24], &rnding, bit);
+    u[25] = half_btf_0_sse4_1(&cospi26, &v[22], &rnding, bit);
+    u[26] = half_btf_0_sse4_1(&cospi22, &v[26], &rnding, bit);
+    u[27] = half_btf_0_sse4_1(&cospi10, &v[20], &rnding, bit);
+    u[28] = half_btf_0_sse4_1(&cospi14, &v[28], &rnding, bit);
+    u[29] = half_btf_0_sse4_1(&cospi18, &v[18], &rnding, bit);
+    u[30] = half_btf_0_sse4_1(&cospi30, &v[30], &rnding, bit);
+    u[31] = half_btf_0_sse4_1(&cospi2, &v[16], &rnding, bit);
+
+    for (i = 32; i < 64; i += 4) {
+      addsub_sse4_1(v[i + 0], v[i + 1], &u[i + 0], &u[i + 1], &clamp_lo,
+                    &clamp_hi);
+      addsub_sse4_1(v[i + 3], v[i + 2], &u[i + 3], &u[i + 2], &clamp_lo,
+                    &clamp_hi);
+    }
+
+    // stage 4
+    v[8] = half_btf_0_sse4_1(&cospi60, &u[8], &rnding, bit);
+    v[9] = half_btf_0_sse4_1(&cospim36, &u[14], &rnding, bit);
+    v[10] = half_btf_0_sse4_1(&cospi44, &u[10], &rnding, bit);
+    v[11] = half_btf_0_sse4_1(&cospim52, &u[12], &rnding, bit);
+    v[12] = half_btf_0_sse4_1(&cospi12, &u[12], &rnding, bit);
+    v[13] = half_btf_0_sse4_1(&cospi20, &u[10], &rnding, bit);
+    v[14] = half_btf_0_sse4_1(&cospi28, &u[14], &rnding, bit);
+    v[15] = half_btf_0_sse4_1(&cospi4, &u[8], &rnding, bit);
+
+    for (i = 16; i < 32; i += 4) {
+      addsub_sse4_1(u[i + 0], u[i + 1], &v[i + 0], &v[i + 1], &clamp_lo,
+                    &clamp_hi);
+      addsub_sse4_1(u[i + 3], u[i + 2], &v[i + 3], &v[i + 2], &clamp_lo,
+                    &clamp_hi);
+    }
+
+    for (i = 32; i < 64; i += 4) {
+      v[i + 0] = u[i + 0];
+      v[i + 3] = u[i + 3];
+    }
+
+    v[33] = half_btf_sse4_1(&cospim4, &u[33], &cospi60, &u[62], &rnding, bit);
+    v[34] = half_btf_sse4_1(&cospim60, &u[34], &cospim4, &u[61], &rnding, bit);
+    v[37] = half_btf_sse4_1(&cospim36, &u[37], &cospi28, &u[58], &rnding, bit);
+    v[38] = half_btf_sse4_1(&cospim28, &u[38], &cospim36, &u[57], &rnding, bit);
+    v[41] = half_btf_sse4_1(&cospim20, &u[41], &cospi44, &u[54], &rnding, bit);
+    v[42] = half_btf_sse4_1(&cospim44, &u[42], &cospim20, &u[53], &rnding, bit);
+    v[45] = half_btf_sse4_1(&cospim52, &u[45], &cospi12, &u[50], &rnding, bit);
+    v[46] = half_btf_sse4_1(&cospim12, &u[46], &cospim52, &u[49], &rnding, bit);
+    v[49] = half_btf_sse4_1(&cospim52, &u[46], &cospi12, &u[49], &rnding, bit);
+    v[50] = half_btf_sse4_1(&cospi12, &u[45], &cospi52, &u[50], &rnding, bit);
+    v[53] = half_btf_sse4_1(&cospim20, &u[42], &cospi44, &u[53], &rnding, bit);
+    v[54] = half_btf_sse4_1(&cospi44, &u[41], &cospi20, &u[54], &rnding, bit);
+    v[57] = half_btf_sse4_1(&cospim36, &u[38], &cospi28, &u[57], &rnding, bit);
+    v[58] = half_btf_sse4_1(&cospi28, &u[37], &cospi36, &u[58], &rnding, bit);
+    v[61] = half_btf_sse4_1(&cospim4, &u[34], &cospi60, &u[61], &rnding, bit);
+    v[62] = half_btf_sse4_1(&cospi60, &u[33], &cospi4, &u[62], &rnding, bit);
+
+    // stage 5
+    u[4] = half_btf_0_sse4_1(&cospi56, &v[4], &rnding, bit);
+    u[5] = half_btf_0_sse4_1(&cospim40, &v[6], &rnding, bit);
+    u[6] = half_btf_0_sse4_1(&cospi24, &v[6], &rnding, bit);
+    u[7] = half_btf_0_sse4_1(&cospi8, &v[4], &rnding, bit);
+
+    for (i = 8; i < 16; i += 4) {
+      addsub_sse4_1(v[i + 0], v[i + 1], &u[i + 0], &u[i + 1], &clamp_lo,
+                    &clamp_hi);
+      addsub_sse4_1(v[i + 3], v[i + 2], &u[i + 3], &u[i + 2], &clamp_lo,
+                    &clamp_hi);
+    }
+
+    for (i = 16; i < 32; i += 4) {
+      u[i + 0] = v[i + 0];
+      u[i + 3] = v[i + 3];
+    }
+
+    u[17] = half_btf_sse4_1(&cospim8, &v[17], &cospi56, &v[30], &rnding, bit);
+    u[18] = half_btf_sse4_1(&cospim56, &v[18], &cospim8, &v[29], &rnding, bit);
+    u[21] = half_btf_sse4_1(&cospim40, &v[21], &cospi24, &v[26], &rnding, bit);
+    u[22] = half_btf_sse4_1(&cospim24, &v[22], &cospim40, &v[25], &rnding, bit);
+    u[25] = half_btf_sse4_1(&cospim40, &v[22], &cospi24, &v[25], &rnding, bit);
+    u[26] = half_btf_sse4_1(&cospi24, &v[21], &cospi40, &v[26], &rnding, bit);
+    u[29] = half_btf_sse4_1(&cospim8, &v[18], &cospi56, &v[29], &rnding, bit);
+    u[30] = half_btf_sse4_1(&cospi56, &v[17], &cospi8, &v[30], &rnding, bit);
+
+    for (i = 32; i < 64; i += 8) {
+      addsub_sse4_1(v[i + 0], v[i + 3], &u[i + 0], &u[i + 3], &clamp_lo,
+                    &clamp_hi);
+      addsub_sse4_1(v[i + 1], v[i + 2], &u[i + 1], &u[i + 2], &clamp_lo,
+                    &clamp_hi);
+
+      addsub_sse4_1(v[i + 7], v[i + 4], &u[i + 7], &u[i + 4], &clamp_lo,
+                    &clamp_hi);
+      addsub_sse4_1(v[i + 6], v[i + 5], &u[i + 6], &u[i + 5], &clamp_lo,
+                    &clamp_hi);
+    }
+
+    // stage 6
+    v[0] = half_btf_0_sse4_1(&cospi32, &u[0], &rnding, bit);
+    v[1] = half_btf_0_sse4_1(&cospi32, &u[0], &rnding, bit);
+    v[2] = half_btf_0_sse4_1(&cospi48, &u[2], &rnding, bit);
+    v[3] = half_btf_0_sse4_1(&cospi16, &u[2], &rnding, bit);
+
+    addsub_sse4_1(u[4], u[5], &v[4], &v[5], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(u[7], u[6], &v[7], &v[6], &clamp_lo, &clamp_hi);
+
+    for (i = 8; i < 16; i += 4) {
+      v[i + 0] = u[i + 0];
+      v[i + 3] = u[i + 3];
+    }
+
+    v[9] = half_btf_sse4_1(&cospim16, &u[9], &cospi48, &u[14], &rnding, bit);
+    v[10] = half_btf_sse4_1(&cospim48, &u[10], &cospim16, &u[13], &rnding, bit);
+    v[13] = half_btf_sse4_1(&cospim16, &u[10], &cospi48, &u[13], &rnding, bit);
+    v[14] = half_btf_sse4_1(&cospi48, &u[9], &cospi16, &u[14], &rnding, bit);
+
+    for (i = 16; i < 32; i += 8) {
+      addsub_sse4_1(u[i + 0], u[i + 3], &v[i + 0], &v[i + 3], &clamp_lo,
+                    &clamp_hi);
+      addsub_sse4_1(u[i + 1], u[i + 2], &v[i + 1], &v[i + 2], &clamp_lo,
+                    &clamp_hi);
+
+      addsub_sse4_1(u[i + 7], u[i + 4], &v[i + 7], &v[i + 4], &clamp_lo,
+                    &clamp_hi);
+      addsub_sse4_1(u[i + 6], u[i + 5], &v[i + 6], &v[i + 5], &clamp_lo,
+                    &clamp_hi);
+    }
+
+    for (i = 32; i < 64; i += 8) {
+      v[i + 0] = u[i + 0];
+      v[i + 1] = u[i + 1];
+      v[i + 6] = u[i + 6];
+      v[i + 7] = u[i + 7];
+    }
+
+    v[34] = half_btf_sse4_1(&cospim8, &u[34], &cospi56, &u[61], &rnding, bit);
+    v[35] = half_btf_sse4_1(&cospim8, &u[35], &cospi56, &u[60], &rnding, bit);
+    v[36] = half_btf_sse4_1(&cospim56, &u[36], &cospim8, &u[59], &rnding, bit);
+    v[37] = half_btf_sse4_1(&cospim56, &u[37], &cospim8, &u[58], &rnding, bit);
+    v[42] = half_btf_sse4_1(&cospim40, &u[42], &cospi24, &u[53], &rnding, bit);
+    v[43] = half_btf_sse4_1(&cospim40, &u[43], &cospi24, &u[52], &rnding, bit);
+    v[44] = half_btf_sse4_1(&cospim24, &u[44], &cospim40, &u[51], &rnding, bit);
+    v[45] = half_btf_sse4_1(&cospim24, &u[45], &cospim40, &u[50], &rnding, bit);
+    v[50] = half_btf_sse4_1(&cospim40, &u[45], &cospi24, &u[50], &rnding, bit);
+    v[51] = half_btf_sse4_1(&cospim40, &u[44], &cospi24, &u[51], &rnding, bit);
+    v[52] = half_btf_sse4_1(&cospi24, &u[43], &cospi40, &u[52], &rnding, bit);
+    v[53] = half_btf_sse4_1(&cospi24, &u[42], &cospi40, &u[53], &rnding, bit);
+    v[58] = half_btf_sse4_1(&cospim8, &u[37], &cospi56, &u[58], &rnding, bit);
+    v[59] = half_btf_sse4_1(&cospim8, &u[36], &cospi56, &u[59], &rnding, bit);
+    v[60] = half_btf_sse4_1(&cospi56, &u[35], &cospi8, &u[60], &rnding, bit);
+    v[61] = half_btf_sse4_1(&cospi56, &u[34], &cospi8, &u[61], &rnding, bit);
+
+    // stage 7
+    addsub_sse4_1(v[0], v[3], &u[0], &u[3], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(v[1], v[2], &u[1], &u[2], &clamp_lo, &clamp_hi);
+
+    u[4] = v[4];
+    u[7] = v[7];
+    u[5] = half_btf_sse4_1(&cospim32, &v[5], &cospi32, &v[6], &rnding, bit);
+    u[6] = half_btf_sse4_1(&cospi32, &v[5], &cospi32, &v[6], &rnding, bit);
+
+    addsub_sse4_1(v[8], v[11], &u[8], &u[11], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(v[9], v[10], &u[9], &u[10], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(v[15], v[12], &u[15], &u[12], &clamp_lo, &clamp_hi);
+    addsub_sse4_1(v[14], v[13], &u[14], &u[13], &clamp_lo, &clamp_hi);
+
+    for (i = 16; i < 32; i += 8) {
+      u[i + 0] = v[i + 0];
+      u[i + 1] = v[i + 1];
+      u[i + 6] = v[i + 6];
+      u[i + 7] = v[i + 7];
+    }
+
+    u[18] = half_btf_sse4_1(&cospim16, &v[18], &cospi48, &v[29], &rnding, bit);
+    u[19] = half_btf_sse4_1(&cospim16, &v[19], &cospi48, &v[28], &rnding, bit);
+    u[20] = half_btf_sse4_1(&cospim48, &v[20], &cospim16, &v[27], &rnding, bit);
+    u[21] = half_btf_sse4_1(&cospim48, &v[21], &cospim16, &v[26], &rnding, bit);
+    u[26] = half_btf_sse4_1(&cospim16, &v[21], &cospi48, &v[26], &rnding, bit);
+    u[27] = half_btf_sse4_1(&cospim16, &v[20], &cospi48, &v[27], &rnding, bit);
+    u[28] = half_btf_sse4_1(&cospi48, &v[19], &cospi16, &v[28], &rnding, bit);
+    u[29] = half_btf_sse4_1(&cospi48, &v[18], &cospi16, &v[29], &rnding, bit);
+
+    for (i = 32; i < 64; i += 16) {
+      for (j = i; j < i + 4; j++) {
+        addsub_sse4_1(v[j], v[j ^ 7], &u[j], &u[j ^ 7], &clamp_lo, &clamp_hi);
+        addsub_sse4_1(v[j ^ 15], v[j ^ 8], &u[j ^ 15], &u[j ^ 8], &clamp_lo,
+                      &clamp_hi);
+      }
+    }
+
+    // stage 8
+    for (i = 0; i < 4; ++i) {
+      addsub_sse4_1(u[i], u[7 - i], &v[i], &v[7 - i], &clamp_lo, &clamp_hi);
+    }
+
+    v[8] = u[8];
+    v[9] = u[9];
+    v[14] = u[14];
+    v[15] = u[15];
+
+    v[10] = half_btf_sse4_1(&cospim32, &u[10], &cospi32, &u[13], &rnding, bit);
+    v[11] = half_btf_sse4_1(&cospim32, &u[11], &cospi32, &u[12], &rnding, bit);
+    v[12] = half_btf_sse4_1(&cospi32, &u[11], &cospi32, &u[12], &rnding, bit);
+    v[13] = half_btf_sse4_1(&cospi32, &u[10], &cospi32, &u[13], &rnding, bit);
+
+    for (i = 16; i < 20; ++i) {
+      addsub_sse4_1(u[i], u[i ^ 7], &v[i], &v[i ^ 7], &clamp_lo, &clamp_hi);
+      addsub_sse4_1(u[i ^ 15], u[i ^ 8], &v[i ^ 15], &v[i ^ 8], &clamp_lo,
+                    &clamp_hi);
+    }
+
+    for (i = 32; i < 36; ++i) {
+      v[i] = u[i];
+      v[i + 12] = u[i + 12];
+      v[i + 16] = u[i + 16];
+      v[i + 28] = u[i + 28];
+    }
+
+    v[36] = half_btf_sse4_1(&cospim16, &u[36], &cospi48, &u[59], &rnding, bit);
+    v[37] = half_btf_sse4_1(&cospim16, &u[37], &cospi48, &u[58], &rnding, bit);
+    v[38] = half_btf_sse4_1(&cospim16, &u[38], &cospi48, &u[57], &rnding, bit);
+    v[39] = half_btf_sse4_1(&cospim16, &u[39], &cospi48, &u[56], &rnding, bit);
+    v[40] = half_btf_sse4_1(&cospim48, &u[40], &cospim16, &u[55], &rnding, bit);
+    v[41] = half_btf_sse4_1(&cospim48, &u[41], &cospim16, &u[54], &rnding, bit);
+    v[42] = half_btf_sse4_1(&cospim48, &u[42], &cospim16, &u[53], &rnding, bit);
+    v[43] = half_btf_sse4_1(&cospim48, &u[43], &cospim16, &u[52], &rnding, bit);
+    v[52] = half_btf_sse4_1(&cospim16, &u[43], &cospi48, &u[52], &rnding, bit);
+    v[53] = half_btf_sse4_1(&cospim16, &u[42], &cospi48, &u[53], &rnding, bit);
+    v[54] = half_btf_sse4_1(&cospim16, &u[41], &cospi48, &u[54], &rnding, bit);
+    v[55] = half_btf_sse4_1(&cospim16, &u[40], &cospi48, &u[55], &rnding, bit);
+    v[56] = half_btf_sse4_1(&cospi48, &u[39], &cospi16, &u[56], &rnding, bit);
+    v[57] = half_btf_sse4_1(&cospi48, &u[38], &cospi16, &u[57], &rnding, bit);
+    v[58] = half_btf_sse4_1(&cospi48, &u[37], &cospi16, &u[58], &rnding, bit);
+    v[59] = half_btf_sse4_1(&cospi48, &u[36], &cospi16, &u[59], &rnding, bit);
+
+    // stage 9
+    for (i = 0; i < 8; ++i) {
+      addsub_sse4_1(v[i], v[15 - i], &u[i], &u[15 - i], &clamp_lo, &clamp_hi);
+    }
+
+    for (i = 16; i < 20; ++i) {
+      u[i] = v[i];
+      u[i + 12] = v[i + 12];
+    }
+
+    u[20] = half_btf_sse4_1(&cospim32, &v[20], &cospi32, &v[27], &rnding, bit);
+    u[21] = half_btf_sse4_1(&cospim32, &v[21], &cospi32, &v[26], &rnding, bit);
+    u[22] = half_btf_sse4_1(&cospim32, &v[22], &cospi32, &v[25], &rnding, bit);
+    u[23] = half_btf_sse4_1(&cospim32, &v[23], &cospi32, &v[24], &rnding, bit);
+    u[24] = half_btf_sse4_1(&cospi32, &v[23], &cospi32, &v[24], &rnding, bit);
+    u[25] = half_btf_sse4_1(&cospi32, &v[22], &cospi32, &v[25], &rnding, bit);
+    u[26] = half_btf_sse4_1(&cospi32, &v[21], &cospi32, &v[26], &rnding, bit);
+    u[27] = half_btf_sse4_1(&cospi32, &v[20], &cospi32, &v[27], &rnding, bit);
+
+    for (i = 32; i < 40; i++) {
+      addsub_sse4_1(v[i], v[i ^ 15], &u[i], &u[i ^ 15], &clamp_lo, &clamp_hi);
+    }
+
+    for (i = 48; i < 56; i++) {
+      addsub_sse4_1(v[i ^ 15], v[i], &u[i ^ 15], &u[i], &clamp_lo, &clamp_hi);
+    }
+
+    // stage 10
+    for (i = 0; i < 16; i++) {
+      addsub_sse4_1(u[i], u[31 - i], &v[i], &v[31 - i], &clamp_lo, &clamp_hi);
+    }
+
+    for (i = 32; i < 40; i++) v[i] = u[i];
+
+    v[40] = half_btf_sse4_1(&cospim32, &u[40], &cospi32, &u[55], &rnding, bit);
+    v[41] = half_btf_sse4_1(&cospim32, &u[41], &cospi32, &u[54], &rnding, bit);
+    v[42] = half_btf_sse4_1(&cospim32, &u[42], &cospi32, &u[53], &rnding, bit);
+    v[43] = half_btf_sse4_1(&cospim32, &u[43], &cospi32, &u[52], &rnding, bit);
+    v[44] = half_btf_sse4_1(&cospim32, &u[44], &cospi32, &u[51], &rnding, bit);
+    v[45] = half_btf_sse4_1(&cospim32, &u[45], &cospi32, &u[50], &rnding, bit);
+    v[46] = half_btf_sse4_1(&cospim32, &u[46], &cospi32, &u[49], &rnding, bit);
+    v[47] = half_btf_sse4_1(&cospim32, &u[47], &cospi32, &u[48], &rnding, bit);
+    v[48] = half_btf_sse4_1(&cospi32, &u[47], &cospi32, &u[48], &rnding, bit);
+    v[49] = half_btf_sse4_1(&cospi32, &u[46], &cospi32, &u[49], &rnding, bit);
+    v[50] = half_btf_sse4_1(&cospi32, &u[45], &cospi32, &u[50], &rnding, bit);
+    v[51] = half_btf_sse4_1(&cospi32, &u[44], &cospi32, &u[51], &rnding, bit);
+    v[52] = half_btf_sse4_1(&cospi32, &u[43], &cospi32, &u[52], &rnding, bit);
+    v[53] = half_btf_sse4_1(&cospi32, &u[42], &cospi32, &u[53], &rnding, bit);
+    v[54] = half_btf_sse4_1(&cospi32, &u[41], &cospi32, &u[54], &rnding, bit);
+    v[55] = half_btf_sse4_1(&cospi32, &u[40], &cospi32, &u[55], &rnding, bit);
+
+    for (i = 56; i < 64; i++) v[i] = u[i];
+
+    // stage 11
+    if (do_cols) {
+      for (i = 0; i < 32; i++) {
+        addsub_no_clamp_sse4_1(v[i], v[63 - i], &out[(i)], &out[(63 - i)]);
+      }
+    } else {
+      const int log_range_out = AOMMAX(16, bd + 6);
+      const __m128i clamp_lo_out = _mm_set1_epi32(AOMMAX(
+          -(1 << (log_range_out - 1)), -(1 << (log_range - 1 - out_shift))));
+      const __m128i clamp_hi_out = _mm_set1_epi32(AOMMIN(
+          (1 << (log_range_out - 1)) - 1, (1 << (log_range - 1 - out_shift))));
+
+      for (i = 0; i < 32; i++) {
+        addsub_shift_sse4_1(v[i], v[63 - i], &out[(i)], &out[(63 - i)],
+                            &clamp_lo_out, &clamp_hi_out, out_shift);
+      }
+    }
+  }
+}
+
+static void idct32x32_low1_sse4_1(__m128i *in, __m128i *out, int bit,
+                                  int do_cols, int bd, int out_shift) {
+  const int32_t *cospi = cospi_arr(bit);
+  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
+  const __m128i rounding = _mm_set1_epi32(1 << (bit - 1));
+  const int log_range = AOMMAX(16, bd + (do_cols ? 6 : 8));
+  const __m128i clamp_lo = _mm_set1_epi32(-(1 << (log_range - 1)));
+  const __m128i clamp_hi = _mm_set1_epi32((1 << (log_range - 1)) - 1);
+  __m128i bf1;
+
+  // stage 0
+  // stage 1
+  bf1 = in[0];
+
+  // stage 2
+  // stage 3
+  // stage 4
+  // stage 5
+  bf1 = half_btf_0_sse4_1(&cospi32, &bf1, &rounding, bit);
+
+  // stage 6
+  // stage 7
+  // stage 8
+  // stage 9
+  if (do_cols) {
+    bf1 = _mm_max_epi32(bf1, clamp_lo);
+    bf1 = _mm_min_epi32(bf1, clamp_hi);
+  } else {
+    const int log_range_out = AOMMAX(16, bd + 6);
+    const __m128i clamp_lo_out = _mm_set1_epi32(AOMMAX(
+        -(1 << (log_range_out - 1)), -(1 << (log_range - 1 - out_shift))));
+    const __m128i clamp_hi_out = _mm_set1_epi32(AOMMIN(
+        (1 << (log_range_out - 1)) - 1, (1 << (log_range - 1 - out_shift))));
+
+    __m128i offset = _mm_set1_epi32((1 << out_shift) >> 1);
+    bf1 = _mm_add_epi32(bf1, offset);
+    bf1 = _mm_sra_epi32(bf1, _mm_cvtsi32_si128(out_shift));
+    bf1 = _mm_max_epi32(bf1, clamp_lo_out);
+    bf1 = _mm_min_epi32(bf1, clamp_hi_out);
+  }
+  out[0] = bf1;
+  out[1] = bf1;
+  out[2] = bf1;
+  out[3] = bf1;
+  out[4] = bf1;
+  out[5] = bf1;
+  out[6] = bf1;
+  out[7] = bf1;
+  out[8] = bf1;
+  out[9] = bf1;
+  out[10] = bf1;
+  out[11] = bf1;
+  out[12] = bf1;
+  out[13] = bf1;
+  out[14] = bf1;
+  out[15] = bf1;
+  out[16] = bf1;
+  out[17] = bf1;
+  out[18] = bf1;
+  out[19] = bf1;
+  out[20] = bf1;
+  out[21] = bf1;
+  out[22] = bf1;
+  out[23] = bf1;
+  out[24] = bf1;
+  out[25] = bf1;
+  out[26] = bf1;
+  out[27] = bf1;
+  out[28] = bf1;
+  out[29] = bf1;
+  out[30] = bf1;
+  out[31] = bf1;
+}
+
+static void idct32x32_low8_sse4_1(__m128i *in, __m128i *out, int bit,
+                                  int do_cols, int bd, int out_shift) {
+  const int32_t *cospi = cospi_arr(bit);
+  const __m128i cospi62 = _mm_set1_epi32(cospi[62]);
+  const __m128i cospi14 = _mm_set1_epi32(cospi[14]);
+  const __m128i cospi54 = _mm_set1_epi32(cospi[54]);
+  const __m128i cospi6 = _mm_set1_epi32(cospi[6]);
+  const __m128i cospi10 = _mm_set1_epi32(cospi[10]);
+  const __m128i cospi2 = _mm_set1_epi32(cospi[2]);
+  const __m128i cospim58 = _mm_set1_epi32(-cospi[58]);
+  const __m128i cospim50 = _mm_set1_epi32(-cospi[50]);
+  const __m128i cospi60 = _mm_set1_epi32(cospi[60]);
+  const __m128i cospi12 = _mm_set1_epi32(cospi[12]);
+  const __m128i cospi4 = _mm_set1_epi32(cospi[4]);
+  const __m128i cospim52 = _mm_set1_epi32(-cospi[52]);
+  const __m128i cospi56 = _mm_set1_epi32(cospi[56]);
+  const __m128i cospi24 = _mm_set1_epi32(cospi[24]);
+  const __m128i cospi40 = _mm_set1_epi32(cospi[40]);
+  const __m128i cospi8 = _mm_set1_epi32(cospi[8]);
+  const __m128i cospim40 = _mm_set1_epi32(-cospi[40]);
+  const __m128i cospim8 = _mm_set1_epi32(-cospi[8]);
+  const __m128i cospim56 = _mm_set1_epi32(-cospi[56]);
+  const __m128i cospim24 = _mm_set1_epi32(-cospi[24]);
+  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
+  const __m128i cospim32 = _mm_set1_epi32(-cospi[32]);
+  const __m128i cospi48 = _mm_set1_epi32(cospi[48]);
+  const __m128i cospim48 = _mm_set1_epi32(-cospi[48]);
+  const __m128i cospi16 = _mm_set1_epi32(cospi[16]);
+  const __m128i cospim16 = _mm_set1_epi32(-cospi[16]);
+  const __m128i rounding = _mm_set1_epi32(1 << (bit - 1));
+  const int log_range = AOMMAX(16, bd + (do_cols ? 6 : 8));
+  const __m128i clamp_lo = _mm_set1_epi32(-(1 << (log_range - 1)));
+  const __m128i clamp_hi = _mm_set1_epi32((1 << (log_range - 1)) - 1);
+  __m128i bf1[32];
+
+  // stage 0
+  // stage 1
+  bf1[0] = in[0];
+  bf1[4] = in[4];
+  bf1[8] = in[2];
+  bf1[12] = in[6];
+  bf1[16] = in[1];
+  bf1[20] = in[5];
+  bf1[24] = in[3];
+  bf1[28] = in[7];
+
+  // stage 2
+  bf1[31] = half_btf_0_sse4_1(&cospi2, &bf1[16], &rounding, bit);
+  bf1[16] = half_btf_0_sse4_1(&cospi62, &bf1[16], &rounding, bit);
+  bf1[19] = half_btf_0_sse4_1(&cospim50, &bf1[28], &rounding, bit);
+  bf1[28] = half_btf_0_sse4_1(&cospi14, &bf1[28], &rounding, bit);
+  bf1[27] = half_btf_0_sse4_1(&cospi10, &bf1[20], &rounding, bit);
+  bf1[20] = half_btf_0_sse4_1(&cospi54, &bf1[20], &rounding, bit);
+  bf1[23] = half_btf_0_sse4_1(&cospim58, &bf1[24], &rounding, bit);
+  bf1[24] = half_btf_0_sse4_1(&cospi6, &bf1[24], &rounding, bit);
+
+  // stage 3
+  bf1[15] = half_btf_0_sse4_1(&cospi4, &bf1[8], &rounding, bit);
+  bf1[8] = half_btf_0_sse4_1(&cospi60, &bf1[8], &rounding, bit);
+
+  bf1[11] = half_btf_0_sse4_1(&cospim52, &bf1[12], &rounding, bit);
+  bf1[12] = half_btf_0_sse4_1(&cospi12, &bf1[12], &rounding, bit);
+  bf1[17] = bf1[16];
+  bf1[18] = bf1[19];
+  bf1[21] = bf1[20];
+  bf1[22] = bf1[23];
+  bf1[25] = bf1[24];
+  bf1[26] = bf1[27];
+  bf1[29] = bf1[28];
+  bf1[30] = bf1[31];
+
+  // stage 4 :
+  bf1[7] = half_btf_0_sse4_1(&cospi8, &bf1[4], &rounding, bit);
+  bf1[4] = half_btf_0_sse4_1(&cospi56, &bf1[4], &rounding, bit);
+
+  bf1[9] = bf1[8];
+  bf1[10] = bf1[11];
+  bf1[13] = bf1[12];
+  bf1[14] = bf1[15];
+
+  idct32_stage4_sse4_1(bf1, &cospim8, &cospi56, &cospi8, &cospim56, &cospim40,
+                       &cospi24, &cospi40, &cospim24, &rounding, bit);
+
+  // stage 5
+  bf1[0] = half_btf_0_sse4_1(&cospi32, &bf1[0], &rounding, bit);
+  bf1[1] = bf1[0];
+  bf1[5] = bf1[4];
+  bf1[6] = bf1[7];
+
+  idct32_stage5_sse4_1(bf1, &cospim16, &cospi48, &cospi16, &cospim48, &clamp_lo,
+                       &clamp_hi, &rounding, bit);
+
+  // stage 6
+  bf1[3] = bf1[0];
+  bf1[2] = bf1[1];
+
+  idct32_stage6_sse4_1(bf1, &cospim32, &cospi32, &cospim16, &cospi48, &cospi16,
+                       &cospim48, &clamp_lo, &clamp_hi, &rounding, bit);
+
+  // stage 7
+  idct32_stage7_sse4_1(bf1, &cospim32, &cospi32, &clamp_lo, &clamp_hi,
+                       &rounding, bit);
+
+  // stage 8
+  idct32_stage8_sse4_1(bf1, &cospim32, &cospi32, &clamp_lo, &clamp_hi,
+                       &rounding, bit);
+
+  // stage 9
+  idct32_stage9_sse4_1(bf1, out, do_cols, bd, out_shift, log_range);
+}
+
+static void idct32x32_low16_sse4_1(__m128i *in, __m128i *out, int bit,
+                                   int do_cols, int bd, int out_shift) {
+  const int32_t *cospi = cospi_arr(bit);
+  const __m128i cospi62 = _mm_set1_epi32(cospi[62]);
+  const __m128i cospi30 = _mm_set1_epi32(cospi[30]);
+  const __m128i cospi46 = _mm_set1_epi32(cospi[46]);
+  const __m128i cospi14 = _mm_set1_epi32(cospi[14]);
+  const __m128i cospi54 = _mm_set1_epi32(cospi[54]);
+  const __m128i cospi22 = _mm_set1_epi32(cospi[22]);
+  const __m128i cospi38 = _mm_set1_epi32(cospi[38]);
+  const __m128i cospi6 = _mm_set1_epi32(cospi[6]);
+  const __m128i cospi26 = _mm_set1_epi32(cospi[26]);
+  const __m128i cospi10 = _mm_set1_epi32(cospi[10]);
+  const __m128i cospi18 = _mm_set1_epi32(cospi[18]);
+  const __m128i cospi2 = _mm_set1_epi32(cospi[2]);
+  const __m128i cospim58 = _mm_set1_epi32(-cospi[58]);
+  const __m128i cospim42 = _mm_set1_epi32(-cospi[42]);
+  const __m128i cospim50 = _mm_set1_epi32(-cospi[50]);
+  const __m128i cospim34 = _mm_set1_epi32(-cospi[34]);
+  const __m128i cospi60 = _mm_set1_epi32(cospi[60]);
+  const __m128i cospi28 = _mm_set1_epi32(cospi[28]);
+  const __m128i cospi44 = _mm_set1_epi32(cospi[44]);
+  const __m128i cospi12 = _mm_set1_epi32(cospi[12]);
+  const __m128i cospi20 = _mm_set1_epi32(cospi[20]);
+  const __m128i cospi4 = _mm_set1_epi32(cospi[4]);
+  const __m128i cospim52 = _mm_set1_epi32(-cospi[52]);
+  const __m128i cospim36 = _mm_set1_epi32(-cospi[36]);
+  const __m128i cospi56 = _mm_set1_epi32(cospi[56]);
+  const __m128i cospi24 = _mm_set1_epi32(cospi[24]);
+  const __m128i cospi40 = _mm_set1_epi32(cospi[40]);
+  const __m128i cospi8 = _mm_set1_epi32(cospi[8]);
+  const __m128i cospim40 = _mm_set1_epi32(-cospi[40]);
+  const __m128i cospim8 = _mm_set1_epi32(-cospi[8]);
+  const __m128i cospim56 = _mm_set1_epi32(-cospi[56]);
+  const __m128i cospim24 = _mm_set1_epi32(-cospi[24]);
+  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
+  const __m128i cospim32 = _mm_set1_epi32(-cospi[32]);
+  const __m128i cospi48 = _mm_set1_epi32(cospi[48]);
+  const __m128i cospim48 = _mm_set1_epi32(-cospi[48]);
+  const __m128i cospi16 = _mm_set1_epi32(cospi[16]);
+  const __m128i cospim16 = _mm_set1_epi32(-cospi[16]);
+  const __m128i rounding = _mm_set1_epi32(1 << (bit - 1));
+  const int log_range = AOMMAX(16, bd + (do_cols ? 6 : 8));
+  const __m128i clamp_lo = _mm_set1_epi32(-(1 << (log_range - 1)));
+  const __m128i clamp_hi = _mm_set1_epi32((1 << (log_range - 1)) - 1);
+  __m128i bf1[32];
+
+  // stage 0
+  // stage 1
+
+  bf1[0] = in[0];
+  bf1[2] = in[8];
+  bf1[4] = in[4];
+  bf1[6] = in[12];
+  bf1[8] = in[2];
+  bf1[10] = in[10];
+  bf1[12] = in[6];
+  bf1[14] = in[14];
+  bf1[16] = in[1];
+  bf1[18] = in[9];
+  bf1[20] = in[5];
+  bf1[22] = in[13];
+  bf1[24] = in[3];
+  bf1[26] = in[11];
+  bf1[28] = in[7];
+  bf1[30] = in[15];
+
+  // stage 2
+  bf1[31] = half_btf_0_sse4_1(&cospi2, &bf1[16], &rounding, bit);
+  bf1[16] = half_btf_0_sse4_1(&cospi62, &bf1[16], &rounding, bit);
+  bf1[17] = half_btf_0_sse4_1(&cospim34, &bf1[30], &rounding, bit);
+  bf1[30] = half_btf_0_sse4_1(&cospi30, &bf1[30], &rounding, bit);
+  bf1[29] = half_btf_0_sse4_1(&cospi18, &bf1[18], &rounding, bit);
+  bf1[18] = half_btf_0_sse4_1(&cospi46, &bf1[18], &rounding, bit);
+  bf1[19] = half_btf_0_sse4_1(&cospim50, &bf1[28], &rounding, bit);
+  bf1[28] = half_btf_0_sse4_1(&cospi14, &bf1[28], &rounding, bit);
+  bf1[27] = half_btf_0_sse4_1(&cospi10, &bf1[20], &rounding, bit);
+  bf1[20] = half_btf_0_sse4_1(&cospi54, &bf1[20], &rounding, bit);
+  bf1[21] = half_btf_0_sse4_1(&cospim42, &bf1[26], &rounding, bit);
+  bf1[26] = half_btf_0_sse4_1(&cospi22, &bf1[26], &rounding, bit);
+  bf1[25] = half_btf_0_sse4_1(&cospi26, &bf1[22], &rounding, bit);
+  bf1[22] = half_btf_0_sse4_1(&cospi38, &bf1[22], &rounding, bit);
+  bf1[23] = half_btf_0_sse4_1(&cospim58, &bf1[24], &rounding, bit);
+  bf1[24] = half_btf_0_sse4_1(&cospi6, &bf1[24], &rounding, bit);
+
+  // stage 3
+  bf1[15] = half_btf_0_sse4_1(&cospi4, &bf1[8], &rounding, bit);
+  bf1[8] = half_btf_0_sse4_1(&cospi60, &bf1[8], &rounding, bit);
+  bf1[9] = half_btf_0_sse4_1(&cospim36, &bf1[14], &rounding, bit);
+  bf1[14] = half_btf_0_sse4_1(&cospi28, &bf1[14], &rounding, bit);
+  bf1[13] = half_btf_0_sse4_1(&cospi20, &bf1[10], &rounding, bit);
+  bf1[10] = half_btf_0_sse4_1(&cospi44, &bf1[10], &rounding, bit);
+  bf1[11] = half_btf_0_sse4_1(&cospim52, &bf1[12], &rounding, bit);
+  bf1[12] = half_btf_0_sse4_1(&cospi12, &bf1[12], &rounding, bit);
+
+  addsub_sse4_1(bf1[16], bf1[17], bf1 + 16, bf1 + 17, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf1[19], bf1[18], bf1 + 19, bf1 + 18, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf1[20], bf1[21], bf1 + 20, bf1 + 21, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf1[23], bf1[22], bf1 + 23, bf1 + 22, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf1[24], bf1[25], bf1 + 24, bf1 + 25, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf1[27], bf1[26], bf1 + 27, bf1 + 26, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf1[28], bf1[29], bf1 + 28, bf1 + 29, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf1[31], bf1[30], bf1 + 31, bf1 + 30, &clamp_lo, &clamp_hi);
+  // stage 4
+  bf1[7] = half_btf_0_sse4_1(&cospi8, &bf1[4], &rounding, bit);
+  bf1[4] = half_btf_0_sse4_1(&cospi56, &bf1[4], &rounding, bit);
+  bf1[5] = half_btf_0_sse4_1(&cospim40, &bf1[6], &rounding, bit);
+  bf1[6] = half_btf_0_sse4_1(&cospi24, &bf1[6], &rounding, bit);
+
+  addsub_sse4_1(bf1[8], bf1[9], bf1 + 8, bf1 + 9, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf1[11], bf1[10], bf1 + 11, bf1 + 10, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf1[12], bf1[13], bf1 + 12, bf1 + 13, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf1[15], bf1[14], bf1 + 15, bf1 + 14, &clamp_lo, &clamp_hi);
+
+  idct32_stage4_sse4_1(bf1, &cospim8, &cospi56, &cospi8, &cospim56, &cospim40,
+                       &cospi24, &cospi40, &cospim24, &rounding, bit);
+
+  // stage 5
+  bf1[0] = half_btf_0_sse4_1(&cospi32, &bf1[0], &rounding, bit);
+  bf1[1] = bf1[0];
+  bf1[3] = half_btf_0_sse4_1(&cospi16, &bf1[2], &rounding, bit);
+  bf1[2] = half_btf_0_sse4_1(&cospi48, &bf1[2], &rounding, bit);
+
+  addsub_sse4_1(bf1[4], bf1[5], bf1 + 4, bf1 + 5, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf1[7], bf1[6], bf1 + 7, bf1 + 6, &clamp_lo, &clamp_hi);
+
+  idct32_stage5_sse4_1(bf1, &cospim16, &cospi48, &cospi16, &cospim48, &clamp_lo,
+                       &clamp_hi, &rounding, bit);
+
+  // stage 6
+  addsub_sse4_1(bf1[0], bf1[3], bf1 + 0, bf1 + 3, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf1[1], bf1[2], bf1 + 1, bf1 + 2, &clamp_lo, &clamp_hi);
+
+  idct32_stage6_sse4_1(bf1, &cospim32, &cospi32, &cospim16, &cospi48, &cospi16,
+                       &cospim48, &clamp_lo, &clamp_hi, &rounding, bit);
+
+  // stage 7
+  idct32_stage7_sse4_1(bf1, &cospim32, &cospi32, &clamp_lo, &clamp_hi,
+                       &rounding, bit);
+
+  // stage 8
+  idct32_stage8_sse4_1(bf1, &cospim32, &cospi32, &clamp_lo, &clamp_hi,
+                       &rounding, bit);
+
+  // stage 9
+  idct32_stage9_sse4_1(bf1, out, do_cols, bd, out_shift, log_range);
+}
+
+static void idct32x32_sse4_1(__m128i *in, __m128i *out, int bit, int do_cols,
+                             int bd, int out_shift) {
+  const int32_t *cospi = cospi_arr(bit);
+  const __m128i cospi62 = _mm_set1_epi32(cospi[62]);
+  const __m128i cospi30 = _mm_set1_epi32(cospi[30]);
+  const __m128i cospi46 = _mm_set1_epi32(cospi[46]);
+  const __m128i cospi14 = _mm_set1_epi32(cospi[14]);
+  const __m128i cospi54 = _mm_set1_epi32(cospi[54]);
+  const __m128i cospi22 = _mm_set1_epi32(cospi[22]);
+  const __m128i cospi38 = _mm_set1_epi32(cospi[38]);
+  const __m128i cospi6 = _mm_set1_epi32(cospi[6]);
+  const __m128i cospi58 = _mm_set1_epi32(cospi[58]);
+  const __m128i cospi26 = _mm_set1_epi32(cospi[26]);
+  const __m128i cospi42 = _mm_set1_epi32(cospi[42]);
+  const __m128i cospi10 = _mm_set1_epi32(cospi[10]);
+  const __m128i cospi50 = _mm_set1_epi32(cospi[50]);
+  const __m128i cospi18 = _mm_set1_epi32(cospi[18]);
+  const __m128i cospi34 = _mm_set1_epi32(cospi[34]);
+  const __m128i cospi2 = _mm_set1_epi32(cospi[2]);
+  const __m128i cospim58 = _mm_set1_epi32(-cospi[58]);
+  const __m128i cospim26 = _mm_set1_epi32(-cospi[26]);
+  const __m128i cospim42 = _mm_set1_epi32(-cospi[42]);
+  const __m128i cospim10 = _mm_set1_epi32(-cospi[10]);
+  const __m128i cospim50 = _mm_set1_epi32(-cospi[50]);
+  const __m128i cospim18 = _mm_set1_epi32(-cospi[18]);
+  const __m128i cospim34 = _mm_set1_epi32(-cospi[34]);
+  const __m128i cospim2 = _mm_set1_epi32(-cospi[2]);
+  const __m128i cospi60 = _mm_set1_epi32(cospi[60]);
+  const __m128i cospi28 = _mm_set1_epi32(cospi[28]);
+  const __m128i cospi44 = _mm_set1_epi32(cospi[44]);
+  const __m128i cospi12 = _mm_set1_epi32(cospi[12]);
+  const __m128i cospi52 = _mm_set1_epi32(cospi[52]);
+  const __m128i cospi20 = _mm_set1_epi32(cospi[20]);
+  const __m128i cospi36 = _mm_set1_epi32(cospi[36]);
+  const __m128i cospi4 = _mm_set1_epi32(cospi[4]);
+  const __m128i cospim52 = _mm_set1_epi32(-cospi[52]);
+  const __m128i cospim20 = _mm_set1_epi32(-cospi[20]);
+  const __m128i cospim36 = _mm_set1_epi32(-cospi[36]);
+  const __m128i cospim4 = _mm_set1_epi32(-cospi[4]);
+  const __m128i cospi56 = _mm_set1_epi32(cospi[56]);
+  const __m128i cospi24 = _mm_set1_epi32(cospi[24]);
+  const __m128i cospi40 = _mm_set1_epi32(cospi[40]);
+  const __m128i cospi8 = _mm_set1_epi32(cospi[8]);
+  const __m128i cospim40 = _mm_set1_epi32(-cospi[40]);
+  const __m128i cospim8 = _mm_set1_epi32(-cospi[8]);
+  const __m128i cospim56 = _mm_set1_epi32(-cospi[56]);
+  const __m128i cospim24 = _mm_set1_epi32(-cospi[24]);
+  const __m128i cospi32 = _mm_set1_epi32(cospi[32]);
+  const __m128i cospim32 = _mm_set1_epi32(-cospi[32]);
+  const __m128i cospi48 = _mm_set1_epi32(cospi[48]);
+  const __m128i cospim48 = _mm_set1_epi32(-cospi[48]);
+  const __m128i cospi16 = _mm_set1_epi32(cospi[16]);
+  const __m128i cospim16 = _mm_set1_epi32(-cospi[16]);
+  const __m128i rounding = _mm_set1_epi32(1 << (bit - 1));
+  const int log_range = AOMMAX(16, bd + (do_cols ? 6 : 8));
+  const __m128i clamp_lo = _mm_set1_epi32(-(1 << (log_range - 1)));
+  const __m128i clamp_hi = _mm_set1_epi32((1 << (log_range - 1)) - 1);
+  __m128i bf1[32], bf0[32];
+
+  // stage 0
+  // stage 1
+  bf1[0] = in[0];
+  bf1[1] = in[16];
+  bf1[2] = in[8];
+  bf1[3] = in[24];
+  bf1[4] = in[4];
+  bf1[5] = in[20];
+  bf1[6] = in[12];
+  bf1[7] = in[28];
+  bf1[8] = in[2];
+  bf1[9] = in[18];
+  bf1[10] = in[10];
+  bf1[11] = in[26];
+  bf1[12] = in[6];
+  bf1[13] = in[22];
+  bf1[14] = in[14];
+  bf1[15] = in[30];
+  bf1[16] = in[1];
+  bf1[17] = in[17];
+  bf1[18] = in[9];
+  bf1[19] = in[25];
+  bf1[20] = in[5];
+  bf1[21] = in[21];
+  bf1[22] = in[13];
+  bf1[23] = in[29];
+  bf1[24] = in[3];
+  bf1[25] = in[19];
+  bf1[26] = in[11];
+  bf1[27] = in[27];
+  bf1[28] = in[7];
+  bf1[29] = in[23];
+  bf1[30] = in[15];
+  bf1[31] = in[31];
+
+  // stage 2
+  bf0[0] = bf1[0];
+  bf0[1] = bf1[1];
+  bf0[2] = bf1[2];
+  bf0[3] = bf1[3];
+  bf0[4] = bf1[4];
+  bf0[5] = bf1[5];
+  bf0[6] = bf1[6];
+  bf0[7] = bf1[7];
+  bf0[8] = bf1[8];
+  bf0[9] = bf1[9];
+  bf0[10] = bf1[10];
+  bf0[11] = bf1[11];
+  bf0[12] = bf1[12];
+  bf0[13] = bf1[13];
+  bf0[14] = bf1[14];
+  bf0[15] = bf1[15];
+  bf0[16] =
+      half_btf_sse4_1(&cospi62, &bf1[16], &cospim2, &bf1[31], &rounding, bit);
+  bf0[17] =
+      half_btf_sse4_1(&cospi30, &bf1[17], &cospim34, &bf1[30], &rounding, bit);
+  bf0[18] =
+      half_btf_sse4_1(&cospi46, &bf1[18], &cospim18, &bf1[29], &rounding, bit);
+  bf0[19] =
+      half_btf_sse4_1(&cospi14, &bf1[19], &cospim50, &bf1[28], &rounding, bit);
+  bf0[20] =
+      half_btf_sse4_1(&cospi54, &bf1[20], &cospim10, &bf1[27], &rounding, bit);
+  bf0[21] =
+      half_btf_sse4_1(&cospi22, &bf1[21], &cospim42, &bf1[26], &rounding, bit);
+  bf0[22] =
+      half_btf_sse4_1(&cospi38, &bf1[22], &cospim26, &bf1[25], &rounding, bit);
+  bf0[23] =
+      half_btf_sse4_1(&cospi6, &bf1[23], &cospim58, &bf1[24], &rounding, bit);
+  bf0[24] =
+      half_btf_sse4_1(&cospi58, &bf1[23], &cospi6, &bf1[24], &rounding, bit);
+  bf0[25] =
+      half_btf_sse4_1(&cospi26, &bf1[22], &cospi38, &bf1[25], &rounding, bit);
+  bf0[26] =
+      half_btf_sse4_1(&cospi42, &bf1[21], &cospi22, &bf1[26], &rounding, bit);
+  bf0[27] =
+      half_btf_sse4_1(&cospi10, &bf1[20], &cospi54, &bf1[27], &rounding, bit);
+  bf0[28] =
+      half_btf_sse4_1(&cospi50, &bf1[19], &cospi14, &bf1[28], &rounding, bit);
+  bf0[29] =
+      half_btf_sse4_1(&cospi18, &bf1[18], &cospi46, &bf1[29], &rounding, bit);
+  bf0[30] =
+      half_btf_sse4_1(&cospi34, &bf1[17], &cospi30, &bf1[30], &rounding, bit);
+  bf0[31] =
+      half_btf_sse4_1(&cospi2, &bf1[16], &cospi62, &bf1[31], &rounding, bit);
+
+  // stage 3
+  bf1[0] = bf0[0];
+  bf1[1] = bf0[1];
+  bf1[2] = bf0[2];
+  bf1[3] = bf0[3];
+  bf1[4] = bf0[4];
+  bf1[5] = bf0[5];
+  bf1[6] = bf0[6];
+  bf1[7] = bf0[7];
+  bf1[8] =
+      half_btf_sse4_1(&cospi60, &bf0[8], &cospim4, &bf0[15], &rounding, bit);
+  bf1[9] =
+      half_btf_sse4_1(&cospi28, &bf0[9], &cospim36, &bf0[14], &rounding, bit);
+  bf1[10] =
+      half_btf_sse4_1(&cospi44, &bf0[10], &cospim20, &bf0[13], &rounding, bit);
+  bf1[11] =
+      half_btf_sse4_1(&cospi12, &bf0[11], &cospim52, &bf0[12], &rounding, bit);
+  bf1[12] =
+      half_btf_sse4_1(&cospi52, &bf0[11], &cospi12, &bf0[12], &rounding, bit);
+  bf1[13] =
+      half_btf_sse4_1(&cospi20, &bf0[10], &cospi44, &bf0[13], &rounding, bit);
+  bf1[14] =
+      half_btf_sse4_1(&cospi36, &bf0[9], &cospi28, &bf0[14], &rounding, bit);
+  bf1[15] =
+      half_btf_sse4_1(&cospi4, &bf0[8], &cospi60, &bf0[15], &rounding, bit);
+
+  addsub_sse4_1(bf0[16], bf0[17], bf1 + 16, bf1 + 17, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf0[19], bf0[18], bf1 + 19, bf1 + 18, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf0[20], bf0[21], bf1 + 20, bf1 + 21, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf0[23], bf0[22], bf1 + 23, bf1 + 22, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf0[24], bf0[25], bf1 + 24, bf1 + 25, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf0[27], bf0[26], bf1 + 27, bf1 + 26, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf0[28], bf0[29], bf1 + 28, bf1 + 29, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf0[31], bf0[30], bf1 + 31, bf1 + 30, &clamp_lo, &clamp_hi);
+
+  // stage 4
+  bf0[0] = bf1[0];
+  bf0[1] = bf1[1];
+  bf0[2] = bf1[2];
+  bf0[3] = bf1[3];
+  bf0[4] =
+      half_btf_sse4_1(&cospi56, &bf1[4], &cospim8, &bf1[7], &rounding, bit);
+  bf0[5] =
+      half_btf_sse4_1(&cospi24, &bf1[5], &cospim40, &bf1[6], &rounding, bit);
+  bf0[6] =
+      half_btf_sse4_1(&cospi40, &bf1[5], &cospi24, &bf1[6], &rounding, bit);
+  bf0[7] = half_btf_sse4_1(&cospi8, &bf1[4], &cospi56, &bf1[7], &rounding, bit);
+
+  addsub_sse4_1(bf1[8], bf1[9], bf0 + 8, bf0 + 9, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf1[11], bf1[10], bf0 + 11, bf0 + 10, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf1[12], bf1[13], bf0 + 12, bf0 + 13, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf1[15], bf1[14], bf0 + 15, bf0 + 14, &clamp_lo, &clamp_hi);
+
+  bf0[16] = bf1[16];
+  bf0[17] =
+      half_btf_sse4_1(&cospim8, &bf1[17], &cospi56, &bf1[30], &rounding, bit);
+  bf0[18] =
+      half_btf_sse4_1(&cospim56, &bf1[18], &cospim8, &bf1[29], &rounding, bit);
+  bf0[19] = bf1[19];
+  bf0[20] = bf1[20];
+  bf0[21] =
+      half_btf_sse4_1(&cospim40, &bf1[21], &cospi24, &bf1[26], &rounding, bit);
+  bf0[22] =
+      half_btf_sse4_1(&cospim24, &bf1[22], &cospim40, &bf1[25], &rounding, bit);
+  bf0[23] = bf1[23];
+  bf0[24] = bf1[24];
+  bf0[25] =
+      half_btf_sse4_1(&cospim40, &bf1[22], &cospi24, &bf1[25], &rounding, bit);
+  bf0[26] =
+      half_btf_sse4_1(&cospi24, &bf1[21], &cospi40, &bf1[26], &rounding, bit);
+  bf0[27] = bf1[27];
+  bf0[28] = bf1[28];
+  bf0[29] =
+      half_btf_sse4_1(&cospim8, &bf1[18], &cospi56, &bf1[29], &rounding, bit);
+  bf0[30] =
+      half_btf_sse4_1(&cospi56, &bf1[17], &cospi8, &bf1[30], &rounding, bit);
+  bf0[31] = bf1[31];
+
+  // stage 5
+  bf1[0] =
+      half_btf_sse4_1(&cospi32, &bf0[0], &cospi32, &bf0[1], &rounding, bit);
+  bf1[1] =
+      half_btf_sse4_1(&cospi32, &bf0[0], &cospim32, &bf0[1], &rounding, bit);
+  bf1[2] =
+      half_btf_sse4_1(&cospi48, &bf0[2], &cospim16, &bf0[3], &rounding, bit);
+  bf1[3] =
+      half_btf_sse4_1(&cospi16, &bf0[2], &cospi48, &bf0[3], &rounding, bit);
+  addsub_sse4_1(bf0[4], bf0[5], bf1 + 4, bf1 + 5, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf0[7], bf0[6], bf1 + 7, bf1 + 6, &clamp_lo, &clamp_hi);
+  bf1[8] = bf0[8];
+  bf1[9] =
+      half_btf_sse4_1(&cospim16, &bf0[9], &cospi48, &bf0[14], &rounding, bit);
+  bf1[10] =
+      half_btf_sse4_1(&cospim48, &bf0[10], &cospim16, &bf0[13], &rounding, bit);
+  bf1[11] = bf0[11];
+  bf1[12] = bf0[12];
+  bf1[13] =
+      half_btf_sse4_1(&cospim16, &bf0[10], &cospi48, &bf0[13], &rounding, bit);
+  bf1[14] =
+      half_btf_sse4_1(&cospi48, &bf0[9], &cospi16, &bf0[14], &rounding, bit);
+  bf1[15] = bf0[15];
+  addsub_sse4_1(bf0[16], bf0[19], bf1 + 16, bf1 + 19, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf0[17], bf0[18], bf1 + 17, bf1 + 18, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf0[23], bf0[20], bf1 + 23, bf1 + 20, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf0[22], bf0[21], bf1 + 22, bf1 + 21, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf0[24], bf0[27], bf1 + 24, bf1 + 27, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf0[25], bf0[26], bf1 + 25, bf1 + 26, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf0[31], bf0[28], bf1 + 31, bf1 + 28, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf0[30], bf0[29], bf1 + 30, bf1 + 29, &clamp_lo, &clamp_hi);
+
+  // stage 6
+  addsub_sse4_1(bf1[0], bf1[3], bf0 + 0, bf0 + 3, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf1[1], bf1[2], bf0 + 1, bf0 + 2, &clamp_lo, &clamp_hi);
+  bf0[4] = bf1[4];
+  bf0[5] =
+      half_btf_sse4_1(&cospim32, &bf1[5], &cospi32, &bf1[6], &rounding, bit);
+  bf0[6] =
+      half_btf_sse4_1(&cospi32, &bf1[5], &cospi32, &bf1[6], &rounding, bit);
+  bf0[7] = bf1[7];
+  addsub_sse4_1(bf1[8], bf1[11], bf0 + 8, bf0 + 11, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf1[9], bf1[10], bf0 + 9, bf0 + 10, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf1[15], bf1[12], bf0 + 15, bf0 + 12, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf1[14], bf1[13], bf0 + 14, bf0 + 13, &clamp_lo, &clamp_hi);
+  bf0[16] = bf1[16];
+  bf0[17] = bf1[17];
+  bf0[18] =
+      half_btf_sse4_1(&cospim16, &bf1[18], &cospi48, &bf1[29], &rounding, bit);
+  bf0[19] =
+      half_btf_sse4_1(&cospim16, &bf1[19], &cospi48, &bf1[28], &rounding, bit);
+  bf0[20] =
+      half_btf_sse4_1(&cospim48, &bf1[20], &cospim16, &bf1[27], &rounding, bit);
+  bf0[21] =
+      half_btf_sse4_1(&cospim48, &bf1[21], &cospim16, &bf1[26], &rounding, bit);
+  bf0[22] = bf1[22];
+  bf0[23] = bf1[23];
+  bf0[24] = bf1[24];
+  bf0[25] = bf1[25];
+  bf0[26] =
+      half_btf_sse4_1(&cospim16, &bf1[21], &cospi48, &bf1[26], &rounding, bit);
+  bf0[27] =
+      half_btf_sse4_1(&cospim16, &bf1[20], &cospi48, &bf1[27], &rounding, bit);
+  bf0[28] =
+      half_btf_sse4_1(&cospi48, &bf1[19], &cospi16, &bf1[28], &rounding, bit);
+  bf0[29] =
+      half_btf_sse4_1(&cospi48, &bf1[18], &cospi16, &bf1[29], &rounding, bit);
+  bf0[30] = bf1[30];
+  bf0[31] = bf1[31];
+
+  // stage 7
+  addsub_sse4_1(bf0[0], bf0[7], bf1 + 0, bf1 + 7, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf0[1], bf0[6], bf1 + 1, bf1 + 6, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf0[2], bf0[5], bf1 + 2, bf1 + 5, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf0[3], bf0[4], bf1 + 3, bf1 + 4, &clamp_lo, &clamp_hi);
+  bf1[8] = bf0[8];
+  bf1[9] = bf0[9];
+  bf1[10] =
+      half_btf_sse4_1(&cospim32, &bf0[10], &cospi32, &bf0[13], &rounding, bit);
+  bf1[11] =
+      half_btf_sse4_1(&cospim32, &bf0[11], &cospi32, &bf0[12], &rounding, bit);
+  bf1[12] =
+      half_btf_sse4_1(&cospi32, &bf0[11], &cospi32, &bf0[12], &rounding, bit);
+  bf1[13] =
+      half_btf_sse4_1(&cospi32, &bf0[10], &cospi32, &bf0[13], &rounding, bit);
+  bf1[14] = bf0[14];
+  bf1[15] = bf0[15];
+  addsub_sse4_1(bf0[16], bf0[23], bf1 + 16, bf1 + 23, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf0[17], bf0[22], bf1 + 17, bf1 + 22, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf0[18], bf0[21], bf1 + 18, bf1 + 21, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf0[19], bf0[20], bf1 + 19, bf1 + 20, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf0[31], bf0[24], bf1 + 31, bf1 + 24, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf0[30], bf0[25], bf1 + 30, bf1 + 25, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf0[29], bf0[26], bf1 + 29, bf1 + 26, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf0[28], bf0[27], bf1 + 28, bf1 + 27, &clamp_lo, &clamp_hi);
+
+  // stage 8
+  addsub_sse4_1(bf1[0], bf1[15], bf0 + 0, bf0 + 15, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf1[1], bf1[14], bf0 + 1, bf0 + 14, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf1[2], bf1[13], bf0 + 2, bf0 + 13, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf1[3], bf1[12], bf0 + 3, bf0 + 12, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf1[4], bf1[11], bf0 + 4, bf0 + 11, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf1[5], bf1[10], bf0 + 5, bf0 + 10, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf1[6], bf1[9], bf0 + 6, bf0 + 9, &clamp_lo, &clamp_hi);
+  addsub_sse4_1(bf1[7], bf1[8], bf0 + 7, bf0 + 8, &clamp_lo, &clamp_hi);
+  bf0[16] = bf1[16];
+  bf0[17] = bf1[17];
+  bf0[18] = bf1[18];
+  bf0[19] = bf1[19];
+  bf0[20] =
+      half_btf_sse4_1(&cospim32, &bf1[20], &cospi32, &bf1[27], &rounding, bit);
+  bf0[21] =
+      half_btf_sse4_1(&cospim32, &bf1[21], &cospi32, &bf1[26], &rounding, bit);
+  bf0[22] =
+      half_btf_sse4_1(&cospim32, &bf1[22], &cospi32, &bf1[25], &rounding, bit);
+  bf0[23] =
+      half_btf_sse4_1(&cospim32, &bf1[23], &cospi32, &bf1[24], &rounding, bit);
+  bf0[24] =
+      half_btf_sse4_1(&cospi32, &bf1[23], &cospi32, &bf1[24], &rounding, bit);
+  bf0[25] =
+      half_btf_sse4_1(&cospi32, &bf1[22], &cospi32, &bf1[25], &rounding, bit);
+  bf0[26] =
+      half_btf_sse4_1(&cospi32, &bf1[21], &cospi32, &bf1[26], &rounding, bit);
+  bf0[27] =
+      half_btf_sse4_1(&cospi32, &bf1[20], &cospi32, &bf1[27], &rounding, bit);
+  bf0[28] = bf1[28];
+  bf0[29] = bf1[29];
+  bf0[30] = bf1[30];
+  bf0[31] = bf1[31];
+
+  // stage 9
+  if (do_cols) {
+    addsub_no_clamp_sse4_1(bf0[0], bf0[31], out + 0, out + 31);
+    addsub_no_clamp_sse4_1(bf0[1], bf0[30], out + 1, out + 30);
+    addsub_no_clamp_sse4_1(bf0[2], bf0[29], out + 2, out + 29);
+    addsub_no_clamp_sse4_1(bf0[3], bf0[28], out + 3, out + 28);
+    addsub_no_clamp_sse4_1(bf0[4], bf0[27], out + 4, out + 27);
+    addsub_no_clamp_sse4_1(bf0[5], bf0[26], out + 5, out + 26);
+    addsub_no_clamp_sse4_1(bf0[6], bf0[25], out + 6, out + 25);
+    addsub_no_clamp_sse4_1(bf0[7], bf0[24], out + 7, out + 24);
+    addsub_no_clamp_sse4_1(bf0[8], bf0[23], out + 8, out + 23);
+    addsub_no_clamp_sse4_1(bf0[9], bf0[22], out + 9, out + 22);
+    addsub_no_clamp_sse4_1(bf0[10], bf0[21], out + 10, out + 21);
+    addsub_no_clamp_sse4_1(bf0[11], bf0[20], out + 11, out + 20);
+    addsub_no_clamp_sse4_1(bf0[12], bf0[19], out + 12, out + 19);
+    addsub_no_clamp_sse4_1(bf0[13], bf0[18], out + 13, out + 18);
+    addsub_no_clamp_sse4_1(bf0[14], bf0[17], out + 14, out + 17);
+    addsub_no_clamp_sse4_1(bf0[15], bf0[16], out + 15, out + 16);
+  } else {
+    const int log_range_out = AOMMAX(16, bd + 6);
+    const __m128i clamp_lo_out = _mm_set1_epi32(AOMMAX(
+        -(1 << (log_range_out - 1)), -(1 << (log_range - 1 - out_shift))));
+    const __m128i clamp_hi_out = _mm_set1_epi32(AOMMIN(
+        (1 << (log_range_out - 1)) - 1, (1 << (log_range - 1 - out_shift))));
+
+    addsub_shift_sse4_1(bf0[0], bf0[31], out + 0, out + 31, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+    addsub_shift_sse4_1(bf0[1], bf0[30], out + 1, out + 30, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+    addsub_shift_sse4_1(bf0[2], bf0[29], out + 2, out + 29, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+    addsub_shift_sse4_1(bf0[3], bf0[28], out + 3, out + 28, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+    addsub_shift_sse4_1(bf0[4], bf0[27], out + 4, out + 27, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+    addsub_shift_sse4_1(bf0[5], bf0[26], out + 5, out + 26, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+    addsub_shift_sse4_1(bf0[6], bf0[25], out + 6, out + 25, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+    addsub_shift_sse4_1(bf0[7], bf0[24], out + 7, out + 24, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+    addsub_shift_sse4_1(bf0[8], bf0[23], out + 8, out + 23, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+    addsub_shift_sse4_1(bf0[9], bf0[22], out + 9, out + 22, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+    addsub_shift_sse4_1(bf0[10], bf0[21], out + 10, out + 21, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+    addsub_shift_sse4_1(bf0[11], bf0[20], out + 11, out + 20, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+    addsub_shift_sse4_1(bf0[12], bf0[19], out + 12, out + 19, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+    addsub_shift_sse4_1(bf0[13], bf0[18], out + 13, out + 18, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+    addsub_shift_sse4_1(bf0[14], bf0[17], out + 14, out + 17, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+    addsub_shift_sse4_1(bf0[15], bf0[16], out + 15, out + 16, &clamp_lo_out,
+                        &clamp_hi_out, out_shift);
+  }
+}
+
+void av1_highbd_inv_txfm_add_8x8_sse4_1(const tran_low_t *input, uint8_t *dest,
+                                        int stride,
+                                        const TxfmParam *txfm_param) {
+  int bd = txfm_param->bd;
+  const TX_TYPE tx_type = txfm_param->tx_type;
+  const int32_t *src = cast_to_int32(input);
+  switch (tx_type) {
+      // Assembly version doesn't support some transform types, so use C version
+      // for those.
