Add m-esr52 at 52.6.0

26 files changed, 16845 insertions, 0 deletions

diff --git a/media/libvpx/vp9/encoder/x86/vp9_avg_intrin_sse2.c b/media/libvpx/vp9/encoder/x86/vp9_avg_intrin_sse2.c
new file mode 100644
index 000000000..56a91ed2d
--- /dev/null
+++ b/media/libvpx/vp9/encoder/x86/vp9_avg_intrin_sse2.c
@@ -0,0 +1,423 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <emmintrin.h>
+
+#include "./vp9_rtcd.h"
+#include "vpx_ports/mem.h"
+
+void vp9_minmax_8x8_sse2(const uint8_t *s, int p, const uint8_t *d, int dp,
+                         int *min, int *max) {
+  __m128i u0, s0, d0, diff, maxabsdiff, minabsdiff, negdiff, absdiff0, absdiff;
+  u0  = _mm_setzero_si128();
+  // Row 0
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s)), u0);
+  d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d)), u0);
+  diff = _mm_subs_epi16(s0, d0);
+  negdiff = _mm_subs_epi16(u0, diff);
+  absdiff0 = _mm_max_epi16(diff, negdiff);
+  // Row 1
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + p)), u0);
+  d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + dp)), u0);
+  diff = _mm_subs_epi16(s0, d0);
+  negdiff = _mm_subs_epi16(u0, diff);
+  absdiff = _mm_max_epi16(diff, negdiff);
+  maxabsdiff = _mm_max_epi16(absdiff0, absdiff);
+  minabsdiff = _mm_min_epi16(absdiff0, absdiff);
+  // Row 2
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 2 * p)), u0);
+  d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 2 * dp)), u0);
+  diff = _mm_subs_epi16(s0, d0);
+  negdiff = _mm_subs_epi16(u0, diff);
+  absdiff = _mm_max_epi16(diff, negdiff);
+  maxabsdiff = _mm_max_epi16(maxabsdiff, absdiff);
+  minabsdiff = _mm_min_epi16(minabsdiff, absdiff);
+  // Row 3
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 3 * p)), u0);
+  d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 3 * dp)), u0);
+  diff = _mm_subs_epi16(s0, d0);
+  negdiff = _mm_subs_epi16(u0, diff);
+  absdiff = _mm_max_epi16(diff, negdiff);
+  maxabsdiff = _mm_max_epi16(maxabsdiff, absdiff);
+  minabsdiff = _mm_min_epi16(minabsdiff, absdiff);
+  // Row 4
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 4 * p)), u0);
+  d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 4 * dp)), u0);
+  diff = _mm_subs_epi16(s0, d0);
+  negdiff = _mm_subs_epi16(u0, diff);
+  absdiff = _mm_max_epi16(diff, negdiff);
+  maxabsdiff = _mm_max_epi16(maxabsdiff, absdiff);
+  minabsdiff = _mm_min_epi16(minabsdiff, absdiff);
+  // Row 5
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 5 * p)), u0);
+  d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 5 * dp)), u0);
+  diff = _mm_subs_epi16(s0, d0);
+  negdiff = _mm_subs_epi16(u0, diff);
+  absdiff = _mm_max_epi16(diff, negdiff);
+  maxabsdiff = _mm_max_epi16(maxabsdiff, absdiff);
+  minabsdiff = _mm_min_epi16(minabsdiff, absdiff);
+  // Row 6
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 6 * p)), u0);
+  d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 6 * dp)), u0);
+  diff = _mm_subs_epi16(s0, d0);
+  negdiff = _mm_subs_epi16(u0, diff);
+  absdiff = _mm_max_epi16(diff, negdiff);
+  maxabsdiff = _mm_max_epi16(maxabsdiff, absdiff);
+  minabsdiff = _mm_min_epi16(minabsdiff, absdiff);
+  // Row 7
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 7 * p)), u0);
+  d0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(d + 7 * dp)), u0);
+  diff = _mm_subs_epi16(s0, d0);
+  negdiff = _mm_subs_epi16(u0, diff);
+  absdiff = _mm_max_epi16(diff, negdiff);
+  maxabsdiff = _mm_max_epi16(maxabsdiff, absdiff);
+  minabsdiff = _mm_min_epi16(minabsdiff, absdiff);
+
+  maxabsdiff = _mm_max_epi16(maxabsdiff, _mm_srli_si128(maxabsdiff, 8));
+  maxabsdiff = _mm_max_epi16(maxabsdiff, _mm_srli_epi64(maxabsdiff, 32));
+  maxabsdiff = _mm_max_epi16(maxabsdiff, _mm_srli_epi64(maxabsdiff, 16));
+  *max = _mm_extract_epi16(maxabsdiff, 0);
+
+  minabsdiff = _mm_min_epi16(minabsdiff, _mm_srli_si128(minabsdiff, 8));
+  minabsdiff = _mm_min_epi16(minabsdiff, _mm_srli_epi64(minabsdiff, 32));
+  minabsdiff = _mm_min_epi16(minabsdiff, _mm_srli_epi64(minabsdiff, 16));
+  *min = _mm_extract_epi16(minabsdiff, 0);
+}
+
+unsigned int vp9_avg_8x8_sse2(const uint8_t *s, int p) {
+  __m128i s0, s1, u0;
+  unsigned int avg = 0;
+  u0  = _mm_setzero_si128();
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s)), u0);
+  s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + p)), u0);
+  s0 = _mm_adds_epu16(s0, s1);
+  s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 2 * p)), u0);
+  s0 = _mm_adds_epu16(s0, s1);
+  s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 3 * p)), u0);
+  s0 = _mm_adds_epu16(s0, s1);
+  s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 4 * p)), u0);
+  s0 = _mm_adds_epu16(s0, s1);
+  s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 5 * p)), u0);
+  s0 = _mm_adds_epu16(s0, s1);
+  s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 6 * p)), u0);
+  s0 = _mm_adds_epu16(s0, s1);
+  s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 7 * p)), u0);
+  s0 = _mm_adds_epu16(s0, s1);
+
+  s0 = _mm_adds_epu16(s0, _mm_srli_si128(s0, 8));
+  s0 = _mm_adds_epu16(s0, _mm_srli_epi64(s0, 32));
+  s0 = _mm_adds_epu16(s0, _mm_srli_epi64(s0, 16));
+  avg = _mm_extract_epi16(s0, 0);
+  return (avg + 32) >> 6;
+}
+
+unsigned int vp9_avg_4x4_sse2(const uint8_t *s, int p) {
+  __m128i s0, s1, u0;
+  unsigned int avg = 0;
+  u0  = _mm_setzero_si128();
+  s0 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s)), u0);
+  s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + p)), u0);
+  s0 = _mm_adds_epu16(s0, s1);
+  s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 2 * p)), u0);
+  s0 = _mm_adds_epu16(s0, s1);
+  s1 = _mm_unpacklo_epi8(_mm_loadl_epi64((const __m128i *)(s + 3 * p)), u0);
+  s0 = _mm_adds_epu16(s0, s1);
+
+  s0 = _mm_adds_epu16(s0, _mm_srli_si128(s0, 4));
+  s0 = _mm_adds_epu16(s0, _mm_srli_epi64(s0, 16));
+  avg = _mm_extract_epi16(s0, 0);
+  return (avg + 8) >> 4;
+}
+
+static void hadamard_col8_sse2(__m128i *in, int iter) {
+  __m128i a0 = in[0];
+  __m128i a1 = in[1];
+  __m128i a2 = in[2];
+  __m128i a3 = in[3];
+  __m128i a4 = in[4];
+  __m128i a5 = in[5];
+  __m128i a6 = in[6];
+  __m128i a7 = in[7];
+
+  __m128i b0 = _mm_add_epi16(a0, a1);
+  __m128i b1 = _mm_sub_epi16(a0, a1);
+  __m128i b2 = _mm_add_epi16(a2, a3);
+  __m128i b3 = _mm_sub_epi16(a2, a3);
+  __m128i b4 = _mm_add_epi16(a4, a5);
+  __m128i b5 = _mm_sub_epi16(a4, a5);
+  __m128i b6 = _mm_add_epi16(a6, a7);
+  __m128i b7 = _mm_sub_epi16(a6, a7);
+
+  a0 = _mm_add_epi16(b0, b2);
+  a1 = _mm_add_epi16(b1, b3);
+  a2 = _mm_sub_epi16(b0, b2);
+  a3 = _mm_sub_epi16(b1, b3);
+  a4 = _mm_add_epi16(b4, b6);
+  a5 = _mm_add_epi16(b5, b7);
+  a6 = _mm_sub_epi16(b4, b6);
+  a7 = _mm_sub_epi16(b5, b7);
+
+  if (iter == 0) {
+    b0 = _mm_add_epi16(a0, a4);
+    b7 = _mm_add_epi16(a1, a5);
+    b3 = _mm_add_epi16(a2, a6);
+    b4 = _mm_add_epi16(a3, a7);
+    b2 = _mm_sub_epi16(a0, a4);
+    b6 = _mm_sub_epi16(a1, a5);
+    b1 = _mm_sub_epi16(a2, a6);
+    b5 = _mm_sub_epi16(a3, a7);
+
+    a0 = _mm_unpacklo_epi16(b0, b1);
+    a1 = _mm_unpacklo_epi16(b2, b3);
+    a2 = _mm_unpackhi_epi16(b0, b1);
+    a3 = _mm_unpackhi_epi16(b2, b3);
+    a4 = _mm_unpacklo_epi16(b4, b5);
+    a5 = _mm_unpacklo_epi16(b6, b7);
+    a6 = _mm_unpackhi_epi16(b4, b5);
+    a7 = _mm_unpackhi_epi16(b6, b7);
+
+    b0 = _mm_unpacklo_epi32(a0, a1);
+    b1 = _mm_unpacklo_epi32(a4, a5);
+    b2 = _mm_unpackhi_epi32(a0, a1);
+    b3 = _mm_unpackhi_epi32(a4, a5);
+    b4 = _mm_unpacklo_epi32(a2, a3);
+    b5 = _mm_unpacklo_epi32(a6, a7);
+    b6 = _mm_unpackhi_epi32(a2, a3);
+    b7 = _mm_unpackhi_epi32(a6, a7);
+
+    in[0] = _mm_unpacklo_epi64(b0, b1);
+    in[1] = _mm_unpackhi_epi64(b0, b1);
+    in[2] = _mm_unpacklo_epi64(b2, b3);
+    in[3] = _mm_unpackhi_epi64(b2, b3);
+    in[4] = _mm_unpacklo_epi64(b4, b5);
+    in[5] = _mm_unpackhi_epi64(b4, b5);
+    in[6] = _mm_unpacklo_epi64(b6, b7);
+    in[7] = _mm_unpackhi_epi64(b6, b7);
+  } else {
+    in[0] = _mm_add_epi16(a0, a4);
+    in[7] = _mm_add_epi16(a1, a5);
+    in[3] = _mm_add_epi16(a2, a6);
+    in[4] = _mm_add_epi16(a3, a7);
+    in[2] = _mm_sub_epi16(a0, a4);
+    in[6] = _mm_sub_epi16(a1, a5);
+    in[1] = _mm_sub_epi16(a2, a6);
+    in[5] = _mm_sub_epi16(a3, a7);
+  }
+}
+
+void vp9_hadamard_8x8_sse2(int16_t const *src_diff, int src_stride,
+                           int16_t *coeff) {
+  __m128i src[8];
+  src[0] = _mm_load_si128((const __m128i *)src_diff);
+  src[1] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
+  src[2] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
+  src[3] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
+  src[4] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
+  src[5] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
+  src[6] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
+  src[7] = _mm_load_si128((const __m128i *)(src_diff += src_stride));
+
+  hadamard_col8_sse2(src, 0);
+  hadamard_col8_sse2(src, 1);
+
+  _mm_store_si128((__m128i *)coeff, src[0]);
+  coeff += 8;
+  _mm_store_si128((__m128i *)coeff, src[1]);
+  coeff += 8;
+  _mm_store_si128((__m128i *)coeff, src[2]);
+  coeff += 8;
+  _mm_store_si128((__m128i *)coeff, src[3]);
+  coeff += 8;
+  _mm_store_si128((__m128i *)coeff, src[4]);
+  coeff += 8;
+  _mm_store_si128((__m128i *)coeff, src[5]);
+  coeff += 8;
+  _mm_store_si128((__m128i *)coeff, src[6]);
+  coeff += 8;
+  _mm_store_si128((__m128i *)coeff, src[7]);
+}
+
+void vp9_hadamard_16x16_sse2(int16_t const *src_diff, int src_stride,
+                             int16_t *coeff) {
+  int idx;
+  for (idx = 0; idx < 4; ++idx) {
+    int16_t const *src_ptr = src_diff + (idx >> 1) * 8 * src_stride
+                                + (idx & 0x01) * 8;
+    vp9_hadamard_8x8_sse2(src_ptr, src_stride, coeff + idx * 64);
+  }
+
+  for (idx = 0; idx < 64; idx += 8) {
+    __m128i coeff0 = _mm_load_si128((const __m128i *)coeff);
+    __m128i coeff1 = _mm_load_si128((const __m128i *)(coeff + 64));
+    __m128i coeff2 = _mm_load_si128((const __m128i *)(coeff + 128));
+    __m128i coeff3 = _mm_load_si128((const __m128i *)(coeff + 192));
+
+    __m128i b0 = _mm_add_epi16(coeff0, coeff1);
+    __m128i b1 = _mm_sub_epi16(coeff0, coeff1);
+    __m128i b2 = _mm_add_epi16(coeff2, coeff3);
+    __m128i b3 = _mm_sub_epi16(coeff2, coeff3);
+
+    coeff0 = _mm_add_epi16(b0, b2);
+    coeff1 = _mm_add_epi16(b1, b3);
+    coeff0 = _mm_srai_epi16(coeff0, 1);
+    coeff1 = _mm_srai_epi16(coeff1, 1);
+    _mm_store_si128((__m128i *)coeff, coeff0);
+    _mm_store_si128((__m128i *)(coeff + 64), coeff1);
+
+    coeff2 = _mm_sub_epi16(b0, b2);
+    coeff3 = _mm_sub_epi16(b1, b3);
+    coeff2 = _mm_srai_epi16(coeff2, 1);
+    coeff3 = _mm_srai_epi16(coeff3, 1);
+    _mm_store_si128((__m128i *)(coeff + 128), coeff2);
+    _mm_store_si128((__m128i *)(coeff + 192), coeff3);
+
+    coeff += 8;
+  }
+}
+
+int16_t vp9_satd_sse2(const int16_t *coeff, int length) {
+  int i;
+  __m128i sum = _mm_load_si128((const __m128i *)coeff);
+  __m128i sign = _mm_srai_epi16(sum, 15);
+  __m128i val = _mm_xor_si128(sum, sign);
+  sum = _mm_sub_epi16(val, sign);
+  coeff += 8;
+
+  for (i = 8; i < length; i += 8) {
+    __m128i src_line = _mm_load_si128((const __m128i *)coeff);
+    sign = _mm_srai_epi16(src_line, 15);
+    val = _mm_xor_si128(src_line, sign);
+    val = _mm_sub_epi16(val, sign);
+    sum = _mm_add_epi16(sum, val);
+    coeff += 8;
+  }
+
+  val = _mm_srli_si128(sum, 8);
+  sum = _mm_add_epi16(sum, val);
+  val = _mm_srli_epi64(sum, 32);
+  sum = _mm_add_epi16(sum, val);
+  val = _mm_srli_epi32(sum, 16);
+  sum = _mm_add_epi16(sum, val);
+
+  return _mm_extract_epi16(sum, 0);
+}
+
+void vp9_int_pro_row_sse2(int16_t *hbuf, uint8_t const*ref,
+                          const int ref_stride, const int height) {
+  int idx;
+  __m128i zero = _mm_setzero_si128();
+  __m128i src_line = _mm_loadu_si128((const __m128i *)ref);
+  __m128i s0 = _mm_unpacklo_epi8(src_line, zero);
+  __m128i s1 = _mm_unpackhi_epi8(src_line, zero);
+  __m128i t0, t1;
+  int height_1 = height - 1;
+  ref += ref_stride;
+
+  for (idx = 1; idx < height_1; idx += 2) {
+    src_line = _mm_loadu_si128((const __m128i *)ref);
+    t0 = _mm_unpacklo_epi8(src_line, zero);
+    t1 = _mm_unpackhi_epi8(src_line, zero);
+    s0 = _mm_adds_epu16(s0, t0);
+    s1 = _mm_adds_epu16(s1, t1);
+    ref += ref_stride;
+
+    src_line = _mm_loadu_si128((const __m128i *)ref);
+    t0 = _mm_unpacklo_epi8(src_line, zero);
+    t1 = _mm_unpackhi_epi8(src_line, zero);
+    s0 = _mm_adds_epu16(s0, t0);
+    s1 = _mm_adds_epu16(s1, t1);
+    ref += ref_stride;
+  }
+
+  src_line = _mm_loadu_si128((const __m128i *)ref);
+  t0 = _mm_unpacklo_epi8(src_line, zero);
+  t1 = _mm_unpackhi_epi8(src_line, zero);
+  s0 = _mm_adds_epu16(s0, t0);
+  s1 = _mm_adds_epu16(s1, t1);
+
+  if (height == 64) {
+    s0 = _mm_srai_epi16(s0, 5);
+    s1 = _mm_srai_epi16(s1, 5);
+  } else if (height == 32) {
+    s0 = _mm_srai_epi16(s0, 4);
+    s1 = _mm_srai_epi16(s1, 4);
+  } else {
+    s0 = _mm_srai_epi16(s0, 3);
+    s1 = _mm_srai_epi16(s1, 3);
+  }
+
+  _mm_storeu_si128((__m128i *)hbuf, s0);
+  hbuf += 8;
+  _mm_storeu_si128((__m128i *)hbuf, s1);
+}
+
+int16_t vp9_int_pro_col_sse2(uint8_t const *ref, const int width) {
+  __m128i zero = _mm_setzero_si128();
+  __m128i src_line = _mm_load_si128((const __m128i *)ref);
+  __m128i s0 = _mm_sad_epu8(src_line, zero);
+  __m128i s1;
+  int i;
+
+  for (i = 16; i < width; i += 16) {
+    ref += 16;
+    src_line = _mm_load_si128((const __m128i *)ref);
+    s1 = _mm_sad_epu8(src_line, zero);
+    s0 = _mm_adds_epu16(s0, s1);
+  }
+
+  s1 = _mm_srli_si128(s0, 8);
+  s0 = _mm_adds_epu16(s0, s1);
+
+  return _mm_extract_epi16(s0, 0);
+}
+
+int vp9_vector_var_sse2(int16_t const *ref, int16_t const *src,
+                        const int bwl) {
+  int idx;
+  int width = 4 << bwl;
+  int16_t mean;
+  __m128i v0 = _mm_loadu_si128((const __m128i *)ref);
+  __m128i v1 = _mm_load_si128((const __m128i *)src);
+  __m128i diff = _mm_subs_epi16(v0, v1);
+  __m128i sum = diff;
+  __m128i sse = _mm_madd_epi16(diff, diff);
+
+  ref += 8;
+  src += 8;
+
+  for (idx = 8; idx < width; idx += 8) {
+    v0 = _mm_loadu_si128((const __m128i *)ref);
+    v1 = _mm_load_si128((const __m128i *)src);
+    diff = _mm_subs_epi16(v0, v1);
+
+    sum = _mm_add_epi16(sum, diff);
+    v0  = _mm_madd_epi16(diff, diff);
+    sse = _mm_add_epi32(sse, v0);
+
+    ref += 8;
+    src += 8;
+  }
+
+  v0  = _mm_srli_si128(sum, 8);
+  sum = _mm_add_epi16(sum, v0);
+  v0  = _mm_srli_epi64(sum, 32);
+  sum = _mm_add_epi16(sum, v0);
+  v0  = _mm_srli_epi32(sum, 16);
+  sum = _mm_add_epi16(sum, v0);
+
+  v1  = _mm_srli_si128(sse, 8);
+  sse = _mm_add_epi32(sse, v1);
+  v1  = _mm_srli_epi64(sse, 32);
+  sse = _mm_add_epi32(sse, v1);
+
+  mean = _mm_extract_epi16(sum, 0);
+
+  return _mm_cvtsi128_si32(sse) - ((mean * mean) >> (bwl + 2));
+}
diff --git a/media/libvpx/vp9/encoder/x86/vp9_dct32x32_avx2_impl.h b/media/libvpx/vp9/encoder/x86/vp9_dct32x32_avx2_impl.h
new file mode 100644
index 000000000..ae6bfe5fa
--- /dev/null
+++ b/media/libvpx/vp9/encoder/x86/vp9_dct32x32_avx2_impl.h
@@ -0,0 +1,2713 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <immintrin.h>  // AVX2
+
+#include "./vp9_rtcd.h"
+#include "vp9/common/vp9_idct.h"  // for cospi constants
+#include "vpx_ports/mem.h"
+
+#define pair256_set_epi16(a, b) \
+  _mm256_set_epi16((int16_t)(b), (int16_t)(a), (int16_t)(b), (int16_t)(a), \
+                   (int16_t)(b), (int16_t)(a), (int16_t)(b), (int16_t)(a), \
+                   (int16_t)(b), (int16_t)(a), (int16_t)(b), (int16_t)(a), \
+                   (int16_t)(b), (int16_t)(a), (int16_t)(b), (int16_t)(a))
+
+#define pair256_set_epi32(a, b) \
+  _mm256_set_epi32((int)(b), (int)(a), (int)(b), (int)(a), \
+                   (int)(b), (int)(a), (int)(b), (int)(a))
+
+#if FDCT32x32_HIGH_PRECISION
+static INLINE __m256i k_madd_epi32_avx2(__m256i a, __m256i b) {
+  __m256i buf0, buf1;
+  buf0 = _mm256_mul_epu32(a, b);
+  a = _mm256_srli_epi64(a, 32);
+  b = _mm256_srli_epi64(b, 32);
+  buf1 = _mm256_mul_epu32(a, b);
+  return _mm256_add_epi64(buf0, buf1);
+}
+
+static INLINE __m256i k_packs_epi64_avx2(__m256i a, __m256i b) {
+  __m256i buf0 = _mm256_shuffle_epi32(a, _MM_SHUFFLE(0, 0, 2, 0));
+  __m256i buf1 = _mm256_shuffle_epi32(b, _MM_SHUFFLE(0, 0, 2, 0));
+  return _mm256_unpacklo_epi64(buf0, buf1);
+}
+#endif
+
+void FDCT32x32_2D_AVX2(const int16_t *input,
+                  int16_t *output_org, int stride) {
+  // Calculate pre-multiplied strides
+  const int str1 = stride;
+  const int str2 = 2 * stride;
+  const int str3 = 2 * stride + str1;
+  // We need an intermediate buffer between passes.
+  DECLARE_ALIGNED(32, int16_t, intermediate[32 * 32]);
+  // Constants
+  //    When we use them, in one case, they are all the same. In all others
+  //    it's a pair of them that we need to repeat four times. This is done
+  //    by constructing the 32 bit constant corresponding to that pair.
+  const __m256i k__cospi_p16_p16 = _mm256_set1_epi16((int16_t)cospi_16_64);
+  const __m256i k__cospi_p16_m16 = pair256_set_epi16(+cospi_16_64, -cospi_16_64);
+  const __m256i k__cospi_m08_p24 = pair256_set_epi16(-cospi_8_64,   cospi_24_64);
+  const __m256i k__cospi_m24_m08 = pair256_set_epi16(-cospi_24_64, -cospi_8_64);
+  const __m256i k__cospi_p24_p08 = pair256_set_epi16(+cospi_24_64,  cospi_8_64);
+  const __m256i k__cospi_p12_p20 = pair256_set_epi16(+cospi_12_64,  cospi_20_64);
+  const __m256i k__cospi_m20_p12 = pair256_set_epi16(-cospi_20_64,  cospi_12_64);
+  const __m256i k__cospi_m04_p28 = pair256_set_epi16(-cospi_4_64,   cospi_28_64);
+  const __m256i k__cospi_p28_p04 = pair256_set_epi16(+cospi_28_64,  cospi_4_64);
+  const __m256i k__cospi_m28_m04 = pair256_set_epi16(-cospi_28_64, -cospi_4_64);
+  const __m256i k__cospi_m12_m20 = pair256_set_epi16(-cospi_12_64, -cospi_20_64);
+  const __m256i k__cospi_p30_p02 = pair256_set_epi16(+cospi_30_64,  cospi_2_64);
+  const __m256i k__cospi_p14_p18 = pair256_set_epi16(+cospi_14_64,  cospi_18_64);
+  const __m256i k__cospi_p22_p10 = pair256_set_epi16(+cospi_22_64,  cospi_10_64);
+  const __m256i k__cospi_p06_p26 = pair256_set_epi16(+cospi_6_64,   cospi_26_64);
+  const __m256i k__cospi_m26_p06 = pair256_set_epi16(-cospi_26_64,  cospi_6_64);
+  const __m256i k__cospi_m10_p22 = pair256_set_epi16(-cospi_10_64,  cospi_22_64);
+  const __m256i k__cospi_m18_p14 = pair256_set_epi16(-cospi_18_64,  cospi_14_64);
+  const __m256i k__cospi_m02_p30 = pair256_set_epi16(-cospi_2_64,   cospi_30_64);
+  const __m256i k__cospi_p31_p01 = pair256_set_epi16(+cospi_31_64,  cospi_1_64);
+  const __m256i k__cospi_p15_p17 = pair256_set_epi16(+cospi_15_64,  cospi_17_64);
+  const __m256i k__cospi_p23_p09 = pair256_set_epi16(+cospi_23_64,  cospi_9_64);
+  const __m256i k__cospi_p07_p25 = pair256_set_epi16(+cospi_7_64,   cospi_25_64);
+  const __m256i k__cospi_m25_p07 = pair256_set_epi16(-cospi_25_64,  cospi_7_64);
+  const __m256i k__cospi_m09_p23 = pair256_set_epi16(-cospi_9_64,   cospi_23_64);
+  const __m256i k__cospi_m17_p15 = pair256_set_epi16(-cospi_17_64,  cospi_15_64);
+  const __m256i k__cospi_m01_p31 = pair256_set_epi16(-cospi_1_64,   cospi_31_64);
+  const __m256i k__cospi_p27_p05 = pair256_set_epi16(+cospi_27_64,  cospi_5_64);
+  const __m256i k__cospi_p11_p21 = pair256_set_epi16(+cospi_11_64,  cospi_21_64);
+  const __m256i k__cospi_p19_p13 = pair256_set_epi16(+cospi_19_64,  cospi_13_64);
+  const __m256i k__cospi_p03_p29 = pair256_set_epi16(+cospi_3_64,   cospi_29_64);
+  const __m256i k__cospi_m29_p03 = pair256_set_epi16(-cospi_29_64,  cospi_3_64);
+  const __m256i k__cospi_m13_p19 = pair256_set_epi16(-cospi_13_64,  cospi_19_64);
+  const __m256i k__cospi_m21_p11 = pair256_set_epi16(-cospi_21_64,  cospi_11_64);
+  const __m256i k__cospi_m05_p27 = pair256_set_epi16(-cospi_5_64,   cospi_27_64);
+  const __m256i k__DCT_CONST_ROUNDING = _mm256_set1_epi32(DCT_CONST_ROUNDING);
+  const __m256i kZero = _mm256_set1_epi16(0);
+  const __m256i kOne  = _mm256_set1_epi16(1);
+  // Do the two transform/transpose passes
+  int pass;
+  for (pass = 0; pass < 2; ++pass) {
+    // We process sixteen columns (transposed rows in second pass) at a time.
+    int column_start;
+    for (column_start = 0; column_start < 32; column_start += 16) {
+      __m256i step1[32];
+      __m256i step2[32];
+      __m256i step3[32];
+      __m256i out[32];
+      // Stage 1
+      // Note: even though all the loads below are aligned, using the aligned
+      //       intrinsic make the code slightly slower.
+      if (0 == pass) {
+        const int16_t *in  = &input[column_start];
+        // step1[i] =  (in[ 0 * stride] + in[(32 -  1) * stride]) << 2;
+        // Note: the next four blocks could be in a loop. That would help the
+        //       instruction cache but is actually slower.