+    case V_DCT:
+    case H_DCT:
+    case V_ADST:
+    case H_ADST:
+    case V_FLIPADST:
+    case H_FLIPADST:
+    case IDTX:
+      av1_inv_txfm2d_add_8x8_c(src, CONVERT_TO_SHORTPTR(dest), stride, tx_type,
+                               bd);
+      break;
+    default:
+      av1_inv_txfm2d_add_8x8_sse4_1(src, CONVERT_TO_SHORTPTR(dest), stride,
+                                    tx_type, bd);
+      break;
+  }
+}
+
+void av1_highbd_inv_txfm_add_16x8_sse4_1(const tran_low_t *input, uint8_t *dest,
+                                         int stride,
+                                         const TxfmParam *txfm_param) {
+  int bd = txfm_param->bd;
+  const TX_TYPE tx_type = txfm_param->tx_type;
+  const int32_t *src = cast_to_int32(input);
+  switch (tx_type) {
+      // Assembly version doesn't support some transform types, so use C version
+      // for those.
+    case V_DCT:
+    case H_DCT:
+    case V_ADST:
+    case H_ADST:
+    case V_FLIPADST:
+    case H_FLIPADST:
+    case IDTX:
+      av1_inv_txfm2d_add_16x8_c(src, CONVERT_TO_SHORTPTR(dest), stride,
+                                txfm_param->tx_type, txfm_param->bd);
+      break;
+    default:
+      av1_highbd_inv_txfm2d_add_universe_sse4_1(input, dest, stride, tx_type,
+                                                txfm_param->tx_size,
+                                                txfm_param->eob, bd);
+      break;
+  }
+}
+
+void av1_highbd_inv_txfm_add_8x16_sse4_1(const tran_low_t *input, uint8_t *dest,
+                                         int stride,
+                                         const TxfmParam *txfm_param) {
+  int bd = txfm_param->bd;
+  const TX_TYPE tx_type = txfm_param->tx_type;
+  const int32_t *src = cast_to_int32(input);
+  switch (tx_type) {
+      // Assembly version doesn't support some transform types, so use C version
+      // for those.
+    case V_DCT:
+    case H_DCT:
+    case V_ADST:
+    case H_ADST:
+    case V_FLIPADST:
+    case H_FLIPADST:
+    case IDTX:
+      av1_inv_txfm2d_add_8x16_c(src, CONVERT_TO_SHORTPTR(dest), stride,
+                                txfm_param->tx_type, txfm_param->bd);
+      break;
+    default:
+      av1_highbd_inv_txfm2d_add_universe_sse4_1(input, dest, stride, tx_type,
+                                                txfm_param->tx_size,
+                                                txfm_param->eob, bd);
+      break;
+  }
+}
+
+void av1_highbd_inv_txfm_add_16x16_sse4_1(const tran_low_t *input,
+                                          uint8_t *dest, int stride,
+                                          const TxfmParam *txfm_param) {
+  int bd = txfm_param->bd;
+  const TX_TYPE tx_type = txfm_param->tx_type;
+  const int32_t *src = cast_to_int32(input);
+  switch (tx_type) {
+      // Assembly version doesn't support some transform types, so use C version
+      // for those.
+    case V_DCT:
+    case H_DCT:
+    case V_ADST:
+    case H_ADST:
+    case V_FLIPADST:
+    case H_FLIPADST:
+    case IDTX:
+      av1_inv_txfm2d_add_16x16_c(src, CONVERT_TO_SHORTPTR(dest), stride,
+                                 tx_type, bd);
+      break;
+    default:
+      av1_highbd_inv_txfm2d_add_universe_sse4_1(input, dest, stride, tx_type,
+                                                txfm_param->tx_size,
+                                                txfm_param->eob, bd);
+      break;
+  }
+}
+
+void av1_highbd_inv_txfm_add_32x32_sse4_1(const tran_low_t *input,
+                                          uint8_t *dest, int stride,
+                                          const TxfmParam *txfm_param) {
+  int bd = txfm_param->bd;
+  const TX_TYPE tx_type = txfm_param->tx_type;
+  const int32_t *src = cast_to_int32(input);
+  switch (tx_type) {
+    case DCT_DCT:
+      av1_highbd_inv_txfm2d_add_universe_sse4_1(input, dest, stride, tx_type,
+                                                txfm_param->tx_size,
+                                                txfm_param->eob, bd);
+      break;
+      // Assembly version doesn't support IDTX, so use C version for it.
+    case IDTX:
+      av1_inv_txfm2d_add_32x32_c(src, CONVERT_TO_SHORTPTR(dest), stride,
+                                 tx_type, bd);
+      break;
+    default: assert(0);
+  }
+}
+
+void av1_highbd_inv_txfm_add_4x4_sse4_1(const tran_low_t *input, uint8_t *dest,
+                                        int stride,
+                                        const TxfmParam *txfm_param) {
+  assert(av1_ext_tx_used[txfm_param->tx_set_type][txfm_param->tx_type]);
+  int eob = txfm_param->eob;
+  int bd = txfm_param->bd;
+  int lossless = txfm_param->lossless;
+  const int32_t *src = cast_to_int32(input);
+  const TX_TYPE tx_type = txfm_param->tx_type;
+  if (lossless) {
+    assert(tx_type == DCT_DCT);
+    av1_highbd_iwht4x4_add(input, dest, stride, eob, bd);
+    return;
+  }
+  switch (tx_type) {
+      // Assembly version doesn't support some transform types, so use C version
+      // for those.
+    case V_DCT:
+    case H_DCT:
+    case V_ADST:
+    case H_ADST:
+    case V_FLIPADST:
+    case H_FLIPADST:
+    case IDTX:
+      av1_inv_txfm2d_add_4x4_c(src, CONVERT_TO_SHORTPTR(dest), stride, tx_type,
+                               bd);
+      break;
+    default:
+      av1_inv_txfm2d_add_4x4_sse4_1(src, CONVERT_TO_SHORTPTR(dest), stride,
+                                    tx_type, bd);
+      break;
+  }
+}
+
+static const transform_1d_sse4_1
+    highbd_txfm_all_1d_zeros_w8_arr[TX_SIZES][ITX_TYPES_1D][4] = {
+      {
+          { NULL, NULL, NULL, NULL },
+          { NULL, NULL, NULL, NULL },
+          { NULL, NULL, NULL, NULL },
+      },
+      { { idct8x8_low1_sse4_1, idct8x8_new_sse4_1, NULL, NULL },
+        { iadst8x8_low1_sse4_1, iadst8x8_new_sse4_1, NULL, NULL },
+        { NULL, NULL, NULL, NULL } },
+      {
+          { idct16x16_low1_sse4_1, idct16x16_low8_sse4_1, idct16x16_sse4_1,
+            NULL },
+          { iadst16x16_low1_sse4_1, iadst16x16_low8_sse4_1, iadst16x16_sse4_1,
+            NULL },
+          { NULL, NULL, NULL, NULL },
+      },
+      { { idct32x32_low1_sse4_1, idct32x32_low8_sse4_1, idct32x32_low16_sse4_1,
+          idct32x32_sse4_1 },
+        { NULL, NULL, NULL, NULL },
+        { NULL, NULL, NULL, NULL } },
+      { { idct64x64_low1_sse4_1, idct64x64_low8_sse4_1, idct64x64_low16_sse4_1,
+          idct64x64_sse4_1 },
+        { NULL, NULL, NULL, NULL },
+        { NULL, NULL, NULL, NULL } }
+    };
+
+static void highbd_inv_txfm2d_add_no_identity_sse41(const int32_t *input,
+                                                    uint16_t *output,
+                                                    int stride, TX_TYPE tx_type,
+                                                    TX_SIZE tx_size, int eob,
+                                                    const int bd) {
+  __m128i buf1[64 * 16];
+  int eobx, eoby;
+  get_eobx_eoby_scan_default(&eobx, &eoby, tx_size, eob);
+  const int8_t *shift = inv_txfm_shift_ls[tx_size];
+  const int txw_idx = get_txw_idx(tx_size);
+  const int txh_idx = get_txh_idx(tx_size);
+  const int txfm_size_col = tx_size_wide[tx_size];
+  const int txfm_size_row = tx_size_high[tx_size];
+  const int buf_size_w_div8 = txfm_size_col >> 2;
+  const int buf_size_nonzero_w_div8 = (eobx + 8) >> 3;
+  const int buf_size_nonzero_h_div8 = (eoby + 8) >> 3;
+  const int input_stride = AOMMIN(32, txfm_size_col);
+  const int rect_type = get_rect_tx_log_ratio(txfm_size_col, txfm_size_row);
+
+  const int fun_idx_x = lowbd_txfm_all_1d_zeros_idx[eobx];
+  const int fun_idx_y = lowbd_txfm_all_1d_zeros_idx[eoby];
+  const transform_1d_sse4_1 row_txfm =
+      highbd_txfm_all_1d_zeros_w8_arr[txw_idx][hitx_1d_tab[tx_type]][fun_idx_x];
+  const transform_1d_sse4_1 col_txfm =
+      highbd_txfm_all_1d_zeros_w8_arr[txh_idx][vitx_1d_tab[tx_type]][fun_idx_y];
+
+  assert(col_txfm != NULL);
+  assert(row_txfm != NULL);
+  int ud_flip, lr_flip;
+  get_flip_cfg(tx_type, &ud_flip, &lr_flip);
+
+  // 1st stage: column transform
+  for (int i = 0; i < buf_size_nonzero_h_div8 << 1; i++) {
+    __m128i buf0[64];
+    const int32_t *input_row = input + i * input_stride * 4;
+    for (int j = 0; j < buf_size_nonzero_w_div8 << 1; ++j) {
+      __m128i *buf0_cur = buf0 + j * 4;
+      load_buffer_32bit_input(input_row + j * 4, input_stride, buf0_cur, 4);
+
+      TRANSPOSE_4X4(buf0_cur[0], buf0_cur[1], buf0_cur[2], buf0_cur[3],
+                    buf0_cur[0], buf0_cur[1], buf0_cur[2], buf0_cur[3]);
+    }
+    if (rect_type == 1 || rect_type == -1) {
+      av1_round_shift_rect_array_32_sse4_1(
+          buf0, buf0, buf_size_nonzero_w_div8 << 3, 0, NewInvSqrt2);
+    }
+    row_txfm(buf0, buf0, inv_cos_bit_row[txw_idx][txh_idx], 0, bd, -shift[0]);
+
+    __m128i *_buf1 = buf1 + i * 4;
+    if (lr_flip) {
+      for (int j = 0; j < buf_size_w_div8; ++j) {
+        TRANSPOSE_4X4(buf0[4 * j + 3], buf0[4 * j + 2], buf0[4 * j + 1],
+                      buf0[4 * j],
+                      _buf1[txfm_size_row * (buf_size_w_div8 - 1 - j) + 0],
+                      _buf1[txfm_size_row * (buf_size_w_div8 - 1 - j) + 1],
+                      _buf1[txfm_size_row * (buf_size_w_div8 - 1 - j) + 2],
+                      _buf1[txfm_size_row * (buf_size_w_div8 - 1 - j) + 3]);
+      }
+    } else {
+      for (int j = 0; j < buf_size_w_div8; ++j) {
+        TRANSPOSE_4X4(
+            buf0[j * 4 + 0], buf0[j * 4 + 1], buf0[j * 4 + 2], buf0[j * 4 + 3],
+            _buf1[j * txfm_size_row + 0], _buf1[j * txfm_size_row + 1],
+            _buf1[j * txfm_size_row + 2], _buf1[j * txfm_size_row + 3]);
+      }
+    }
+  }
+  // 2nd stage: column transform
+  for (int i = 0; i < buf_size_w_div8; i++) {
+    col_txfm(buf1 + i * txfm_size_row, buf1 + i * txfm_size_row,
+             inv_cos_bit_col[txw_idx][txh_idx], 1, bd, 0);
+
+    av1_round_shift_array_32_sse4_1(buf1 + i * txfm_size_row,
+                                    buf1 + i * txfm_size_row, txfm_size_row,
+                                    -shift[1]);
+  }
+
+  // write to buffer
+  {
+    for (int i = 0; i < (txfm_size_col >> 3); i++) {
+      highbd_write_buffer_8xn_sse4_1(buf1 + i * txfm_size_row * 2,
+                                     output + 8 * i, stride, ud_flip,
+                                     txfm_size_row, bd);
+    }
+  }
+}
+
+void av1_highbd_inv_txfm2d_add_universe_sse4_1(const int32_t *input,
+                                               uint8_t *output, int stride,
+                                               TX_TYPE tx_type, TX_SIZE tx_size,
+                                               int eob, const int bd) {
+  switch (tx_type) {
+    case DCT_DCT:
+    case ADST_DCT:
+    case DCT_ADST:
+    case ADST_ADST:
+    case FLIPADST_DCT:
+    case DCT_FLIPADST:
+    case FLIPADST_FLIPADST:
+    case ADST_FLIPADST:
+    case FLIPADST_ADST:
+      highbd_inv_txfm2d_add_no_identity_sse41(
+          input, CONVERT_TO_SHORTPTR(output), stride, tx_type, tx_size, eob,
+          bd);
+      break;
+    default: assert(0); break;
+  }
+}
+
+void av1_highbd_inv_txfm_add_sse4_1(const tran_low_t *input, uint8_t *dest,
+                                    int stride, const TxfmParam *txfm_param) {
+  assert(av1_ext_tx_used[txfm_param->tx_set_type][txfm_param->tx_type]);
+  const TX_SIZE tx_size = txfm_param->tx_size;
+  switch (tx_size) {
+    case TX_32X32:
+      av1_highbd_inv_txfm_add_32x32_sse4_1(input, dest, stride, txfm_param);
+      break;
+    case TX_16X16:
+      av1_highbd_inv_txfm_add_16x16_sse4_1(input, dest, stride, txfm_param);
+      break;
+    case TX_8X8:
+      av1_highbd_inv_txfm_add_8x8_sse4_1(input, dest, stride, txfm_param);
+      break;
+    case TX_4X8:
+      av1_highbd_inv_txfm_add_4x8(input, dest, stride, txfm_param);
+      break;
+    case TX_8X4:
+      av1_highbd_inv_txfm_add_8x4(input, dest, stride, txfm_param);
+      break;
+    case TX_8X16:
+      av1_highbd_inv_txfm_add_8x16_sse4_1(input, dest, stride, txfm_param);
+      break;
+    case TX_16X8:
+      av1_highbd_inv_txfm_add_16x8_sse4_1(input, dest, stride, txfm_param);
+      break;
+    case TX_16X32:
+      av1_highbd_inv_txfm_add_16x32(input, dest, stride, txfm_param);
+      break;
+    case TX_32X16:
+      av1_highbd_inv_txfm_add_32x16(input, dest, stride, txfm_param);
+      break;
+    case TX_32X64:
+      av1_highbd_inv_txfm_add_32x64(input, dest, stride, txfm_param);
+      break;
+    case TX_64X32:
+      av1_highbd_inv_txfm_add_64x32(input, dest, stride, txfm_param);
+      break;
+    case TX_4X4:
+      av1_highbd_inv_txfm_add_4x4_sse4_1(input, dest, stride, txfm_param);
+      break;
+    case TX_16X4:
+      av1_highbd_inv_txfm_add_16x4(input, dest, stride, txfm_param);
+      break;
+    case TX_4X16:
+      av1_highbd_inv_txfm_add_4x16(input, dest, stride, txfm_param);
+      break;
+    case TX_8X32:
+      av1_highbd_inv_txfm_add_8x32(input, dest, stride, txfm_param);
+      break;
+    case TX_32X8:
+      av1_highbd_inv_txfm_add_32x8(input, dest, stride, txfm_param);
+      break;
+    case TX_64X64:
+    case TX_16X64:
+    case TX_64X16:
+      av1_highbd_inv_txfm2d_add_universe_sse4_1(
+          input, dest, stride, txfm_param->tx_type, txfm_param->tx_size,
+          txfm_param->eob, txfm_param->bd);
+      break;
+    default: assert(0 && "Invalid transform size"); break;
+  }
+}
diff --git a/third_party/aom/av1/common/x86/highbd_jnt_convolve_avx2.c b/third_party/aom/av1/common/x86/highbd_jnt_convolve_avx2.c
new file mode 100644
index 000000000..e298cf653
--- /dev/null
+++ b/third_party/aom/av1/common/x86/highbd_jnt_convolve_avx2.c
@@ -0,0 +1,846 @@
+/*
+ * 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 <assert.h>
+
+#include "config/aom_dsp_rtcd.h"
+
+#include "aom_dsp/x86/convolve_avx2.h"
+#include "aom_dsp/x86/convolve_common_intrin.h"
+#include "aom_dsp/x86/convolve_sse4_1.h"
+#include "aom_dsp/x86/synonyms.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/aom_filter.h"
+#include "av1/common/convolve.h"
+
+void av1_highbd_jnt_convolve_2d_copy_avx2(
+    const uint16_t *src, int src_stride, uint16_t *dst0, int dst_stride0, int w,
+    int h, const InterpFilterParams *filter_params_x,
+    const InterpFilterParams *filter_params_y, const int subpel_x_q4,
+    const int subpel_y_q4, ConvolveParams *conv_params, int bd) {
+  CONV_BUF_TYPE *dst = conv_params->dst;
+  int dst_stride = conv_params->dst_stride;
+  (void)filter_params_x;
+  (void)filter_params_y;
+  (void)subpel_x_q4;
+  (void)subpel_y_q4;
+
+  const int bits =
+      FILTER_BITS * 2 - conv_params->round_1 - conv_params->round_0;
+  const __m128i left_shift = _mm_cvtsi32_si128(bits);
+  const int do_average = conv_params->do_average;
+  const int use_jnt_comp_avg = conv_params->use_jnt_comp_avg;
+  const int w0 = conv_params->fwd_offset;
+  const int w1 = conv_params->bck_offset;
+  const __m256i wt0 = _mm256_set1_epi32(w0);
+  const __m256i wt1 = _mm256_set1_epi32(w1);
+  const __m256i zero = _mm256_setzero_si256();
+  int i, j;
+
+  const int offset_0 =
+      bd + 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+  const int offset = (1 << offset_0) + (1 << (offset_0 - 1));
+  const __m256i offset_const = _mm256_set1_epi32(offset);
+  const __m256i offset_const_16b = _mm256_set1_epi16(offset);
+  const int rounding_shift =
+      2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+  const __m256i rounding_const = _mm256_set1_epi32((1 << rounding_shift) >> 1);
+  const __m256i clip_pixel_to_bd =
+      _mm256_set1_epi16(bd == 10 ? 1023 : (bd == 12 ? 4095 : 255));
+
+  assert(bits <= 4);
+
+  if (!(w % 16)) {
+    for (i = 0; i < h; i += 1) {
+      for (j = 0; j < w; j += 16) {
+        const __m256i src_16bit =
+            _mm256_loadu_si256((__m256i *)(&src[i * src_stride + j]));
+
+        const __m256i res = _mm256_sll_epi16(src_16bit, left_shift);
+
+        if (do_average) {
+          const __m256i data_0 =
+              _mm256_loadu_si256((__m256i *)(&dst[i * dst_stride + j]));
+
+          const __m256i data_ref_0_lo = _mm256_unpacklo_epi16(data_0, zero);
+          const __m256i data_ref_0_hi = _mm256_unpackhi_epi16(data_0, zero);
+
+          const __m256i res_32b_lo = _mm256_unpacklo_epi16(res, zero);
+          const __m256i res_unsigned_lo =
+              _mm256_add_epi32(res_32b_lo, offset_const);
+
+          const __m256i comp_avg_res_lo = highbd_comp_avg(
+              &data_ref_0_lo, &res_unsigned_lo, &wt0, &wt1, use_jnt_comp_avg);
+
+          const __m256i res_32b_hi = _mm256_unpackhi_epi16(res, zero);
+          const __m256i res_unsigned_hi =
+              _mm256_add_epi32(res_32b_hi, offset_const);
+
+          const __m256i comp_avg_res_hi = highbd_comp_avg(
+              &data_ref_0_hi, &res_unsigned_hi, &wt0, &wt1, use_jnt_comp_avg);
+
+          const __m256i round_result_lo = highbd_convolve_rounding(
+              &comp_avg_res_lo, &offset_const, &rounding_const, rounding_shift);
+          const __m256i round_result_hi = highbd_convolve_rounding(
+              &comp_avg_res_hi, &offset_const, &rounding_const, rounding_shift);
+
+          const __m256i res_16b =
+              _mm256_packus_epi32(round_result_lo, round_result_hi);
+          const __m256i res_clip = _mm256_min_epi16(res_16b, clip_pixel_to_bd);
+
+          _mm256_store_si256((__m256i *)(&dst0[i * dst_stride0 + j]), res_clip);
+        } else {
+          const __m256i res_unsigned_16b =
+              _mm256_adds_epu16(res, offset_const_16b);
+
+          _mm256_store_si256((__m256i *)(&dst[i * dst_stride + j]),
+                             res_unsigned_16b);
+        }
+      }
+    }
+  } else if (!(w % 4)) {
+    for (i = 0; i < h; i += 2) {
+      for (j = 0; j < w; j += 8) {
+        const __m128i src_row_0 =
+            _mm_loadu_si128((__m128i *)(&src[i * src_stride + j]));
+        const __m128i src_row_1 =
+            _mm_loadu_si128((__m128i *)(&src[i * src_stride + j + src_stride]));
+        // since not all compilers yet support _mm256_set_m128i()
+        const __m256i src_10 = _mm256_insertf128_si256(
+            _mm256_castsi128_si256(src_row_0), src_row_1, 1);
+
+        const __m256i res = _mm256_sll_epi16(src_10, left_shift);
+
+        if (w - j < 8) {
+          if (do_average) {
+            const __m256i data_0 = _mm256_castsi128_si256(
+                _mm_loadl_epi64((__m128i *)(&dst[i * dst_stride + j])));
+            const __m256i data_1 = _mm256_castsi128_si256(_mm_loadl_epi64(
+                (__m128i *)(&dst[i * dst_stride + j + dst_stride])));
+            const __m256i data_01 =
+                _mm256_permute2x128_si256(data_0, data_1, 0x20);
+
+            const __m256i data_ref_0 = _mm256_unpacklo_epi16(data_01, zero);
+
+            const __m256i res_32b = _mm256_unpacklo_epi16(res, zero);
+            const __m256i res_unsigned_lo =
+                _mm256_add_epi32(res_32b, offset_const);
+
+            const __m256i comp_avg_res = highbd_comp_avg(
+                &data_ref_0, &res_unsigned_lo, &wt0, &wt1, use_jnt_comp_avg);
+
+            const __m256i round_result = highbd_convolve_rounding(
+                &comp_avg_res, &offset_const, &rounding_const, rounding_shift);
+
+            const __m256i res_16b =
+                _mm256_packus_epi32(round_result, round_result);
+            const __m256i res_clip =
+                _mm256_min_epi16(res_16b, clip_pixel_to_bd);
+
+            const __m128i res_0 = _mm256_castsi256_si128(res_clip);
+            const __m128i res_1 = _mm256_extracti128_si256(res_clip, 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 __m256i res_unsigned_16b =
+                _mm256_adds_epu16(res, offset_const_16b);
+
+            const __m128i res_0 = _mm256_castsi256_si128(res_unsigned_16b);
+            const __m128i res_1 = _mm256_extracti128_si256(res_unsigned_16b, 1);
+
+            _mm_storel_epi64((__m128i *)(&dst[i * dst_stride + j]), res_0);
+            _mm_storel_epi64((__m128i *)(&dst[i * dst_stride + j + dst_stride]),
+                             res_1);
+          }
+        } else {
+          if (do_average) {
+            const __m256i data_0 = _mm256_castsi128_si256(
+                _mm_loadu_si128((__m128i *)(&dst[i * dst_stride + j])));
+            const __m256i data_1 = _mm256_castsi128_si256(_mm_loadu_si128(
+                (__m128i *)(&dst[i * dst_stride + j + dst_stride])));
+            const __m256i data_01 =
+                _mm256_permute2x128_si256(data_0, data_1, 0x20);
+
+            const __m256i data_ref_0_lo = _mm256_unpacklo_epi16(data_01, zero);
+            const __m256i data_ref_0_hi = _mm256_unpackhi_epi16(data_01, zero);
+
+            const __m256i res_32b_lo = _mm256_unpacklo_epi16(res, zero);
+            const __m256i res_unsigned_lo =
+                _mm256_add_epi32(res_32b_lo, offset_const);
+
+            const __m256i comp_avg_res_lo = highbd_comp_avg(
+                &data_ref_0_lo, &res_unsigned_lo, &wt0, &wt1, use_jnt_comp_avg);
+
+            const __m256i res_32b_hi = _mm256_unpackhi_epi16(res, zero);
+            const __m256i res_unsigned_hi =
+                _mm256_add_epi32(res_32b_hi, offset_const);
+
+            const __m256i comp_avg_res_hi = highbd_comp_avg(
+                &data_ref_0_hi, &res_unsigned_hi, &wt0, &wt1, use_jnt_comp_avg);
+
+            const __m256i round_result_lo =
+                highbd_convolve_rounding(&comp_avg_res_lo, &offset_const,
+                                         &rounding_const, rounding_shift);
+            const __m256i round_result_hi =
+                highbd_convolve_rounding(&comp_avg_res_hi, &offset_const,
+                                         &rounding_const, rounding_shift);
+
+            const __m256i res_16b =
+                _mm256_packus_epi32(round_result_lo, round_result_hi);
+            const __m256i res_clip =
+                _mm256_min_epi16(res_16b, clip_pixel_to_bd);
+
+            const __m128i res_0 = _mm256_castsi256_si128(res_clip);
+            const __m128i res_1 = _mm256_extracti128_si256(res_clip, 1);
+
+            _mm_store_si128((__m128i *)(&dst0[i * dst_stride0 + j]), res_0);
+            _mm_store_si128(
+                (__m128i *)(&dst0[i * dst_stride0 + j + dst_stride0]), res_1);
+          } else {
+            const __m256i res_unsigned_16b =
+                _mm256_adds_epu16(res, offset_const_16b);
+            const __m128i res_0 = _mm256_castsi256_si128(res_unsigned_16b);
+            const __m128i res_1 = _mm256_extracti128_si256(res_unsigned_16b, 1);
+
+            _mm_store_si128((__m128i *)(&dst[i * dst_stride + j]), res_0);
+            _mm_store_si128((__m128i *)(&dst[i * dst_stride + j + dst_stride]),
+                            res_1);
+          }
+        }
+      }
+    }
+  }
+}
+
+void av1_highbd_jnt_convolve_2d_avx2(
+    const uint16_t *src, int src_stride, uint16_t *dst0, int dst_stride0, int w,
+    int h, const InterpFilterParams *filter_params_x,
+    const InterpFilterParams *filter_params_y, const int subpel_x_q4,
+    const int subpel_y_q4, ConvolveParams *conv_params, int bd) {
+  DECLARE_ALIGNED(32, int16_t, im_block[(MAX_SB_SIZE + MAX_FILTER_TAP) * 8]);
+  CONV_BUF_TYPE *dst = conv_params->dst;
+  int dst_stride = conv_params->dst_stride;
+  int im_h = h + filter_params_y->taps - 1;
+  int im_stride = 8;
+  int i, j;
+  const int fo_vert = filter_params_y->taps / 2 - 1;
+  const int fo_horiz = filter_params_x->taps / 2 - 1;
+  const uint16_t *const src_ptr = src - fo_vert * src_stride - 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[8], coeffs_y[4], coeffs_x[4];
+  const int do_average = conv_params->do_average;
+  const int use_jnt_comp_avg = conv_params->use_jnt_comp_avg;
+
+  const int w0 = conv_params->fwd_offset;
+  const int w1 = conv_params->bck_offset;
+  const __m256i wt0 = _mm256_set1_epi32(w0);
+  const __m256i wt1 = _mm256_set1_epi32(w1);
+  const __m256i zero = _mm256_setzero_si256();
+
+  const __m256i round_const_x = _mm256_set1_epi32(
+      ((1 << conv_params->round_0) >> 1) + (1 << (bd + FILTER_BITS - 1)));
+  const __m128i round_shift_x = _mm_cvtsi32_si128(conv_params->round_0);
+
+  const __m256i round_const_y = _mm256_set1_epi32(
+      ((1 << conv_params->round_1) >> 1) -
+      (1 << (bd + 2 * FILTER_BITS - conv_params->round_0 - 1)));
+  const __m128i round_shift_y = _mm_cvtsi32_si128(conv_params->round_1);
+
+  const int offset_0 =
+      bd + 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+  const int offset = (1 << offset_0) + (1 << (offset_0 - 1));
+  const __m256i offset_const = _mm256_set1_epi32(offset);
+  const int rounding_shift =
+      2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+  const __m256i rounding_const = _mm256_set1_epi32((1 << rounding_shift) >> 1);
+
+  const __m256i clip_pixel_to_bd =
+      _mm256_set1_epi16(bd == 10 ? 1023 : (bd == 12 ? 4095 : 255));
+
+  prepare_coeffs(filter_params_x, subpel_x_q4, coeffs_x);
+  prepare_coeffs(filter_params_y, subpel_y_q4, coeffs_y);
+
+  for (j = 0; j < w; j += 8) {
+    /* Horizontal filter */
+    {
+      for (i = 0; i < im_h; i += 2) {
+        const __m256i row0 =
+            _mm256_loadu_si256((__m256i *)&src_ptr[i * src_stride + j]);
+        __m256i row1 = _mm256_set1_epi16(0);
+        if (i + 1 < im_h)
+          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);
+
+        __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);
+
+        _mm256_store_si256((__m256i *)&im_block[i * im_stride], res);
+      }
+    }
+
+    /* Vertical filter */
+    {
+      __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_y), round_shift_y);
+
+        const __m256i res_unsigned_lo =
+            _mm256_add_epi32(res_a_round, offset_const);
+
+        if (w - j < 8) {
+          if (do_average) {
+            const __m256i data_0 = _mm256_castsi128_si256(
+                _mm_loadl_epi64((__m128i *)(&dst[i * dst_stride + j])));
+            const __m256i data_1 = _mm256_castsi128_si256(_mm_loadl_epi64(
+                (__m128i *)(&dst[i * dst_stride + j + dst_stride])));
+            const __m256i data_01 =
+                _mm256_permute2x128_si256(data_0, data_1, 0x20);
+
+            const __m256i data_ref_0 = _mm256_unpacklo_epi16(data_01, zero);
+
+            const __m256i comp_avg_res = highbd_comp_avg(
+                &data_ref_0, &res_unsigned_lo, &wt0, &wt1, use_jnt_comp_avg);
+
+            const __m256i round_result = highbd_convolve_rounding(
+                &comp_avg_res, &offset_const, &rounding_const, rounding_shift);
+
+            const __m256i res_16b =
+                _mm256_packus_epi32(round_result, round_result);
+            const __m256i res_clip =
+                _mm256_min_epi16(res_16b, clip_pixel_to_bd);
+
+            const __m128i res_0 = _mm256_castsi256_si128(res_clip);
+            const __m128i res_1 = _mm256_extracti128_si256(res_clip, 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 {
+            __m256i res_16b =
+                _mm256_packus_epi32(res_unsigned_lo, res_unsigned_lo);
+            const __m128i res_0 = _mm256_castsi256_si128(res_16b);
+            const __m128i res_1 = _mm256_extracti128_si256(res_16b, 1);
+
+            _mm_storel_epi64((__m128i *)(&dst[i * dst_stride + j]), res_0);
+            _mm_storel_epi64((__m128i *)(&dst[i * dst_stride + j + dst_stride]),
+                             res_1);
+          }
+        } else {
+          const __m256i res_b = convolve(s + 4, coeffs_y);
+          const __m256i res_b_round = _mm256_sra_epi32(
+              _mm256_add_epi32(res_b, round_const_y), round_shift_y);
+
+          __m256i res_unsigned_hi = _mm256_add_epi32(res_b_round, offset_const);
+
+          if (do_average) {
+            const __m256i data_0 = _mm256_castsi128_si256(
+                _mm_loadu_si128((__m128i *)(&dst[i * dst_stride + j])));
+            const __m256i data_1 = _mm256_castsi128_si256(_mm_loadu_si128(
+                (__m128i *)(&dst[i * dst_stride + j + dst_stride])));
+            const __m256i data_01 =
+                _mm256_permute2x128_si256(data_0, data_1, 0x20);
+
+            const __m256i data_ref_0_lo = _mm256_unpacklo_epi16(data_01, zero);
+            const __m256i data_ref_0_hi = _mm256_unpackhi_epi16(data_01, zero);
+
+            const __m256i comp_avg_res_lo = highbd_comp_avg(
+                &data_ref_0_lo, &res_unsigned_lo, &wt0, &wt1, use_jnt_comp_avg);
+            const __m256i comp_avg_res_hi = highbd_comp_avg(
+                &data_ref_0_hi, &res_unsigned_hi, &wt0, &wt1, use_jnt_comp_avg);
+
+            const __m256i round_result_lo =
+                highbd_convolve_rounding(&comp_avg_res_lo, &offset_const,
+                                         &rounding_const, rounding_shift);
+            const __m256i round_result_hi =
+                highbd_convolve_rounding(&comp_avg_res_hi, &offset_const,
+                                         &rounding_const, rounding_shift);
+
+            const __m256i res_16b =
+                _mm256_packus_epi32(round_result_lo, round_result_hi);
+            const __m256i res_clip =
+                _mm256_min_epi16(res_16b, clip_pixel_to_bd);
+
+            const __m128i res_0 = _mm256_castsi256_si128(res_clip);
+            const __m128i res_1 = _mm256_extracti128_si256(res_clip, 1);
+
+            _mm_store_si128((__m128i *)(&dst0[i * dst_stride0 + j]), res_0);
+            _mm_store_si128(
+                (__m128i *)(&dst0[i * dst_stride0 + j + dst_stride0]), res_1);
+          } else {
+            __m256i res_16b =
+                _mm256_packus_epi32(res_unsigned_lo, res_unsigned_hi);
+            const __m128i res_0 = _mm256_castsi256_si128(res_16b);
+            const __m128i res_1 = _mm256_extracti128_si256(res_16b, 1);
+
+            _mm_store_si128((__m128i *)(&dst[i * dst_stride + j]), res_0);
+            _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];
+      }
+    }
+  }
+}
+
+void av1_highbd_jnt_convolve_x_avx2(
+    const uint16_t *src, int src_stride, uint16_t *dst0, int dst_stride0, int w,
+    int h, const InterpFilterParams *filter_params_x,
+    const InterpFilterParams *filter_params_y, const int subpel_x_q4,
+    const int subpel_y_q4, ConvolveParams *conv_params, int bd) {
+  CONV_BUF_TYPE *dst = conv_params->dst;
+  int dst_stride = conv_params->dst_stride;
+  const int fo_horiz = filter_params_x->taps / 2 - 1;
+  const uint16_t *const src_ptr = src - fo_horiz;
+  const int bits = FILTER_BITS - conv_params->round_1;
+  (void)filter_params_y;
+  (void)subpel_y_q4;
+
+  int i, j;
+  __m256i s[4], coeffs_x[4];
+
+  const int do_average = conv_params->do_average;
+  const int use_jnt_comp_avg = conv_params->use_jnt_comp_avg;
+  const int w0 = conv_params->fwd_offset;
+  const int w1 = conv_params->bck_offset;
+  const __m256i wt0 = _mm256_set1_epi32(w0);
+  const __m256i wt1 = _mm256_set1_epi32(w1);
+  const __m256i zero = _mm256_setzero_si256();
+
+  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 __m128i round_shift_bits = _mm_cvtsi32_si128(bits);
+
+  const int offset_0 =
+      bd + 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+  const int offset = (1 << offset_0) + (1 << (offset_0 - 1));
+  const __m256i offset_const = _mm256_set1_epi32(offset);
+  const int rounding_shift =
+      2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+  const __m256i rounding_const = _mm256_set1_epi32((1 << rounding_shift) >> 1);
+  const __m256i clip_pixel_to_bd =
+      _mm256_set1_epi16(bd == 10 ? 1023 : (bd == 12 ? 4095 : 255));
+
+  assert(bits >= 0);
+  prepare_coeffs(filter_params_x, subpel_x_q4, 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_sll_epi32(res_even, round_shift_bits);
+      res_odd = _mm256_sll_epi32(res_odd, round_shift_bits);
+
+      __m256i res1 = _mm256_unpacklo_epi32(res_even, res_odd);
+
+      __m256i res_unsigned_lo = _mm256_add_epi32(res1, offset_const);
+
+      if (w - j < 8) {
+        if (do_average) {
+          const __m256i data_0 = _mm256_castsi128_si256(
+              _mm_loadl_epi64((__m128i *)(&dst[i * dst_stride + j])));
+          const __m256i data_1 = _mm256_castsi128_si256(_mm_loadl_epi64(
+              (__m128i *)(&dst[i * dst_stride + j + dst_stride])));
+          const __m256i data_01 =
+              _mm256_permute2x128_si256(data_0, data_1, 0x20);
+
+          const __m256i data_ref_0 = _mm256_unpacklo_epi16(data_01, zero);
+
+          const __m256i comp_avg_res = highbd_comp_avg(
+              &data_ref_0, &res_unsigned_lo, &wt0, &wt1, use_jnt_comp_avg);
+
+          const __m256i round_result = highbd_convolve_rounding(
+              &comp_avg_res, &offset_const, &rounding_const, rounding_shift);
+
+          const __m256i res_16b =
+              _mm256_packus_epi32(round_result, round_result);
+          const __m256i res_clip = _mm256_min_epi16(res_16b, clip_pixel_to_bd);
+
+          const __m128i res_0 = _mm256_castsi256_si128(res_clip);
+          const __m128i res_1 = _mm256_extracti128_si256(res_clip, 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 {
+          __m256i res_16b =
+              _mm256_packus_epi32(res_unsigned_lo, res_unsigned_lo);
+          const __m128i res_0 = _mm256_castsi256_si128(res_16b);
+          const __m128i res_1 = _mm256_extracti128_si256(res_16b, 1);
+
+          _mm_storel_epi64((__m128i *)(&dst[i * dst_stride + j]), res_0);
+          _mm_storel_epi64((__m128i *)(&dst[i * dst_stride + j + dst_stride]),
+                           res_1);
+        }
+      } else {
+        __m256i res2 = _mm256_unpackhi_epi32(res_even, res_odd);
+        __m256i res_unsigned_hi = _mm256_add_epi32(res2, offset_const);
+
+        if (do_average) {
+          const __m256i data_0 = _mm256_castsi128_si256(
+              _mm_loadu_si128((__m128i *)(&dst[i * dst_stride + j])));
+          const __m256i data_1 = _mm256_castsi128_si256(_mm_loadu_si128(
+              (__m128i *)(&dst[i * dst_stride + j + dst_stride])));
+          const __m256i data_01 =
+              _mm256_permute2x128_si256(data_0, data_1, 0x20);
+
+          const __m256i data_ref_0_lo = _mm256_unpacklo_epi16(data_01, zero);
+          const __m256i data_ref_0_hi = _mm256_unpackhi_epi16(data_01, zero);
+
+          const __m256i comp_avg_res_lo = highbd_comp_avg(
+              &data_ref_0_lo, &res_unsigned_lo, &wt0, &wt1, use_jnt_comp_avg);
+          const __m256i comp_avg_res_hi = highbd_comp_avg(
+              &data_ref_0_hi, &res_unsigned_hi, &wt0, &wt1, use_jnt_comp_avg);
+
+          const __m256i round_result_lo = highbd_convolve_rounding(
+              &comp_avg_res_lo, &offset_const, &rounding_const, rounding_shift);
+          const __m256i round_result_hi = highbd_convolve_rounding(
+              &comp_avg_res_hi, &offset_const, &rounding_const, rounding_shift);
+
+          const __m256i res_16b =
+              _mm256_packus_epi32(round_result_lo, round_result_hi);
+          const __m256i res_clip = _mm256_min_epi16(res_16b, clip_pixel_to_bd);
+
+          const __m128i res_0 = _mm256_castsi256_si128(res_clip);
+          const __m128i res_1 = _mm256_extracti128_si256(res_clip, 1);
+
+          _mm_store_si128((__m128i *)(&dst0[i * dst_stride0 + j]), res_0);
+          _mm_store_si128((__m128i *)(&dst0[i * dst_stride0 + j + dst_stride0]),
+                          res_1);
+        } else {
+          __m256i res_16b =
+              _mm256_packus_epi32(res_unsigned_lo, res_unsigned_hi);
+          const __m128i res_0 = _mm256_castsi256_si128(res_16b);
+          const __m128i res_1 = _mm256_extracti128_si256(res_16b, 1);
+
+          _mm_store_si128((__m128i *)(&dst[i * dst_stride + j]), res_0);
+          _mm_store_si128((__m128i *)(&dst[i * dst_stride + j + dst_stride]),
+                          res_1);
+        }
+      }
+    }
+  }
+}
+
+void av1_highbd_jnt_convolve_y_avx2(
+    const uint16_t *src, int src_stride, uint16_t *dst0, int dst_stride0, int w,
+    int h, const InterpFilterParams *filter_params_x,
+    const InterpFilterParams *filter_params_y, const int subpel_x_q4,
+    const int subpel_y_q4, ConvolveParams *conv_params, int bd) {
+  CONV_BUF_TYPE *dst = conv_params->dst;
+  int dst_stride = conv_params->dst_stride;
+  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 - conv_params->round_0;
+  (void)filter_params_x;
+  (void)subpel_x_q4;
+
+  assert(bits >= 0);
+  int i, j;
+  __m256i s[8], coeffs_y[4];
+  const int do_average = conv_params->do_average;
+  const int use_jnt_comp_avg = conv_params->use_jnt_comp_avg;
+
+  const int w0 = conv_params->fwd_offset;
+  const int w1 = conv_params->bck_offset;
+  const __m256i wt0 = _mm256_set1_epi32(w0);
+  const __m256i wt1 = _mm256_set1_epi32(w1);
+  const __m256i round_const_y =
+      _mm256_set1_epi32(((1 << conv_params->round_1) >> 1));
+  const __m128i round_shift_y = _mm_cvtsi32_si128(conv_params->round_1);
+  const __m128i round_shift_bits = _mm_cvtsi32_si128(bits);
+
+  const int offset_0 =
+      bd + 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+  const int offset = (1 << offset_0) + (1 << (offset_0 - 1));
+  const __m256i offset_const = _mm256_set1_epi32(offset);
+  const int rounding_shift =
+      2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+  const __m256i rounding_const = _mm256_set1_epi32((1 << rounding_shift) >> 1);
+  const __m256i clip_pixel_to_bd =
+      _mm256_set1_epi16(bd == 10 ? 1023 : (bd == 12 ? 4095 : 255));
+  const __m256i zero = _mm256_setzero_si256();
+
+  prepare_coeffs(filter_params_y, subpel_y_q4, 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_sll_epi32(res_a, round_shift_bits);
+        res_a_round = _mm256_sra_epi32(
+            _mm256_add_epi32(res_a_round, round_const_y), round_shift_y);
+
+        __m256i res_unsigned_lo = _mm256_add_epi32(res_a_round, offset_const);
+
+        if (w - j < 8) {
+          if (do_average) {
+            const __m256i data_0 = _mm256_castsi128_si256(
+                _mm_loadl_epi64((__m128i *)(&dst[i * dst_stride + j])));
+            const __m256i data_1 = _mm256_castsi128_si256(_mm_loadl_epi64(
+                (__m128i *)(&dst[i * dst_stride + j + dst_stride])));
+            const __m256i data_01 =
+                _mm256_permute2x128_si256(data_0, data_1, 0x20);
+
+            const __m256i data_ref_0 = _mm256_unpacklo_epi16(data_01, zero);
+
+            const __m256i comp_avg_res = highbd_comp_avg(
+                &data_ref_0, &res_unsigned_lo, &wt0, &wt1, use_jnt_comp_avg);
+
+            const __m256i round_result = highbd_convolve_rounding(
+                &comp_avg_res, &offset_const, &rounding_const, rounding_shift);
+
+            const __m256i res_16b =
+                _mm256_packus_epi32(round_result, round_result);
+            const __m256i res_clip =
+                _mm256_min_epi16(res_16b, clip_pixel_to_bd);
+
+            const __m128i res_0 = _mm256_castsi256_si128(res_clip);
+            const __m128i res_1 = _mm256_extracti128_si256(res_clip, 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 {
+            __m256i res_16b =
+                _mm256_packus_epi32(res_unsigned_lo, res_unsigned_lo);
+            const __m128i res_0 = _mm256_castsi256_si128(res_16b);
+            const __m128i res_1 = _mm256_extracti128_si256(res_16b, 1);
+
+            _mm_storel_epi64((__m128i *)(&dst[i * dst_stride + j]), res_0);
+            _mm_storel_epi64((__m128i *)(&dst[i * dst_stride + j + dst_stride]),
+                             res_1);
+          }
+        } else {
+          const __m256i res_b = convolve(s + 4, coeffs_y);
+          __m256i res_b_round = _mm256_sll_epi32(res_b, round_shift_bits);
+          res_b_round = _mm256_sra_epi32(
+              _mm256_add_epi32(res_b_round, round_const_y), round_shift_y);
+
+          __m256i res_unsigned_hi = _mm256_add_epi32(res_b_round, offset_const);
+
+          if (do_average) {
+            const __m256i data_0 = _mm256_castsi128_si256(
+                _mm_loadu_si128((__m128i *)(&dst[i * dst_stride + j])));
+            const __m256i data_1 = _mm256_castsi128_si256(_mm_loadu_si128(
+                (__m128i *)(&dst[i * dst_stride + j + dst_stride])));
+            const __m256i data_01 =
+                _mm256_permute2x128_si256(data_0, data_1, 0x20);
+
+            const __m256i data_ref_0_lo = _mm256_unpacklo_epi16(data_01, zero);
+            const __m256i data_ref_0_hi = _mm256_unpackhi_epi16(data_01, zero);
+
+            const __m256i comp_avg_res_lo = highbd_comp_avg(
+                &data_ref_0_lo, &res_unsigned_lo, &wt0, &wt1, use_jnt_comp_avg);
+            const __m256i comp_avg_res_hi = highbd_comp_avg(
+                &data_ref_0_hi, &res_unsigned_hi, &wt0, &wt1, use_jnt_comp_avg);
+
+            const __m256i round_result_lo =
+                highbd_convolve_rounding(&comp_avg_res_lo, &offset_const,
+                                         &rounding_const, rounding_shift);
+            const __m256i round_result_hi =
+                highbd_convolve_rounding(&comp_avg_res_hi, &offset_const,
+                                         &rounding_const, rounding_shift);
+
+            const __m256i res_16b =
+                _mm256_packus_epi32(round_result_lo, round_result_hi);
+            const __m256i res_clip =
+                _mm256_min_epi16(res_16b, clip_pixel_to_bd);
+
+            const __m128i res_0 = _mm256_castsi256_si128(res_clip);
+            const __m128i res_1 = _mm256_extracti128_si256(res_clip, 1);
+
+            _mm_store_si128((__m128i *)(&dst0[i * dst_stride0 + j]), res_0);
+            _mm_store_si128(
+                (__m128i *)(&dst0[i * dst_stride0 + j + dst_stride0]), res_1);
+          } else {
+            __m256i res_16b =
+                _mm256_packus_epi32(res_unsigned_lo, res_unsigned_hi);
+            const __m128i res_0 = _mm256_castsi256_si128(res_16b);
+            const __m128i res_1 = _mm256_extracti128_si256(res_16b, 1);
+
+            _mm_store_si128((__m128i *)(&dst[i * dst_stride + j]), res_0);
+            _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];
+      }
+    }
+  }
+}
diff --git a/third_party/aom/av1/common/x86/highbd_jnt_convolve_sse4.c b/third_party/aom/av1/common/x86/highbd_jnt_convolve_sse4.c
new file mode 100644
index 000000000..1a29985b5
--- /dev/null
+++ b/third_party/aom/av1/common/x86/highbd_jnt_convolve_sse4.c
@@ -0,0 +1,383 @@
+/*
+ * 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 <assert.h>
+
+#include "config/aom_dsp_rtcd.h"
+
+#include "aom_dsp/x86/convolve_sse2.h"
+#include "aom_dsp/x86/convolve_sse4_1.h"
+
+void av1_highbd_jnt_convolve_y_sse4_1(
+    const uint16_t *src, int src_stride, uint16_t *dst0, int dst_stride0, int w,
+    int h, const InterpFilterParams *filter_params_x,
+    const InterpFilterParams *filter_params_y, const int subpel_x_q4,
+    const int subpel_y_q4, ConvolveParams *conv_params, int bd) {
+  CONV_BUF_TYPE *dst = conv_params->dst;
+  int dst_stride = conv_params->dst_stride;
+  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 - conv_params->round_0;
+  (void)filter_params_x;
+  (void)subpel_x_q4;
+
+  assert(bits >= 0);
+  int i, j;
+  const int do_average = conv_params->do_average;
+  const int use_jnt_comp_avg = conv_params->use_jnt_comp_avg;
+
+  const int w0 = conv_params->fwd_offset;
+  const int w1 = conv_params->bck_offset;
+  const __m128i wt0 = _mm_set1_epi32(w0);
+  const __m128i wt1 = _mm_set1_epi32(w1);
+  const __m128i round_const_y =
+      _mm_set1_epi32(((1 << conv_params->round_1) >> 1));
+  const __m128i round_shift_y = _mm_cvtsi32_si128(conv_params->round_1);
+  const __m128i round_shift_bits = _mm_cvtsi32_si128(bits);
+
+  const int offset_0 =
+      bd + 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+  const int offset = (1 << offset_0) + (1 << (offset_0 - 1));
+  const __m128i offset_const = _mm_set1_epi32(offset);
+  const int rounding_shift =
+      2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+  const __m128i rounding_const = _mm_set1_epi32((1 << rounding_shift) >> 1);
+  const __m128i clip_pixel_to_bd =
+      _mm_set1_epi16(bd == 10 ? 1023 : (bd == 12 ? 4095 : 255));
+  const __m128i zero = _mm_setzero_si128();
+  __m128i s[16], coeffs_y[4];
+
+  prepare_coeffs(filter_params_y, subpel_y_q4, 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_sll_epi32(res_a0, round_shift_bits);
+        res_a_round0 = _mm_sra_epi32(_mm_add_epi32(res_a_round0, round_const_y),
+                                     round_shift_y);
+
+        const __m128i res_a1 = convolve(s + 8, coeffs_y);
+        __m128i res_a_round1 = _mm_sll_epi32(res_a1, round_shift_bits);
+        res_a_round1 = _mm_sra_epi32(_mm_add_epi32(res_a_round1, round_const_y),
+                                     round_shift_y);
+
+        __m128i res_unsigned_lo_0 = _mm_add_epi32(res_a_round0, offset_const);
+        __m128i res_unsigned_lo_1 = _mm_add_epi32(res_a_round1, offset_const);
+
+        if (w - j < 8) {
+          if (do_average) {
+            const __m128i data_0 =
+                _mm_loadl_epi64((__m128i *)(&dst[i * dst_stride + j]));
+            const __m128i data_1 = _mm_loadl_epi64(
+                (__m128i *)(&dst[i * dst_stride + j + dst_stride]));
+
+            const __m128i data_ref_0 = _mm_unpacklo_epi16(data_0, zero);
+            const __m128i data_ref_1 = _mm_unpacklo_epi16(data_1, zero);
+
+            const __m128i comp_avg_res_0 = highbd_comp_avg_sse4_1(
+                &data_ref_0, &res_unsigned_lo_0, &wt0, &wt1, use_jnt_comp_avg);
+            const __m128i comp_avg_res_1 = highbd_comp_avg_sse4_1(
+                &data_ref_1, &res_unsigned_lo_1, &wt0, &wt1, use_jnt_comp_avg);
+
+            const __m128i round_result_0 =
+                highbd_convolve_rounding_sse2(&comp_avg_res_0, &offset_const,
+                                              &rounding_const, rounding_shift);
+            const __m128i round_result_1 =
+                highbd_convolve_rounding_sse2(&comp_avg_res_1, &offset_const,
+                                              &rounding_const, rounding_shift);
+
+            const __m128i res_16b_0 =
+                _mm_packus_epi32(round_result_0, round_result_0);
+            const __m128i res_clip_0 =
+                _mm_min_epi16(res_16b_0, clip_pixel_to_bd);
+            const __m128i res_16b_1 =
+                _mm_packus_epi32(round_result_1, round_result_1);
+            const __m128i res_clip_1 =
+                _mm_min_epi16(res_16b_1, clip_pixel_to_bd);
+
+            _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]),
+                             res_clip_0);
+            _mm_storel_epi64(
+                (__m128i *)(&dst0[i * dst_stride0 + j + dst_stride0]),
+                res_clip_1);
+
+          } else {
+            __m128i res_16b_0 =
+                _mm_packus_epi32(res_unsigned_lo_0, res_unsigned_lo_0);
+
+            __m128i res_16b_1 =
+                _mm_packus_epi32(res_unsigned_lo_1, res_unsigned_lo_1);
+
+            _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j], res_16b_0);
+            _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j + dst_stride],
+                             res_16b_1);
+          }
+        } else {
+          const __m128i res_b0 = convolve(s + 4, coeffs_y);
+          __m128i res_b_round0 = _mm_sll_epi32(res_b0, round_shift_bits);
+          res_b_round0 = _mm_sra_epi32(
+              _mm_add_epi32(res_b_round0, round_const_y), round_shift_y);
+
+          const __m128i res_b1 = convolve(s + 4 + 8, coeffs_y);
+          __m128i res_b_round1 = _mm_sll_epi32(res_b1, round_shift_bits);
+          res_b_round1 = _mm_sra_epi32(
+              _mm_add_epi32(res_b_round1, round_const_y), round_shift_y);
+
+          __m128i res_unsigned_hi_0 = _mm_add_epi32(res_b_round0, offset_const);
+          __m128i res_unsigned_hi_1 = _mm_add_epi32(res_b_round1, offset_const);
+
+          if (do_average) {
+            const __m128i data_0 =
+                _mm_loadu_si128((__m128i *)(&dst[i * dst_stride + j]));
+            const __m128i data_1 = _mm_loadu_si128(
+                (__m128i *)(&dst[i * dst_stride + j + dst_stride]));
+            const __m128i data_ref_0_lo_0 = _mm_unpacklo_epi16(data_0, zero);
+            const __m128i data_ref_0_lo_1 = _mm_unpacklo_epi16(data_1, zero);
+
+            const __m128i data_ref_0_hi_0 = _mm_unpackhi_epi16(data_0, zero);
+            const __m128i data_ref_0_hi_1 = _mm_unpackhi_epi16(data_1, zero);
+
+            const __m128i comp_avg_res_lo_0 =
+                highbd_comp_avg_sse4_1(&data_ref_0_lo_0, &res_unsigned_lo_0,
+                                       &wt0, &wt1, use_jnt_comp_avg);
+            const __m128i comp_avg_res_lo_1 =
+                highbd_comp_avg_sse4_1(&data_ref_0_lo_1, &res_unsigned_lo_1,
+                                       &wt0, &wt1, use_jnt_comp_avg);
+            const __m128i comp_avg_res_hi_0 =
+                highbd_comp_avg_sse4_1(&data_ref_0_hi_0, &res_unsigned_hi_0,
+                                       &wt0, &wt1, use_jnt_comp_avg);
+            const __m128i comp_avg_res_hi_1 =
+                highbd_comp_avg_sse4_1(&data_ref_0_hi_1, &res_unsigned_hi_1,
+                                       &wt0, &wt1, use_jnt_comp_avg);
+
+            const __m128i round_result_lo_0 =
+                highbd_convolve_rounding_sse2(&comp_avg_res_lo_0, &offset_const,
+                                              &rounding_const, rounding_shift);
+            const __m128i round_result_lo_1 =
+                highbd_convolve_rounding_sse2(&comp_avg_res_lo_1, &offset_const,
+                                              &rounding_const, rounding_shift);
+            const __m128i round_result_hi_0 =
+                highbd_convolve_rounding_sse2(&comp_avg_res_hi_0, &offset_const,
+                                              &rounding_const, rounding_shift);
+            const __m128i round_result_hi_1 =
+                highbd_convolve_rounding_sse2(&comp_avg_res_hi_1, &offset_const,
+                                              &rounding_const, rounding_shift);
+
+            const __m128i res_16b_0 =
+                _mm_packus_epi32(round_result_lo_0, round_result_hi_0);
+            const __m128i res_clip_0 =
+                _mm_min_epi16(res_16b_0, clip_pixel_to_bd);
+
+            const __m128i res_16b_1 =
+                _mm_packus_epi32(round_result_lo_1, round_result_hi_1);
+            const __m128i res_clip_1 =
+                _mm_min_epi16(res_16b_1, clip_pixel_to_bd);
+
+            _mm_store_si128((__m128i *)(&dst0[i * dst_stride0 + j]),
+                            res_clip_0);
+            _mm_store_si128(
+                (__m128i *)(&dst0[i * dst_stride0 + j + dst_stride0]),
+                res_clip_1);
+          } else {
+            __m128i res_16bit0 =
+                _mm_packus_epi32(res_unsigned_lo_0, res_unsigned_hi_0);
+            __m128i res_16bit1 =
+                _mm_packus_epi32(res_unsigned_lo_1, res_unsigned_hi_1);
+            _mm_store_si128((__m128i *)(&dst[i * dst_stride + j]), res_16bit0);
+            _mm_store_si128((__m128i *)(&dst[i * dst_stride + j + dst_stride]),
+                            res_16bit1);
+          }
+        }
+        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_jnt_convolve_x_sse4_1(
+    const uint16_t *src, int src_stride, uint16_t *dst0, int dst_stride0, int w,
+    int h, const InterpFilterParams *filter_params_x,
+    const InterpFilterParams *filter_params_y, const int subpel_x_q4,
+    const int subpel_y_q4, ConvolveParams *conv_params, int bd) {
+  CONV_BUF_TYPE *dst = conv_params->dst;
+  int dst_stride = conv_params->dst_stride;
+  const int fo_horiz = filter_params_x->taps / 2 - 1;
+  const uint16_t *const src_ptr = src - fo_horiz;
+  const int bits = FILTER_BITS - conv_params->round_1;
+  (void)filter_params_y;
+  (void)subpel_y_q4;
+
+  int i, j;
+  __m128i s[4], coeffs_x[4];
+
+  const int do_average = conv_params->do_average;
+  const int use_jnt_comp_avg = conv_params->use_jnt_comp_avg;
+  const int w0 = conv_params->fwd_offset;
+  const int w1 = conv_params->bck_offset;
+  const __m128i wt0 = _mm_set1_epi32(w0);
+  const __m128i wt1 = _mm_set1_epi32(w1);
+  const __m128i zero = _mm_setzero_si128();
+
+  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 __m128i round_shift_bits = _mm_cvtsi32_si128(bits);
+
+  const int offset_0 =
+      bd + 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+  const int offset = (1 << offset_0) + (1 << (offset_0 - 1));
+  const __m128i offset_const = _mm_set1_epi32(offset);
+  const int rounding_shift =
+      2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+  const __m128i rounding_const = _mm_set1_epi32((1 << rounding_shift) >> 1);
+  const __m128i clip_pixel_to_bd =
+      _mm_set1_epi16(bd == 10 ? 1023 : (bd == 12 ? 4095 : 255));
+
+  assert(bits >= 0);
+  prepare_coeffs(filter_params_x, subpel_x_q4, 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_sll_epi32(res_even, round_shift_bits);
+      res_odd = _mm_sll_epi32(res_odd, round_shift_bits);
+
+      __m128i res1 = _mm_unpacklo_epi32(res_even, res_odd);
+      __m128i res_unsigned_lo = _mm_add_epi32(res1, offset_const);
+      if (w - j < 8) {
+        if (do_average) {
+          const __m128i data_0 =
+              _mm_loadl_epi64((__m128i *)(&dst[i * dst_stride + j]));
+          const __m128i data_ref_0 = _mm_unpacklo_epi16(data_0, zero);
+
+          const __m128i comp_avg_res = highbd_comp_avg_sse4_1(
+              &data_ref_0, &res_unsigned_lo, &wt0, &wt1, use_jnt_comp_avg);
+          const __m128i round_result = highbd_convolve_rounding_sse2(
+              &comp_avg_res, &offset_const, &rounding_const, rounding_shift);
+
+          const __m128i res_16b = _mm_packus_epi32(round_result, round_result);
+          const __m128i res_clip = _mm_min_epi16(res_16b, clip_pixel_to_bd);
+          _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_clip);
+        } else {
+          __m128i res_16b = _mm_packus_epi32(res_unsigned_lo, res_unsigned_lo);
+          _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j], res_16b);
+        }
+      } else {
+        __m128i res2 = _mm_unpackhi_epi32(res_even, res_odd);
+        __m128i res_unsigned_hi = _mm_add_epi32(res2, offset_const);
+        if (do_average) {
+          const __m128i data_0 =
+              _mm_loadu_si128((__m128i *)(&dst[i * dst_stride + j]));
+          const __m128i data_ref_0_lo = _mm_unpacklo_epi16(data_0, zero);
+          const __m128i data_ref_0_hi = _mm_unpackhi_epi16(data_0, zero);
+
+          const __m128i comp_avg_res_lo = highbd_comp_avg_sse4_1(
+              &data_ref_0_lo, &res_unsigned_lo, &wt0, &wt1, use_jnt_comp_avg);
+          const __m128i comp_avg_res_hi = highbd_comp_avg_sse4_1(
+              &data_ref_0_hi, &res_unsigned_hi, &wt0, &wt1, use_jnt_comp_avg);
+
+          const __m128i round_result_lo = highbd_convolve_rounding_sse2(
+              &comp_avg_res_lo, &offset_const, &rounding_const, rounding_shift);
+          const __m128i round_result_hi = highbd_convolve_rounding_sse2(
+              &comp_avg_res_hi, &offset_const, &rounding_const, rounding_shift);
+
+          const __m128i res_16b =
+              _mm_packus_epi32(round_result_lo, round_result_hi);
+          const __m128i res_clip = _mm_min_epi16(res_16b, clip_pixel_to_bd);
+          _mm_store_si128((__m128i *)(&dst0[i * dst_stride0 + j]), res_clip);
+        } else {
+          __m128i res_16b = _mm_packus_epi32(res_unsigned_lo, res_unsigned_hi);
+          _mm_store_si128((__m128i *)(&dst[i * dst_stride + j]), res_16b);
+        }
+      }
+    }
+  }
+}
diff --git a/third_party/aom/av1/common/x86/highbd_txfm_utility_sse4.h b/third_party/aom/av1/common/x86/highbd_txfm_utility_sse4.h
new file mode 100644
index 000000000..6f24e5948
--- /dev/null
+++ b/third_party/aom/av1/common/x86/highbd_txfm_utility_sse4.h
@@ -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.