+        {
+          const int16_t *ina =  in +  0 * str1;
+          const int16_t *inb =  in + 31 * str1;
+          __m256i *step1a = &step1[ 0];
+          __m256i *step1b = &step1[31];
+          const __m256i ina0  = _mm256_loadu_si256((const __m256i *)(ina));
+          const __m256i ina1  = _mm256_loadu_si256((const __m256i *)(ina + str1));
+          const __m256i ina2  = _mm256_loadu_si256((const __m256i *)(ina + str2));
+          const __m256i ina3  = _mm256_loadu_si256((const __m256i *)(ina + str3));
+          const __m256i inb3  = _mm256_loadu_si256((const __m256i *)(inb - str3));
+          const __m256i inb2  = _mm256_loadu_si256((const __m256i *)(inb - str2));
+          const __m256i inb1  = _mm256_loadu_si256((const __m256i *)(inb - str1));
+          const __m256i inb0  = _mm256_loadu_si256((const __m256i *)(inb));
+          step1a[ 0] = _mm256_add_epi16(ina0, inb0);
+          step1a[ 1] = _mm256_add_epi16(ina1, inb1);
+          step1a[ 2] = _mm256_add_epi16(ina2, inb2);
+          step1a[ 3] = _mm256_add_epi16(ina3, inb3);
+          step1b[-3] = _mm256_sub_epi16(ina3, inb3);
+          step1b[-2] = _mm256_sub_epi16(ina2, inb2);
+          step1b[-1] = _mm256_sub_epi16(ina1, inb1);
+          step1b[-0] = _mm256_sub_epi16(ina0, inb0);
+          step1a[ 0] = _mm256_slli_epi16(step1a[ 0], 2);
+          step1a[ 1] = _mm256_slli_epi16(step1a[ 1], 2);
+          step1a[ 2] = _mm256_slli_epi16(step1a[ 2], 2);
+          step1a[ 3] = _mm256_slli_epi16(step1a[ 3], 2);
+          step1b[-3] = _mm256_slli_epi16(step1b[-3], 2);
+          step1b[-2] = _mm256_slli_epi16(step1b[-2], 2);
+          step1b[-1] = _mm256_slli_epi16(step1b[-1], 2);
+          step1b[-0] = _mm256_slli_epi16(step1b[-0], 2);
+        }
+        {
+          const int16_t *ina =  in +  4 * str1;
+          const int16_t *inb =  in + 27 * str1;
+          __m256i *step1a = &step1[ 4];
+          __m256i *step1b = &step1[27];
+          const __m256i ina0  = _mm256_loadu_si256((const __m256i *)(ina));
+          const __m256i ina1  = _mm256_loadu_si256((const __m256i *)(ina + str1));
+          const __m256i ina2  = _mm256_loadu_si256((const __m256i *)(ina + str2));
+          const __m256i ina3  = _mm256_loadu_si256((const __m256i *)(ina + str3));
+          const __m256i inb3  = _mm256_loadu_si256((const __m256i *)(inb - str3));
+          const __m256i inb2  = _mm256_loadu_si256((const __m256i *)(inb - str2));
+          const __m256i inb1  = _mm256_loadu_si256((const __m256i *)(inb - str1));
+          const __m256i inb0  = _mm256_loadu_si256((const __m256i *)(inb));
+          step1a[ 0] = _mm256_add_epi16(ina0, inb0);
+          step1a[ 1] = _mm256_add_epi16(ina1, inb1);
+          step1a[ 2] = _mm256_add_epi16(ina2, inb2);
+          step1a[ 3] = _mm256_add_epi16(ina3, inb3);
+          step1b[-3] = _mm256_sub_epi16(ina3, inb3);
+          step1b[-2] = _mm256_sub_epi16(ina2, inb2);
+          step1b[-1] = _mm256_sub_epi16(ina1, inb1);
+          step1b[-0] = _mm256_sub_epi16(ina0, inb0);
+          step1a[ 0] = _mm256_slli_epi16(step1a[ 0], 2);
+          step1a[ 1] = _mm256_slli_epi16(step1a[ 1], 2);
+          step1a[ 2] = _mm256_slli_epi16(step1a[ 2], 2);
+          step1a[ 3] = _mm256_slli_epi16(step1a[ 3], 2);
+          step1b[-3] = _mm256_slli_epi16(step1b[-3], 2);
+          step1b[-2] = _mm256_slli_epi16(step1b[-2], 2);
+          step1b[-1] = _mm256_slli_epi16(step1b[-1], 2);
+          step1b[-0] = _mm256_slli_epi16(step1b[-0], 2);
+        }
+        {
+          const int16_t *ina =  in +  8 * str1;
+          const int16_t *inb =  in + 23 * str1;
+          __m256i *step1a = &step1[ 8];
+          __m256i *step1b = &step1[23];
+          const __m256i ina0  = _mm256_loadu_si256((const __m256i *)(ina));
+          const __m256i ina1  = _mm256_loadu_si256((const __m256i *)(ina + str1));
+          const __m256i ina2  = _mm256_loadu_si256((const __m256i *)(ina + str2));
+          const __m256i ina3  = _mm256_loadu_si256((const __m256i *)(ina + str3));
+          const __m256i inb3  = _mm256_loadu_si256((const __m256i *)(inb - str3));
+          const __m256i inb2  = _mm256_loadu_si256((const __m256i *)(inb - str2));
+          const __m256i inb1  = _mm256_loadu_si256((const __m256i *)(inb - str1));
+          const __m256i inb0  = _mm256_loadu_si256((const __m256i *)(inb));
+          step1a[ 0] = _mm256_add_epi16(ina0, inb0);
+          step1a[ 1] = _mm256_add_epi16(ina1, inb1);
+          step1a[ 2] = _mm256_add_epi16(ina2, inb2);
+          step1a[ 3] = _mm256_add_epi16(ina3, inb3);
+          step1b[-3] = _mm256_sub_epi16(ina3, inb3);
+          step1b[-2] = _mm256_sub_epi16(ina2, inb2);
+          step1b[-1] = _mm256_sub_epi16(ina1, inb1);
+          step1b[-0] = _mm256_sub_epi16(ina0, inb0);
+          step1a[ 0] = _mm256_slli_epi16(step1a[ 0], 2);
+          step1a[ 1] = _mm256_slli_epi16(step1a[ 1], 2);
+          step1a[ 2] = _mm256_slli_epi16(step1a[ 2], 2);
+          step1a[ 3] = _mm256_slli_epi16(step1a[ 3], 2);
+          step1b[-3] = _mm256_slli_epi16(step1b[-3], 2);
+          step1b[-2] = _mm256_slli_epi16(step1b[-2], 2);
+          step1b[-1] = _mm256_slli_epi16(step1b[-1], 2);
+          step1b[-0] = _mm256_slli_epi16(step1b[-0], 2);
+        }
+        {
+          const int16_t *ina =  in + 12 * str1;
+          const int16_t *inb =  in + 19 * str1;
+          __m256i *step1a = &step1[12];
+          __m256i *step1b = &step1[19];
+          const __m256i ina0  = _mm256_loadu_si256((const __m256i *)(ina));
+          const __m256i ina1  = _mm256_loadu_si256((const __m256i *)(ina + str1));
+          const __m256i ina2  = _mm256_loadu_si256((const __m256i *)(ina + str2));
+          const __m256i ina3  = _mm256_loadu_si256((const __m256i *)(ina + str3));
+          const __m256i inb3  = _mm256_loadu_si256((const __m256i *)(inb - str3));
+          const __m256i inb2  = _mm256_loadu_si256((const __m256i *)(inb - str2));
+          const __m256i inb1  = _mm256_loadu_si256((const __m256i *)(inb - str1));
+          const __m256i inb0  = _mm256_loadu_si256((const __m256i *)(inb));
+          step1a[ 0] = _mm256_add_epi16(ina0, inb0);
+          step1a[ 1] = _mm256_add_epi16(ina1, inb1);
+          step1a[ 2] = _mm256_add_epi16(ina2, inb2);
+          step1a[ 3] = _mm256_add_epi16(ina3, inb3);
+          step1b[-3] = _mm256_sub_epi16(ina3, inb3);
+          step1b[-2] = _mm256_sub_epi16(ina2, inb2);
+          step1b[-1] = _mm256_sub_epi16(ina1, inb1);
+          step1b[-0] = _mm256_sub_epi16(ina0, inb0);
+          step1a[ 0] = _mm256_slli_epi16(step1a[ 0], 2);
+          step1a[ 1] = _mm256_slli_epi16(step1a[ 1], 2);
+          step1a[ 2] = _mm256_slli_epi16(step1a[ 2], 2);
+          step1a[ 3] = _mm256_slli_epi16(step1a[ 3], 2);
+          step1b[-3] = _mm256_slli_epi16(step1b[-3], 2);
+          step1b[-2] = _mm256_slli_epi16(step1b[-2], 2);
+          step1b[-1] = _mm256_slli_epi16(step1b[-1], 2);
+          step1b[-0] = _mm256_slli_epi16(step1b[-0], 2);
+        }
+      } else {
+        int16_t *in = &intermediate[column_start];
+        // step1[i] =  in[ 0 * 32] + in[(32 -  1) * 32];
+        // Note: using the same approach as above to have common offset is
+        //       counter-productive as all offsets can be calculated at compile
+        //       time.
+        // Note: the next four blocks could be in a loop. That would help the
+        //       instruction cache but is actually slower.
+        {
+          __m256i in00  = _mm256_loadu_si256((const __m256i *)(in +  0 * 32));
+          __m256i in01  = _mm256_loadu_si256((const __m256i *)(in +  1 * 32));
+          __m256i in02  = _mm256_loadu_si256((const __m256i *)(in +  2 * 32));
+          __m256i in03  = _mm256_loadu_si256((const __m256i *)(in +  3 * 32));
+          __m256i in28  = _mm256_loadu_si256((const __m256i *)(in + 28 * 32));
+          __m256i in29  = _mm256_loadu_si256((const __m256i *)(in + 29 * 32));
+          __m256i in30  = _mm256_loadu_si256((const __m256i *)(in + 30 * 32));
+          __m256i in31  = _mm256_loadu_si256((const __m256i *)(in + 31 * 32));
+          step1[ 0] = _mm256_add_epi16(in00, in31);
+          step1[ 1] = _mm256_add_epi16(in01, in30);
+          step1[ 2] = _mm256_add_epi16(in02, in29);
+          step1[ 3] = _mm256_add_epi16(in03, in28);
+          step1[28] = _mm256_sub_epi16(in03, in28);
+          step1[29] = _mm256_sub_epi16(in02, in29);
+          step1[30] = _mm256_sub_epi16(in01, in30);
+          step1[31] = _mm256_sub_epi16(in00, in31);
+        }
+        {
+          __m256i in04  = _mm256_loadu_si256((const __m256i *)(in +  4 * 32));
+          __m256i in05  = _mm256_loadu_si256((const __m256i *)(in +  5 * 32));
+          __m256i in06  = _mm256_loadu_si256((const __m256i *)(in +  6 * 32));
+          __m256i in07  = _mm256_loadu_si256((const __m256i *)(in +  7 * 32));
+          __m256i in24  = _mm256_loadu_si256((const __m256i *)(in + 24 * 32));
+          __m256i in25  = _mm256_loadu_si256((const __m256i *)(in + 25 * 32));
+          __m256i in26  = _mm256_loadu_si256((const __m256i *)(in + 26 * 32));
+          __m256i in27  = _mm256_loadu_si256((const __m256i *)(in + 27 * 32));
+          step1[ 4] = _mm256_add_epi16(in04, in27);
+          step1[ 5] = _mm256_add_epi16(in05, in26);
+          step1[ 6] = _mm256_add_epi16(in06, in25);
+          step1[ 7] = _mm256_add_epi16(in07, in24);
+          step1[24] = _mm256_sub_epi16(in07, in24);
+          step1[25] = _mm256_sub_epi16(in06, in25);
+          step1[26] = _mm256_sub_epi16(in05, in26);
+          step1[27] = _mm256_sub_epi16(in04, in27);
+        }
+        {
+          __m256i in08  = _mm256_loadu_si256((const __m256i *)(in +  8 * 32));
+          __m256i in09  = _mm256_loadu_si256((const __m256i *)(in +  9 * 32));
+          __m256i in10  = _mm256_loadu_si256((const __m256i *)(in + 10 * 32));
+          __m256i in11  = _mm256_loadu_si256((const __m256i *)(in + 11 * 32));
+          __m256i in20  = _mm256_loadu_si256((const __m256i *)(in + 20 * 32));
+          __m256i in21  = _mm256_loadu_si256((const __m256i *)(in + 21 * 32));
+          __m256i in22  = _mm256_loadu_si256((const __m256i *)(in + 22 * 32));
+          __m256i in23  = _mm256_loadu_si256((const __m256i *)(in + 23 * 32));
+          step1[ 8] = _mm256_add_epi16(in08, in23);
+          step1[ 9] = _mm256_add_epi16(in09, in22);
+          step1[10] = _mm256_add_epi16(in10, in21);
+          step1[11] = _mm256_add_epi16(in11, in20);
+          step1[20] = _mm256_sub_epi16(in11, in20);
+          step1[21] = _mm256_sub_epi16(in10, in21);
+          step1[22] = _mm256_sub_epi16(in09, in22);
+          step1[23] = _mm256_sub_epi16(in08, in23);
+        }
+        {
+          __m256i in12  = _mm256_loadu_si256((const __m256i *)(in + 12 * 32));
+          __m256i in13  = _mm256_loadu_si256((const __m256i *)(in + 13 * 32));
+          __m256i in14  = _mm256_loadu_si256((const __m256i *)(in + 14 * 32));
+          __m256i in15  = _mm256_loadu_si256((const __m256i *)(in + 15 * 32));
+          __m256i in16  = _mm256_loadu_si256((const __m256i *)(in + 16 * 32));
+          __m256i in17  = _mm256_loadu_si256((const __m256i *)(in + 17 * 32));
+          __m256i in18  = _mm256_loadu_si256((const __m256i *)(in + 18 * 32));
+          __m256i in19  = _mm256_loadu_si256((const __m256i *)(in + 19 * 32));
+          step1[12] = _mm256_add_epi16(in12, in19);
+          step1[13] = _mm256_add_epi16(in13, in18);
+          step1[14] = _mm256_add_epi16(in14, in17);
+          step1[15] = _mm256_add_epi16(in15, in16);
+          step1[16] = _mm256_sub_epi16(in15, in16);
+          step1[17] = _mm256_sub_epi16(in14, in17);
+          step1[18] = _mm256_sub_epi16(in13, in18);
+          step1[19] = _mm256_sub_epi16(in12, in19);
+        }
+      }
+      // Stage 2
+      {
+        step2[ 0] = _mm256_add_epi16(step1[0], step1[15]);
+        step2[ 1] = _mm256_add_epi16(step1[1], step1[14]);
+        step2[ 2] = _mm256_add_epi16(step1[2], step1[13]);
+        step2[ 3] = _mm256_add_epi16(step1[3], step1[12]);
+        step2[ 4] = _mm256_add_epi16(step1[4], step1[11]);
+        step2[ 5] = _mm256_add_epi16(step1[5], step1[10]);
+        step2[ 6] = _mm256_add_epi16(step1[6], step1[ 9]);
+        step2[ 7] = _mm256_add_epi16(step1[7], step1[ 8]);
+        step2[ 8] = _mm256_sub_epi16(step1[7], step1[ 8]);
+        step2[ 9] = _mm256_sub_epi16(step1[6], step1[ 9]);
+        step2[10] = _mm256_sub_epi16(step1[5], step1[10]);
+        step2[11] = _mm256_sub_epi16(step1[4], step1[11]);
+        step2[12] = _mm256_sub_epi16(step1[3], step1[12]);
+        step2[13] = _mm256_sub_epi16(step1[2], step1[13]);
+        step2[14] = _mm256_sub_epi16(step1[1], step1[14]);
+        step2[15] = _mm256_sub_epi16(step1[0], step1[15]);
+      }
+      {
+        const __m256i s2_20_0 = _mm256_unpacklo_epi16(step1[27], step1[20]);
+        const __m256i s2_20_1 = _mm256_unpackhi_epi16(step1[27], step1[20]);
+        const __m256i s2_21_0 = _mm256_unpacklo_epi16(step1[26], step1[21]);
+        const __m256i s2_21_1 = _mm256_unpackhi_epi16(step1[26], step1[21]);
+        const __m256i s2_22_0 = _mm256_unpacklo_epi16(step1[25], step1[22]);
+        const __m256i s2_22_1 = _mm256_unpackhi_epi16(step1[25], step1[22]);
+        const __m256i s2_23_0 = _mm256_unpacklo_epi16(step1[24], step1[23]);
+        const __m256i s2_23_1 = _mm256_unpackhi_epi16(step1[24], step1[23]);
+        const __m256i s2_20_2 = _mm256_madd_epi16(s2_20_0, k__cospi_p16_m16);
+        const __m256i s2_20_3 = _mm256_madd_epi16(s2_20_1, k__cospi_p16_m16);
+        const __m256i s2_21_2 = _mm256_madd_epi16(s2_21_0, k__cospi_p16_m16);
+        const __m256i s2_21_3 = _mm256_madd_epi16(s2_21_1, k__cospi_p16_m16);
+        const __m256i s2_22_2 = _mm256_madd_epi16(s2_22_0, k__cospi_p16_m16);
+        const __m256i s2_22_3 = _mm256_madd_epi16(s2_22_1, k__cospi_p16_m16);
+        const __m256i s2_23_2 = _mm256_madd_epi16(s2_23_0, k__cospi_p16_m16);
+        const __m256i s2_23_3 = _mm256_madd_epi16(s2_23_1, k__cospi_p16_m16);
+        const __m256i s2_24_2 = _mm256_madd_epi16(s2_23_0, k__cospi_p16_p16);
+        const __m256i s2_24_3 = _mm256_madd_epi16(s2_23_1, k__cospi_p16_p16);
+        const __m256i s2_25_2 = _mm256_madd_epi16(s2_22_0, k__cospi_p16_p16);
+        const __m256i s2_25_3 = _mm256_madd_epi16(s2_22_1, k__cospi_p16_p16);
+        const __m256i s2_26_2 = _mm256_madd_epi16(s2_21_0, k__cospi_p16_p16);
+        const __m256i s2_26_3 = _mm256_madd_epi16(s2_21_1, k__cospi_p16_p16);
+        const __m256i s2_27_2 = _mm256_madd_epi16(s2_20_0, k__cospi_p16_p16);
+        const __m256i s2_27_3 = _mm256_madd_epi16(s2_20_1, k__cospi_p16_p16);
+        // dct_const_round_shift
+        const __m256i s2_20_4 = _mm256_add_epi32(s2_20_2, k__DCT_CONST_ROUNDING);
+        const __m256i s2_20_5 = _mm256_add_epi32(s2_20_3, k__DCT_CONST_ROUNDING);
+        const __m256i s2_21_4 = _mm256_add_epi32(s2_21_2, k__DCT_CONST_ROUNDING);
+        const __m256i s2_21_5 = _mm256_add_epi32(s2_21_3, k__DCT_CONST_ROUNDING);
+        const __m256i s2_22_4 = _mm256_add_epi32(s2_22_2, k__DCT_CONST_ROUNDING);
+        const __m256i s2_22_5 = _mm256_add_epi32(s2_22_3, k__DCT_CONST_ROUNDING);
+        const __m256i s2_23_4 = _mm256_add_epi32(s2_23_2, k__DCT_CONST_ROUNDING);
+        const __m256i s2_23_5 = _mm256_add_epi32(s2_23_3, k__DCT_CONST_ROUNDING);
+        const __m256i s2_24_4 = _mm256_add_epi32(s2_24_2, k__DCT_CONST_ROUNDING);
+        const __m256i s2_24_5 = _mm256_add_epi32(s2_24_3, k__DCT_CONST_ROUNDING);
+        const __m256i s2_25_4 = _mm256_add_epi32(s2_25_2, k__DCT_CONST_ROUNDING);
+        const __m256i s2_25_5 = _mm256_add_epi32(s2_25_3, k__DCT_CONST_ROUNDING);
+        const __m256i s2_26_4 = _mm256_add_epi32(s2_26_2, k__DCT_CONST_ROUNDING);
+        const __m256i s2_26_5 = _mm256_add_epi32(s2_26_3, k__DCT_CONST_ROUNDING);
+        const __m256i s2_27_4 = _mm256_add_epi32(s2_27_2, k__DCT_CONST_ROUNDING);
+        const __m256i s2_27_5 = _mm256_add_epi32(s2_27_3, k__DCT_CONST_ROUNDING);
+        const __m256i s2_20_6 = _mm256_srai_epi32(s2_20_4, DCT_CONST_BITS);
+        const __m256i s2_20_7 = _mm256_srai_epi32(s2_20_5, DCT_CONST_BITS);
+        const __m256i s2_21_6 = _mm256_srai_epi32(s2_21_4, DCT_CONST_BITS);
+        const __m256i s2_21_7 = _mm256_srai_epi32(s2_21_5, DCT_CONST_BITS);
+        const __m256i s2_22_6 = _mm256_srai_epi32(s2_22_4, DCT_CONST_BITS);
+        const __m256i s2_22_7 = _mm256_srai_epi32(s2_22_5, DCT_CONST_BITS);
+        const __m256i s2_23_6 = _mm256_srai_epi32(s2_23_4, DCT_CONST_BITS);
+        const __m256i s2_23_7 = _mm256_srai_epi32(s2_23_5, DCT_CONST_BITS);
+        const __m256i s2_24_6 = _mm256_srai_epi32(s2_24_4, DCT_CONST_BITS);
+        const __m256i s2_24_7 = _mm256_srai_epi32(s2_24_5, DCT_CONST_BITS);
+        const __m256i s2_25_6 = _mm256_srai_epi32(s2_25_4, DCT_CONST_BITS);
+        const __m256i s2_25_7 = _mm256_srai_epi32(s2_25_5, DCT_CONST_BITS);
+        const __m256i s2_26_6 = _mm256_srai_epi32(s2_26_4, DCT_CONST_BITS);
+        const __m256i s2_26_7 = _mm256_srai_epi32(s2_26_5, DCT_CONST_BITS);
+        const __m256i s2_27_6 = _mm256_srai_epi32(s2_27_4, DCT_CONST_BITS);
+        const __m256i s2_27_7 = _mm256_srai_epi32(s2_27_5, DCT_CONST_BITS);
+        // Combine
+        step2[20] = _mm256_packs_epi32(s2_20_6, s2_20_7);
+        step2[21] = _mm256_packs_epi32(s2_21_6, s2_21_7);
+        step2[22] = _mm256_packs_epi32(s2_22_6, s2_22_7);
+        step2[23] = _mm256_packs_epi32(s2_23_6, s2_23_7);
+        step2[24] = _mm256_packs_epi32(s2_24_6, s2_24_7);
+        step2[25] = _mm256_packs_epi32(s2_25_6, s2_25_7);
+        step2[26] = _mm256_packs_epi32(s2_26_6, s2_26_7);
+        step2[27] = _mm256_packs_epi32(s2_27_6, s2_27_7);
+      }
+
+#if !FDCT32x32_HIGH_PRECISION
+      // dump the magnitude by half, hence the intermediate values are within
+      // the range of 16 bits.