+ */
+
+#ifndef AOM_AV1_COMMON_X86_HIGHBD_TXFM_UTILITY_SSE4_H_
+#define AOM_AV1_COMMON_X86_HIGHBD_TXFM_UTILITY_SSE4_H_
+
+#include <smmintrin.h> /* SSE4.1 */
+
+#define TRANSPOSE_4X4(x0, x1, x2, x3, y0, y1, y2, y3) \
+  do {                                                \
+    __m128i u0, u1, u2, u3;                           \
+    u0 = _mm_unpacklo_epi32(x0, x1);                  \
+    u1 = _mm_unpackhi_epi32(x0, x1);                  \
+    u2 = _mm_unpacklo_epi32(x2, x3);                  \
+    u3 = _mm_unpackhi_epi32(x2, x3);                  \
+    y0 = _mm_unpacklo_epi64(u0, u2);                  \
+    y1 = _mm_unpackhi_epi64(u0, u2);                  \
+    y2 = _mm_unpacklo_epi64(u1, u3);                  \
+    y3 = _mm_unpackhi_epi64(u1, u3);                  \
+  } while (0)
+
+static INLINE void transpose_8x8(const __m128i *in, __m128i *out) {
+  TRANSPOSE_4X4(in[0], in[2], in[4], in[6], out[0], out[2], out[4], out[6]);
+  TRANSPOSE_4X4(in[1], in[3], in[5], in[7], out[8], out[10], out[12], out[14]);
+  TRANSPOSE_4X4(in[8], in[10], in[12], in[14], out[1], out[3], out[5], out[7]);
+  TRANSPOSE_4X4(in[9], in[11], in[13], in[15], out[9], out[11], out[13],
+                out[15]);
+}
+
+static INLINE void transpose_16x16(const __m128i *in, __m128i *out) {
+  // Upper left 8x8
+  TRANSPOSE_4X4(in[0], in[4], in[8], in[12], out[0], out[4], out[8], out[12]);
+  TRANSPOSE_4X4(in[1], in[5], in[9], in[13], out[16], out[20], out[24],
+                out[28]);
+  TRANSPOSE_4X4(in[16], in[20], in[24], in[28], out[1], out[5], out[9],
+                out[13]);
+  TRANSPOSE_4X4(in[17], in[21], in[25], in[29], out[17], out[21], out[25],
+                out[29]);
+
+  // Upper right 8x8
+  TRANSPOSE_4X4(in[2], in[6], in[10], in[14], out[32], out[36], out[40],
+                out[44]);
+  TRANSPOSE_4X4(in[3], in[7], in[11], in[15], out[48], out[52], out[56],
+                out[60]);
+  TRANSPOSE_4X4(in[18], in[22], in[26], in[30], out[33], out[37], out[41],
+                out[45]);
+  TRANSPOSE_4X4(in[19], in[23], in[27], in[31], out[49], out[53], out[57],
+                out[61]);
+
+  // Lower left 8x8
+  TRANSPOSE_4X4(in[32], in[36], in[40], in[44], out[2], out[6], out[10],
+                out[14]);
+  TRANSPOSE_4X4(in[33], in[37], in[41], in[45], out[18], out[22], out[26],
+                out[30]);
+  TRANSPOSE_4X4(in[48], in[52], in[56], in[60], out[3], out[7], out[11],
+                out[15]);
+  TRANSPOSE_4X4(in[49], in[53], in[57], in[61], out[19], out[23], out[27],
+                out[31]);
+  // Lower right 8x8
+  TRANSPOSE_4X4(in[34], in[38], in[42], in[46], out[34], out[38], out[42],
+                out[46]);
+  TRANSPOSE_4X4(in[35], in[39], in[43], in[47], out[50], out[54], out[58],
+                out[62]);
+  TRANSPOSE_4X4(in[50], in[54], in[58], in[62], out[35], out[39], out[43],
+                out[47]);
+  TRANSPOSE_4X4(in[51], in[55], in[59], in[63], out[51], out[55], out[59],
+                out[63]);
+}
+
+static INLINE void transpose_32x32(const __m128i *input, __m128i *output) {
+  for (int j = 0; j < 8; j++) {
+    for (int i = 0; i < 8; i++) {
+      TRANSPOSE_4X4(input[i * 32 + j + 0], input[i * 32 + j + 8],
+                    input[i * 32 + j + 16], input[i * 32 + j + 24],
+                    output[j * 32 + i + 0], output[j * 32 + i + 8],
+                    output[j * 32 + i + 16], output[j * 32 + i + 24]);
+    }
+  }
+}
+
+// Note:
+//  rounding = 1 << (bit - 1)
+static INLINE __m128i half_btf_sse4_1(const __m128i *w0, const __m128i *n0,
+                                      const __m128i *w1, const __m128i *n1,
+                                      const __m128i *rounding, int bit) {
+  __m128i x, y;
+
+  x = _mm_mullo_epi32(*w0, *n0);
+  y = _mm_mullo_epi32(*w1, *n1);
+  x = _mm_add_epi32(x, y);
+  x = _mm_add_epi32(x, *rounding);
+  x = _mm_srai_epi32(x, bit);
+  return x;
+}
+
+static INLINE __m128i half_btf_0_sse4_1(const __m128i *w0, const __m128i *n0,
+                                        const __m128i *rounding, int bit) {
+  __m128i x;
+
+  x = _mm_mullo_epi32(*w0, *n0);
+  x = _mm_add_epi32(x, *rounding);
+  x = _mm_srai_epi32(x, bit);
+  return x;
+}
+
+typedef void (*transform_1d_sse4_1)(__m128i *in, __m128i *out, int bit,
+                                    int do_cols, int bd, int out_shift);
+
+typedef void (*fwd_transform_1d_sse4_1)(__m128i *in, __m128i *out, int bit,
+                                        const int num_cols);
+
+void av1_highbd_inv_txfm2d_add_universe_sse4_1(const int32_t *input,
+                                               uint8_t *output, int stride,
+                                               TX_TYPE tx_type, TX_SIZE tx_size,
+                                               int eob, const int bd);
+
+#endif  // AOM_AV1_COMMON_X86_HIGHBD_TXFM_UTILITY_SSE4_H_
diff --git a/third_party/aom/av1/common/x86/highbd_warp_plane_sse4.c b/third_party/aom/av1/common/x86/highbd_warp_plane_sse4.c
new file mode 100644
index 000000000..4bcab0564
--- /dev/null
+++ b/third_party/aom/av1/common/x86/highbd_warp_plane_sse4.c
@@ -0,0 +1,624 @@
+/*
+ * 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>
+
+#include "config/av1_rtcd.h"
+
+#include "av1/common/warped_motion.h"
+
+static const uint8_t warp_highbd_arrange_bytes[16] = {
+  0, 2, 4, 6, 8, 10, 12, 14, 1, 3, 5, 7, 9, 11, 13, 15
+};
+
+static const uint8_t highbd_shuffle_alpha0_mask0[16] = {
+  0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3, 0, 1, 2, 3
+};
+static const uint8_t highbd_shuffle_alpha0_mask1[16] = {
+  4, 5, 6, 7, 4, 5, 6, 7, 4, 5, 6, 7, 4, 5, 6, 7
+};
+static const uint8_t highbd_shuffle_alpha0_mask2[16] = {
+  8, 9, 10, 11, 8, 9, 10, 11, 8, 9, 10, 11, 8, 9, 10, 11
+};
+static const uint8_t highbd_shuffle_alpha0_mask3[16] = {
+  12, 13, 14, 15, 12, 13, 14, 15, 12, 13, 14, 15, 12, 13, 14, 15
+};
+
+static INLINE void highbd_prepare_horizontal_filter_coeff(int alpha, int sx,
+                                                          __m128i *coeff) {
+  // Filter even-index pixels
+  const __m128i tmp_0 = _mm_loadu_si128(
+      (__m128i *)(warped_filter + ((sx + 0 * alpha) >> WARPEDDIFF_PREC_BITS)));
+  const __m128i tmp_2 = _mm_loadu_si128(
+      (__m128i *)(warped_filter + ((sx + 2 * alpha) >> WARPEDDIFF_PREC_BITS)));
+  const __m128i tmp_4 = _mm_loadu_si128(
+      (__m128i *)(warped_filter + ((sx + 4 * alpha) >> WARPEDDIFF_PREC_BITS)));
+  const __m128i tmp_6 = _mm_loadu_si128(
+      (__m128i *)(warped_filter + ((sx + 6 * alpha) >> WARPEDDIFF_PREC_BITS)));
+
+  // coeffs 0 1 0 1 2 3 2 3 for pixels 0, 2
+  const __m128i tmp_8 = _mm_unpacklo_epi32(tmp_0, tmp_2);
+  // coeffs 0 1 0 1 2 3 2 3 for pixels 4, 6
+  const __m128i tmp_10 = _mm_unpacklo_epi32(tmp_4, tmp_6);
+  // coeffs 4 5 4 5 6 7 6 7 for pixels 0, 2
+  const __m128i tmp_12 = _mm_unpackhi_epi32(tmp_0, tmp_2);
+  // coeffs 4 5 4 5 6 7 6 7 for pixels 4, 6
+  const __m128i tmp_14 = _mm_unpackhi_epi32(tmp_4, tmp_6);
+
+  // coeffs 0 1 0 1 0 1 0 1 for pixels 0, 2, 4, 6
+  coeff[0] = _mm_unpacklo_epi64(tmp_8, tmp_10);
+  // coeffs 2 3 2 3 2 3 2 3 for pixels 0, 2, 4, 6
+  coeff[2] = _mm_unpackhi_epi64(tmp_8, tmp_10);
+  // coeffs 4 5 4 5 4 5 4 5 for pixels 0, 2, 4, 6
+  coeff[4] = _mm_unpacklo_epi64(tmp_12, tmp_14);
+  // coeffs 6 7 6 7 6 7 6 7 for pixels 0, 2, 4, 6
+  coeff[6] = _mm_unpackhi_epi64(tmp_12, tmp_14);
+
+  // Filter odd-index pixels
+  const __m128i tmp_1 = _mm_loadu_si128(
+      (__m128i *)(warped_filter + ((sx + 1 * alpha) >> WARPEDDIFF_PREC_BITS)));
+  const __m128i tmp_3 = _mm_loadu_si128(
+      (__m128i *)(warped_filter + ((sx + 3 * alpha) >> WARPEDDIFF_PREC_BITS)));
+  const __m128i tmp_5 = _mm_loadu_si128(
+      (__m128i *)(warped_filter + ((sx + 5 * alpha) >> WARPEDDIFF_PREC_BITS)));
+  const __m128i tmp_7 = _mm_loadu_si128(
+      (__m128i *)(warped_filter + ((sx + 7 * alpha) >> WARPEDDIFF_PREC_BITS)));
+
+  const __m128i tmp_9 = _mm_unpacklo_epi32(tmp_1, tmp_3);
+  const __m128i tmp_11 = _mm_unpacklo_epi32(tmp_5, tmp_7);
+  const __m128i tmp_13 = _mm_unpackhi_epi32(tmp_1, tmp_3);
+  const __m128i tmp_15 = _mm_unpackhi_epi32(tmp_5, tmp_7);
+
+  coeff[1] = _mm_unpacklo_epi64(tmp_9, tmp_11);
+  coeff[3] = _mm_unpackhi_epi64(tmp_9, tmp_11);
+  coeff[5] = _mm_unpacklo_epi64(tmp_13, tmp_15);
+  coeff[7] = _mm_unpackhi_epi64(tmp_13, tmp_15);
+}
+
+static INLINE void highbd_prepare_horizontal_filter_coeff_alpha0(
+    int sx, __m128i *coeff) {
+  // Filter coeff
+  const __m128i tmp_0 = _mm_loadu_si128(
+      (__m128i *)(warped_filter + (sx >> WARPEDDIFF_PREC_BITS)));
+
+  coeff[0] = _mm_shuffle_epi8(
+      tmp_0, _mm_loadu_si128((__m128i *)highbd_shuffle_alpha0_mask0));
+  coeff[2] = _mm_shuffle_epi8(
+      tmp_0, _mm_loadu_si128((__m128i *)highbd_shuffle_alpha0_mask1));
+  coeff[4] = _mm_shuffle_epi8(
+      tmp_0, _mm_loadu_si128((__m128i *)highbd_shuffle_alpha0_mask2));
+  coeff[6] = _mm_shuffle_epi8(
+      tmp_0, _mm_loadu_si128((__m128i *)highbd_shuffle_alpha0_mask3));
+
+  coeff[1] = coeff[0];
+  coeff[3] = coeff[2];
+  coeff[5] = coeff[4];
+  coeff[7] = coeff[6];
+}
+
+static INLINE void highbd_filter_src_pixels(
+    const __m128i *src, const __m128i *src2, __m128i *tmp, __m128i *coeff,
+    const int offset_bits_horiz, const int reduce_bits_horiz, int k) {
+  const __m128i src_1 = *src;
+  const __m128i src2_1 = *src2;
+
+  const __m128i round_const = _mm_set1_epi32((1 << offset_bits_horiz) +
+                                             ((1 << reduce_bits_horiz) >> 1));
+
+  const __m128i res_0 = _mm_madd_epi16(src_1, coeff[0]);
+  const __m128i res_2 =
+      _mm_madd_epi16(_mm_alignr_epi8(src2_1, src_1, 4), coeff[2]);
+  const __m128i res_4 =
+      _mm_madd_epi16(_mm_alignr_epi8(src2_1, src_1, 8), coeff[4]);
+  const __m128i res_6 =
+      _mm_madd_epi16(_mm_alignr_epi8(src2_1, src_1, 12), coeff[6]);
+
+  __m128i res_even =
+      _mm_add_epi32(_mm_add_epi32(res_0, res_4), _mm_add_epi32(res_2, res_6));
+  res_even = _mm_sra_epi32(_mm_add_epi32(res_even, round_const),
+                           _mm_cvtsi32_si128(reduce_bits_horiz));
+
+  const __m128i res_1 =
+      _mm_madd_epi16(_mm_alignr_epi8(src2_1, src_1, 2), coeff[1]);
+  const __m128i res_3 =
+      _mm_madd_epi16(_mm_alignr_epi8(src2_1, src_1, 6), coeff[3]);
+  const __m128i res_5 =
+      _mm_madd_epi16(_mm_alignr_epi8(src2_1, src_1, 10), coeff[5]);
+  const __m128i res_7 =
+      _mm_madd_epi16(_mm_alignr_epi8(src2_1, src_1, 14), coeff[7]);
+
+  __m128i res_odd =
+      _mm_add_epi32(_mm_add_epi32(res_1, res_5), _mm_add_epi32(res_3, res_7));
+  res_odd = _mm_sra_epi32(_mm_add_epi32(res_odd, round_const),
+                          _mm_cvtsi32_si128(reduce_bits_horiz));
+
+  // Combine results into one register.
+  // We store the columns in the order 0, 2, 4, 6, 1, 3, 5, 7
+  // as this order helps with the vertical filter.
+  tmp[k + 7] = _mm_packs_epi32(res_even, res_odd);
+}
+
+static INLINE void highbd_horiz_filter(const __m128i *src, const __m128i *src2,
+                                       __m128i *tmp, int sx, int alpha, int k,
+                                       const int offset_bits_horiz,
+                                       const int reduce_bits_horiz) {
+  __m128i coeff[8];
+  highbd_prepare_horizontal_filter_coeff(alpha, sx, coeff);
+  highbd_filter_src_pixels(src, src2, tmp, coeff, offset_bits_horiz,
+                           reduce_bits_horiz, k);
+}
+
+static INLINE void highbd_warp_horizontal_filter_alpha0_beta0(
+    const uint16_t *ref, __m128i *tmp, int stride, int32_t ix4, int32_t iy4,
+    int32_t sx4, int alpha, int beta, int p_height, int height, int i,
+    const int offset_bits_horiz, const int reduce_bits_horiz) {
+  (void)beta;
+  (void)alpha;
+  int k;
+
+  __m128i coeff[8];
+  highbd_prepare_horizontal_filter_coeff_alpha0(sx4, coeff);
+
+  for (k = -7; k < AOMMIN(8, p_height - i); ++k) {
+    int iy = iy4 + k;
+    if (iy < 0)
+      iy = 0;
+    else if (iy > height - 1)
+      iy = height - 1;
+
+    // Load source pixels
+    const __m128i src =
+        _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7));
+    const __m128i src2 =
+        _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 + 1));
+    highbd_filter_src_pixels(&src, &src2, tmp, coeff, offset_bits_horiz,
+                             reduce_bits_horiz, k);
+  }
+}
+
+static INLINE void highbd_warp_horizontal_filter_alpha0(
+    const uint16_t *ref, __m128i *tmp, int stride, int32_t ix4, int32_t iy4,
+    int32_t sx4, int alpha, int beta, int p_height, int height, int i,
+    const int offset_bits_horiz, const int reduce_bits_horiz) {
+  (void)alpha;
+  int k;
+  for (k = -7; k < AOMMIN(8, p_height - i); ++k) {
+    int iy = iy4 + k;
+    if (iy < 0)
+      iy = 0;
+    else if (iy > height - 1)
+      iy = height - 1;
+    int sx = sx4 + beta * (k + 4);
+
+    // Load source pixels
+    const __m128i src =
+        _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7));
+    const __m128i src2 =
+        _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 + 1));
+
+    __m128i coeff[8];
+    highbd_prepare_horizontal_filter_coeff_alpha0(sx, coeff);
+    highbd_filter_src_pixels(&src, &src2, tmp, coeff, offset_bits_horiz,
+                             reduce_bits_horiz, k);
+  }
+}
+
+static INLINE void highbd_warp_horizontal_filter_beta0(
+    const uint16_t *ref, __m128i *tmp, int stride, int32_t ix4, int32_t iy4,
+    int32_t sx4, int alpha, int beta, int p_height, int height, int i,
+    const int offset_bits_horiz, const int reduce_bits_horiz) {
+  (void)beta;
+  int k;
+  __m128i coeff[8];
+  highbd_prepare_horizontal_filter_coeff(alpha, sx4, coeff);
+
+  for (k = -7; k < AOMMIN(8, p_height - i); ++k) {
+    int iy = iy4 + k;
+    if (iy < 0)
+      iy = 0;
+    else if (iy > height - 1)
+      iy = height - 1;
+
+    // Load source pixels
+    const __m128i src =
+        _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7));
+    const __m128i src2 =
+        _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 + 1));
+    highbd_filter_src_pixels(&src, &src2, tmp, coeff, offset_bits_horiz,
+                             reduce_bits_horiz, k);
+  }
+}
+
+static INLINE void highbd_warp_horizontal_filter(
+    const uint16_t *ref, __m128i *tmp, int stride, int32_t ix4, int32_t iy4,
+    int32_t sx4, int alpha, int beta, int p_height, int height, int i,
+    const int offset_bits_horiz, const int reduce_bits_horiz) {
+  int k;
+  for (k = -7; k < AOMMIN(8, p_height - i); ++k) {
+    int iy = iy4 + k;
+    if (iy < 0)
+      iy = 0;
+    else if (iy > height - 1)
+      iy = height - 1;
+    int sx = sx4 + beta * (k + 4);
+
+    // Load source pixels
+    const __m128i src =
+        _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7));
+    const __m128i src2 =
+        _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 + 1));
+
+    highbd_horiz_filter(&src, &src2, tmp, sx, alpha, k, offset_bits_horiz,
+                        reduce_bits_horiz);
+  }
+}
+
+static INLINE void highbd_prepare_warp_horizontal_filter(
+    const uint16_t *ref, __m128i *tmp, int stride, int32_t ix4, int32_t iy4,
+    int32_t sx4, int alpha, int beta, int p_height, int height, int i,
+    const int offset_bits_horiz, const int reduce_bits_horiz) {
+  if (alpha == 0 && beta == 0)
+    highbd_warp_horizontal_filter_alpha0_beta0(
+        ref, tmp, stride, ix4, iy4, sx4, alpha, beta, p_height, height, i,
+        offset_bits_horiz, reduce_bits_horiz);
+
+  else if (alpha == 0 && beta != 0)
+    highbd_warp_horizontal_filter_alpha0(ref, tmp, stride, ix4, iy4, sx4, alpha,
+                                         beta, p_height, height, i,
+                                         offset_bits_horiz, reduce_bits_horiz);
+
+  else if (alpha != 0 && beta == 0)
+    highbd_warp_horizontal_filter_beta0(ref, tmp, stride, ix4, iy4, sx4, alpha,
+                                        beta, p_height, height, i,
+                                        offset_bits_horiz, reduce_bits_horiz);
+  else
+    highbd_warp_horizontal_filter(ref, tmp, stride, ix4, iy4, sx4, alpha, beta,
+                                  p_height, height, i, offset_bits_horiz,
+                                  reduce_bits_horiz);
+}
+
+void av1_highbd_warp_affine_sse4_1(const int32_t *mat, const uint16_t *ref,
+                                   int width, int height, int stride,
+                                   uint16_t *pred, int p_col, int p_row,
+                                   int p_width, int p_height, int p_stride,
+                                   int subsampling_x, int subsampling_y, int bd,
+                                   ConvolveParams *conv_params, int16_t alpha,
+                                   int16_t beta, int16_t gamma, int16_t delta) {
+  __m128i tmp[15];
+  int i, j, k;
+  const int reduce_bits_horiz =
+      conv_params->round_0 +
+      AOMMAX(bd + FILTER_BITS - conv_params->round_0 - 14, 0);
+  const int reduce_bits_vert = conv_params->is_compound
+                                   ? conv_params->round_1
+                                   : 2 * FILTER_BITS - reduce_bits_horiz;
+  const int offset_bits_horiz = bd + FILTER_BITS - 1;
+  assert(IMPLIES(conv_params->is_compound, conv_params->dst != NULL));
+  assert(!(bd == 12 && reduce_bits_horiz < 5));
+  assert(IMPLIES(conv_params->do_average, conv_params->is_compound));
+
+  const int offset_bits_vert = bd + 2 * FILTER_BITS - reduce_bits_horiz;
+  const __m128i clip_pixel =
+      _mm_set1_epi16(bd == 10 ? 1023 : (bd == 12 ? 4095 : 255));
+  const __m128i reduce_bits_vert_shift = _mm_cvtsi32_si128(reduce_bits_vert);
+  const __m128i reduce_bits_vert_const =
+      _mm_set1_epi32(((1 << reduce_bits_vert) >> 1));
+  const __m128i res_add_const = _mm_set1_epi32(1 << offset_bits_vert);
+  const int round_bits =
+      2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+  const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0;
+  const __m128i res_sub_const =
+      _mm_set1_epi32(-(1 << (offset_bits - conv_params->round_1)) -
+                     (1 << (offset_bits - conv_params->round_1 - 1)));
+  __m128i round_bits_shift = _mm_cvtsi32_si128(round_bits);
+  __m128i round_bits_const = _mm_set1_epi32(((1 << round_bits) >> 1));
+
+  const int w0 = conv_params->fwd_offset;
+  const int w1 = conv_params->bck_offset;
+  const __m128i wt0 = _mm_set1_epi32(w0);
+  const __m128i wt1 = _mm_set1_epi32(w1);
+
+  /* Note: For this code to work, the left/right frame borders need to be
+  extended by at least 13 pixels each. By the time we get here, other
+  code will have set up this border, but we allow an explicit check
+  for debugging purposes.
+  */
+  /*for (i = 0; i < height; ++i) {
+  for (j = 0; j < 13; ++j) {
+  assert(ref[i * stride - 13 + j] == ref[i * stride]);
+  assert(ref[i * stride + width + j] == ref[i * stride + (width - 1)]);
+  }
+  }*/
+
+  for (i = 0; i < p_height; i += 8) {
+    for (j = 0; j < p_width; j += 8) {
+      const int32_t src_x = (p_col + j + 4) << subsampling_x;
+      const int32_t src_y = (p_row + i + 4) << subsampling_y;
+      const int32_t dst_x = mat[2] * src_x + mat[3] * src_y + mat[0];
+      const int32_t dst_y = mat[4] * src_x + mat[5] * src_y + mat[1];
+      const int32_t x4 = dst_x >> subsampling_x;
+      const int32_t y4 = dst_y >> subsampling_y;
+
+      int32_t ix4 = x4 >> WARPEDMODEL_PREC_BITS;
+      int32_t sx4 = x4 & ((1 << WARPEDMODEL_PREC_BITS) - 1);
+      int32_t iy4 = y4 >> WARPEDMODEL_PREC_BITS;
+      int32_t sy4 = y4 & ((1 << WARPEDMODEL_PREC_BITS) - 1);
+
+      // Add in all the constant terms, including rounding and offset
+      sx4 += alpha * (-4) + beta * (-4) + (1 << (WARPEDDIFF_PREC_BITS - 1)) +
+             (WARPEDPIXEL_PREC_SHIFTS << WARPEDDIFF_PREC_BITS);
+      sy4 += gamma * (-4) + delta * (-4) + (1 << (WARPEDDIFF_PREC_BITS - 1)) +
+             (WARPEDPIXEL_PREC_SHIFTS << WARPEDDIFF_PREC_BITS);
+
+      sx4 &= ~((1 << WARP_PARAM_REDUCE_BITS) - 1);
+      sy4 &= ~((1 << WARP_PARAM_REDUCE_BITS) - 1);
+
+      // Horizontal filter
+      // If the block is aligned such that, after clamping, every sample
+      // would be taken from the leftmost/rightmost column, then we can
+      // skip the expensive horizontal filter.
+      if (ix4 <= -7) {
+        for (k = -7; k < AOMMIN(8, p_height - i); ++k) {
+          int iy = iy4 + k;
+          if (iy < 0)
+            iy = 0;
+          else if (iy > height - 1)
+            iy = height - 1;
+          tmp[k + 7] = _mm_set1_epi16(
+              (1 << (bd + FILTER_BITS - reduce_bits_horiz - 1)) +
+              ref[iy * stride] * (1 << (FILTER_BITS - reduce_bits_horiz)));
+        }
+      } else if (ix4 >= width + 6) {
+        for (k = -7; k < AOMMIN(8, p_height - i); ++k) {
+          int iy = iy4 + k;
+          if (iy < 0)
+            iy = 0;
+          else if (iy > height - 1)
+            iy = height - 1;
+          tmp[k + 7] =
+              _mm_set1_epi16((1 << (bd + FILTER_BITS - reduce_bits_horiz - 1)) +
+                             ref[iy * stride + (width - 1)] *
+                                 (1 << (FILTER_BITS - reduce_bits_horiz)));
+        }
+      } else if (((ix4 - 7) < 0) || ((ix4 + 9) > width)) {
+        const int out_of_boundary_left = -(ix4 - 6);
+        const int out_of_boundary_right = (ix4 + 8) - width;
+
+        for (k = -7; k < AOMMIN(8, p_height - i); ++k) {
+          int iy = iy4 + k;
+          if (iy < 0)
+            iy = 0;
+          else if (iy > height - 1)
+            iy = height - 1;
+          int sx = sx4 + beta * (k + 4);
+
+          // Load source pixels
+          const __m128i src =
+              _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7));
+          const __m128i src2 =
+              _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 + 1));
+
+          const __m128i src_01 = _mm_shuffle_epi8(
+              src, _mm_loadu_si128((__m128i *)warp_highbd_arrange_bytes));
+          const __m128i src2_01 = _mm_shuffle_epi8(
+              src2, _mm_loadu_si128((__m128i *)warp_highbd_arrange_bytes));
+
+          __m128i src_lo = _mm_unpacklo_epi64(src_01, src2_01);
+          __m128i src_hi = _mm_unpackhi_epi64(src_01, src2_01);
+
+          if (out_of_boundary_left >= 0) {
+            const __m128i shuffle_reg_left =
+                _mm_loadu_si128((__m128i *)warp_pad_left[out_of_boundary_left]);
+            src_lo = _mm_shuffle_epi8(src_lo, shuffle_reg_left);
+            src_hi = _mm_shuffle_epi8(src_hi, shuffle_reg_left);
+          }
+
+          if (out_of_boundary_right >= 0) {
+            const __m128i shuffle_reg_right = _mm_loadu_si128(
+                (__m128i *)warp_pad_right[out_of_boundary_right]);
+            src_lo = _mm_shuffle_epi8(src_lo, shuffle_reg_right);
+            src_hi = _mm_shuffle_epi8(src_hi, shuffle_reg_right);
+          }
+
+          const __m128i src_padded = _mm_unpacklo_epi8(src_lo, src_hi);
+          const __m128i src2_padded = _mm_unpackhi_epi8(src_lo, src_hi);
+
+          highbd_horiz_filter(&src_padded, &src2_padded, tmp, sx, alpha, k,
+                              offset_bits_horiz, reduce_bits_horiz);
+        }
+      } else {
+        highbd_prepare_warp_horizontal_filter(
+            ref, tmp, stride, ix4, iy4, sx4, alpha, beta, p_height, height, i,
+            offset_bits_horiz, reduce_bits_horiz);
+      }
+
+      // Vertical filter
+      for (k = -4; k < AOMMIN(4, p_height - i - 4); ++k) {
+        int sy = sy4 + delta * (k + 4);
+
+        // Load from tmp and rearrange pairs of consecutive rows into the
+        // column order 0 0 2 2 4 4 6 6; 1 1 3 3 5 5 7 7
+        const __m128i *src = tmp + (k + 4);
+        const __m128i src_0 = _mm_unpacklo_epi16(src[0], src[1]);
+        const __m128i src_2 = _mm_unpacklo_epi16(src[2], src[3]);
+        const __m128i src_4 = _mm_unpacklo_epi16(src[4], src[5]);
+        const __m128i src_6 = _mm_unpacklo_epi16(src[6], src[7]);
+
+        // Filter even-index pixels
+        const __m128i tmp_0 = _mm_loadu_si128(
+            (__m128i *)(warped_filter +
+                        ((sy + 0 * gamma) >> WARPEDDIFF_PREC_BITS)));
+        const __m128i tmp_2 = _mm_loadu_si128(
+            (__m128i *)(warped_filter +
+                        ((sy + 2 * gamma) >> WARPEDDIFF_PREC_BITS)));
+        const __m128i tmp_4 = _mm_loadu_si128(
+            (__m128i *)(warped_filter +
+                        ((sy + 4 * gamma) >> WARPEDDIFF_PREC_BITS)));
+        const __m128i tmp_6 = _mm_loadu_si128(
+            (__m128i *)(warped_filter +
+                        ((sy + 6 * gamma) >> WARPEDDIFF_PREC_BITS)));
+
+        const __m128i tmp_8 = _mm_unpacklo_epi32(tmp_0, tmp_2);
+        const __m128i tmp_10 = _mm_unpacklo_epi32(tmp_4, tmp_6);
+        const __m128i tmp_12 = _mm_unpackhi_epi32(tmp_0, tmp_2);
+        const __m128i tmp_14 = _mm_unpackhi_epi32(tmp_4, tmp_6);
+
+        const __m128i coeff_0 = _mm_unpacklo_epi64(tmp_8, tmp_10);
+        const __m128i coeff_2 = _mm_unpackhi_epi64(tmp_8, tmp_10);
+        const __m128i coeff_4 = _mm_unpacklo_epi64(tmp_12, tmp_14);
+        const __m128i coeff_6 = _mm_unpackhi_epi64(tmp_12, tmp_14);
+
+        const __m128i res_0 = _mm_madd_epi16(src_0, coeff_0);
+        const __m128i res_2 = _mm_madd_epi16(src_2, coeff_2);
+        const __m128i res_4 = _mm_madd_epi16(src_4, coeff_4);
+        const __m128i res_6 = _mm_madd_epi16(src_6, coeff_6);
+
+        const __m128i res_even = _mm_add_epi32(_mm_add_epi32(res_0, res_2),
+                                               _mm_add_epi32(res_4, res_6));
+
+        // Filter odd-index pixels
+        const __m128i src_1 = _mm_unpackhi_epi16(src[0], src[1]);
+        const __m128i src_3 = _mm_unpackhi_epi16(src[2], src[3]);
+        const __m128i src_5 = _mm_unpackhi_epi16(src[4], src[5]);
+        const __m128i src_7 = _mm_unpackhi_epi16(src[6], src[7]);
+
+        const __m128i tmp_1 = _mm_loadu_si128(
+            (__m128i *)(warped_filter +
+                        ((sy + 1 * gamma) >> WARPEDDIFF_PREC_BITS)));
+        const __m128i tmp_3 = _mm_loadu_si128(
+            (__m128i *)(warped_filter +
+                        ((sy + 3 * gamma) >> WARPEDDIFF_PREC_BITS)));
+        const __m128i tmp_5 = _mm_loadu_si128(
+            (__m128i *)(warped_filter +
+                        ((sy + 5 * gamma) >> WARPEDDIFF_PREC_BITS)));
+        const __m128i tmp_7 = _mm_loadu_si128(
+            (__m128i *)(warped_filter +
+                        ((sy + 7 * gamma) >> WARPEDDIFF_PREC_BITS)));
+
+        const __m128i tmp_9 = _mm_unpacklo_epi32(tmp_1, tmp_3);
+        const __m128i tmp_11 = _mm_unpacklo_epi32(tmp_5, tmp_7);
+        const __m128i tmp_13 = _mm_unpackhi_epi32(tmp_1, tmp_3);
+        const __m128i tmp_15 = _mm_unpackhi_epi32(tmp_5, tmp_7);
+
+        const __m128i coeff_1 = _mm_unpacklo_epi64(tmp_9, tmp_11);
+        const __m128i coeff_3 = _mm_unpackhi_epi64(tmp_9, tmp_11);
+        const __m128i coeff_5 = _mm_unpacklo_epi64(tmp_13, tmp_15);
+        const __m128i coeff_7 = _mm_unpackhi_epi64(tmp_13, tmp_15);
+
+        const __m128i res_1 = _mm_madd_epi16(src_1, coeff_1);
+        const __m128i res_3 = _mm_madd_epi16(src_3, coeff_3);
+        const __m128i res_5 = _mm_madd_epi16(src_5, coeff_5);
+        const __m128i res_7 = _mm_madd_epi16(src_7, coeff_7);
+
+        const __m128i res_odd = _mm_add_epi32(_mm_add_epi32(res_1, res_3),
+                                              _mm_add_epi32(res_5, res_7));
+
+        // Rearrange pixels back into the order 0 ... 7
+        __m128i res_lo = _mm_unpacklo_epi32(res_even, res_odd);
+        __m128i res_hi = _mm_unpackhi_epi32(res_even, res_odd);
+
+        if (conv_params->is_compound) {
+          __m128i *const p =
+              (__m128i *)&conv_params
+                  ->dst[(i + k + 4) * conv_params->dst_stride + j];
+          res_lo = _mm_add_epi32(res_lo, res_add_const);
+          res_lo = _mm_sra_epi32(_mm_add_epi32(res_lo, reduce_bits_vert_const),
+                                 reduce_bits_vert_shift);
+
+          if (conv_params->do_average) {
+            __m128i *const dst16 = (__m128i *)&pred[(i + k + 4) * p_stride + j];
+            __m128i p_32 = _mm_cvtepu16_epi32(_mm_loadl_epi64(p));
+
+            if (conv_params->use_jnt_comp_avg) {
+              res_lo = _mm_add_epi32(_mm_mullo_epi32(p_32, wt0),
+                                     _mm_mullo_epi32(res_lo, wt1));
+              res_lo = _mm_srai_epi32(res_lo, DIST_PRECISION_BITS);
+            } else {
+              res_lo = _mm_srai_epi32(_mm_add_epi32(p_32, res_lo), 1);
+            }
+
+            __m128i res32_lo = _mm_add_epi32(res_lo, res_sub_const);
+            res32_lo = _mm_sra_epi32(_mm_add_epi32(res32_lo, round_bits_const),
+                                     round_bits_shift);
+
+            __m128i res16_lo = _mm_packus_epi32(res32_lo, res32_lo);
+            res16_lo = _mm_min_epi16(res16_lo, clip_pixel);
+            _mm_storel_epi64(dst16, res16_lo);
+          } else {
+            res_lo = _mm_packus_epi32(res_lo, res_lo);
+            _mm_storel_epi64(p, res_lo);
+          }
+          if (p_width > 4) {
+            __m128i *const p4 =
+                (__m128i *)&conv_params
+                    ->dst[(i + k + 4) * conv_params->dst_stride + j + 4];
+
+            res_hi = _mm_add_epi32(res_hi, res_add_const);
+            res_hi =
+                _mm_sra_epi32(_mm_add_epi32(res_hi, reduce_bits_vert_const),
+                              reduce_bits_vert_shift);
+            if (conv_params->do_average) {
+              __m128i *const dst16_4 =
+                  (__m128i *)&pred[(i + k + 4) * p_stride + j + 4];
+              __m128i p4_32 = _mm_cvtepu16_epi32(_mm_loadl_epi64(p4));
+
+              if (conv_params->use_jnt_comp_avg) {
+                res_hi = _mm_add_epi32(_mm_mullo_epi32(p4_32, wt0),
+                                       _mm_mullo_epi32(res_hi, wt1));
+                res_hi = _mm_srai_epi32(res_hi, DIST_PRECISION_BITS);
+              } else {
+                res_hi = _mm_srai_epi32(_mm_add_epi32(p4_32, res_hi), 1);
+              }
+
+              __m128i res32_hi = _mm_add_epi32(res_hi, res_sub_const);
+              res32_hi = _mm_sra_epi32(
+                  _mm_add_epi32(res32_hi, round_bits_const), round_bits_shift);
+              __m128i res16_hi = _mm_packus_epi32(res32_hi, res32_hi);
+              res16_hi = _mm_min_epi16(res16_hi, clip_pixel);
+              _mm_storel_epi64(dst16_4, res16_hi);
+            } else {
+              res_hi = _mm_packus_epi32(res_hi, res_hi);
+              _mm_storel_epi64(p4, res_hi);
+            }
+          }
+        } else {
+          // Round and pack into 8 bits
+          const __m128i round_const =
+              _mm_set1_epi32(-(1 << (bd + reduce_bits_vert - 1)) +
+                             ((1 << reduce_bits_vert) >> 1));
+
+          const __m128i res_lo_round = _mm_srai_epi32(
+              _mm_add_epi32(res_lo, round_const), reduce_bits_vert);
+          const __m128i res_hi_round = _mm_srai_epi32(
+              _mm_add_epi32(res_hi, round_const), reduce_bits_vert);
+
+          __m128i res_16bit = _mm_packs_epi32(res_lo_round, res_hi_round);
+          // Clamp res_16bit to the range [0, 2^bd - 1]
+          const __m128i max_val = _mm_set1_epi16((1 << bd) - 1);
+          const __m128i zero = _mm_setzero_si128();
+          res_16bit = _mm_max_epi16(_mm_min_epi16(res_16bit, max_val), zero);
+
+          // Store, blending with 'pred' if needed
+          __m128i *const p = (__m128i *)&pred[(i + k + 4) * p_stride + j];
+
+          // Note: If we're outputting a 4x4 block, we need to be very careful
+          // to only output 4 pixels at this point, to avoid encode/decode
+          // mismatches when encoding with multiple threads.
+          if (p_width == 4) {
+            _mm_storel_epi64(p, res_16bit);
+          } else {
+            _mm_storeu_si128(p, res_16bit);
+          }
+        }
+      }
+    }
+  }
+}
diff --git a/third_party/aom/av1/common/x86/highbd_wiener_convolve_avx2.c b/third_party/aom/av1/common/x86/highbd_wiener_convolve_avx2.c
new file mode 100644
index 000000000..0c8a8505b
--- /dev/null
+++ b/third_party/aom/av1/common/x86/highbd_wiener_convolve_avx2.c
@@ -0,0 +1,245 @@
+/*
+ * 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 <assert.h>
+
+#include "config/aom_dsp_rtcd.h"
+
+#include "av1/common/convolve.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/aom_filter.h"
+#include "aom_dsp/x86/synonyms.h"
+#include "aom_dsp/x86/synonyms_avx2.h"
+
+// 128-bit xmmwords are written as [ ... ] with the MSB on the left.
+// 256-bit ymmwords are written as two xmmwords, [ ... ][ ... ] with the MSB
+// on the left.
+// A row of, say, 16-bit pixels with values p0, p1, p2, ..., p14, p15 will be
+// loaded and stored as [ p15 ... p9 p8 ][ p7 ... p1 p0 ].
+void av1_highbd_wiener_convolve_add_src_avx2(
+    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,
+    const ConvolveParams *conv_params, int bd) {
+  assert(x_step_q4 == 16 && y_step_q4 == 16);
+  assert(!(w & 7));
+  assert(bd + FILTER_BITS - conv_params->round_0 + 2 <= 16);
+  (void)x_step_q4;
+  (void)y_step_q4;
+
+  const uint16_t *const src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *const dst = CONVERT_TO_SHORTPTR(dst8);
+
+  DECLARE_ALIGNED(32, uint16_t,
+                  temp[(MAX_SB_SIZE + SUBPEL_TAPS - 1) * MAX_SB_SIZE]);
+  int intermediate_height = h + SUBPEL_TAPS - 1;
+  const int center_tap = ((SUBPEL_TAPS - 1) / 2);
+  const uint16_t *const src_ptr = src - center_tap * src_stride - center_tap;
+
+  const __m128i zero_128 = _mm_setzero_si128();
+  const __m256i zero_256 = _mm256_setzero_si256();
+
+  // Add an offset to account for the "add_src" part of the convolve function.
+  const __m128i offset = _mm_insert_epi16(zero_128, 1 << FILTER_BITS, 3);
+
+  const __m256i clamp_low = zero_256;
+
+  /* Horizontal filter */
+  {
+    const __m256i clamp_high_ep =
+        _mm256_set1_epi16(WIENER_CLAMP_LIMIT(conv_params->round_0, bd) - 1);
+
+    // coeffs [ f7 f6 f5 f4 f3 f2 f1 f0 ]
+    const __m128i coeffs_x = _mm_add_epi16(xx_loadu_128(filter_x), offset);
+
+    // coeffs [ f3 f2 f3 f2 f1 f0 f1 f0 ]
+    const __m128i coeffs_0123 = _mm_unpacklo_epi32(coeffs_x, coeffs_x);
+    // coeffs [ f7 f6 f7 f6 f5 f4 f5 f4 ]
+    const __m128i coeffs_4567 = _mm_unpackhi_epi32(coeffs_x, coeffs_x);
+
+    // coeffs [ f1 f0 f1 f0 f1 f0 f1 f0 ]
+    const __m128i coeffs_01_128 = _mm_unpacklo_epi64(coeffs_0123, coeffs_0123);
+    // coeffs [ f3 f2 f3 f2 f3 f2 f3 f2 ]
+    const __m128i coeffs_23_128 = _mm_unpackhi_epi64(coeffs_0123, coeffs_0123);
+    // coeffs [ f5 f4 f5 f4 f5 f4 f5 f4 ]
+    const __m128i coeffs_45_128 = _mm_unpacklo_epi64(coeffs_4567, coeffs_4567);
+    // coeffs [ f7 f6 f7 f6 f7 f6 f7 f6 ]
+    const __m128i coeffs_67_128 = _mm_unpackhi_epi64(coeffs_4567, coeffs_4567);
+
+    // coeffs [ f1 f0 f1 f0 f1 f0 f1 f0 ][ f1 f0 f1 f0 f1 f0 f1 f0 ]
+    const __m256i coeffs_01 = yy_set_m128i(coeffs_01_128, coeffs_01_128);
+    // coeffs [ f3 f2 f3 f2 f3 f2 f3 f2 ][ f3 f2 f3 f2 f3 f2 f3 f2 ]
+    const __m256i coeffs_23 = yy_set_m128i(coeffs_23_128, coeffs_23_128);
+    // coeffs [ f5 f4 f5 f4 f5 f4 f5 f4 ][ f5 f4 f5 f4 f5 f4 f5 f4 ]
+    const __m256i coeffs_45 = yy_set_m128i(coeffs_45_128, coeffs_45_128);
+    // coeffs [ f7 f6 f7 f6 f7 f6 f7 f6 ][ f7 f6 f7 f6 f7 f6 f7 f6 ]
+    const __m256i coeffs_67 = yy_set_m128i(coeffs_67_128, coeffs_67_128);
+
+    const __m256i round_const = _mm256_set1_epi32(
+        (1 << (conv_params->round_0 - 1)) + (1 << (bd + FILTER_BITS - 1)));
+
+    for (int i = 0; i < intermediate_height; ++i) {
+      for (int j = 0; j < w; j += 16) {
+        const uint16_t *src_ij = src_ptr + i * src_stride + j;
+
+        // Load 16-bit src data
+        const __m256i src_0 = yy_loadu_256(src_ij + 0);
+        const __m256i src_1 = yy_loadu_256(src_ij + 1);
+        const __m256i src_2 = yy_loadu_256(src_ij + 2);
+        const __m256i src_3 = yy_loadu_256(src_ij + 3);
+        const __m256i src_4 = yy_loadu_256(src_ij + 4);
+        const __m256i src_5 = yy_loadu_256(src_ij + 5);
+        const __m256i src_6 = yy_loadu_256(src_ij + 6);
+        const __m256i src_7 = yy_loadu_256(src_ij + 7);
+
+        // Multiply src data by filter coeffs and sum pairs
+        const __m256i res_0 = _mm256_madd_epi16(src_0, coeffs_01);
+        const __m256i res_1 = _mm256_madd_epi16(src_1, coeffs_01);
+        const __m256i res_2 = _mm256_madd_epi16(src_2, coeffs_23);
+        const __m256i res_3 = _mm256_madd_epi16(src_3, coeffs_23);
+        const __m256i res_4 = _mm256_madd_epi16(src_4, coeffs_45);
+        const __m256i res_5 = _mm256_madd_epi16(src_5, coeffs_45);
+        const __m256i res_6 = _mm256_madd_epi16(src_6, coeffs_67);
+        const __m256i res_7 = _mm256_madd_epi16(src_7, coeffs_67);
+
+        // Calculate scalar product for even- and odd-indices separately,
+        // increasing to 32-bit precision
+        const __m256i res_even_sum = _mm256_add_epi32(
+            _mm256_add_epi32(res_0, res_4), _mm256_add_epi32(res_2, res_6));
+        const __m256i res_even = _mm256_srai_epi32(
+            _mm256_add_epi32(res_even_sum, round_const), conv_params->round_0);
+
+        const __m256i res_odd_sum = _mm256_add_epi32(
+            _mm256_add_epi32(res_1, res_5), _mm256_add_epi32(res_3, res_7));
+        const __m256i res_odd = _mm256_srai_epi32(
+            _mm256_add_epi32(res_odd_sum, round_const), conv_params->round_0);
+
+        // Reduce to 16-bit precision and pack even- and odd-index results
+        // back into one register. The _mm256_packs_epi32 intrinsic returns
+        // a register with the pixels ordered as follows:
+        // [ 15 13 11 9 14 12 10 8 ] [ 7 5 3 1 6 4 2 0 ]
+        const __m256i res = _mm256_packs_epi32(res_even, res_odd);
+        const __m256i res_clamped =
+            _mm256_min_epi16(_mm256_max_epi16(res, clamp_low), clamp_high_ep);
+
+        // Store in a temporary array
+        yy_storeu_256(temp + i * MAX_SB_SIZE + j, res_clamped);
+      }
+    }
+  }
+
+  /* Vertical filter */
+  {
+    const __m256i clamp_high = _mm256_set1_epi16((1 << bd) - 1);
+
+    // coeffs [ f7 f6 f5 f4 f3 f2 f1 f0 ]
+    const __m128i coeffs_y = _mm_add_epi16(xx_loadu_128(filter_y), offset);
+
+    // coeffs [ f3 f2 f3 f2 f1 f0 f1 f0 ]
+    const __m128i coeffs_0123 = _mm_unpacklo_epi32(coeffs_y, coeffs_y);
+    // coeffs [ f7 f6 f7 f6 f5 f4 f5 f4 ]
+    const __m128i coeffs_4567 = _mm_unpackhi_epi32(coeffs_y, coeffs_y);
+
+    // coeffs [ f1 f0 f1 f0 f1 f0 f1 f0 ]
+    const __m128i coeffs_01_128 = _mm_unpacklo_epi64(coeffs_0123, coeffs_0123);
+    // coeffs [ f3 f2 f3 f2 f3 f2 f3 f2 ]
+    const __m128i coeffs_23_128 = _mm_unpackhi_epi64(coeffs_0123, coeffs_0123);
+    // coeffs [ f5 f4 f5 f4 f5 f4 f5 f4 ]
+    const __m128i coeffs_45_128 = _mm_unpacklo_epi64(coeffs_4567, coeffs_4567);
+    // coeffs [ f7 f6 f7 f6 f7 f6 f7 f6 ]
+    const __m128i coeffs_67_128 = _mm_unpackhi_epi64(coeffs_4567, coeffs_4567);
+
+    // coeffs [ f1 f0 f1 f0 f1 f0 f1 f0 ][ f1 f0 f1 f0 f1 f0 f1 f0 ]
+    const __m256i coeffs_01 = yy_set_m128i(coeffs_01_128, coeffs_01_128);
+    // coeffs [ f3 f2 f3 f2 f3 f2 f3 f2 ][ f3 f2 f3 f2 f3 f2 f3 f2 ]
+    const __m256i coeffs_23 = yy_set_m128i(coeffs_23_128, coeffs_23_128);
+    // coeffs [ f5 f4 f5 f4 f5 f4 f5 f4 ][ f5 f4 f5 f4 f5 f4 f5 f4 ]
+    const __m256i coeffs_45 = yy_set_m128i(coeffs_45_128, coeffs_45_128);
+    // coeffs [ f7 f6 f7 f6 f7 f6 f7 f6 ][ f7 f6 f7 f6 f7 f6 f7 f6 ]
+    const __m256i coeffs_67 = yy_set_m128i(coeffs_67_128, coeffs_67_128);
+
+    const __m256i round_const =
+        _mm256_set1_epi32((1 << (conv_params->round_1 - 1)) -
+                          (1 << (bd + conv_params->round_1 - 1)));
+
+    for (int i = 0; i < h; ++i) {
+      for (int j = 0; j < w; j += 16) {
+        const uint16_t *temp_ij = temp + i * MAX_SB_SIZE + j;
+
+        // Load 16-bit data from the output of the horizontal filter in
+        // which the pixels are ordered as follows:
+        // [ 15 13 11 9 14 12 10 8 ] [ 7 5 3 1 6 4 2 0 ]
+        const __m256i data_0 = yy_loadu_256(temp_ij + 0 * MAX_SB_SIZE);
+        const __m256i data_1 = yy_loadu_256(temp_ij + 1 * MAX_SB_SIZE);
+        const __m256i data_2 = yy_loadu_256(temp_ij + 2 * MAX_SB_SIZE);
+        const __m256i data_3 = yy_loadu_256(temp_ij + 3 * MAX_SB_SIZE);
+        const __m256i data_4 = yy_loadu_256(temp_ij + 4 * MAX_SB_SIZE);
+        const __m256i data_5 = yy_loadu_256(temp_ij + 5 * MAX_SB_SIZE);
+        const __m256i data_6 = yy_loadu_256(temp_ij + 6 * MAX_SB_SIZE);
+        const __m256i data_7 = yy_loadu_256(temp_ij + 7 * MAX_SB_SIZE);
+
+        // Filter the even-indices, increasing to 32-bit precision
+        const __m256i src_0 = _mm256_unpacklo_epi16(data_0, data_1);
+        const __m256i src_2 = _mm256_unpacklo_epi16(data_2, data_3);
+        const __m256i src_4 = _mm256_unpacklo_epi16(data_4, data_5);
+        const __m256i src_6 = _mm256_unpacklo_epi16(data_6, data_7);
+
+        const __m256i res_0 = _mm256_madd_epi16(src_0, coeffs_01);
+        const __m256i res_2 = _mm256_madd_epi16(src_2, coeffs_23);
+        const __m256i res_4 = _mm256_madd_epi16(src_4, coeffs_45);
+        const __m256i res_6 = _mm256_madd_epi16(src_6, coeffs_67);
+
+        const __m256i res_even = _mm256_add_epi32(
+            _mm256_add_epi32(res_0, res_2), _mm256_add_epi32(res_4, res_6));
+
+        // Filter the odd-indices, increasing to 32-bit precision
+        const __m256i src_1 = _mm256_unpackhi_epi16(data_0, data_1);
+        const __m256i src_3 = _mm256_unpackhi_epi16(data_2, data_3);
+        const __m256i src_5 = _mm256_unpackhi_epi16(data_4, data_5);
+        const __m256i src_7 = _mm256_unpackhi_epi16(data_6, data_7);
+
+        const __m256i res_1 = _mm256_madd_epi16(src_1, coeffs_01);
+        const __m256i res_3 = _mm256_madd_epi16(src_3, coeffs_23);
+        const __m256i res_5 = _mm256_madd_epi16(src_5, coeffs_45);
+        const __m256i res_7 = _mm256_madd_epi16(src_7, coeffs_67);
+
+        const __m256i res_odd = _mm256_add_epi32(
+            _mm256_add_epi32(res_1, res_3), _mm256_add_epi32(res_5, res_7));
+
+        // Pixels are currently in the following order:
+        // res_even order: [ 14 12 10 8 ] [ 6 4 2 0 ]
+        // res_odd order:  [ 15 13 11 9 ] [ 7 5 3 1 ]
+        //
+        // Rearrange the pixels into the following order:
+        // res_lo order: [ 11 10  9  8 ] [ 3 2 1 0 ]
+        // res_hi order: [ 15 14 13 12 ] [ 7 6 5 4 ]
+        const __m256i res_lo = _mm256_unpacklo_epi32(res_even, res_odd);
+        const __m256i res_hi = _mm256_unpackhi_epi32(res_even, res_odd);
+
+        const __m256i res_lo_round = _mm256_srai_epi32(
+            _mm256_add_epi32(res_lo, round_const), conv_params->round_1);
+        const __m256i res_hi_round = _mm256_srai_epi32(
+            _mm256_add_epi32(res_hi, round_const), conv_params->round_1);
+
+        // Reduce to 16-bit precision and pack into the correct order:
+        // [ 15 14 13 12 11 10 9 8 ][ 7 6 5 4 3 2 1 0 ]
+        const __m256i res_16bit =
+            _mm256_packs_epi32(res_lo_round, res_hi_round);
+        const __m256i res_16bit_clamped = _mm256_min_epi16(
+            _mm256_max_epi16(res_16bit, clamp_low), clamp_high);
+
+        // Store in the dst array
+        yy_storeu_256(dst + i * dst_stride + j, res_16bit_clamped);
+      }
+    }
+  }
+}
diff --git a/third_party/aom/av1/common/x86/highbd_wiener_convolve_ssse3.c b/third_party/aom/av1/common/x86/highbd_wiener_convolve_ssse3.c
new file mode 100644
index 000000000..818b1099c
--- /dev/null
+++ b/third_party/aom/av1/common/x86/highbd_wiener_convolve_ssse3.c
@@ -0,0 +1,202 @@
+/*
+ * 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 <assert.h>
+
+#include "config/aom_dsp_rtcd.h"
+
+#include "av1/common/convolve.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/aom_filter.h"
+
+void av1_highbd_wiener_convolve_add_src_ssse3(
+    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,
+    const ConvolveParams *conv_params, int bd) {
+  assert(x_step_q4 == 16 && y_step_q4 == 16);
+  assert(!(w & 7));
+  assert(bd + FILTER_BITS - conv_params->round_0 + 2 <= 16);
+  (void)x_step_q4;
+  (void)y_step_q4;
+
+  const uint16_t *const src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *const dst = CONVERT_TO_SHORTPTR(dst8);
+
+  DECLARE_ALIGNED(16, uint16_t,
+                  temp[(MAX_SB_SIZE + SUBPEL_TAPS - 1) * MAX_SB_SIZE]);
+  int intermediate_height = h + SUBPEL_TAPS - 1;
+  int i, j;
+  const int center_tap = ((SUBPEL_TAPS - 1) / 2);
+  const uint16_t *const src_ptr = src - center_tap * src_stride - center_tap;
+
+  const __m128i zero = _mm_setzero_si128();
+  // Add an offset to account for the "add_src" part of the convolve function.
+  const __m128i offset = _mm_insert_epi16(zero, 1 << FILTER_BITS, 3);
+
+  /* Horizontal filter */
+  {
+    const __m128i coeffs_x =
+        _mm_add_epi16(_mm_loadu_si128((__m128i *)filter_x), offset);
+
+    // coeffs 0 1 0 1 2 3 2 3
+    const __m128i tmp_0 = _mm_unpacklo_epi32(coeffs_x, coeffs_x);
+    // coeffs 4 5 4 5 6 7 6 7
+    const __m128i tmp_1 = _mm_unpackhi_epi32(coeffs_x, coeffs_x);
+
+    // coeffs 0 1 0 1 0 1 0 1
+    const __m128i coeff_01 = _mm_unpacklo_epi64(tmp_0, tmp_0);
+    // coeffs 2 3 2 3 2 3 2 3
+    const __m128i coeff_23 = _mm_unpackhi_epi64(tmp_0, tmp_0);
+    // coeffs 4 5 4 5 4 5 4 5
+    const __m128i coeff_45 = _mm_unpacklo_epi64(tmp_1, tmp_1);
+    // coeffs 6 7 6 7 6 7 6 7
+    const __m128i coeff_67 = _mm_unpackhi_epi64(tmp_1, tmp_1);
+
+    const __m128i round_const = _mm_set1_epi32(
+        (1 << (conv_params->round_0 - 1)) + (1 << (bd + FILTER_BITS - 1)));
+
+    for (i = 0; i < intermediate_height; ++i) {
+      for (j = 0; j < w; j += 8) {
+        const __m128i data =
+            _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + j]);
+        const __m128i data2 =
+            _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + j + 8]);
+
+        // Filter even-index pixels
+        const __m128i res_0 = _mm_madd_epi16(data, coeff_01);
+        const __m128i res_2 =
+            _mm_madd_epi16(_mm_alignr_epi8(data2, data, 4), coeff_23);
+        const __m128i res_4 =
+            _mm_madd_epi16(_mm_alignr_epi8(data2, data, 8), coeff_45);
+        const __m128i res_6 =
+            _mm_madd_epi16(_mm_alignr_epi8(data2, data, 12), coeff_67);
+
+        __m128i res_even = _mm_add_epi32(_mm_add_epi32(res_0, res_4),
+                                         _mm_add_epi32(res_2, res_6));
+        res_even = _mm_srai_epi32(_mm_add_epi32(res_even, round_const),
+                                  conv_params->round_0);
+
+        // Filter odd-index pixels
+        const __m128i res_1 =
+            _mm_madd_epi16(_mm_alignr_epi8(data2, data, 2), coeff_01);
+        const __m128i res_3 =
+            _mm_madd_epi16(_mm_alignr_epi8(data2, data, 6), coeff_23);
+        const __m128i res_5 =
+            _mm_madd_epi16(_mm_alignr_epi8(data2, data, 10), coeff_45);
+        const __m128i res_7 =
+            _mm_madd_epi16(_mm_alignr_epi8(data2, data, 14), coeff_67);
+
+        __m128i res_odd = _mm_add_epi32(_mm_add_epi32(res_1, res_5),
+                                        _mm_add_epi32(res_3, res_7));
+        res_odd = _mm_srai_epi32(_mm_add_epi32(res_odd, round_const),
+                                 conv_params->round_0);
+
+        // Pack in the column order 0, 2, 4, 6, 1, 3, 5, 7
+        const __m128i maxval =
+            _mm_set1_epi16((WIENER_CLAMP_LIMIT(conv_params->round_0, bd)) - 1);
+        __m128i res = _mm_packs_epi32(res_even, res_odd);
+        res = _mm_min_epi16(_mm_max_epi16(res, zero), maxval);
+        _mm_storeu_si128((__m128i *)&temp[i * MAX_SB_SIZE + j], res);
+      }
+    }
+  }
+
+  /* Vertical filter */
+  {
+    const __m128i coeffs_y =
+        _mm_add_epi16(_mm_loadu_si128((__m128i *)filter_y), offset);
+
+    // coeffs 0 1 0 1 2 3 2 3
+    const __m128i tmp_0 = _mm_unpacklo_epi32(coeffs_y, coeffs_y);
+    // coeffs 4 5 4 5 6 7 6 7
+    const __m128i tmp_1 = _mm_unpackhi_epi32(coeffs_y, coeffs_y);
+
+    // coeffs 0 1 0 1 0 1 0 1
+    const __m128i coeff_01 = _mm_unpacklo_epi64(tmp_0, tmp_0);
+    // coeffs 2 3 2 3 2 3 2 3
+    const __m128i coeff_23 = _mm_unpackhi_epi64(tmp_0, tmp_0);
+    // coeffs 4 5 4 5 4 5 4 5
+    const __m128i coeff_45 = _mm_unpacklo_epi64(tmp_1, tmp_1);
+    // coeffs 6 7 6 7 6 7 6 7
+    const __m128i coeff_67 = _mm_unpackhi_epi64(tmp_1, tmp_1);
+
+    const __m128i round_const =
+        _mm_set1_epi32((1 << (conv_params->round_1 - 1)) -
+                       (1 << (bd + conv_params->round_1 - 1)));
+
+    for (i = 0; i < h; ++i) {
+      for (j = 0; j < w; j += 8) {
+        // Filter even-index pixels
+        const uint16_t *data = &temp[i * MAX_SB_SIZE + j];
+        const __m128i src_0 =
+            _mm_unpacklo_epi16(*(__m128i *)(data + 0 * MAX_SB_SIZE),
+                               *(__m128i *)(data + 1 * MAX_SB_SIZE));
+        const __m128i src_2 =
+            _mm_unpacklo_epi16(*(__m128i *)(data + 2 * MAX_SB_SIZE),
+                               *(__m128i *)(data + 3 * MAX_SB_SIZE));
+        const __m128i src_4 =
+            _mm_unpacklo_epi16(*(__m128i *)(data + 4 * MAX_SB_SIZE),
+                               *(__m128i *)(data + 5 * MAX_SB_SIZE));
+        const __m128i src_6 =
+            _mm_unpacklo_epi16(*(__m128i *)(data + 6 * MAX_SB_SIZE),
+                               *(__m128i *)(data + 7 * MAX_SB_SIZE));
+
+        const __m128i res_0 = _mm_madd_epi16(src_0, coeff_01);
+        const __m128i res_2 = _mm_madd_epi16(src_2, coeff_23);
+        const __m128i res_4 = _mm_madd_epi16(src_4, coeff_45);
+        const __m128i res_6 = _mm_madd_epi16(src_6, coeff_67);
+
+        const __m128i res_even = _mm_add_epi32(_mm_add_epi32(res_0, res_2),
+                                               _mm_add_epi32(res_4, res_6));
+
+        // Filter odd-index pixels
+        const __m128i src_1 =
+            _mm_unpackhi_epi16(*(__m128i *)(data + 0 * MAX_SB_SIZE),
+                               *(__m128i *)(data + 1 * MAX_SB_SIZE));
+        const __m128i src_3 =
+            _mm_unpackhi_epi16(*(__m128i *)(data + 2 * MAX_SB_SIZE),
+                               *(__m128i *)(data + 3 * MAX_SB_SIZE));
+        const __m128i src_5 =
+            _mm_unpackhi_epi16(*(__m128i *)(data + 4 * MAX_SB_SIZE),
+                               *(__m128i *)(data + 5 * MAX_SB_SIZE));
+        const __m128i src_7 =
+            _mm_unpackhi_epi16(*(__m128i *)(data + 6 * MAX_SB_SIZE),
+                               *(__m128i *)(data + 7 * MAX_SB_SIZE));
+
+        const __m128i res_1 = _mm_madd_epi16(src_1, coeff_01);
+        const __m128i res_3 = _mm_madd_epi16(src_3, coeff_23);
+        const __m128i res_5 = _mm_madd_epi16(src_5, coeff_45);
+        const __m128i res_7 = _mm_madd_epi16(src_7, coeff_67);
+
+        const __m128i res_odd = _mm_add_epi32(_mm_add_epi32(res_1, res_3),
+                                              _mm_add_epi32(res_5, res_7));
+
+        // Rearrange pixels back into the order 0 ... 7
+        const __m128i res_lo = _mm_unpacklo_epi32(res_even, res_odd);
+        const __m128i res_hi = _mm_unpackhi_epi32(res_even, res_odd);
+
+        const __m128i res_lo_round = _mm_srai_epi32(
+            _mm_add_epi32(res_lo, round_const), conv_params->round_1);
+        const __m128i res_hi_round = _mm_srai_epi32(
+            _mm_add_epi32(res_hi, round_const), conv_params->round_1);
+
+        const __m128i maxval = _mm_set1_epi16((1 << bd) - 1);
+        __m128i res_16bit = _mm_packs_epi32(res_lo_round, res_hi_round);
+        res_16bit = _mm_min_epi16(_mm_max_epi16(res_16bit, zero), maxval);
+
+        __m128i *const p = (__m128i *)&dst[i * dst_stride + j];
+        _mm_storeu_si128(p, res_16bit);
+      }
+    }
+  }
+}
diff --git a/third_party/aom/av1/common/x86/intra_edge_sse4.c b/third_party/aom/av1/common/x86/intra_edge_sse4.c
new file mode 100644
index 000000000..0c857b583
--- /dev/null
+++ b/third_party/aom/av1/common/x86/intra_edge_sse4.c
@@ -0,0 +1,318 @@
+/*
+ * 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 <assert.h>
+#include <smmintrin.h>
+
+#include "config/aom_config.h"
+#include "config/av1_rtcd.h"
+
+void av1_filter_intra_edge_sse4_1(uint8_t *p, int sz, int strength) {
+  if (!strength) return;
+
+  DECLARE_ALIGNED(16, static const int8_t, kern[3][16]) = {
+    { 4, 8, 4, 0, 4, 8, 4, 0, 4, 8, 4, 0, 4, 8, 4, 0 },  // strength 1: 4,8,4
+    { 5, 6, 5, 0, 5, 6, 5, 0, 5, 6, 5, 0, 5, 6, 5, 0 },  // strength 2: 5,6,5
+    { 2, 4, 4, 4, 2, 0, 0, 0, 2, 4, 4, 4, 2, 0, 0, 0 }  // strength 3: 2,4,4,4,2
+  };
+
+  DECLARE_ALIGNED(16, static const int8_t, v_const[5][16]) = {
+    { 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 },
+    { 0, 1, 2, 3, 4, 5, 6, 7, 1, 2, 3, 4, 5, 6, 7, 8 },
+    { 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15 },
+  };
+
+  // Extend the first and last samples to simplify the loop for the 5-tap case
+  p[-1] = p[0];
+  __m128i last = _mm_set1_epi8(p[sz - 1]);
+  _mm_storeu_si128((__m128i *)&p[sz], last);
+
+  // Adjust input pointer for filter support area
+  uint8_t *in = (strength == 3) ? p - 1 : p;
+
+  // Avoid modifying first sample
+  uint8_t *out = p + 1;
+  int len = sz - 1;
+
+  const int use_3tap_filter = (strength < 3);
+
+  if (use_3tap_filter) {
+    __m128i coef0 = _mm_lddqu_si128((__m128i const *)kern[strength - 1]);
+    __m128i shuf0 = _mm_lddqu_si128((__m128i const *)v_const[0]);
+    __m128i shuf1 = _mm_lddqu_si128((__m128i const *)v_const[1]);
+    __m128i iden = _mm_lddqu_si128((__m128i *)v_const[3]);
+    __m128i in0 = _mm_lddqu_si128((__m128i *)in);
+    while (len > 0) {
+      int n_out = (len < 8) ? len : 8;
+      __m128i d0 = _mm_shuffle_epi8(in0, shuf0);
+      __m128i d1 = _mm_shuffle_epi8(in0, shuf1);
+      d0 = _mm_maddubs_epi16(d0, coef0);
+      d1 = _mm_maddubs_epi16(d1, coef0);
+      d0 = _mm_hadd_epi16(d0, d1);
+      __m128i eight = _mm_set1_epi16(8);
+      d0 = _mm_add_epi16(d0, eight);
+      d0 = _mm_srai_epi16(d0, 4);
+      d0 = _mm_packus_epi16(d0, d0);
+      __m128i out0 = _mm_lddqu_si128((__m128i *)out);
+      __m128i n0 = _mm_set1_epi8(n_out);
+      __m128i mask = _mm_cmpgt_epi8(n0, iden);
+      out0 = _mm_blendv_epi8(out0, d0, mask);
+      _mm_storel_epi64((__m128i *)out, out0);
+      __m128i in1 = _mm_lddqu_si128((__m128i *)(in + 16));
+      in0 = _mm_alignr_epi8(in1, in0, 8);
+      in += 8;
+      out += 8;
+      len -= n_out;
+    }
+  } else {  // 5-tap filter
+    __m128i coef0 = _mm_lddqu_si128((__m128i const *)kern[strength - 1]);
+    __m128i two = _mm_set1_epi8(2);
+    __m128i shuf_a = _mm_lddqu_si128((__m128i const *)v_const[2]);
+    __m128i shuf_b = _mm_add_epi8(shuf_a, two);
+    __m128i shuf_c = _mm_add_epi8(shuf_b, two);
+    __m128i shuf_d = _mm_add_epi8(shuf_c, two);
+    __m128i iden = _mm_lddqu_si128((__m128i *)v_const[3]);
+    __m128i in0 = _mm_lddqu_si128((__m128i *)in);
+    while (len > 0) {
+      int n_out = (len < 8) ? len : 8;
+      __m128i d0 = _mm_shuffle_epi8(in0, shuf_a);
+      __m128i d1 = _mm_shuffle_epi8(in0, shuf_b);
+      __m128i d2 = _mm_shuffle_epi8(in0, shuf_c);
+      __m128i d3 = _mm_shuffle_epi8(in0, shuf_d);
+      d0 = _mm_maddubs_epi16(d0, coef0);
+      d1 = _mm_maddubs_epi16(d1, coef0);
+      d2 = _mm_maddubs_epi16(d2, coef0);
+      d3 = _mm_maddubs_epi16(d3, coef0);
+      d0 = _mm_hadd_epi16(d0, d1);
+      d2 = _mm_hadd_epi16(d2, d3);
+      d0 = _mm_hadd_epi16(d0, d2);
+      __m128i eight = _mm_set1_epi16(8);
+      d0 = _mm_add_epi16(d0, eight);
+      d0 = _mm_srai_epi16(d0, 4);
+      d0 = _mm_packus_epi16(d0, d0);
+      __m128i out0 = _mm_lddqu_si128((__m128i *)out);
+      __m128i n0 = _mm_set1_epi8(n_out);
+      __m128i mask = _mm_cmpgt_epi8(n0, iden);
+      out0 = _mm_blendv_epi8(out0, d0, mask);
+      _mm_storel_epi64((__m128i *)out, out0);
+      __m128i in1 = _mm_lddqu_si128((__m128i *)(in + 16));
+      in0 = _mm_alignr_epi8(in1, in0, 8);
+      in += 8;
+      out += 8;
+      len -= n_out;
+    }
+  }
+}
+
+void av1_filter_intra_edge_high_sse4_1(uint16_t *p, int sz, int strength) {
+  if (!strength) return;
+
+  DECLARE_ALIGNED(16, static const int16_t, kern[3][8]) = {
+    { 4, 8, 4, 8, 4, 8, 4, 8 },  // strength 1: 4,8,4
+    { 5, 6, 5, 6, 5, 6, 5, 6 },  // strength 2: 5,6,5
+    { 2, 4, 2, 4, 2, 4, 2, 4 }   // strength 3: 2,4,4,4,2
+  };
+
+  DECLARE_ALIGNED(16, static const int16_t,
+                  v_const[1][8]) = { { 0, 1, 2, 3, 4, 5, 6, 7 } };
+
+  // Extend the first and last samples to simplify the loop for the 5-tap case
+  p[-1] = p[0];
+  __m128i last = _mm_set1_epi16(p[sz - 1]);
+  _mm_storeu_si128((__m128i *)&p[sz], last);
+
+  // Adjust input pointer for filter support area
+  uint16_t *in = (strength == 3) ? p - 1 : p;
+
+  // Avoid modifying first sample
+  uint16_t *out = p + 1;
+  int len = sz - 1;
+
+  const int use_3tap_filter = (strength < 3);
+
+  if (use_3tap_filter) {
+    __m128i coef0 = _mm_lddqu_si128((__m128i const *)kern[strength - 1]);
+    __m128i iden = _mm_lddqu_si128((__m128i *)v_const[0]);
+    __m128i in0 = _mm_lddqu_si128((__m128i *)&in[0]);
+    __m128i in8 = _mm_lddqu_si128((__m128i *)&in[8]);
+    while (len > 0) {
+      int n_out = (len < 8) ? len : 8;
+      __m128i in1 = _mm_alignr_epi8(in8, in0, 2);
+      __m128i in2 = _mm_alignr_epi8(in8, in0, 4);
+      __m128i in02 = _mm_add_epi16(in0, in2);
+      __m128i d0 = _mm_unpacklo_epi16(in02, in1);
+      __m128i d1 = _mm_unpackhi_epi16(in02, in1);
+      d0 = _mm_mullo_epi16(d0, coef0);
+      d1 = _mm_mullo_epi16(d1, coef0);
+      d0 = _mm_hadd_epi16(d0, d1);
+      __m128i eight = _mm_set1_epi16(8);
+      d0 = _mm_add_epi16(d0, eight);
+      d0 = _mm_srli_epi16(d0, 4);
+      __m128i out0 = _mm_lddqu_si128((__m128i *)out);
+      __m128i n0 = _mm_set1_epi16(n_out);
+      __m128i mask = _mm_cmpgt_epi16(n0, iden);
+      out0 = _mm_blendv_epi8(out0, d0, mask);
+      _mm_storeu_si128((__m128i *)out, out0);
+      in += 8;
+      in0 = in8;
+      in8 = _mm_lddqu_si128((__m128i *)&in[8]);
+      out += 8;
+      len -= n_out;
+    }
+  } else {  // 5-tap filter
+    __m128i coef0 = _mm_lddqu_si128((__m128i const *)kern[strength - 1]);
+    __m128i iden = _mm_lddqu_si128((__m128i *)v_const[0]);
+    __m128i in0 = _mm_lddqu_si128((__m128i *)&in[0]);
+    __m128i in8 = _mm_lddqu_si128((__m128i *)&in[8]);
+    while (len > 0) {
+      int n_out = (len < 8) ? len : 8;
+      __m128i in1 = _mm_alignr_epi8(in8, in0, 2);
+      __m128i in2 = _mm_alignr_epi8(in8, in0, 4);
+      __m128i in3 = _mm_alignr_epi8(in8, in0, 6);
+      __m128i in4 = _mm_alignr_epi8(in8, in0, 8);
+      __m128i in04 = _mm_add_epi16(in0, in4);
+      __m128i in123 = _mm_add_epi16(in1, in2);
+      in123 = _mm_add_epi16(in123, in3);
+      __m128i d0 = _mm_unpacklo_epi16(in04, in123);
+      __m128i d1 = _mm_unpackhi_epi16(in04, in123);
+      d0 = _mm_mullo_epi16(d0, coef0);
+      d1 = _mm_mullo_epi16(d1, coef0);
+      d0 = _mm_hadd_epi16(d0, d1);
+      __m128i eight = _mm_set1_epi16(8);
+      d0 = _mm_add_epi16(d0, eight);
+      d0 = _mm_srli_epi16(d0, 4);
+      __m128i out0 = _mm_lddqu_si128((__m128i *)out);
+      __m128i n0 = _mm_set1_epi16(n_out);
+      __m128i mask = _mm_cmpgt_epi16(n0, iden);
+      out0 = _mm_blendv_epi8(out0, d0, mask);
+      _mm_storeu_si128((__m128i *)out, out0);
+      in += 8;
+      in0 = in8;
+      in8 = _mm_lddqu_si128((__m128i *)&in[8]);
+      out += 8;
+      len -= n_out;
+    }
+  }
+}
+
+void av1_upsample_intra_edge_sse4_1(uint8_t *p, int sz) {
+  // interpolate half-sample positions
+  assert(sz <= 24);
+
+  DECLARE_ALIGNED(16, static const int8_t, kernel[1][16]) = {
+    { -1, 9, 9, -1, -1, 9, 9, -1, -1, 9, 9, -1, -1, 9, 9, -1 }
+  };
+
+  DECLARE_ALIGNED(16, static const int8_t, v_const[2][16]) = {
+    { 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 }
+  };
+
+  // Extend first/last samples (upper-left p[-1], last p[sz-1])
+  // to support 4-tap filter
+  p[-2] = p[-1];
+  p[sz] = p[sz - 1];
+
+  uint8_t *in = &p[-2];
+  uint8_t *out = &p[-2];
+
+  int n = sz + 1;  // Input length including upper-left sample
+
+  __m128i in0 = _mm_lddqu_si128((__m128i *)&in[0]);
+  __m128i in16 = _mm_lddqu_si128((__m128i *)&in[16]);
+
+  __m128i coef0 = _mm_lddqu_si128((__m128i *)kernel[0]);
+  __m128i shuf0 = _mm_lddqu_si128((__m128i *)v_const[0]);
+  __m128i shuf1 = _mm_lddqu_si128((__m128i *)v_const[1]);
+
+  while (n > 0) {
+    __m128i in8 = _mm_alignr_epi8(in16, in0, 8);
+    __m128i d0 = _mm_shuffle_epi8(in0, shuf0);
+    __m128i d1 = _mm_shuffle_epi8(in0, shuf1);
+    __m128i d2 = _mm_shuffle_epi8(in8, shuf0);
+    __m128i d3 = _mm_shuffle_epi8(in8, shuf1);
+    d0 = _mm_maddubs_epi16(d0, coef0);
+    d1 = _mm_maddubs_epi16(d1, coef0);
+    d2 = _mm_maddubs_epi16(d2, coef0);
+    d3 = _mm_maddubs_epi16(d3, coef0);
+    d0 = _mm_hadd_epi16(d0, d1);
+    d2 = _mm_hadd_epi16(d2, d3);
+    __m128i eight = _mm_set1_epi16(8);
+    d0 = _mm_add_epi16(d0, eight);
+    d2 = _mm_add_epi16(d2, eight);
+    d0 = _mm_srai_epi16(d0, 4);
+    d2 = _mm_srai_epi16(d2, 4);
+    d0 = _mm_packus_epi16(d0, d2);
+    __m128i in1 = _mm_alignr_epi8(in16, in0, 1);
+    __m128i out0 = _mm_unpacklo_epi8(in1, d0);
+    __m128i out1 = _mm_unpackhi_epi8(in1, d0);
+    _mm_storeu_si128((__m128i *)&out[0], out0);
+    _mm_storeu_si128((__m128i *)&out[16], out1);
+    in0 = in16;
+    in16 = _mm_setzero_si128();
+    out += 32;
+    n -= 16;
+  }
+}
+
+void av1_upsample_intra_edge_high_sse4_1(uint16_t *p, int sz, int bd) {
+  // interpolate half-sample positions
+  assert(sz <= 24);
+
+  DECLARE_ALIGNED(16, static const int16_t,
+                  kernel[1][8]) = { { -1, 9, -1, 9, -1, 9, -1, 9 } };
+
+  // Extend first/last samples (upper-left p[-1], last p[sz-1])
+  // to support 4-tap filter
+  p[-2] = p[-1];
+  p[sz] = p[sz - 1];
+
+  uint16_t *in = &p[-2];
+  uint16_t *out = in;
+  int n = sz + 1;
+
+  __m128i in0 = _mm_lddqu_si128((__m128i *)&in[0]);
+  __m128i in8 = _mm_lddqu_si128((__m128i *)&in[8]);
+  __m128i in16 = _mm_lddqu_si128((__m128i *)&in[16]);
+  __m128i in24 = _mm_lddqu_si128((__m128i *)&in[24]);
+
+  while (n > 0) {
+    __m128i in1 = _mm_alignr_epi8(in8, in0, 2);
+    __m128i in2 = _mm_alignr_epi8(in8, in0, 4);
+    __m128i in3 = _mm_alignr_epi8(in8, in0, 6);
+    __m128i sum0 = _mm_add_epi16(in0, in3);
+    __m128i sum1 = _mm_add_epi16(in1, in2);
+    __m128i d0 = _mm_unpacklo_epi16(sum0, sum1);
+    __m128i d1 = _mm_unpackhi_epi16(sum0, sum1);
+    __m128i coef0 = _mm_lddqu_si128((__m128i *)kernel[0]);
+    d0 = _mm_madd_epi16(d0, coef0);
+    d1 = _mm_madd_epi16(d1, coef0);
+    __m128i eight = _mm_set1_epi32(8);
+    d0 = _mm_add_epi32(d0, eight);
+    d1 = _mm_add_epi32(d1, eight);
+    d0 = _mm_srai_epi32(d0, 4);
+    d1 = _mm_srai_epi32(d1, 4);
+    d0 = _mm_packus_epi32(d0, d1);
+    __m128i max0 = _mm_set1_epi16((1 << bd) - 1);
+    d0 = _mm_min_epi16(d0, max0);
+    __m128i out0 = _mm_unpacklo_epi16(in1, d0);
+    __m128i out1 = _mm_unpackhi_epi16(in1, d0);
+    _mm_storeu_si128((__m128i *)&out[0], out0);
+    _mm_storeu_si128((__m128i *)&out[8], out1);
+    in0 = in8;
+    in8 = in16;
+    in16 = in24;
+    in24 = _mm_setzero_si128();
+    out += 16;
+    n -= 8;
+  }
+}
diff --git a/third_party/aom/av1/common/x86/jnt_convolve_avx2.c b/third_party/aom/av1/common/x86/jnt_convolve_avx2.c
new file mode 100644
index 000000000..9f2e2b457
--- /dev/null
+++ b/third_party/aom/av1/common/x86/jnt_convolve_avx2.c
@@ -0,0 +1,633 @@
+/*
+ * 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/x86/convolve_avx2.h"
+#include "aom_dsp/x86/convolve_common_intrin.h"
+#include "aom_dsp/x86/convolve_sse4_1.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/aom_filter.h"
+#include "av1/common/convolve.h"
+
+static INLINE __m256i unpack_weights_avx2(ConvolveParams *conv_params) {
+  const int w0 = conv_params->fwd_offset;
+  const int w1 = conv_params->bck_offset;
+  const __m256i wt0 = _mm256_set1_epi16(w0);
+  const __m256i wt1 = _mm256_set1_epi16(w1);
+  const __m256i wt = _mm256_unpacklo_epi16(wt0, wt1);
+  return wt;
+}
+
+static INLINE __m256i load_line2_avx2(const void *a, const void *b) {
+  return _mm256_permute2x128_si256(
+      _mm256_castsi128_si256(_mm_loadu_si128((__m128i *)a)),
+      _mm256_castsi128_si256(_mm_loadu_si128((__m128i *)b)), 0x20);
+}
+
+void av1_jnt_convolve_x_avx2(const uint8_t *src, int src_stride, uint8_t *dst0,
+                             int dst_stride0, int w, int h,
+                             const InterpFilterParams *filter_params_x,
+                             const InterpFilterParams *filter_params_y,
+                             const int subpel_x_q4, const int subpel_y_q4,
+                             ConvolveParams *conv_params) {
+  CONV_BUF_TYPE *dst = conv_params->dst;
+  int dst_stride = conv_params->dst_stride;
+  const int bd = 8;
+  int i, j;
+  const int fo_horiz = filter_params_x->taps / 2 - 1;
+  const uint8_t *const src_ptr = src - fo_horiz;
+  const int bits = FILTER_BITS - conv_params->round_1;
+  const __m256i wt = unpack_weights_avx2(conv_params);
+  const int do_average = conv_params->do_average;
+  const int use_jnt_comp_avg = conv_params->use_jnt_comp_avg;
+  const int offset_0 =
+      bd + 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+  const int offset = (1 << offset_0) + (1 << (offset_0 - 1));
+  const __m256i offset_const = _mm256_set1_epi16(offset);
+  const int rounding_shift =
+      2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+  const __m256i rounding_const = _mm256_set1_epi16((1 << rounding_shift) >> 1);
+  __m256i filt[4], coeffs[4];
+
+  assert(bits >= 0);
+  assert(conv_params->round_0 > 0);
+
+  filt[0] = _mm256_load_si256((__m256i const *)filt_global_avx2);
+  filt[1] = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32));
+  filt[2] = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32 * 2));
+  filt[3] = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32 * 3));
+
+  prepare_coeffs_lowbd(filter_params_x, subpel_x_q4, coeffs);
+
+  const __m256i round_const =
+      _mm256_set1_epi16((1 << (conv_params->round_0 - 1)) >> 1);
+  const __m128i round_shift = _mm_cvtsi32_si128(conv_params->round_0 - 1);
+
+  (void)filter_params_y;
+  (void)subpel_y_q4;
+
+  for (i = 0; i < h; i += 2) {
+    const uint8_t *src_data = src_ptr + i * src_stride;
+    CONV_BUF_TYPE *dst_data = dst + i * dst_stride;
+    for (j = 0; j < w; j += 8) {
+      const __m256i data =
+          load_line2_avx2(&src_data[j], &src_data[j + src_stride]);
+
+      __m256i res = convolve_lowbd_x(data, coeffs, filt);
+
+      res = _mm256_sra_epi16(_mm256_add_epi16(res, round_const), round_shift);
+
+      res = _mm256_slli_epi16(res, bits);
+
+      const __m256i res_unsigned = _mm256_add_epi16(res, offset_const);
+
+      // Accumulate values into the destination buffer
+      if (do_average) {
+        const __m256i data_ref_0 =
+            load_line2_avx2(&dst_data[j], &dst_data[j + dst_stride]);
+        const __m256i comp_avg_res =
+            comp_avg(&data_ref_0, &res_unsigned, &wt, use_jnt_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);
+
+        if (w > 4) {
+          _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 {
+          *(uint32_t *)(&dst0[i * dst_stride0 + j]) = _mm_cvtsi128_si32(res_0);
+          *(uint32_t *)(&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);
+      }
+    }
+  }
+}
+
+void av1_jnt_convolve_y_avx2(const uint8_t *src, int src_stride, uint8_t *dst0,
+                             int dst_stride0, int w, int h,
+                             const InterpFilterParams *filter_params_x,
+                             const InterpFilterParams *filter_params_y,
+                             const int subpel_x_q4, const int subpel_y_q4,
+                             ConvolveParams *conv_params) {
+  CONV_BUF_TYPE *dst = conv_params->dst;
+  int dst_stride = conv_params->dst_stride;
+  const int bd = 8;
+  int i, j;
+  const int fo_vert = filter_params_y->taps / 2 - 1;
+  const uint8_t *const src_ptr = src - fo_vert * src_stride;
+  // +1 to compensate for dividing the filter coeffs by 2
+  const int left_shift = FILTER_BITS - conv_params->round_0 + 1;
+  const __m256i round_const =
+      _mm256_set1_epi32((1 << conv_params->round_1) >> 1);
+  const __m128i round_shift = _mm_cvtsi32_si128(conv_params->round_1);
+  const __m256i wt = unpack_weights_avx2(conv_params);
+  const int do_average = conv_params->do_average;
+  const int use_jnt_comp_avg = conv_params->use_jnt_comp_avg;
+  const int offset_0 =
+      bd + 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+  const int offset = (1 << offset_0) + (1 << (offset_0 - 1));
+  const __m256i offset_const = _mm256_set1_epi16(offset);
+  const int offset_1 = (1 << (bd + FILTER_BITS - 2));
+  const __m256i offset_const_1 = _mm256_set1_epi16(offset_1);
+  const __m256i offset_const_2 = _mm256_set1_epi16((1 << offset_0));
+  const int rounding_shift =
+      2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+  const __m256i rounding_const = _mm256_set1_epi16((1 << rounding_shift) >> 1);
+  const __m256i zero = _mm256_setzero_si256();
+  __m256i coeffs[4], s[8];
+
+  assert((FILTER_BITS - conv_params->round_0) >= 0);
+
+  prepare_coeffs_lowbd(filter_params_y, subpel_y_q4, coeffs);
+
+  (void)conv_params;
+  (void)filter_params_x;
+  (void)subpel_x_q4;
+
+  for (j = 0; j < w; j += 16) {
+    const uint8_t *data = &src_ptr[j];
+    __m256i src6;
+    // Load lines a and b. Line a to lower 128, line b to upper 128
+    {
+      __m256i src_ab[7];
+      __m256i src_a[7];
+      src_a[0] = _mm256_castsi128_si256(_mm_loadu_si128((__m128i *)data));
+      for (int kk = 0; kk < 6; ++kk) {
+        data += src_stride;
+        src_a[kk + 1] =
+            _mm256_castsi128_si256(_mm_loadu_si128((__m128i *)data));
+        src_ab[kk] = _mm256_permute2x128_si256(src_a[kk], src_a[kk + 1], 0x20);
+      }
+      src6 = src_a[6];
+      s[0] = _mm256_unpacklo_epi8(src_ab[0], src_ab[1]);
+      s[1] = _mm256_unpacklo_epi8(src_ab[2], src_ab[3]);
+      s[2] = _mm256_unpacklo_epi8(src_ab[4], src_ab[5]);
+      s[4] = _mm256_unpackhi_epi8(src_ab[0], src_ab[1]);
+      s[5] = _mm256_unpackhi_epi8(src_ab[2], src_ab[3]);
+      s[6] = _mm256_unpackhi_epi8(src_ab[4], src_ab[5]);
+    }
+
+    for (i = 0; i < h; i += 2) {
+      data = &src_ptr[(i + 7) * src_stride + j];
+      const __m256i src7 =
+          _mm256_castsi128_si256(_mm_loadu_si128((__m128i *)data));
+      const __m256i src_67a = _mm256_permute2x128_si256(src6, src7, 0x20);
+
+      src6 = _mm256_castsi128_si256(
+          _mm_loadu_si128((__m128i *)(data + src_stride)));
+      const __m256i src_78a = _mm256_permute2x128_si256(src7, src6, 0x20);
+
+      s[3] = _mm256_unpacklo_epi8(src_67a, src_78a);
+      s[7] = _mm256_unpackhi_epi8(src_67a, src_78a);
+
+      __m256i res_lo = convolve_lowbd(s, coeffs);
+
+      res_lo = _mm256_add_epi16(res_lo, offset_const_1);
+
+      const __m256i res_lo_0_32b = _mm256_unpacklo_epi16(res_lo, zero);
+      const __m256i res_lo_0_shift =
+          _mm256_slli_epi32(res_lo_0_32b, left_shift);
+      const __m256i res_lo_0_round = _mm256_sra_epi32(
+          _mm256_add_epi32(res_lo_0_shift, round_const), round_shift);
+
+      const __m256i res_lo_1_32b = _mm256_unpackhi_epi16(res_lo, zero);
+      const __m256i res_lo_1_shift =
+          _mm256_slli_epi32(res_lo_1_32b, left_shift);
+      const __m256i res_lo_1_round = _mm256_sra_epi32(
+          _mm256_add_epi32(res_lo_1_shift, round_const), round_shift);
+
+      const __m256i res_lo_round =
+          _mm256_packs_epi32(res_lo_0_round, res_lo_1_round);
+
+      const __m256i res_lo_unsigned =
+          _mm256_add_epi16(res_lo_round, offset_const_2);
+
+      if (w - j < 16) {
+        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_lo_unsigned, &wt, use_jnt_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);
+
+          if (w - j > 4) {
+            _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 {
+            *(uint32_t *)(&dst0[i * dst_stride0 + j]) =
+                _mm_cvtsi128_si32(res_0);
+            *(uint32_t *)(&dst0[i * dst_stride0 + j + dst_stride0]) =
+                _mm_cvtsi128_si32(res_1);
+          }
+        } else {
+          const __m128i res_0 = _mm256_castsi256_si128(res_lo_unsigned);
+          _mm_store_si128((__m128i *)(&dst[i * dst_stride + j]), res_0);
+
+          const __m128i res_1 = _mm256_extracti128_si256(res_lo_unsigned, 1);
+          _mm_store_si128((__m128i *)(&dst[i * dst_stride + j + dst_stride]),
+                          res_1);
+        }
+      } else {
+        __m256i res_hi = convolve_lowbd(s + 4, coeffs);
+
+        res_hi = _mm256_add_epi16(res_hi, offset_const_1);
+
+        const __m256i res_hi_0_32b = _mm256_unpacklo_epi16(res_hi, zero);
+        const __m256i res_hi_0_shift =
+            _mm256_slli_epi32(res_hi_0_32b, left_shift);
+        const __m256i res_hi_0_round = _mm256_sra_epi32(
+            _mm256_add_epi32(res_hi_0_shift, round_const), round_shift);
+
+        const __m256i res_hi_1_32b = _mm256_unpackhi_epi16(res_hi, zero);
+        const __m256i res_hi_1_shift =
+            _mm256_slli_epi32(res_hi_1_32b, left_shift);
+        const __m256i res_hi_1_round = _mm256_sra_epi32(
+            _mm256_add_epi32(res_hi_1_shift, round_const), round_shift);
+
+        const __m256i res_hi_round =
+            _mm256_packs_epi32(res_hi_0_round, res_hi_1_round);
+
+        const __m256i res_hi_unsigned =
+            _mm256_add_epi16(res_hi_round, offset_const_2);
+
+        if (do_average) {
+          const __m256i data_ref_0_lo = load_line2_avx2(
+              &dst[i * dst_stride + j], &dst[i * dst_stride + j + dst_stride]);
+
+          const __m256i data_ref_0_hi =
+              load_line2_avx2(&dst[i * dst_stride + j + 8],
+                              &dst[i * dst_stride + j + 8 + dst_stride]);
+
+          const __m256i comp_avg_res_lo =
+              comp_avg(&data_ref_0_lo, &res_lo_unsigned, &wt, use_jnt_comp_avg);
+
+          const __m256i comp_avg_res_hi =
+              comp_avg(&data_ref_0_hi, &res_hi_unsigned, &wt, use_jnt_comp_avg);
+
+          const __m256i round_result_lo = convolve_rounding(
+              &comp_avg_res_lo, &offset_const, &rounding_const, rounding_shift);
+
+          const __m256i round_result_hi = convolve_rounding(
+              &comp_avg_res_hi, &offset_const, &rounding_const, rounding_shift);
+
+          const __m256i res_8 =
+              _mm256_packus_epi16(round_result_lo, round_result_hi);
+          const __m128i res_0 = _mm256_castsi256_si128(res_8);
+          const __m128i res_1 = _mm256_extracti128_si256(res_8, 1);
+
+          _mm_store_si128((__m128i *)(&dst0[i * dst_stride0 + j]), res_0);
+          _mm_store_si128(
+              (__m128i *)((&dst0[i * dst_stride0 + j + dst_stride0])), res_1);
+
+        } else {
+          const __m128i res_lo_0 = _mm256_castsi256_si128(res_lo_unsigned);
+          _mm_store_si128((__m128i *)(&dst[i * dst_stride + j]), res_lo_0);
+
+          const __m128i res_lo_1 = _mm256_extracti128_si256(res_lo_unsigned, 1);
+          _mm_store_si128((__m128i *)(&dst[i * dst_stride + j + dst_stride]),
+                          res_lo_1);
+
+          const __m128i res_hi_0 = _mm256_castsi256_si128(res_hi_unsigned);
+          _mm_store_si128((__m128i *)(&dst[i * dst_stride + j + 8]), res_hi_0);
+
+          const __m128i res_hi_1 = _mm256_extracti128_si256(res_hi_unsigned, 1);
+          _mm_store_si128(
+              (__m128i *)(&dst[i * dst_stride + j + 8 + dst_stride]), res_hi_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_jnt_convolve_2d_avx2(const uint8_t *src, int src_stride, uint8_t *dst0,
+                              int dst_stride0, int w, int h,
+                              const InterpFilterParams *filter_params_x,
+                              const InterpFilterParams *filter_params_y,
+                              const int subpel_x_q4, const int subpel_y_q4,
+                              ConvolveParams *conv_params) {
+  CONV_BUF_TYPE *dst = conv_params->dst;
+  int dst_stride = conv_params->dst_stride;
+  const int bd = 8;
+
+  DECLARE_ALIGNED(32, int16_t, im_block[(MAX_SB_SIZE + MAX_FILTER_TAP) * 8]);
+  int im_h = h + filter_params_y->taps - 1;
+  int im_stride = 8;
+  int i, j;
+  const int fo_vert = filter_params_y->taps / 2 - 1;
+  const int fo_horiz = filter_params_x->taps / 2 - 1;
+  const uint8_t *const src_ptr = src - fo_vert * src_stride - fo_horiz;
+  const __m256i wt = unpack_weights_avx2(conv_params);
+  const int do_average = conv_params->do_average;
+  const int use_jnt_comp_avg = conv_params->use_jnt_comp_avg;
+  const int offset_0 =
+      bd + 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+  const int offset = (1 << offset_0) + (1 << (offset_0 - 1));
+  const __m256i offset_const = _mm256_set1_epi16(offset);
+  const int rounding_shift =
+      2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+  const __m256i rounding_const = _mm256_set1_epi16((1 << rounding_shift) >> 1);
+  __m256i filt[4], s[8], coeffs_x[4], coeffs_y[4];
+
+  assert(conv_params->round_0 > 0);
+
+  filt[0] = _mm256_load_si256((__m256i const *)filt_global_avx2);
+  filt[1] = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32));
+  filt[2] = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32 * 2));
+  filt[3] = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32 * 3));
+
+  prepare_coeffs_lowbd(filter_params_x, subpel_x_q4, coeffs_x);
+  prepare_coeffs(filter_params_y, subpel_y_q4, coeffs_y);
+
+  const __m256i round_const_h = _mm256_set1_epi16(
+      ((1 << (conv_params->round_0 - 1)) >> 1) + (1 << (bd + FILTER_BITS - 2)));
+  const __m128i round_shift_h = _mm_cvtsi32_si128(conv_params->round_0 - 1);
+
+  const __m256i round_const_v = _mm256_set1_epi32(
+      ((1 << conv_params->round_1) >> 1) -
+      (1 << (bd + 2 * FILTER_BITS - conv_params->round_0 - 1)));
+  const __m128i round_shift_v = _mm_cvtsi32_si128(conv_params->round_1);
+
+  for (j = 0; j < w; j += 8) {
+    /* Horizontal filter */
+    {
+      const uint8_t *src_h = src_ptr + j;
+      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);
+      }
+    }
+
+    /* Vertical filter */
+    {
+      __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_jnt_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_jnt_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);
+
+            *(uint32_t *)(&dst0[i * dst_stride0 + j]) =
+                _mm_cvtsi128_si32(res_0);
+            *(uint32_t *)(&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];
+      }
+    }
+  }
+}
+
+void av1_jnt_convolve_2d_copy_avx2(const uint8_t *src, int src_stride,
+                                   uint8_t *dst0, int dst_stride0, int w, int h,
+                                   const InterpFilterParams *filter_params_x,
+                                   const InterpFilterParams *filter_params_y,
+                                   const int subpel_x_q4, const int subpel_y_q4,
+                                   ConvolveParams *conv_params) {
+  const int bd = 8;
+  CONV_BUF_TYPE *dst = conv_params->dst;
+  int dst_stride = conv_params->dst_stride;
+  (void)filter_params_x;
+  (void)filter_params_y;
+  (void)subpel_x_q4;
+  (void)subpel_y_q4;
+
+  const int bits =
+      FILTER_BITS * 2 - conv_params->round_1 - conv_params->round_0;
+  const __m128i left_shift = _mm_cvtsi32_si128(bits);
+  const int do_average = conv_params->do_average;
+  const int use_jnt_comp_avg = conv_params->use_jnt_comp_avg;
+  const __m256i wt = unpack_weights_avx2(conv_params);
+  const __m256i zero = _mm256_setzero_si256();
+
+  const int offset_0 =
+      bd + 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+  const int offset = (1 << offset_0) + (1 << (offset_0 - 1));
+  const __m256i offset_const = _mm256_set1_epi16(offset);
+  const int rounding_shift =
+      2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+  const __m256i rounding_const = _mm256_set1_epi16((1 << rounding_shift) >> 1);
+  int i, j;
+
+  if (!(w % 16)) {
+    for (i = 0; i < h; i += 1) {
+      for (j = 0; j < w; j += 16) {
+        const __m256i src_16bit = _mm256_cvtepu8_epi16(
+            _mm_loadu_si128((__m128i *)(&src[i * src_stride + j])));
+
+        const __m256i res = _mm256_sll_epi16(src_16bit, left_shift);
+        const __m256i res_unsigned = _mm256_add_epi16(res, offset_const);
+
+        if (do_average) {
+          const __m256i data_ref_0 =
+              _mm256_loadu_si256((__m256i *)(&dst[i * dst_stride + j]));
+
+          const __m256i comp_avg_res =
+              comp_avg(&data_ref_0, &res_unsigned, &wt, use_jnt_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 __m256i res_0 = _mm256_permute4x64_epi64(res_8, 0xD8);
+
+          _mm_store_si128((__m128i *)(&dst0[i * dst_stride0 + j]),
+                          _mm256_castsi256_si128(res_0));
+        } else {
+          _mm256_store_si256((__m256i *)(&dst[i * dst_stride + j]),
+                             res_unsigned);
+        }
+      }
+    }
+  } else if (!(w % 4)) {
+    for (i = 0; i < h; i += 2) {
+      for (j = 0; j < w; j += 8) {
+        const __m128i src_row_0 =
+            _mm_loadl_epi64((__m128i *)(&src[i * src_stride + j]));
+        const __m128i src_row_1 =
+            _mm_loadl_epi64((__m128i *)(&src[i * src_stride + j + src_stride]));
+        // since not all compilers yet support _mm256_set_m128i()
+        const __m256i src_10 = _mm256_insertf128_si256(
+            _mm256_castsi128_si256(src_row_0), src_row_1, 1);
+
+        const __m256i src_16bit = _mm256_unpacklo_epi8(src_10, zero);
+
+        const __m256i res = _mm256_sll_epi16(src_16bit, left_shift);
+
+        const __m256i res_unsigned = _mm256_add_epi16(res, offset_const);
+
+        // Accumulate values into the destination buffer
+        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_jnt_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);
+
+          if (w > 4) {
+            _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 {
+            *(uint32_t *)(&dst0[i * dst_stride0 + j]) =
+                _mm_cvtsi128_si32(res_0);