+      if (1 == pass) {
+        __m256i s3_00_0 = _mm256_cmpgt_epi16(kZero,step2[ 0]);
+        __m256i s3_01_0 = _mm256_cmpgt_epi16(kZero,step2[ 1]);
+        __m256i s3_02_0 = _mm256_cmpgt_epi16(kZero,step2[ 2]);
+        __m256i s3_03_0 = _mm256_cmpgt_epi16(kZero,step2[ 3]);
+        __m256i s3_04_0 = _mm256_cmpgt_epi16(kZero,step2[ 4]);
+        __m256i s3_05_0 = _mm256_cmpgt_epi16(kZero,step2[ 5]);
+        __m256i s3_06_0 = _mm256_cmpgt_epi16(kZero,step2[ 6]);
+        __m256i s3_07_0 = _mm256_cmpgt_epi16(kZero,step2[ 7]);
+        __m256i s2_08_0 = _mm256_cmpgt_epi16(kZero,step2[ 8]);
+        __m256i s2_09_0 = _mm256_cmpgt_epi16(kZero,step2[ 9]);
+        __m256i s3_10_0 = _mm256_cmpgt_epi16(kZero,step2[10]);
+        __m256i s3_11_0 = _mm256_cmpgt_epi16(kZero,step2[11]);
+        __m256i s3_12_0 = _mm256_cmpgt_epi16(kZero,step2[12]);
+        __m256i s3_13_0 = _mm256_cmpgt_epi16(kZero,step2[13]);
+        __m256i s2_14_0 = _mm256_cmpgt_epi16(kZero,step2[14]);
+        __m256i s2_15_0 = _mm256_cmpgt_epi16(kZero,step2[15]);
+        __m256i s3_16_0 = _mm256_cmpgt_epi16(kZero,step1[16]);
+        __m256i s3_17_0 = _mm256_cmpgt_epi16(kZero,step1[17]);
+        __m256i s3_18_0 = _mm256_cmpgt_epi16(kZero,step1[18]);
+        __m256i s3_19_0 = _mm256_cmpgt_epi16(kZero,step1[19]);
+        __m256i s3_20_0 = _mm256_cmpgt_epi16(kZero,step2[20]);
+        __m256i s3_21_0 = _mm256_cmpgt_epi16(kZero,step2[21]);
+        __m256i s3_22_0 = _mm256_cmpgt_epi16(kZero,step2[22]);
+        __m256i s3_23_0 = _mm256_cmpgt_epi16(kZero,step2[23]);
+        __m256i s3_24_0 = _mm256_cmpgt_epi16(kZero,step2[24]);
+        __m256i s3_25_0 = _mm256_cmpgt_epi16(kZero,step2[25]);
+        __m256i s3_26_0 = _mm256_cmpgt_epi16(kZero,step2[26]);
+        __m256i s3_27_0 = _mm256_cmpgt_epi16(kZero,step2[27]);
+        __m256i s3_28_0 = _mm256_cmpgt_epi16(kZero,step1[28]);
+        __m256i s3_29_0 = _mm256_cmpgt_epi16(kZero,step1[29]);
+        __m256i s3_30_0 = _mm256_cmpgt_epi16(kZero,step1[30]);
+        __m256i s3_31_0 = _mm256_cmpgt_epi16(kZero,step1[31]);
+
+        step2[ 0] = _mm256_sub_epi16(step2[ 0], s3_00_0);
+        step2[ 1] = _mm256_sub_epi16(step2[ 1], s3_01_0);
+        step2[ 2] = _mm256_sub_epi16(step2[ 2], s3_02_0);
+        step2[ 3] = _mm256_sub_epi16(step2[ 3], s3_03_0);
+        step2[ 4] = _mm256_sub_epi16(step2[ 4], s3_04_0);
+        step2[ 5] = _mm256_sub_epi16(step2[ 5], s3_05_0);
+        step2[ 6] = _mm256_sub_epi16(step2[ 6], s3_06_0);
+        step2[ 7] = _mm256_sub_epi16(step2[ 7], s3_07_0);
+        step2[ 8] = _mm256_sub_epi16(step2[ 8], s2_08_0);
+        step2[ 9] = _mm256_sub_epi16(step2[ 9], s2_09_0);
+        step2[10] = _mm256_sub_epi16(step2[10], s3_10_0);
+        step2[11] = _mm256_sub_epi16(step2[11], s3_11_0);
+        step2[12] = _mm256_sub_epi16(step2[12], s3_12_0);
+        step2[13] = _mm256_sub_epi16(step2[13], s3_13_0);
+        step2[14] = _mm256_sub_epi16(step2[14], s2_14_0);
+        step2[15] = _mm256_sub_epi16(step2[15], s2_15_0);
+        step1[16] = _mm256_sub_epi16(step1[16], s3_16_0);
+        step1[17] = _mm256_sub_epi16(step1[17], s3_17_0);
+        step1[18] = _mm256_sub_epi16(step1[18], s3_18_0);
+        step1[19] = _mm256_sub_epi16(step1[19], s3_19_0);
+        step2[20] = _mm256_sub_epi16(step2[20], s3_20_0);
+        step2[21] = _mm256_sub_epi16(step2[21], s3_21_0);
+        step2[22] = _mm256_sub_epi16(step2[22], s3_22_0);
+        step2[23] = _mm256_sub_epi16(step2[23], s3_23_0);
+        step2[24] = _mm256_sub_epi16(step2[24], s3_24_0);
+        step2[25] = _mm256_sub_epi16(step2[25], s3_25_0);
+        step2[26] = _mm256_sub_epi16(step2[26], s3_26_0);
+        step2[27] = _mm256_sub_epi16(step2[27], s3_27_0);
+        step1[28] = _mm256_sub_epi16(step1[28], s3_28_0);
+        step1[29] = _mm256_sub_epi16(step1[29], s3_29_0);
+        step1[30] = _mm256_sub_epi16(step1[30], s3_30_0);
+        step1[31] = _mm256_sub_epi16(step1[31], s3_31_0);
+
+        step2[ 0] = _mm256_add_epi16(step2[ 0], kOne);
+        step2[ 1] = _mm256_add_epi16(step2[ 1], kOne);
+        step2[ 2] = _mm256_add_epi16(step2[ 2], kOne);
+        step2[ 3] = _mm256_add_epi16(step2[ 3], kOne);
+        step2[ 4] = _mm256_add_epi16(step2[ 4], kOne);
+        step2[ 5] = _mm256_add_epi16(step2[ 5], kOne);
+        step2[ 6] = _mm256_add_epi16(step2[ 6], kOne);
+        step2[ 7] = _mm256_add_epi16(step2[ 7], kOne);
+        step2[ 8] = _mm256_add_epi16(step2[ 8], kOne);
+        step2[ 9] = _mm256_add_epi16(step2[ 9], kOne);
+        step2[10] = _mm256_add_epi16(step2[10], kOne);
+        step2[11] = _mm256_add_epi16(step2[11], kOne);
+        step2[12] = _mm256_add_epi16(step2[12], kOne);
+        step2[13] = _mm256_add_epi16(step2[13], kOne);
+        step2[14] = _mm256_add_epi16(step2[14], kOne);
+        step2[15] = _mm256_add_epi16(step2[15], kOne);
+        step1[16] = _mm256_add_epi16(step1[16], kOne);
+        step1[17] = _mm256_add_epi16(step1[17], kOne);
+        step1[18] = _mm256_add_epi16(step1[18], kOne);
+        step1[19] = _mm256_add_epi16(step1[19], kOne);
+        step2[20] = _mm256_add_epi16(step2[20], kOne);
+        step2[21] = _mm256_add_epi16(step2[21], kOne);
+        step2[22] = _mm256_add_epi16(step2[22], kOne);
+        step2[23] = _mm256_add_epi16(step2[23], kOne);
+        step2[24] = _mm256_add_epi16(step2[24], kOne);
+        step2[25] = _mm256_add_epi16(step2[25], kOne);
+        step2[26] = _mm256_add_epi16(step2[26], kOne);
+        step2[27] = _mm256_add_epi16(step2[27], kOne);
+        step1[28] = _mm256_add_epi16(step1[28], kOne);
+        step1[29] = _mm256_add_epi16(step1[29], kOne);
+        step1[30] = _mm256_add_epi16(step1[30], kOne);
+        step1[31] = _mm256_add_epi16(step1[31], kOne);
+
+        step2[ 0] = _mm256_srai_epi16(step2[ 0], 2);
+        step2[ 1] = _mm256_srai_epi16(step2[ 1], 2);
+        step2[ 2] = _mm256_srai_epi16(step2[ 2], 2);
+        step2[ 3] = _mm256_srai_epi16(step2[ 3], 2);
+        step2[ 4] = _mm256_srai_epi16(step2[ 4], 2);
+        step2[ 5] = _mm256_srai_epi16(step2[ 5], 2);
+        step2[ 6] = _mm256_srai_epi16(step2[ 6], 2);
+        step2[ 7] = _mm256_srai_epi16(step2[ 7], 2);
+        step2[ 8] = _mm256_srai_epi16(step2[ 8], 2);
+        step2[ 9] = _mm256_srai_epi16(step2[ 9], 2);
+        step2[10] = _mm256_srai_epi16(step2[10], 2);
+        step2[11] = _mm256_srai_epi16(step2[11], 2);
+        step2[12] = _mm256_srai_epi16(step2[12], 2);
+        step2[13] = _mm256_srai_epi16(step2[13], 2);
+        step2[14] = _mm256_srai_epi16(step2[14], 2);
+        step2[15] = _mm256_srai_epi16(step2[15], 2);
+        step1[16] = _mm256_srai_epi16(step1[16], 2);
+        step1[17] = _mm256_srai_epi16(step1[17], 2);
+        step1[18] = _mm256_srai_epi16(step1[18], 2);
+        step1[19] = _mm256_srai_epi16(step1[19], 2);
+        step2[20] = _mm256_srai_epi16(step2[20], 2);
+        step2[21] = _mm256_srai_epi16(step2[21], 2);
+        step2[22] = _mm256_srai_epi16(step2[22], 2);
+        step2[23] = _mm256_srai_epi16(step2[23], 2);
+        step2[24] = _mm256_srai_epi16(step2[24], 2);
+        step2[25] = _mm256_srai_epi16(step2[25], 2);
+        step2[26] = _mm256_srai_epi16(step2[26], 2);
+        step2[27] = _mm256_srai_epi16(step2[27], 2);
+        step1[28] = _mm256_srai_epi16(step1[28], 2);
+        step1[29] = _mm256_srai_epi16(step1[29], 2);
+        step1[30] = _mm256_srai_epi16(step1[30], 2);
+        step1[31] = _mm256_srai_epi16(step1[31], 2);
+      }
+#endif
+
+#if FDCT32x32_HIGH_PRECISION
+      if (pass == 0) {
+#endif
+      // Stage 3
+      {
+        step3[0] = _mm256_add_epi16(step2[(8 - 1)], step2[0]);
+        step3[1] = _mm256_add_epi16(step2[(8 - 2)], step2[1]);
+        step3[2] = _mm256_add_epi16(step2[(8 - 3)], step2[2]);
+        step3[3] = _mm256_add_epi16(step2[(8 - 4)], step2[3]);
+        step3[4] = _mm256_sub_epi16(step2[(8 - 5)], step2[4]);
+        step3[5] = _mm256_sub_epi16(step2[(8 - 6)], step2[5]);
+        step3[6] = _mm256_sub_epi16(step2[(8 - 7)], step2[6]);
+        step3[7] = _mm256_sub_epi16(step2[(8 - 8)], step2[7]);
+      }
+      {
+        const __m256i s3_10_0 = _mm256_unpacklo_epi16(step2[13], step2[10]);
+        const __m256i s3_10_1 = _mm256_unpackhi_epi16(step2[13], step2[10]);
+        const __m256i s3_11_0 = _mm256_unpacklo_epi16(step2[12], step2[11]);
+        const __m256i s3_11_1 = _mm256_unpackhi_epi16(step2[12], step2[11]);
+        const __m256i s3_10_2 = _mm256_madd_epi16(s3_10_0, k__cospi_p16_m16);
+        const __m256i s3_10_3 = _mm256_madd_epi16(s3_10_1, k__cospi_p16_m16);
+        const __m256i s3_11_2 = _mm256_madd_epi16(s3_11_0, k__cospi_p16_m16);
+        const __m256i s3_11_3 = _mm256_madd_epi16(s3_11_1, k__cospi_p16_m16);
+        const __m256i s3_12_2 = _mm256_madd_epi16(s3_11_0, k__cospi_p16_p16);
+        const __m256i s3_12_3 = _mm256_madd_epi16(s3_11_1, k__cospi_p16_p16);
+        const __m256i s3_13_2 = _mm256_madd_epi16(s3_10_0, k__cospi_p16_p16);
+        const __m256i s3_13_3 = _mm256_madd_epi16(s3_10_1, k__cospi_p16_p16);
+        // dct_const_round_shift
+        const __m256i s3_10_4 = _mm256_add_epi32(s3_10_2, k__DCT_CONST_ROUNDING);
+        const __m256i s3_10_5 = _mm256_add_epi32(s3_10_3, k__DCT_CONST_ROUNDING);
+        const __m256i s3_11_4 = _mm256_add_epi32(s3_11_2, k__DCT_CONST_ROUNDING);
+        const __m256i s3_11_5 = _mm256_add_epi32(s3_11_3, k__DCT_CONST_ROUNDING);
+        const __m256i s3_12_4 = _mm256_add_epi32(s3_12_2, k__DCT_CONST_ROUNDING);
+        const __m256i s3_12_5 = _mm256_add_epi32(s3_12_3, k__DCT_CONST_ROUNDING);
+        const __m256i s3_13_4 = _mm256_add_epi32(s3_13_2, k__DCT_CONST_ROUNDING);
+        const __m256i s3_13_5 = _mm256_add_epi32(s3_13_3, k__DCT_CONST_ROUNDING);
+        const __m256i s3_10_6 = _mm256_srai_epi32(s3_10_4, DCT_CONST_BITS);
+        const __m256i s3_10_7 = _mm256_srai_epi32(s3_10_5, DCT_CONST_BITS);
+        const __m256i s3_11_6 = _mm256_srai_epi32(s3_11_4, DCT_CONST_BITS);
+        const __m256i s3_11_7 = _mm256_srai_epi32(s3_11_5, DCT_CONST_BITS);
+        const __m256i s3_12_6 = _mm256_srai_epi32(s3_12_4, DCT_CONST_BITS);
+        const __m256i s3_12_7 = _mm256_srai_epi32(s3_12_5, DCT_CONST_BITS);
+        const __m256i s3_13_6 = _mm256_srai_epi32(s3_13_4, DCT_CONST_BITS);
+        const __m256i s3_13_7 = _mm256_srai_epi32(s3_13_5, DCT_CONST_BITS);
+        // Combine
+        step3[10] = _mm256_packs_epi32(s3_10_6, s3_10_7);
+        step3[11] = _mm256_packs_epi32(s3_11_6, s3_11_7);
+        step3[12] = _mm256_packs_epi32(s3_12_6, s3_12_7);
+        step3[13] = _mm256_packs_epi32(s3_13_6, s3_13_7);
+      }
+      {
+        step3[16] = _mm256_add_epi16(step2[23], step1[16]);
+        step3[17] = _mm256_add_epi16(step2[22], step1[17]);
+        step3[18] = _mm256_add_epi16(step2[21], step1[18]);
+        step3[19] = _mm256_add_epi16(step2[20], step1[19]);
+        step3[20] = _mm256_sub_epi16(step1[19], step2[20]);
+        step3[21] = _mm256_sub_epi16(step1[18], step2[21]);
+        step3[22] = _mm256_sub_epi16(step1[17], step2[22]);
+        step3[23] = _mm256_sub_epi16(step1[16], step2[23]);
+        step3[24] = _mm256_sub_epi16(step1[31], step2[24]);
+        step3[25] = _mm256_sub_epi16(step1[30], step2[25]);
+        step3[26] = _mm256_sub_epi16(step1[29], step2[26]);
+        step3[27] = _mm256_sub_epi16(step1[28], step2[27]);
+        step3[28] = _mm256_add_epi16(step2[27], step1[28]);
+        step3[29] = _mm256_add_epi16(step2[26], step1[29]);
+        step3[30] = _mm256_add_epi16(step2[25], step1[30]);
+        step3[31] = _mm256_add_epi16(step2[24], step1[31]);
+      }
+
+      // Stage 4
+      {
+        step1[ 0] = _mm256_add_epi16(step3[ 3], step3[ 0]);
+        step1[ 1] = _mm256_add_epi16(step3[ 2], step3[ 1]);
+        step1[ 2] = _mm256_sub_epi16(step3[ 1], step3[ 2]);
+        step1[ 3] = _mm256_sub_epi16(step3[ 0], step3[ 3]);
+        step1[ 8] = _mm256_add_epi16(step3[11], step2[ 8]);
+        step1[ 9] = _mm256_add_epi16(step3[10], step2[ 9]);
+        step1[10] = _mm256_sub_epi16(step2[ 9], step3[10]);
+        step1[11] = _mm256_sub_epi16(step2[ 8], step3[11]);
+        step1[12] = _mm256_sub_epi16(step2[15], step3[12]);
+        step1[13] = _mm256_sub_epi16(step2[14], step3[13]);
+        step1[14] = _mm256_add_epi16(step3[13], step2[14]);
+        step1[15] = _mm256_add_epi16(step3[12], step2[15]);
+      }
+      {
+        const __m256i s1_05_0 = _mm256_unpacklo_epi16(step3[6], step3[5]);
+        const __m256i s1_05_1 = _mm256_unpackhi_epi16(step3[6], step3[5]);
+        const __m256i s1_05_2 = _mm256_madd_epi16(s1_05_0, k__cospi_p16_m16);
+        const __m256i s1_05_3 = _mm256_madd_epi16(s1_05_1, k__cospi_p16_m16);
+        const __m256i s1_06_2 = _mm256_madd_epi16(s1_05_0, k__cospi_p16_p16);
+        const __m256i s1_06_3 = _mm256_madd_epi16(s1_05_1, k__cospi_p16_p16);
+        // dct_const_round_shift
+        const __m256i s1_05_4 = _mm256_add_epi32(s1_05_2, k__DCT_CONST_ROUNDING);
+        const __m256i s1_05_5 = _mm256_add_epi32(s1_05_3, k__DCT_CONST_ROUNDING);
+        const __m256i s1_06_4 = _mm256_add_epi32(s1_06_2, k__DCT_CONST_ROUNDING);
+        const __m256i s1_06_5 = _mm256_add_epi32(s1_06_3, k__DCT_CONST_ROUNDING);
+        const __m256i s1_05_6 = _mm256_srai_epi32(s1_05_4, DCT_CONST_BITS);
+        const __m256i s1_05_7 = _mm256_srai_epi32(s1_05_5, DCT_CONST_BITS);
+        const __m256i s1_06_6 = _mm256_srai_epi32(s1_06_4, DCT_CONST_BITS);
+        const __m256i s1_06_7 = _mm256_srai_epi32(s1_06_5, DCT_CONST_BITS);
+        // Combine
+        step1[5] = _mm256_packs_epi32(s1_05_6, s1_05_7);
+        step1[6] = _mm256_packs_epi32(s1_06_6, s1_06_7);
+      }
+      {
+        const __m256i s1_18_0 = _mm256_unpacklo_epi16(step3[18], step3[29]);
+        const __m256i s1_18_1 = _mm256_unpackhi_epi16(step3[18], step3[29]);
+        const __m256i s1_19_0 = _mm256_unpacklo_epi16(step3[19], step3[28]);
+        const __m256i s1_19_1 = _mm256_unpackhi_epi16(step3[19], step3[28]);
+        const __m256i s1_20_0 = _mm256_unpacklo_epi16(step3[20], step3[27]);
+        const __m256i s1_20_1 = _mm256_unpackhi_epi16(step3[20], step3[27]);
+        const __m256i s1_21_0 = _mm256_unpacklo_epi16(step3[21], step3[26]);
+        const __m256i s1_21_1 = _mm256_unpackhi_epi16(step3[21], step3[26]);
+        const __m256i s1_18_2 = _mm256_madd_epi16(s1_18_0, k__cospi_m08_p24);
+        const __m256i s1_18_3 = _mm256_madd_epi16(s1_18_1, k__cospi_m08_p24);
+        const __m256i s1_19_2 = _mm256_madd_epi16(s1_19_0, k__cospi_m08_p24);
+        const __m256i s1_19_3 = _mm256_madd_epi16(s1_19_1, k__cospi_m08_p24);
+        const __m256i s1_20_2 = _mm256_madd_epi16(s1_20_0, k__cospi_m24_m08);
+        const __m256i s1_20_3 = _mm256_madd_epi16(s1_20_1, k__cospi_m24_m08);
+        const __m256i s1_21_2 = _mm256_madd_epi16(s1_21_0, k__cospi_m24_m08);
+        const __m256i s1_21_3 = _mm256_madd_epi16(s1_21_1, k__cospi_m24_m08);
+        const __m256i s1_26_2 = _mm256_madd_epi16(s1_21_0, k__cospi_m08_p24);
+        const __m256i s1_26_3 = _mm256_madd_epi16(s1_21_1, k__cospi_m08_p24);
+        const __m256i s1_27_2 = _mm256_madd_epi16(s1_20_0, k__cospi_m08_p24);
+        const __m256i s1_27_3 = _mm256_madd_epi16(s1_20_1, k__cospi_m08_p24);
+        const __m256i s1_28_2 = _mm256_madd_epi16(s1_19_0, k__cospi_p24_p08);
+        const __m256i s1_28_3 = _mm256_madd_epi16(s1_19_1, k__cospi_p24_p08);
+        const __m256i s1_29_2 = _mm256_madd_epi16(s1_18_0, k__cospi_p24_p08);
+        const __m256i s1_29_3 = _mm256_madd_epi16(s1_18_1, k__cospi_p24_p08);
+        // dct_const_round_shift
+        const __m256i s1_18_4 = _mm256_add_epi32(s1_18_2, k__DCT_CONST_ROUNDING);
+        const __m256i s1_18_5 = _mm256_add_epi32(s1_18_3, k__DCT_CONST_ROUNDING);
+        const __m256i s1_19_4 = _mm256_add_epi32(s1_19_2, k__DCT_CONST_ROUNDING);
+        const __m256i s1_19_5 = _mm256_add_epi32(s1_19_3, k__DCT_CONST_ROUNDING);
+        const __m256i s1_20_4 = _mm256_add_epi32(s1_20_2, k__DCT_CONST_ROUNDING);
+        const __m256i s1_20_5 = _mm256_add_epi32(s1_20_3, k__DCT_CONST_ROUNDING);
+        const __m256i s1_21_4 = _mm256_add_epi32(s1_21_2, k__DCT_CONST_ROUNDING);
+        const __m256i s1_21_5 = _mm256_add_epi32(s1_21_3, k__DCT_CONST_ROUNDING);
+        const __m256i s1_26_4 = _mm256_add_epi32(s1_26_2, k__DCT_CONST_ROUNDING);
+        const __m256i s1_26_5 = _mm256_add_epi32(s1_26_3, k__DCT_CONST_ROUNDING);
+        const __m256i s1_27_4 = _mm256_add_epi32(s1_27_2, k__DCT_CONST_ROUNDING);
+        const __m256i s1_27_5 = _mm256_add_epi32(s1_27_3, k__DCT_CONST_ROUNDING);
+        const __m256i s1_28_4 = _mm256_add_epi32(s1_28_2, k__DCT_CONST_ROUNDING);
+        const __m256i s1_28_5 = _mm256_add_epi32(s1_28_3, k__DCT_CONST_ROUNDING);
+        const __m256i s1_29_4 = _mm256_add_epi32(s1_29_2, k__DCT_CONST_ROUNDING);
+        const __m256i s1_29_5 = _mm256_add_epi32(s1_29_3, k__DCT_CONST_ROUNDING);
+        const __m256i s1_18_6 = _mm256_srai_epi32(s1_18_4, DCT_CONST_BITS);
+        const __m256i s1_18_7 = _mm256_srai_epi32(s1_18_5, DCT_CONST_BITS);
+        const __m256i s1_19_6 = _mm256_srai_epi32(s1_19_4, DCT_CONST_BITS);
+        const __m256i s1_19_7 = _mm256_srai_epi32(s1_19_5, DCT_CONST_BITS);
+        const __m256i s1_20_6 = _mm256_srai_epi32(s1_20_4, DCT_CONST_BITS);
+        const __m256i s1_20_7 = _mm256_srai_epi32(s1_20_5, DCT_CONST_BITS);
+        const __m256i s1_21_6 = _mm256_srai_epi32(s1_21_4, DCT_CONST_BITS);
+        const __m256i s1_21_7 = _mm256_srai_epi32(s1_21_5, DCT_CONST_BITS);
+        const __m256i s1_26_6 = _mm256_srai_epi32(s1_26_4, DCT_CONST_BITS);
+        const __m256i s1_26_7 = _mm256_srai_epi32(s1_26_5, DCT_CONST_BITS);
+        const __m256i s1_27_6 = _mm256_srai_epi32(s1_27_4, DCT_CONST_BITS);
+        const __m256i s1_27_7 = _mm256_srai_epi32(s1_27_5, DCT_CONST_BITS);
+        const __m256i s1_28_6 = _mm256_srai_epi32(s1_28_4, DCT_CONST_BITS);
+        const __m256i s1_28_7 = _mm256_srai_epi32(s1_28_5, DCT_CONST_BITS);
+        const __m256i s1_29_6 = _mm256_srai_epi32(s1_29_4, DCT_CONST_BITS);
+        const __m256i s1_29_7 = _mm256_srai_epi32(s1_29_5, DCT_CONST_BITS);
+        // Combine
+        step1[18] = _mm256_packs_epi32(s1_18_6, s1_18_7);
+        step1[19] = _mm256_packs_epi32(s1_19_6, s1_19_7);
+        step1[20] = _mm256_packs_epi32(s1_20_6, s1_20_7);
+        step1[21] = _mm256_packs_epi32(s1_21_6, s1_21_7);
+        step1[26] = _mm256_packs_epi32(s1_26_6, s1_26_7);
+        step1[27] = _mm256_packs_epi32(s1_27_6, s1_27_7);
+        step1[28] = _mm256_packs_epi32(s1_28_6, s1_28_7);
+        step1[29] = _mm256_packs_epi32(s1_29_6, s1_29_7);
+      }
+      // Stage 5
+      {
+        step2[4] = _mm256_add_epi16(step1[5], step3[4]);
+        step2[5] = _mm256_sub_epi16(step3[4], step1[5]);
+        step2[6] = _mm256_sub_epi16(step3[7], step1[6]);
+        step2[7] = _mm256_add_epi16(step1[6], step3[7]);
+      }
+      {
+        const __m256i out_00_0 = _mm256_unpacklo_epi16(step1[0], step1[1]);
+        const __m256i out_00_1 = _mm256_unpackhi_epi16(step1[0], step1[1]);
+        const __m256i out_08_0 = _mm256_unpacklo_epi16(step1[2], step1[3]);
+        const __m256i out_08_1 = _mm256_unpackhi_epi16(step1[2], step1[3]);
+        const __m256i out_00_2 = _mm256_madd_epi16(out_00_0, k__cospi_p16_p16);
+        const __m256i out_00_3 = _mm256_madd_epi16(out_00_1, k__cospi_p16_p16);
+        const __m256i out_16_2 = _mm256_madd_epi16(out_00_0, k__cospi_p16_m16);
+        const __m256i out_16_3 = _mm256_madd_epi16(out_00_1, k__cospi_p16_m16);
+        const __m256i out_08_2 = _mm256_madd_epi16(out_08_0, k__cospi_p24_p08);
+        const __m256i out_08_3 = _mm256_madd_epi16(out_08_1, k__cospi_p24_p08);
+        const __m256i out_24_2 = _mm256_madd_epi16(out_08_0, k__cospi_m08_p24);
+        const __m256i out_24_3 = _mm256_madd_epi16(out_08_1, k__cospi_m08_p24);
+        // dct_const_round_shift
+        const __m256i out_00_4 = _mm256_add_epi32(out_00_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_00_5 = _mm256_add_epi32(out_00_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_16_4 = _mm256_add_epi32(out_16_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_16_5 = _mm256_add_epi32(out_16_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_08_4 = _mm256_add_epi32(out_08_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_08_5 = _mm256_add_epi32(out_08_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_24_4 = _mm256_add_epi32(out_24_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_24_5 = _mm256_add_epi32(out_24_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_00_6 = _mm256_srai_epi32(out_00_4, DCT_CONST_BITS);
+        const __m256i out_00_7 = _mm256_srai_epi32(out_00_5, DCT_CONST_BITS);
+        const __m256i out_16_6 = _mm256_srai_epi32(out_16_4, DCT_CONST_BITS);
+        const __m256i out_16_7 = _mm256_srai_epi32(out_16_5, DCT_CONST_BITS);
+        const __m256i out_08_6 = _mm256_srai_epi32(out_08_4, DCT_CONST_BITS);
+        const __m256i out_08_7 = _mm256_srai_epi32(out_08_5, DCT_CONST_BITS);
+        const __m256i out_24_6 = _mm256_srai_epi32(out_24_4, DCT_CONST_BITS);
+        const __m256i out_24_7 = _mm256_srai_epi32(out_24_5, DCT_CONST_BITS);
+        // Combine
+        out[ 0] = _mm256_packs_epi32(out_00_6, out_00_7);
+        out[16] = _mm256_packs_epi32(out_16_6, out_16_7);
+        out[ 8] = _mm256_packs_epi32(out_08_6, out_08_7);
+        out[24] = _mm256_packs_epi32(out_24_6, out_24_7);
+      }
+      {
+        const __m256i s2_09_0 = _mm256_unpacklo_epi16(step1[ 9], step1[14]);
+        const __m256i s2_09_1 = _mm256_unpackhi_epi16(step1[ 9], step1[14]);
+        const __m256i s2_10_0 = _mm256_unpacklo_epi16(step1[10], step1[13]);
+        const __m256i s2_10_1 = _mm256_unpackhi_epi16(step1[10], step1[13]);
+        const __m256i s2_09_2 = _mm256_madd_epi16(s2_09_0, k__cospi_m08_p24);
+        const __m256i s2_09_3 = _mm256_madd_epi16(s2_09_1, k__cospi_m08_p24);
+        const __m256i s2_10_2 = _mm256_madd_epi16(s2_10_0, k__cospi_m24_m08);
+        const __m256i s2_10_3 = _mm256_madd_epi16(s2_10_1, k__cospi_m24_m08);
+        const __m256i s2_13_2 = _mm256_madd_epi16(s2_10_0, k__cospi_m08_p24);
+        const __m256i s2_13_3 = _mm256_madd_epi16(s2_10_1, k__cospi_m08_p24);
+        const __m256i s2_14_2 = _mm256_madd_epi16(s2_09_0, k__cospi_p24_p08);
+        const __m256i s2_14_3 = _mm256_madd_epi16(s2_09_1, k__cospi_p24_p08);
+        // dct_const_round_shift
+        const __m256i s2_09_4 = _mm256_add_epi32(s2_09_2, k__DCT_CONST_ROUNDING);
+        const __m256i s2_09_5 = _mm256_add_epi32(s2_09_3, k__DCT_CONST_ROUNDING);
+        const __m256i s2_10_4 = _mm256_add_epi32(s2_10_2, k__DCT_CONST_ROUNDING);
+        const __m256i s2_10_5 = _mm256_add_epi32(s2_10_3, k__DCT_CONST_ROUNDING);
+        const __m256i s2_13_4 = _mm256_add_epi32(s2_13_2, k__DCT_CONST_ROUNDING);
+        const __m256i s2_13_5 = _mm256_add_epi32(s2_13_3, k__DCT_CONST_ROUNDING);
+        const __m256i s2_14_4 = _mm256_add_epi32(s2_14_2, k__DCT_CONST_ROUNDING);
+        const __m256i s2_14_5 = _mm256_add_epi32(s2_14_3, k__DCT_CONST_ROUNDING);
+        const __m256i s2_09_6 = _mm256_srai_epi32(s2_09_4, DCT_CONST_BITS);
+        const __m256i s2_09_7 = _mm256_srai_epi32(s2_09_5, DCT_CONST_BITS);
+        const __m256i s2_10_6 = _mm256_srai_epi32(s2_10_4, DCT_CONST_BITS);
+        const __m256i s2_10_7 = _mm256_srai_epi32(s2_10_5, DCT_CONST_BITS);
+        const __m256i s2_13_6 = _mm256_srai_epi32(s2_13_4, DCT_CONST_BITS);
+        const __m256i s2_13_7 = _mm256_srai_epi32(s2_13_5, DCT_CONST_BITS);
+        const __m256i s2_14_6 = _mm256_srai_epi32(s2_14_4, DCT_CONST_BITS);
+        const __m256i s2_14_7 = _mm256_srai_epi32(s2_14_5, DCT_CONST_BITS);
+        // Combine
+        step2[ 9] = _mm256_packs_epi32(s2_09_6, s2_09_7);
+        step2[10] = _mm256_packs_epi32(s2_10_6, s2_10_7);
+        step2[13] = _mm256_packs_epi32(s2_13_6, s2_13_7);
+        step2[14] = _mm256_packs_epi32(s2_14_6, s2_14_7);
+      }
+      {
+        step2[16] = _mm256_add_epi16(step1[19], step3[16]);
+        step2[17] = _mm256_add_epi16(step1[18], step3[17]);
+        step2[18] = _mm256_sub_epi16(step3[17], step1[18]);
+        step2[19] = _mm256_sub_epi16(step3[16], step1[19]);
+        step2[20] = _mm256_sub_epi16(step3[23], step1[20]);
+        step2[21] = _mm256_sub_epi16(step3[22], step1[21]);
+        step2[22] = _mm256_add_epi16(step1[21], step3[22]);
+        step2[23] = _mm256_add_epi16(step1[20], step3[23]);
+        step2[24] = _mm256_add_epi16(step1[27], step3[24]);
+        step2[25] = _mm256_add_epi16(step1[26], step3[25]);
+        step2[26] = _mm256_sub_epi16(step3[25], step1[26]);
+        step2[27] = _mm256_sub_epi16(step3[24], step1[27]);
+        step2[28] = _mm256_sub_epi16(step3[31], step1[28]);
+        step2[29] = _mm256_sub_epi16(step3[30], step1[29]);
+        step2[30] = _mm256_add_epi16(step1[29], step3[30]);
+        step2[31] = _mm256_add_epi16(step1[28], step3[31]);
+      }
+      // Stage 6
+      {
+        const __m256i out_04_0 = _mm256_unpacklo_epi16(step2[4], step2[7]);
+        const __m256i out_04_1 = _mm256_unpackhi_epi16(step2[4], step2[7]);
+        const __m256i out_20_0 = _mm256_unpacklo_epi16(step2[5], step2[6]);
+        const __m256i out_20_1 = _mm256_unpackhi_epi16(step2[5], step2[6]);
+        const __m256i out_12_0 = _mm256_unpacklo_epi16(step2[5], step2[6]);
+        const __m256i out_12_1 = _mm256_unpackhi_epi16(step2[5], step2[6]);
+        const __m256i out_28_0 = _mm256_unpacklo_epi16(step2[4], step2[7]);
+        const __m256i out_28_1 = _mm256_unpackhi_epi16(step2[4], step2[7]);
+        const __m256i out_04_2 = _mm256_madd_epi16(out_04_0, k__cospi_p28_p04);
+        const __m256i out_04_3 = _mm256_madd_epi16(out_04_1, k__cospi_p28_p04);
+        const __m256i out_20_2 = _mm256_madd_epi16(out_20_0, k__cospi_p12_p20);
+        const __m256i out_20_3 = _mm256_madd_epi16(out_20_1, k__cospi_p12_p20);
+        const __m256i out_12_2 = _mm256_madd_epi16(out_12_0, k__cospi_m20_p12);
+        const __m256i out_12_3 = _mm256_madd_epi16(out_12_1, k__cospi_m20_p12);
+        const __m256i out_28_2 = _mm256_madd_epi16(out_28_0, k__cospi_m04_p28);
+        const __m256i out_28_3 = _mm256_madd_epi16(out_28_1, k__cospi_m04_p28);
+        // dct_const_round_shift
+        const __m256i out_04_4 = _mm256_add_epi32(out_04_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_04_5 = _mm256_add_epi32(out_04_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_20_4 = _mm256_add_epi32(out_20_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_20_5 = _mm256_add_epi32(out_20_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_12_4 = _mm256_add_epi32(out_12_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_12_5 = _mm256_add_epi32(out_12_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_28_4 = _mm256_add_epi32(out_28_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_28_5 = _mm256_add_epi32(out_28_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_04_6 = _mm256_srai_epi32(out_04_4, DCT_CONST_BITS);
+        const __m256i out_04_7 = _mm256_srai_epi32(out_04_5, DCT_CONST_BITS);
+        const __m256i out_20_6 = _mm256_srai_epi32(out_20_4, DCT_CONST_BITS);
+        const __m256i out_20_7 = _mm256_srai_epi32(out_20_5, DCT_CONST_BITS);
+        const __m256i out_12_6 = _mm256_srai_epi32(out_12_4, DCT_CONST_BITS);
+        const __m256i out_12_7 = _mm256_srai_epi32(out_12_5, DCT_CONST_BITS);
+        const __m256i out_28_6 = _mm256_srai_epi32(out_28_4, DCT_CONST_BITS);
+        const __m256i out_28_7 = _mm256_srai_epi32(out_28_5, DCT_CONST_BITS);
+        // Combine
+        out[ 4] = _mm256_packs_epi32(out_04_6, out_04_7);
+        out[20] = _mm256_packs_epi32(out_20_6, out_20_7);
+        out[12] = _mm256_packs_epi32(out_12_6, out_12_7);
+        out[28] = _mm256_packs_epi32(out_28_6, out_28_7);
+      }
+      {
+        step3[ 8] = _mm256_add_epi16(step2[ 9], step1[ 8]);
+        step3[ 9] = _mm256_sub_epi16(step1[ 8], step2[ 9]);
+        step3[10] = _mm256_sub_epi16(step1[11], step2[10]);
+        step3[11] = _mm256_add_epi16(step2[10], step1[11]);
+        step3[12] = _mm256_add_epi16(step2[13], step1[12]);
+        step3[13] = _mm256_sub_epi16(step1[12], step2[13]);
+        step3[14] = _mm256_sub_epi16(step1[15], step2[14]);
+        step3[15] = _mm256_add_epi16(step2[14], step1[15]);
+      }
+      {
+        const __m256i s3_17_0 = _mm256_unpacklo_epi16(step2[17], step2[30]);
+        const __m256i s3_17_1 = _mm256_unpackhi_epi16(step2[17], step2[30]);
+        const __m256i s3_18_0 = _mm256_unpacklo_epi16(step2[18], step2[29]);
+        const __m256i s3_18_1 = _mm256_unpackhi_epi16(step2[18], step2[29]);
+        const __m256i s3_21_0 = _mm256_unpacklo_epi16(step2[21], step2[26]);
+        const __m256i s3_21_1 = _mm256_unpackhi_epi16(step2[21], step2[26]);
+        const __m256i s3_22_0 = _mm256_unpacklo_epi16(step2[22], step2[25]);
+        const __m256i s3_22_1 = _mm256_unpackhi_epi16(step2[22], step2[25]);
+        const __m256i s3_17_2 = _mm256_madd_epi16(s3_17_0, k__cospi_m04_p28);
+        const __m256i s3_17_3 = _mm256_madd_epi16(s3_17_1, k__cospi_m04_p28);
+        const __m256i s3_18_2 = _mm256_madd_epi16(s3_18_0, k__cospi_m28_m04);
+        const __m256i s3_18_3 = _mm256_madd_epi16(s3_18_1, k__cospi_m28_m04);
+        const __m256i s3_21_2 = _mm256_madd_epi16(s3_21_0, k__cospi_m20_p12);
+        const __m256i s3_21_3 = _mm256_madd_epi16(s3_21_1, k__cospi_m20_p12);
+        const __m256i s3_22_2 = _mm256_madd_epi16(s3_22_0, k__cospi_m12_m20);
+        const __m256i s3_22_3 = _mm256_madd_epi16(s3_22_1, k__cospi_m12_m20);
+        const __m256i s3_25_2 = _mm256_madd_epi16(s3_22_0, k__cospi_m20_p12);
+        const __m256i s3_25_3 = _mm256_madd_epi16(s3_22_1, k__cospi_m20_p12);
+        const __m256i s3_26_2 = _mm256_madd_epi16(s3_21_0, k__cospi_p12_p20);
+        const __m256i s3_26_3 = _mm256_madd_epi16(s3_21_1, k__cospi_p12_p20);
+        const __m256i s3_29_2 = _mm256_madd_epi16(s3_18_0, k__cospi_m04_p28);
+        const __m256i s3_29_3 = _mm256_madd_epi16(s3_18_1, k__cospi_m04_p28);
+        const __m256i s3_30_2 = _mm256_madd_epi16(s3_17_0, k__cospi_p28_p04);
+        const __m256i s3_30_3 = _mm256_madd_epi16(s3_17_1, k__cospi_p28_p04);
+        // dct_const_round_shift
+        const __m256i s3_17_4 = _mm256_add_epi32(s3_17_2, k__DCT_CONST_ROUNDING);
+        const __m256i s3_17_5 = _mm256_add_epi32(s3_17_3, k__DCT_CONST_ROUNDING);
+        const __m256i s3_18_4 = _mm256_add_epi32(s3_18_2, k__DCT_CONST_ROUNDING);
+        const __m256i s3_18_5 = _mm256_add_epi32(s3_18_3, k__DCT_CONST_ROUNDING);
+        const __m256i s3_21_4 = _mm256_add_epi32(s3_21_2, k__DCT_CONST_ROUNDING);
+        const __m256i s3_21_5 = _mm256_add_epi32(s3_21_3, k__DCT_CONST_ROUNDING);
+        const __m256i s3_22_4 = _mm256_add_epi32(s3_22_2, k__DCT_CONST_ROUNDING);
+        const __m256i s3_22_5 = _mm256_add_epi32(s3_22_3, k__DCT_CONST_ROUNDING);
+        const __m256i s3_17_6 = _mm256_srai_epi32(s3_17_4, DCT_CONST_BITS);
+        const __m256i s3_17_7 = _mm256_srai_epi32(s3_17_5, DCT_CONST_BITS);
+        const __m256i s3_18_6 = _mm256_srai_epi32(s3_18_4, DCT_CONST_BITS);
+        const __m256i s3_18_7 = _mm256_srai_epi32(s3_18_5, DCT_CONST_BITS);
+        const __m256i s3_21_6 = _mm256_srai_epi32(s3_21_4, DCT_CONST_BITS);
+        const __m256i s3_21_7 = _mm256_srai_epi32(s3_21_5, DCT_CONST_BITS);
+        const __m256i s3_22_6 = _mm256_srai_epi32(s3_22_4, DCT_CONST_BITS);
+        const __m256i s3_22_7 = _mm256_srai_epi32(s3_22_5, DCT_CONST_BITS);
+        const __m256i s3_25_4 = _mm256_add_epi32(s3_25_2, k__DCT_CONST_ROUNDING);
+        const __m256i s3_25_5 = _mm256_add_epi32(s3_25_3, k__DCT_CONST_ROUNDING);
+        const __m256i s3_26_4 = _mm256_add_epi32(s3_26_2, k__DCT_CONST_ROUNDING);
+        const __m256i s3_26_5 = _mm256_add_epi32(s3_26_3, k__DCT_CONST_ROUNDING);
+        const __m256i s3_29_4 = _mm256_add_epi32(s3_29_2, k__DCT_CONST_ROUNDING);
+        const __m256i s3_29_5 = _mm256_add_epi32(s3_29_3, k__DCT_CONST_ROUNDING);
+        const __m256i s3_30_4 = _mm256_add_epi32(s3_30_2, k__DCT_CONST_ROUNDING);
+        const __m256i s3_30_5 = _mm256_add_epi32(s3_30_3, k__DCT_CONST_ROUNDING);
+        const __m256i s3_25_6 = _mm256_srai_epi32(s3_25_4, DCT_CONST_BITS);
+        const __m256i s3_25_7 = _mm256_srai_epi32(s3_25_5, DCT_CONST_BITS);
+        const __m256i s3_26_6 = _mm256_srai_epi32(s3_26_4, DCT_CONST_BITS);
+        const __m256i s3_26_7 = _mm256_srai_epi32(s3_26_5, DCT_CONST_BITS);
+        const __m256i s3_29_6 = _mm256_srai_epi32(s3_29_4, DCT_CONST_BITS);
+        const __m256i s3_29_7 = _mm256_srai_epi32(s3_29_5, DCT_CONST_BITS);
+        const __m256i s3_30_6 = _mm256_srai_epi32(s3_30_4, DCT_CONST_BITS);
+        const __m256i s3_30_7 = _mm256_srai_epi32(s3_30_5, DCT_CONST_BITS);
+        // Combine
+        step3[17] = _mm256_packs_epi32(s3_17_6, s3_17_7);
+        step3[18] = _mm256_packs_epi32(s3_18_6, s3_18_7);
+        step3[21] = _mm256_packs_epi32(s3_21_6, s3_21_7);
+        step3[22] = _mm256_packs_epi32(s3_22_6, s3_22_7);
+        // Combine
+        step3[25] = _mm256_packs_epi32(s3_25_6, s3_25_7);
+        step3[26] = _mm256_packs_epi32(s3_26_6, s3_26_7);
+        step3[29] = _mm256_packs_epi32(s3_29_6, s3_29_7);
+        step3[30] = _mm256_packs_epi32(s3_30_6, s3_30_7);
+      }
+      // Stage 7
+      {
+        const __m256i out_02_0 = _mm256_unpacklo_epi16(step3[ 8], step3[15]);
+        const __m256i out_02_1 = _mm256_unpackhi_epi16(step3[ 8], step3[15]);
+        const __m256i out_18_0 = _mm256_unpacklo_epi16(step3[ 9], step3[14]);
+        const __m256i out_18_1 = _mm256_unpackhi_epi16(step3[ 9], step3[14]);
+        const __m256i out_10_0 = _mm256_unpacklo_epi16(step3[10], step3[13]);
+        const __m256i out_10_1 = _mm256_unpackhi_epi16(step3[10], step3[13]);
+        const __m256i out_26_0 = _mm256_unpacklo_epi16(step3[11], step3[12]);
+        const __m256i out_26_1 = _mm256_unpackhi_epi16(step3[11], step3[12]);
+        const __m256i out_02_2 = _mm256_madd_epi16(out_02_0, k__cospi_p30_p02);
+        const __m256i out_02_3 = _mm256_madd_epi16(out_02_1, k__cospi_p30_p02);
+        const __m256i out_18_2 = _mm256_madd_epi16(out_18_0, k__cospi_p14_p18);
+        const __m256i out_18_3 = _mm256_madd_epi16(out_18_1, k__cospi_p14_p18);
+        const __m256i out_10_2 = _mm256_madd_epi16(out_10_0, k__cospi_p22_p10);
+        const __m256i out_10_3 = _mm256_madd_epi16(out_10_1, k__cospi_p22_p10);
+        const __m256i out_26_2 = _mm256_madd_epi16(out_26_0, k__cospi_p06_p26);
+        const __m256i out_26_3 = _mm256_madd_epi16(out_26_1, k__cospi_p06_p26);
+        const __m256i out_06_2 = _mm256_madd_epi16(out_26_0, k__cospi_m26_p06);
+        const __m256i out_06_3 = _mm256_madd_epi16(out_26_1, k__cospi_m26_p06);
+        const __m256i out_22_2 = _mm256_madd_epi16(out_10_0, k__cospi_m10_p22);
+        const __m256i out_22_3 = _mm256_madd_epi16(out_10_1, k__cospi_m10_p22);
+        const __m256i out_14_2 = _mm256_madd_epi16(out_18_0, k__cospi_m18_p14);
+        const __m256i out_14_3 = _mm256_madd_epi16(out_18_1, k__cospi_m18_p14);
+        const __m256i out_30_2 = _mm256_madd_epi16(out_02_0, k__cospi_m02_p30);
+        const __m256i out_30_3 = _mm256_madd_epi16(out_02_1, k__cospi_m02_p30);
+        // dct_const_round_shift
+        const __m256i out_02_4 = _mm256_add_epi32(out_02_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_02_5 = _mm256_add_epi32(out_02_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_18_4 = _mm256_add_epi32(out_18_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_18_5 = _mm256_add_epi32(out_18_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_10_4 = _mm256_add_epi32(out_10_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_10_5 = _mm256_add_epi32(out_10_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_26_4 = _mm256_add_epi32(out_26_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_26_5 = _mm256_add_epi32(out_26_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_06_4 = _mm256_add_epi32(out_06_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_06_5 = _mm256_add_epi32(out_06_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_22_4 = _mm256_add_epi32(out_22_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_22_5 = _mm256_add_epi32(out_22_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_14_4 = _mm256_add_epi32(out_14_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_14_5 = _mm256_add_epi32(out_14_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_30_4 = _mm256_add_epi32(out_30_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_30_5 = _mm256_add_epi32(out_30_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_02_6 = _mm256_srai_epi32(out_02_4, DCT_CONST_BITS);
+        const __m256i out_02_7 = _mm256_srai_epi32(out_02_5, DCT_CONST_BITS);
+        const __m256i out_18_6 = _mm256_srai_epi32(out_18_4, DCT_CONST_BITS);
+        const __m256i out_18_7 = _mm256_srai_epi32(out_18_5, DCT_CONST_BITS);
+        const __m256i out_10_6 = _mm256_srai_epi32(out_10_4, DCT_CONST_BITS);
+        const __m256i out_10_7 = _mm256_srai_epi32(out_10_5, DCT_CONST_BITS);
+        const __m256i out_26_6 = _mm256_srai_epi32(out_26_4, DCT_CONST_BITS);
+        const __m256i out_26_7 = _mm256_srai_epi32(out_26_5, DCT_CONST_BITS);
+        const __m256i out_06_6 = _mm256_srai_epi32(out_06_4, DCT_CONST_BITS);
+        const __m256i out_06_7 = _mm256_srai_epi32(out_06_5, DCT_CONST_BITS);
+        const __m256i out_22_6 = _mm256_srai_epi32(out_22_4, DCT_CONST_BITS);
+        const __m256i out_22_7 = _mm256_srai_epi32(out_22_5, DCT_CONST_BITS);
+        const __m256i out_14_6 = _mm256_srai_epi32(out_14_4, DCT_CONST_BITS);
+        const __m256i out_14_7 = _mm256_srai_epi32(out_14_5, DCT_CONST_BITS);
+        const __m256i out_30_6 = _mm256_srai_epi32(out_30_4, DCT_CONST_BITS);
+        const __m256i out_30_7 = _mm256_srai_epi32(out_30_5, DCT_CONST_BITS);
+        // Combine
+        out[ 2] = _mm256_packs_epi32(out_02_6, out_02_7);
+        out[18] = _mm256_packs_epi32(out_18_6, out_18_7);
+        out[10] = _mm256_packs_epi32(out_10_6, out_10_7);
+        out[26] = _mm256_packs_epi32(out_26_6, out_26_7);
+        out[ 6] = _mm256_packs_epi32(out_06_6, out_06_7);
+        out[22] = _mm256_packs_epi32(out_22_6, out_22_7);
+        out[14] = _mm256_packs_epi32(out_14_6, out_14_7);
+        out[30] = _mm256_packs_epi32(out_30_6, out_30_7);
+      }
+      {
+        step1[16] = _mm256_add_epi16(step3[17], step2[16]);
+        step1[17] = _mm256_sub_epi16(step2[16], step3[17]);
+        step1[18] = _mm256_sub_epi16(step2[19], step3[18]);
+        step1[19] = _mm256_add_epi16(step3[18], step2[19]);
+        step1[20] = _mm256_add_epi16(step3[21], step2[20]);
+        step1[21] = _mm256_sub_epi16(step2[20], step3[21]);
+        step1[22] = _mm256_sub_epi16(step2[23], step3[22]);
+        step1[23] = _mm256_add_epi16(step3[22], step2[23]);
+        step1[24] = _mm256_add_epi16(step3[25], step2[24]);
+        step1[25] = _mm256_sub_epi16(step2[24], step3[25]);
+        step1[26] = _mm256_sub_epi16(step2[27], step3[26]);
+        step1[27] = _mm256_add_epi16(step3[26], step2[27]);
+        step1[28] = _mm256_add_epi16(step3[29], step2[28]);
+        step1[29] = _mm256_sub_epi16(step2[28], step3[29]);
+        step1[30] = _mm256_sub_epi16(step2[31], step3[30]);
+        step1[31] = _mm256_add_epi16(step3[30], step2[31]);
+      }
+      // Final stage --- outputs indices are bit-reversed.
+      {
+        const __m256i out_01_0 = _mm256_unpacklo_epi16(step1[16], step1[31]);
+        const __m256i out_01_1 = _mm256_unpackhi_epi16(step1[16], step1[31]);
+        const __m256i out_17_0 = _mm256_unpacklo_epi16(step1[17], step1[30]);
+        const __m256i out_17_1 = _mm256_unpackhi_epi16(step1[17], step1[30]);
+        const __m256i out_09_0 = _mm256_unpacklo_epi16(step1[18], step1[29]);
+        const __m256i out_09_1 = _mm256_unpackhi_epi16(step1[18], step1[29]);
+        const __m256i out_25_0 = _mm256_unpacklo_epi16(step1[19], step1[28]);
+        const __m256i out_25_1 = _mm256_unpackhi_epi16(step1[19], step1[28]);
+        const __m256i out_01_2 = _mm256_madd_epi16(out_01_0, k__cospi_p31_p01);
+        const __m256i out_01_3 = _mm256_madd_epi16(out_01_1, k__cospi_p31_p01);
+        const __m256i out_17_2 = _mm256_madd_epi16(out_17_0, k__cospi_p15_p17);
+        const __m256i out_17_3 = _mm256_madd_epi16(out_17_1, k__cospi_p15_p17);
+        const __m256i out_09_2 = _mm256_madd_epi16(out_09_0, k__cospi_p23_p09);
+        const __m256i out_09_3 = _mm256_madd_epi16(out_09_1, k__cospi_p23_p09);
+        const __m256i out_25_2 = _mm256_madd_epi16(out_25_0, k__cospi_p07_p25);
+        const __m256i out_25_3 = _mm256_madd_epi16(out_25_1, k__cospi_p07_p25);
+        const __m256i out_07_2 = _mm256_madd_epi16(out_25_0, k__cospi_m25_p07);
+        const __m256i out_07_3 = _mm256_madd_epi16(out_25_1, k__cospi_m25_p07);
+        const __m256i out_23_2 = _mm256_madd_epi16(out_09_0, k__cospi_m09_p23);
+        const __m256i out_23_3 = _mm256_madd_epi16(out_09_1, k__cospi_m09_p23);
+        const __m256i out_15_2 = _mm256_madd_epi16(out_17_0, k__cospi_m17_p15);
+        const __m256i out_15_3 = _mm256_madd_epi16(out_17_1, k__cospi_m17_p15);
+        const __m256i out_31_2 = _mm256_madd_epi16(out_01_0, k__cospi_m01_p31);
+        const __m256i out_31_3 = _mm256_madd_epi16(out_01_1, k__cospi_m01_p31);
+        // dct_const_round_shift
+        const __m256i out_01_4 = _mm256_add_epi32(out_01_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_01_5 = _mm256_add_epi32(out_01_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_17_4 = _mm256_add_epi32(out_17_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_17_5 = _mm256_add_epi32(out_17_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_09_4 = _mm256_add_epi32(out_09_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_09_5 = _mm256_add_epi32(out_09_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_25_4 = _mm256_add_epi32(out_25_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_25_5 = _mm256_add_epi32(out_25_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_07_4 = _mm256_add_epi32(out_07_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_07_5 = _mm256_add_epi32(out_07_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_23_4 = _mm256_add_epi32(out_23_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_23_5 = _mm256_add_epi32(out_23_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_15_4 = _mm256_add_epi32(out_15_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_15_5 = _mm256_add_epi32(out_15_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_31_4 = _mm256_add_epi32(out_31_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_31_5 = _mm256_add_epi32(out_31_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_01_6 = _mm256_srai_epi32(out_01_4, DCT_CONST_BITS);
+        const __m256i out_01_7 = _mm256_srai_epi32(out_01_5, DCT_CONST_BITS);
+        const __m256i out_17_6 = _mm256_srai_epi32(out_17_4, DCT_CONST_BITS);
+        const __m256i out_17_7 = _mm256_srai_epi32(out_17_5, DCT_CONST_BITS);
+        const __m256i out_09_6 = _mm256_srai_epi32(out_09_4, DCT_CONST_BITS);
+        const __m256i out_09_7 = _mm256_srai_epi32(out_09_5, DCT_CONST_BITS);
+        const __m256i out_25_6 = _mm256_srai_epi32(out_25_4, DCT_CONST_BITS);
+        const __m256i out_25_7 = _mm256_srai_epi32(out_25_5, DCT_CONST_BITS);
+        const __m256i out_07_6 = _mm256_srai_epi32(out_07_4, DCT_CONST_BITS);
+        const __m256i out_07_7 = _mm256_srai_epi32(out_07_5, DCT_CONST_BITS);
+        const __m256i out_23_6 = _mm256_srai_epi32(out_23_4, DCT_CONST_BITS);
+        const __m256i out_23_7 = _mm256_srai_epi32(out_23_5, DCT_CONST_BITS);
+        const __m256i out_15_6 = _mm256_srai_epi32(out_15_4, DCT_CONST_BITS);
+        const __m256i out_15_7 = _mm256_srai_epi32(out_15_5, DCT_CONST_BITS);
+        const __m256i out_31_6 = _mm256_srai_epi32(out_31_4, DCT_CONST_BITS);
+        const __m256i out_31_7 = _mm256_srai_epi32(out_31_5, DCT_CONST_BITS);
+        // Combine
+        out[ 1] = _mm256_packs_epi32(out_01_6, out_01_7);
+        out[17] = _mm256_packs_epi32(out_17_6, out_17_7);
+        out[ 9] = _mm256_packs_epi32(out_09_6, out_09_7);
+        out[25] = _mm256_packs_epi32(out_25_6, out_25_7);
+        out[ 7] = _mm256_packs_epi32(out_07_6, out_07_7);
+        out[23] = _mm256_packs_epi32(out_23_6, out_23_7);
+        out[15] = _mm256_packs_epi32(out_15_6, out_15_7);
+        out[31] = _mm256_packs_epi32(out_31_6, out_31_7);
+      }
+      {
+        const __m256i out_05_0 = _mm256_unpacklo_epi16(step1[20], step1[27]);
+        const __m256i out_05_1 = _mm256_unpackhi_epi16(step1[20], step1[27]);
+        const __m256i out_21_0 = _mm256_unpacklo_epi16(step1[21], step1[26]);
+        const __m256i out_21_1 = _mm256_unpackhi_epi16(step1[21], step1[26]);
+        const __m256i out_13_0 = _mm256_unpacklo_epi16(step1[22], step1[25]);
+        const __m256i out_13_1 = _mm256_unpackhi_epi16(step1[22], step1[25]);
+        const __m256i out_29_0 = _mm256_unpacklo_epi16(step1[23], step1[24]);
+        const __m256i out_29_1 = _mm256_unpackhi_epi16(step1[23], step1[24]);
+        const __m256i out_05_2 = _mm256_madd_epi16(out_05_0, k__cospi_p27_p05);
+        const __m256i out_05_3 = _mm256_madd_epi16(out_05_1, k__cospi_p27_p05);
+        const __m256i out_21_2 = _mm256_madd_epi16(out_21_0, k__cospi_p11_p21);
+        const __m256i out_21_3 = _mm256_madd_epi16(out_21_1, k__cospi_p11_p21);
+        const __m256i out_13_2 = _mm256_madd_epi16(out_13_0, k__cospi_p19_p13);
+        const __m256i out_13_3 = _mm256_madd_epi16(out_13_1, k__cospi_p19_p13);
+        const __m256i out_29_2 = _mm256_madd_epi16(out_29_0, k__cospi_p03_p29);
+        const __m256i out_29_3 = _mm256_madd_epi16(out_29_1, k__cospi_p03_p29);
+        const __m256i out_03_2 = _mm256_madd_epi16(out_29_0, k__cospi_m29_p03);
+        const __m256i out_03_3 = _mm256_madd_epi16(out_29_1, k__cospi_m29_p03);
+        const __m256i out_19_2 = _mm256_madd_epi16(out_13_0, k__cospi_m13_p19);
+        const __m256i out_19_3 = _mm256_madd_epi16(out_13_1, k__cospi_m13_p19);
+        const __m256i out_11_2 = _mm256_madd_epi16(out_21_0, k__cospi_m21_p11);
+        const __m256i out_11_3 = _mm256_madd_epi16(out_21_1, k__cospi_m21_p11);
+        const __m256i out_27_2 = _mm256_madd_epi16(out_05_0, k__cospi_m05_p27);
+        const __m256i out_27_3 = _mm256_madd_epi16(out_05_1, k__cospi_m05_p27);
+        // dct_const_round_shift
+        const __m256i out_05_4 = _mm256_add_epi32(out_05_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_05_5 = _mm256_add_epi32(out_05_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_21_4 = _mm256_add_epi32(out_21_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_21_5 = _mm256_add_epi32(out_21_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_13_4 = _mm256_add_epi32(out_13_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_13_5 = _mm256_add_epi32(out_13_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_29_4 = _mm256_add_epi32(out_29_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_29_5 = _mm256_add_epi32(out_29_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_03_4 = _mm256_add_epi32(out_03_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_03_5 = _mm256_add_epi32(out_03_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_19_4 = _mm256_add_epi32(out_19_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_19_5 = _mm256_add_epi32(out_19_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_11_4 = _mm256_add_epi32(out_11_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_11_5 = _mm256_add_epi32(out_11_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_27_4 = _mm256_add_epi32(out_27_2, k__DCT_CONST_ROUNDING);
+        const __m256i out_27_5 = _mm256_add_epi32(out_27_3, k__DCT_CONST_ROUNDING);
+        const __m256i out_05_6 = _mm256_srai_epi32(out_05_4, DCT_CONST_BITS);
+        const __m256i out_05_7 = _mm256_srai_epi32(out_05_5, DCT_CONST_BITS);
+        const __m256i out_21_6 = _mm256_srai_epi32(out_21_4, DCT_CONST_BITS);
+        const __m256i out_21_7 = _mm256_srai_epi32(out_21_5, DCT_CONST_BITS);
+        const __m256i out_13_6 = _mm256_srai_epi32(out_13_4, DCT_CONST_BITS);
+        const __m256i out_13_7 = _mm256_srai_epi32(out_13_5, DCT_CONST_BITS);
+        const __m256i out_29_6 = _mm256_srai_epi32(out_29_4, DCT_CONST_BITS);
+        const __m256i out_29_7 = _mm256_srai_epi32(out_29_5, DCT_CONST_BITS);
+        const __m256i out_03_6 = _mm256_srai_epi32(out_03_4, DCT_CONST_BITS);
+        const __m256i out_03_7 = _mm256_srai_epi32(out_03_5, DCT_CONST_BITS);
+        const __m256i out_19_6 = _mm256_srai_epi32(out_19_4, DCT_CONST_BITS);
+        const __m256i out_19_7 = _mm256_srai_epi32(out_19_5, DCT_CONST_BITS);
+        const __m256i out_11_6 = _mm256_srai_epi32(out_11_4, DCT_CONST_BITS);
+        const __m256i out_11_7 = _mm256_srai_epi32(out_11_5, DCT_CONST_BITS);
+        const __m256i out_27_6 = _mm256_srai_epi32(out_27_4, DCT_CONST_BITS);
+        const __m256i out_27_7 = _mm256_srai_epi32(out_27_5, DCT_CONST_BITS);