+            *(uint32_t *)(&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);
+        }
+      }
+    }
+  }
+}
diff --git a/third_party/aom/av1/common/x86/jnt_convolve_sse2.c b/third_party/aom/av1/common/x86/jnt_convolve_sse2.c
new file mode 100644
index 000000000..87dc3242e
--- /dev/null
+++ b/third_party/aom/av1/common/x86/jnt_convolve_sse2.c
@@ -0,0 +1,385 @@
+/*
+ * 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/aom_filter.h"
+#include "aom_dsp/x86/convolve_sse2.h"
+
+void av1_jnt_convolve_x_sse2(const uint8_t *src, int src_stride, uint8_t *dst0,
+                             int dst_stride0, int w, int h,
+                             const InterpFilterParams *filter_params_x,
+                             const InterpFilterParams *filter_params_y,
+                             const int subpel_x_q4, const int subpel_y_q4,
+                             ConvolveParams *conv_params) {
+  const int bd = 8;
+  CONV_BUF_TYPE *dst = conv_params->dst;
+  const int dst_stride = conv_params->dst_stride;
+  const int fo_horiz = filter_params_x->taps / 2 - 1;
+  const uint8_t *src_ptr = src - fo_horiz;
+  const int bits = FILTER_BITS - conv_params->round_1;
+  const __m128i left_shift = _mm_cvtsi32_si128(bits);
+  const __m128i round_const = _mm_set1_epi32((1 << conv_params->round_0) >> 1);
+  const __m128i round_shift = _mm_cvtsi32_si128(conv_params->round_0);
+  const int w0 = conv_params->fwd_offset;
+  const int w1 = conv_params->bck_offset;
+  const __m128i wt0 = _mm_set1_epi16(w0);
+  const __m128i wt1 = _mm_set1_epi16(w1);
+  const __m128i wt = _mm_unpacklo_epi16(wt0, wt1);
+  const int do_average = conv_params->do_average;
+  const int use_jnt_comp_avg = conv_params->use_jnt_comp_avg;
+  const int offset_0 =
+      bd + 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+  const int offset = (1 << offset_0) + (1 << (offset_0 - 1));
+  const __m128i offset_const = _mm_set1_epi16(offset);
+  const int rounding_shift =
+      2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+  const __m128i rounding_const = _mm_set1_epi16((1 << rounding_shift) >> 1);
+  __m128i coeffs[4];
+
+  (void)filter_params_y;
+  (void)subpel_y_q4;
+
+  prepare_coeffs(filter_params_x, subpel_x_q4, coeffs);
+
+  if (w == 4) {
+    do {
+      const __m128i data = _mm_loadu_si128((__m128i *)src_ptr);
+      __m128i s[4];
+
+      s[0] = _mm_unpacklo_epi8(data, _mm_srli_si128(data, 1));
+      s[1] =
+          _mm_unpacklo_epi8(_mm_srli_si128(data, 2), _mm_srli_si128(data, 3));
+      s[2] =
+          _mm_unpacklo_epi8(_mm_srli_si128(data, 4), _mm_srli_si128(data, 5));
+      s[3] =
+          _mm_unpacklo_epi8(_mm_srli_si128(data, 6), _mm_srli_si128(data, 7));
+      const __m128i res_lo = convolve_lo_x(s, coeffs);
+      const __m128i res_lo_round =
+          _mm_sra_epi32(_mm_add_epi32(res_lo, round_const), round_shift);
+      const __m128i res_lo_shift = _mm_sll_epi32(res_lo_round, left_shift);
+
+      const __m128i res_16b = _mm_packs_epi32(res_lo_shift, res_lo_shift);
+      const __m128i res_unsigned = _mm_add_epi16(res_16b, offset_const);
+
+      // Accumulate values into the destination buffer
+      if (do_average) {
+        const __m128i data_ref_0 = _mm_loadu_si128((__m128i *)dst);
+
+        const __m128i comp_avg_res =
+            comp_avg(&data_ref_0, &res_unsigned, &wt, use_jnt_comp_avg);
+
+        const __m128i round_result = convolve_rounding(
+            &comp_avg_res, &offset_const, &rounding_const, rounding_shift);
+
+        const __m128i res_8 = _mm_packus_epi16(round_result, round_result);
+        *(uint32_t *)(&dst0[0]) = _mm_cvtsi128_si32(res_8);
+      } else {
+        _mm_store_si128((__m128i *)(&dst[0]), res_unsigned);
+      }
+      src_ptr += src_stride;
+      dst += dst_stride;
+      dst0 += dst_stride0;
+    } while (--h);
+  } else {
+    assert(!(w % 8));
+    int i = 0;
+    do {
+      int j = 0;
+      do {
+        const __m128i data =
+            _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + j]);
+        __m128i s[4];
+
+        // Filter even-index pixels
+        s[0] = data;
+        s[1] = _mm_srli_si128(data, 2);
+        s[2] = _mm_srli_si128(data, 4);
+        s[3] = _mm_srli_si128(data, 6);
+        const __m128i res_even = convolve_lo_x(s, coeffs);
+
+        // Filter odd-index pixels
+        s[0] = _mm_srli_si128(data, 1);
+        s[1] = _mm_srli_si128(data, 3);
+        s[2] = _mm_srli_si128(data, 5);
+        s[3] = _mm_srli_si128(data, 7);
+        const __m128i res_odd = convolve_lo_x(s, coeffs);
+
+        // Rearrange pixels back into the order 0 ... 7
+        const __m128i res_lo = _mm_unpacklo_epi32(res_even, res_odd);
+        const __m128i res_hi = _mm_unpackhi_epi32(res_even, res_odd);
+        const __m128i res_lo_round =
+            _mm_sra_epi32(_mm_add_epi32(res_lo, round_const), round_shift);
+        const __m128i res_hi_round =
+            _mm_sra_epi32(_mm_add_epi32(res_hi, round_const), round_shift);
+        const __m128i res_lo_shift = _mm_sll_epi32(res_lo_round, left_shift);
+        const __m128i res_hi_shift = _mm_sll_epi32(res_hi_round, left_shift);
+
+        const __m128i res_16b = _mm_packs_epi32(res_lo_shift, res_hi_shift);
+        const __m128i res_unsigned = _mm_add_epi16(res_16b, offset_const);
+
+        // Accumulate values into the destination buffer
+        if (do_average) {
+          const __m128i data_ref_0 =
+              _mm_loadu_si128((__m128i *)(&dst[i * dst_stride + j]));
+
+          const __m128i comp_avg_res =
+              comp_avg(&data_ref_0, &res_unsigned, &wt, use_jnt_comp_avg);
+
+          const __m128i round_result = convolve_rounding(
+              &comp_avg_res, &offset_const, &rounding_const, rounding_shift);
+
+          const __m128i res_8 = _mm_packus_epi16(round_result, round_result);
+          _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_8);
+        } else {
+          _mm_store_si128((__m128i *)(&dst[i * dst_stride + j]), res_unsigned);
+        }
+        j += 8;
+      } while (j < w);
+    } while (++i < h);
+  }
+}
+
+void av1_jnt_convolve_y_sse2(const uint8_t *src, int src_stride, uint8_t *dst0,
+                             int dst_stride0, int w, int h,
+                             const InterpFilterParams *filter_params_x,
+                             const InterpFilterParams *filter_params_y,
+                             const int subpel_x_q4, const int subpel_y_q4,
+                             ConvolveParams *conv_params) {
+  const int bd = 8;
+  CONV_BUF_TYPE *dst = conv_params->dst;
+  const int dst_stride = conv_params->dst_stride;
+  const int fo_vert = filter_params_y->taps / 2 - 1;
+  const uint8_t *src_ptr = src - fo_vert * src_stride;
+  const int bits = FILTER_BITS - conv_params->round_0;
+  const __m128i left_shift = _mm_cvtsi32_si128(bits);
+  const __m128i wt0 = _mm_set1_epi16(conv_params->fwd_offset);
+  const __m128i wt1 = _mm_set1_epi16(conv_params->bck_offset);
+  const __m128i wt = _mm_unpacklo_epi16(wt0, wt1);
+  const int do_average = conv_params->do_average;
+  const int use_jnt_comp_avg = conv_params->use_jnt_comp_avg;
+  const int offset_0 =
+      bd + 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+  const int offset = (1 << offset_0) + (1 << (offset_0 - 1));
+  const __m128i offset_const = _mm_set1_epi16(offset);
+  const int rounding_shift =
+      2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+  const __m128i rounding_const = _mm_set1_epi16((1 << rounding_shift) >> 1);
+  const __m128i round_const = _mm_set1_epi32((1 << conv_params->round_1) >> 1);
+  const __m128i round_shift = _mm_cvtsi32_si128(conv_params->round_1);
+  __m128i coeffs[4];
+
+  (void)filter_params_x;
+  (void)subpel_x_q4;
+
+  prepare_coeffs(filter_params_y, subpel_y_q4, coeffs);
+
+  if (w == 4) {
+    __m128i s[8], src6, res, res_shift;
+    src6 = _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 6 * src_stride));
+    s[0] = _mm_unpacklo_epi8(
+        _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 0 * src_stride)),
+        _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 1 * src_stride)));
+    s[1] = _mm_unpacklo_epi8(
+        _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 1 * src_stride)),
+        _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 2 * src_stride)));
+    s[2] = _mm_unpacklo_epi8(
+        _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 2 * src_stride)),
+        _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 3 * src_stride)));
+    s[3] = _mm_unpacklo_epi8(
+        _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 3 * src_stride)),
+        _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 4 * src_stride)));
+    s[4] = _mm_unpacklo_epi8(
+        _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 4 * src_stride)),
+        _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 5 * src_stride)));
+    s[5] = _mm_unpacklo_epi8(
+        _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 5 * src_stride)), src6);
+
+    do {
+      s[6] = _mm_unpacklo_epi8(
+          src6, _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 7 * src_stride)));
+      src6 = _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 8 * src_stride));
+      s[7] = _mm_unpacklo_epi8(
+          _mm_cvtsi32_si128(*(uint32_t *)(src_ptr + 7 * src_stride)), src6);
+
+      res = convolve_lo_y(s + 0, coeffs);
+      res_shift = _mm_sll_epi32(res, left_shift);
+      res_shift =
+          _mm_sra_epi32(_mm_add_epi32(res_shift, round_const), round_shift);
+
+      __m128i res_16b = _mm_packs_epi32(res_shift, res_shift);
+      __m128i res_unsigned = _mm_add_epi16(res_16b, offset_const);
+
+      // Accumulate values into the destination buffer
+      if (do_average) {
+        const __m128i data_ref_0 = _mm_loadu_si128((__m128i *)dst);
+
+        const __m128i comp_avg_res =
+            comp_avg(&data_ref_0, &res_unsigned, &wt, use_jnt_comp_avg);
+
+        const __m128i round_result = convolve_rounding(
+            &comp_avg_res, &offset_const, &rounding_const, rounding_shift);
+
+        const __m128i res_8 = _mm_packus_epi16(round_result, round_result);
+        *(uint32_t *)(&dst0[0]) = _mm_cvtsi128_si32(res_8);
+
+      } else {
+        _mm_store_si128((__m128i *)dst, res_unsigned);
+      }
+
+      src_ptr += src_stride;
+      dst += dst_stride;
+      dst0 += dst_stride0;
+
+      res = convolve_lo_y(s + 1, coeffs);
+      res_shift = _mm_sll_epi32(res, left_shift);
+      res_shift =
+          _mm_sra_epi32(_mm_add_epi32(res_shift, round_const), round_shift);
+
+      res_16b = _mm_packs_epi32(res_shift, res_shift);
+      res_unsigned = _mm_add_epi16(res_16b, offset_const);
+
+      // Accumulate values into the destination buffer
+      if (do_average) {
+        const __m128i data_ref_0 = _mm_loadu_si128((__m128i *)dst);
+
+        const __m128i comp_avg_res =
+            comp_avg(&data_ref_0, &res_unsigned, &wt, use_jnt_comp_avg);
+
+        const __m128i round_result = convolve_rounding(
+            &comp_avg_res, &offset_const, &rounding_const, rounding_shift);
+
+        const __m128i res_8 = _mm_packus_epi16(round_result, round_result);
+        *(uint32_t *)(&dst0[0]) = _mm_cvtsi128_si32(res_8);
+
+      } else {
+        _mm_store_si128((__m128i *)dst, res_unsigned);
+      }
+
+      src_ptr += src_stride;
+      dst += dst_stride;
+      dst0 += dst_stride0;
+
+      s[0] = s[2];
+      s[1] = s[3];
+      s[2] = s[4];
+      s[3] = s[5];
+      s[4] = s[6];
+      s[5] = s[7];
+      h -= 2;
+    } while (h);
+  } else {
+    assert(!(w % 8));
+    int j = 0;
+    do {
+      __m128i s[8], src6, res_lo, res_hi, res_lo_shift, res_hi_shift;
+      const uint8_t *data = &src_ptr[j];
+
+      src6 = _mm_loadl_epi64((__m128i *)(data + 6 * src_stride));
+      s[0] = _mm_unpacklo_epi8(
+          _mm_loadl_epi64((__m128i *)(data + 0 * src_stride)),
+          _mm_loadl_epi64((__m128i *)(data + 1 * src_stride)));
+      s[1] = _mm_unpacklo_epi8(
+          _mm_loadl_epi64((__m128i *)(data + 1 * src_stride)),
+          _mm_loadl_epi64((__m128i *)(data + 2 * src_stride)));
+      s[2] = _mm_unpacklo_epi8(
+          _mm_loadl_epi64((__m128i *)(data + 2 * src_stride)),
+          _mm_loadl_epi64((__m128i *)(data + 3 * src_stride)));
+      s[3] = _mm_unpacklo_epi8(
+          _mm_loadl_epi64((__m128i *)(data + 3 * src_stride)),
+          _mm_loadl_epi64((__m128i *)(data + 4 * src_stride)));
+      s[4] = _mm_unpacklo_epi8(
+          _mm_loadl_epi64((__m128i *)(data + 4 * src_stride)),
+          _mm_loadl_epi64((__m128i *)(data + 5 * src_stride)));
+      s[5] = _mm_unpacklo_epi8(
+          _mm_loadl_epi64((__m128i *)(data + 5 * src_stride)), src6);
+
+      int i = 0;
+      do {
+        data = &src_ptr[i * src_stride + j];
+        s[6] = _mm_unpacklo_epi8(
+            src6, _mm_loadl_epi64((__m128i *)(data + 7 * src_stride)));
+        src6 = _mm_loadl_epi64((__m128i *)(data + 8 * src_stride));
+        s[7] = _mm_unpacklo_epi8(
+            _mm_loadl_epi64((__m128i *)(data + 7 * src_stride)), src6);
+
+        res_lo = convolve_lo_y(s, coeffs);  // Filter low index pixels
+        res_hi = convolve_hi_y(s, coeffs);  // Filter high index pixels
+        res_lo_shift = _mm_sll_epi32(res_lo, left_shift);
+        res_hi_shift = _mm_sll_epi32(res_hi, left_shift);
+        res_lo_shift = _mm_sra_epi32(_mm_add_epi32(res_lo_shift, round_const),
+                                     round_shift);
+        res_hi_shift = _mm_sra_epi32(_mm_add_epi32(res_hi_shift, round_const),
+                                     round_shift);
+
+        __m128i res_16b = _mm_packs_epi32(res_lo_shift, res_hi_shift);
+        __m128i res_unsigned = _mm_add_epi16(res_16b, offset_const);
+
+        // Accumulate values into the destination buffer
+        if (do_average) {
+          const __m128i data_ref_0 =
+              _mm_loadu_si128((__m128i *)(&dst[i * dst_stride + j]));
+
+          const __m128i comp_avg_res =
+              comp_avg(&data_ref_0, &res_unsigned, &wt, use_jnt_comp_avg);
+
+          const __m128i round_result = convolve_rounding(
+              &comp_avg_res, &offset_const, &rounding_const, rounding_shift);
+
+          const __m128i res_8 = _mm_packus_epi16(round_result, round_result);
+          _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_8);
+        } else {
+          _mm_store_si128((__m128i *)(&dst[i * dst_stride + j]), res_unsigned);
+        }
+        i++;
+
+        res_lo = convolve_lo_y(s + 1, coeffs);  // Filter low index pixels
+        res_hi = convolve_hi_y(s + 1, coeffs);  // Filter high index pixels
+        res_lo_shift = _mm_sll_epi32(res_lo, left_shift);
+        res_hi_shift = _mm_sll_epi32(res_hi, left_shift);
+        res_lo_shift = _mm_sra_epi32(_mm_add_epi32(res_lo_shift, round_const),
+                                     round_shift);
+        res_hi_shift = _mm_sra_epi32(_mm_add_epi32(res_hi_shift, round_const),
+                                     round_shift);
+        res_16b = _mm_packs_epi32(res_lo_shift, res_hi_shift);
+        res_unsigned = _mm_add_epi16(res_16b, offset_const);
+
+        // Accumulate values into the destination buffer
+        if (do_average) {
+          __m128i data_ref_0 =
+              _mm_loadu_si128((__m128i *)(&dst[i * dst_stride + j]));
+
+          const __m128i comp_avg_res =
+              comp_avg(&data_ref_0, &res_unsigned, &wt, use_jnt_comp_avg);
+
+          const __m128i round_result = convolve_rounding(
+              &comp_avg_res, &offset_const, &rounding_const, rounding_shift);
+
+          const __m128i res_8 = _mm_packus_epi16(round_result, round_result);
+          _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_8);
+        } else {
+          _mm_store_si128((__m128i *)(&dst[i * dst_stride + j]), res_unsigned);
+        }
+        i++;
+
+        s[0] = s[2];
+        s[1] = s[3];
+        s[2] = s[4];
+        s[3] = s[5];
+        s[4] = s[6];
+        s[5] = s[7];
+      } while (i < h);
+      j += 8;
+    } while (j < w);
+  }
+}
diff --git a/third_party/aom/av1/common/x86/jnt_convolve_ssse3.c b/third_party/aom/av1/common/x86/jnt_convolve_ssse3.c
new file mode 100644
index 000000000..822772782
--- /dev/null
+++ b/third_party/aom/av1/common/x86/jnt_convolve_ssse3.c
@@ -0,0 +1,232 @@
+/*
+ * 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_dsp_rtcd.h"
+
+#include "aom_dsp/aom_filter.h"
+#include "aom_dsp/x86/convolve_sse2.h"
+
+void av1_jnt_convolve_2d_ssse3(const uint8_t *src, int src_stride,
+                               uint8_t *dst0, int dst_stride0, int w, int h,
+                               const InterpFilterParams *filter_params_x,
+                               const InterpFilterParams *filter_params_y,
+                               const int subpel_x_q4, const int subpel_y_q4,
+                               ConvolveParams *conv_params) {
+  CONV_BUF_TYPE *dst = conv_params->dst;
+  int dst_stride = conv_params->dst_stride;
+  const int bd = 8;
+
+  DECLARE_ALIGNED(16, int16_t,
+                  im_block[(MAX_SB_SIZE + MAX_FILTER_TAP - 1) * MAX_SB_SIZE]);
+  int im_h = h + filter_params_y->taps - 1;
+  int im_stride = MAX_SB_SIZE;
+  int i, j;
+  const int fo_vert = filter_params_y->taps / 2 - 1;
+  const int fo_horiz = filter_params_x->taps / 2 - 1;
+  const int do_average = conv_params->do_average;
+  const int use_jnt_comp_avg = conv_params->use_jnt_comp_avg;
+  const uint8_t *const src_ptr = src - fo_vert * src_stride - fo_horiz;
+
+  const __m128i zero = _mm_setzero_si128();
+
+  const int w0 = conv_params->fwd_offset;
+  const int w1 = conv_params->bck_offset;
+  const __m128i wt0 = _mm_set1_epi16(w0);
+  const __m128i wt1 = _mm_set1_epi16(w1);
+  const __m128i wt = _mm_unpacklo_epi16(wt0, wt1);
+
+  const int offset_0 =
+      bd + 2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+  const int offset = (1 << offset_0) + (1 << (offset_0 - 1));
+  const __m128i offset_const = _mm_set1_epi16(offset);
+  const int rounding_shift =
+      2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+  const __m128i rounding_const = _mm_set1_epi16((1 << rounding_shift) >> 1);
+
+  /* Horizontal filter */
+  {
+    const int16_t *x_filter = av1_get_interp_filter_subpel_kernel(
+        filter_params_x, subpel_x_q4 & SUBPEL_MASK);
+    const __m128i coeffs_x = _mm_loadu_si128((__m128i *)x_filter);
+
+    // coeffs 0 1 0 1 2 3 2 3
+    const __m128i tmp_0 = _mm_unpacklo_epi32(coeffs_x, coeffs_x);
+    // coeffs 4 5 4 5 6 7 6 7
+    const __m128i tmp_1 = _mm_unpackhi_epi32(coeffs_x, coeffs_x);
+
+    // coeffs 0 1 0 1 0 1 0 1
+    const __m128i coeff_01 = _mm_unpacklo_epi64(tmp_0, tmp_0);
+    // coeffs 2 3 2 3 2 3 2 3
+    const __m128i coeff_23 = _mm_unpackhi_epi64(tmp_0, tmp_0);
+    // coeffs 4 5 4 5 4 5 4 5
+    const __m128i coeff_45 = _mm_unpacklo_epi64(tmp_1, tmp_1);
+    // coeffs 6 7 6 7 6 7 6 7
+    const __m128i coeff_67 = _mm_unpackhi_epi64(tmp_1, tmp_1);
+
+    const __m128i round_const = _mm_set1_epi32(
+        ((1 << conv_params->round_0) >> 1) + (1 << (bd + FILTER_BITS - 1)));
+    const __m128i round_shift = _mm_cvtsi32_si128(conv_params->round_0);
+
+    for (i = 0; i < im_h; ++i) {
+      for (j = 0; j < w; j += 8) {
+        const __m128i data =
+            _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + j]);
+
+        const __m128i src_lo = _mm_unpacklo_epi8(data, zero);
+        const __m128i src_hi = _mm_unpackhi_epi8(data, zero);
+
+        // Filter even-index pixels
+        const __m128i res_0 = _mm_madd_epi16(src_lo, coeff_01);
+        const __m128i src_2 = _mm_alignr_epi8(src_hi, src_lo, 4);
+        const __m128i res_2 = _mm_madd_epi16(src_2, coeff_23);
+        const __m128i src_4 = _mm_alignr_epi8(src_hi, src_lo, 8);
+        const __m128i res_4 = _mm_madd_epi16(src_4, coeff_45);
+        const __m128i src_6 = _mm_alignr_epi8(src_hi, src_lo, 12);
+        const __m128i res_6 = _mm_madd_epi16(src_6, coeff_67);
+
+        __m128i res_even = _mm_add_epi32(_mm_add_epi32(res_0, res_4),
+                                         _mm_add_epi32(res_2, res_6));
+        res_even =
+            _mm_sra_epi32(_mm_add_epi32(res_even, round_const), round_shift);
+
+        // Filter odd-index pixels
+        const __m128i src_1 = _mm_alignr_epi8(src_hi, src_lo, 2);
+        const __m128i res_1 = _mm_madd_epi16(src_1, coeff_01);
+        const __m128i src_3 = _mm_alignr_epi8(src_hi, src_lo, 6);
+        const __m128i res_3 = _mm_madd_epi16(src_3, coeff_23);
+        const __m128i src_5 = _mm_alignr_epi8(src_hi, src_lo, 10);
+        const __m128i res_5 = _mm_madd_epi16(src_5, coeff_45);
+        const __m128i src_7 = _mm_alignr_epi8(src_hi, src_lo, 14);
+        const __m128i res_7 = _mm_madd_epi16(src_7, coeff_67);
+
+        __m128i res_odd = _mm_add_epi32(_mm_add_epi32(res_1, res_5),
+                                        _mm_add_epi32(res_3, res_7));
+        res_odd =
+            _mm_sra_epi32(_mm_add_epi32(res_odd, round_const), round_shift);
+
+        // Pack in the column order 0, 2, 4, 6, 1, 3, 5, 7
+        __m128i res = _mm_packs_epi32(res_even, res_odd);
+        _mm_store_si128((__m128i *)&im_block[i * im_stride + j], res);
+      }
+    }
+  }
+
+  /* Vertical filter */
+  {
+    const int16_t *y_filter = av1_get_interp_filter_subpel_kernel(
+        filter_params_y, subpel_y_q4 & SUBPEL_MASK);
+    const __m128i coeffs_y = _mm_loadu_si128((__m128i *)y_filter);
+
+    // coeffs 0 1 0 1 2 3 2 3
+    const __m128i tmp_0 = _mm_unpacklo_epi32(coeffs_y, coeffs_y);
+    // coeffs 4 5 4 5 6 7 6 7
+    const __m128i tmp_1 = _mm_unpackhi_epi32(coeffs_y, coeffs_y);
+
+    // coeffs 0 1 0 1 0 1 0 1
+    const __m128i coeff_01 = _mm_unpacklo_epi64(tmp_0, tmp_0);
+    // coeffs 2 3 2 3 2 3 2 3
+    const __m128i coeff_23 = _mm_unpackhi_epi64(tmp_0, tmp_0);
+    // coeffs 4 5 4 5 4 5 4 5
+    const __m128i coeff_45 = _mm_unpacklo_epi64(tmp_1, tmp_1);
+    // coeffs 6 7 6 7 6 7 6 7
+    const __m128i coeff_67 = _mm_unpackhi_epi64(tmp_1, tmp_1);
+
+    const __m128i round_const = _mm_set1_epi32(
+        ((1 << conv_params->round_1) >> 1) -
+        (1 << (bd + 2 * FILTER_BITS - conv_params->round_0 - 1)));
+    const __m128i round_shift = _mm_cvtsi32_si128(conv_params->round_1);
+
+    for (i = 0; i < h; ++i) {
+      for (j = 0; j < w; j += 8) {
+        // Filter even-index pixels
+        const int16_t *data = &im_block[i * im_stride + j];
+        const __m128i src_0 =
+            _mm_unpacklo_epi16(*(__m128i *)(data + 0 * im_stride),
+                               *(__m128i *)(data + 1 * im_stride));
+        const __m128i src_2 =
+            _mm_unpacklo_epi16(*(__m128i *)(data + 2 * im_stride),
+                               *(__m128i *)(data + 3 * im_stride));
+        const __m128i src_4 =
+            _mm_unpacklo_epi16(*(__m128i *)(data + 4 * im_stride),
+                               *(__m128i *)(data + 5 * im_stride));
+        const __m128i src_6 =
+            _mm_unpacklo_epi16(*(__m128i *)(data + 6 * im_stride),
+                               *(__m128i *)(data + 7 * im_stride));
+
+        const __m128i res_0 = _mm_madd_epi16(src_0, coeff_01);
+        const __m128i res_2 = _mm_madd_epi16(src_2, coeff_23);
+        const __m128i res_4 = _mm_madd_epi16(src_4, coeff_45);
+        const __m128i res_6 = _mm_madd_epi16(src_6, coeff_67);
+
+        const __m128i res_even = _mm_add_epi32(_mm_add_epi32(res_0, res_2),
+                                               _mm_add_epi32(res_4, res_6));
+
+        // Filter odd-index pixels
+        const __m128i src_1 =
+            _mm_unpackhi_epi16(*(__m128i *)(data + 0 * im_stride),
+                               *(__m128i *)(data + 1 * im_stride));
+        const __m128i src_3 =
+            _mm_unpackhi_epi16(*(__m128i *)(data + 2 * im_stride),
+                               *(__m128i *)(data + 3 * im_stride));
+        const __m128i src_5 =
+            _mm_unpackhi_epi16(*(__m128i *)(data + 4 * im_stride),
+                               *(__m128i *)(data + 5 * im_stride));
+        const __m128i src_7 =
+            _mm_unpackhi_epi16(*(__m128i *)(data + 6 * im_stride),
+                               *(__m128i *)(data + 7 * im_stride));
+
+        const __m128i res_1 = _mm_madd_epi16(src_1, coeff_01);
+        const __m128i res_3 = _mm_madd_epi16(src_3, coeff_23);
+        const __m128i res_5 = _mm_madd_epi16(src_5, coeff_45);
+        const __m128i res_7 = _mm_madd_epi16(src_7, coeff_67);
+
+        const __m128i res_odd = _mm_add_epi32(_mm_add_epi32(res_1, res_3),
+                                              _mm_add_epi32(res_5, res_7));
+
+        // Rearrange pixels back into the order 0 ... 7
+        const __m128i res_lo = _mm_unpacklo_epi32(res_even, res_odd);
+        const __m128i res_hi = _mm_unpackhi_epi32(res_even, res_odd);
+
+        const __m128i res_lo_round =
+            _mm_sra_epi32(_mm_add_epi32(res_lo, round_const), round_shift);
+        const __m128i res_hi_round =
+            _mm_sra_epi32(_mm_add_epi32(res_hi, round_const), round_shift);
+
+        const __m128i res_16b = _mm_packs_epi32(res_lo_round, res_hi_round);
+        const __m128i res_unsigned = _mm_add_epi16(res_16b, offset_const);
+
+        // Accumulate values into the destination buffer
+        if (do_average) {
+          const __m128i data_ref_0 =
+              _mm_loadu_si128((__m128i *)(&dst[i * dst_stride + j]));
+
+          const __m128i comp_avg_res =
+              comp_avg(&data_ref_0, &res_unsigned, &wt, use_jnt_comp_avg);
+
+          const __m128i round_result = convolve_rounding(
+              &comp_avg_res, &offset_const, &rounding_const, rounding_shift);
+
+          const __m128i res_8 = _mm_packus_epi16(round_result, round_result);
+
+          if (w > 4)
+            _mm_storel_epi64((__m128i *)(&dst0[i * dst_stride0 + j]), res_8);
+          else
+            *(uint32_t *)(&dst0[i * dst_stride0 + j]) =
+                _mm_cvtsi128_si32(res_8);
+        } else {
+          _mm_store_si128((__m128i *)(&dst[i * dst_stride + j]), res_unsigned);
+        }
+      }
+    }
+  }
+}
diff --git a/third_party/aom/av1/common/x86/reconinter_avx2.c b/third_party/aom/av1/common/x86/reconinter_avx2.c
new file mode 100644
index 000000000..f645e0454
--- /dev/null
+++ b/third_party/aom/av1/common/x86/reconinter_avx2.c
@@ -0,0 +1,620 @@
+/*
+ * 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/av1_rtcd.h"
+
+#include "aom/aom_integer.h"
+#include "aom_dsp/blend.h"
+#include "aom_dsp/x86/synonyms.h"
+#include "aom_dsp/x86/synonyms_avx2.h"
+#include "av1/common/blockd.h"
+
+static INLINE __m256i calc_mask_avx2(const __m256i mask_base, const __m256i s0,
+                                     const __m256i s1) {
+  const __m256i diff = _mm256_abs_epi16(_mm256_sub_epi16(s0, s1));
+  return _mm256_abs_epi16(
+      _mm256_add_epi16(mask_base, _mm256_srli_epi16(diff, 4)));
+  // clamp(diff, 0, 64) can be skiped for diff is always in the range ( 38, 54)
+}
+void av1_build_compound_diffwtd_mask_avx2(uint8_t *mask,
+                                          DIFFWTD_MASK_TYPE mask_type,
+                                          const uint8_t *src0, int stride0,
+                                          const uint8_t *src1, int stride1,
+                                          int h, int w) {
+  const int mb = (mask_type == DIFFWTD_38_INV) ? AOM_BLEND_A64_MAX_ALPHA : 0;
+  const __m256i y_mask_base = _mm256_set1_epi16(38 - mb);
+  int i = 0;
+  if (4 == w) {
+    do {
+      const __m128i s0A = xx_loadl_32(src0);
+      const __m128i s0B = xx_loadl_32(src0 + stride0);
+      const __m128i s0C = xx_loadl_32(src0 + stride0 * 2);
+      const __m128i s0D = xx_loadl_32(src0 + stride0 * 3);
+      const __m128i s0AB = _mm_unpacklo_epi32(s0A, s0B);
+      const __m128i s0CD = _mm_unpacklo_epi32(s0C, s0D);
+      const __m128i s0ABCD = _mm_unpacklo_epi64(s0AB, s0CD);
+      const __m256i s0ABCD_w = _mm256_cvtepu8_epi16(s0ABCD);
+
+      const __m128i s1A = xx_loadl_32(src1);
+      const __m128i s1B = xx_loadl_32(src1 + stride1);
+      const __m128i s1C = xx_loadl_32(src1 + stride1 * 2);
+      const __m128i s1D = xx_loadl_32(src1 + stride1 * 3);
+      const __m128i s1AB = _mm_unpacklo_epi32(s1A, s1B);
+      const __m128i s1CD = _mm_unpacklo_epi32(s1C, s1D);
+      const __m128i s1ABCD = _mm_unpacklo_epi64(s1AB, s1CD);
+      const __m256i s1ABCD_w = _mm256_cvtepu8_epi16(s1ABCD);
+      const __m256i m16 = calc_mask_avx2(y_mask_base, s0ABCD_w, s1ABCD_w);
+      const __m256i m8 = _mm256_packus_epi16(m16, _mm256_setzero_si256());
+      const __m128i x_m8 =
+          _mm256_castsi256_si128(_mm256_permute4x64_epi64(m8, 0xd8));
+      xx_storeu_128(mask, x_m8);
+      src0 += (stride0 << 2);
+      src1 += (stride1 << 2);
+      mask += 16;
+      i += 4;
+    } while (i < h);
+  } else if (8 == w) {
+    do {
+      const __m128i s0A = xx_loadl_64(src0);
+      const __m128i s0B = xx_loadl_64(src0 + stride0);
+      const __m128i s0C = xx_loadl_64(src0 + stride0 * 2);
+      const __m128i s0D = xx_loadl_64(src0 + stride0 * 3);
+      const __m256i s0AC_w = _mm256_cvtepu8_epi16(_mm_unpacklo_epi64(s0A, s0C));
+      const __m256i s0BD_w = _mm256_cvtepu8_epi16(_mm_unpacklo_epi64(s0B, s0D));
+      const __m128i s1A = xx_loadl_64(src1);
+      const __m128i s1B = xx_loadl_64(src1 + stride1);
+      const __m128i s1C = xx_loadl_64(src1 + stride1 * 2);
+      const __m128i s1D = xx_loadl_64(src1 + stride1 * 3);
+      const __m256i s1AB_w = _mm256_cvtepu8_epi16(_mm_unpacklo_epi64(s1A, s1C));
+      const __m256i s1CD_w = _mm256_cvtepu8_epi16(_mm_unpacklo_epi64(s1B, s1D));
+      const __m256i m16AC = calc_mask_avx2(y_mask_base, s0AC_w, s1AB_w);
+      const __m256i m16BD = calc_mask_avx2(y_mask_base, s0BD_w, s1CD_w);
+      const __m256i m8 = _mm256_packus_epi16(m16AC, m16BD);
+      yy_storeu_256(mask, m8);
+      src0 += stride0 << 2;
+      src1 += stride1 << 2;
+      mask += 32;
+      i += 4;
+    } while (i < h);
+  } else if (16 == w) {
+    do {
+      const __m128i s0A = xx_load_128(src0);
+      const __m128i s0B = xx_load_128(src0 + stride0);
+      const __m128i s1A = xx_load_128(src1);
+      const __m128i s1B = xx_load_128(src1 + stride1);
+      const __m256i s0AL = _mm256_cvtepu8_epi16(s0A);
+      const __m256i s0BL = _mm256_cvtepu8_epi16(s0B);
+      const __m256i s1AL = _mm256_cvtepu8_epi16(s1A);
+      const __m256i s1BL = _mm256_cvtepu8_epi16(s1B);
+
+      const __m256i m16AL = calc_mask_avx2(y_mask_base, s0AL, s1AL);
+      const __m256i m16BL = calc_mask_avx2(y_mask_base, s0BL, s1BL);
+
+      const __m256i m8 =
+          _mm256_permute4x64_epi64(_mm256_packus_epi16(m16AL, m16BL), 0xd8);
+      yy_storeu_256(mask, m8);
+      src0 += stride0 << 1;
+      src1 += stride1 << 1;
+      mask += 32;
+      i += 2;
+    } while (i < h);
+  } else {
+    do {
+      int j = 0;
+      do {
+        const __m256i s0 = yy_loadu_256(src0 + j);
+        const __m256i s1 = yy_loadu_256(src1 + j);
+        const __m256i s0L = _mm256_cvtepu8_epi16(_mm256_castsi256_si128(s0));
+        const __m256i s1L = _mm256_cvtepu8_epi16(_mm256_castsi256_si128(s1));
+        const __m256i s0H =
+            _mm256_cvtepu8_epi16(_mm256_extracti128_si256(s0, 1));
+        const __m256i s1H =
+            _mm256_cvtepu8_epi16(_mm256_extracti128_si256(s1, 1));
+        const __m256i m16L = calc_mask_avx2(y_mask_base, s0L, s1L);
+        const __m256i m16H = calc_mask_avx2(y_mask_base, s0H, s1H);
+        const __m256i m8 =
+            _mm256_permute4x64_epi64(_mm256_packus_epi16(m16L, m16H), 0xd8);
+        yy_storeu_256(mask + j, m8);
+        j += 32;
+      } while (j < w);
+      src0 += stride0;
+      src1 += stride1;
+      mask += w;
+      i += 1;
+    } while (i < h);
+  }
+}
+
+static INLINE __m256i calc_mask_d16_avx2(const __m256i *data_src0,
+                                         const __m256i *data_src1,
+                                         const __m256i *round_const,
+                                         const __m256i *mask_base_16,
+                                         const __m256i *clip_diff, int round) {
+  const __m256i diffa = _mm256_subs_epu16(*data_src0, *data_src1);
+  const __m256i diffb = _mm256_subs_epu16(*data_src1, *data_src0);
+  const __m256i diff = _mm256_max_epu16(diffa, diffb);
+  const __m256i diff_round =
+      _mm256_srli_epi16(_mm256_adds_epu16(diff, *round_const), round);
+  const __m256i diff_factor = _mm256_srli_epi16(diff_round, DIFF_FACTOR_LOG2);
+  const __m256i diff_mask = _mm256_adds_epi16(diff_factor, *mask_base_16);
+  const __m256i diff_clamp = _mm256_min_epi16(diff_mask, *clip_diff);
+  return diff_clamp;
+}
+
+static INLINE __m256i calc_mask_d16_inv_avx2(const __m256i *data_src0,
+                                             const __m256i *data_src1,
+                                             const __m256i *round_const,
+                                             const __m256i *mask_base_16,
+                                             const __m256i *clip_diff,
+                                             int round) {
+  const __m256i diffa = _mm256_subs_epu16(*data_src0, *data_src1);
+  const __m256i diffb = _mm256_subs_epu16(*data_src1, *data_src0);
+  const __m256i diff = _mm256_max_epu16(diffa, diffb);
+  const __m256i diff_round =
+      _mm256_srli_epi16(_mm256_adds_epu16(diff, *round_const), round);
+  const __m256i diff_factor = _mm256_srli_epi16(diff_round, DIFF_FACTOR_LOG2);
+  const __m256i diff_mask = _mm256_adds_epi16(diff_factor, *mask_base_16);
+  const __m256i diff_clamp = _mm256_min_epi16(diff_mask, *clip_diff);
+  const __m256i diff_const_16 = _mm256_sub_epi16(*clip_diff, diff_clamp);
+  return diff_const_16;
+}
+
+static INLINE void build_compound_diffwtd_mask_d16_avx2(
+    uint8_t *mask, const CONV_BUF_TYPE *src0, int src0_stride,
+    const CONV_BUF_TYPE *src1, int src1_stride, int h, int w, int shift) {
+  const int mask_base = 38;
+  const __m256i _r = _mm256_set1_epi16((1 << shift) >> 1);
+  const __m256i y38 = _mm256_set1_epi16(mask_base);
+  const __m256i y64 = _mm256_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+  int i = 0;
+  if (w == 4) {
+    do {
+      const __m128i s0A = xx_loadl_64(src0);
+      const __m128i s0B = xx_loadl_64(src0 + src0_stride);
+      const __m128i s0C = xx_loadl_64(src0 + src0_stride * 2);
+      const __m128i s0D = xx_loadl_64(src0 + src0_stride * 3);
+      const __m128i s1A = xx_loadl_64(src1);
+      const __m128i s1B = xx_loadl_64(src1 + src1_stride);
+      const __m128i s1C = xx_loadl_64(src1 + src1_stride * 2);
+      const __m128i s1D = xx_loadl_64(src1 + src1_stride * 3);
+      const __m256i s0 = yy_set_m128i(_mm_unpacklo_epi64(s0C, s0D),
+                                      _mm_unpacklo_epi64(s0A, s0B));
+      const __m256i s1 = yy_set_m128i(_mm_unpacklo_epi64(s1C, s1D),
+                                      _mm_unpacklo_epi64(s1A, s1B));
+      const __m256i m16 = calc_mask_d16_avx2(&s0, &s1, &_r, &y38, &y64, shift);
+      const __m256i m8 = _mm256_packus_epi16(m16, _mm256_setzero_si256());
+      xx_storeu_128(mask,
+                    _mm256_castsi256_si128(_mm256_permute4x64_epi64(m8, 0xd8)));
+      src0 += src0_stride << 2;
+      src1 += src1_stride << 2;
+      mask += 16;
+      i += 4;
+    } while (i < h);
+  } else if (w == 8) {
+    do {
+      const __m256i s0AB = yy_loadu2_128(src0 + src0_stride, src0);
+      const __m256i s0CD =
+          yy_loadu2_128(src0 + src0_stride * 3, src0 + src0_stride * 2);
+      const __m256i s1AB = yy_loadu2_128(src1 + src1_stride, src1);
+      const __m256i s1CD =
+          yy_loadu2_128(src1 + src1_stride * 3, src1 + src1_stride * 2);
+      const __m256i m16AB =
+          calc_mask_d16_avx2(&s0AB, &s1AB, &_r, &y38, &y64, shift);
+      const __m256i m16CD =
+          calc_mask_d16_avx2(&s0CD, &s1CD, &_r, &y38, &y64, shift);
+      const __m256i m8 = _mm256_packus_epi16(m16AB, m16CD);
+      yy_storeu_256(mask, _mm256_permute4x64_epi64(m8, 0xd8));
+      src0 += src0_stride << 2;
+      src1 += src1_stride << 2;
+      mask += 32;
+      i += 4;
+    } while (i < h);
+  } else if (w == 16) {
+    do {
+      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);
+      const __m256i m16A =
+          calc_mask_d16_avx2(&s0A, &s1A, &_r, &y38, &y64, shift);
+      const __m256i m16B =
+          calc_mask_d16_avx2(&s0B, &s1B, &_r, &y38, &y64, shift);
+      const __m256i m8 = _mm256_packus_epi16(m16A, m16B);
+      yy_storeu_256(mask, _mm256_permute4x64_epi64(m8, 0xd8));
+      src0 += src0_stride << 1;
+      src1 += src1_stride << 1;
+      mask += 32;
+      i += 2;
+    } while (i < h);
+  } else if (w == 32) {
+    do {
+      const __m256i s0A = yy_loadu_256(src0);
+      const __m256i s0B = yy_loadu_256(src0 + 16);
+      const __m256i s1A = yy_loadu_256(src1);
+      const __m256i s1B = yy_loadu_256(src1 + 16);
+      const __m256i m16A =
+          calc_mask_d16_avx2(&s0A, &s1A, &_r, &y38, &y64, shift);
+      const __m256i m16B =
+          calc_mask_d16_avx2(&s0B, &s1B, &_r, &y38, &y64, shift);
+      const __m256i m8 = _mm256_packus_epi16(m16A, m16B);
+      yy_storeu_256(mask, _mm256_permute4x64_epi64(m8, 0xd8));
+      src0 += src0_stride;
+      src1 += src1_stride;
+      mask += 32;
+      i += 1;
+    } while (i < h);
+  } else if (w == 64) {
+    do {
+      const __m256i s0A = yy_loadu_256(src0);
+      const __m256i s0B = yy_loadu_256(src0 + 16);
+      const __m256i s0C = yy_loadu_256(src0 + 32);
+      const __m256i s0D = yy_loadu_256(src0 + 48);
+      const __m256i s1A = yy_loadu_256(src1);
+      const __m256i s1B = yy_loadu_256(src1 + 16);
+      const __m256i s1C = yy_loadu_256(src1 + 32);
+      const __m256i s1D = yy_loadu_256(src1 + 48);
+      const __m256i m16A =
+          calc_mask_d16_avx2(&s0A, &s1A, &_r, &y38, &y64, shift);
+      const __m256i m16B =
+          calc_mask_d16_avx2(&s0B, &s1B, &_r, &y38, &y64, shift);
+      const __m256i m16C =
+          calc_mask_d16_avx2(&s0C, &s1C, &_r, &y38, &y64, shift);
+      const __m256i m16D =
+          calc_mask_d16_avx2(&s0D, &s1D, &_r, &y38, &y64, shift);
+      const __m256i m8AB = _mm256_packus_epi16(m16A, m16B);
+      const __m256i m8CD = _mm256_packus_epi16(m16C, m16D);
+      yy_storeu_256(mask, _mm256_permute4x64_epi64(m8AB, 0xd8));
+      yy_storeu_256(mask + 32, _mm256_permute4x64_epi64(m8CD, 0xd8));
+      src0 += src0_stride;
+      src1 += src1_stride;
+      mask += 64;
+      i += 1;
+    } while (i < h);
+  } else {
+    do {
+      const __m256i s0A = yy_loadu_256(src0);
+      const __m256i s0B = yy_loadu_256(src0 + 16);
+      const __m256i s0C = yy_loadu_256(src0 + 32);
+      const __m256i s0D = yy_loadu_256(src0 + 48);
+      const __m256i s0E = yy_loadu_256(src0 + 64);
+      const __m256i s0F = yy_loadu_256(src0 + 80);
+      const __m256i s0G = yy_loadu_256(src0 + 96);
+      const __m256i s0H = yy_loadu_256(src0 + 112);
+      const __m256i s1A = yy_loadu_256(src1);
+      const __m256i s1B = yy_loadu_256(src1 + 16);
+      const __m256i s1C = yy_loadu_256(src1 + 32);
+      const __m256i s1D = yy_loadu_256(src1 + 48);
+      const __m256i s1E = yy_loadu_256(src1 + 64);
+      const __m256i s1F = yy_loadu_256(src1 + 80);
+      const __m256i s1G = yy_loadu_256(src1 + 96);
+      const __m256i s1H = yy_loadu_256(src1 + 112);
+      const __m256i m16A =
+          calc_mask_d16_avx2(&s0A, &s1A, &_r, &y38, &y64, shift);
+      const __m256i m16B =
+          calc_mask_d16_avx2(&s0B, &s1B, &_r, &y38, &y64, shift);
+      const __m256i m16C =
+          calc_mask_d16_avx2(&s0C, &s1C, &_r, &y38, &y64, shift);
+      const __m256i m16D =
+          calc_mask_d16_avx2(&s0D, &s1D, &_r, &y38, &y64, shift);
+      const __m256i m16E =
+          calc_mask_d16_avx2(&s0E, &s1E, &_r, &y38, &y64, shift);
+      const __m256i m16F =
+          calc_mask_d16_avx2(&s0F, &s1F, &_r, &y38, &y64, shift);
+      const __m256i m16G =
+          calc_mask_d16_avx2(&s0G, &s1G, &_r, &y38, &y64, shift);
+      const __m256i m16H =
+          calc_mask_d16_avx2(&s0H, &s1H, &_r, &y38, &y64, shift);
+      const __m256i m8AB = _mm256_packus_epi16(m16A, m16B);
+      const __m256i m8CD = _mm256_packus_epi16(m16C, m16D);
+      const __m256i m8EF = _mm256_packus_epi16(m16E, m16F);
+      const __m256i m8GH = _mm256_packus_epi16(m16G, m16H);
+      yy_storeu_256(mask, _mm256_permute4x64_epi64(m8AB, 0xd8));
+      yy_storeu_256(mask + 32, _mm256_permute4x64_epi64(m8CD, 0xd8));
+      yy_storeu_256(mask + 64, _mm256_permute4x64_epi64(m8EF, 0xd8));
+      yy_storeu_256(mask + 96, _mm256_permute4x64_epi64(m8GH, 0xd8));
+      src0 += src0_stride;
+      src1 += src1_stride;
+      mask += 128;
+      i += 1;
+    } while (i < h);
+  }
+}
+
+static INLINE void build_compound_diffwtd_mask_d16_inv_avx2(
+    uint8_t *mask, const CONV_BUF_TYPE *src0, int src0_stride,
+    const CONV_BUF_TYPE *src1, int src1_stride, int h, int w, int shift) {
+  const int mask_base = 38;
+  const __m256i _r = _mm256_set1_epi16((1 << shift) >> 1);
+  const __m256i y38 = _mm256_set1_epi16(mask_base);
+  const __m256i y64 = _mm256_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+  int i = 0;
+  if (w == 4) {
+    do {
+      const __m128i s0A = xx_loadl_64(src0);
+      const __m128i s0B = xx_loadl_64(src0 + src0_stride);
+      const __m128i s0C = xx_loadl_64(src0 + src0_stride * 2);
+      const __m128i s0D = xx_loadl_64(src0 + src0_stride * 3);
+      const __m128i s1A = xx_loadl_64(src1);
+      const __m128i s1B = xx_loadl_64(src1 + src1_stride);
+      const __m128i s1C = xx_loadl_64(src1 + src1_stride * 2);
+      const __m128i s1D = xx_loadl_64(src1 + src1_stride * 3);
+      const __m256i s0 = yy_set_m128i(_mm_unpacklo_epi64(s0C, s0D),
+                                      _mm_unpacklo_epi64(s0A, s0B));
+      const __m256i s1 = yy_set_m128i(_mm_unpacklo_epi64(s1C, s1D),
+                                      _mm_unpacklo_epi64(s1A, s1B));
+      const __m256i m16 =
+          calc_mask_d16_inv_avx2(&s0, &s1, &_r, &y38, &y64, shift);
+      const __m256i m8 = _mm256_packus_epi16(m16, _mm256_setzero_si256());
+      xx_storeu_128(mask,
+                    _mm256_castsi256_si128(_mm256_permute4x64_epi64(m8, 0xd8)));
+      src0 += src0_stride << 2;
+      src1 += src1_stride << 2;
+      mask += 16;
+      i += 4;
+    } while (i < h);
+  } else if (w == 8) {
+    do {
+      const __m256i s0AB = yy_loadu2_128(src0 + src0_stride, src0);
+      const __m256i s0CD =
+          yy_loadu2_128(src0 + src0_stride * 3, src0 + src0_stride * 2);
+      const __m256i s1AB = yy_loadu2_128(src1 + src1_stride, src1);
+      const __m256i s1CD =
+          yy_loadu2_128(src1 + src1_stride * 3, src1 + src1_stride * 2);
+      const __m256i m16AB =
+          calc_mask_d16_inv_avx2(&s0AB, &s1AB, &_r, &y38, &y64, shift);
+      const __m256i m16CD =
+          calc_mask_d16_inv_avx2(&s0CD, &s1CD, &_r, &y38, &y64, shift);
+      const __m256i m8 = _mm256_packus_epi16(m16AB, m16CD);
+      yy_storeu_256(mask, _mm256_permute4x64_epi64(m8, 0xd8));
+      src0 += src0_stride << 2;
+      src1 += src1_stride << 2;
+      mask += 32;
+      i += 4;
+    } while (i < h);
+  } else if (w == 16) {
+    do {
+      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);
+      const __m256i m16A =
+          calc_mask_d16_inv_avx2(&s0A, &s1A, &_r, &y38, &y64, shift);
+      const __m256i m16B =
+          calc_mask_d16_inv_avx2(&s0B, &s1B, &_r, &y38, &y64, shift);
+      const __m256i m8 = _mm256_packus_epi16(m16A, m16B);
+      yy_storeu_256(mask, _mm256_permute4x64_epi64(m8, 0xd8));
+      src0 += src0_stride << 1;
+      src1 += src1_stride << 1;
+      mask += 32;
+      i += 2;
+    } while (i < h);
+  } else if (w == 32) {
+    do {
+      const __m256i s0A = yy_loadu_256(src0);
+      const __m256i s0B = yy_loadu_256(src0 + 16);
+      const __m256i s1A = yy_loadu_256(src1);
+      const __m256i s1B = yy_loadu_256(src1 + 16);
+      const __m256i m16A =
+          calc_mask_d16_inv_avx2(&s0A, &s1A, &_r, &y38, &y64, shift);
+      const __m256i m16B =
+          calc_mask_d16_inv_avx2(&s0B, &s1B, &_r, &y38, &y64, shift);
+      const __m256i m8 = _mm256_packus_epi16(m16A, m16B);
+      yy_storeu_256(mask, _mm256_permute4x64_epi64(m8, 0xd8));
+      src0 += src0_stride;
+      src1 += src1_stride;
+      mask += 32;
+      i += 1;
+    } while (i < h);
+  } else if (w == 64) {
+    do {
+      const __m256i s0A = yy_loadu_256(src0);
+      const __m256i s0B = yy_loadu_256(src0 + 16);
+      const __m256i s0C = yy_loadu_256(src0 + 32);
+      const __m256i s0D = yy_loadu_256(src0 + 48);
+      const __m256i s1A = yy_loadu_256(src1);
+      const __m256i s1B = yy_loadu_256(src1 + 16);
+      const __m256i s1C = yy_loadu_256(src1 + 32);
+      const __m256i s1D = yy_loadu_256(src1 + 48);
+      const __m256i m16A =
+          calc_mask_d16_inv_avx2(&s0A, &s1A, &_r, &y38, &y64, shift);
+      const __m256i m16B =
+          calc_mask_d16_inv_avx2(&s0B, &s1B, &_r, &y38, &y64, shift);
+      const __m256i m16C =
+          calc_mask_d16_inv_avx2(&s0C, &s1C, &_r, &y38, &y64, shift);
+      const __m256i m16D =
+          calc_mask_d16_inv_avx2(&s0D, &s1D, &_r, &y38, &y64, shift);
+      const __m256i m8AB = _mm256_packus_epi16(m16A, m16B);
+      const __m256i m8CD = _mm256_packus_epi16(m16C, m16D);
+      yy_storeu_256(mask, _mm256_permute4x64_epi64(m8AB, 0xd8));
+      yy_storeu_256(mask + 32, _mm256_permute4x64_epi64(m8CD, 0xd8));
+      src0 += src0_stride;
+      src1 += src1_stride;
+      mask += 64;
+      i += 1;
+    } while (i < h);
+  } else {
+    do {
+      const __m256i s0A = yy_loadu_256(src0);
+      const __m256i s0B = yy_loadu_256(src0 + 16);
+      const __m256i s0C = yy_loadu_256(src0 + 32);
+      const __m256i s0D = yy_loadu_256(src0 + 48);
+      const __m256i s0E = yy_loadu_256(src0 + 64);
+      const __m256i s0F = yy_loadu_256(src0 + 80);
+      const __m256i s0G = yy_loadu_256(src0 + 96);
+      const __m256i s0H = yy_loadu_256(src0 + 112);
+      const __m256i s1A = yy_loadu_256(src1);
+      const __m256i s1B = yy_loadu_256(src1 + 16);
+      const __m256i s1C = yy_loadu_256(src1 + 32);
+      const __m256i s1D = yy_loadu_256(src1 + 48);
+      const __m256i s1E = yy_loadu_256(src1 + 64);
+      const __m256i s1F = yy_loadu_256(src1 + 80);
+      const __m256i s1G = yy_loadu_256(src1 + 96);
+      const __m256i s1H = yy_loadu_256(src1 + 112);
+      const __m256i m16A =
+          calc_mask_d16_inv_avx2(&s0A, &s1A, &_r, &y38, &y64, shift);
+      const __m256i m16B =
+          calc_mask_d16_inv_avx2(&s0B, &s1B, &_r, &y38, &y64, shift);
+      const __m256i m16C =
+          calc_mask_d16_inv_avx2(&s0C, &s1C, &_r, &y38, &y64, shift);
+      const __m256i m16D =
+          calc_mask_d16_inv_avx2(&s0D, &s1D, &_r, &y38, &y64, shift);
+      const __m256i m16E =
+          calc_mask_d16_inv_avx2(&s0E, &s1E, &_r, &y38, &y64, shift);
+      const __m256i m16F =
+          calc_mask_d16_inv_avx2(&s0F, &s1F, &_r, &y38, &y64, shift);
+      const __m256i m16G =
+          calc_mask_d16_inv_avx2(&s0G, &s1G, &_r, &y38, &y64, shift);
+      const __m256i m16H =
+          calc_mask_d16_inv_avx2(&s0H, &s1H, &_r, &y38, &y64, shift);
+      const __m256i m8AB = _mm256_packus_epi16(m16A, m16B);
+      const __m256i m8CD = _mm256_packus_epi16(m16C, m16D);
+      const __m256i m8EF = _mm256_packus_epi16(m16E, m16F);
+      const __m256i m8GH = _mm256_packus_epi16(m16G, m16H);
+      yy_storeu_256(mask, _mm256_permute4x64_epi64(m8AB, 0xd8));
+      yy_storeu_256(mask + 32, _mm256_permute4x64_epi64(m8CD, 0xd8));
+      yy_storeu_256(mask + 64, _mm256_permute4x64_epi64(m8EF, 0xd8));
+      yy_storeu_256(mask + 96, _mm256_permute4x64_epi64(m8GH, 0xd8));
+      src0 += src0_stride;
+      src1 += src1_stride;
+      mask += 128;
+      i += 1;
+    } while (i < h);
+  }
+}
+
+void av1_build_compound_diffwtd_mask_d16_avx2(
+    uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const CONV_BUF_TYPE *src0,
+    int src0_stride, const CONV_BUF_TYPE *src1, int src1_stride, int h, int w,
+    ConvolveParams *conv_params, int bd) {
+  const int shift =
+      2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1 + (bd - 8);
+  // When rounding constant is added, there is a possibility of overflow.
+  // However that much precision is not required. Code should very well work for
+  // other values of DIFF_FACTOR_LOG2 and AOM_BLEND_A64_MAX_ALPHA as well. But
+  // there is a possibility of corner case bugs.