+        // Combine
+        out[ 5] = _mm256_packs_epi32(out_05_6, out_05_7);
+        out[21] = _mm256_packs_epi32(out_21_6, out_21_7);
+        out[13] = _mm256_packs_epi32(out_13_6, out_13_7);
+        out[29] = _mm256_packs_epi32(out_29_6, out_29_7);
+        out[ 3] = _mm256_packs_epi32(out_03_6, out_03_7);
+        out[19] = _mm256_packs_epi32(out_19_6, out_19_7);
+        out[11] = _mm256_packs_epi32(out_11_6, out_11_7);
+        out[27] = _mm256_packs_epi32(out_27_6, out_27_7);
+      }
+#if FDCT32x32_HIGH_PRECISION
+      } else {
+        __m256i lstep1[64], lstep2[64], lstep3[64];
+        __m256i u[32], v[32], sign[16];
+        const __m256i K32One = _mm256_set_epi32(1, 1, 1, 1, 1, 1, 1, 1);
+        // start using 32-bit operations
+        // stage 3
+        {
+          // expanding to 32-bit length priori to addition operations
+          lstep2[ 0] = _mm256_unpacklo_epi16(step2[ 0], kZero);
+          lstep2[ 1] = _mm256_unpackhi_epi16(step2[ 0], kZero);
+          lstep2[ 2] = _mm256_unpacklo_epi16(step2[ 1], kZero);
+          lstep2[ 3] = _mm256_unpackhi_epi16(step2[ 1], kZero);
+          lstep2[ 4] = _mm256_unpacklo_epi16(step2[ 2], kZero);
+          lstep2[ 5] = _mm256_unpackhi_epi16(step2[ 2], kZero);
+          lstep2[ 6] = _mm256_unpacklo_epi16(step2[ 3], kZero);
+          lstep2[ 7] = _mm256_unpackhi_epi16(step2[ 3], kZero);
+          lstep2[ 8] = _mm256_unpacklo_epi16(step2[ 4], kZero);
+          lstep2[ 9] = _mm256_unpackhi_epi16(step2[ 4], kZero);
+          lstep2[10] = _mm256_unpacklo_epi16(step2[ 5], kZero);
+          lstep2[11] = _mm256_unpackhi_epi16(step2[ 5], kZero);
+          lstep2[12] = _mm256_unpacklo_epi16(step2[ 6], kZero);
+          lstep2[13] = _mm256_unpackhi_epi16(step2[ 6], kZero);
+          lstep2[14] = _mm256_unpacklo_epi16(step2[ 7], kZero);
+          lstep2[15] = _mm256_unpackhi_epi16(step2[ 7], kZero);
+          lstep2[ 0] = _mm256_madd_epi16(lstep2[ 0], kOne);
+          lstep2[ 1] = _mm256_madd_epi16(lstep2[ 1], kOne);
+          lstep2[ 2] = _mm256_madd_epi16(lstep2[ 2], kOne);
+          lstep2[ 3] = _mm256_madd_epi16(lstep2[ 3], kOne);
+          lstep2[ 4] = _mm256_madd_epi16(lstep2[ 4], kOne);
+          lstep2[ 5] = _mm256_madd_epi16(lstep2[ 5], kOne);
+          lstep2[ 6] = _mm256_madd_epi16(lstep2[ 6], kOne);
+          lstep2[ 7] = _mm256_madd_epi16(lstep2[ 7], kOne);
+          lstep2[ 8] = _mm256_madd_epi16(lstep2[ 8], kOne);
+          lstep2[ 9] = _mm256_madd_epi16(lstep2[ 9], kOne);
+          lstep2[10] = _mm256_madd_epi16(lstep2[10], kOne);
+          lstep2[11] = _mm256_madd_epi16(lstep2[11], kOne);
+          lstep2[12] = _mm256_madd_epi16(lstep2[12], kOne);
+          lstep2[13] = _mm256_madd_epi16(lstep2[13], kOne);
+          lstep2[14] = _mm256_madd_epi16(lstep2[14], kOne);
+          lstep2[15] = _mm256_madd_epi16(lstep2[15], kOne);
+
+          lstep3[ 0] = _mm256_add_epi32(lstep2[14], lstep2[ 0]);
+          lstep3[ 1] = _mm256_add_epi32(lstep2[15], lstep2[ 1]);
+          lstep3[ 2] = _mm256_add_epi32(lstep2[12], lstep2[ 2]);
+          lstep3[ 3] = _mm256_add_epi32(lstep2[13], lstep2[ 3]);
+          lstep3[ 4] = _mm256_add_epi32(lstep2[10], lstep2[ 4]);
+          lstep3[ 5] = _mm256_add_epi32(lstep2[11], lstep2[ 5]);
+          lstep3[ 6] = _mm256_add_epi32(lstep2[ 8], lstep2[ 6]);
+          lstep3[ 7] = _mm256_add_epi32(lstep2[ 9], lstep2[ 7]);
+          lstep3[ 8] = _mm256_sub_epi32(lstep2[ 6], lstep2[ 8]);
+          lstep3[ 9] = _mm256_sub_epi32(lstep2[ 7], lstep2[ 9]);
+          lstep3[10] = _mm256_sub_epi32(lstep2[ 4], lstep2[10]);
+          lstep3[11] = _mm256_sub_epi32(lstep2[ 5], lstep2[11]);
+          lstep3[12] = _mm256_sub_epi32(lstep2[ 2], lstep2[12]);
+          lstep3[13] = _mm256_sub_epi32(lstep2[ 3], lstep2[13]);
+          lstep3[14] = _mm256_sub_epi32(lstep2[ 0], lstep2[14]);
+          lstep3[15] = _mm256_sub_epi32(lstep2[ 1], lstep2[15]);
+        }
+        {
+          const __m256i s3_10_0 = _mm256_unpacklo_epi16(step2[13], step2[10]);
+          const __m256i s3_10_1 = _mm256_unpackhi_epi16(step2[13], step2[10]);
+          const __m256i s3_11_0 = _mm256_unpacklo_epi16(step2[12], step2[11]);
+          const __m256i s3_11_1 = _mm256_unpackhi_epi16(step2[12], step2[11]);
+          const __m256i s3_10_2 = _mm256_madd_epi16(s3_10_0, k__cospi_p16_m16);
+          const __m256i s3_10_3 = _mm256_madd_epi16(s3_10_1, k__cospi_p16_m16);
+          const __m256i s3_11_2 = _mm256_madd_epi16(s3_11_0, k__cospi_p16_m16);
+          const __m256i s3_11_3 = _mm256_madd_epi16(s3_11_1, k__cospi_p16_m16);
+          const __m256i s3_12_2 = _mm256_madd_epi16(s3_11_0, k__cospi_p16_p16);
+          const __m256i s3_12_3 = _mm256_madd_epi16(s3_11_1, k__cospi_p16_p16);
+          const __m256i s3_13_2 = _mm256_madd_epi16(s3_10_0, k__cospi_p16_p16);
+          const __m256i s3_13_3 = _mm256_madd_epi16(s3_10_1, k__cospi_p16_p16);
+          // dct_const_round_shift
+          const __m256i s3_10_4 = _mm256_add_epi32(s3_10_2, k__DCT_CONST_ROUNDING);
+          const __m256i s3_10_5 = _mm256_add_epi32(s3_10_3, k__DCT_CONST_ROUNDING);
+          const __m256i s3_11_4 = _mm256_add_epi32(s3_11_2, k__DCT_CONST_ROUNDING);
+          const __m256i s3_11_5 = _mm256_add_epi32(s3_11_3, k__DCT_CONST_ROUNDING);
+          const __m256i s3_12_4 = _mm256_add_epi32(s3_12_2, k__DCT_CONST_ROUNDING);
+          const __m256i s3_12_5 = _mm256_add_epi32(s3_12_3, k__DCT_CONST_ROUNDING);
+          const __m256i s3_13_4 = _mm256_add_epi32(s3_13_2, k__DCT_CONST_ROUNDING);
+          const __m256i s3_13_5 = _mm256_add_epi32(s3_13_3, k__DCT_CONST_ROUNDING);
+          lstep3[20] = _mm256_srai_epi32(s3_10_4, DCT_CONST_BITS);
+          lstep3[21] = _mm256_srai_epi32(s3_10_5, DCT_CONST_BITS);
+          lstep3[22] = _mm256_srai_epi32(s3_11_4, DCT_CONST_BITS);
+          lstep3[23] = _mm256_srai_epi32(s3_11_5, DCT_CONST_BITS);
+          lstep3[24] = _mm256_srai_epi32(s3_12_4, DCT_CONST_BITS);
+          lstep3[25] = _mm256_srai_epi32(s3_12_5, DCT_CONST_BITS);
+          lstep3[26] = _mm256_srai_epi32(s3_13_4, DCT_CONST_BITS);
+          lstep3[27] = _mm256_srai_epi32(s3_13_5, DCT_CONST_BITS);
+        }
+        {
+          lstep2[40] = _mm256_unpacklo_epi16(step2[20], kZero);
+          lstep2[41] = _mm256_unpackhi_epi16(step2[20], kZero);
+          lstep2[42] = _mm256_unpacklo_epi16(step2[21], kZero);
+          lstep2[43] = _mm256_unpackhi_epi16(step2[21], kZero);
+          lstep2[44] = _mm256_unpacklo_epi16(step2[22], kZero);
+          lstep2[45] = _mm256_unpackhi_epi16(step2[22], kZero);
+          lstep2[46] = _mm256_unpacklo_epi16(step2[23], kZero);
+          lstep2[47] = _mm256_unpackhi_epi16(step2[23], kZero);
+          lstep2[48] = _mm256_unpacklo_epi16(step2[24], kZero);
+          lstep2[49] = _mm256_unpackhi_epi16(step2[24], kZero);
+          lstep2[50] = _mm256_unpacklo_epi16(step2[25], kZero);
+          lstep2[51] = _mm256_unpackhi_epi16(step2[25], kZero);
+          lstep2[52] = _mm256_unpacklo_epi16(step2[26], kZero);
+          lstep2[53] = _mm256_unpackhi_epi16(step2[26], kZero);
+          lstep2[54] = _mm256_unpacklo_epi16(step2[27], kZero);
+          lstep2[55] = _mm256_unpackhi_epi16(step2[27], kZero);
+          lstep2[40] = _mm256_madd_epi16(lstep2[40], kOne);
+          lstep2[41] = _mm256_madd_epi16(lstep2[41], kOne);
+          lstep2[42] = _mm256_madd_epi16(lstep2[42], kOne);
+          lstep2[43] = _mm256_madd_epi16(lstep2[43], kOne);
+          lstep2[44] = _mm256_madd_epi16(lstep2[44], kOne);
+          lstep2[45] = _mm256_madd_epi16(lstep2[45], kOne);
+          lstep2[46] = _mm256_madd_epi16(lstep2[46], kOne);
+          lstep2[47] = _mm256_madd_epi16(lstep2[47], kOne);
+          lstep2[48] = _mm256_madd_epi16(lstep2[48], kOne);
+          lstep2[49] = _mm256_madd_epi16(lstep2[49], kOne);
+          lstep2[50] = _mm256_madd_epi16(lstep2[50], kOne);
+          lstep2[51] = _mm256_madd_epi16(lstep2[51], kOne);
+          lstep2[52] = _mm256_madd_epi16(lstep2[52], kOne);
+          lstep2[53] = _mm256_madd_epi16(lstep2[53], kOne);
+          lstep2[54] = _mm256_madd_epi16(lstep2[54], kOne);
+          lstep2[55] = _mm256_madd_epi16(lstep2[55], kOne);
+
+          lstep1[32] = _mm256_unpacklo_epi16(step1[16], kZero);
+          lstep1[33] = _mm256_unpackhi_epi16(step1[16], kZero);
+          lstep1[34] = _mm256_unpacklo_epi16(step1[17], kZero);
+          lstep1[35] = _mm256_unpackhi_epi16(step1[17], kZero);
+          lstep1[36] = _mm256_unpacklo_epi16(step1[18], kZero);
+          lstep1[37] = _mm256_unpackhi_epi16(step1[18], kZero);
+          lstep1[38] = _mm256_unpacklo_epi16(step1[19], kZero);
+          lstep1[39] = _mm256_unpackhi_epi16(step1[19], kZero);
+          lstep1[56] = _mm256_unpacklo_epi16(step1[28], kZero);
+          lstep1[57] = _mm256_unpackhi_epi16(step1[28], kZero);
+          lstep1[58] = _mm256_unpacklo_epi16(step1[29], kZero);
+          lstep1[59] = _mm256_unpackhi_epi16(step1[29], kZero);
+          lstep1[60] = _mm256_unpacklo_epi16(step1[30], kZero);
+          lstep1[61] = _mm256_unpackhi_epi16(step1[30], kZero);
+          lstep1[62] = _mm256_unpacklo_epi16(step1[31], kZero);
+          lstep1[63] = _mm256_unpackhi_epi16(step1[31], kZero);
+          lstep1[32] = _mm256_madd_epi16(lstep1[32], kOne);
+          lstep1[33] = _mm256_madd_epi16(lstep1[33], kOne);
+          lstep1[34] = _mm256_madd_epi16(lstep1[34], kOne);
+          lstep1[35] = _mm256_madd_epi16(lstep1[35], kOne);
+          lstep1[36] = _mm256_madd_epi16(lstep1[36], kOne);
+          lstep1[37] = _mm256_madd_epi16(lstep1[37], kOne);
+          lstep1[38] = _mm256_madd_epi16(lstep1[38], kOne);
+          lstep1[39] = _mm256_madd_epi16(lstep1[39], kOne);
+          lstep1[56] = _mm256_madd_epi16(lstep1[56], kOne);
+          lstep1[57] = _mm256_madd_epi16(lstep1[57], kOne);
+          lstep1[58] = _mm256_madd_epi16(lstep1[58], kOne);
+          lstep1[59] = _mm256_madd_epi16(lstep1[59], kOne);
+          lstep1[60] = _mm256_madd_epi16(lstep1[60], kOne);
+          lstep1[61] = _mm256_madd_epi16(lstep1[61], kOne);
+          lstep1[62] = _mm256_madd_epi16(lstep1[62], kOne);
+          lstep1[63] = _mm256_madd_epi16(lstep1[63], kOne);
+
+          lstep3[32] = _mm256_add_epi32(lstep2[46], lstep1[32]);
+          lstep3[33] = _mm256_add_epi32(lstep2[47], lstep1[33]);
+
+          lstep3[34] = _mm256_add_epi32(lstep2[44], lstep1[34]);
+          lstep3[35] = _mm256_add_epi32(lstep2[45], lstep1[35]);
+          lstep3[36] = _mm256_add_epi32(lstep2[42], lstep1[36]);
+          lstep3[37] = _mm256_add_epi32(lstep2[43], lstep1[37]);
+          lstep3[38] = _mm256_add_epi32(lstep2[40], lstep1[38]);
+          lstep3[39] = _mm256_add_epi32(lstep2[41], lstep1[39]);
+          lstep3[40] = _mm256_sub_epi32(lstep1[38], lstep2[40]);
+          lstep3[41] = _mm256_sub_epi32(lstep1[39], lstep2[41]);
+          lstep3[42] = _mm256_sub_epi32(lstep1[36], lstep2[42]);
+          lstep3[43] = _mm256_sub_epi32(lstep1[37], lstep2[43]);
+          lstep3[44] = _mm256_sub_epi32(lstep1[34], lstep2[44]);
+          lstep3[45] = _mm256_sub_epi32(lstep1[35], lstep2[45]);
+          lstep3[46] = _mm256_sub_epi32(lstep1[32], lstep2[46]);
+          lstep3[47] = _mm256_sub_epi32(lstep1[33], lstep2[47]);
+          lstep3[48] = _mm256_sub_epi32(lstep1[62], lstep2[48]);
+          lstep3[49] = _mm256_sub_epi32(lstep1[63], lstep2[49]);
+          lstep3[50] = _mm256_sub_epi32(lstep1[60], lstep2[50]);
+          lstep3[51] = _mm256_sub_epi32(lstep1[61], lstep2[51]);
+          lstep3[52] = _mm256_sub_epi32(lstep1[58], lstep2[52]);
+          lstep3[53] = _mm256_sub_epi32(lstep1[59], lstep2[53]);
+          lstep3[54] = _mm256_sub_epi32(lstep1[56], lstep2[54]);
+          lstep3[55] = _mm256_sub_epi32(lstep1[57], lstep2[55]);
+          lstep3[56] = _mm256_add_epi32(lstep2[54], lstep1[56]);
+          lstep3[57] = _mm256_add_epi32(lstep2[55], lstep1[57]);
+          lstep3[58] = _mm256_add_epi32(lstep2[52], lstep1[58]);
+          lstep3[59] = _mm256_add_epi32(lstep2[53], lstep1[59]);
+          lstep3[60] = _mm256_add_epi32(lstep2[50], lstep1[60]);
+          lstep3[61] = _mm256_add_epi32(lstep2[51], lstep1[61]);
+          lstep3[62] = _mm256_add_epi32(lstep2[48], lstep1[62]);
+          lstep3[63] = _mm256_add_epi32(lstep2[49], lstep1[63]);
+        }
+
+        // stage 4
+        {
+          // expanding to 32-bit length priori to addition operations
+          lstep2[16] = _mm256_unpacklo_epi16(step2[ 8], kZero);
+          lstep2[17] = _mm256_unpackhi_epi16(step2[ 8], kZero);
+          lstep2[18] = _mm256_unpacklo_epi16(step2[ 9], kZero);
+          lstep2[19] = _mm256_unpackhi_epi16(step2[ 9], kZero);
+          lstep2[28] = _mm256_unpacklo_epi16(step2[14], kZero);
+          lstep2[29] = _mm256_unpackhi_epi16(step2[14], kZero);
+          lstep2[30] = _mm256_unpacklo_epi16(step2[15], kZero);
+          lstep2[31] = _mm256_unpackhi_epi16(step2[15], kZero);
+          lstep2[16] = _mm256_madd_epi16(lstep2[16], kOne);
+          lstep2[17] = _mm256_madd_epi16(lstep2[17], kOne);
+          lstep2[18] = _mm256_madd_epi16(lstep2[18], kOne);
+          lstep2[19] = _mm256_madd_epi16(lstep2[19], kOne);
+          lstep2[28] = _mm256_madd_epi16(lstep2[28], kOne);
+          lstep2[29] = _mm256_madd_epi16(lstep2[29], kOne);
+          lstep2[30] = _mm256_madd_epi16(lstep2[30], kOne);
+          lstep2[31] = _mm256_madd_epi16(lstep2[31], kOne);
+
+          lstep1[ 0] = _mm256_add_epi32(lstep3[ 6], lstep3[ 0]);
+          lstep1[ 1] = _mm256_add_epi32(lstep3[ 7], lstep3[ 1]);
+          lstep1[ 2] = _mm256_add_epi32(lstep3[ 4], lstep3[ 2]);
+          lstep1[ 3] = _mm256_add_epi32(lstep3[ 5], lstep3[ 3]);
+          lstep1[ 4] = _mm256_sub_epi32(lstep3[ 2], lstep3[ 4]);
+          lstep1[ 5] = _mm256_sub_epi32(lstep3[ 3], lstep3[ 5]);
+          lstep1[ 6] = _mm256_sub_epi32(lstep3[ 0], lstep3[ 6]);
+          lstep1[ 7] = _mm256_sub_epi32(lstep3[ 1], lstep3[ 7]);
+          lstep1[16] = _mm256_add_epi32(lstep3[22], lstep2[16]);
+          lstep1[17] = _mm256_add_epi32(lstep3[23], lstep2[17]);
+          lstep1[18] = _mm256_add_epi32(lstep3[20], lstep2[18]);
+          lstep1[19] = _mm256_add_epi32(lstep3[21], lstep2[19]);
+          lstep1[20] = _mm256_sub_epi32(lstep2[18], lstep3[20]);
+          lstep1[21] = _mm256_sub_epi32(lstep2[19], lstep3[21]);
+          lstep1[22] = _mm256_sub_epi32(lstep2[16], lstep3[22]);
+          lstep1[23] = _mm256_sub_epi32(lstep2[17], lstep3[23]);
+          lstep1[24] = _mm256_sub_epi32(lstep2[30], lstep3[24]);
+          lstep1[25] = _mm256_sub_epi32(lstep2[31], lstep3[25]);
+          lstep1[26] = _mm256_sub_epi32(lstep2[28], lstep3[26]);
+          lstep1[27] = _mm256_sub_epi32(lstep2[29], lstep3[27]);
+          lstep1[28] = _mm256_add_epi32(lstep3[26], lstep2[28]);
+          lstep1[29] = _mm256_add_epi32(lstep3[27], lstep2[29]);
+          lstep1[30] = _mm256_add_epi32(lstep3[24], lstep2[30]);
+          lstep1[31] = _mm256_add_epi32(lstep3[25], lstep2[31]);
+        }
+        {
+        // to be continued...
+        //
+        const __m256i k32_p16_p16 = pair256_set_epi32(cospi_16_64, cospi_16_64);
+        const __m256i k32_p16_m16 = pair256_set_epi32(cospi_16_64, -cospi_16_64);
+
+        u[0] = _mm256_unpacklo_epi32(lstep3[12], lstep3[10]);
+        u[1] = _mm256_unpackhi_epi32(lstep3[12], lstep3[10]);
+        u[2] = _mm256_unpacklo_epi32(lstep3[13], lstep3[11]);
+        u[3] = _mm256_unpackhi_epi32(lstep3[13], lstep3[11]);
+
+        // TODO(jingning): manually inline k_madd_epi32_avx2_ to further hide
+        // instruction latency.
+        v[ 0] = k_madd_epi32_avx2(u[0], k32_p16_m16);
+        v[ 1] = k_madd_epi32_avx2(u[1], k32_p16_m16);
+        v[ 2] = k_madd_epi32_avx2(u[2], k32_p16_m16);
+        v[ 3] = k_madd_epi32_avx2(u[3], k32_p16_m16);
+        v[ 4] = k_madd_epi32_avx2(u[0], k32_p16_p16);
+        v[ 5] = k_madd_epi32_avx2(u[1], k32_p16_p16);
+        v[ 6] = k_madd_epi32_avx2(u[2], k32_p16_p16);
+        v[ 7] = k_madd_epi32_avx2(u[3], k32_p16_p16);
+
+        u[0] = k_packs_epi64_avx2(v[0], v[1]);
+        u[1] = k_packs_epi64_avx2(v[2], v[3]);
+        u[2] = k_packs_epi64_avx2(v[4], v[5]);
+        u[3] = k_packs_epi64_avx2(v[6], v[7]);
+
+        v[0] = _mm256_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+        v[1] = _mm256_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+        v[2] = _mm256_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+        v[3] = _mm256_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+
+        lstep1[10] = _mm256_srai_epi32(v[0], DCT_CONST_BITS);
+        lstep1[11] = _mm256_srai_epi32(v[1], DCT_CONST_BITS);
+        lstep1[12] = _mm256_srai_epi32(v[2], DCT_CONST_BITS);
+        lstep1[13] = _mm256_srai_epi32(v[3], DCT_CONST_BITS);
+        }
+        {
+          const __m256i k32_m08_p24 = pair256_set_epi32(-cospi_8_64, cospi_24_64);
+          const __m256i k32_m24_m08 = pair256_set_epi32(-cospi_24_64, -cospi_8_64);
+          const __m256i k32_p24_p08 = pair256_set_epi32(cospi_24_64, cospi_8_64);
+
+          u[ 0] = _mm256_unpacklo_epi32(lstep3[36], lstep3[58]);
+          u[ 1] = _mm256_unpackhi_epi32(lstep3[36], lstep3[58]);
+          u[ 2] = _mm256_unpacklo_epi32(lstep3[37], lstep3[59]);
+          u[ 3] = _mm256_unpackhi_epi32(lstep3[37], lstep3[59]);
+          u[ 4] = _mm256_unpacklo_epi32(lstep3[38], lstep3[56]);
+          u[ 5] = _mm256_unpackhi_epi32(lstep3[38], lstep3[56]);
+          u[ 6] = _mm256_unpacklo_epi32(lstep3[39], lstep3[57]);
+          u[ 7] = _mm256_unpackhi_epi32(lstep3[39], lstep3[57]);
+          u[ 8] = _mm256_unpacklo_epi32(lstep3[40], lstep3[54]);
+          u[ 9] = _mm256_unpackhi_epi32(lstep3[40], lstep3[54]);
+          u[10] = _mm256_unpacklo_epi32(lstep3[41], lstep3[55]);
+          u[11] = _mm256_unpackhi_epi32(lstep3[41], lstep3[55]);
+          u[12] = _mm256_unpacklo_epi32(lstep3[42], lstep3[52]);
+          u[13] = _mm256_unpackhi_epi32(lstep3[42], lstep3[52]);
+          u[14] = _mm256_unpacklo_epi32(lstep3[43], lstep3[53]);
+          u[15] = _mm256_unpackhi_epi32(lstep3[43], lstep3[53]);
+
+          v[ 0] = k_madd_epi32_avx2(u[ 0], k32_m08_p24);
+          v[ 1] = k_madd_epi32_avx2(u[ 1], k32_m08_p24);
+          v[ 2] = k_madd_epi32_avx2(u[ 2], k32_m08_p24);
+          v[ 3] = k_madd_epi32_avx2(u[ 3], k32_m08_p24);
+          v[ 4] = k_madd_epi32_avx2(u[ 4], k32_m08_p24);
+          v[ 5] = k_madd_epi32_avx2(u[ 5], k32_m08_p24);
+          v[ 6] = k_madd_epi32_avx2(u[ 6], k32_m08_p24);
+          v[ 7] = k_madd_epi32_avx2(u[ 7], k32_m08_p24);
+          v[ 8] = k_madd_epi32_avx2(u[ 8], k32_m24_m08);
+          v[ 9] = k_madd_epi32_avx2(u[ 9], k32_m24_m08);
+          v[10] = k_madd_epi32_avx2(u[10], k32_m24_m08);
+          v[11] = k_madd_epi32_avx2(u[11], k32_m24_m08);
+          v[12] = k_madd_epi32_avx2(u[12], k32_m24_m08);
+          v[13] = k_madd_epi32_avx2(u[13], k32_m24_m08);
+          v[14] = k_madd_epi32_avx2(u[14], k32_m24_m08);
+          v[15] = k_madd_epi32_avx2(u[15], k32_m24_m08);
+          v[16] = k_madd_epi32_avx2(u[12], k32_m08_p24);
+          v[17] = k_madd_epi32_avx2(u[13], k32_m08_p24);
+          v[18] = k_madd_epi32_avx2(u[14], k32_m08_p24);
+          v[19] = k_madd_epi32_avx2(u[15], k32_m08_p24);
+          v[20] = k_madd_epi32_avx2(u[ 8], k32_m08_p24);
+          v[21] = k_madd_epi32_avx2(u[ 9], k32_m08_p24);
+          v[22] = k_madd_epi32_avx2(u[10], k32_m08_p24);
+          v[23] = k_madd_epi32_avx2(u[11], k32_m08_p24);
+          v[24] = k_madd_epi32_avx2(u[ 4], k32_p24_p08);
+          v[25] = k_madd_epi32_avx2(u[ 5], k32_p24_p08);
+          v[26] = k_madd_epi32_avx2(u[ 6], k32_p24_p08);
+          v[27] = k_madd_epi32_avx2(u[ 7], k32_p24_p08);
+          v[28] = k_madd_epi32_avx2(u[ 0], k32_p24_p08);
+          v[29] = k_madd_epi32_avx2(u[ 1], k32_p24_p08);
+          v[30] = k_madd_epi32_avx2(u[ 2], k32_p24_p08);
+          v[31] = k_madd_epi32_avx2(u[ 3], k32_p24_p08);
+
+          u[ 0] = k_packs_epi64_avx2(v[ 0], v[ 1]);
+          u[ 1] = k_packs_epi64_avx2(v[ 2], v[ 3]);
+          u[ 2] = k_packs_epi64_avx2(v[ 4], v[ 5]);
+          u[ 3] = k_packs_epi64_avx2(v[ 6], v[ 7]);
+          u[ 4] = k_packs_epi64_avx2(v[ 8], v[ 9]);
+          u[ 5] = k_packs_epi64_avx2(v[10], v[11]);
+          u[ 6] = k_packs_epi64_avx2(v[12], v[13]);
+          u[ 7] = k_packs_epi64_avx2(v[14], v[15]);
+          u[ 8] = k_packs_epi64_avx2(v[16], v[17]);
+          u[ 9] = k_packs_epi64_avx2(v[18], v[19]);
+          u[10] = k_packs_epi64_avx2(v[20], v[21]);
+          u[11] = k_packs_epi64_avx2(v[22], v[23]);
+          u[12] = k_packs_epi64_avx2(v[24], v[25]);
+          u[13] = k_packs_epi64_avx2(v[26], v[27]);
+          u[14] = k_packs_epi64_avx2(v[28], v[29]);
+          u[15] = k_packs_epi64_avx2(v[30], v[31]);
+
+          v[ 0] = _mm256_add_epi32(u[ 0], k__DCT_CONST_ROUNDING);
+          v[ 1] = _mm256_add_epi32(u[ 1], k__DCT_CONST_ROUNDING);
+          v[ 2] = _mm256_add_epi32(u[ 2], k__DCT_CONST_ROUNDING);
+          v[ 3] = _mm256_add_epi32(u[ 3], k__DCT_CONST_ROUNDING);
+          v[ 4] = _mm256_add_epi32(u[ 4], k__DCT_CONST_ROUNDING);
+          v[ 5] = _mm256_add_epi32(u[ 5], k__DCT_CONST_ROUNDING);
+          v[ 6] = _mm256_add_epi32(u[ 6], k__DCT_CONST_ROUNDING);
+          v[ 7] = _mm256_add_epi32(u[ 7], k__DCT_CONST_ROUNDING);
+          v[ 8] = _mm256_add_epi32(u[ 8], k__DCT_CONST_ROUNDING);
+          v[ 9] = _mm256_add_epi32(u[ 9], k__DCT_CONST_ROUNDING);
+          v[10] = _mm256_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+          v[11] = _mm256_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+          v[12] = _mm256_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+          v[13] = _mm256_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+          v[14] = _mm256_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+          v[15] = _mm256_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+          lstep1[36] = _mm256_srai_epi32(v[ 0], DCT_CONST_BITS);
+          lstep1[37] = _mm256_srai_epi32(v[ 1], DCT_CONST_BITS);
+          lstep1[38] = _mm256_srai_epi32(v[ 2], DCT_CONST_BITS);
+          lstep1[39] = _mm256_srai_epi32(v[ 3], DCT_CONST_BITS);
+          lstep1[40] = _mm256_srai_epi32(v[ 4], DCT_CONST_BITS);
+          lstep1[41] = _mm256_srai_epi32(v[ 5], DCT_CONST_BITS);
+          lstep1[42] = _mm256_srai_epi32(v[ 6], DCT_CONST_BITS);
+          lstep1[43] = _mm256_srai_epi32(v[ 7], DCT_CONST_BITS);
+          lstep1[52] = _mm256_srai_epi32(v[ 8], DCT_CONST_BITS);
+          lstep1[53] = _mm256_srai_epi32(v[ 9], DCT_CONST_BITS);
+          lstep1[54] = _mm256_srai_epi32(v[10], DCT_CONST_BITS);
+          lstep1[55] = _mm256_srai_epi32(v[11], DCT_CONST_BITS);
+          lstep1[56] = _mm256_srai_epi32(v[12], DCT_CONST_BITS);
+          lstep1[57] = _mm256_srai_epi32(v[13], DCT_CONST_BITS);
+          lstep1[58] = _mm256_srai_epi32(v[14], DCT_CONST_BITS);
+          lstep1[59] = _mm256_srai_epi32(v[15], DCT_CONST_BITS);
+        }
+        // stage 5
+        {
+          lstep2[ 8] = _mm256_add_epi32(lstep1[10], lstep3[ 8]);
+          lstep2[ 9] = _mm256_add_epi32(lstep1[11], lstep3[ 9]);
+          lstep2[10] = _mm256_sub_epi32(lstep3[ 8], lstep1[10]);
+          lstep2[11] = _mm256_sub_epi32(lstep3[ 9], lstep1[11]);
+          lstep2[12] = _mm256_sub_epi32(lstep3[14], lstep1[12]);
+          lstep2[13] = _mm256_sub_epi32(lstep3[15], lstep1[13]);
+          lstep2[14] = _mm256_add_epi32(lstep1[12], lstep3[14]);
+          lstep2[15] = _mm256_add_epi32(lstep1[13], lstep3[15]);
+        }
+        {
+          const __m256i k32_p16_p16 = pair256_set_epi32(cospi_16_64, cospi_16_64);
+          const __m256i k32_p16_m16 = pair256_set_epi32(cospi_16_64, -cospi_16_64);
+          const __m256i k32_p24_p08 = pair256_set_epi32(cospi_24_64, cospi_8_64);
+          const __m256i k32_m08_p24 = pair256_set_epi32(-cospi_8_64, cospi_24_64);
+
+          u[0] = _mm256_unpacklo_epi32(lstep1[0], lstep1[2]);
+          u[1] = _mm256_unpackhi_epi32(lstep1[0], lstep1[2]);
+          u[2] = _mm256_unpacklo_epi32(lstep1[1], lstep1[3]);
+          u[3] = _mm256_unpackhi_epi32(lstep1[1], lstep1[3]);
+          u[4] = _mm256_unpacklo_epi32(lstep1[4], lstep1[6]);
+          u[5] = _mm256_unpackhi_epi32(lstep1[4], lstep1[6]);
+          u[6] = _mm256_unpacklo_epi32(lstep1[5], lstep1[7]);
+          u[7] = _mm256_unpackhi_epi32(lstep1[5], lstep1[7]);
+
+          // TODO(jingning): manually inline k_madd_epi32_avx2_ to further hide
+          // instruction latency.