+  assert(DIFF_FACTOR_LOG2 == 4);
+  assert(AOM_BLEND_A64_MAX_ALPHA == 64);
+
+  if (mask_type == DIFFWTD_38) {
+    build_compound_diffwtd_mask_d16_avx2(mask, src0, src0_stride, src1,
+                                         src1_stride, h, w, shift);
+  } else {
+    build_compound_diffwtd_mask_d16_inv_avx2(mask, src0, src0_stride, src1,
+                                             src1_stride, h, w, shift);
+  }
+}
+
+void av1_build_compound_diffwtd_mask_highbd_avx2(
+    uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0,
+    int src0_stride, const uint8_t *src1, int src1_stride, int h, int w,
+    int bd) {
+  if (w < 16) {
+    av1_build_compound_diffwtd_mask_highbd_ssse3(
+        mask, mask_type, src0, src0_stride, src1, src1_stride, h, w, bd);
+  } else {
+    assert(mask_type == DIFFWTD_38 || mask_type == DIFFWTD_38_INV);
+    assert(bd >= 8);
+    assert((w % 16) == 0);
+    const __m256i y0 = _mm256_setzero_si256();
+    const __m256i yAOM_BLEND_A64_MAX_ALPHA =
+        _mm256_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+    const int mask_base = 38;
+    const __m256i ymask_base = _mm256_set1_epi16(mask_base);
+    const uint16_t *ssrc0 = CONVERT_TO_SHORTPTR(src0);
+    const uint16_t *ssrc1 = CONVERT_TO_SHORTPTR(src1);
+    if (bd == 8) {
+      if (mask_type == DIFFWTD_38_INV) {
+        for (int i = 0; i < h; ++i) {
+          for (int j = 0; j < w; j += 16) {
+            __m256i s0 = _mm256_loadu_si256((const __m256i *)&ssrc0[j]);
+            __m256i s1 = _mm256_loadu_si256((const __m256i *)&ssrc1[j]);
+            __m256i diff = _mm256_srai_epi16(
+                _mm256_abs_epi16(_mm256_sub_epi16(s0, s1)), DIFF_FACTOR_LOG2);
+            __m256i m = _mm256_min_epi16(
+                _mm256_max_epi16(y0, _mm256_add_epi16(diff, ymask_base)),
+                yAOM_BLEND_A64_MAX_ALPHA);
+            m = _mm256_sub_epi16(yAOM_BLEND_A64_MAX_ALPHA, m);
+            m = _mm256_packus_epi16(m, m);
+            m = _mm256_permute4x64_epi64(m, _MM_SHUFFLE(0, 0, 2, 0));
+            __m128i m0 = _mm256_castsi256_si128(m);
+            _mm_storeu_si128((__m128i *)&mask[j], m0);
+          }
+          ssrc0 += src0_stride;
+          ssrc1 += src1_stride;
+          mask += w;
+        }
+      } else {
+        for (int i = 0; i < h; ++i) {
+          for (int j = 0; j < w; j += 16) {
+            __m256i s0 = _mm256_loadu_si256((const __m256i *)&ssrc0[j]);
+            __m256i s1 = _mm256_loadu_si256((const __m256i *)&ssrc1[j]);
+            __m256i diff = _mm256_srai_epi16(
+                _mm256_abs_epi16(_mm256_sub_epi16(s0, s1)), DIFF_FACTOR_LOG2);
+            __m256i m = _mm256_min_epi16(
+                _mm256_max_epi16(y0, _mm256_add_epi16(diff, ymask_base)),
+                yAOM_BLEND_A64_MAX_ALPHA);
+            m = _mm256_packus_epi16(m, m);
+            m = _mm256_permute4x64_epi64(m, _MM_SHUFFLE(0, 0, 2, 0));
+            __m128i m0 = _mm256_castsi256_si128(m);
+            _mm_storeu_si128((__m128i *)&mask[j], m0);
+          }
+          ssrc0 += src0_stride;
+          ssrc1 += src1_stride;
+          mask += w;
+        }
+      }
+    } else {
+      const __m128i xshift = xx_set1_64_from_32i(bd - 8 + DIFF_FACTOR_LOG2);
+      if (mask_type == DIFFWTD_38_INV) {
+        for (int i = 0; i < h; ++i) {
+          for (int j = 0; j < w; j += 16) {
+            __m256i s0 = _mm256_loadu_si256((const __m256i *)&ssrc0[j]);
+            __m256i s1 = _mm256_loadu_si256((const __m256i *)&ssrc1[j]);
+            __m256i diff = _mm256_sra_epi16(
+                _mm256_abs_epi16(_mm256_sub_epi16(s0, s1)), xshift);
+            __m256i m = _mm256_min_epi16(
+                _mm256_max_epi16(y0, _mm256_add_epi16(diff, ymask_base)),
+                yAOM_BLEND_A64_MAX_ALPHA);
+            m = _mm256_sub_epi16(yAOM_BLEND_A64_MAX_ALPHA, m);
+            m = _mm256_packus_epi16(m, m);
+            m = _mm256_permute4x64_epi64(m, _MM_SHUFFLE(0, 0, 2, 0));
+            __m128i m0 = _mm256_castsi256_si128(m);
+            _mm_storeu_si128((__m128i *)&mask[j], m0);
+          }
+          ssrc0 += src0_stride;
+          ssrc1 += src1_stride;
+          mask += w;
+        }
+      } else {
+        for (int i = 0; i < h; ++i) {
+          for (int j = 0; j < w; j += 16) {
+            __m256i s0 = _mm256_loadu_si256((const __m256i *)&ssrc0[j]);
+            __m256i s1 = _mm256_loadu_si256((const __m256i *)&ssrc1[j]);
+            __m256i diff = _mm256_sra_epi16(
+                _mm256_abs_epi16(_mm256_sub_epi16(s0, s1)), xshift);
+            __m256i m = _mm256_min_epi16(
+                _mm256_max_epi16(y0, _mm256_add_epi16(diff, ymask_base)),
+                yAOM_BLEND_A64_MAX_ALPHA);
+            m = _mm256_packus_epi16(m, m);
+            m = _mm256_permute4x64_epi64(m, _MM_SHUFFLE(0, 0, 2, 0));
+            __m128i m0 = _mm256_castsi256_si128(m);
+            _mm_storeu_si128((__m128i *)&mask[j], m0);
+          }
+          ssrc0 += src0_stride;
+          ssrc1 += src1_stride;
+          mask += w;
+        }
+      }
+    }
+  }
+}
diff --git a/third_party/aom/av1/common/x86/reconinter_sse4.c b/third_party/aom/av1/common/x86/reconinter_sse4.c
new file mode 100644
index 000000000..5171ca493
--- /dev/null
+++ b/third_party/aom/av1/common/x86/reconinter_sse4.c
@@ -0,0 +1,153 @@
+/*
+ * 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 <smmintrin.h>  /* SSE4.1 */
+
+#include "aom/aom_integer.h"
+#include "aom_dsp/blend.h"
+#include "av1/common/blockd.h"
+
+static INLINE __m128i calc_mask(const __m128i mask_base, const __m128i s0,
+                                const __m128i s1) {
+  const __m128i diff = _mm_abs_epi16(_mm_sub_epi16(s0, s1));
+  return _mm_abs_epi16(_mm_add_epi16(mask_base, _mm_srli_epi16(diff, 4)));
+  // clamp(diff, 0, 64) can be skiped for diff is always in the range ( 38, 54)
+}
+
+void av1_build_compound_diffwtd_mask_sse4_1(uint8_t *mask,
+                                            DIFFWTD_MASK_TYPE mask_type,
+                                            const uint8_t *src0, int stride0,
+                                            const uint8_t *src1, int stride1,
+                                            int h, int w) {
+  const int mb = (mask_type == DIFFWTD_38_INV) ? AOM_BLEND_A64_MAX_ALPHA : 0;
+  const __m128i mask_base = _mm_set1_epi16(38 - mb);
+  int i = 0;
+  if (4 == w) {
+    do {
+      const __m128i s0A = _mm_cvtsi32_si128(*(uint32_t *)src0);
+      const __m128i s0B = _mm_cvtsi32_si128(*(uint32_t *)(src0 + stride0));
+      const __m128i s0AB = _mm_unpacklo_epi32(s0A, s0B);
+      const __m128i s0 = _mm_cvtepu8_epi16(s0AB);
+
+      const __m128i s1A = _mm_cvtsi32_si128(*(uint32_t *)src1);
+      const __m128i s1B = _mm_cvtsi32_si128(*(uint32_t *)(src1 + stride1));
+      const __m128i s1AB = _mm_unpacklo_epi32(s1A, s1B);
+      const __m128i s1 = _mm_cvtepu8_epi16(s1AB);
+
+      const __m128i m16 = calc_mask(mask_base, s0, s1);
+      const __m128i m8 = _mm_packus_epi16(m16, m16);
+
+      *(uint32_t *)mask = _mm_cvtsi128_si32(m8);
+      *(uint32_t *)(mask + w) = _mm_extract_epi32(m8, 1);
+      src0 += (stride0 << 1);
+      src1 += (stride1 << 1);
+      mask += 8;
+      i += 2;
+    } while (i < h);
+  } else if (8 == w) {
+    do {
+      __m128i s0 = _mm_loadl_epi64((__m128i const *)src0);
+      __m128i s1 = _mm_loadl_epi64((__m128i const *)src1);
+      s0 = _mm_cvtepu8_epi16(s0);
+      s1 = _mm_cvtepu8_epi16(s1);
+      const __m128i m16 = calc_mask(mask_base, s0, s1);
+      const __m128i m8 = _mm_packus_epi16(m16, m16);
+      _mm_storel_epi64((__m128i *)mask, m8);
+      src0 += stride0;
+      src1 += stride1;
+      mask += 8;
+      i += 1;
+    } while (i < h);
+  } else {
+    const __m128i zero = _mm_setzero_si128();
+    do {
+      int j = 0;
+      do {
+        const __m128i s0 = _mm_load_si128((__m128i const *)(src0 + j));
+        const __m128i s1 = _mm_load_si128((__m128i const *)(src1 + j));
+        const __m128i s0L = _mm_cvtepu8_epi16(s0);
+        const __m128i s1L = _mm_cvtepu8_epi16(s1);
+        const __m128i s0H = _mm_unpackhi_epi8(s0, zero);
+        const __m128i s1H = _mm_unpackhi_epi8(s1, zero);
+
+        const __m128i m16L = calc_mask(mask_base, s0L, s1L);
+        const __m128i m16H = calc_mask(mask_base, s0H, s1H);
+
+        const __m128i m8 = _mm_packus_epi16(m16L, m16H);
+        _mm_store_si128((__m128i *)(mask + j), m8);
+        j += 16;
+      } while (j < w);
+      src0 += stride0;
+      src1 += stride1;
+      mask += w;
+      i += 1;
+    } while (i < h);
+  }
+}
+
+void av1_build_compound_diffwtd_mask_d16_sse4_1(
+    uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const CONV_BUF_TYPE *src0,
+    int src0_stride, const CONV_BUF_TYPE *src1, int src1_stride, int h, int w,
+    ConvolveParams *conv_params, int bd) {
+  const int which_inverse = (mask_type == DIFFWTD_38) ? 0 : 1;
+  const int mask_base = 38;
+  int round =
+      2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1 + (bd - 8);
+  const __m128i round_const = _mm_set1_epi16((1 << round) >> 1);
+  const __m128i mask_base_16 = _mm_set1_epi16(mask_base);
+  const __m128i clip_diff = _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+  const __m128i add_const =
+      _mm_set1_epi16((which_inverse ? AOM_BLEND_A64_MAX_ALPHA : 0));
+  const __m128i add_sign = _mm_set1_epi16((which_inverse ? -1 : 1));
+
+  int i, j;
+  // When rounding constant is added, there is a possibility of overflow.
+  // However that much precision is not required. Code should very well work for
+  // other values of DIFF_FACTOR_LOG2 and AOM_BLEND_A64_MAX_ALPHA as well. But
+  // there is a possibility of corner case bugs.
+  assert(DIFF_FACTOR_LOG2 == 4);
+  assert(AOM_BLEND_A64_MAX_ALPHA == 64);
+  for (i = 0; i < h; ++i) {
+    for (j = 0; j < w; j += 8) {
+      const __m128i data_src0 =
+          _mm_loadu_si128((__m128i *)&src0[(i * src0_stride) + j]);
+      const __m128i data_src1 =
+          _mm_loadu_si128((__m128i *)&src1[(i * src1_stride) + j]);
+
+      const __m128i diffa = _mm_subs_epu16(data_src0, data_src1);
+      const __m128i diffb = _mm_subs_epu16(data_src1, data_src0);
+      const __m128i diff = _mm_max_epu16(diffa, diffb);
+      const __m128i diff_round =
+          _mm_srli_epi16(_mm_adds_epu16(diff, round_const), round);
+      const __m128i diff_factor = _mm_srli_epi16(diff_round, DIFF_FACTOR_LOG2);
+      const __m128i diff_mask = _mm_adds_epi16(diff_factor, mask_base_16);
+      __m128i diff_clamp = _mm_min_epi16(diff_mask, clip_diff);
+      // clamp to 0 can be skipped since we are using add and saturate
+      // instruction
+
+      const __m128i diff_sign = _mm_sign_epi16(diff_clamp, add_sign);
+      const __m128i diff_const_16 = _mm_add_epi16(diff_sign, add_const);
+
+      // 8 bit conversion and saturation to uint8
+      const __m128i res_8 = _mm_packus_epi16(diff_const_16, diff_const_16);
+
+      // Store values into the destination buffer
+      __m128i *const dst = (__m128i *)&mask[i * w + j];
+
+      if ((w - j) > 4) {
+        _mm_storel_epi64(dst, res_8);
+      } else {  // w==4
+        *(uint32_t *)dst = _mm_cvtsi128_si32(res_8);
+      }
+    }
+  }
+}
diff --git a/third_party/aom/av1/common/x86/reconinter_ssse3.c b/third_party/aom/av1/common/x86/reconinter_ssse3.c
new file mode 100644
index 000000000..cf684447c
--- /dev/null
+++ b/third_party/aom/av1/common/x86/reconinter_ssse3.c
@@ -0,0 +1,116 @@
+/*
+ * 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/av1_rtcd.h"
+
+#include "aom/aom_integer.h"
+#include "aom_dsp/blend.h"
+#include "aom_dsp/x86/synonyms.h"
+#include "av1/common/blockd.h"
+
+void av1_build_compound_diffwtd_mask_highbd_ssse3(
+    uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0,
+    int src0_stride, const uint8_t *src1, int src1_stride, int h, int w,
+    int bd) {
+  if (w < 8) {
+    av1_build_compound_diffwtd_mask_highbd_c(mask, mask_type, src0, src0_stride,
+                                             src1, src1_stride, h, w, bd);
+  } else {
+    assert(bd >= 8);
+    assert((w % 8) == 0);
+    assert(mask_type == DIFFWTD_38 || mask_type == DIFFWTD_38_INV);
+    const __m128i x0 = _mm_setzero_si128();
+    const __m128i xAOM_BLEND_A64_MAX_ALPHA =
+        _mm_set1_epi16(AOM_BLEND_A64_MAX_ALPHA);
+    const int mask_base = 38;
+    const __m128i xmask_base = _mm_set1_epi16(mask_base);
+    const uint16_t *ssrc0 = CONVERT_TO_SHORTPTR(src0);
+    const uint16_t *ssrc1 = CONVERT_TO_SHORTPTR(src1);
+    if (bd == 8) {
+      if (mask_type == DIFFWTD_38_INV) {
+        for (int i = 0; i < h; ++i) {
+          for (int j = 0; j < w; j += 8) {
+            __m128i s0 = _mm_loadu_si128((const __m128i *)&ssrc0[j]);
+            __m128i s1 = _mm_loadu_si128((const __m128i *)&ssrc1[j]);
+            __m128i diff = _mm_srai_epi16(_mm_abs_epi16(_mm_sub_epi16(s0, s1)),
+                                          DIFF_FACTOR_LOG2);
+            __m128i m = _mm_min_epi16(
+                _mm_max_epi16(x0, _mm_add_epi16(diff, xmask_base)),
+                xAOM_BLEND_A64_MAX_ALPHA);
+            m = _mm_sub_epi16(xAOM_BLEND_A64_MAX_ALPHA, m);
+            m = _mm_packus_epi16(m, m);
+            _mm_storel_epi64((__m128i *)&mask[j], m);
+          }
+          ssrc0 += src0_stride;
+          ssrc1 += src1_stride;
+          mask += w;
+        }
+      } else {
+        for (int i = 0; i < h; ++i) {
+          for (int j = 0; j < w; j += 8) {
+            __m128i s0 = _mm_loadu_si128((const __m128i *)&ssrc0[j]);
+            __m128i s1 = _mm_loadu_si128((const __m128i *)&ssrc1[j]);
+            __m128i diff = _mm_srai_epi16(_mm_abs_epi16(_mm_sub_epi16(s0, s1)),
+                                          DIFF_FACTOR_LOG2);
+            __m128i m = _mm_min_epi16(
+                _mm_max_epi16(x0, _mm_add_epi16(diff, xmask_base)),
+                xAOM_BLEND_A64_MAX_ALPHA);
+            m = _mm_packus_epi16(m, m);
+            _mm_storel_epi64((__m128i *)&mask[j], m);
+          }
+          ssrc0 += src0_stride;
+          ssrc1 += src1_stride;
+          mask += w;
+        }
+      }
+    } else {
+      const __m128i xshift = xx_set1_64_from_32i(bd - 8 + DIFF_FACTOR_LOG2);
+      if (mask_type == DIFFWTD_38_INV) {
+        for (int i = 0; i < h; ++i) {
+          for (int j = 0; j < w; j += 8) {
+            __m128i s0 = _mm_loadu_si128((const __m128i *)&ssrc0[j]);
+            __m128i s1 = _mm_loadu_si128((const __m128i *)&ssrc1[j]);
+            __m128i diff =
+                _mm_sra_epi16(_mm_abs_epi16(_mm_sub_epi16(s0, s1)), xshift);
+            __m128i m = _mm_min_epi16(
+                _mm_max_epi16(x0, _mm_add_epi16(diff, xmask_base)),
+                xAOM_BLEND_A64_MAX_ALPHA);
+            m = _mm_sub_epi16(xAOM_BLEND_A64_MAX_ALPHA, m);
+            m = _mm_packus_epi16(m, m);
+            _mm_storel_epi64((__m128i *)&mask[j], m);
+          }
+          ssrc0 += src0_stride;
+          ssrc1 += src1_stride;
+          mask += w;
+        }
+      } else {
+        for (int i = 0; i < h; ++i) {
+          for (int j = 0; j < w; j += 8) {
+            __m128i s0 = _mm_loadu_si128((const __m128i *)&ssrc0[j]);
+            __m128i s1 = _mm_loadu_si128((const __m128i *)&ssrc1[j]);
+            __m128i diff =
+                _mm_sra_epi16(_mm_abs_epi16(_mm_sub_epi16(s0, s1)), xshift);
+            __m128i m = _mm_min_epi16(
+                _mm_max_epi16(x0, _mm_add_epi16(diff, xmask_base)),
+                xAOM_BLEND_A64_MAX_ALPHA);
+            m = _mm_packus_epi16(m, m);
+            _mm_storel_epi64((__m128i *)&mask[j], m);
+          }
+          ssrc0 += src0_stride;
+          ssrc1 += src1_stride;
+          mask += w;
+        }
+      }
+    }
+  }
+}
diff --git a/third_party/aom/av1/common/x86/selfguided_avx2.c b/third_party/aom/av1/common/x86/selfguided_avx2.c
new file mode 100644
index 000000000..0aaf1f454
--- /dev/null
+++ b/third_party/aom/av1/common/x86/selfguided_avx2.c
@@ -0,0 +1,724 @@
+/*
+ * 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_config.h"
+#include "config/av1_rtcd.h"
+
+#include "av1/common/restoration.h"
+#include "aom_dsp/x86/synonyms.h"
+#include "aom_dsp/x86/synonyms_avx2.h"
+
+// Load 8 bytes from the possibly-misaligned pointer p, extend each byte to
+// 32-bit precision and return them in an AVX2 register.
+static __m256i yy256_load_extend_8_32(const void *p) {
+  return _mm256_cvtepu8_epi32(xx_loadl_64(p));
+}
+
+// Load 8 halfwords from the possibly-misaligned pointer p, extend each
+// halfword to 32-bit precision and return them in an AVX2 register.
+static __m256i yy256_load_extend_16_32(const void *p) {
+  return _mm256_cvtepu16_epi32(xx_loadu_128(p));
+}
+
+// Compute the scan of an AVX2 register holding 8 32-bit integers. If the
+// register holds x0..x7 then the scan will hold x0, x0+x1, x0+x1+x2, ...,
+// x0+x1+...+x7
+//
+// Let [...] represent a 128-bit block, and let a, ..., h be 32-bit integers
+// (assumed small enough to be able to add them without overflow).
+//
+// Use -> as shorthand for summing, i.e. h->a = h + g + f + e + d + c + b + a.
+//
+// x   = [h g f e][d c b a]
+// x01 = [g f e 0][c b a 0]
+// x02 = [g+h f+g e+f e][c+d b+c a+b a]
+// x03 = [e+f e 0 0][a+b a 0 0]
+// x04 = [e->h e->g e->f e][a->d a->c a->b a]
+// s   = a->d
+// s01 = [a->d a->d a->d a->d]
+// s02 = [a->d a->d a->d a->d][0 0 0 0]
+// ret = [a->h a->g a->f a->e][a->d a->c a->b a]
+static __m256i scan_32(__m256i x) {
+  const __m256i x01 = _mm256_slli_si256(x, 4);
+  const __m256i x02 = _mm256_add_epi32(x, x01);
+  const __m256i x03 = _mm256_slli_si256(x02, 8);
+  const __m256i x04 = _mm256_add_epi32(x02, x03);
+  const int32_t s = _mm256_extract_epi32(x04, 3);
+  const __m128i s01 = _mm_set1_epi32(s);
+  const __m256i s02 = _mm256_insertf128_si256(_mm256_setzero_si256(), s01, 1);
+  return _mm256_add_epi32(x04, s02);
+}
+
+// Compute two integral images from src. B sums elements; A sums their
+// squares. The images are offset by one pixel, so will have width and height
+// equal to width + 1, height + 1 and the first row and column will be zero.
+//
+// A+1 and B+1 should be aligned to 32 bytes. buf_stride should be a multiple
+// of 8.
+
+static void *memset_zero_avx(int32_t *dest, const __m256i *zero, size_t count) {
+  unsigned int i = 0;
+  for (i = 0; i < (count & 0xffffffe0); i += 32) {
+    _mm256_storeu_si256((__m256i *)(dest + i), *zero);
+    _mm256_storeu_si256((__m256i *)(dest + i + 8), *zero);
+    _mm256_storeu_si256((__m256i *)(dest + i + 16), *zero);
+    _mm256_storeu_si256((__m256i *)(dest + i + 24), *zero);
+  }
+  for (; i < (count & 0xfffffff8); i += 8) {
+    _mm256_storeu_si256((__m256i *)(dest + i), *zero);
+  }
+  for (; i < count; i++) {
+    dest[i] = 0;
+  }
+  return dest;
+}
+
+static void integral_images(const uint8_t *src, int src_stride, int width,
+                            int height, int32_t *A, int32_t *B,
+                            int buf_stride) {
+  const __m256i zero = _mm256_setzero_si256();
+  // Write out the zero top row
+  memset_zero_avx(A, &zero, (width + 8));
+  memset_zero_avx(B, &zero, (width + 8));
+  for (int i = 0; i < height; ++i) {
+    // Zero the left column.
+    A[(i + 1) * buf_stride] = B[(i + 1) * buf_stride] = 0;
+
+    // ldiff is the difference H - D where H is the output sample immediately
+    // to the left and D is the output sample above it. These are scalars,
+    // replicated across the eight lanes.
+    __m256i ldiff1 = zero, ldiff2 = zero;
+    for (int j = 0; j < width; j += 8) {
+      const int ABj = 1 + j;
+
+      const __m256i above1 = yy_load_256(B + ABj + i * buf_stride);
+      const __m256i above2 = yy_load_256(A + ABj + i * buf_stride);
+
+      const __m256i x1 = yy256_load_extend_8_32(src + j + i * src_stride);
+      const __m256i x2 = _mm256_madd_epi16(x1, x1);
+
+      const __m256i sc1 = scan_32(x1);
+      const __m256i sc2 = scan_32(x2);
+
+      const __m256i row1 =
+          _mm256_add_epi32(_mm256_add_epi32(sc1, above1), ldiff1);
+      const __m256i row2 =
+          _mm256_add_epi32(_mm256_add_epi32(sc2, above2), ldiff2);
+
+      yy_store_256(B + ABj + (i + 1) * buf_stride, row1);
+      yy_store_256(A + ABj + (i + 1) * buf_stride, row2);
+
+      // Calculate the new H - D.
+      ldiff1 = _mm256_set1_epi32(
+          _mm256_extract_epi32(_mm256_sub_epi32(row1, above1), 7));
+      ldiff2 = _mm256_set1_epi32(
+          _mm256_extract_epi32(_mm256_sub_epi32(row2, above2), 7));
+    }
+  }
+}
+
+// Compute two integral images from src. B sums elements; A sums their squares
+//
+// A and B should be aligned to 32 bytes. buf_stride should be a multiple of 8.
+static void integral_images_highbd(const uint16_t *src, int src_stride,
+                                   int width, int height, int32_t *A,
+                                   int32_t *B, int buf_stride) {
+  const __m256i zero = _mm256_setzero_si256();
+  // Write out the zero top row
+  memset_zero_avx(A, &zero, (width + 8));
+  memset_zero_avx(B, &zero, (width + 8));
+
+  for (int i = 0; i < height; ++i) {
+    // Zero the left column.
+    A[(i + 1) * buf_stride] = B[(i + 1) * buf_stride] = 0;
+
+    // ldiff is the difference H - D where H is the output sample immediately
+    // to the left and D is the output sample above it. These are scalars,
+    // replicated across the eight lanes.
+    __m256i ldiff1 = zero, ldiff2 = zero;
+    for (int j = 0; j < width; j += 8) {
+      const int ABj = 1 + j;
+
+      const __m256i above1 = yy_load_256(B + ABj + i * buf_stride);
+      const __m256i above2 = yy_load_256(A + ABj + i * buf_stride);
+
+      const __m256i x1 = yy256_load_extend_16_32(src + j + i * src_stride);
+      const __m256i x2 = _mm256_madd_epi16(x1, x1);
+
+      const __m256i sc1 = scan_32(x1);
+      const __m256i sc2 = scan_32(x2);
+
+      const __m256i row1 =
+          _mm256_add_epi32(_mm256_add_epi32(sc1, above1), ldiff1);
+      const __m256i row2 =
+          _mm256_add_epi32(_mm256_add_epi32(sc2, above2), ldiff2);
+
+      yy_store_256(B + ABj + (i + 1) * buf_stride, row1);
+      yy_store_256(A + ABj + (i + 1) * buf_stride, row2);
+
+      // Calculate the new H - D.
+      ldiff1 = _mm256_set1_epi32(
+          _mm256_extract_epi32(_mm256_sub_epi32(row1, above1), 7));
+      ldiff2 = _mm256_set1_epi32(
+          _mm256_extract_epi32(_mm256_sub_epi32(row2, above2), 7));
+    }
+  }
+}
+
+// Compute 8 values of boxsum from the given integral image. ii should point
+// at the middle of the box (for the first value). r is the box radius.
+static INLINE __m256i boxsum_from_ii(const int32_t *ii, int stride, int r) {
+  const __m256i tl = yy_loadu_256(ii - (r + 1) - (r + 1) * stride);
+  const __m256i tr = yy_loadu_256(ii + (r + 0) - (r + 1) * stride);
+  const __m256i bl = yy_loadu_256(ii - (r + 1) + r * stride);
+  const __m256i br = yy_loadu_256(ii + (r + 0) + r * stride);
+  const __m256i u = _mm256_sub_epi32(tr, tl);
+  const __m256i v = _mm256_sub_epi32(br, bl);
+  return _mm256_sub_epi32(v, u);
+}
+
+static __m256i round_for_shift(unsigned shift) {
+  return _mm256_set1_epi32((1 << shift) >> 1);
+}
+
+static __m256i compute_p(__m256i sum1, __m256i sum2, int bit_depth, int n) {
+  __m256i an, bb;
+  if (bit_depth > 8) {
+    const __m256i rounding_a = round_for_shift(2 * (bit_depth - 8));
+    const __m256i rounding_b = round_for_shift(bit_depth - 8);
+    const __m128i shift_a = _mm_cvtsi32_si128(2 * (bit_depth - 8));
+    const __m128i shift_b = _mm_cvtsi32_si128(bit_depth - 8);
+    const __m256i a =
+        _mm256_srl_epi32(_mm256_add_epi32(sum2, rounding_a), shift_a);
+    const __m256i b =
+        _mm256_srl_epi32(_mm256_add_epi32(sum1, rounding_b), shift_b);
+    // b < 2^14, so we can use a 16-bit madd rather than a 32-bit
+    // mullo to square it
+    bb = _mm256_madd_epi16(b, b);
+    an = _mm256_max_epi32(_mm256_mullo_epi32(a, _mm256_set1_epi32(n)), bb);
+  } else {
+    bb = _mm256_madd_epi16(sum1, sum1);
+    an = _mm256_mullo_epi32(sum2, _mm256_set1_epi32(n));
+  }
+  return _mm256_sub_epi32(an, bb);
+}
+
+// Assumes that C, D are integral images for the original buffer which has been
+// extended to have a padding of SGRPROJ_BORDER_VERT/SGRPROJ_BORDER_HORZ pixels
+// on the sides. A, B, C, D point at logical position (0, 0).
+static void calc_ab(int32_t *A, int32_t *B, const int32_t *C, const int32_t *D,
+                    int width, int height, int buf_stride, int bit_depth,
+                    int sgr_params_idx, int radius_idx) {
+  const sgr_params_type *const params = &sgr_params[sgr_params_idx];
+  const int r = params->r[radius_idx];
+  const int n = (2 * r + 1) * (2 * r + 1);
+  const __m256i s = _mm256_set1_epi32(params->s[radius_idx]);
+  // one_over_n[n-1] is 2^12/n, so easily fits in an int16
+  const __m256i one_over_n = _mm256_set1_epi32(one_by_x[n - 1]);
+
+  const __m256i rnd_z = round_for_shift(SGRPROJ_MTABLE_BITS);
+  const __m256i rnd_res = round_for_shift(SGRPROJ_RECIP_BITS);
+
+  // Set up masks
+  const __m128i ones32 = _mm_set_epi32(0, 0, 0xffffffff, 0xffffffff);
+  __m256i mask[8];
+  for (int idx = 0; idx < 8; idx++) {
+    const __m128i shift = _mm_cvtsi32_si128(8 * (8 - idx));
+    mask[idx] = _mm256_cvtepi8_epi32(_mm_srl_epi64(ones32, shift));
+  }
+
+  for (int i = -1; i < height + 1; ++i) {
+    for (int j = -1; j < width + 1; j += 8) {
+      const int32_t *Cij = C + i * buf_stride + j;
+      const int32_t *Dij = D + i * buf_stride + j;
+
+      __m256i sum1 = boxsum_from_ii(Dij, buf_stride, r);
+      __m256i sum2 = boxsum_from_ii(Cij, buf_stride, r);
+
+      // When width + 2 isn't a multiple of 8, sum1 and sum2 will contain
+      // some uninitialised data in their upper words. We use a mask to
+      // ensure that these bits are set to 0.
+      int idx = AOMMIN(8, width + 1 - j);
+      assert(idx >= 1);
+
+      if (idx < 8) {
+        sum1 = _mm256_and_si256(mask[idx], sum1);
+        sum2 = _mm256_and_si256(mask[idx], sum2);
+      }
+
+      const __m256i p = compute_p(sum1, sum2, bit_depth, n);
+
+      const __m256i z = _mm256_min_epi32(
+          _mm256_srli_epi32(_mm256_add_epi32(_mm256_mullo_epi32(p, s), rnd_z),
+                            SGRPROJ_MTABLE_BITS),
+          _mm256_set1_epi32(255));
+
+      const __m256i a_res = _mm256_i32gather_epi32(x_by_xplus1, z, 4);
+
+      yy_storeu_256(A + i * buf_stride + j, a_res);
+
+      const __m256i a_complement =
+          _mm256_sub_epi32(_mm256_set1_epi32(SGRPROJ_SGR), a_res);
+
+      // sum1 might have lanes greater than 2^15, so we can't use madd to do
+      // multiplication involving sum1. However, a_complement and one_over_n
+      // are both less than 256, so we can multiply them first.
+      const __m256i a_comp_over_n = _mm256_madd_epi16(a_complement, one_over_n);
+      const __m256i b_int = _mm256_mullo_epi32(a_comp_over_n, sum1);
+      const __m256i b_res = _mm256_srli_epi32(_mm256_add_epi32(b_int, rnd_res),
+                                              SGRPROJ_RECIP_BITS);
+
+      yy_storeu_256(B + i * buf_stride + j, b_res);
+    }
+  }
+}
+
+// Calculate 8 values of the "cross sum" starting at buf. This is a 3x3 filter
+// where the outer four corners have weight 3 and all other pixels have weight
+// 4.
+//
+// Pixels are indexed as follows:
+// xtl  xt   xtr
+// xl    x   xr
+// xbl  xb   xbr
+//
+// buf points to x
+//
+// fours = xl + xt + xr + xb + x
+// threes = xtl + xtr + xbr + xbl
+// cross_sum = 4 * fours + 3 * threes
+//           = 4 * (fours + threes) - threes
+//           = (fours + threes) << 2 - threes
+static INLINE __m256i cross_sum(const int32_t *buf, int stride) {
+  const __m256i xtl = yy_loadu_256(buf - 1 - stride);
+  const __m256i xt = yy_loadu_256(buf - stride);
+  const __m256i xtr = yy_loadu_256(buf + 1 - stride);
+  const __m256i xl = yy_loadu_256(buf - 1);
+  const __m256i x = yy_loadu_256(buf);
+  const __m256i xr = yy_loadu_256(buf + 1);
+  const __m256i xbl = yy_loadu_256(buf - 1 + stride);
+  const __m256i xb = yy_loadu_256(buf + stride);
+  const __m256i xbr = yy_loadu_256(buf + 1 + stride);
+
+  const __m256i fours = _mm256_add_epi32(
+      xl, _mm256_add_epi32(xt, _mm256_add_epi32(xr, _mm256_add_epi32(xb, x))));
+  const __m256i threes =
+      _mm256_add_epi32(xtl, _mm256_add_epi32(xtr, _mm256_add_epi32(xbr, xbl)));
+
+  return _mm256_sub_epi32(_mm256_slli_epi32(_mm256_add_epi32(fours, threes), 2),
+                          threes);
+}
+
+// The final filter for self-guided restoration. Computes a weighted average
+// across A, B with "cross sums" (see cross_sum implementation above).
+static void final_filter(int32_t *dst, int dst_stride, const int32_t *A,
+                         const int32_t *B, int buf_stride, const void *dgd8,
+                         int dgd_stride, int width, int height, int highbd) {
+  const int nb = 5;
+  const __m256i rounding =
+      round_for_shift(SGRPROJ_SGR_BITS + nb - SGRPROJ_RST_BITS);
+  const uint8_t *dgd_real =
+      highbd ? (const uint8_t *)CONVERT_TO_SHORTPTR(dgd8) : dgd8;
+
+  for (int i = 0; i < height; ++i) {
+    for (int j = 0; j < width; j += 8) {
+      const __m256i a = cross_sum(A + i * buf_stride + j, buf_stride);
+      const __m256i b = cross_sum(B + i * buf_stride + j, buf_stride);
+
+      const __m128i raw =
+          xx_loadu_128(dgd_real + ((i * dgd_stride + j) << highbd));
+      const __m256i src =
+          highbd ? _mm256_cvtepu16_epi32(raw) : _mm256_cvtepu8_epi32(raw);
+
+      __m256i v = _mm256_add_epi32(_mm256_madd_epi16(a, src), b);
+      __m256i w = _mm256_srai_epi32(_mm256_add_epi32(v, rounding),
+                                    SGRPROJ_SGR_BITS + nb - SGRPROJ_RST_BITS);
+
+      yy_storeu_256(dst + i * dst_stride + j, w);
+    }
+  }
+}
+
+// Assumes that C, D are integral images for the original buffer which has been
+// extended to have a padding of SGRPROJ_BORDER_VERT/SGRPROJ_BORDER_HORZ pixels
+// on the sides. A, B, C, D point at logical position (0, 0).
+static void calc_ab_fast(int32_t *A, int32_t *B, const int32_t *C,
+                         const int32_t *D, int width, int height,
+                         int buf_stride, int bit_depth, int sgr_params_idx,
+                         int radius_idx) {
+  const sgr_params_type *const params = &sgr_params[sgr_params_idx];
+  const int r = params->r[radius_idx];
+  const int n = (2 * r + 1) * (2 * r + 1);
+  const __m256i s = _mm256_set1_epi32(params->s[radius_idx]);
+  // one_over_n[n-1] is 2^12/n, so easily fits in an int16
+  const __m256i one_over_n = _mm256_set1_epi32(one_by_x[n - 1]);
+
+  const __m256i rnd_z = round_for_shift(SGRPROJ_MTABLE_BITS);
+  const __m256i rnd_res = round_for_shift(SGRPROJ_RECIP_BITS);
+
+  // Set up masks
+  const __m128i ones32 = _mm_set_epi32(0, 0, 0xffffffff, 0xffffffff);
+  __m256i mask[8];
+  for (int idx = 0; idx < 8; idx++) {
+    const __m128i shift = _mm_cvtsi32_si128(8 * (8 - idx));
+    mask[idx] = _mm256_cvtepi8_epi32(_mm_srl_epi64(ones32, shift));
+  }
+
+  for (int i = -1; i < height + 1; i += 2) {
+    for (int j = -1; j < width + 1; j += 8) {
+      const int32_t *Cij = C + i * buf_stride + j;
+      const int32_t *Dij = D + i * buf_stride + j;
+
+      __m256i sum1 = boxsum_from_ii(Dij, buf_stride, r);
+      __m256i sum2 = boxsum_from_ii(Cij, buf_stride, r);
+
+      // When width + 2 isn't a multiple of 8, sum1 and sum2 will contain
+      // some uninitialised data in their upper words. We use a mask to
+      // ensure that these bits are set to 0.
+      int idx = AOMMIN(8, width + 1 - j);
+      assert(idx >= 1);
+
+      if (idx < 8) {
+        sum1 = _mm256_and_si256(mask[idx], sum1);
+        sum2 = _mm256_and_si256(mask[idx], sum2);
+      }
+
+      const __m256i p = compute_p(sum1, sum2, bit_depth, n);
+
+      const __m256i z = _mm256_min_epi32(
+          _mm256_srli_epi32(_mm256_add_epi32(_mm256_mullo_epi32(p, s), rnd_z),
+                            SGRPROJ_MTABLE_BITS),
+          _mm256_set1_epi32(255));
+
+      const __m256i a_res = _mm256_i32gather_epi32(x_by_xplus1, z, 4);
+
+      yy_storeu_256(A + i * buf_stride + j, a_res);
+
+      const __m256i a_complement =
+          _mm256_sub_epi32(_mm256_set1_epi32(SGRPROJ_SGR), a_res);
+
+      // sum1 might have lanes greater than 2^15, so we can't use madd to do
+      // multiplication involving sum1. However, a_complement and one_over_n
+      // are both less than 256, so we can multiply them first.
+      const __m256i a_comp_over_n = _mm256_madd_epi16(a_complement, one_over_n);
+      const __m256i b_int = _mm256_mullo_epi32(a_comp_over_n, sum1);
+      const __m256i b_res = _mm256_srli_epi32(_mm256_add_epi32(b_int, rnd_res),
+                                              SGRPROJ_RECIP_BITS);
+
+      yy_storeu_256(B + i * buf_stride + j, b_res);
+    }
+  }
+}
+
+// Calculate 8 values of the "cross sum" starting at buf.
+//
+// Pixels are indexed like this:
+// xtl  xt   xtr
+//  -   buf   -
+// xbl  xb   xbr
+//
+// Pixels are weighted like this:
+//  5    6    5
+//  0    0    0
+//  5    6    5
+//
+// fives = xtl + xtr + xbl + xbr
+// sixes = xt + xb
+// cross_sum = 6 * sixes + 5 * fives
+//           = 5 * (fives + sixes) - sixes
+//           = (fives + sixes) << 2 + (fives + sixes) + sixes
+static INLINE __m256i cross_sum_fast_even_row(const int32_t *buf, int stride) {
+  const __m256i xtl = yy_loadu_256(buf - 1 - stride);
+  const __m256i xt = yy_loadu_256(buf - stride);
+  const __m256i xtr = yy_loadu_256(buf + 1 - stride);
+  const __m256i xbl = yy_loadu_256(buf - 1 + stride);
+  const __m256i xb = yy_loadu_256(buf + stride);
+  const __m256i xbr = yy_loadu_256(buf + 1 + stride);
+
+  const __m256i fives =
+      _mm256_add_epi32(xtl, _mm256_add_epi32(xtr, _mm256_add_epi32(xbr, xbl)));
+  const __m256i sixes = _mm256_add_epi32(xt, xb);
+  const __m256i fives_plus_sixes = _mm256_add_epi32(fives, sixes);
+
+  return _mm256_add_epi32(
+      _mm256_add_epi32(_mm256_slli_epi32(fives_plus_sixes, 2),
+                       fives_plus_sixes),
+      sixes);
+}
+
+// Calculate 8 values of the "cross sum" starting at buf.
+//
+// Pixels are indexed like this:
+// xl    x   xr
+//
+// Pixels are weighted like this:
+//  5    6    5
+//
+// buf points to x
+//
+// fives = xl + xr
+// sixes = x
+// cross_sum = 5 * fives + 6 * sixes
+//           = 4 * (fives + sixes) + (fives + sixes) + sixes
+//           = (fives + sixes) << 2 + (fives + sixes) + sixes
+static INLINE __m256i cross_sum_fast_odd_row(const int32_t *buf) {
+  const __m256i xl = yy_loadu_256(buf - 1);
+  const __m256i x = yy_loadu_256(buf);
+  const __m256i xr = yy_loadu_256(buf + 1);
+
+  const __m256i fives = _mm256_add_epi32(xl, xr);
+  const __m256i sixes = x;
+
+  const __m256i fives_plus_sixes = _mm256_add_epi32(fives, sixes);
+
+  return _mm256_add_epi32(
+      _mm256_add_epi32(_mm256_slli_epi32(fives_plus_sixes, 2),
+                       fives_plus_sixes),
+      sixes);
+}
+
+// The final filter for the self-guided restoration. Computes a
+// weighted average across A, B with "cross sums" (see cross_sum_...
+// implementations above).
+static void final_filter_fast(int32_t *dst, int dst_stride, const int32_t *A,
+                              const int32_t *B, int buf_stride,
+                              const void *dgd8, int dgd_stride, int width,
+                              int height, int highbd) {
+  const int nb0 = 5;
+  const int nb1 = 4;
+
+  const __m256i rounding0 =
+      round_for_shift(SGRPROJ_SGR_BITS + nb0 - SGRPROJ_RST_BITS);
+  const __m256i rounding1 =
+      round_for_shift(SGRPROJ_SGR_BITS + nb1 - SGRPROJ_RST_BITS);
+
+  const uint8_t *dgd_real =
+      highbd ? (const uint8_t *)CONVERT_TO_SHORTPTR(dgd8) : dgd8;
+
+  for (int i = 0; i < height; ++i) {
+    if (!(i & 1)) {  // even row
+      for (int j = 0; j < width; j += 8) {
+        const __m256i a =
+            cross_sum_fast_even_row(A + i * buf_stride + j, buf_stride);
+        const __m256i b =
+            cross_sum_fast_even_row(B + i * buf_stride + j, buf_stride);
+
+        const __m128i raw =
+            xx_loadu_128(dgd_real + ((i * dgd_stride + j) << highbd));
+        const __m256i src =
+            highbd ? _mm256_cvtepu16_epi32(raw) : _mm256_cvtepu8_epi32(raw);
+
+        __m256i v = _mm256_add_epi32(_mm256_madd_epi16(a, src), b);
+        __m256i w =
+            _mm256_srai_epi32(_mm256_add_epi32(v, rounding0),
+                              SGRPROJ_SGR_BITS + nb0 - SGRPROJ_RST_BITS);
+
+        yy_storeu_256(dst + i * dst_stride + j, w);
+      }
+    } else {  // odd row
+      for (int j = 0; j < width; j += 8) {
+        const __m256i a = cross_sum_fast_odd_row(A + i * buf_stride + j);
+        const __m256i b = cross_sum_fast_odd_row(B + i * buf_stride + j);
+
+        const __m128i raw =
+            xx_loadu_128(dgd_real + ((i * dgd_stride + j) << highbd));
+        const __m256i src =
+            highbd ? _mm256_cvtepu16_epi32(raw) : _mm256_cvtepu8_epi32(raw);
+
+        __m256i v = _mm256_add_epi32(_mm256_madd_epi16(a, src), b);
+        __m256i w =
+            _mm256_srai_epi32(_mm256_add_epi32(v, rounding1),
+                              SGRPROJ_SGR_BITS + nb1 - SGRPROJ_RST_BITS);
+
+        yy_storeu_256(dst + i * dst_stride + j, w);
+      }
+    }
+  }
+}
+
+int av1_selfguided_restoration_avx2(const uint8_t *dgd8, int width, int height,
+                                    int dgd_stride, int32_t *flt0,
+                                    int32_t *flt1, int flt_stride,
+                                    int sgr_params_idx, int bit_depth,
+                                    int highbd) {
+  // The ALIGN_POWER_OF_TWO macro here ensures that column 1 of Atl, Btl,
+  // Ctl and Dtl is 32-byte aligned.
+  const int buf_elts = ALIGN_POWER_OF_TWO(RESTORATION_PROC_UNIT_PELS, 3);
+
+  int32_t *buf = aom_memalign(
+      32, 4 * sizeof(*buf) * ALIGN_POWER_OF_TWO(RESTORATION_PROC_UNIT_PELS, 3));
+  if (!buf) return -1;
+
+  const int width_ext = width + 2 * SGRPROJ_BORDER_HORZ;
+  const int height_ext = height + 2 * SGRPROJ_BORDER_VERT;
+
+  // Adjusting the stride of A and B here appears to avoid bad cache effects,
+  // leading to a significant speed improvement.
+  // We also align the stride to a multiple of 32 bytes for efficiency.
+  int buf_stride = ALIGN_POWER_OF_TWO(width_ext + 16, 3);
+
+  // The "tl" pointers point at the top-left of the initialised data for the
+  // array.
+  int32_t *Atl = buf + 0 * buf_elts + 7;
+  int32_t *Btl = buf + 1 * buf_elts + 7;
+  int32_t *Ctl = buf + 2 * buf_elts + 7;
+  int32_t *Dtl = buf + 3 * buf_elts + 7;
+
+  // The "0" pointers are (- SGRPROJ_BORDER_VERT, -SGRPROJ_BORDER_HORZ). Note
+  // there's a zero row and column in A, B (integral images), so we move down
+  // and right one for them.
+  const int buf_diag_border =
+      SGRPROJ_BORDER_HORZ + buf_stride * SGRPROJ_BORDER_VERT;
+
+  int32_t *A0 = Atl + 1 + buf_stride;
+  int32_t *B0 = Btl + 1 + buf_stride;
+  int32_t *C0 = Ctl + 1 + buf_stride;
+  int32_t *D0 = Dtl + 1 + buf_stride;
+
+  // Finally, A, B, C, D point at position (0, 0).
+  int32_t *A = A0 + buf_diag_border;
+  int32_t *B = B0 + buf_diag_border;
+  int32_t *C = C0 + buf_diag_border;
+  int32_t *D = D0 + buf_diag_border;
+
+  const int dgd_diag_border =
+      SGRPROJ_BORDER_HORZ + dgd_stride * SGRPROJ_BORDER_VERT;
+  const uint8_t *dgd0 = dgd8 - dgd_diag_border;
+
+  // Generate integral images from the input. C will contain sums of squares; D
+  // will contain just sums
+  if (highbd)
+    integral_images_highbd(CONVERT_TO_SHORTPTR(dgd0), dgd_stride, width_ext,
+                           height_ext, Ctl, Dtl, buf_stride);
+  else
+    integral_images(dgd0, dgd_stride, width_ext, height_ext, Ctl, Dtl,
+                    buf_stride);
+
+  const sgr_params_type *const params = &sgr_params[sgr_params_idx];
+  // Write to flt0 and flt1
+  // If params->r == 0 we skip the corresponding filter. We only allow one of
+  // the radii to be 0, as having both equal to 0 would be equivalent to
+  // skipping SGR entirely.
+  assert(!(params->r[0] == 0 && params->r[1] == 0));
+  assert(params->r[0] < AOMMIN(SGRPROJ_BORDER_VERT, SGRPROJ_BORDER_HORZ));
+  assert(params->r[1] < AOMMIN(SGRPROJ_BORDER_VERT, SGRPROJ_BORDER_HORZ));
+
+  if (params->r[0] > 0) {
+    calc_ab_fast(A, B, C, D, width, height, buf_stride, bit_depth,
+                 sgr_params_idx, 0);
+    final_filter_fast(flt0, flt_stride, A, B, buf_stride, dgd8, dgd_stride,
+                      width, height, highbd);
+  }
+
+  if (params->r[1] > 0) {
+    calc_ab(A, B, C, D, width, height, buf_stride, bit_depth, sgr_params_idx,
+            1);
+    final_filter(flt1, flt_stride, A, B, buf_stride, dgd8, dgd_stride, width,
+                 height, highbd);
+  }
+  aom_free(buf);
+  return 0;
+}
+
+void apply_selfguided_restoration_avx2(const uint8_t *dat8, int width,
+                                       int height, int stride, int eps,
+                                       const int *xqd, uint8_t *dst8,
+                                       int dst_stride, int32_t *tmpbuf,
+                                       int bit_depth, int highbd) {
+  int32_t *flt0 = tmpbuf;
+  int32_t *flt1 = flt0 + RESTORATION_UNITPELS_MAX;
+  assert(width * height <= RESTORATION_UNITPELS_MAX);
+  const int ret = av1_selfguided_restoration_avx2(
+      dat8, width, height, stride, flt0, flt1, width, eps, bit_depth, highbd);
+  (void)ret;
+  assert(!ret);
+  const sgr_params_type *const params = &sgr_params[eps];
+  int xq[2];
+  decode_xq(xqd, xq, params);
+
+  __m256i xq0 = _mm256_set1_epi32(xq[0]);
+  __m256i xq1 = _mm256_set1_epi32(xq[1]);
+
+  for (int i = 0; i < height; ++i) {
+    // Calculate output in batches of 16 pixels
+    for (int j = 0; j < width; j += 16) {
+      const int k = i * width + j;
+      const int m = i * dst_stride + j;
+
+      const uint8_t *dat8ij = dat8 + i * stride + j;
+      __m256i ep_0, ep_1;
+      __m128i src_0, src_1;
+      if (highbd) {
+        src_0 = xx_loadu_128(CONVERT_TO_SHORTPTR(dat8ij));
+        src_1 = xx_loadu_128(CONVERT_TO_SHORTPTR(dat8ij + 8));
+        ep_0 = _mm256_cvtepu16_epi32(src_0);
+        ep_1 = _mm256_cvtepu16_epi32(src_1);
+      } else {
+        src_0 = xx_loadu_128(dat8ij);
+        ep_0 = _mm256_cvtepu8_epi32(src_0);
+        ep_1 = _mm256_cvtepu8_epi32(_mm_srli_si128(src_0, 8));
+      }
+
+      const __m256i u_0 = _mm256_slli_epi32(ep_0, SGRPROJ_RST_BITS);
+      const __m256i u_1 = _mm256_slli_epi32(ep_1, SGRPROJ_RST_BITS);
+
+      __m256i v_0 = _mm256_slli_epi32(u_0, SGRPROJ_PRJ_BITS);
+      __m256i v_1 = _mm256_slli_epi32(u_1, SGRPROJ_PRJ_BITS);
+
+      if (params->r[0] > 0) {
+        const __m256i f1_0 = _mm256_sub_epi32(yy_loadu_256(&flt0[k]), u_0);
+        v_0 = _mm256_add_epi32(v_0, _mm256_mullo_epi32(xq0, f1_0));
+
+        const __m256i f1_1 = _mm256_sub_epi32(yy_loadu_256(&flt0[k + 8]), u_1);
+        v_1 = _mm256_add_epi32(v_1, _mm256_mullo_epi32(xq0, f1_1));
+      }
+
+      if (params->r[1] > 0) {
+        const __m256i f2_0 = _mm256_sub_epi32(yy_loadu_256(&flt1[k]), u_0);
+        v_0 = _mm256_add_epi32(v_0, _mm256_mullo_epi32(xq1, f2_0));
+
+        const __m256i f2_1 = _mm256_sub_epi32(yy_loadu_256(&flt1[k + 8]), u_1);
+        v_1 = _mm256_add_epi32(v_1, _mm256_mullo_epi32(xq1, f2_1));
+      }
+
+      const __m256i rounding =
+          round_for_shift(SGRPROJ_PRJ_BITS + SGRPROJ_RST_BITS);
+      const __m256i w_0 = _mm256_srai_epi32(
+          _mm256_add_epi32(v_0, rounding), SGRPROJ_PRJ_BITS + SGRPROJ_RST_BITS);
+      const __m256i w_1 = _mm256_srai_epi32(
+          _mm256_add_epi32(v_1, rounding), SGRPROJ_PRJ_BITS + SGRPROJ_RST_BITS);
+
+      if (highbd) {
+        // Pack into 16 bits and clamp to [0, 2^bit_depth)
+        // Note that packing into 16 bits messes up the order of the bits,
+        // so we use a permute function to correct this
+        const __m256i tmp = _mm256_packus_epi32(w_0, w_1);
+        const __m256i tmp2 = _mm256_permute4x64_epi64(tmp, 0xd8);
+        const __m256i max = _mm256_set1_epi16((1 << bit_depth) - 1);
+        const __m256i res = _mm256_min_epi16(tmp2, max);
+        yy_storeu_256(CONVERT_TO_SHORTPTR(dst8 + m), res);
+      } else {
+        // Pack into 8 bits and clamp to [0, 256)
+        // Note that each pack messes up the order of the bits,
+        // so we use a permute function to correct this
+        const __m256i tmp = _mm256_packs_epi32(w_0, w_1);
+        const __m256i tmp2 = _mm256_permute4x64_epi64(tmp, 0xd8);
+        const __m256i res =
+            _mm256_packus_epi16(tmp2, tmp2 /* "don't care" value */);
+        const __m128i res2 =
+            _mm256_castsi256_si128(_mm256_permute4x64_epi64(res, 0xd8));
+        xx_storeu_128(dst8 + m, res2);
+      }
+    }
+  }
+}
diff --git a/third_party/aom/av1/common/x86/selfguided_sse4.c b/third_party/aom/av1/common/x86/selfguided_sse4.c
new file mode 100644
index 000000000..ea3f6d942
--- /dev/null
+++ b/third_party/aom/av1/common/x86/selfguided_sse4.c
@@ -0,0 +1,660 @@
+/*
+ * 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 "config/aom_config.h"
+#include "config/av1_rtcd.h"
+
+#include "av1/common/restoration.h"
+#include "aom_dsp/x86/synonyms.h"
+
+// Load 4 bytes from the possibly-misaligned pointer p, extend each byte to
+// 32-bit precision and return them in an SSE register.
+static __m128i xx_load_extend_8_32(const void *p) {
+  return _mm_cvtepu8_epi32(xx_loadl_32(p));
+}
+
+// Load 4 halfwords from the possibly-misaligned pointer p, extend each
+// halfword to 32-bit precision and return them in an SSE register.
+static __m128i xx_load_extend_16_32(const void *p) {
+  return _mm_cvtepu16_epi32(xx_loadl_64(p));
+}
+
+// Compute the scan of an SSE register holding 4 32-bit integers. If the
+// register holds x0..x3 then the scan will hold x0, x0+x1, x0+x1+x2,
+// x0+x1+x2+x3
+static __m128i scan_32(__m128i x) {
+  const __m128i x01 = _mm_add_epi32(x, _mm_slli_si128(x, 4));
+  return _mm_add_epi32(x01, _mm_slli_si128(x01, 8));
+}
+
+// Compute two integral images from src. B sums elements; A sums their
+// squares. The images are offset by one pixel, so will have width and height
+// equal to width + 1, height + 1 and the first row and column will be zero.
+//
+// A+1 and B+1 should be aligned to 16 bytes. buf_stride should be a multiple
+// of 4.
+static void integral_images(const uint8_t *src, int src_stride, int width,
+                            int height, int32_t *A, int32_t *B,
+                            int buf_stride) {
+  // Write out the zero top row
+  memset(A, 0, sizeof(*A) * (width + 1));
+  memset(B, 0, sizeof(*B) * (width + 1));
+
+  const __m128i zero = _mm_setzero_si128();
+  for (int i = 0; i < height; ++i) {
+    // Zero the left column.
+    A[(i + 1) * buf_stride] = B[(i + 1) * buf_stride] = 0;
+
+    // ldiff is the difference H - D where H is the output sample immediately
+    // to the left and D is the output sample above it. These are scalars,
+    // replicated across the four lanes.
+    __m128i ldiff1 = zero, ldiff2 = zero;
+    for (int j = 0; j < width; j += 4) {
+      const int ABj = 1 + j;
+
+      const __m128i above1 = xx_load_128(B + ABj + i * buf_stride);
+      const __m128i above2 = xx_load_128(A + ABj + i * buf_stride);
+
+      const __m128i x1 = xx_load_extend_8_32(src + j + i * src_stride);
+      const __m128i x2 = _mm_madd_epi16(x1, x1);
+
+      const __m128i sc1 = scan_32(x1);
+      const __m128i sc2 = scan_32(x2);
+
+      const __m128i row1 = _mm_add_epi32(_mm_add_epi32(sc1, above1), ldiff1);
+      const __m128i row2 = _mm_add_epi32(_mm_add_epi32(sc2, above2), ldiff2);
+
+      xx_store_128(B + ABj + (i + 1) * buf_stride, row1);
+      xx_store_128(A + ABj + (i + 1) * buf_stride, row2);
+
+      // Calculate the new H - D.
+      ldiff1 = _mm_shuffle_epi32(_mm_sub_epi32(row1, above1), 0xff);
+      ldiff2 = _mm_shuffle_epi32(_mm_sub_epi32(row2, above2), 0xff);
+    }
+  }
+}
+
+// Compute two integral images from src. B sums elements; A sums their squares
+//
+// A and B should be aligned to 16 bytes. buf_stride should be a multiple of 4.
+static void integral_images_highbd(const uint16_t *src, int src_stride,
+                                   int width, int height, int32_t *A,
+                                   int32_t *B, int buf_stride) {
+  // Write out the zero top row
+  memset(A, 0, sizeof(*A) * (width + 1));
+  memset(B, 0, sizeof(*B) * (width + 1));
+
+  const __m128i zero = _mm_setzero_si128();
+  for (int i = 0; i < height; ++i) {
+    // Zero the left column.
+    A[(i + 1) * buf_stride] = B[(i + 1) * buf_stride] = 0;
+
+    // ldiff is the difference H - D where H is the output sample immediately
+    // to the left and D is the output sample above it. These are scalars,
+    // replicated across the four lanes.
+    __m128i ldiff1 = zero, ldiff2 = zero;
+    for (int j = 0; j < width; j += 4) {
+      const int ABj = 1 + j;
+
+      const __m128i above1 = xx_load_128(B + ABj + i * buf_stride);
+      const __m128i above2 = xx_load_128(A + ABj + i * buf_stride);
+
+      const __m128i x1 = xx_load_extend_16_32(src + j + i * src_stride);
+      const __m128i x2 = _mm_madd_epi16(x1, x1);
+
+      const __m128i sc1 = scan_32(x1);
+      const __m128i sc2 = scan_32(x2);
+
+      const __m128i row1 = _mm_add_epi32(_mm_add_epi32(sc1, above1), ldiff1);
+      const __m128i row2 = _mm_add_epi32(_mm_add_epi32(sc2, above2), ldiff2);
+
+      xx_store_128(B + ABj + (i + 1) * buf_stride, row1);
+      xx_store_128(A + ABj + (i + 1) * buf_stride, row2);
+
+      // Calculate the new H - D.
+      ldiff1 = _mm_shuffle_epi32(_mm_sub_epi32(row1, above1), 0xff);
+      ldiff2 = _mm_shuffle_epi32(_mm_sub_epi32(row2, above2), 0xff);
+    }
+  }
+}
+
+// Compute 4 values of boxsum from the given integral image. ii should point
+// at the middle of the box (for the first value). r is the box radius.