+          v[ 0] = k_madd_epi32_avx2(u[0], k32_p16_p16);
+          v[ 1] = k_madd_epi32_avx2(u[1], k32_p16_p16);
+          v[ 2] = k_madd_epi32_avx2(u[2], k32_p16_p16);
+          v[ 3] = k_madd_epi32_avx2(u[3], k32_p16_p16);
+          v[ 4] = k_madd_epi32_avx2(u[0], k32_p16_m16);
+          v[ 5] = k_madd_epi32_avx2(u[1], k32_p16_m16);
+          v[ 6] = k_madd_epi32_avx2(u[2], k32_p16_m16);
+          v[ 7] = k_madd_epi32_avx2(u[3], k32_p16_m16);
+          v[ 8] = k_madd_epi32_avx2(u[4], k32_p24_p08);
+          v[ 9] = k_madd_epi32_avx2(u[5], k32_p24_p08);
+          v[10] = k_madd_epi32_avx2(u[6], k32_p24_p08);
+          v[11] = k_madd_epi32_avx2(u[7], k32_p24_p08);
+          v[12] = k_madd_epi32_avx2(u[4], k32_m08_p24);
+          v[13] = k_madd_epi32_avx2(u[5], k32_m08_p24);
+          v[14] = k_madd_epi32_avx2(u[6], k32_m08_p24);
+          v[15] = k_madd_epi32_avx2(u[7], k32_m08_p24);
+
+          u[0] = k_packs_epi64_avx2(v[0], v[1]);
+          u[1] = k_packs_epi64_avx2(v[2], v[3]);
+          u[2] = k_packs_epi64_avx2(v[4], v[5]);
+          u[3] = k_packs_epi64_avx2(v[6], v[7]);
+          u[4] = k_packs_epi64_avx2(v[8], v[9]);
+          u[5] = k_packs_epi64_avx2(v[10], v[11]);
+          u[6] = k_packs_epi64_avx2(v[12], v[13]);
+          u[7] = k_packs_epi64_avx2(v[14], v[15]);
+
+          v[0] = _mm256_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+          v[1] = _mm256_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+          v[2] = _mm256_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+          v[3] = _mm256_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+          v[4] = _mm256_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+          v[5] = _mm256_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+          v[6] = _mm256_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+          v[7] = _mm256_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+
+          u[0] = _mm256_srai_epi32(v[0], DCT_CONST_BITS);
+          u[1] = _mm256_srai_epi32(v[1], DCT_CONST_BITS);
+          u[2] = _mm256_srai_epi32(v[2], DCT_CONST_BITS);
+          u[3] = _mm256_srai_epi32(v[3], DCT_CONST_BITS);
+          u[4] = _mm256_srai_epi32(v[4], DCT_CONST_BITS);
+          u[5] = _mm256_srai_epi32(v[5], DCT_CONST_BITS);
+          u[6] = _mm256_srai_epi32(v[6], DCT_CONST_BITS);
+          u[7] = _mm256_srai_epi32(v[7], DCT_CONST_BITS);
+
+          sign[0] = _mm256_cmpgt_epi32(kZero,u[0]);
+          sign[1] = _mm256_cmpgt_epi32(kZero,u[1]);
+          sign[2] = _mm256_cmpgt_epi32(kZero,u[2]);
+          sign[3] = _mm256_cmpgt_epi32(kZero,u[3]);
+          sign[4] = _mm256_cmpgt_epi32(kZero,u[4]);
+          sign[5] = _mm256_cmpgt_epi32(kZero,u[5]);
+          sign[6] = _mm256_cmpgt_epi32(kZero,u[6]);
+          sign[7] = _mm256_cmpgt_epi32(kZero,u[7]);
+
+          u[0] = _mm256_sub_epi32(u[0], sign[0]);
+          u[1] = _mm256_sub_epi32(u[1], sign[1]);
+          u[2] = _mm256_sub_epi32(u[2], sign[2]);
+          u[3] = _mm256_sub_epi32(u[3], sign[3]);
+          u[4] = _mm256_sub_epi32(u[4], sign[4]);
+          u[5] = _mm256_sub_epi32(u[5], sign[5]);
+          u[6] = _mm256_sub_epi32(u[6], sign[6]);
+          u[7] = _mm256_sub_epi32(u[7], sign[7]);
+
+          u[0] = _mm256_add_epi32(u[0], K32One);
+          u[1] = _mm256_add_epi32(u[1], K32One);
+          u[2] = _mm256_add_epi32(u[2], K32One);
+          u[3] = _mm256_add_epi32(u[3], K32One);
+          u[4] = _mm256_add_epi32(u[4], K32One);
+          u[5] = _mm256_add_epi32(u[5], K32One);
+          u[6] = _mm256_add_epi32(u[6], K32One);
+          u[7] = _mm256_add_epi32(u[7], K32One);
+
+          u[0] = _mm256_srai_epi32(u[0], 2);
+          u[1] = _mm256_srai_epi32(u[1], 2);
+          u[2] = _mm256_srai_epi32(u[2], 2);
+          u[3] = _mm256_srai_epi32(u[3], 2);
+          u[4] = _mm256_srai_epi32(u[4], 2);
+          u[5] = _mm256_srai_epi32(u[5], 2);
+          u[6] = _mm256_srai_epi32(u[6], 2);
+          u[7] = _mm256_srai_epi32(u[7], 2);
+
+          // Combine
+          out[ 0] = _mm256_packs_epi32(u[0], u[1]);
+          out[16] = _mm256_packs_epi32(u[2], u[3]);
+          out[ 8] = _mm256_packs_epi32(u[4], u[5]);
+          out[24] = _mm256_packs_epi32(u[6], u[7]);
+        }
+        {
+          const __m256i k32_m08_p24 = pair256_set_epi32(-cospi_8_64, cospi_24_64);
+          const __m256i k32_m24_m08 = pair256_set_epi32(-cospi_24_64, -cospi_8_64);
+          const __m256i k32_p24_p08 = pair256_set_epi32(cospi_24_64, cospi_8_64);
+
+          u[0] = _mm256_unpacklo_epi32(lstep1[18], lstep1[28]);
+          u[1] = _mm256_unpackhi_epi32(lstep1[18], lstep1[28]);
+          u[2] = _mm256_unpacklo_epi32(lstep1[19], lstep1[29]);
+          u[3] = _mm256_unpackhi_epi32(lstep1[19], lstep1[29]);
+          u[4] = _mm256_unpacklo_epi32(lstep1[20], lstep1[26]);
+          u[5] = _mm256_unpackhi_epi32(lstep1[20], lstep1[26]);
+          u[6] = _mm256_unpacklo_epi32(lstep1[21], lstep1[27]);
+          u[7] = _mm256_unpackhi_epi32(lstep1[21], lstep1[27]);
+
+          v[0] = k_madd_epi32_avx2(u[0], k32_m08_p24);
+          v[1] = k_madd_epi32_avx2(u[1], k32_m08_p24);
+          v[2] = k_madd_epi32_avx2(u[2], k32_m08_p24);
+          v[3] = k_madd_epi32_avx2(u[3], k32_m08_p24);
+          v[4] = k_madd_epi32_avx2(u[4], k32_m24_m08);
+          v[5] = k_madd_epi32_avx2(u[5], k32_m24_m08);
+          v[6] = k_madd_epi32_avx2(u[6], k32_m24_m08);
+          v[7] = k_madd_epi32_avx2(u[7], k32_m24_m08);
+          v[ 8] = k_madd_epi32_avx2(u[4], k32_m08_p24);
+          v[ 9] = k_madd_epi32_avx2(u[5], k32_m08_p24);
+          v[10] = k_madd_epi32_avx2(u[6], k32_m08_p24);
+          v[11] = k_madd_epi32_avx2(u[7], k32_m08_p24);
+          v[12] = k_madd_epi32_avx2(u[0], k32_p24_p08);
+          v[13] = k_madd_epi32_avx2(u[1], k32_p24_p08);
+          v[14] = k_madd_epi32_avx2(u[2], k32_p24_p08);
+          v[15] = k_madd_epi32_avx2(u[3], k32_p24_p08);
+
+          u[0] = k_packs_epi64_avx2(v[0], v[1]);
+          u[1] = k_packs_epi64_avx2(v[2], v[3]);
+          u[2] = k_packs_epi64_avx2(v[4], v[5]);
+          u[3] = k_packs_epi64_avx2(v[6], v[7]);
+          u[4] = k_packs_epi64_avx2(v[8], v[9]);
+          u[5] = k_packs_epi64_avx2(v[10], v[11]);
+          u[6] = k_packs_epi64_avx2(v[12], v[13]);
+          u[7] = k_packs_epi64_avx2(v[14], v[15]);
+
+          u[0] = _mm256_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+          u[1] = _mm256_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+          u[2] = _mm256_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+          u[3] = _mm256_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+          u[4] = _mm256_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+          u[5] = _mm256_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+          u[6] = _mm256_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+          u[7] = _mm256_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+
+          lstep2[18] = _mm256_srai_epi32(u[0], DCT_CONST_BITS);
+          lstep2[19] = _mm256_srai_epi32(u[1], DCT_CONST_BITS);
+          lstep2[20] = _mm256_srai_epi32(u[2], DCT_CONST_BITS);
+          lstep2[21] = _mm256_srai_epi32(u[3], DCT_CONST_BITS);
+          lstep2[26] = _mm256_srai_epi32(u[4], DCT_CONST_BITS);
+          lstep2[27] = _mm256_srai_epi32(u[5], DCT_CONST_BITS);
+          lstep2[28] = _mm256_srai_epi32(u[6], DCT_CONST_BITS);
+          lstep2[29] = _mm256_srai_epi32(u[7], DCT_CONST_BITS);
+        }
+        {
+          lstep2[32] = _mm256_add_epi32(lstep1[38], lstep3[32]);
+          lstep2[33] = _mm256_add_epi32(lstep1[39], lstep3[33]);
+          lstep2[34] = _mm256_add_epi32(lstep1[36], lstep3[34]);
+          lstep2[35] = _mm256_add_epi32(lstep1[37], lstep3[35]);
+          lstep2[36] = _mm256_sub_epi32(lstep3[34], lstep1[36]);
+          lstep2[37] = _mm256_sub_epi32(lstep3[35], lstep1[37]);
+          lstep2[38] = _mm256_sub_epi32(lstep3[32], lstep1[38]);
+          lstep2[39] = _mm256_sub_epi32(lstep3[33], lstep1[39]);
+          lstep2[40] = _mm256_sub_epi32(lstep3[46], lstep1[40]);
+          lstep2[41] = _mm256_sub_epi32(lstep3[47], lstep1[41]);
+          lstep2[42] = _mm256_sub_epi32(lstep3[44], lstep1[42]);
+          lstep2[43] = _mm256_sub_epi32(lstep3[45], lstep1[43]);
+          lstep2[44] = _mm256_add_epi32(lstep1[42], lstep3[44]);
+          lstep2[45] = _mm256_add_epi32(lstep1[43], lstep3[45]);
+          lstep2[46] = _mm256_add_epi32(lstep1[40], lstep3[46]);
+          lstep2[47] = _mm256_add_epi32(lstep1[41], lstep3[47]);
+          lstep2[48] = _mm256_add_epi32(lstep1[54], lstep3[48]);
+          lstep2[49] = _mm256_add_epi32(lstep1[55], lstep3[49]);
+          lstep2[50] = _mm256_add_epi32(lstep1[52], lstep3[50]);
+          lstep2[51] = _mm256_add_epi32(lstep1[53], lstep3[51]);
+          lstep2[52] = _mm256_sub_epi32(lstep3[50], lstep1[52]);
+          lstep2[53] = _mm256_sub_epi32(lstep3[51], lstep1[53]);
+          lstep2[54] = _mm256_sub_epi32(lstep3[48], lstep1[54]);
+          lstep2[55] = _mm256_sub_epi32(lstep3[49], lstep1[55]);
+          lstep2[56] = _mm256_sub_epi32(lstep3[62], lstep1[56]);
+          lstep2[57] = _mm256_sub_epi32(lstep3[63], lstep1[57]);
+          lstep2[58] = _mm256_sub_epi32(lstep3[60], lstep1[58]);
+          lstep2[59] = _mm256_sub_epi32(lstep3[61], lstep1[59]);
+          lstep2[60] = _mm256_add_epi32(lstep1[58], lstep3[60]);
+          lstep2[61] = _mm256_add_epi32(lstep1[59], lstep3[61]);
+          lstep2[62] = _mm256_add_epi32(lstep1[56], lstep3[62]);
+          lstep2[63] = _mm256_add_epi32(lstep1[57], lstep3[63]);
+        }
+        // stage 6
+        {
+          const __m256i k32_p28_p04 = pair256_set_epi32(cospi_28_64, cospi_4_64);
+          const __m256i k32_p12_p20 = pair256_set_epi32(cospi_12_64, cospi_20_64);
+          const __m256i k32_m20_p12 = pair256_set_epi32(-cospi_20_64, cospi_12_64);
+          const __m256i k32_m04_p28 = pair256_set_epi32(-cospi_4_64, cospi_28_64);
+
+          u[0] = _mm256_unpacklo_epi32(lstep2[ 8], lstep2[14]);
+          u[1] = _mm256_unpackhi_epi32(lstep2[ 8], lstep2[14]);
+          u[2] = _mm256_unpacklo_epi32(lstep2[ 9], lstep2[15]);
+          u[3] = _mm256_unpackhi_epi32(lstep2[ 9], lstep2[15]);
+          u[4] = _mm256_unpacklo_epi32(lstep2[10], lstep2[12]);
+          u[5] = _mm256_unpackhi_epi32(lstep2[10], lstep2[12]);
+          u[6] = _mm256_unpacklo_epi32(lstep2[11], lstep2[13]);
+          u[7] = _mm256_unpackhi_epi32(lstep2[11], lstep2[13]);
+          u[8] = _mm256_unpacklo_epi32(lstep2[10], lstep2[12]);
+          u[9] = _mm256_unpackhi_epi32(lstep2[10], lstep2[12]);
+          u[10] = _mm256_unpacklo_epi32(lstep2[11], lstep2[13]);
+          u[11] = _mm256_unpackhi_epi32(lstep2[11], lstep2[13]);
+          u[12] = _mm256_unpacklo_epi32(lstep2[ 8], lstep2[14]);
+          u[13] = _mm256_unpackhi_epi32(lstep2[ 8], lstep2[14]);
+          u[14] = _mm256_unpacklo_epi32(lstep2[ 9], lstep2[15]);
+          u[15] = _mm256_unpackhi_epi32(lstep2[ 9], lstep2[15]);
+
+          v[0] = k_madd_epi32_avx2(u[0], k32_p28_p04);
+          v[1] = k_madd_epi32_avx2(u[1], k32_p28_p04);
+          v[2] = k_madd_epi32_avx2(u[2], k32_p28_p04);
+          v[3] = k_madd_epi32_avx2(u[3], k32_p28_p04);
+          v[4] = k_madd_epi32_avx2(u[4], k32_p12_p20);
+          v[5] = k_madd_epi32_avx2(u[5], k32_p12_p20);
+          v[6] = k_madd_epi32_avx2(u[6], k32_p12_p20);
+          v[7] = k_madd_epi32_avx2(u[7], k32_p12_p20);
+          v[ 8] = k_madd_epi32_avx2(u[ 8], k32_m20_p12);
+          v[ 9] = k_madd_epi32_avx2(u[ 9], k32_m20_p12);
+          v[10] = k_madd_epi32_avx2(u[10], k32_m20_p12);
+          v[11] = k_madd_epi32_avx2(u[11], k32_m20_p12);
+          v[12] = k_madd_epi32_avx2(u[12], k32_m04_p28);
+          v[13] = k_madd_epi32_avx2(u[13], k32_m04_p28);
+          v[14] = k_madd_epi32_avx2(u[14], k32_m04_p28);
+          v[15] = k_madd_epi32_avx2(u[15], k32_m04_p28);
+
+          u[0] = k_packs_epi64_avx2(v[0], v[1]);
+          u[1] = k_packs_epi64_avx2(v[2], v[3]);
+          u[2] = k_packs_epi64_avx2(v[4], v[5]);
+          u[3] = k_packs_epi64_avx2(v[6], v[7]);
+          u[4] = k_packs_epi64_avx2(v[8], v[9]);
+          u[5] = k_packs_epi64_avx2(v[10], v[11]);
+          u[6] = k_packs_epi64_avx2(v[12], v[13]);
+          u[7] = k_packs_epi64_avx2(v[14], v[15]);
+
+          v[0] = _mm256_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+          v[1] = _mm256_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+          v[2] = _mm256_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+          v[3] = _mm256_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+          v[4] = _mm256_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+          v[5] = _mm256_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+          v[6] = _mm256_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+          v[7] = _mm256_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+
+          u[0] = _mm256_srai_epi32(v[0], DCT_CONST_BITS);
+          u[1] = _mm256_srai_epi32(v[1], DCT_CONST_BITS);
+          u[2] = _mm256_srai_epi32(v[2], DCT_CONST_BITS);
+          u[3] = _mm256_srai_epi32(v[3], DCT_CONST_BITS);
+          u[4] = _mm256_srai_epi32(v[4], DCT_CONST_BITS);
+          u[5] = _mm256_srai_epi32(v[5], DCT_CONST_BITS);
+          u[6] = _mm256_srai_epi32(v[6], DCT_CONST_BITS);
+          u[7] = _mm256_srai_epi32(v[7], DCT_CONST_BITS);
+
+          sign[0] = _mm256_cmpgt_epi32(kZero,u[0]);
+          sign[1] = _mm256_cmpgt_epi32(kZero,u[1]);
+          sign[2] = _mm256_cmpgt_epi32(kZero,u[2]);
+          sign[3] = _mm256_cmpgt_epi32(kZero,u[3]);
+          sign[4] = _mm256_cmpgt_epi32(kZero,u[4]);
+          sign[5] = _mm256_cmpgt_epi32(kZero,u[5]);
+          sign[6] = _mm256_cmpgt_epi32(kZero,u[6]);
+          sign[7] = _mm256_cmpgt_epi32(kZero,u[7]);
+
+          u[0] = _mm256_sub_epi32(u[0], sign[0]);
+          u[1] = _mm256_sub_epi32(u[1], sign[1]);
+          u[2] = _mm256_sub_epi32(u[2], sign[2]);
+          u[3] = _mm256_sub_epi32(u[3], sign[3]);
+          u[4] = _mm256_sub_epi32(u[4], sign[4]);
+          u[5] = _mm256_sub_epi32(u[5], sign[5]);
+          u[6] = _mm256_sub_epi32(u[6], sign[6]);
+          u[7] = _mm256_sub_epi32(u[7], sign[7]);
+
+          u[0] = _mm256_add_epi32(u[0], K32One);
+          u[1] = _mm256_add_epi32(u[1], K32One);
+          u[2] = _mm256_add_epi32(u[2], K32One);
+          u[3] = _mm256_add_epi32(u[3], K32One);
+          u[4] = _mm256_add_epi32(u[4], K32One);
+          u[5] = _mm256_add_epi32(u[5], K32One);
+          u[6] = _mm256_add_epi32(u[6], K32One);
+          u[7] = _mm256_add_epi32(u[7], K32One);
+
+          u[0] = _mm256_srai_epi32(u[0], 2);
+          u[1] = _mm256_srai_epi32(u[1], 2);
+          u[2] = _mm256_srai_epi32(u[2], 2);
+          u[3] = _mm256_srai_epi32(u[3], 2);
+          u[4] = _mm256_srai_epi32(u[4], 2);
+          u[5] = _mm256_srai_epi32(u[5], 2);
+          u[6] = _mm256_srai_epi32(u[6], 2);
+          u[7] = _mm256_srai_epi32(u[7], 2);
+
+          out[ 4] = _mm256_packs_epi32(u[0], u[1]);
+          out[20] = _mm256_packs_epi32(u[2], u[3]);
+          out[12] = _mm256_packs_epi32(u[4], u[5]);
+          out[28] = _mm256_packs_epi32(u[6], u[7]);
+        }
+        {
+          lstep3[16] = _mm256_add_epi32(lstep2[18], lstep1[16]);
+          lstep3[17] = _mm256_add_epi32(lstep2[19], lstep1[17]);
+          lstep3[18] = _mm256_sub_epi32(lstep1[16], lstep2[18]);
+          lstep3[19] = _mm256_sub_epi32(lstep1[17], lstep2[19]);
+          lstep3[20] = _mm256_sub_epi32(lstep1[22], lstep2[20]);
+          lstep3[21] = _mm256_sub_epi32(lstep1[23], lstep2[21]);
+          lstep3[22] = _mm256_add_epi32(lstep2[20], lstep1[22]);
+          lstep3[23] = _mm256_add_epi32(lstep2[21], lstep1[23]);
+          lstep3[24] = _mm256_add_epi32(lstep2[26], lstep1[24]);
+          lstep3[25] = _mm256_add_epi32(lstep2[27], lstep1[25]);
+          lstep3[26] = _mm256_sub_epi32(lstep1[24], lstep2[26]);
+          lstep3[27] = _mm256_sub_epi32(lstep1[25], lstep2[27]);
+          lstep3[28] = _mm256_sub_epi32(lstep1[30], lstep2[28]);
+          lstep3[29] = _mm256_sub_epi32(lstep1[31], lstep2[29]);
+          lstep3[30] = _mm256_add_epi32(lstep2[28], lstep1[30]);
+          lstep3[31] = _mm256_add_epi32(lstep2[29], lstep1[31]);
+        }
+        {
+          const __m256i k32_m04_p28 = pair256_set_epi32(-cospi_4_64, cospi_28_64);
+          const __m256i k32_m28_m04 = pair256_set_epi32(-cospi_28_64, -cospi_4_64);
+          const __m256i k32_m20_p12 = pair256_set_epi32(-cospi_20_64, cospi_12_64);
+          const __m256i k32_m12_m20 = pair256_set_epi32(-cospi_12_64,
+                                                     -cospi_20_64);
+          const __m256i k32_p12_p20 = pair256_set_epi32(cospi_12_64, cospi_20_64);
+          const __m256i k32_p28_p04 = pair256_set_epi32(cospi_28_64, cospi_4_64);
+
+          u[ 0] = _mm256_unpacklo_epi32(lstep2[34], lstep2[60]);
+          u[ 1] = _mm256_unpackhi_epi32(lstep2[34], lstep2[60]);
+          u[ 2] = _mm256_unpacklo_epi32(lstep2[35], lstep2[61]);
+          u[ 3] = _mm256_unpackhi_epi32(lstep2[35], lstep2[61]);
+          u[ 4] = _mm256_unpacklo_epi32(lstep2[36], lstep2[58]);
+          u[ 5] = _mm256_unpackhi_epi32(lstep2[36], lstep2[58]);
+          u[ 6] = _mm256_unpacklo_epi32(lstep2[37], lstep2[59]);
+          u[ 7] = _mm256_unpackhi_epi32(lstep2[37], lstep2[59]);
+          u[ 8] = _mm256_unpacklo_epi32(lstep2[42], lstep2[52]);
+          u[ 9] = _mm256_unpackhi_epi32(lstep2[42], lstep2[52]);
+          u[10] = _mm256_unpacklo_epi32(lstep2[43], lstep2[53]);
+          u[11] = _mm256_unpackhi_epi32(lstep2[43], lstep2[53]);
+          u[12] = _mm256_unpacklo_epi32(lstep2[44], lstep2[50]);
+          u[13] = _mm256_unpackhi_epi32(lstep2[44], lstep2[50]);
+          u[14] = _mm256_unpacklo_epi32(lstep2[45], lstep2[51]);
+          u[15] = _mm256_unpackhi_epi32(lstep2[45], lstep2[51]);
+
+          v[ 0] = k_madd_epi32_avx2(u[ 0], k32_m04_p28);
+          v[ 1] = k_madd_epi32_avx2(u[ 1], k32_m04_p28);
+          v[ 2] = k_madd_epi32_avx2(u[ 2], k32_m04_p28);
+          v[ 3] = k_madd_epi32_avx2(u[ 3], k32_m04_p28);
+          v[ 4] = k_madd_epi32_avx2(u[ 4], k32_m28_m04);
+          v[ 5] = k_madd_epi32_avx2(u[ 5], k32_m28_m04);
+          v[ 6] = k_madd_epi32_avx2(u[ 6], k32_m28_m04);
+          v[ 7] = k_madd_epi32_avx2(u[ 7], k32_m28_m04);
+          v[ 8] = k_madd_epi32_avx2(u[ 8], k32_m20_p12);
+          v[ 9] = k_madd_epi32_avx2(u[ 9], k32_m20_p12);
+          v[10] = k_madd_epi32_avx2(u[10], k32_m20_p12);
+          v[11] = k_madd_epi32_avx2(u[11], k32_m20_p12);
+          v[12] = k_madd_epi32_avx2(u[12], k32_m12_m20);
+          v[13] = k_madd_epi32_avx2(u[13], k32_m12_m20);
+          v[14] = k_madd_epi32_avx2(u[14], k32_m12_m20);
+          v[15] = k_madd_epi32_avx2(u[15], k32_m12_m20);
+          v[16] = k_madd_epi32_avx2(u[12], k32_m20_p12);
+          v[17] = k_madd_epi32_avx2(u[13], k32_m20_p12);
+          v[18] = k_madd_epi32_avx2(u[14], k32_m20_p12);
+          v[19] = k_madd_epi32_avx2(u[15], k32_m20_p12);
+          v[20] = k_madd_epi32_avx2(u[ 8], k32_p12_p20);
+          v[21] = k_madd_epi32_avx2(u[ 9], k32_p12_p20);
+          v[22] = k_madd_epi32_avx2(u[10], k32_p12_p20);
+          v[23] = k_madd_epi32_avx2(u[11], k32_p12_p20);
+          v[24] = k_madd_epi32_avx2(u[ 4], k32_m04_p28);
+          v[25] = k_madd_epi32_avx2(u[ 5], k32_m04_p28);
+          v[26] = k_madd_epi32_avx2(u[ 6], k32_m04_p28);
+          v[27] = k_madd_epi32_avx2(u[ 7], k32_m04_p28);
+          v[28] = k_madd_epi32_avx2(u[ 0], k32_p28_p04);
+          v[29] = k_madd_epi32_avx2(u[ 1], k32_p28_p04);
+          v[30] = k_madd_epi32_avx2(u[ 2], k32_p28_p04);
+          v[31] = k_madd_epi32_avx2(u[ 3], k32_p28_p04);
+
+          u[ 0] = k_packs_epi64_avx2(v[ 0], v[ 1]);
+          u[ 1] = k_packs_epi64_avx2(v[ 2], v[ 3]);
+          u[ 2] = k_packs_epi64_avx2(v[ 4], v[ 5]);
+          u[ 3] = k_packs_epi64_avx2(v[ 6], v[ 7]);
+          u[ 4] = k_packs_epi64_avx2(v[ 8], v[ 9]);
+          u[ 5] = k_packs_epi64_avx2(v[10], v[11]);
+          u[ 6] = k_packs_epi64_avx2(v[12], v[13]);
+          u[ 7] = k_packs_epi64_avx2(v[14], v[15]);
+          u[ 8] = k_packs_epi64_avx2(v[16], v[17]);
+          u[ 9] = k_packs_epi64_avx2(v[18], v[19]);
+          u[10] = k_packs_epi64_avx2(v[20], v[21]);
+          u[11] = k_packs_epi64_avx2(v[22], v[23]);
+          u[12] = k_packs_epi64_avx2(v[24], v[25]);
+          u[13] = k_packs_epi64_avx2(v[26], v[27]);
+          u[14] = k_packs_epi64_avx2(v[28], v[29]);
+          u[15] = k_packs_epi64_avx2(v[30], v[31]);
+
+          v[ 0] = _mm256_add_epi32(u[ 0], k__DCT_CONST_ROUNDING);
+          v[ 1] = _mm256_add_epi32(u[ 1], k__DCT_CONST_ROUNDING);
+          v[ 2] = _mm256_add_epi32(u[ 2], k__DCT_CONST_ROUNDING);
+          v[ 3] = _mm256_add_epi32(u[ 3], k__DCT_CONST_ROUNDING);
+          v[ 4] = _mm256_add_epi32(u[ 4], k__DCT_CONST_ROUNDING);
+          v[ 5] = _mm256_add_epi32(u[ 5], k__DCT_CONST_ROUNDING);
+          v[ 6] = _mm256_add_epi32(u[ 6], k__DCT_CONST_ROUNDING);
+          v[ 7] = _mm256_add_epi32(u[ 7], k__DCT_CONST_ROUNDING);
+          v[ 8] = _mm256_add_epi32(u[ 8], k__DCT_CONST_ROUNDING);
+          v[ 9] = _mm256_add_epi32(u[ 9], k__DCT_CONST_ROUNDING);
+          v[10] = _mm256_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+          v[11] = _mm256_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+          v[12] = _mm256_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+          v[13] = _mm256_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+          v[14] = _mm256_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+          v[15] = _mm256_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+          lstep3[34] = _mm256_srai_epi32(v[ 0], DCT_CONST_BITS);
+          lstep3[35] = _mm256_srai_epi32(v[ 1], DCT_CONST_BITS);
+          lstep3[36] = _mm256_srai_epi32(v[ 2], DCT_CONST_BITS);
+          lstep3[37] = _mm256_srai_epi32(v[ 3], DCT_CONST_BITS);
+          lstep3[42] = _mm256_srai_epi32(v[ 4], DCT_CONST_BITS);
+          lstep3[43] = _mm256_srai_epi32(v[ 5], DCT_CONST_BITS);
+          lstep3[44] = _mm256_srai_epi32(v[ 6], DCT_CONST_BITS);
+          lstep3[45] = _mm256_srai_epi32(v[ 7], DCT_CONST_BITS);
+          lstep3[50] = _mm256_srai_epi32(v[ 8], DCT_CONST_BITS);
+          lstep3[51] = _mm256_srai_epi32(v[ 9], DCT_CONST_BITS);
+          lstep3[52] = _mm256_srai_epi32(v[10], DCT_CONST_BITS);
+          lstep3[53] = _mm256_srai_epi32(v[11], DCT_CONST_BITS);
+          lstep3[58] = _mm256_srai_epi32(v[12], DCT_CONST_BITS);
+          lstep3[59] = _mm256_srai_epi32(v[13], DCT_CONST_BITS);
+          lstep3[60] = _mm256_srai_epi32(v[14], DCT_CONST_BITS);
+          lstep3[61] = _mm256_srai_epi32(v[15], DCT_CONST_BITS);
+        }
+        // stage 7
+        {
+          const __m256i k32_p30_p02 = pair256_set_epi32(cospi_30_64, cospi_2_64);
+          const __m256i k32_p14_p18 = pair256_set_epi32(cospi_14_64, cospi_18_64);
+          const __m256i k32_p22_p10 = pair256_set_epi32(cospi_22_64, cospi_10_64);
+          const __m256i k32_p06_p26 = pair256_set_epi32(cospi_6_64,  cospi_26_64);
+          const __m256i k32_m26_p06 = pair256_set_epi32(-cospi_26_64, cospi_6_64);
+          const __m256i k32_m10_p22 = pair256_set_epi32(-cospi_10_64, cospi_22_64);
+          const __m256i k32_m18_p14 = pair256_set_epi32(-cospi_18_64, cospi_14_64);
+          const __m256i k32_m02_p30 = pair256_set_epi32(-cospi_2_64, cospi_30_64);
+
+          u[ 0] = _mm256_unpacklo_epi32(lstep3[16], lstep3[30]);
+          u[ 1] = _mm256_unpackhi_epi32(lstep3[16], lstep3[30]);
+          u[ 2] = _mm256_unpacklo_epi32(lstep3[17], lstep3[31]);
+          u[ 3] = _mm256_unpackhi_epi32(lstep3[17], lstep3[31]);
+          u[ 4] = _mm256_unpacklo_epi32(lstep3[18], lstep3[28]);
+          u[ 5] = _mm256_unpackhi_epi32(lstep3[18], lstep3[28]);
+          u[ 6] = _mm256_unpacklo_epi32(lstep3[19], lstep3[29]);
+          u[ 7] = _mm256_unpackhi_epi32(lstep3[19], lstep3[29]);
+          u[ 8] = _mm256_unpacklo_epi32(lstep3[20], lstep3[26]);
+          u[ 9] = _mm256_unpackhi_epi32(lstep3[20], lstep3[26]);
+          u[10] = _mm256_unpacklo_epi32(lstep3[21], lstep3[27]);
+          u[11] = _mm256_unpackhi_epi32(lstep3[21], lstep3[27]);
+          u[12] = _mm256_unpacklo_epi32(lstep3[22], lstep3[24]);
+          u[13] = _mm256_unpackhi_epi32(lstep3[22], lstep3[24]);
+          u[14] = _mm256_unpacklo_epi32(lstep3[23], lstep3[25]);
+          u[15] = _mm256_unpackhi_epi32(lstep3[23], lstep3[25]);
+
+          v[ 0] = k_madd_epi32_avx2(u[ 0], k32_p30_p02);
+          v[ 1] = k_madd_epi32_avx2(u[ 1], k32_p30_p02);
+          v[ 2] = k_madd_epi32_avx2(u[ 2], k32_p30_p02);
+          v[ 3] = k_madd_epi32_avx2(u[ 3], k32_p30_p02);
+          v[ 4] = k_madd_epi32_avx2(u[ 4], k32_p14_p18);
+          v[ 5] = k_madd_epi32_avx2(u[ 5], k32_p14_p18);
+          v[ 6] = k_madd_epi32_avx2(u[ 6], k32_p14_p18);
+          v[ 7] = k_madd_epi32_avx2(u[ 7], k32_p14_p18);
+          v[ 8] = k_madd_epi32_avx2(u[ 8], k32_p22_p10);
+          v[ 9] = k_madd_epi32_avx2(u[ 9], k32_p22_p10);
+          v[10] = k_madd_epi32_avx2(u[10], k32_p22_p10);
+          v[11] = k_madd_epi32_avx2(u[11], k32_p22_p10);
+          v[12] = k_madd_epi32_avx2(u[12], k32_p06_p26);
+          v[13] = k_madd_epi32_avx2(u[13], k32_p06_p26);
+          v[14] = k_madd_epi32_avx2(u[14], k32_p06_p26);
+          v[15] = k_madd_epi32_avx2(u[15], k32_p06_p26);
+          v[16] = k_madd_epi32_avx2(u[12], k32_m26_p06);
+          v[17] = k_madd_epi32_avx2(u[13], k32_m26_p06);
+          v[18] = k_madd_epi32_avx2(u[14], k32_m26_p06);
+          v[19] = k_madd_epi32_avx2(u[15], k32_m26_p06);
+          v[20] = k_madd_epi32_avx2(u[ 8], k32_m10_p22);
+          v[21] = k_madd_epi32_avx2(u[ 9], k32_m10_p22);
+          v[22] = k_madd_epi32_avx2(u[10], k32_m10_p22);
+          v[23] = k_madd_epi32_avx2(u[11], k32_m10_p22);
+          v[24] = k_madd_epi32_avx2(u[ 4], k32_m18_p14);
+          v[25] = k_madd_epi32_avx2(u[ 5], k32_m18_p14);
+          v[26] = k_madd_epi32_avx2(u[ 6], k32_m18_p14);
+          v[27] = k_madd_epi32_avx2(u[ 7], k32_m18_p14);
+          v[28] = k_madd_epi32_avx2(u[ 0], k32_m02_p30);
+          v[29] = k_madd_epi32_avx2(u[ 1], k32_m02_p30);
+          v[30] = k_madd_epi32_avx2(u[ 2], k32_m02_p30);
+          v[31] = k_madd_epi32_avx2(u[ 3], k32_m02_p30);
+
+          u[ 0] = k_packs_epi64_avx2(v[ 0], v[ 1]);
+          u[ 1] = k_packs_epi64_avx2(v[ 2], v[ 3]);
+          u[ 2] = k_packs_epi64_avx2(v[ 4], v[ 5]);
+          u[ 3] = k_packs_epi64_avx2(v[ 6], v[ 7]);
+          u[ 4] = k_packs_epi64_avx2(v[ 8], v[ 9]);
+          u[ 5] = k_packs_epi64_avx2(v[10], v[11]);
+          u[ 6] = k_packs_epi64_avx2(v[12], v[13]);
+          u[ 7] = k_packs_epi64_avx2(v[14], v[15]);
+          u[ 8] = k_packs_epi64_avx2(v[16], v[17]);
+          u[ 9] = k_packs_epi64_avx2(v[18], v[19]);
+          u[10] = k_packs_epi64_avx2(v[20], v[21]);
+          u[11] = k_packs_epi64_avx2(v[22], v[23]);
+          u[12] = k_packs_epi64_avx2(v[24], v[25]);
+          u[13] = k_packs_epi64_avx2(v[26], v[27]);
+          u[14] = k_packs_epi64_avx2(v[28], v[29]);
+          u[15] = k_packs_epi64_avx2(v[30], v[31]);
+
+          v[ 0] = _mm256_add_epi32(u[ 0], k__DCT_CONST_ROUNDING);
+          v[ 1] = _mm256_add_epi32(u[ 1], k__DCT_CONST_ROUNDING);
+          v[ 2] = _mm256_add_epi32(u[ 2], k__DCT_CONST_ROUNDING);
+          v[ 3] = _mm256_add_epi32(u[ 3], k__DCT_CONST_ROUNDING);
+          v[ 4] = _mm256_add_epi32(u[ 4], k__DCT_CONST_ROUNDING);
+          v[ 5] = _mm256_add_epi32(u[ 5], k__DCT_CONST_ROUNDING);
+          v[ 6] = _mm256_add_epi32(u[ 6], k__DCT_CONST_ROUNDING);
+          v[ 7] = _mm256_add_epi32(u[ 7], k__DCT_CONST_ROUNDING);
+          v[ 8] = _mm256_add_epi32(u[ 8], k__DCT_CONST_ROUNDING);
+          v[ 9] = _mm256_add_epi32(u[ 9], k__DCT_CONST_ROUNDING);
+          v[10] = _mm256_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+          v[11] = _mm256_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+          v[12] = _mm256_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+          v[13] = _mm256_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+          v[14] = _mm256_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+          v[15] = _mm256_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+          u[ 0] = _mm256_srai_epi32(v[ 0], DCT_CONST_BITS);
+          u[ 1] = _mm256_srai_epi32(v[ 1], DCT_CONST_BITS);
+          u[ 2] = _mm256_srai_epi32(v[ 2], DCT_CONST_BITS);
+          u[ 3] = _mm256_srai_epi32(v[ 3], DCT_CONST_BITS);
+          u[ 4] = _mm256_srai_epi32(v[ 4], DCT_CONST_BITS);
+          u[ 5] = _mm256_srai_epi32(v[ 5], DCT_CONST_BITS);
+          u[ 6] = _mm256_srai_epi32(v[ 6], DCT_CONST_BITS);
+          u[ 7] = _mm256_srai_epi32(v[ 7], DCT_CONST_BITS);
+          u[ 8] = _mm256_srai_epi32(v[ 8], DCT_CONST_BITS);
+          u[ 9] = _mm256_srai_epi32(v[ 9], DCT_CONST_BITS);
+          u[10] = _mm256_srai_epi32(v[10], DCT_CONST_BITS);
+          u[11] = _mm256_srai_epi32(v[11], DCT_CONST_BITS);
+          u[12] = _mm256_srai_epi32(v[12], DCT_CONST_BITS);
+          u[13] = _mm256_srai_epi32(v[13], DCT_CONST_BITS);
+          u[14] = _mm256_srai_epi32(v[14], DCT_CONST_BITS);
+          u[15] = _mm256_srai_epi32(v[15], DCT_CONST_BITS);
+
+          v[ 0] = _mm256_cmpgt_epi32(kZero,u[ 0]);
+          v[ 1] = _mm256_cmpgt_epi32(kZero,u[ 1]);
+          v[ 2] = _mm256_cmpgt_epi32(kZero,u[ 2]);
+          v[ 3] = _mm256_cmpgt_epi32(kZero,u[ 3]);
+          v[ 4] = _mm256_cmpgt_epi32(kZero,u[ 4]);
+          v[ 5] = _mm256_cmpgt_epi32(kZero,u[ 5]);
+          v[ 6] = _mm256_cmpgt_epi32(kZero,u[ 6]);
+          v[ 7] = _mm256_cmpgt_epi32(kZero,u[ 7]);
+          v[ 8] = _mm256_cmpgt_epi32(kZero,u[ 8]);
+          v[ 9] = _mm256_cmpgt_epi32(kZero,u[ 9]);
+          v[10] = _mm256_cmpgt_epi32(kZero,u[10]);
+          v[11] = _mm256_cmpgt_epi32(kZero,u[11]);
+          v[12] = _mm256_cmpgt_epi32(kZero,u[12]);
+          v[13] = _mm256_cmpgt_epi32(kZero,u[13]);
+          v[14] = _mm256_cmpgt_epi32(kZero,u[14]);
+          v[15] = _mm256_cmpgt_epi32(kZero,u[15]);
+
+          u[ 0] = _mm256_sub_epi32(u[ 0], v[ 0]);
+          u[ 1] = _mm256_sub_epi32(u[ 1], v[ 1]);
+          u[ 2] = _mm256_sub_epi32(u[ 2], v[ 2]);
+          u[ 3] = _mm256_sub_epi32(u[ 3], v[ 3]);
+          u[ 4] = _mm256_sub_epi32(u[ 4], v[ 4]);
+          u[ 5] = _mm256_sub_epi32(u[ 5], v[ 5]);
+          u[ 6] = _mm256_sub_epi32(u[ 6], v[ 6]);
+          u[ 7] = _mm256_sub_epi32(u[ 7], v[ 7]);
+          u[ 8] = _mm256_sub_epi32(u[ 8], v[ 8]);
+          u[ 9] = _mm256_sub_epi32(u[ 9], v[ 9]);
+          u[10] = _mm256_sub_epi32(u[10], v[10]);
+          u[11] = _mm256_sub_epi32(u[11], v[11]);
+          u[12] = _mm256_sub_epi32(u[12], v[12]);
+          u[13] = _mm256_sub_epi32(u[13], v[13]);
+          u[14] = _mm256_sub_epi32(u[14], v[14]);
+          u[15] = _mm256_sub_epi32(u[15], v[15]);
+
+          v[ 0] = _mm256_add_epi32(u[ 0], K32One);
+          v[ 1] = _mm256_add_epi32(u[ 1], K32One);
+          v[ 2] = _mm256_add_epi32(u[ 2], K32One);
+          v[ 3] = _mm256_add_epi32(u[ 3], K32One);
+          v[ 4] = _mm256_add_epi32(u[ 4], K32One);
+          v[ 5] = _mm256_add_epi32(u[ 5], K32One);
+          v[ 6] = _mm256_add_epi32(u[ 6], K32One);
+          v[ 7] = _mm256_add_epi32(u[ 7], K32One);
+          v[ 8] = _mm256_add_epi32(u[ 8], K32One);
+          v[ 9] = _mm256_add_epi32(u[ 9], K32One);
+          v[10] = _mm256_add_epi32(u[10], K32One);
+          v[11] = _mm256_add_epi32(u[11], K32One);
+          v[12] = _mm256_add_epi32(u[12], K32One);
+          v[13] = _mm256_add_epi32(u[13], K32One);
+          v[14] = _mm256_add_epi32(u[14], K32One);
+          v[15] = _mm256_add_epi32(u[15], K32One);
+
+          u[ 0] = _mm256_srai_epi32(v[ 0], 2);
+          u[ 1] = _mm256_srai_epi32(v[ 1], 2);
+          u[ 2] = _mm256_srai_epi32(v[ 2], 2);
+          u[ 3] = _mm256_srai_epi32(v[ 3], 2);
+          u[ 4] = _mm256_srai_epi32(v[ 4], 2);
+          u[ 5] = _mm256_srai_epi32(v[ 5], 2);
+          u[ 6] = _mm256_srai_epi32(v[ 6], 2);
+          u[ 7] = _mm256_srai_epi32(v[ 7], 2);
+          u[ 8] = _mm256_srai_epi32(v[ 8], 2);
+          u[ 9] = _mm256_srai_epi32(v[ 9], 2);
+          u[10] = _mm256_srai_epi32(v[10], 2);
+          u[11] = _mm256_srai_epi32(v[11], 2);
+          u[12] = _mm256_srai_epi32(v[12], 2);
+          u[13] = _mm256_srai_epi32(v[13], 2);
+          u[14] = _mm256_srai_epi32(v[14], 2);
+          u[15] = _mm256_srai_epi32(v[15], 2);
+
+          out[ 2] = _mm256_packs_epi32(u[0], u[1]);
+          out[18] = _mm256_packs_epi32(u[2], u[3]);
+          out[10] = _mm256_packs_epi32(u[4], u[5]);
+          out[26] = _mm256_packs_epi32(u[6], u[7]);
+          out[ 6] = _mm256_packs_epi32(u[8], u[9]);
+          out[22] = _mm256_packs_epi32(u[10], u[11]);
+          out[14] = _mm256_packs_epi32(u[12], u[13]);
+          out[30] = _mm256_packs_epi32(u[14], u[15]);
+        }
+        {
+          lstep1[32] = _mm256_add_epi32(lstep3[34], lstep2[32]);
+          lstep1[33] = _mm256_add_epi32(lstep3[35], lstep2[33]);
+          lstep1[34] = _mm256_sub_epi32(lstep2[32], lstep3[34]);
+          lstep1[35] = _mm256_sub_epi32(lstep2[33], lstep3[35]);
+          lstep1[36] = _mm256_sub_epi32(lstep2[38], lstep3[36]);
+          lstep1[37] = _mm256_sub_epi32(lstep2[39], lstep3[37]);
+          lstep1[38] = _mm256_add_epi32(lstep3[36], lstep2[38]);
+          lstep1[39] = _mm256_add_epi32(lstep3[37], lstep2[39]);
+          lstep1[40] = _mm256_add_epi32(lstep3[42], lstep2[40]);
+          lstep1[41] = _mm256_add_epi32(lstep3[43], lstep2[41]);
+          lstep1[42] = _mm256_sub_epi32(lstep2[40], lstep3[42]);
+          lstep1[43] = _mm256_sub_epi32(lstep2[41], lstep3[43]);
+          lstep1[44] = _mm256_sub_epi32(lstep2[46], lstep3[44]);
+          lstep1[45] = _mm256_sub_epi32(lstep2[47], lstep3[45]);
+          lstep1[46] = _mm256_add_epi32(lstep3[44], lstep2[46]);
+          lstep1[47] = _mm256_add_epi32(lstep3[45], lstep2[47]);
+          lstep1[48] = _mm256_add_epi32(lstep3[50], lstep2[48]);
+          lstep1[49] = _mm256_add_epi32(lstep3[51], lstep2[49]);
+          lstep1[50] = _mm256_sub_epi32(lstep2[48], lstep3[50]);
+          lstep1[51] = _mm256_sub_epi32(lstep2[49], lstep3[51]);
+          lstep1[52] = _mm256_sub_epi32(lstep2[54], lstep3[52]);
+          lstep1[53] = _mm256_sub_epi32(lstep2[55], lstep3[53]);
+          lstep1[54] = _mm256_add_epi32(lstep3[52], lstep2[54]);
+          lstep1[55] = _mm256_add_epi32(lstep3[53], lstep2[55]);
+          lstep1[56] = _mm256_add_epi32(lstep3[58], lstep2[56]);
+          lstep1[57] = _mm256_add_epi32(lstep3[59], lstep2[57]);
+          lstep1[58] = _mm256_sub_epi32(lstep2[56], lstep3[58]);
+          lstep1[59] = _mm256_sub_epi32(lstep2[57], lstep3[59]);
+          lstep1[60] = _mm256_sub_epi32(lstep2[62], lstep3[60]);
+          lstep1[61] = _mm256_sub_epi32(lstep2[63], lstep3[61]);
+          lstep1[62] = _mm256_add_epi32(lstep3[60], lstep2[62]);
+          lstep1[63] = _mm256_add_epi32(lstep3[61], lstep2[63]);
+        }
+        // stage 8
+        {
+          const __m256i k32_p31_p01 = pair256_set_epi32(cospi_31_64, cospi_1_64);
+          const __m256i k32_p15_p17 = pair256_set_epi32(cospi_15_64, cospi_17_64);
+          const __m256i k32_p23_p09 = pair256_set_epi32(cospi_23_64, cospi_9_64);
+          const __m256i k32_p07_p25 = pair256_set_epi32(cospi_7_64, cospi_25_64);
+          const __m256i k32_m25_p07 = pair256_set_epi32(-cospi_25_64, cospi_7_64);
+          const __m256i k32_m09_p23 = pair256_set_epi32(-cospi_9_64, cospi_23_64);
+          const __m256i k32_m17_p15 = pair256_set_epi32(-cospi_17_64, cospi_15_64);
+          const __m256i k32_m01_p31 = pair256_set_epi32(-cospi_1_64, cospi_31_64);
+
+          u[ 0] = _mm256_unpacklo_epi32(lstep1[32], lstep1[62]);
+          u[ 1] = _mm256_unpackhi_epi32(lstep1[32], lstep1[62]);
+          u[ 2] = _mm256_unpacklo_epi32(lstep1[33], lstep1[63]);
+          u[ 3] = _mm256_unpackhi_epi32(lstep1[33], lstep1[63]);
+          u[ 4] = _mm256_unpacklo_epi32(lstep1[34], lstep1[60]);
+          u[ 5] = _mm256_unpackhi_epi32(lstep1[34], lstep1[60]);
+          u[ 6] = _mm256_unpacklo_epi32(lstep1[35], lstep1[61]);
+          u[ 7] = _mm256_unpackhi_epi32(lstep1[35], lstep1[61]);
+          u[ 8] = _mm256_unpacklo_epi32(lstep1[36], lstep1[58]);
+          u[ 9] = _mm256_unpackhi_epi32(lstep1[36], lstep1[58]);
+          u[10] = _mm256_unpacklo_epi32(lstep1[37], lstep1[59]);
+          u[11] = _mm256_unpackhi_epi32(lstep1[37], lstep1[59]);
+          u[12] = _mm256_unpacklo_epi32(lstep1[38], lstep1[56]);
+          u[13] = _mm256_unpackhi_epi32(lstep1[38], lstep1[56]);
+          u[14] = _mm256_unpacklo_epi32(lstep1[39], lstep1[57]);
+          u[15] = _mm256_unpackhi_epi32(lstep1[39], lstep1[57]);
+
+          v[ 0] = k_madd_epi32_avx2(u[ 0], k32_p31_p01);
+          v[ 1] = k_madd_epi32_avx2(u[ 1], k32_p31_p01);
+          v[ 2] = k_madd_epi32_avx2(u[ 2], k32_p31_p01);
+          v[ 3] = k_madd_epi32_avx2(u[ 3], k32_p31_p01);
+          v[ 4] = k_madd_epi32_avx2(u[ 4], k32_p15_p17);
+          v[ 5] = k_madd_epi32_avx2(u[ 5], k32_p15_p17);
+          v[ 6] = k_madd_epi32_avx2(u[ 6], k32_p15_p17);
+          v[ 7] = k_madd_epi32_avx2(u[ 7], k32_p15_p17);
+          v[ 8] = k_madd_epi32_avx2(u[ 8], k32_p23_p09);
+          v[ 9] = k_madd_epi32_avx2(u[ 9], k32_p23_p09);
+          v[10] = k_madd_epi32_avx2(u[10], k32_p23_p09);
+          v[11] = k_madd_epi32_avx2(u[11], k32_p23_p09);
+          v[12] = k_madd_epi32_avx2(u[12], k32_p07_p25);
+          v[13] = k_madd_epi32_avx2(u[13], k32_p07_p25);
+          v[14] = k_madd_epi32_avx2(u[14], k32_p07_p25);
+          v[15] = k_madd_epi32_avx2(u[15], k32_p07_p25);
+          v[16] = k_madd_epi32_avx2(u[12], k32_m25_p07);
+          v[17] = k_madd_epi32_avx2(u[13], k32_m25_p07);
+          v[18] = k_madd_epi32_avx2(u[14], k32_m25_p07);
+          v[19] = k_madd_epi32_avx2(u[15], k32_m25_p07);
+          v[20] = k_madd_epi32_avx2(u[ 8], k32_m09_p23);
+          v[21] = k_madd_epi32_avx2(u[ 9], k32_m09_p23);
+          v[22] = k_madd_epi32_avx2(u[10], k32_m09_p23);
+          v[23] = k_madd_epi32_avx2(u[11], k32_m09_p23);
+          v[24] = k_madd_epi32_avx2(u[ 4], k32_m17_p15);
+          v[25] = k_madd_epi32_avx2(u[ 5], k32_m17_p15);
+          v[26] = k_madd_epi32_avx2(u[ 6], k32_m17_p15);
+          v[27] = k_madd_epi32_avx2(u[ 7], k32_m17_p15);
+          v[28] = k_madd_epi32_avx2(u[ 0], k32_m01_p31);
+          v[29] = k_madd_epi32_avx2(u[ 1], k32_m01_p31);
+          v[30] = k_madd_epi32_avx2(u[ 2], k32_m01_p31);
+          v[31] = k_madd_epi32_avx2(u[ 3], k32_m01_p31);
+
+          u[ 0] = k_packs_epi64_avx2(v[ 0], v[ 1]);
+          u[ 1] = k_packs_epi64_avx2(v[ 2], v[ 3]);
+          u[ 2] = k_packs_epi64_avx2(v[ 4], v[ 5]);
+          u[ 3] = k_packs_epi64_avx2(v[ 6], v[ 7]);
+          u[ 4] = k_packs_epi64_avx2(v[ 8], v[ 9]);
+          u[ 5] = k_packs_epi64_avx2(v[10], v[11]);
+          u[ 6] = k_packs_epi64_avx2(v[12], v[13]);
+          u[ 7] = k_packs_epi64_avx2(v[14], v[15]);
+          u[ 8] = k_packs_epi64_avx2(v[16], v[17]);
+          u[ 9] = k_packs_epi64_avx2(v[18], v[19]);
+          u[10] = k_packs_epi64_avx2(v[20], v[21]);
+          u[11] = k_packs_epi64_avx2(v[22], v[23]);
+          u[12] = k_packs_epi64_avx2(v[24], v[25]);
+          u[13] = k_packs_epi64_avx2(v[26], v[27]);
+          u[14] = k_packs_epi64_avx2(v[28], v[29]);
+          u[15] = k_packs_epi64_avx2(v[30], v[31]);
+
+          v[ 0] = _mm256_add_epi32(u[ 0], k__DCT_CONST_ROUNDING);
+          v[ 1] = _mm256_add_epi32(u[ 1], k__DCT_CONST_ROUNDING);
+          v[ 2] = _mm256_add_epi32(u[ 2], k__DCT_CONST_ROUNDING);
+          v[ 3] = _mm256_add_epi32(u[ 3], k__DCT_CONST_ROUNDING);
+          v[ 4] = _mm256_add_epi32(u[ 4], k__DCT_CONST_ROUNDING);
+          v[ 5] = _mm256_add_epi32(u[ 5], k__DCT_CONST_ROUNDING);
+          v[ 6] = _mm256_add_epi32(u[ 6], k__DCT_CONST_ROUNDING);
+          v[ 7] = _mm256_add_epi32(u[ 7], k__DCT_CONST_ROUNDING);
+          v[ 8] = _mm256_add_epi32(u[ 8], k__DCT_CONST_ROUNDING);
+          v[ 9] = _mm256_add_epi32(u[ 9], k__DCT_CONST_ROUNDING);
+          v[10] = _mm256_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+          v[11] = _mm256_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+          v[12] = _mm256_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+          v[13] = _mm256_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+          v[14] = _mm256_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+          v[15] = _mm256_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+          u[ 0] = _mm256_srai_epi32(v[ 0], DCT_CONST_BITS);
+          u[ 1] = _mm256_srai_epi32(v[ 1], DCT_CONST_BITS);
+          u[ 2] = _mm256_srai_epi32(v[ 2], DCT_CONST_BITS);
+          u[ 3] = _mm256_srai_epi32(v[ 3], DCT_CONST_BITS);
+          u[ 4] = _mm256_srai_epi32(v[ 4], DCT_CONST_BITS);
+          u[ 5] = _mm256_srai_epi32(v[ 5], DCT_CONST_BITS);
+          u[ 6] = _mm256_srai_epi32(v[ 6], DCT_CONST_BITS);
+          u[ 7] = _mm256_srai_epi32(v[ 7], DCT_CONST_BITS);
+          u[ 8] = _mm256_srai_epi32(v[ 8], DCT_CONST_BITS);
+          u[ 9] = _mm256_srai_epi32(v[ 9], DCT_CONST_BITS);
+          u[10] = _mm256_srai_epi32(v[10], DCT_CONST_BITS);
+          u[11] = _mm256_srai_epi32(v[11], DCT_CONST_BITS);
+          u[12] = _mm256_srai_epi32(v[12], DCT_CONST_BITS);
+          u[13] = _mm256_srai_epi32(v[13], DCT_CONST_BITS);
+          u[14] = _mm256_srai_epi32(v[14], DCT_CONST_BITS);
+          u[15] = _mm256_srai_epi32(v[15], DCT_CONST_BITS);
+
+          v[ 0] = _mm256_cmpgt_epi32(kZero,u[ 0]);
+          v[ 1] = _mm256_cmpgt_epi32(kZero,u[ 1]);
+          v[ 2] = _mm256_cmpgt_epi32(kZero,u[ 2]);
+          v[ 3] = _mm256_cmpgt_epi32(kZero,u[ 3]);
+          v[ 4] = _mm256_cmpgt_epi32(kZero,u[ 4]);
+          v[ 5] = _mm256_cmpgt_epi32(kZero,u[ 5]);
+          v[ 6] = _mm256_cmpgt_epi32(kZero,u[ 6]);
+          v[ 7] = _mm256_cmpgt_epi32(kZero,u[ 7]);
+          v[ 8] = _mm256_cmpgt_epi32(kZero,u[ 8]);
+          v[ 9] = _mm256_cmpgt_epi32(kZero,u[ 9]);
+          v[10] = _mm256_cmpgt_epi32(kZero,u[10]);
+          v[11] = _mm256_cmpgt_epi32(kZero,u[11]);
+          v[12] = _mm256_cmpgt_epi32(kZero,u[12]);