+static INLINE __m128i boxsum_from_ii(const int32_t *ii, int stride, int r) {
+  const __m128i tl = xx_loadu_128(ii - (r + 1) - (r + 1) * stride);
+  const __m128i tr = xx_loadu_128(ii + (r + 0) - (r + 1) * stride);
+  const __m128i bl = xx_loadu_128(ii - (r + 1) + r * stride);
+  const __m128i br = xx_loadu_128(ii + (r + 0) + r * stride);
+  const __m128i u = _mm_sub_epi32(tr, tl);
+  const __m128i v = _mm_sub_epi32(br, bl);
+  return _mm_sub_epi32(v, u);
+}
+
+static __m128i round_for_shift(unsigned shift) {
+  return _mm_set1_epi32((1 << shift) >> 1);
+}
+
+static __m128i compute_p(__m128i sum1, __m128i sum2, int bit_depth, int n) {
+  __m128i an, bb;
+  if (bit_depth > 8) {
+    const __m128i rounding_a = round_for_shift(2 * (bit_depth - 8));
+    const __m128i rounding_b = round_for_shift(bit_depth - 8);
+    const __m128i shift_a = _mm_cvtsi32_si128(2 * (bit_depth - 8));
+    const __m128i shift_b = _mm_cvtsi32_si128(bit_depth - 8);
+    const __m128i a = _mm_srl_epi32(_mm_add_epi32(sum2, rounding_a), shift_a);
+    const __m128i b = _mm_srl_epi32(_mm_add_epi32(sum1, rounding_b), shift_b);
+    // b < 2^14, so we can use a 16-bit madd rather than a 32-bit
+    // mullo to square it
+    bb = _mm_madd_epi16(b, b);
+    an = _mm_max_epi32(_mm_mullo_epi32(a, _mm_set1_epi32(n)), bb);
+  } else {
+    bb = _mm_madd_epi16(sum1, sum1);
+    an = _mm_mullo_epi32(sum2, _mm_set1_epi32(n));
+  }
+  return _mm_sub_epi32(an, bb);
+}
+
+// Assumes that C, D are integral images for the original buffer which has been
+// extended to have a padding of SGRPROJ_BORDER_VERT/SGRPROJ_BORDER_HORZ pixels
+// on the sides. A, B, C, D point at logical position (0, 0).
+static void calc_ab(int32_t *A, int32_t *B, const int32_t *C, const int32_t *D,
+                    int width, int height, int buf_stride, int bit_depth,
+                    int sgr_params_idx, int radius_idx) {
+  const sgr_params_type *const params = &sgr_params[sgr_params_idx];
+  const int r = params->r[radius_idx];
+  const int n = (2 * r + 1) * (2 * r + 1);
+  const __m128i s = _mm_set1_epi32(params->s[radius_idx]);
+  // one_over_n[n-1] is 2^12/n, so easily fits in an int16
+  const __m128i one_over_n = _mm_set1_epi32(one_by_x[n - 1]);
+
+  const __m128i rnd_z = round_for_shift(SGRPROJ_MTABLE_BITS);
+  const __m128i rnd_res = round_for_shift(SGRPROJ_RECIP_BITS);
+
+  // Set up masks
+  const __m128i ones32 = _mm_set_epi32(0, 0, 0xffffffff, 0xffffffff);
+  __m128i mask[4];
+  for (int idx = 0; idx < 4; idx++) {
+    const __m128i shift = _mm_cvtsi32_si128(8 * (4 - idx));
+    mask[idx] = _mm_cvtepi8_epi32(_mm_srl_epi64(ones32, shift));
+  }
+
+  for (int i = -1; i < height + 1; ++i) {
+    for (int j = -1; j < width + 1; j += 4) {
+      const int32_t *Cij = C + i * buf_stride + j;
+      const int32_t *Dij = D + i * buf_stride + j;
+
+      __m128i sum1 = boxsum_from_ii(Dij, buf_stride, r);
+      __m128i sum2 = boxsum_from_ii(Cij, buf_stride, r);
+
+      // When width + 2 isn't a multiple of 4, sum1 and sum2 will contain
+      // some uninitialised data in their upper words. We use a mask to
+      // ensure that these bits are set to 0.
+      int idx = AOMMIN(4, width + 1 - j);
+      assert(idx >= 1);
+
+      if (idx < 4) {
+        sum1 = _mm_and_si128(mask[idx], sum1);
+        sum2 = _mm_and_si128(mask[idx], sum2);
+      }
+
+      const __m128i p = compute_p(sum1, sum2, bit_depth, n);
+
+      const __m128i z = _mm_min_epi32(
+          _mm_srli_epi32(_mm_add_epi32(_mm_mullo_epi32(p, s), rnd_z),
+                         SGRPROJ_MTABLE_BITS),
+          _mm_set1_epi32(255));
+
+      // 'Gather' type instructions are not available pre-AVX2, so synthesize a
+      // gather using scalar loads.
+      const __m128i a_res = _mm_set_epi32(x_by_xplus1[_mm_extract_epi32(z, 3)],
+                                          x_by_xplus1[_mm_extract_epi32(z, 2)],
+                                          x_by_xplus1[_mm_extract_epi32(z, 1)],
+                                          x_by_xplus1[_mm_extract_epi32(z, 0)]);
+
+      xx_storeu_128(A + i * buf_stride + j, a_res);
+
+      const __m128i a_complement =
+          _mm_sub_epi32(_mm_set1_epi32(SGRPROJ_SGR), a_res);
+
+      // sum1 might have lanes greater than 2^15, so we can't use madd to do
+      // multiplication involving sum1. However, a_complement and one_over_n
+      // are both less than 256, so we can multiply them first.
+      const __m128i a_comp_over_n = _mm_madd_epi16(a_complement, one_over_n);
+      const __m128i b_int = _mm_mullo_epi32(a_comp_over_n, sum1);
+      const __m128i b_res =
+          _mm_srli_epi32(_mm_add_epi32(b_int, rnd_res), SGRPROJ_RECIP_BITS);
+
+      xx_storeu_128(B + i * buf_stride + j, b_res);
+    }
+  }
+}
+
+// Calculate 4 values of the "cross sum" starting at buf. This is a 3x3 filter
+// where the outer four corners have weight 3 and all other pixels have weight
+// 4.
+//
+// Pixels are indexed like this:
+// xtl  xt   xtr
+// xl    x   xr
+// xbl  xb   xbr
+//
+// buf points to x
+//
+// fours = xl + xt + xr + xb + x
+// threes = xtl + xtr + xbr + xbl
+// cross_sum = 4 * fours + 3 * threes
+//           = 4 * (fours + threes) - threes
+//           = (fours + threes) << 2 - threes
+static INLINE __m128i cross_sum(const int32_t *buf, int stride) {
+  const __m128i xtl = xx_loadu_128(buf - 1 - stride);
+  const __m128i xt = xx_loadu_128(buf - stride);
+  const __m128i xtr = xx_loadu_128(buf + 1 - stride);
+  const __m128i xl = xx_loadu_128(buf - 1);
+  const __m128i x = xx_loadu_128(buf);
+  const __m128i xr = xx_loadu_128(buf + 1);
+  const __m128i xbl = xx_loadu_128(buf - 1 + stride);
+  const __m128i xb = xx_loadu_128(buf + stride);
+  const __m128i xbr = xx_loadu_128(buf + 1 + stride);
+
+  const __m128i fours = _mm_add_epi32(
+      xl, _mm_add_epi32(xt, _mm_add_epi32(xr, _mm_add_epi32(xb, x))));
+  const __m128i threes =
+      _mm_add_epi32(xtl, _mm_add_epi32(xtr, _mm_add_epi32(xbr, xbl)));
+
+  return _mm_sub_epi32(_mm_slli_epi32(_mm_add_epi32(fours, threes), 2), threes);
+}
+
+// The final filter for self-guided restoration. Computes a weighted average
+// across A, B with "cross sums" (see cross_sum implementation above).
+static void final_filter(int32_t *dst, int dst_stride, const int32_t *A,
+                         const int32_t *B, int buf_stride, const void *dgd8,
+                         int dgd_stride, int width, int height, int highbd) {
+  const int nb = 5;
+  const __m128i rounding =
+      round_for_shift(SGRPROJ_SGR_BITS + nb - SGRPROJ_RST_BITS);
+  const uint8_t *dgd_real =
+      highbd ? (const uint8_t *)CONVERT_TO_SHORTPTR(dgd8) : dgd8;
+
+  for (int i = 0; i < height; ++i) {
+    for (int j = 0; j < width; j += 4) {
+      const __m128i a = cross_sum(A + i * buf_stride + j, buf_stride);
+      const __m128i b = cross_sum(B + i * buf_stride + j, buf_stride);
+      const __m128i raw =
+          xx_loadl_64(dgd_real + ((i * dgd_stride + j) << highbd));
+      const __m128i src =
+          highbd ? _mm_cvtepu16_epi32(raw) : _mm_cvtepu8_epi32(raw);
+
+      __m128i v = _mm_add_epi32(_mm_madd_epi16(a, src), b);
+      __m128i w = _mm_srai_epi32(_mm_add_epi32(v, rounding),
+                                 SGRPROJ_SGR_BITS + nb - SGRPROJ_RST_BITS);
+
+      xx_storeu_128(dst + i * dst_stride + j, w);
+    }
+  }
+}
+
+// Assumes that C, D are integral images for the original buffer which has been
+// extended to have a padding of SGRPROJ_BORDER_VERT/SGRPROJ_BORDER_HORZ pixels
+// on the sides. A, B, C, D point at logical position (0, 0).
+static void calc_ab_fast(int32_t *A, int32_t *B, const int32_t *C,
+                         const int32_t *D, int width, int height,
+                         int buf_stride, int bit_depth, int sgr_params_idx,
+                         int radius_idx) {
+  const sgr_params_type *const params = &sgr_params[sgr_params_idx];
+  const int r = params->r[radius_idx];
+  const int n = (2 * r + 1) * (2 * r + 1);
+  const __m128i s = _mm_set1_epi32(params->s[radius_idx]);
+  // one_over_n[n-1] is 2^12/n, so easily fits in an int16
+  const __m128i one_over_n = _mm_set1_epi32(one_by_x[n - 1]);
+
+  const __m128i rnd_z = round_for_shift(SGRPROJ_MTABLE_BITS);
+  const __m128i rnd_res = round_for_shift(SGRPROJ_RECIP_BITS);
+
+  // Set up masks
+  const __m128i ones32 = _mm_set_epi32(0, 0, 0xffffffff, 0xffffffff);
+  __m128i mask[4];
+  for (int idx = 0; idx < 4; idx++) {
+    const __m128i shift = _mm_cvtsi32_si128(8 * (4 - idx));
+    mask[idx] = _mm_cvtepi8_epi32(_mm_srl_epi64(ones32, shift));
+  }
+
+  for (int i = -1; i < height + 1; i += 2) {
+    for (int j = -1; j < width + 1; j += 4) {
+      const int32_t *Cij = C + i * buf_stride + j;
+      const int32_t *Dij = D + i * buf_stride + j;
+
+      __m128i sum1 = boxsum_from_ii(Dij, buf_stride, r);
+      __m128i sum2 = boxsum_from_ii(Cij, buf_stride, r);
+
+      // When width + 2 isn't a multiple of 4, sum1 and sum2 will contain
+      // some uninitialised data in their upper words. We use a mask to
+      // ensure that these bits are set to 0.
+      int idx = AOMMIN(4, width + 1 - j);
+      assert(idx >= 1);
+
+      if (idx < 4) {
+        sum1 = _mm_and_si128(mask[idx], sum1);
+        sum2 = _mm_and_si128(mask[idx], sum2);
+      }
+
+      const __m128i p = compute_p(sum1, sum2, bit_depth, n);
+
+      const __m128i z = _mm_min_epi32(
+          _mm_srli_epi32(_mm_add_epi32(_mm_mullo_epi32(p, s), rnd_z),
+                         SGRPROJ_MTABLE_BITS),
+          _mm_set1_epi32(255));
+
+      // 'Gather' type instructions are not available pre-AVX2, so synthesize a
+      // gather using scalar loads.
+      const __m128i a_res = _mm_set_epi32(x_by_xplus1[_mm_extract_epi32(z, 3)],
+                                          x_by_xplus1[_mm_extract_epi32(z, 2)],
+                                          x_by_xplus1[_mm_extract_epi32(z, 1)],
+                                          x_by_xplus1[_mm_extract_epi32(z, 0)]);
+
+      xx_storeu_128(A + i * buf_stride + j, a_res);
+
+      const __m128i a_complement =
+          _mm_sub_epi32(_mm_set1_epi32(SGRPROJ_SGR), a_res);
+
+      // sum1 might have lanes greater than 2^15, so we can't use madd to do
+      // multiplication involving sum1. However, a_complement and one_over_n
+      // are both less than 256, so we can multiply them first.
+      const __m128i a_comp_over_n = _mm_madd_epi16(a_complement, one_over_n);
+      const __m128i b_int = _mm_mullo_epi32(a_comp_over_n, sum1);
+      const __m128i b_res =
+          _mm_srli_epi32(_mm_add_epi32(b_int, rnd_res), SGRPROJ_RECIP_BITS);
+
+      xx_storeu_128(B + i * buf_stride + j, b_res);
+    }
+  }
+}
+
+// Calculate 4 values of the "cross sum" starting at buf.
+//
+// Pixels are indexed like this:
+// xtl  xt   xtr
+//  -   buf   -
+// xbl  xb   xbr
+//
+// Pixels are weighted like this:
+//  5    6    5
+//  0    0    0
+//  5    6    5
+//
+// fives = xtl + xtr + xbl + xbr
+// sixes = xt + xb
+// cross_sum = 6 * sixes + 5 * fives
+//           = 5 * (fives + sixes) - sixes
+//           = (fives + sixes) << 2 + (fives + sixes) + sixes
+static INLINE __m128i cross_sum_fast_even_row(const int32_t *buf, int stride) {
+  const __m128i xtl = xx_loadu_128(buf - 1 - stride);
+  const __m128i xt = xx_loadu_128(buf - stride);
+  const __m128i xtr = xx_loadu_128(buf + 1 - stride);
+  const __m128i xbl = xx_loadu_128(buf - 1 + stride);
+  const __m128i xb = xx_loadu_128(buf + stride);
+  const __m128i xbr = xx_loadu_128(buf + 1 + stride);
+
+  const __m128i fives =
+      _mm_add_epi32(xtl, _mm_add_epi32(xtr, _mm_add_epi32(xbr, xbl)));
+  const __m128i sixes = _mm_add_epi32(xt, xb);
+  const __m128i fives_plus_sixes = _mm_add_epi32(fives, sixes);
+
+  return _mm_add_epi32(
+      _mm_add_epi32(_mm_slli_epi32(fives_plus_sixes, 2), fives_plus_sixes),
+      sixes);
+}
+
+// Calculate 4 values of the "cross sum" starting at buf.
+//
+// Pixels are indexed like this:
+// xl    x   xr
+//
+// Pixels are weighted like this:
+//  5    6    5
+//
+// buf points to x
+//
+// fives = xl + xr
+// sixes = x
+// cross_sum = 5 * fives + 6 * sixes
+//           = 4 * (fives + sixes) + (fives + sixes) + sixes
+//           = (fives + sixes) << 2 + (fives + sixes) + sixes
+static INLINE __m128i cross_sum_fast_odd_row(const int32_t *buf) {
+  const __m128i xl = xx_loadu_128(buf - 1);
+  const __m128i x = xx_loadu_128(buf);
+  const __m128i xr = xx_loadu_128(buf + 1);
+
+  const __m128i fives = _mm_add_epi32(xl, xr);
+  const __m128i sixes = x;
+
+  const __m128i fives_plus_sixes = _mm_add_epi32(fives, sixes);
+
+  return _mm_add_epi32(
+      _mm_add_epi32(_mm_slli_epi32(fives_plus_sixes, 2), fives_plus_sixes),
+      sixes);
+}
+
+// The final filter for the self-guided restoration. Computes a
+// weighted average across A, B with "cross sums" (see cross_sum_...
+// implementations above).
+static void final_filter_fast(int32_t *dst, int dst_stride, const int32_t *A,
+                              const int32_t *B, int buf_stride,
+                              const void *dgd8, int dgd_stride, int width,
+                              int height, int highbd) {
+  const int nb0 = 5;
+  const int nb1 = 4;
+
+  const __m128i rounding0 =
+      round_for_shift(SGRPROJ_SGR_BITS + nb0 - SGRPROJ_RST_BITS);
+  const __m128i rounding1 =
+      round_for_shift(SGRPROJ_SGR_BITS + nb1 - SGRPROJ_RST_BITS);
+
+  const uint8_t *dgd_real =
+      highbd ? (const uint8_t *)CONVERT_TO_SHORTPTR(dgd8) : dgd8;
+
+  for (int i = 0; i < height; ++i) {
+    if (!(i & 1)) {  // even row
+      for (int j = 0; j < width; j += 4) {
+        const __m128i a =
+            cross_sum_fast_even_row(A + i * buf_stride + j, buf_stride);
+        const __m128i b =
+            cross_sum_fast_even_row(B + i * buf_stride + j, buf_stride);
+        const __m128i raw =
+            xx_loadl_64(dgd_real + ((i * dgd_stride + j) << highbd));
+        const __m128i src =
+            highbd ? _mm_cvtepu16_epi32(raw) : _mm_cvtepu8_epi32(raw);
+
+        __m128i v = _mm_add_epi32(_mm_madd_epi16(a, src), b);
+        __m128i w = _mm_srai_epi32(_mm_add_epi32(v, rounding0),
+                                   SGRPROJ_SGR_BITS + nb0 - SGRPROJ_RST_BITS);
+
+        xx_storeu_128(dst + i * dst_stride + j, w);
+      }
+    } else {  // odd row
+      for (int j = 0; j < width; j += 4) {
+        const __m128i a = cross_sum_fast_odd_row(A + i * buf_stride + j);
+        const __m128i b = cross_sum_fast_odd_row(B + i * buf_stride + j);
+        const __m128i raw =
+            xx_loadl_64(dgd_real + ((i * dgd_stride + j) << highbd));
+        const __m128i src =
+            highbd ? _mm_cvtepu16_epi32(raw) : _mm_cvtepu8_epi32(raw);
+
+        __m128i v = _mm_add_epi32(_mm_madd_epi16(a, src), b);
+        __m128i w = _mm_srai_epi32(_mm_add_epi32(v, rounding1),
+                                   SGRPROJ_SGR_BITS + nb1 - SGRPROJ_RST_BITS);
+
+        xx_storeu_128(dst + i * dst_stride + j, w);
+      }
+    }
+  }
+}
+
+int av1_selfguided_restoration_sse4_1(const uint8_t *dgd8, int width,
+                                      int height, int dgd_stride, int32_t *flt0,
+                                      int32_t *flt1, int flt_stride,
+                                      int sgr_params_idx, int bit_depth,
+                                      int highbd) {
+  int32_t *buf = (int32_t *)aom_memalign(
+      16, 4 * sizeof(*buf) * RESTORATION_PROC_UNIT_PELS);
+  if (!buf) return -1;
+  memset(buf, 0, 4 * sizeof(*buf) * RESTORATION_PROC_UNIT_PELS);
+
+  const int width_ext = width + 2 * SGRPROJ_BORDER_HORZ;
+  const int height_ext = height + 2 * SGRPROJ_BORDER_VERT;
+
+  // Adjusting the stride of A and B here appears to avoid bad cache effects,
+  // leading to a significant speed improvement.
+  // We also align the stride to a multiple of 16 bytes for efficiency.
+  int buf_stride = ((width_ext + 3) & ~3) + 16;
+
+  // The "tl" pointers point at the top-left of the initialised data for the
+  // array. Adding 3 here ensures that column 1 is 16-byte aligned.
+  int32_t *Atl = buf + 0 * RESTORATION_PROC_UNIT_PELS + 3;
+  int32_t *Btl = buf + 1 * RESTORATION_PROC_UNIT_PELS + 3;
+  int32_t *Ctl = buf + 2 * RESTORATION_PROC_UNIT_PELS + 3;
+  int32_t *Dtl = buf + 3 * RESTORATION_PROC_UNIT_PELS + 3;
+
+  // The "0" pointers are (- SGRPROJ_BORDER_VERT, -SGRPROJ_BORDER_HORZ). Note
+  // there's a zero row and column in A, B (integral images), so we move down
+  // and right one for them.
+  const int buf_diag_border =
+      SGRPROJ_BORDER_HORZ + buf_stride * SGRPROJ_BORDER_VERT;
+
+  int32_t *A0 = Atl + 1 + buf_stride;
+  int32_t *B0 = Btl + 1 + buf_stride;
+  int32_t *C0 = Ctl + 1 + buf_stride;
+  int32_t *D0 = Dtl + 1 + buf_stride;
+
+  // Finally, A, B, C, D point at position (0, 0).
+  int32_t *A = A0 + buf_diag_border;
+  int32_t *B = B0 + buf_diag_border;
+  int32_t *C = C0 + buf_diag_border;
+  int32_t *D = D0 + buf_diag_border;
+
+  const int dgd_diag_border =
+      SGRPROJ_BORDER_HORZ + dgd_stride * SGRPROJ_BORDER_VERT;
+  const uint8_t *dgd0 = dgd8 - dgd_diag_border;
+
+  // Generate integral images from the input. C will contain sums of squares; D
+  // will contain just sums
+  if (highbd)
+    integral_images_highbd(CONVERT_TO_SHORTPTR(dgd0), dgd_stride, width_ext,
+                           height_ext, Ctl, Dtl, buf_stride);
+  else
+    integral_images(dgd0, dgd_stride, width_ext, height_ext, Ctl, Dtl,
+                    buf_stride);
+
+  const sgr_params_type *const params = &sgr_params[sgr_params_idx];
+  // Write to flt0 and flt1
+  // If params->r == 0 we skip the corresponding filter. We only allow one of
+  // the radii to be 0, as having both equal to 0 would be equivalent to
+  // skipping SGR entirely.
+  assert(!(params->r[0] == 0 && params->r[1] == 0));
+  assert(params->r[0] < AOMMIN(SGRPROJ_BORDER_VERT, SGRPROJ_BORDER_HORZ));
+  assert(params->r[1] < AOMMIN(SGRPROJ_BORDER_VERT, SGRPROJ_BORDER_HORZ));
+
+  if (params->r[0] > 0) {
+    calc_ab_fast(A, B, C, D, width, height, buf_stride, bit_depth,
+                 sgr_params_idx, 0);
+    final_filter_fast(flt0, flt_stride, A, B, buf_stride, dgd8, dgd_stride,
+                      width, height, highbd);
+  }
+
+  if (params->r[1] > 0) {
+    calc_ab(A, B, C, D, width, height, buf_stride, bit_depth, sgr_params_idx,
+            1);
+    final_filter(flt1, flt_stride, A, B, buf_stride, dgd8, dgd_stride, width,
+                 height, highbd);
+  }
+  aom_free(buf);
+  return 0;
+}
+
+void apply_selfguided_restoration_sse4_1(const uint8_t *dat8, int width,
+                                         int height, int stride, int eps,
+                                         const int *xqd, uint8_t *dst8,
+                                         int dst_stride, int32_t *tmpbuf,
+                                         int bit_depth, int highbd) {
+  int32_t *flt0 = tmpbuf;
+  int32_t *flt1 = flt0 + RESTORATION_UNITPELS_MAX;
+  assert(width * height <= RESTORATION_UNITPELS_MAX);
+  const int ret = av1_selfguided_restoration_sse4_1(
+      dat8, width, height, stride, flt0, flt1, width, eps, bit_depth, highbd);
+  (void)ret;
+  assert(!ret);
+  const sgr_params_type *const params = &sgr_params[eps];
+  int xq[2];
+  decode_xq(xqd, xq, params);
+
+  __m128i xq0 = _mm_set1_epi32(xq[0]);
+  __m128i xq1 = _mm_set1_epi32(xq[1]);
+
+  for (int i = 0; i < height; ++i) {
+    // Calculate output in batches of 8 pixels
+    for (int j = 0; j < width; j += 8) {
+      const int k = i * width + j;
+      const int m = i * dst_stride + j;
+
+      const uint8_t *dat8ij = dat8 + i * stride + j;
+      __m128i src;
+      if (highbd) {
+        src = xx_loadu_128(CONVERT_TO_SHORTPTR(dat8ij));
+      } else {
+        src = _mm_cvtepu8_epi16(xx_loadl_64(dat8ij));
+      }
+
+      const __m128i u = _mm_slli_epi16(src, SGRPROJ_RST_BITS);
+      const __m128i u_0 = _mm_cvtepu16_epi32(u);
+      const __m128i u_1 = _mm_cvtepu16_epi32(_mm_srli_si128(u, 8));
+
+      __m128i v_0 = _mm_slli_epi32(u_0, SGRPROJ_PRJ_BITS);
+      __m128i v_1 = _mm_slli_epi32(u_1, SGRPROJ_PRJ_BITS);
+
+      if (params->r[0] > 0) {
+        const __m128i f1_0 = _mm_sub_epi32(xx_loadu_128(&flt0[k]), u_0);
+        v_0 = _mm_add_epi32(v_0, _mm_mullo_epi32(xq0, f1_0));
+
+        const __m128i f1_1 = _mm_sub_epi32(xx_loadu_128(&flt0[k + 4]), u_1);
+        v_1 = _mm_add_epi32(v_1, _mm_mullo_epi32(xq0, f1_1));
+      }
+
+      if (params->r[1] > 0) {
+        const __m128i f2_0 = _mm_sub_epi32(xx_loadu_128(&flt1[k]), u_0);
+        v_0 = _mm_add_epi32(v_0, _mm_mullo_epi32(xq1, f2_0));
+
+        const __m128i f2_1 = _mm_sub_epi32(xx_loadu_128(&flt1[k + 4]), u_1);
+        v_1 = _mm_add_epi32(v_1, _mm_mullo_epi32(xq1, f2_1));
+      }
+
+      const __m128i rounding =
+          round_for_shift(SGRPROJ_PRJ_BITS + SGRPROJ_RST_BITS);
+      const __m128i w_0 = _mm_srai_epi32(_mm_add_epi32(v_0, rounding),
+                                         SGRPROJ_PRJ_BITS + SGRPROJ_RST_BITS);
+      const __m128i w_1 = _mm_srai_epi32(_mm_add_epi32(v_1, rounding),
+                                         SGRPROJ_PRJ_BITS + SGRPROJ_RST_BITS);
+
+      if (highbd) {
+        // Pack into 16 bits and clamp to [0, 2^bit_depth)
+        const __m128i tmp = _mm_packus_epi32(w_0, w_1);
+        const __m128i max = _mm_set1_epi16((1 << bit_depth) - 1);
+        const __m128i res = _mm_min_epi16(tmp, max);
+        xx_storeu_128(CONVERT_TO_SHORTPTR(dst8 + m), res);
+      } else {
+        // Pack into 8 bits and clamp to [0, 256)
+        const __m128i tmp = _mm_packs_epi32(w_0, w_1);
+        const __m128i res = _mm_packus_epi16(tmp, tmp /* "don't care" value */);
+        xx_storel_64(dst8 + m, res);
+      }
+    }
+  }
+}
diff --git a/third_party/aom/av1/common/x86/warp_plane_sse4.c b/third_party/aom/av1/common/x86/warp_plane_sse4.c
new file mode 100644
index 000000000..b810cea2e
--- /dev/null
+++ b/third_party/aom/av1/common/x86/warp_plane_sse4.c
@@ -0,0 +1,942 @@
+/*
+ * 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 <smmintrin.h>
+
+#include "config/av1_rtcd.h"
+
+#include "av1/common/warped_motion.h"
+
+/* This is a modified version of 'warped_filter' from warped_motion.c:
+   * Each coefficient is stored in 8 bits instead of 16 bits
+   * The coefficients are rearranged in the column order 0, 2, 4, 6, 1, 3, 5, 7
+
+     This is done in order to avoid overflow: Since the tap with the largest
+     coefficient could be any of taps 2, 3, 4 or 5, we can't use the summation
+     order ((0 + 1) + (4 + 5)) + ((2 + 3) + (6 + 7)) used in the regular
+     convolve functions.
+
+     Instead, we use the summation order
+     ((0 + 2) + (4 + 6)) + ((1 + 3) + (5 + 7)).
+     The rearrangement of coefficients in this table is so that we can get the
+     coefficients into the correct order more quickly.
+*/
+/* clang-format off */
+DECLARE_ALIGNED(8, static const int8_t,
+                filter_8bit[WARPEDPIXEL_PREC_SHIFTS * 3 + 1][8]) = {
+#if WARPEDPIXEL_PREC_BITS == 6
+  // [-1, 0)
+  { 0, 127,   0, 0,   0,   1, 0, 0}, { 0, 127,   0, 0,  -1,   2, 0, 0},
+  { 1, 127,  -1, 0,  -3,   4, 0, 0}, { 1, 126,  -2, 0,  -4,   6, 1, 0},
+  { 1, 126,  -3, 0,  -5,   8, 1, 0}, { 1, 125,  -4, 0,  -6,  11, 1, 0},
+  { 1, 124,  -4, 0,  -7,  13, 1, 0}, { 2, 123,  -5, 0,  -8,  15, 1, 0},
+  { 2, 122,  -6, 0,  -9,  18, 1, 0}, { 2, 121,  -6, 0, -10,  20, 1, 0},
+  { 2, 120,  -7, 0, -11,  22, 2, 0}, { 2, 119,  -8, 0, -12,  25, 2, 0},
+  { 3, 117,  -8, 0, -13,  27, 2, 0}, { 3, 116,  -9, 0, -13,  29, 2, 0},
+  { 3, 114, -10, 0, -14,  32, 3, 0}, { 3, 113, -10, 0, -15,  35, 2, 0},
+  { 3, 111, -11, 0, -15,  37, 3, 0}, { 3, 109, -11, 0, -16,  40, 3, 0},
+  { 3, 108, -12, 0, -16,  42, 3, 0}, { 4, 106, -13, 0, -17,  45, 3, 0},
+  { 4, 104, -13, 0, -17,  47, 3, 0}, { 4, 102, -14, 0, -17,  50, 3, 0},
+  { 4, 100, -14, 0, -17,  52, 3, 0}, { 4,  98, -15, 0, -18,  55, 4, 0},
+  { 4,  96, -15, 0, -18,  58, 3, 0}, { 4,  94, -16, 0, -18,  60, 4, 0},
+  { 4,  91, -16, 0, -18,  63, 4, 0}, { 4,  89, -16, 0, -18,  65, 4, 0},
+  { 4,  87, -17, 0, -18,  68, 4, 0}, { 4,  85, -17, 0, -18,  70, 4, 0},
+  { 4,  82, -17, 0, -18,  73, 4, 0}, { 4,  80, -17, 0, -18,  75, 4, 0},
+  { 4,  78, -18, 0, -18,  78, 4, 0}, { 4,  75, -18, 0, -17,  80, 4, 0},
+  { 4,  73, -18, 0, -17,  82, 4, 0}, { 4,  70, -18, 0, -17,  85, 4, 0},
+  { 4,  68, -18, 0, -17,  87, 4, 0}, { 4,  65, -18, 0, -16,  89, 4, 0},
+  { 4,  63, -18, 0, -16,  91, 4, 0}, { 4,  60, -18, 0, -16,  94, 4, 0},
+  { 3,  58, -18, 0, -15,  96, 4, 0}, { 4,  55, -18, 0, -15,  98, 4, 0},
+  { 3,  52, -17, 0, -14, 100, 4, 0}, { 3,  50, -17, 0, -14, 102, 4, 0},
+  { 3,  47, -17, 0, -13, 104, 4, 0}, { 3,  45, -17, 0, -13, 106, 4, 0},
+  { 3,  42, -16, 0, -12, 108, 3, 0}, { 3,  40, -16, 0, -11, 109, 3, 0},
+  { 3,  37, -15, 0, -11, 111, 3, 0}, { 2,  35, -15, 0, -10, 113, 3, 0},
+  { 3,  32, -14, 0, -10, 114, 3, 0}, { 2,  29, -13, 0,  -9, 116, 3, 0},
+  { 2,  27, -13, 0,  -8, 117, 3, 0}, { 2,  25, -12, 0,  -8, 119, 2, 0},
+  { 2,  22, -11, 0,  -7, 120, 2, 0}, { 1,  20, -10, 0,  -6, 121, 2, 0},
+  { 1,  18,  -9, 0,  -6, 122, 2, 0}, { 1,  15,  -8, 0,  -5, 123, 2, 0},
+  { 1,  13,  -7, 0,  -4, 124, 1, 0}, { 1,  11,  -6, 0,  -4, 125, 1, 0},
+  { 1,   8,  -5, 0,  -3, 126, 1, 0}, { 1,   6,  -4, 0,  -2, 126, 1, 0},
+  { 0,   4,  -3, 0,  -1, 127, 1, 0}, { 0,   2,  -1, 0,   0, 127, 0, 0},
+  // [0, 1)
+  { 0,   0,   1, 0, 0, 127,   0,  0}, { 0,  -1,   2, 0, 0, 127,   0,  0},
+  { 0,  -3,   4, 1, 1, 127,  -2,  0}, { 0,  -5,   6, 1, 1, 127,  -2,  0},
+  { 0,  -6,   8, 1, 2, 126,  -3,  0}, {-1,  -7,  11, 2, 2, 126,  -4, -1},
+  {-1,  -8,  13, 2, 3, 125,  -5, -1}, {-1, -10,  16, 3, 3, 124,  -6, -1},
+  {-1, -11,  18, 3, 4, 123,  -7, -1}, {-1, -12,  20, 3, 4, 122,  -7, -1},
+  {-1, -13,  23, 3, 4, 121,  -8, -1}, {-2, -14,  25, 4, 5, 120,  -9, -1},
+  {-1, -15,  27, 4, 5, 119, -10, -1}, {-1, -16,  30, 4, 5, 118, -11, -1},
+  {-2, -17,  33, 5, 6, 116, -12, -1}, {-2, -17,  35, 5, 6, 114, -12, -1},
+  {-2, -18,  38, 5, 6, 113, -13, -1}, {-2, -19,  41, 6, 7, 111, -14, -2},
+  {-2, -19,  43, 6, 7, 110, -15, -2}, {-2, -20,  46, 6, 7, 108, -15, -2},
+  {-2, -20,  49, 6, 7, 106, -16, -2}, {-2, -21,  51, 7, 7, 104, -16, -2},
+  {-2, -21,  54, 7, 7, 102, -17, -2}, {-2, -21,  56, 7, 8, 100, -18, -2},
+  {-2, -22,  59, 7, 8,  98, -18, -2}, {-2, -22,  62, 7, 8,  96, -19, -2},
+  {-2, -22,  64, 7, 8,  94, -19, -2}, {-2, -22,  67, 8, 8,  91, -20, -2},
+  {-2, -22,  69, 8, 8,  89, -20, -2}, {-2, -22,  72, 8, 8,  87, -21, -2},
+  {-2, -21,  74, 8, 8,  84, -21, -2}, {-2, -22,  77, 8, 8,  82, -21, -2},
+  {-2, -21,  79, 8, 8,  79, -21, -2}, {-2, -21,  82, 8, 8,  77, -22, -2},
+  {-2, -21,  84, 8, 8,  74, -21, -2}, {-2, -21,  87, 8, 8,  72, -22, -2},
+  {-2, -20,  89, 8, 8,  69, -22, -2}, {-2, -20,  91, 8, 8,  67, -22, -2},
+  {-2, -19,  94, 8, 7,  64, -22, -2}, {-2, -19,  96, 8, 7,  62, -22, -2},
+  {-2, -18,  98, 8, 7,  59, -22, -2}, {-2, -18, 100, 8, 7,  56, -21, -2},
+  {-2, -17, 102, 7, 7,  54, -21, -2}, {-2, -16, 104, 7, 7,  51, -21, -2},
+  {-2, -16, 106, 7, 6,  49, -20, -2}, {-2, -15, 108, 7, 6,  46, -20, -2},
+  {-2, -15, 110, 7, 6,  43, -19, -2}, {-2, -14, 111, 7, 6,  41, -19, -2},
+  {-1, -13, 113, 6, 5,  38, -18, -2}, {-1, -12, 114, 6, 5,  35, -17, -2},
+  {-1, -12, 116, 6, 5,  33, -17, -2}, {-1, -11, 118, 5, 4,  30, -16, -1},
+  {-1, -10, 119, 5, 4,  27, -15, -1}, {-1,  -9, 120, 5, 4,  25, -14, -2},
+  {-1,  -8, 121, 4, 3,  23, -13, -1}, {-1,  -7, 122, 4, 3,  20, -12, -1},
+  {-1,  -7, 123, 4, 3,  18, -11, -1}, {-1,  -6, 124, 3, 3,  16, -10, -1},
+  {-1,  -5, 125, 3, 2,  13,  -8, -1}, {-1,  -4, 126, 2, 2,  11,  -7, -1},
+  { 0,  -3, 126, 2, 1,   8,  -6,  0}, { 0,  -2, 127, 1, 1,   6,  -5,  0},
+  { 0,  -2, 127, 1, 1,   4,  -3,  0}, { 0,   0, 127, 0, 0,   2,  -1,  0},
+  // [1, 2)
+  { 0, 0, 127,   0, 0,   1,   0, 0}, { 0, 0, 127,   0, 0,  -1,   2, 0},
+  { 0, 1, 127,  -1, 0,  -3,   4, 0}, { 0, 1, 126,  -2, 0,  -4,   6, 1},
+  { 0, 1, 126,  -3, 0,  -5,   8, 1}, { 0, 1, 125,  -4, 0,  -6,  11, 1},
+  { 0, 1, 124,  -4, 0,  -7,  13, 1}, { 0, 2, 123,  -5, 0,  -8,  15, 1},
+  { 0, 2, 122,  -6, 0,  -9,  18, 1}, { 0, 2, 121,  -6, 0, -10,  20, 1},
+  { 0, 2, 120,  -7, 0, -11,  22, 2}, { 0, 2, 119,  -8, 0, -12,  25, 2},
+  { 0, 3, 117,  -8, 0, -13,  27, 2}, { 0, 3, 116,  -9, 0, -13,  29, 2},
+  { 0, 3, 114, -10, 0, -14,  32, 3}, { 0, 3, 113, -10, 0, -15,  35, 2},
+  { 0, 3, 111, -11, 0, -15,  37, 3}, { 0, 3, 109, -11, 0, -16,  40, 3},
+  { 0, 3, 108, -12, 0, -16,  42, 3}, { 0, 4, 106, -13, 0, -17,  45, 3},
+  { 0, 4, 104, -13, 0, -17,  47, 3}, { 0, 4, 102, -14, 0, -17,  50, 3},
+  { 0, 4, 100, -14, 0, -17,  52, 3}, { 0, 4,  98, -15, 0, -18,  55, 4},
+  { 0, 4,  96, -15, 0, -18,  58, 3}, { 0, 4,  94, -16, 0, -18,  60, 4},
+  { 0, 4,  91, -16, 0, -18,  63, 4}, { 0, 4,  89, -16, 0, -18,  65, 4},
+  { 0, 4,  87, -17, 0, -18,  68, 4}, { 0, 4,  85, -17, 0, -18,  70, 4},
+  { 0, 4,  82, -17, 0, -18,  73, 4}, { 0, 4,  80, -17, 0, -18,  75, 4},
+  { 0, 4,  78, -18, 0, -18,  78, 4}, { 0, 4,  75, -18, 0, -17,  80, 4},
+  { 0, 4,  73, -18, 0, -17,  82, 4}, { 0, 4,  70, -18, 0, -17,  85, 4},
+  { 0, 4,  68, -18, 0, -17,  87, 4}, { 0, 4,  65, -18, 0, -16,  89, 4},
+  { 0, 4,  63, -18, 0, -16,  91, 4}, { 0, 4,  60, -18, 0, -16,  94, 4},
+  { 0, 3,  58, -18, 0, -15,  96, 4}, { 0, 4,  55, -18, 0, -15,  98, 4},
+  { 0, 3,  52, -17, 0, -14, 100, 4}, { 0, 3,  50, -17, 0, -14, 102, 4},
+  { 0, 3,  47, -17, 0, -13, 104, 4}, { 0, 3,  45, -17, 0, -13, 106, 4},
+  { 0, 3,  42, -16, 0, -12, 108, 3}, { 0, 3,  40, -16, 0, -11, 109, 3},
+  { 0, 3,  37, -15, 0, -11, 111, 3}, { 0, 2,  35, -15, 0, -10, 113, 3},
+  { 0, 3,  32, -14, 0, -10, 114, 3}, { 0, 2,  29, -13, 0,  -9, 116, 3},
+  { 0, 2,  27, -13, 0,  -8, 117, 3}, { 0, 2,  25, -12, 0,  -8, 119, 2},
+  { 0, 2,  22, -11, 0,  -7, 120, 2}, { 0, 1,  20, -10, 0,  -6, 121, 2},
+  { 0, 1,  18,  -9, 0,  -6, 122, 2}, { 0, 1,  15,  -8, 0,  -5, 123, 2},
+  { 0, 1,  13,  -7, 0,  -4, 124, 1}, { 0, 1,  11,  -6, 0,  -4, 125, 1},
+  { 0, 1,   8,  -5, 0,  -3, 126, 1}, { 0, 1,   6,  -4, 0,  -2, 126, 1},
+  { 0, 0,   4,  -3, 0,  -1, 127, 1}, { 0, 0,   2,  -1, 0,   0, 127, 0},
+  // dummy (replicate row index 191)
+  { 0, 0,   2,  -1, 0,   0, 127, 0},
+
+#else
+  // [-1, 0)
+  { 0, 127,   0, 0,   0,   1, 0, 0}, { 1, 127,  -1, 0,  -3,   4, 0, 0},
+  { 1, 126,  -3, 0,  -5,   8, 1, 0}, { 1, 124,  -4, 0,  -7,  13, 1, 0},
+  { 2, 122,  -6, 0,  -9,  18, 1, 0}, { 2, 120,  -7, 0, -11,  22, 2, 0},
+  { 3, 117,  -8, 0, -13,  27, 2, 0}, { 3, 114, -10, 0, -14,  32, 3, 0},
+  { 3, 111, -11, 0, -15,  37, 3, 0}, { 3, 108, -12, 0, -16,  42, 3, 0},
+  { 4, 104, -13, 0, -17,  47, 3, 0}, { 4, 100, -14, 0, -17,  52, 3, 0},
+  { 4,  96, -15, 0, -18,  58, 3, 0}, { 4,  91, -16, 0, -18,  63, 4, 0},
+  { 4,  87, -17, 0, -18,  68, 4, 0}, { 4,  82, -17, 0, -18,  73, 4, 0},
+  { 4,  78, -18, 0, -18,  78, 4, 0}, { 4,  73, -18, 0, -17,  82, 4, 0},
+  { 4,  68, -18, 0, -17,  87, 4, 0}, { 4,  63, -18, 0, -16,  91, 4, 0},
+  { 3,  58, -18, 0, -15,  96, 4, 0}, { 3,  52, -17, 0, -14, 100, 4, 0},
+  { 3,  47, -17, 0, -13, 104, 4, 0}, { 3,  42, -16, 0, -12, 108, 3, 0},
+  { 3,  37, -15, 0, -11, 111, 3, 0}, { 3,  32, -14, 0, -10, 114, 3, 0},
+  { 2,  27, -13, 0,  -8, 117, 3, 0}, { 2,  22, -11, 0,  -7, 120, 2, 0},
+  { 1,  18,  -9, 0,  -6, 122, 2, 0}, { 1,  13,  -7, 0,  -4, 124, 1, 0},
+  { 1,   8,  -5, 0,  -3, 126, 1, 0}, { 0,   4,  -3, 0,  -1, 127, 1, 0},
+  // [0, 1)
+  { 0,   0,   1, 0, 0, 127,   0,  0}, { 0,  -3,   4, 1, 1, 127,  -2,  0},
+  { 0,  -6,   8, 1, 2, 126,  -3,  0}, {-1,  -8,  13, 2, 3, 125,  -5, -1},
+  {-1, -11,  18, 3, 4, 123,  -7, -1}, {-1, -13,  23, 3, 4, 121,  -8, -1},
+  {-1, -15,  27, 4, 5, 119, -10, -1}, {-2, -17,  33, 5, 6, 116, -12, -1},
+  {-2, -18,  38, 5, 6, 113, -13, -1}, {-2, -19,  43, 6, 7, 110, -15, -2},
+  {-2, -20,  49, 6, 7, 106, -16, -2}, {-2, -21,  54, 7, 7, 102, -17, -2},
+  {-2, -22,  59, 7, 8,  98, -18, -2}, {-2, -22,  64, 7, 8,  94, -19, -2},
+  {-2, -22,  69, 8, 8,  89, -20, -2}, {-2, -21,  74, 8, 8,  84, -21, -2},
+  {-2, -21,  79, 8, 8,  79, -21, -2}, {-2, -21,  84, 8, 8,  74, -21, -2},
+  {-2, -20,  89, 8, 8,  69, -22, -2}, {-2, -19,  94, 8, 7,  64, -22, -2},
+  {-2, -18,  98, 8, 7,  59, -22, -2}, {-2, -17, 102, 7, 7,  54, -21, -2},
+  {-2, -16, 106, 7, 6,  49, -20, -2}, {-2, -15, 110, 7, 6,  43, -19, -2},
+  {-1, -13, 113, 6, 5,  38, -18, -2}, {-1, -12, 116, 6, 5,  33, -17, -2},
+  {-1, -10, 119, 5, 4,  27, -15, -1}, {-1,  -8, 121, 4, 3,  23, -13, -1},
+  {-1,  -7, 123, 4, 3,  18, -11, -1}, {-1,  -5, 125, 3, 2,  13,  -8, -1},
+  { 0,  -3, 126, 2, 1,   8,  -6,  0}, { 0,  -2, 127, 1, 1,   4,  -3,  0},
+  // [1, 2)
+  { 0,  0, 127,   0, 0,   1,   0, 0}, { 0, 1, 127,  -1, 0,  -3,   4, 0},
+  { 0,  1, 126,  -3, 0,  -5,   8, 1}, { 0, 1, 124,  -4, 0,  -7,  13, 1},
+  { 0,  2, 122,  -6, 0,  -9,  18, 1}, { 0, 2, 120,  -7, 0, -11,  22, 2},
+  { 0,  3, 117,  -8, 0, -13,  27, 2}, { 0, 3, 114, -10, 0, -14,  32, 3},
+  { 0,  3, 111, -11, 0, -15,  37, 3}, { 0, 3, 108, -12, 0, -16,  42, 3},
+  { 0,  4, 104, -13, 0, -17,  47, 3}, { 0, 4, 100, -14, 0, -17,  52, 3},
+  { 0,  4,  96, -15, 0, -18,  58, 3}, { 0, 4,  91, -16, 0, -18,  63, 4},
+  { 0,  4,  87, -17, 0, -18,  68, 4}, { 0, 4,  82, -17, 0, -18,  73, 4},
+  { 0,  4,  78, -18, 0, -18,  78, 4}, { 0, 4,  73, -18, 0, -17,  82, 4},
+  { 0,  4,  68, -18, 0, -17,  87, 4}, { 0, 4,  63, -18, 0, -16,  91, 4},
+  { 0,  3,  58, -18, 0, -15,  96, 4}, { 0, 3,  52, -17, 0, -14, 100, 4},
+  { 0,  3,  47, -17, 0, -13, 104, 4}, { 0, 3,  42, -16, 0, -12, 108, 3},
+  { 0,  3,  37, -15, 0, -11, 111, 3}, { 0, 3,  32, -14, 0, -10, 114, 3},
+  { 0,  2,  27, -13, 0,  -8, 117, 3}, { 0, 2,  22, -11, 0,  -7, 120, 2},
+  { 0,  1,  18,  -9, 0,  -6, 122, 2}, { 0, 1,  13,  -7, 0,  -4, 124, 1},
+  { 0,  1,   8,  -5, 0,  -3, 126, 1}, { 0, 0,   4,  -3, 0,  -1, 127, 1},
+  // dummy (replicate row index 95)
+  { 0, 0,   4,  -3, 0,  -1, 127, 1},
+#endif  // WARPEDPIXEL_PREC_BITS == 6
+};
+/* clang-format on */
+
+// Shuffle masks: we want to convert a sequence of bytes 0, 1, 2, ..., 15
+// in an SSE register into two sequences:
+// 0, 2, 2, 4, ..., 12, 12, 14, <don't care>
+// 1, 3, 3, 5, ..., 13, 13, 15, <don't care>
+static const uint8_t even_mask[16] = { 0, 2,  2,  4,  4,  6,  6,  8,
+                                       8, 10, 10, 12, 12, 14, 14, 0 };
+static const uint8_t odd_mask[16] = { 1, 3,  3,  5,  5,  7,  7,  9,
+                                      9, 11, 11, 13, 13, 15, 15, 0 };
+
+static const uint8_t shuffle_alpha0_mask01[16] = { 0, 1, 0, 1, 0, 1, 0, 1,
+                                                   0, 1, 0, 1, 0, 1, 0, 1 };
+
+static const uint8_t shuffle_alpha0_mask23[16] = { 2, 3, 2, 3, 2, 3, 2, 3,
+                                                   2, 3, 2, 3, 2, 3, 2, 3 };
+
+static const uint8_t shuffle_alpha0_mask45[16] = { 4, 5, 4, 5, 4, 5, 4, 5,
+                                                   4, 5, 4, 5, 4, 5, 4, 5 };
+
+static const uint8_t shuffle_alpha0_mask67[16] = { 6, 7, 6, 7, 6, 7, 6, 7,
+                                                   6, 7, 6, 7, 6, 7, 6, 7 };
+
+static const uint8_t shuffle_gamma0_mask0[16] = { 0, 1, 2, 3, 0, 1, 2, 3,
+                                                  0, 1, 2, 3, 0, 1, 2, 3 };
+static const uint8_t shuffle_gamma0_mask1[16] = { 4, 5, 6, 7, 4, 5, 6, 7,
+                                                  4, 5, 6, 7, 4, 5, 6, 7 };
+static const uint8_t shuffle_gamma0_mask2[16] = { 8, 9, 10, 11, 8, 9, 10, 11,
+                                                  8, 9, 10, 11, 8, 9, 10, 11 };
+static const uint8_t shuffle_gamma0_mask3[16] = {
+  12, 13, 14, 15, 12, 13, 14, 15, 12, 13, 14, 15, 12, 13, 14, 15
+};
+
+static INLINE void filter_src_pixels(__m128i src, __m128i *tmp, __m128i *coeff,
+                                     const int offset_bits_horiz,
+                                     const int reduce_bits_horiz, int k) {
+  const __m128i src_even =
+      _mm_shuffle_epi8(src, _mm_loadu_si128((__m128i *)even_mask));
+  const __m128i src_odd =
+      _mm_shuffle_epi8(src, _mm_loadu_si128((__m128i *)odd_mask));
+  // The pixel order we need for 'src' is:
+  // 0 2 2 4 4 6 6 8 1 3 3 5 5 7 7 9
+  const __m128i src_02 = _mm_unpacklo_epi64(src_even, src_odd);
+  const __m128i res_02 = _mm_maddubs_epi16(src_02, coeff[0]);
+  // 4 6 6 8 8 10 10 12 5 7 7 9 9 11 11 13
+  const __m128i src_46 = _mm_unpacklo_epi64(_mm_srli_si128(src_even, 4),
+                                            _mm_srli_si128(src_odd, 4));
+  const __m128i res_46 = _mm_maddubs_epi16(src_46, coeff[1]);
+  // 1 3 3 5 5 7 7 9 2 4 4 6 6 8 8 10
+  const __m128i src_13 =
+      _mm_unpacklo_epi64(src_odd, _mm_srli_si128(src_even, 2));
+  const __m128i res_13 = _mm_maddubs_epi16(src_13, coeff[2]);
+  // 5 7 7 9 9 11 11 13 6 8 8 10 10 12 12 14
+  const __m128i src_57 = _mm_unpacklo_epi64(_mm_srli_si128(src_odd, 4),
+                                            _mm_srli_si128(src_even, 6));
+  const __m128i res_57 = _mm_maddubs_epi16(src_57, coeff[3]);
+
+  const __m128i round_const = _mm_set1_epi16((1 << offset_bits_horiz) +
+                                             ((1 << reduce_bits_horiz) >> 1));
+
+  // Note: The values res_02 + res_46 and res_13 + res_57 both
+  // fit into int16s at this point, but their sum may be too wide to fit
+  // into an int16. However, once we also add round_const, the sum of
+  // all of these fits into a uint16.
+  //
+  // The wrapping behaviour of _mm_add_* is used here to make sure we
+  // get the correct result despite converting between different
+  // (implicit) types.
+  const __m128i res_even = _mm_add_epi16(res_02, res_46);
+  const __m128i res_odd = _mm_add_epi16(res_13, res_57);
+  const __m128i res =
+      _mm_add_epi16(_mm_add_epi16(res_even, res_odd), round_const);
+  tmp[k + 7] = _mm_srl_epi16(res, _mm_cvtsi32_si128(reduce_bits_horiz));
+}
+
+static INLINE void prepare_horizontal_filter_coeff(int alpha, int sx,
+                                                   __m128i *coeff) {
+  // Filter even-index pixels
+  const __m128i tmp_0 = _mm_loadl_epi64(
+      (__m128i *)&filter_8bit[(sx + 0 * alpha) >> WARPEDDIFF_PREC_BITS]);
+  const __m128i tmp_1 = _mm_loadl_epi64(
+      (__m128i *)&filter_8bit[(sx + 1 * alpha) >> WARPEDDIFF_PREC_BITS]);
+  const __m128i tmp_2 = _mm_loadl_epi64(
+      (__m128i *)&filter_8bit[(sx + 2 * alpha) >> WARPEDDIFF_PREC_BITS]);
+  const __m128i tmp_3 = _mm_loadl_epi64(
+      (__m128i *)&filter_8bit[(sx + 3 * alpha) >> WARPEDDIFF_PREC_BITS]);
+  const __m128i tmp_4 = _mm_loadl_epi64(
+      (__m128i *)&filter_8bit[(sx + 4 * alpha) >> WARPEDDIFF_PREC_BITS]);
+  const __m128i tmp_5 = _mm_loadl_epi64(
+      (__m128i *)&filter_8bit[(sx + 5 * alpha) >> WARPEDDIFF_PREC_BITS]);
+  const __m128i tmp_6 = _mm_loadl_epi64(
+      (__m128i *)&filter_8bit[(sx + 6 * alpha) >> WARPEDDIFF_PREC_BITS]);
+  const __m128i tmp_7 = _mm_loadl_epi64(
+      (__m128i *)&filter_8bit[(sx + 7 * alpha) >> WARPEDDIFF_PREC_BITS]);
+
+  // Coeffs 0 2 0 2 4 6 4 6 1 3 1 3 5 7 5 7 for pixels 0 2
+  const __m128i tmp_8 = _mm_unpacklo_epi16(tmp_0, tmp_2);
+  // Coeffs 0 2 0 2 4 6 4 6 1 3 1 3 5 7 5 7 for pixels 1 3
+  const __m128i tmp_9 = _mm_unpacklo_epi16(tmp_1, tmp_3);
+  // Coeffs 0 2 0 2 4 6 4 6 1 3 1 3 5 7 5 7 for pixels 4 6
+  const __m128i tmp_10 = _mm_unpacklo_epi16(tmp_4, tmp_6);
+  // Coeffs 0 2 0 2 4 6 4 6 1 3 1 3 5 7 5 7 for pixels 5 7
+  const __m128i tmp_11 = _mm_unpacklo_epi16(tmp_5, tmp_7);
+
+  // Coeffs 0 2 0 2 0 2 0 2 4 6 4 6 4 6 4 6 for pixels 0 2 4 6
+  const __m128i tmp_12 = _mm_unpacklo_epi32(tmp_8, tmp_10);
+  // Coeffs 1 3 1 3 1 3 1 3 5 7 5 7 5 7 5 7 for pixels 0 2 4 6
+  const __m128i tmp_13 = _mm_unpackhi_epi32(tmp_8, tmp_10);
+  // Coeffs 0 2 0 2 0 2 0 2 4 6 4 6 4 6 4 6 for pixels 1 3 5 7
+  const __m128i tmp_14 = _mm_unpacklo_epi32(tmp_9, tmp_11);
+  // Coeffs 1 3 1 3 1 3 1 3 5 7 5 7 5 7 5 7 for pixels 1 3 5 7
+  const __m128i tmp_15 = _mm_unpackhi_epi32(tmp_9, tmp_11);
+
+  // Coeffs 0 2 for pixels 0 2 4 6 1 3 5 7
+  coeff[0] = _mm_unpacklo_epi64(tmp_12, tmp_14);
+  // Coeffs 4 6 for pixels 0 2 4 6 1 3 5 7
+  coeff[1] = _mm_unpackhi_epi64(tmp_12, tmp_14);
+  // Coeffs 1 3 for pixels 0 2 4 6 1 3 5 7
+  coeff[2] = _mm_unpacklo_epi64(tmp_13, tmp_15);
+  // Coeffs 5 7 for pixels 0 2 4 6 1 3 5 7
+  coeff[3] = _mm_unpackhi_epi64(tmp_13, tmp_15);
+}
+
+static INLINE void prepare_horizontal_filter_coeff_alpha0(int sx,
+                                                          __m128i *coeff) {
+  // Filter even-index pixels
+  const __m128i tmp_0 =
+      _mm_loadl_epi64((__m128i *)&filter_8bit[sx >> WARPEDDIFF_PREC_BITS]);
+
+  // Coeffs 0 2 for pixels 0 2 4 6 1 3 5 7
+  coeff[0] = _mm_shuffle_epi8(
+      tmp_0, _mm_loadu_si128((__m128i *)shuffle_alpha0_mask01));
+  // Coeffs 4 6 for pixels 0 2 4 6 1 3 5 7
+  coeff[1] = _mm_shuffle_epi8(
+      tmp_0, _mm_loadu_si128((__m128i *)shuffle_alpha0_mask23));
+  // Coeffs 1 3 for pixels 0 2 4 6 1 3 5 7
+  coeff[2] = _mm_shuffle_epi8(
+      tmp_0, _mm_loadu_si128((__m128i *)shuffle_alpha0_mask45));
+  // Coeffs 5 7 for pixels 0 2 4 6 1 3 5 7
+  coeff[3] = _mm_shuffle_epi8(
+      tmp_0, _mm_loadu_si128((__m128i *)shuffle_alpha0_mask67));
+}
+
+static INLINE void horizontal_filter(__m128i src, __m128i *tmp, int sx,
+                                     int alpha, int k,
+                                     const int offset_bits_horiz,
+                                     const int reduce_bits_horiz) {
+  __m128i coeff[4];
+  prepare_horizontal_filter_coeff(alpha, sx, coeff);
+  filter_src_pixels(src, tmp, coeff, offset_bits_horiz, reduce_bits_horiz, k);
+}
+
+static INLINE void warp_horizontal_filter(const uint8_t *ref, __m128i *tmp,
+                                          int stride, int32_t ix4, int32_t iy4,
+                                          int32_t sx4, int alpha, int beta,
+                                          int p_height, int height, int i,
+                                          const int offset_bits_horiz,
+                                          const int reduce_bits_horiz) {
+  int k;
+  for (k = -7; k < AOMMIN(8, p_height - i); ++k) {
+    int iy = iy4 + k;
+    if (iy < 0)
+      iy = 0;
+    else if (iy > height - 1)
+      iy = height - 1;
+    int sx = sx4 + beta * (k + 4);
+
+    // Load source pixels
+    const __m128i src =
+        _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7));
+    horizontal_filter(src, tmp, sx, alpha, k, offset_bits_horiz,
+                      reduce_bits_horiz);
+  }
+}
+
+static INLINE void warp_horizontal_filter_alpha0(
+    const uint8_t *ref, __m128i *tmp, int stride, int32_t ix4, int32_t iy4,
+    int32_t sx4, int alpha, int beta, int p_height, int height, int i,
+    const int offset_bits_horiz, const int reduce_bits_horiz) {
+  (void)alpha;
+  int k;
+  for (k = -7; k < AOMMIN(8, p_height - i); ++k) {
+    int iy = iy4 + k;
+    if (iy < 0)
+      iy = 0;
+    else if (iy > height - 1)
+      iy = height - 1;
+    int sx = sx4 + beta * (k + 4);
+
+    // Load source pixels
+    const __m128i src =
+        _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7));
+
+    __m128i coeff[4];
+    prepare_horizontal_filter_coeff_alpha0(sx, coeff);
+    filter_src_pixels(src, tmp, coeff, offset_bits_horiz, reduce_bits_horiz, k);
+  }
+}
+
+static INLINE void warp_horizontal_filter_beta0(
+    const uint8_t *ref, __m128i *tmp, int stride, int32_t ix4, int32_t iy4,
+    int32_t sx4, int alpha, int beta, int p_height, int height, int i,
+    const int offset_bits_horiz, const int reduce_bits_horiz) {
+  (void)beta;
+  int k;
+  __m128i coeff[4];
+  prepare_horizontal_filter_coeff(alpha, sx4, coeff);
+
+  for (k = -7; k < AOMMIN(8, p_height - i); ++k) {
+    int iy = iy4 + k;
+    if (iy < 0)
+      iy = 0;
+    else if (iy > height - 1)
+      iy = height - 1;
+
+    // Load source pixels
+    const __m128i src =
+        _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7));
+    filter_src_pixels(src, tmp, coeff, offset_bits_horiz, reduce_bits_horiz, k);
+  }
+}
+
+static INLINE void warp_horizontal_filter_alpha0_beta0(
+    const uint8_t *ref, __m128i *tmp, int stride, int32_t ix4, int32_t iy4,
+    int32_t sx4, int alpha, int beta, int p_height, int height, int i,
+    const int offset_bits_horiz, const int reduce_bits_horiz) {
+  (void)beta;
+  (void)alpha;
+  int k;
+
+  __m128i coeff[4];
+  prepare_horizontal_filter_coeff_alpha0(sx4, coeff);
+
+  for (k = -7; k < AOMMIN(8, p_height - i); ++k) {
+    int iy = iy4 + k;
+    if (iy < 0)
+      iy = 0;
+    else if (iy > height - 1)
+      iy = height - 1;
+
+    // Load source pixels
+    const __m128i src =
+        _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7));
+    filter_src_pixels(src, tmp, coeff, offset_bits_horiz, reduce_bits_horiz, k);
+  }
+}
+
+static INLINE void unpack_weights_and_set_round_const(
+    ConvolveParams *conv_params, const int round_bits, const int offset_bits,
+    __m128i *res_sub_const, __m128i *round_bits_const, __m128i *wt) {
+  *res_sub_const =
+      _mm_set1_epi16(-(1 << (offset_bits - conv_params->round_1)) -
+                     (1 << (offset_bits - conv_params->round_1 - 1)));
+  *round_bits_const = _mm_set1_epi16(((1 << round_bits) >> 1));
+
+  const int w0 = conv_params->fwd_offset;
+  const int w1 = conv_params->bck_offset;
+  const __m128i wt0 = _mm_set1_epi16(w0);
+  const __m128i wt1 = _mm_set1_epi16(w1);
+  *wt = _mm_unpacklo_epi16(wt0, wt1);
+}
+
+static INLINE void prepare_vertical_filter_coeffs(int gamma, int sy,
+                                                  __m128i *coeffs) {
+  const __m128i tmp_0 = _mm_loadu_si128(
+      (__m128i *)(warped_filter + ((sy + 0 * gamma) >> WARPEDDIFF_PREC_BITS)));
+  const __m128i tmp_2 = _mm_loadu_si128(
+      (__m128i *)(warped_filter + ((sy + 2 * gamma) >> WARPEDDIFF_PREC_BITS)));
+  const __m128i tmp_4 = _mm_loadu_si128(
+      (__m128i *)(warped_filter + ((sy + 4 * gamma) >> WARPEDDIFF_PREC_BITS)));
+  const __m128i tmp_6 = _mm_loadu_si128(
+      (__m128i *)(warped_filter + ((sy + 6 * gamma) >> WARPEDDIFF_PREC_BITS)));
+
+  const __m128i tmp_8 = _mm_unpacklo_epi32(tmp_0, tmp_2);
+  const __m128i tmp_10 = _mm_unpacklo_epi32(tmp_4, tmp_6);
+  const __m128i tmp_12 = _mm_unpackhi_epi32(tmp_0, tmp_2);
+  const __m128i tmp_14 = _mm_unpackhi_epi32(tmp_4, tmp_6);
+
+  // even coeffs
+  coeffs[0] = _mm_unpacklo_epi64(tmp_8, tmp_10);
+  coeffs[1] = _mm_unpackhi_epi64(tmp_8, tmp_10);
+  coeffs[2] = _mm_unpacklo_epi64(tmp_12, tmp_14);
+  coeffs[3] = _mm_unpackhi_epi64(tmp_12, tmp_14);
+
+  const __m128i tmp_1 = _mm_loadu_si128(
+      (__m128i *)(warped_filter + ((sy + 1 * gamma) >> WARPEDDIFF_PREC_BITS)));
+  const __m128i tmp_3 = _mm_loadu_si128(
+      (__m128i *)(warped_filter + ((sy + 3 * gamma) >> WARPEDDIFF_PREC_BITS)));
+  const __m128i tmp_5 = _mm_loadu_si128(
+      (__m128i *)(warped_filter + ((sy + 5 * gamma) >> WARPEDDIFF_PREC_BITS)));
+  const __m128i tmp_7 = _mm_loadu_si128(
+      (__m128i *)(warped_filter + ((sy + 7 * gamma) >> WARPEDDIFF_PREC_BITS)));
+
+  const __m128i tmp_9 = _mm_unpacklo_epi32(tmp_1, tmp_3);
+  const __m128i tmp_11 = _mm_unpacklo_epi32(tmp_5, tmp_7);
+  const __m128i tmp_13 = _mm_unpackhi_epi32(tmp_1, tmp_3);
+  const __m128i tmp_15 = _mm_unpackhi_epi32(tmp_5, tmp_7);
+
+  // odd coeffs
+  coeffs[4] = _mm_unpacklo_epi64(tmp_9, tmp_11);
+  coeffs[5] = _mm_unpackhi_epi64(tmp_9, tmp_11);
+  coeffs[6] = _mm_unpacklo_epi64(tmp_13, tmp_15);
+  coeffs[7] = _mm_unpackhi_epi64(tmp_13, tmp_15);
+}
+
+static INLINE void prepare_vertical_filter_coeffs_gamma0(int sy,
+                                                         __m128i *coeffs) {
+  const __m128i tmp_0 = _mm_loadu_si128(
+      (__m128i *)(warped_filter + (sy >> WARPEDDIFF_PREC_BITS)));
+
+  // even coeffs
+  coeffs[0] =
+      _mm_shuffle_epi8(tmp_0, _mm_loadu_si128((__m128i *)shuffle_gamma0_mask0));
+  coeffs[1] =
+      _mm_shuffle_epi8(tmp_0, _mm_loadu_si128((__m128i *)shuffle_gamma0_mask1));
+  coeffs[2] =
+      _mm_shuffle_epi8(tmp_0, _mm_loadu_si128((__m128i *)shuffle_gamma0_mask2));
+  coeffs[3] =
+      _mm_shuffle_epi8(tmp_0, _mm_loadu_si128((__m128i *)shuffle_gamma0_mask3));
+
+  // odd coeffs
+  coeffs[4] = coeffs[0];
+  coeffs[5] = coeffs[1];
+  coeffs[6] = coeffs[2];
+  coeffs[7] = coeffs[3];
+}
+
+static INLINE void filter_src_pixels_vertical(__m128i *tmp, __m128i *coeffs,
+                                              __m128i *res_lo, __m128i *res_hi,
+                                              int k) {
+  // Load from tmp and rearrange pairs of consecutive rows into the
+  // column order 0 0 2 2 4 4 6 6; 1 1 3 3 5 5 7 7
+  const __m128i *src = tmp + (k + 4);
+  const __m128i src_0 = _mm_unpacklo_epi16(src[0], src[1]);
+  const __m128i src_2 = _mm_unpacklo_epi16(src[2], src[3]);
+  const __m128i src_4 = _mm_unpacklo_epi16(src[4], src[5]);
+  const __m128i src_6 = _mm_unpacklo_epi16(src[6], src[7]);
+
+  const __m128i res_0 = _mm_madd_epi16(src_0, coeffs[0]);
+  const __m128i res_2 = _mm_madd_epi16(src_2, coeffs[1]);
+  const __m128i res_4 = _mm_madd_epi16(src_4, coeffs[2]);
+  const __m128i res_6 = _mm_madd_epi16(src_6, coeffs[3]);
+
+  const __m128i res_even =
+      _mm_add_epi32(_mm_add_epi32(res_0, res_2), _mm_add_epi32(res_4, res_6));
+
+  // Filter odd-index pixels
+  const __m128i src_1 = _mm_unpackhi_epi16(src[0], src[1]);
+  const __m128i src_3 = _mm_unpackhi_epi16(src[2], src[3]);
+  const __m128i src_5 = _mm_unpackhi_epi16(src[4], src[5]);
+  const __m128i src_7 = _mm_unpackhi_epi16(src[6], src[7]);
+
+  const __m128i res_1 = _mm_madd_epi16(src_1, coeffs[4]);
+  const __m128i res_3 = _mm_madd_epi16(src_3, coeffs[5]);
+  const __m128i res_5 = _mm_madd_epi16(src_5, coeffs[6]);
+  const __m128i res_7 = _mm_madd_epi16(src_7, coeffs[7]);
+
+  const __m128i res_odd =
+      _mm_add_epi32(_mm_add_epi32(res_1, res_3), _mm_add_epi32(res_5, res_7));
+
+  // Rearrange pixels back into the order 0 ... 7
+  *res_lo = _mm_unpacklo_epi32(res_even, res_odd);
+  *res_hi = _mm_unpackhi_epi32(res_even, res_odd);
+}
+
+static INLINE void store_vertical_filter_output(
+    __m128i *res_lo, __m128i *res_hi, const __m128i *res_add_const,
+    const __m128i *wt, const __m128i *res_sub_const, __m128i *round_bits_const,
+    uint8_t *pred, ConvolveParams *conv_params, int i, int j, int k,
+    const int reduce_bits_vert, int p_stride, int p_width,
+    const int round_bits) {
+  __m128i res_lo_1 = *res_lo;
+  __m128i res_hi_1 = *res_hi;
+
+  if (conv_params->is_compound) {
+    __m128i *const p =
+        (__m128i *)&conv_params->dst[(i + k + 4) * conv_params->dst_stride + j];
+    res_lo_1 = _mm_srai_epi32(_mm_add_epi32(res_lo_1, *res_add_const),
+                              reduce_bits_vert);
+    const __m128i temp_lo_16 = _mm_packus_epi32(res_lo_1, res_lo_1);
+    __m128i res_lo_16;
+    if (conv_params->do_average) {
+      __m128i *const dst8 = (__m128i *)&pred[(i + k + 4) * p_stride + j];
+      const __m128i p_16 = _mm_loadl_epi64(p);
+
+      if (conv_params->use_jnt_comp_avg) {
+        const __m128i p_16_lo = _mm_unpacklo_epi16(p_16, temp_lo_16);
+        const __m128i wt_res_lo = _mm_madd_epi16(p_16_lo, *wt);
+        const __m128i shifted_32 =
+            _mm_srai_epi32(wt_res_lo, DIST_PRECISION_BITS);
+        res_lo_16 = _mm_packus_epi32(shifted_32, shifted_32);
+      } else {
+        res_lo_16 = _mm_srai_epi16(_mm_add_epi16(p_16, temp_lo_16), 1);
+      }
+
+      res_lo_16 = _mm_add_epi16(res_lo_16, *res_sub_const);
+
+      res_lo_16 = _mm_srai_epi16(_mm_add_epi16(res_lo_16, *round_bits_const),
+                                 round_bits);
+      __m128i res_8_lo = _mm_packus_epi16(res_lo_16, res_lo_16);
+      *(uint32_t *)dst8 = _mm_cvtsi128_si32(res_8_lo);
+    } else {
+      _mm_storel_epi64(p, temp_lo_16);
+    }
+    if (p_width > 4) {
+      __m128i *const p4 =
+          (__m128i *)&conv_params
+              ->dst[(i + k + 4) * conv_params->dst_stride + j + 4];
+      res_hi_1 = _mm_srai_epi32(_mm_add_epi32(res_hi_1, *res_add_const),
+                                reduce_bits_vert);
+      const __m128i temp_hi_16 = _mm_packus_epi32(res_hi_1, res_hi_1);
+      __m128i res_hi_16;
+
+      if (conv_params->do_average) {
+        __m128i *const dst8_4 =
+            (__m128i *)&pred[(i + k + 4) * p_stride + j + 4];
+        const __m128i p4_16 = _mm_loadl_epi64(p4);
+
+        if (conv_params->use_jnt_comp_avg) {
+          const __m128i p_16_hi = _mm_unpacklo_epi16(p4_16, temp_hi_16);
+          const __m128i wt_res_hi = _mm_madd_epi16(p_16_hi, *wt);
+          const __m128i shifted_32 =
+              _mm_srai_epi32(wt_res_hi, DIST_PRECISION_BITS);
+          res_hi_16 = _mm_packus_epi32(shifted_32, shifted_32);
+        } else {
+          res_hi_16 = _mm_srai_epi16(_mm_add_epi16(p4_16, temp_hi_16), 1);
+        }
+        res_hi_16 = _mm_add_epi16(res_hi_16, *res_sub_const);
+
+        res_hi_16 = _mm_srai_epi16(_mm_add_epi16(res_hi_16, *round_bits_const),
+                                   round_bits);
+        __m128i res_8_hi = _mm_packus_epi16(res_hi_16, res_hi_16);
+        *(uint32_t *)dst8_4 = _mm_cvtsi128_si32(res_8_hi);
+
+      } else {
+        _mm_storel_epi64(p4, temp_hi_16);
+      }
+    }
+  } else {
+    const __m128i res_lo_round = _mm_srai_epi32(
+        _mm_add_epi32(res_lo_1, *res_add_const), reduce_bits_vert);
+    const __m128i res_hi_round = _mm_srai_epi32(
+        _mm_add_epi32(res_hi_1, *res_add_const), reduce_bits_vert);
+
+    const __m128i res_16bit = _mm_packs_epi32(res_lo_round, res_hi_round);
+    __m128i res_8bit = _mm_packus_epi16(res_16bit, res_16bit);
+
+    // Store, blending with 'pred' if needed
+    __m128i *const p = (__m128i *)&pred[(i + k + 4) * p_stride + j];
+
+    // Note: If we're outputting a 4x4 block, we need to be very careful
+    // to only output 4 pixels at this point, to avoid encode/decode
+    // mismatches when encoding with multiple threads.
+    if (p_width == 4) {
+      *(uint32_t *)p = _mm_cvtsi128_si32(res_8bit);
+    } else {
+      _mm_storel_epi64(p, res_8bit);
+    }
+  }
+}
+
+static INLINE void warp_vertical_filter(
+    uint8_t *pred, __m128i *tmp, ConvolveParams *conv_params, int16_t gamma,
+    int16_t delta, int p_height, int p_stride, int p_width, int i, int j,
+    int sy4, const int reduce_bits_vert, const __m128i *res_add_const,
+    const int round_bits, const int offset_bits) {
+  int k;
+  __m128i res_sub_const, round_bits_const, wt;
+  unpack_weights_and_set_round_const(conv_params, round_bits, offset_bits,
+                                     &res_sub_const, &round_bits_const, &wt);
+  // Vertical filter
+  for (k = -4; k < AOMMIN(4, p_height - i - 4); ++k) {
+    int sy = sy4 + delta * (k + 4);
+
+    __m128i coeffs[8];
+    prepare_vertical_filter_coeffs(gamma, sy, coeffs);
+
+    __m128i res_lo;
+    __m128i res_hi;
+    filter_src_pixels_vertical(tmp, coeffs, &res_lo, &res_hi, k);
+
+    store_vertical_filter_output(&res_lo, &res_hi, res_add_const, &wt,
+                                 &res_sub_const, &round_bits_const, pred,
+                                 conv_params, i, j, k, reduce_bits_vert,
+                                 p_stride, p_width, round_bits);
+  }
+}
+
+static INLINE void warp_vertical_filter_gamma0(
+    uint8_t *pred, __m128i *tmp, ConvolveParams *conv_params, int16_t gamma,
+    int16_t delta, int p_height, int p_stride, int p_width, int i, int j,
+    int sy4, const int reduce_bits_vert, const __m128i *res_add_const,
+    const int round_bits, const int offset_bits) {
+  int k;
+  (void)gamma;
+  __m128i res_sub_const, round_bits_const, wt;
+  unpack_weights_and_set_round_const(conv_params, round_bits, offset_bits,
+                                     &res_sub_const, &round_bits_const, &wt);
+  // Vertical filter
+  for (k = -4; k < AOMMIN(4, p_height - i - 4); ++k) {
+    int sy = sy4 + delta * (k + 4);
+
+    __m128i coeffs[8];
+    prepare_vertical_filter_coeffs_gamma0(sy, coeffs);
+
+    __m128i res_lo;
+    __m128i res_hi;
+    filter_src_pixels_vertical(tmp, coeffs, &res_lo, &res_hi, k);
+
+    store_vertical_filter_output(&res_lo, &res_hi, res_add_const, &wt,
+                                 &res_sub_const, &round_bits_const, pred,
+                                 conv_params, i, j, k, reduce_bits_vert,
+                                 p_stride, p_width, round_bits);
+  }
+}
+
+static INLINE void warp_vertical_filter_delta0(
+    uint8_t *pred, __m128i *tmp, ConvolveParams *conv_params, int16_t gamma,
+    int16_t delta, int p_height, int p_stride, int p_width, int i, int j,
+    int sy4, const int reduce_bits_vert, const __m128i *res_add_const,
+    const int round_bits, const int offset_bits) {
+  (void)delta;
+  int k;
+  __m128i res_sub_const, round_bits_const, wt;
+  unpack_weights_and_set_round_const(conv_params, round_bits, offset_bits,
+                                     &res_sub_const, &round_bits_const, &wt);
+
+  __m128i coeffs[8];
+  prepare_vertical_filter_coeffs(gamma, sy4, coeffs);
+  // Vertical filter
+  for (k = -4; k < AOMMIN(4, p_height - i - 4); ++k) {
+    __m128i res_lo;
+    __m128i res_hi;
+    filter_src_pixels_vertical(tmp, coeffs, &res_lo, &res_hi, k);
+
+    store_vertical_filter_output(&res_lo, &res_hi, res_add_const, &wt,
+                                 &res_sub_const, &round_bits_const, pred,
+                                 conv_params, i, j, k, reduce_bits_vert,
+                                 p_stride, p_width, round_bits);
+  }
+}
+
+static INLINE void warp_vertical_filter_gamma0_delta0(
+    uint8_t *pred, __m128i *tmp, ConvolveParams *conv_params, int16_t gamma,
+    int16_t delta, int p_height, int p_stride, int p_width, int i, int j,
+    int sy4, const int reduce_bits_vert, const __m128i *res_add_const,
+    const int round_bits, const int offset_bits) {
+  (void)delta;
+  (void)gamma;
+  int k;
+  __m128i res_sub_const, round_bits_const, wt;
+  unpack_weights_and_set_round_const(conv_params, round_bits, offset_bits,
+                                     &res_sub_const, &round_bits_const, &wt);
+
+  __m128i coeffs[8];
+  prepare_vertical_filter_coeffs_gamma0(sy4, coeffs);
+  // Vertical filter
+  for (k = -4; k < AOMMIN(4, p_height - i - 4); ++k) {
+    __m128i res_lo;
+    __m128i res_hi;
+    filter_src_pixels_vertical(tmp, coeffs, &res_lo, &res_hi, k);
+
+    store_vertical_filter_output(&res_lo, &res_hi, res_add_const, &wt,
+                                 &res_sub_const, &round_bits_const, pred,
+                                 conv_params, i, j, k, reduce_bits_vert,
+                                 p_stride, p_width, round_bits);
+  }
+}
+
+static INLINE void prepare_warp_vertical_filter(
+    uint8_t *pred, __m128i *tmp, ConvolveParams *conv_params, int16_t gamma,
+    int16_t delta, int p_height, int p_stride, int p_width, int i, int j,
+    int sy4, const int reduce_bits_vert, const __m128i *res_add_const,
+    const int round_bits, const int offset_bits) {
+  if (gamma == 0 && delta == 0)
+    warp_vertical_filter_gamma0_delta0(
+        pred, tmp, conv_params, gamma, delta, p_height, p_stride, p_width, i, j,
+        sy4, reduce_bits_vert, res_add_const, round_bits, offset_bits);
+  else if (gamma == 0 && delta != 0)
+    warp_vertical_filter_gamma0(pred, tmp, conv_params, gamma, delta, p_height,
+                                p_stride, p_width, i, j, sy4, reduce_bits_vert,
+                                res_add_const, round_bits, offset_bits);
+  else if (gamma != 0 && delta == 0)
+    warp_vertical_filter_delta0(pred, tmp, conv_params, gamma, delta, p_height,
+                                p_stride, p_width, i, j, sy4, reduce_bits_vert,
+                                res_add_const, round_bits, offset_bits);
+  else
+    warp_vertical_filter(pred, tmp, conv_params, gamma, delta, p_height,
+                         p_stride, p_width, i, j, sy4, reduce_bits_vert,
+                         res_add_const, round_bits, offset_bits);
+}
+
+static INLINE void prepare_warp_horizontal_filter(
+    const uint8_t *ref, __m128i *tmp, int stride, int32_t ix4, int32_t iy4,
+    int32_t sx4, int alpha, int beta, int p_height, int height, int i,
+    const int offset_bits_horiz, const int reduce_bits_horiz) {
+  if (alpha == 0 && beta == 0)
+    warp_horizontal_filter_alpha0_beta0(ref, tmp, stride, ix4, iy4, sx4, alpha,
+                                        beta, p_height, height, i,
+                                        offset_bits_horiz, reduce_bits_horiz);
+  else if (alpha == 0 && beta != 0)
+    warp_horizontal_filter_alpha0(ref, tmp, stride, ix4, iy4, sx4, alpha, beta,
+                                  p_height, height, i, offset_bits_horiz,
+                                  reduce_bits_horiz);
+  else if (alpha != 0 && beta == 0)
+    warp_horizontal_filter_beta0(ref, tmp, stride, ix4, iy4, sx4, alpha, beta,
+                                 p_height, height, i, offset_bits_horiz,
+                                 reduce_bits_horiz);
+  else
+    warp_horizontal_filter(ref, tmp, stride, ix4, iy4, sx4, alpha, beta,
+                           p_height, height, i, offset_bits_horiz,
+                           reduce_bits_horiz);
+}
+
+void av1_warp_affine_sse4_1(const int32_t *mat, const uint8_t *ref, int width,
+                            int height, int stride, uint8_t *pred, int p_col,
+                            int p_row, int p_width, int p_height, int p_stride,
+                            int subsampling_x, int subsampling_y,
+                            ConvolveParams *conv_params, int16_t alpha,
+                            int16_t beta, int16_t gamma, int16_t delta) {
+  __m128i tmp[15];
+  int i, j, k;
+  const int bd = 8;
+  const int reduce_bits_horiz = conv_params->round_0;
+  const int reduce_bits_vert = conv_params->is_compound
+                                   ? conv_params->round_1
+                                   : 2 * FILTER_BITS - reduce_bits_horiz;
+  const int offset_bits_horiz = bd + FILTER_BITS - 1;
+  assert(IMPLIES(conv_params->is_compound, conv_params->dst != NULL));
+
+  const int offset_bits_vert = bd + 2 * FILTER_BITS - reduce_bits_horiz;
+  const __m128i reduce_bits_vert_const =
+      _mm_set1_epi32(((1 << reduce_bits_vert) >> 1));
+  const __m128i res_add_const = _mm_set1_epi32(1 << offset_bits_vert);
+  const int round_bits =
+      2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+  const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0;
+  assert(IMPLIES(conv_params->do_average, conv_params->is_compound));
+
+  /* Note: For this code to work, the left/right frame borders need to be
+  extended by at least 13 pixels each. By the time we get here, other
+  code will have set up this border, but we allow an explicit check
+  for debugging purposes.
+  */
+  /*for (i = 0; i < height; ++i) {
+  for (j = 0; j < 13; ++j) {
+  assert(ref[i * stride - 13 + j] == ref[i * stride]);
+  assert(ref[i * stride + width + j] == ref[i * stride + (width - 1)]);
+  }
+  }*/
+  __m128i res_add_const_1;
+  if (conv_params->is_compound == 1) {
+    res_add_const_1 = _mm_add_epi32(reduce_bits_vert_const, res_add_const);
+  } else {
+    res_add_const_1 = _mm_set1_epi32(-(1 << (bd + reduce_bits_vert - 1)) +
+                                     ((1 << reduce_bits_vert) >> 1));
+  }
+
+  for (i = 0; i < p_height; i += 8) {
+    for (j = 0; j < p_width; j += 8) {
+      const int32_t src_x = (p_col + j + 4) << subsampling_x;
+      const int32_t src_y = (p_row + i + 4) << subsampling_y;
+      const int32_t dst_x = mat[2] * src_x + mat[3] * src_y + mat[0];
+      const int32_t dst_y = mat[4] * src_x + mat[5] * src_y + mat[1];
+      const int32_t x4 = dst_x >> subsampling_x;
+      const int32_t y4 = dst_y >> subsampling_y;
+
+      int32_t ix4 = x4 >> WARPEDMODEL_PREC_BITS;
+      int32_t sx4 = x4 & ((1 << WARPEDMODEL_PREC_BITS) - 1);
+      int32_t iy4 = y4 >> WARPEDMODEL_PREC_BITS;
+      int32_t sy4 = y4 & ((1 << WARPEDMODEL_PREC_BITS) - 1);
+
+      // Add in all the constant terms, including rounding and offset
+      sx4 += alpha * (-4) + beta * (-4) + (1 << (WARPEDDIFF_PREC_BITS - 1)) +
+             (WARPEDPIXEL_PREC_SHIFTS << WARPEDDIFF_PREC_BITS);
+      sy4 += gamma * (-4) + delta * (-4) + (1 << (WARPEDDIFF_PREC_BITS - 1)) +
+             (WARPEDPIXEL_PREC_SHIFTS << WARPEDDIFF_PREC_BITS);
+
+      sx4 &= ~((1 << WARP_PARAM_REDUCE_BITS) - 1);
+      sy4 &= ~((1 << WARP_PARAM_REDUCE_BITS) - 1);
+
+      // Horizontal filter
+      // If the block is aligned such that, after clamping, every sample
+      // would be taken from the leftmost/rightmost column, then we can
+      // skip the expensive horizontal filter.
+      if (ix4 <= -7) {
+        for (k = -7; k < AOMMIN(8, p_height - i); ++k) {
+          int iy = iy4 + k;
+          if (iy < 0)
+            iy = 0;
+          else if (iy > height - 1)
+            iy = height - 1;
+          tmp[k + 7] = _mm_set1_epi16(
+              (1 << (bd + FILTER_BITS - reduce_bits_horiz - 1)) +
+              ref[iy * stride] * (1 << (FILTER_BITS - reduce_bits_horiz)));
+        }
+      } else if (ix4 >= width + 6) {
+        for (k = -7; k < AOMMIN(8, p_height - i); ++k) {
+          int iy = iy4 + k;
+          if (iy < 0)
+            iy = 0;
+          else if (iy > height - 1)
+            iy = height - 1;
+          tmp[k + 7] =
+              _mm_set1_epi16((1 << (bd + FILTER_BITS - reduce_bits_horiz - 1)) +
+                             ref[iy * stride + (width - 1)] *
+                                 (1 << (FILTER_BITS - reduce_bits_horiz)));
+        }
+      } else if (((ix4 - 7) < 0) || ((ix4 + 9) > width)) {
+        const int out_of_boundary_left = -(ix4 - 6);
+        const int out_of_boundary_right = (ix4 + 8) - width;
+        for (k = -7; k < AOMMIN(8, p_height - i); ++k) {
+          int iy = iy4 + k;
+          if (iy < 0)
+            iy = 0;
+          else if (iy > height - 1)
+            iy = height - 1;
+          int sx = sx4 + beta * (k + 4);
+
+          // Load source pixels
+          __m128i src =
+              _mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7));
+          if (out_of_boundary_left >= 0) {
+            const __m128i shuffle_reg_left =
+                _mm_loadu_si128((__m128i *)warp_pad_left[out_of_boundary_left]);
+            src = _mm_shuffle_epi8(src, shuffle_reg_left);
+          }
+          if (out_of_boundary_right >= 0) {
+            const __m128i shuffle_reg_right = _mm_loadu_si128(
+                (__m128i *)warp_pad_right[out_of_boundary_right]);
+            src = _mm_shuffle_epi8(src, shuffle_reg_right);
+          }
+          horizontal_filter(src, tmp, sx, alpha, k, offset_bits_horiz,
+                            reduce_bits_horiz);
+        }
+      } else {
+        prepare_warp_horizontal_filter(ref, tmp, stride, ix4, iy4, sx4, alpha,
+                                       beta, p_height, height, i,
+                                       offset_bits_horiz, reduce_bits_horiz);
+      }
+
+      // Vertical filter
+      prepare_warp_vertical_filter(
+          pred, tmp, conv_params, gamma, delta, p_height, p_stride, p_width, i,
+          j, sy4, reduce_bits_vert, &res_add_const_1, round_bits, offset_bits);
+    }
+  }
+}
diff --git a/third_party/aom/av1/common/x86/wiener_convolve_avx2.c b/third_party/aom/av1/common/x86/wiener_convolve_avx2.c
new file mode 100644
index 000000000..87a6e1239
--- /dev/null
+++ b/third_party/aom/av1/common/x86/wiener_convolve_avx2.c
@@ -0,0 +1,261 @@
+/*
+ * 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 <assert.h>
+
+#include "config/av1_rtcd.h"
+
+#include "av1/common/convolve.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/aom_filter.h"
+#include "aom_dsp/x86/synonyms.h"
+#include "aom_dsp/x86/synonyms_avx2.h"
+
+// 128-bit xmmwords are written as [ ... ] with the MSB on the left.
+// 256-bit ymmwords are written as two xmmwords, [ ... ][ ... ] with the MSB
+// on the left.
+// A row of, say, 8-bit pixels with values p0, p1, p2, ..., p30, p31 will be
+// loaded and stored as [ p31 ... p17 p16 ][ p15 ... p1 p0 ].
+void av1_wiener_convolve_add_src_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,
+                                      const ConvolveParams *conv_params) {
+  const int bd = 8;
+  assert(x_step_q4 == 16 && y_step_q4 == 16);
+  assert(!(w & 7));
+  (void)x_step_q4;
+  (void)y_step_q4;
+
+  DECLARE_ALIGNED(32, uint16_t,
+                  temp[(MAX_SB_SIZE + SUBPEL_TAPS - 1) * MAX_SB_SIZE]);
+  int intermediate_height = h + SUBPEL_TAPS - 2;
+  memset(temp + (intermediate_height * MAX_SB_SIZE), 0, MAX_SB_SIZE);
+  const int center_tap = ((SUBPEL_TAPS - 1) / 2);
+  const uint8_t *const src_ptr = src - center_tap * src_stride - center_tap;
+
+  const __m128i zero_128 = _mm_setzero_si128();
+  const __m256i zero_256 = _mm256_setzero_si256();
+
+  // Add an offset to account for the "add_src" part of the convolve function.
+  const __m128i offset = _mm_insert_epi16(zero_128, 1 << FILTER_BITS, 3);
+
+  const __m256i clamp_low = zero_256;
+  const __m256i clamp_high =
+      _mm256_set1_epi16(WIENER_CLAMP_LIMIT(conv_params->round_0, bd) - 1);
+
+  /* Horizontal filter */
+  {
+    // coeffs [ f7 f6 f5 f4 f3 f2 f1 f0 ]
+    const __m128i coeffs_x = _mm_add_epi16(xx_loadu_128(filter_x), offset);
+
+    // coeffs [ f3 f2 f3 f2 f1 f0 f1 f0 ]
+    const __m128i coeffs_0123 = _mm_unpacklo_epi32(coeffs_x, coeffs_x);
+    // coeffs [ f7 f6 f7 f6 f5 f4 f5 f4 ]
+    const __m128i coeffs_4567 = _mm_unpackhi_epi32(coeffs_x, coeffs_x);
+
+    // coeffs [ f1 f0 f1 f0 f1 f0 f1 f0 ]
+    const __m128i coeffs_01_128 = _mm_unpacklo_epi64(coeffs_0123, coeffs_0123);
+    // coeffs [ f3 f2 f3 f2 f3 f2 f3 f2 ]
+    const __m128i coeffs_23_128 = _mm_unpackhi_epi64(coeffs_0123, coeffs_0123);
+    // coeffs [ f5 f4 f5 f4 f5 f4 f5 f4 ]
+    const __m128i coeffs_45_128 = _mm_unpacklo_epi64(coeffs_4567, coeffs_4567);
+    // coeffs [ f7 f6 f7 f6 f7 f6 f7 f6 ]
+    const __m128i coeffs_67_128 = _mm_unpackhi_epi64(coeffs_4567, coeffs_4567);
+
+    // coeffs [ f1 f0 f1 f0 f1 f0 f1 f0 ][ f1 f0 f1 f0 f1 f0 f1 f0 ]
+    const __m256i coeffs_01 = yy_set_m128i(coeffs_01_128, coeffs_01_128);
+    // coeffs [ f3 f2 f3 f2 f3 f2 f3 f2 ][ f3 f2 f3 f2 f3 f2 f3 f2 ]
+    const __m256i coeffs_23 = yy_set_m128i(coeffs_23_128, coeffs_23_128);
+    // coeffs [ f5 f4 f5 f4 f5 f4 f5 f4 ][ f5 f4 f5 f4 f5 f4 f5 f4 ]
+    const __m256i coeffs_45 = yy_set_m128i(coeffs_45_128, coeffs_45_128);
+    // coeffs [ f7 f6 f7 f6 f7 f6 f7 f6 ][ f7 f6 f7 f6 f7 f6 f7 f6 ]
+    const __m256i coeffs_67 = yy_set_m128i(coeffs_67_128, coeffs_67_128);
+
+    const __m256i round_const = _mm256_set1_epi32(
+        (1 << (conv_params->round_0 - 1)) + (1 << (bd + FILTER_BITS - 1)));
+
+    for (int i = 0; i < intermediate_height; ++i) {
+      for (int j = 0; j < w; j += 16) {
+        const uint8_t *data_ij = src_ptr + i * src_stride + j;
+
+        // Load 8-bit src data
+        const __m128i data_0 = xx_loadu_128(data_ij + 0);
+        const __m128i data_1 = xx_loadu_128(data_ij + 1);
+        const __m128i data_2 = xx_loadu_128(data_ij + 2);
+        const __m128i data_3 = xx_loadu_128(data_ij + 3);
+        const __m128i data_4 = xx_loadu_128(data_ij + 4);
+        const __m128i data_5 = xx_loadu_128(data_ij + 5);
+        const __m128i data_6 = xx_loadu_128(data_ij + 6);
+        const __m128i data_7 = xx_loadu_128(data_ij + 7);
+
+        // (Zero-)Extend 8-bit data to 16-bit data
+        const __m256i src_0 = _mm256_cvtepu8_epi16(data_0);
+        const __m256i src_1 = _mm256_cvtepu8_epi16(data_1);
+        const __m256i src_2 = _mm256_cvtepu8_epi16(data_2);
+        const __m256i src_3 = _mm256_cvtepu8_epi16(data_3);
+        const __m256i src_4 = _mm256_cvtepu8_epi16(data_4);
+        const __m256i src_5 = _mm256_cvtepu8_epi16(data_5);
+        const __m256i src_6 = _mm256_cvtepu8_epi16(data_6);
+        const __m256i src_7 = _mm256_cvtepu8_epi16(data_7);
+
+        // Multiply src data by filter coeffs and sum pairs
+        const __m256i res_0 = _mm256_madd_epi16(src_0, coeffs_01);
+        const __m256i res_1 = _mm256_madd_epi16(src_1, coeffs_01);
+        const __m256i res_2 = _mm256_madd_epi16(src_2, coeffs_23);
+        const __m256i res_3 = _mm256_madd_epi16(src_3, coeffs_23);
+        const __m256i res_4 = _mm256_madd_epi16(src_4, coeffs_45);
+        const __m256i res_5 = _mm256_madd_epi16(src_5, coeffs_45);
+        const __m256i res_6 = _mm256_madd_epi16(src_6, coeffs_67);
+        const __m256i res_7 = _mm256_madd_epi16(src_7, coeffs_67);
+
+        // Calculate scalar product for even- and odd-indices separately,
+        // increasing to 32-bit precision
+        const __m256i res_even_sum = _mm256_add_epi32(
+            _mm256_add_epi32(res_0, res_4), _mm256_add_epi32(res_2, res_6));
+        const __m256i res_odd_sum = _mm256_add_epi32(
+            _mm256_add_epi32(res_1, res_5), _mm256_add_epi32(res_3, res_7));
+
+        const __m256i res_even = _mm256_srai_epi32(
+            _mm256_add_epi32(res_even_sum, round_const), conv_params->round_0);
+        const __m256i res_odd = _mm256_srai_epi32(
+            _mm256_add_epi32(res_odd_sum, round_const), conv_params->round_0);
+
+        // Reduce to 16-bit precision and pack even- and odd-index results
+        // back into one register. The _mm256_packs_epi32 intrinsic returns
+        // a register with the pixels ordered as follows:
+        // [ 15 13 11 9 14 12 10 8 ] [ 7 5 3 1 6 4 2 0 ]
+        const __m256i res = _mm256_packs_epi32(res_even, res_odd);
+        const __m256i res_clamped =
+            _mm256_min_epi16(_mm256_max_epi16(res, clamp_low), clamp_high);
+
+        // Store in a temporary array
+        yy_storeu_256(temp + i * MAX_SB_SIZE + j, res_clamped);
+      }
+    }
+  }
+
+  /* Vertical filter */
+  {
+    // coeffs [ g7 g6 g5 g4 g3 g2 g1 g0 ]
+    const __m128i coeffs_y = _mm_add_epi16(xx_loadu_128(filter_y), offset);
+
+    // coeffs [ g3 g2 g3 g2 g1 g0 g1 g0 ]
+    const __m128i coeffs_0123 = _mm_unpacklo_epi32(coeffs_y, coeffs_y);
+    // coeffs [ g7 g6 g7 g6 g5 g4 g5 g4 ]
+    const __m128i coeffs_4567 = _mm_unpackhi_epi32(coeffs_y, coeffs_y);
+
+    // coeffs [ g1 g0 g1 g0 g1 g0 g1 g0 ]
+    const __m128i coeffs_01_128 = _mm_unpacklo_epi64(coeffs_0123, coeffs_0123);
+    // coeffs [ g3 g2 g3 g2 g3 g2 g3 g2 ]
+    const __m128i coeffs_23_128 = _mm_unpackhi_epi64(coeffs_0123, coeffs_0123);
+    // coeffs [ g5 g4 g5 g4 g5 g4 g5 g4 ]
+    const __m128i coeffs_45_128 = _mm_unpacklo_epi64(coeffs_4567, coeffs_4567);
+    // coeffs [ g7 g6 g7 g6 g7 g6 g7 g6 ]
+    const __m128i coeffs_67_128 = _mm_unpackhi_epi64(coeffs_4567, coeffs_4567);
+
+    // coeffs [ g1 g0 g1 g0 g1 g0 g1 g0 ][ g1 g0 g1 g0 g1 g0 g1 g0 ]
+    const __m256i coeffs_01 = yy_set_m128i(coeffs_01_128, coeffs_01_128);
+    // coeffs [ g3 g2 g3 g2 g3 g2 g3 g2 ][ g3 g2 g3 g2 g3 g2 g3 g2 ]
+    const __m256i coeffs_23 = yy_set_m128i(coeffs_23_128, coeffs_23_128);
+    // coeffs [ g5 g4 g5 g4 g5 g4 g5 g4 ][ g5 g4 g5 g4 g5 g4 g5 g4 ]
+    const __m256i coeffs_45 = yy_set_m128i(coeffs_45_128, coeffs_45_128);
+    // coeffs [ g7 g6 g7 g6 g7 g6 g7 g6 ][ g7 g6 g7 g6 g7 g6 g7 g6 ]
+    const __m256i coeffs_67 = yy_set_m128i(coeffs_67_128, coeffs_67_128);
+
+    const __m256i round_const =
+        _mm256_set1_epi32((1 << (conv_params->round_1 - 1)) -
+                          (1 << (bd + conv_params->round_1 - 1)));
+
+    for (int i = 0; i < h; ++i) {
+      for (int j = 0; j < w; j += 16) {
+        const uint16_t *data_ij = temp + i * MAX_SB_SIZE + j;
+
+        // Load 16-bit data from the output of the horizontal filter in
+        // which the pixels are ordered as follows:
+        // [ 15 13 11 9 14 12 10 8 ] [ 7 5 3 1 6 4 2 0 ]
+        const __m256i data_0 = yy_loadu_256(data_ij + 0 * MAX_SB_SIZE);
+        const __m256i data_1 = yy_loadu_256(data_ij + 1 * MAX_SB_SIZE);
+        const __m256i data_2 = yy_loadu_256(data_ij + 2 * MAX_SB_SIZE);
+        const __m256i data_3 = yy_loadu_256(data_ij + 3 * MAX_SB_SIZE);
+        const __m256i data_4 = yy_loadu_256(data_ij + 4 * MAX_SB_SIZE);
+        const __m256i data_5 = yy_loadu_256(data_ij + 5 * MAX_SB_SIZE);
+        const __m256i data_6 = yy_loadu_256(data_ij + 6 * MAX_SB_SIZE);
+        const __m256i data_7 = yy_loadu_256(data_ij + 7 * MAX_SB_SIZE);
+
+        // Filter the even-indices, increasing to 32-bit precision
+        const __m256i src_0 = _mm256_unpacklo_epi16(data_0, data_1);
+        const __m256i src_2 = _mm256_unpacklo_epi16(data_2, data_3);
+        const __m256i src_4 = _mm256_unpacklo_epi16(data_4, data_5);
+        const __m256i src_6 = _mm256_unpacklo_epi16(data_6, data_7);
+
+        const __m256i res_0 = _mm256_madd_epi16(src_0, coeffs_01);
+        const __m256i res_2 = _mm256_madd_epi16(src_2, coeffs_23);
+        const __m256i res_4 = _mm256_madd_epi16(src_4, coeffs_45);
+        const __m256i res_6 = _mm256_madd_epi16(src_6, coeffs_67);
+
+        const __m256i res_even = _mm256_add_epi32(
+            _mm256_add_epi32(res_0, res_2), _mm256_add_epi32(res_4, res_6));
+
+        // Filter the odd-indices, increasing to 32-bit precision
+        const __m256i src_1 = _mm256_unpackhi_epi16(data_0, data_1);
+        const __m256i src_3 = _mm256_unpackhi_epi16(data_2, data_3);
+        const __m256i src_5 = _mm256_unpackhi_epi16(data_4, data_5);
+        const __m256i src_7 = _mm256_unpackhi_epi16(data_6, data_7);
+
+        const __m256i res_1 = _mm256_madd_epi16(src_1, coeffs_01);
+        const __m256i res_3 = _mm256_madd_epi16(src_3, coeffs_23);
+        const __m256i res_5 = _mm256_madd_epi16(src_5, coeffs_45);
+        const __m256i res_7 = _mm256_madd_epi16(src_7, coeffs_67);
+
+        const __m256i res_odd = _mm256_add_epi32(
+            _mm256_add_epi32(res_1, res_3), _mm256_add_epi32(res_5, res_7));
+
+        // Pixels are currently in the following order:
+        // res_even order: [ 14 12 10 8 ] [ 6 4 2 0 ]
+        // res_odd order:  [ 15 13 11 9 ] [ 7 5 3 1 ]
+        //
+        // Rearrange the pixels into the following order:
+        // res_lo order: [ 11 10  9  8 ] [ 3 2 1 0 ]
+        // res_hi order: [ 15 14 13 12 ] [ 7 6 5 4 ]
+        const __m256i res_lo = _mm256_unpacklo_epi32(res_even, res_odd);
+        const __m256i res_hi = _mm256_unpackhi_epi32(res_even, res_odd);
+
+        const __m256i res_lo_round = _mm256_srai_epi32(
+            _mm256_add_epi32(res_lo, round_const), conv_params->round_1);
+        const __m256i res_hi_round = _mm256_srai_epi32(
+            _mm256_add_epi32(res_hi, round_const), conv_params->round_1);
+
+        // Reduce to 16-bit precision and pack into the correct order:
+        // [ 15 14 13 12 11 10 9 8 ][ 7 6 5 4 3 2 1 0 ]
+        const __m256i res_16bit =
+            _mm256_packs_epi32(res_lo_round, res_hi_round);
+
+        // Reduce to 8-bit precision. This messes up the order:
+        // [ - - - - - - - - 15 14 13 12 11 10 9 8 ]
+        // [ - - - - - - - - 7 6 5 4 3 2 1 0 ]
+        const __m256i res_8bit =
+            _mm256_packus_epi16(res_16bit, zero_256 /* don't care value */);
+
+        // Swap the two central 32-bit values to get the order:
+        // [ - - - - - - - - - - - - - - - - ]
+        // [ 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0 ]
+        const __m256i res_8bit2 = _mm256_permute4x64_epi64(res_8bit, 0xd8);
+
+        // Store the lower 128-bit lane in the dst array
+        xx_storeu_128(dst + i * dst_stride + j,
+                      _mm256_castsi256_si128(res_8bit2));
+      }
+    }
+  }
+}
diff --git a/third_party/aom/av1/common/x86/wiener_convolve_sse2.c b/third_party/aom/av1/common/x86/wiener_convolve_sse2.c
new file mode 100644
index 000000000..f9d00b733
--- /dev/null
+++ b/third_party/aom/av1/common/x86/wiener_convolve_sse2.c
@@ -0,0 +1,199 @@
+/*
+ * 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 <assert.h>
+
+#include "config/av1_rtcd.h"
+
+#include "av1/common/convolve.h"
+#include "aom_dsp/aom_dsp_common.h"
+#include "aom_dsp/aom_filter.h"
+
+void av1_wiener_convolve_add_src_sse2(const uint8_t *src, ptrdiff_t src_stride,
+                                      uint8_t *dst, ptrdiff_t dst_stride,
+                                      const int16_t *filter_x, int x_step_q4,
+                                      const int16_t *filter_y, int y_step_q4,
+                                      int w, int h,
+                                      const ConvolveParams *conv_params) {
+  const int bd = 8;
+  assert(x_step_q4 == 16 && y_step_q4 == 16);
+  assert(!(w & 7));
+  (void)x_step_q4;
+  (void)y_step_q4;
+
+  DECLARE_ALIGNED(16, uint16_t,
+                  temp[(MAX_SB_SIZE + SUBPEL_TAPS - 1) * MAX_SB_SIZE]);
+  int intermediate_height = h + SUBPEL_TAPS - 2;
+  memset(temp + (intermediate_height * MAX_SB_SIZE), 0, MAX_SB_SIZE);
+  int i, j;
+  const int center_tap = ((SUBPEL_TAPS - 1) / 2);
+  const uint8_t *const src_ptr = src - center_tap * src_stride - center_tap;
+
+  const __m128i zero = _mm_setzero_si128();
+  // Add an offset to account for the "add_src" part of the convolve function.
+  const __m128i offset = _mm_insert_epi16(zero, 1 << FILTER_BITS, 3);
+
+  /* Horizontal filter */
+  {
+    const __m128i coeffs_x =
+        _mm_add_epi16(_mm_loadu_si128((__m128i *)filter_x), offset);
+
+    // coeffs 0 1 0 1 2 3 2 3
+    const __m128i tmp_0 = _mm_unpacklo_epi32(coeffs_x, coeffs_x);
+    // coeffs 4 5 4 5 6 7 6 7
+    const __m128i tmp_1 = _mm_unpackhi_epi32(coeffs_x, coeffs_x);
+
+    // coeffs 0 1 0 1 0 1 0 1
+    const __m128i coeff_01 = _mm_unpacklo_epi64(tmp_0, tmp_0);
+    // coeffs 2 3 2 3 2 3 2 3
+    const __m128i coeff_23 = _mm_unpackhi_epi64(tmp_0, tmp_0);
+    // coeffs 4 5 4 5 4 5 4 5
+    const __m128i coeff_45 = _mm_unpacklo_epi64(tmp_1, tmp_1);
+    // coeffs 6 7 6 7 6 7 6 7
+    const __m128i coeff_67 = _mm_unpackhi_epi64(tmp_1, tmp_1);
+
+    const __m128i round_const = _mm_set1_epi32(
+        (1 << (conv_params->round_0 - 1)) + (1 << (bd + FILTER_BITS - 1)));
+
+    for (i = 0; i < intermediate_height; ++i) {
+      for (j = 0; j < w; j += 8) {
+        const __m128i data =
+            _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + j]);
+
+        // Filter even-index pixels
+        const __m128i src_0 = _mm_unpacklo_epi8(data, zero);
+        const __m128i res_0 = _mm_madd_epi16(src_0, coeff_01);
+        const __m128i src_2 = _mm_unpacklo_epi8(_mm_srli_si128(data, 2), zero);
+        const __m128i res_2 = _mm_madd_epi16(src_2, coeff_23);
+        const __m128i src_4 = _mm_unpacklo_epi8(_mm_srli_si128(data, 4), zero);
+        const __m128i res_4 = _mm_madd_epi16(src_4, coeff_45);
+        const __m128i src_6 = _mm_unpacklo_epi8(_mm_srli_si128(data, 6), zero);
+        const __m128i res_6 = _mm_madd_epi16(src_6, coeff_67);
+
+        __m128i res_even = _mm_add_epi32(_mm_add_epi32(res_0, res_4),
+                                         _mm_add_epi32(res_2, res_6));
+        res_even = _mm_srai_epi32(_mm_add_epi32(res_even, round_const),
+                                  conv_params->round_0);
+
+        // Filter odd-index pixels
+        const __m128i src_1 = _mm_unpacklo_epi8(_mm_srli_si128(data, 1), zero);
+        const __m128i res_1 = _mm_madd_epi16(src_1, coeff_01);
+        const __m128i src_3 = _mm_unpacklo_epi8(_mm_srli_si128(data, 3), zero);
+        const __m128i res_3 = _mm_madd_epi16(src_3, coeff_23);
+        const __m128i src_5 = _mm_unpacklo_epi8(_mm_srli_si128(data, 5), zero);
+        const __m128i res_5 = _mm_madd_epi16(src_5, coeff_45);
+        const __m128i src_7 = _mm_unpacklo_epi8(_mm_srli_si128(data, 7), zero);
+        const __m128i res_7 = _mm_madd_epi16(src_7, coeff_67);
+
+        __m128i res_odd = _mm_add_epi32(_mm_add_epi32(res_1, res_5),
+                                        _mm_add_epi32(res_3, res_7));
+        res_odd = _mm_srai_epi32(_mm_add_epi32(res_odd, round_const),
+                                 conv_params->round_0);
+
+        // Pack in the column order 0, 2, 4, 6, 1, 3, 5, 7
+        __m128i res = _mm_packs_epi32(res_even, res_odd);
+        res = _mm_min_epi16(
+            _mm_max_epi16(res, zero),
+            _mm_set1_epi16(WIENER_CLAMP_LIMIT(conv_params->round_0, bd) - 1));
+        _mm_storeu_si128((__m128i *)&temp[i * MAX_SB_SIZE + j], res);
+      }
+    }
+  }
+
+  /* Vertical filter */
+  {
+    const __m128i coeffs_y =
+        _mm_add_epi16(_mm_loadu_si128((__m128i *)filter_y), offset);
+
+    // coeffs 0 1 0 1 2 3 2 3
+    const __m128i tmp_0 = _mm_unpacklo_epi32(coeffs_y, coeffs_y);
+    // coeffs 4 5 4 5 6 7 6 7
+    const __m128i tmp_1 = _mm_unpackhi_epi32(coeffs_y, coeffs_y);
+
+    // coeffs 0 1 0 1 0 1 0 1
+    const __m128i coeff_01 = _mm_unpacklo_epi64(tmp_0, tmp_0);
+    // coeffs 2 3 2 3 2 3 2 3
+    const __m128i coeff_23 = _mm_unpackhi_epi64(tmp_0, tmp_0);
+    // coeffs 4 5 4 5 4 5 4 5
+    const __m128i coeff_45 = _mm_unpacklo_epi64(tmp_1, tmp_1);
+    // coeffs 6 7 6 7 6 7 6 7
+    const __m128i coeff_67 = _mm_unpackhi_epi64(tmp_1, tmp_1);
+
+    const __m128i round_const =
+        _mm_set1_epi32((1 << (conv_params->round_1 - 1)) -
+                       (1 << (bd + conv_params->round_1 - 1)));
+
+    for (i = 0; i < h; ++i) {
+      for (j = 0; j < w; j += 8) {
+        // Filter even-index pixels
+        const uint16_t *data = &temp[i * MAX_SB_SIZE + j];
+        const __m128i src_0 =
+            _mm_unpacklo_epi16(*(__m128i *)(data + 0 * MAX_SB_SIZE),
+                               *(__m128i *)(data + 1 * MAX_SB_SIZE));
+        const __m128i src_2 =
+            _mm_unpacklo_epi16(*(__m128i *)(data + 2 * MAX_SB_SIZE),
+                               *(__m128i *)(data + 3 * MAX_SB_SIZE));
+        const __m128i src_4 =
+            _mm_unpacklo_epi16(*(__m128i *)(data + 4 * MAX_SB_SIZE),
+                               *(__m128i *)(data + 5 * MAX_SB_SIZE));
+        const __m128i src_6 =
+            _mm_unpacklo_epi16(*(__m128i *)(data + 6 * MAX_SB_SIZE),
+                               *(__m128i *)(data + 7 * MAX_SB_SIZE));
+
+        const __m128i res_0 = _mm_madd_epi16(src_0, coeff_01);
+        const __m128i res_2 = _mm_madd_epi16(src_2, coeff_23);
+        const __m128i res_4 = _mm_madd_epi16(src_4, coeff_45);
+        const __m128i res_6 = _mm_madd_epi16(src_6, coeff_67);
+
+        const __m128i res_even = _mm_add_epi32(_mm_add_epi32(res_0, res_2),
+                                               _mm_add_epi32(res_4, res_6));
+
+        // Filter odd-index pixels
+        const __m128i src_1 =
+            _mm_unpackhi_epi16(*(__m128i *)(data + 0 * MAX_SB_SIZE),
+                               *(__m128i *)(data + 1 * MAX_SB_SIZE));
+        const __m128i src_3 =
+            _mm_unpackhi_epi16(*(__m128i *)(data + 2 * MAX_SB_SIZE),
+                               *(__m128i *)(data + 3 * MAX_SB_SIZE));
+        const __m128i src_5 =
+            _mm_unpackhi_epi16(*(__m128i *)(data + 4 * MAX_SB_SIZE),
+                               *(__m128i *)(data + 5 * MAX_SB_SIZE));
+        const __m128i src_7 =
+            _mm_unpackhi_epi16(*(__m128i *)(data + 6 * MAX_SB_SIZE),
+                               *(__m128i *)(data + 7 * MAX_SB_SIZE));
+
+        const __m128i res_1 = _mm_madd_epi16(src_1, coeff_01);
+        const __m128i res_3 = _mm_madd_epi16(src_3, coeff_23);
+        const __m128i res_5 = _mm_madd_epi16(src_5, coeff_45);
+        const __m128i res_7 = _mm_madd_epi16(src_7, coeff_67);
+
+        const __m128i res_odd = _mm_add_epi32(_mm_add_epi32(res_1, res_3),
+                                              _mm_add_epi32(res_5, res_7));
+
+        // Rearrange pixels back into the order 0 ... 7
+        const __m128i res_lo = _mm_unpacklo_epi32(res_even, res_odd);
+        const __m128i res_hi = _mm_unpackhi_epi32(res_even, res_odd);
+
+        const __m128i res_lo_round = _mm_srai_epi32(
+            _mm_add_epi32(res_lo, round_const), conv_params->round_1);
+        const __m128i res_hi_round = _mm_srai_epi32(
+            _mm_add_epi32(res_hi, round_const), conv_params->round_1);
+
+        const __m128i res_16bit = _mm_packs_epi32(res_lo_round, res_hi_round);
+        __m128i res_8bit = _mm_packus_epi16(res_16bit, res_16bit);
+
+        __m128i *const p = (__m128i *)&dst[i * dst_stride + j];
+        _mm_storel_epi64(p, res_8bit);
+      }
+    }
+  }
+}