+          v[13] = _mm256_cmpgt_epi32(kZero,u[13]);
+          v[14] = _mm256_cmpgt_epi32(kZero,u[14]);
+          v[15] = _mm256_cmpgt_epi32(kZero,u[15]);
+
+          u[ 0] = _mm256_sub_epi32(u[ 0], v[ 0]);
+          u[ 1] = _mm256_sub_epi32(u[ 1], v[ 1]);
+          u[ 2] = _mm256_sub_epi32(u[ 2], v[ 2]);
+          u[ 3] = _mm256_sub_epi32(u[ 3], v[ 3]);
+          u[ 4] = _mm256_sub_epi32(u[ 4], v[ 4]);
+          u[ 5] = _mm256_sub_epi32(u[ 5], v[ 5]);
+          u[ 6] = _mm256_sub_epi32(u[ 6], v[ 6]);
+          u[ 7] = _mm256_sub_epi32(u[ 7], v[ 7]);
+          u[ 8] = _mm256_sub_epi32(u[ 8], v[ 8]);
+          u[ 9] = _mm256_sub_epi32(u[ 9], v[ 9]);
+          u[10] = _mm256_sub_epi32(u[10], v[10]);
+          u[11] = _mm256_sub_epi32(u[11], v[11]);
+          u[12] = _mm256_sub_epi32(u[12], v[12]);
+          u[13] = _mm256_sub_epi32(u[13], v[13]);
+          u[14] = _mm256_sub_epi32(u[14], v[14]);
+          u[15] = _mm256_sub_epi32(u[15], v[15]);
+
+          v[0] = _mm256_add_epi32(u[0], K32One);
+          v[1] = _mm256_add_epi32(u[1], K32One);
+          v[2] = _mm256_add_epi32(u[2], K32One);
+          v[3] = _mm256_add_epi32(u[3], K32One);
+          v[4] = _mm256_add_epi32(u[4], K32One);
+          v[5] = _mm256_add_epi32(u[5], K32One);
+          v[6] = _mm256_add_epi32(u[6], K32One);
+          v[7] = _mm256_add_epi32(u[7], K32One);
+          v[8] = _mm256_add_epi32(u[8], K32One);
+          v[9] = _mm256_add_epi32(u[9], K32One);
+          v[10] = _mm256_add_epi32(u[10], K32One);
+          v[11] = _mm256_add_epi32(u[11], K32One);
+          v[12] = _mm256_add_epi32(u[12], K32One);
+          v[13] = _mm256_add_epi32(u[13], K32One);
+          v[14] = _mm256_add_epi32(u[14], K32One);
+          v[15] = _mm256_add_epi32(u[15], K32One);
+
+          u[0] = _mm256_srai_epi32(v[0], 2);
+          u[1] = _mm256_srai_epi32(v[1], 2);
+          u[2] = _mm256_srai_epi32(v[2], 2);
+          u[3] = _mm256_srai_epi32(v[3], 2);
+          u[4] = _mm256_srai_epi32(v[4], 2);
+          u[5] = _mm256_srai_epi32(v[5], 2);
+          u[6] = _mm256_srai_epi32(v[6], 2);
+          u[7] = _mm256_srai_epi32(v[7], 2);
+          u[8] = _mm256_srai_epi32(v[8], 2);
+          u[9] = _mm256_srai_epi32(v[9], 2);
+          u[10] = _mm256_srai_epi32(v[10], 2);
+          u[11] = _mm256_srai_epi32(v[11], 2);
+          u[12] = _mm256_srai_epi32(v[12], 2);
+          u[13] = _mm256_srai_epi32(v[13], 2);
+          u[14] = _mm256_srai_epi32(v[14], 2);
+          u[15] = _mm256_srai_epi32(v[15], 2);
+
+          out[ 1] = _mm256_packs_epi32(u[0], u[1]);
+          out[17] = _mm256_packs_epi32(u[2], u[3]);
+          out[ 9] = _mm256_packs_epi32(u[4], u[5]);
+          out[25] = _mm256_packs_epi32(u[6], u[7]);
+          out[ 7] = _mm256_packs_epi32(u[8], u[9]);
+          out[23] = _mm256_packs_epi32(u[10], u[11]);
+          out[15] = _mm256_packs_epi32(u[12], u[13]);
+          out[31] = _mm256_packs_epi32(u[14], u[15]);
+        }
+        {
+          const __m256i k32_p27_p05 = pair256_set_epi32(cospi_27_64, cospi_5_64);
+          const __m256i k32_p11_p21 = pair256_set_epi32(cospi_11_64, cospi_21_64);
+          const __m256i k32_p19_p13 = pair256_set_epi32(cospi_19_64, cospi_13_64);
+          const __m256i k32_p03_p29 = pair256_set_epi32(cospi_3_64, cospi_29_64);
+          const __m256i k32_m29_p03 = pair256_set_epi32(-cospi_29_64, cospi_3_64);
+          const __m256i k32_m13_p19 = pair256_set_epi32(-cospi_13_64, cospi_19_64);
+          const __m256i k32_m21_p11 = pair256_set_epi32(-cospi_21_64, cospi_11_64);
+          const __m256i k32_m05_p27 = pair256_set_epi32(-cospi_5_64, cospi_27_64);
+
+          u[ 0] = _mm256_unpacklo_epi32(lstep1[40], lstep1[54]);
+          u[ 1] = _mm256_unpackhi_epi32(lstep1[40], lstep1[54]);
+          u[ 2] = _mm256_unpacklo_epi32(lstep1[41], lstep1[55]);
+          u[ 3] = _mm256_unpackhi_epi32(lstep1[41], lstep1[55]);
+          u[ 4] = _mm256_unpacklo_epi32(lstep1[42], lstep1[52]);
+          u[ 5] = _mm256_unpackhi_epi32(lstep1[42], lstep1[52]);
+          u[ 6] = _mm256_unpacklo_epi32(lstep1[43], lstep1[53]);
+          u[ 7] = _mm256_unpackhi_epi32(lstep1[43], lstep1[53]);
+          u[ 8] = _mm256_unpacklo_epi32(lstep1[44], lstep1[50]);
+          u[ 9] = _mm256_unpackhi_epi32(lstep1[44], lstep1[50]);
+          u[10] = _mm256_unpacklo_epi32(lstep1[45], lstep1[51]);
+          u[11] = _mm256_unpackhi_epi32(lstep1[45], lstep1[51]);
+          u[12] = _mm256_unpacklo_epi32(lstep1[46], lstep1[48]);
+          u[13] = _mm256_unpackhi_epi32(lstep1[46], lstep1[48]);
+          u[14] = _mm256_unpacklo_epi32(lstep1[47], lstep1[49]);
+          u[15] = _mm256_unpackhi_epi32(lstep1[47], lstep1[49]);
+
+          v[ 0] = k_madd_epi32_avx2(u[ 0], k32_p27_p05);
+          v[ 1] = k_madd_epi32_avx2(u[ 1], k32_p27_p05);
+          v[ 2] = k_madd_epi32_avx2(u[ 2], k32_p27_p05);
+          v[ 3] = k_madd_epi32_avx2(u[ 3], k32_p27_p05);
+          v[ 4] = k_madd_epi32_avx2(u[ 4], k32_p11_p21);
+          v[ 5] = k_madd_epi32_avx2(u[ 5], k32_p11_p21);
+          v[ 6] = k_madd_epi32_avx2(u[ 6], k32_p11_p21);
+          v[ 7] = k_madd_epi32_avx2(u[ 7], k32_p11_p21);
+          v[ 8] = k_madd_epi32_avx2(u[ 8], k32_p19_p13);
+          v[ 9] = k_madd_epi32_avx2(u[ 9], k32_p19_p13);
+          v[10] = k_madd_epi32_avx2(u[10], k32_p19_p13);
+          v[11] = k_madd_epi32_avx2(u[11], k32_p19_p13);
+          v[12] = k_madd_epi32_avx2(u[12], k32_p03_p29);
+          v[13] = k_madd_epi32_avx2(u[13], k32_p03_p29);
+          v[14] = k_madd_epi32_avx2(u[14], k32_p03_p29);
+          v[15] = k_madd_epi32_avx2(u[15], k32_p03_p29);
+          v[16] = k_madd_epi32_avx2(u[12], k32_m29_p03);
+          v[17] = k_madd_epi32_avx2(u[13], k32_m29_p03);
+          v[18] = k_madd_epi32_avx2(u[14], k32_m29_p03);
+          v[19] = k_madd_epi32_avx2(u[15], k32_m29_p03);
+          v[20] = k_madd_epi32_avx2(u[ 8], k32_m13_p19);
+          v[21] = k_madd_epi32_avx2(u[ 9], k32_m13_p19);
+          v[22] = k_madd_epi32_avx2(u[10], k32_m13_p19);
+          v[23] = k_madd_epi32_avx2(u[11], k32_m13_p19);
+          v[24] = k_madd_epi32_avx2(u[ 4], k32_m21_p11);
+          v[25] = k_madd_epi32_avx2(u[ 5], k32_m21_p11);
+          v[26] = k_madd_epi32_avx2(u[ 6], k32_m21_p11);
+          v[27] = k_madd_epi32_avx2(u[ 7], k32_m21_p11);
+          v[28] = k_madd_epi32_avx2(u[ 0], k32_m05_p27);
+          v[29] = k_madd_epi32_avx2(u[ 1], k32_m05_p27);
+          v[30] = k_madd_epi32_avx2(u[ 2], k32_m05_p27);
+          v[31] = k_madd_epi32_avx2(u[ 3], k32_m05_p27);
+
+          u[ 0] = k_packs_epi64_avx2(v[ 0], v[ 1]);
+          u[ 1] = k_packs_epi64_avx2(v[ 2], v[ 3]);
+          u[ 2] = k_packs_epi64_avx2(v[ 4], v[ 5]);
+          u[ 3] = k_packs_epi64_avx2(v[ 6], v[ 7]);
+          u[ 4] = k_packs_epi64_avx2(v[ 8], v[ 9]);
+          u[ 5] = k_packs_epi64_avx2(v[10], v[11]);
+          u[ 6] = k_packs_epi64_avx2(v[12], v[13]);
+          u[ 7] = k_packs_epi64_avx2(v[14], v[15]);
+          u[ 8] = k_packs_epi64_avx2(v[16], v[17]);
+          u[ 9] = k_packs_epi64_avx2(v[18], v[19]);
+          u[10] = k_packs_epi64_avx2(v[20], v[21]);
+          u[11] = k_packs_epi64_avx2(v[22], v[23]);
+          u[12] = k_packs_epi64_avx2(v[24], v[25]);
+          u[13] = k_packs_epi64_avx2(v[26], v[27]);
+          u[14] = k_packs_epi64_avx2(v[28], v[29]);
+          u[15] = k_packs_epi64_avx2(v[30], v[31]);
+
+          v[ 0] = _mm256_add_epi32(u[ 0], k__DCT_CONST_ROUNDING);
+          v[ 1] = _mm256_add_epi32(u[ 1], k__DCT_CONST_ROUNDING);
+          v[ 2] = _mm256_add_epi32(u[ 2], k__DCT_CONST_ROUNDING);
+          v[ 3] = _mm256_add_epi32(u[ 3], k__DCT_CONST_ROUNDING);
+          v[ 4] = _mm256_add_epi32(u[ 4], k__DCT_CONST_ROUNDING);
+          v[ 5] = _mm256_add_epi32(u[ 5], k__DCT_CONST_ROUNDING);
+          v[ 6] = _mm256_add_epi32(u[ 6], k__DCT_CONST_ROUNDING);
+          v[ 7] = _mm256_add_epi32(u[ 7], k__DCT_CONST_ROUNDING);
+          v[ 8] = _mm256_add_epi32(u[ 8], k__DCT_CONST_ROUNDING);
+          v[ 9] = _mm256_add_epi32(u[ 9], k__DCT_CONST_ROUNDING);
+          v[10] = _mm256_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+          v[11] = _mm256_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+          v[12] = _mm256_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+          v[13] = _mm256_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+          v[14] = _mm256_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+          v[15] = _mm256_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+          u[ 0] = _mm256_srai_epi32(v[ 0], DCT_CONST_BITS);
+          u[ 1] = _mm256_srai_epi32(v[ 1], DCT_CONST_BITS);
+          u[ 2] = _mm256_srai_epi32(v[ 2], DCT_CONST_BITS);
+          u[ 3] = _mm256_srai_epi32(v[ 3], DCT_CONST_BITS);
+          u[ 4] = _mm256_srai_epi32(v[ 4], DCT_CONST_BITS);
+          u[ 5] = _mm256_srai_epi32(v[ 5], DCT_CONST_BITS);
+          u[ 6] = _mm256_srai_epi32(v[ 6], DCT_CONST_BITS);
+          u[ 7] = _mm256_srai_epi32(v[ 7], DCT_CONST_BITS);
+          u[ 8] = _mm256_srai_epi32(v[ 8], DCT_CONST_BITS);
+          u[ 9] = _mm256_srai_epi32(v[ 9], DCT_CONST_BITS);
+          u[10] = _mm256_srai_epi32(v[10], DCT_CONST_BITS);
+          u[11] = _mm256_srai_epi32(v[11], DCT_CONST_BITS);
+          u[12] = _mm256_srai_epi32(v[12], DCT_CONST_BITS);
+          u[13] = _mm256_srai_epi32(v[13], DCT_CONST_BITS);
+          u[14] = _mm256_srai_epi32(v[14], DCT_CONST_BITS);
+          u[15] = _mm256_srai_epi32(v[15], DCT_CONST_BITS);
+
+          v[ 0] = _mm256_cmpgt_epi32(kZero,u[ 0]);
+          v[ 1] = _mm256_cmpgt_epi32(kZero,u[ 1]);
+          v[ 2] = _mm256_cmpgt_epi32(kZero,u[ 2]);
+          v[ 3] = _mm256_cmpgt_epi32(kZero,u[ 3]);
+          v[ 4] = _mm256_cmpgt_epi32(kZero,u[ 4]);
+          v[ 5] = _mm256_cmpgt_epi32(kZero,u[ 5]);
+          v[ 6] = _mm256_cmpgt_epi32(kZero,u[ 6]);
+          v[ 7] = _mm256_cmpgt_epi32(kZero,u[ 7]);
+          v[ 8] = _mm256_cmpgt_epi32(kZero,u[ 8]);
+          v[ 9] = _mm256_cmpgt_epi32(kZero,u[ 9]);
+          v[10] = _mm256_cmpgt_epi32(kZero,u[10]);
+          v[11] = _mm256_cmpgt_epi32(kZero,u[11]);
+          v[12] = _mm256_cmpgt_epi32(kZero,u[12]);
+          v[13] = _mm256_cmpgt_epi32(kZero,u[13]);
+          v[14] = _mm256_cmpgt_epi32(kZero,u[14]);
+          v[15] = _mm256_cmpgt_epi32(kZero,u[15]);
+
+          u[ 0] = _mm256_sub_epi32(u[ 0], v[ 0]);
+          u[ 1] = _mm256_sub_epi32(u[ 1], v[ 1]);
+          u[ 2] = _mm256_sub_epi32(u[ 2], v[ 2]);
+          u[ 3] = _mm256_sub_epi32(u[ 3], v[ 3]);
+          u[ 4] = _mm256_sub_epi32(u[ 4], v[ 4]);
+          u[ 5] = _mm256_sub_epi32(u[ 5], v[ 5]);
+          u[ 6] = _mm256_sub_epi32(u[ 6], v[ 6]);
+          u[ 7] = _mm256_sub_epi32(u[ 7], v[ 7]);
+          u[ 8] = _mm256_sub_epi32(u[ 8], v[ 8]);
+          u[ 9] = _mm256_sub_epi32(u[ 9], v[ 9]);
+          u[10] = _mm256_sub_epi32(u[10], v[10]);
+          u[11] = _mm256_sub_epi32(u[11], v[11]);
+          u[12] = _mm256_sub_epi32(u[12], v[12]);
+          u[13] = _mm256_sub_epi32(u[13], v[13]);
+          u[14] = _mm256_sub_epi32(u[14], v[14]);
+          u[15] = _mm256_sub_epi32(u[15], v[15]);
+
+          v[0] = _mm256_add_epi32(u[0], K32One);
+          v[1] = _mm256_add_epi32(u[1], K32One);
+          v[2] = _mm256_add_epi32(u[2], K32One);
+          v[3] = _mm256_add_epi32(u[3], K32One);
+          v[4] = _mm256_add_epi32(u[4], K32One);
+          v[5] = _mm256_add_epi32(u[5], K32One);
+          v[6] = _mm256_add_epi32(u[6], K32One);
+          v[7] = _mm256_add_epi32(u[7], K32One);
+          v[8] = _mm256_add_epi32(u[8], K32One);
+          v[9] = _mm256_add_epi32(u[9], K32One);
+          v[10] = _mm256_add_epi32(u[10], K32One);
+          v[11] = _mm256_add_epi32(u[11], K32One);
+          v[12] = _mm256_add_epi32(u[12], K32One);
+          v[13] = _mm256_add_epi32(u[13], K32One);
+          v[14] = _mm256_add_epi32(u[14], K32One);
+          v[15] = _mm256_add_epi32(u[15], K32One);
+
+          u[0] = _mm256_srai_epi32(v[0], 2);
+          u[1] = _mm256_srai_epi32(v[1], 2);
+          u[2] = _mm256_srai_epi32(v[2], 2);
+          u[3] = _mm256_srai_epi32(v[3], 2);
+          u[4] = _mm256_srai_epi32(v[4], 2);
+          u[5] = _mm256_srai_epi32(v[5], 2);
+          u[6] = _mm256_srai_epi32(v[6], 2);
+          u[7] = _mm256_srai_epi32(v[7], 2);
+          u[8] = _mm256_srai_epi32(v[8], 2);
+          u[9] = _mm256_srai_epi32(v[9], 2);
+          u[10] = _mm256_srai_epi32(v[10], 2);
+          u[11] = _mm256_srai_epi32(v[11], 2);
+          u[12] = _mm256_srai_epi32(v[12], 2);
+          u[13] = _mm256_srai_epi32(v[13], 2);
+          u[14] = _mm256_srai_epi32(v[14], 2);
+          u[15] = _mm256_srai_epi32(v[15], 2);
+
+          out[ 5] = _mm256_packs_epi32(u[0], u[1]);
+          out[21] = _mm256_packs_epi32(u[2], u[3]);
+          out[13] = _mm256_packs_epi32(u[4], u[5]);
+          out[29] = _mm256_packs_epi32(u[6], u[7]);
+          out[ 3] = _mm256_packs_epi32(u[8], u[9]);
+          out[19] = _mm256_packs_epi32(u[10], u[11]);
+          out[11] = _mm256_packs_epi32(u[12], u[13]);
+          out[27] = _mm256_packs_epi32(u[14], u[15]);
+        }
+      }
+#endif
+      // Transpose the results, do it as four 8x8 transposes.
+      {
+        int transpose_block;
+        int16_t *output_currStep,*output_nextStep;
+        if (0 == pass){
+                 output_currStep = &intermediate[column_start * 32];
+                 output_nextStep = &intermediate[(column_start + 8) * 32];
+        } else{
+                 output_currStep = &output_org[column_start * 32];
+                 output_nextStep = &output_org[(column_start + 8) * 32];
+        }
+        for (transpose_block = 0; transpose_block < 4; ++transpose_block) {
+          __m256i *this_out = &out[8 * transpose_block];
+          // 00  01  02  03  04  05  06  07  08  09  10  11  12  13  14  15
+          // 20  21  22  23  24  25  26  27  28  29  30  31  32  33  34  35
+          // 40  41  42  43  44  45  46  47  48  49  50  51  52  53  54  55
+          // 60  61  62  63  64  65  66  67  68  69  70  71  72  73  74  75
+          // 80  81  82  83  84  85  86  87  88  89  90  91  92  93  94  95
+          // 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115
+          // 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135
+          // 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155
+          const __m256i tr0_0 = _mm256_unpacklo_epi16(this_out[0], this_out[1]);
+          const __m256i tr0_1 = _mm256_unpacklo_epi16(this_out[2], this_out[3]);
+          const __m256i tr0_2 = _mm256_unpackhi_epi16(this_out[0], this_out[1]);
+          const __m256i tr0_3 = _mm256_unpackhi_epi16(this_out[2], this_out[3]);
+          const __m256i tr0_4 = _mm256_unpacklo_epi16(this_out[4], this_out[5]);
+          const __m256i tr0_5 = _mm256_unpacklo_epi16(this_out[6], this_out[7]);
+          const __m256i tr0_6 = _mm256_unpackhi_epi16(this_out[4], this_out[5]);
+          const __m256i tr0_7 = _mm256_unpackhi_epi16(this_out[6], this_out[7]);
+          // 00  20  01  21  02  22  03  23  08  28  09  29  10  30  11  31
+          // 40  60  41  61  42  62  43  63  48  68  49  69  50  70  51  71
+          // 04  24  05  25  06  26  07  27  12  32  13  33  14  34  15  35
+          // 44  64  45  65  46  66  47  67  52  72  53  73  54  74  55  75
+          // 80  100 81  101 82  102 83  103 88  108 89  109 90  110 91  101
+          // 120 140 121 141 122 142 123 143 128 148 129 149 130 150 131 151
+          // 84  104 85  105 86  106 87  107 92  112 93  113 94  114 95  115
+          // 124 144 125 145 126 146 127 147 132 152 133 153 134 154 135 155
+
+          const __m256i tr1_0 = _mm256_unpacklo_epi32(tr0_0, tr0_1);
+          const __m256i tr1_1 = _mm256_unpacklo_epi32(tr0_2, tr0_3);
+          const __m256i tr1_2 = _mm256_unpackhi_epi32(tr0_0, tr0_1);
+          const __m256i tr1_3 = _mm256_unpackhi_epi32(tr0_2, tr0_3);
+          const __m256i tr1_4 = _mm256_unpacklo_epi32(tr0_4, tr0_5);
+          const __m256i tr1_5 = _mm256_unpacklo_epi32(tr0_6, tr0_7);
+          const __m256i tr1_6 = _mm256_unpackhi_epi32(tr0_4, tr0_5);
+          const __m256i tr1_7 = _mm256_unpackhi_epi32(tr0_6, tr0_7);
+          // 00 20  40  60  01 21  41  61  08 28  48  68  09 29  49  69
+          // 04 24  44  64  05 25  45  65  12 32  52  72  13 33  53  73
+          // 02 22  42  62  03 23  43  63  10 30  50  70  11 31  51  71
+          // 06 26  46  66  07 27  47  67  14 34  54  74  15 35  55  75
+          // 80 100 120 140 81 101 121 141 88 108 128 148 89 109 129 149
+          // 84 104 124 144 85 105 125 145 92 112 132 152 93 113 133 153
+          // 82 102 122 142 83 103 123 143 90 110 130 150 91 101 131 151
+          // 86 106 126 146 87 107 127 147 94 114 134 154 95 115 135 155
+          __m256i tr2_0 = _mm256_unpacklo_epi64(tr1_0, tr1_4);
+          __m256i tr2_1 = _mm256_unpackhi_epi64(tr1_0, tr1_4);
+          __m256i tr2_2 = _mm256_unpacklo_epi64(tr1_2, tr1_6);
+          __m256i tr2_3 = _mm256_unpackhi_epi64(tr1_2, tr1_6);
+          __m256i tr2_4 = _mm256_unpacklo_epi64(tr1_1, tr1_5);
+          __m256i tr2_5 = _mm256_unpackhi_epi64(tr1_1, tr1_5);
+          __m256i tr2_6 = _mm256_unpacklo_epi64(tr1_3, tr1_7);
+          __m256i tr2_7 = _mm256_unpackhi_epi64(tr1_3, tr1_7);
+          // 00 20 40 60 80 100 120 140 08 28 48 68 88 108 128 148
+          // 01 21 41 61 81 101 121 141 09 29 49 69 89 109 129 149
+          // 02 22 42 62 82 102 122 142 10 30 50 70 90 110 130 150
+          // 03 23 43 63 83 103 123 143 11 31 51 71 91 101 131 151
+          // 04 24 44 64 84 104 124 144 12 32 52 72 92 112 132 152
+          // 05 25 45 65 85 105 125 145 13 33 53 73 93 113 133 153
+          // 06 26 46 66 86 106 126 146 14 34 54 74 94 114 134 154
+          // 07 27 47 67 87 107 127 147 15 35 55 75 95 115 135 155
+          if (0 == pass) {
+            // output[j] = (output[j] + 1 + (output[j] > 0)) >> 2;
+            // TODO(cd): see quality impact of only doing
+            //           output[j] = (output[j] + 1) >> 2;
+            //           which would remove the code between here ...
+            __m256i tr2_0_0 = _mm256_cmpgt_epi16(tr2_0, kZero);
+            __m256i tr2_1_0 = _mm256_cmpgt_epi16(tr2_1, kZero);
+            __m256i tr2_2_0 = _mm256_cmpgt_epi16(tr2_2, kZero);
+            __m256i tr2_3_0 = _mm256_cmpgt_epi16(tr2_3, kZero);
+            __m256i tr2_4_0 = _mm256_cmpgt_epi16(tr2_4, kZero);
+            __m256i tr2_5_0 = _mm256_cmpgt_epi16(tr2_5, kZero);
+            __m256i tr2_6_0 = _mm256_cmpgt_epi16(tr2_6, kZero);
+            __m256i tr2_7_0 = _mm256_cmpgt_epi16(tr2_7, kZero);
+            tr2_0 = _mm256_sub_epi16(tr2_0, tr2_0_0);
+            tr2_1 = _mm256_sub_epi16(tr2_1, tr2_1_0);
+            tr2_2 = _mm256_sub_epi16(tr2_2, tr2_2_0);
+            tr2_3 = _mm256_sub_epi16(tr2_3, tr2_3_0);
+            tr2_4 = _mm256_sub_epi16(tr2_4, tr2_4_0);
+            tr2_5 = _mm256_sub_epi16(tr2_5, tr2_5_0);
+            tr2_6 = _mm256_sub_epi16(tr2_6, tr2_6_0);
+            tr2_7 = _mm256_sub_epi16(tr2_7, tr2_7_0);
+            //           ... and here.
+            //           PS: also change code in vp9/encoder/vp9_dct.c
+            tr2_0 = _mm256_add_epi16(tr2_0, kOne);
+            tr2_1 = _mm256_add_epi16(tr2_1, kOne);
+            tr2_2 = _mm256_add_epi16(tr2_2, kOne);
+            tr2_3 = _mm256_add_epi16(tr2_3, kOne);
+            tr2_4 = _mm256_add_epi16(tr2_4, kOne);
+            tr2_5 = _mm256_add_epi16(tr2_5, kOne);
+            tr2_6 = _mm256_add_epi16(tr2_6, kOne);
+            tr2_7 = _mm256_add_epi16(tr2_7, kOne);
+            tr2_0 = _mm256_srai_epi16(tr2_0, 2);
+            tr2_1 = _mm256_srai_epi16(tr2_1, 2);
+            tr2_2 = _mm256_srai_epi16(tr2_2, 2);
+            tr2_3 = _mm256_srai_epi16(tr2_3, 2);
+            tr2_4 = _mm256_srai_epi16(tr2_4, 2);
+            tr2_5 = _mm256_srai_epi16(tr2_5, 2);
+            tr2_6 = _mm256_srai_epi16(tr2_6, 2);
+            tr2_7 = _mm256_srai_epi16(tr2_7, 2);
+          }
+          // Note: even though all these stores are aligned, using the aligned
+          //       intrinsic make the code slightly slower.
+          _mm_storeu_si128((__m128i *)(output_currStep + 0 * 32), _mm256_castsi256_si128(tr2_0));
+          _mm_storeu_si128((__m128i *)(output_currStep + 1 * 32), _mm256_castsi256_si128(tr2_1));
+          _mm_storeu_si128((__m128i *)(output_currStep + 2 * 32), _mm256_castsi256_si128(tr2_2));
+          _mm_storeu_si128((__m128i *)(output_currStep + 3 * 32), _mm256_castsi256_si128(tr2_3));
+          _mm_storeu_si128((__m128i *)(output_currStep + 4 * 32), _mm256_castsi256_si128(tr2_4));
+          _mm_storeu_si128((__m128i *)(output_currStep + 5 * 32), _mm256_castsi256_si128(tr2_5));
+          _mm_storeu_si128((__m128i *)(output_currStep + 6 * 32), _mm256_castsi256_si128(tr2_6));
+          _mm_storeu_si128((__m128i *)(output_currStep + 7 * 32), _mm256_castsi256_si128(tr2_7));
+
+          _mm_storeu_si128((__m128i *)(output_nextStep + 0 * 32), _mm256_extractf128_si256(tr2_0,1));
+          _mm_storeu_si128((__m128i *)(output_nextStep + 1 * 32), _mm256_extractf128_si256(tr2_1,1));
+          _mm_storeu_si128((__m128i *)(output_nextStep + 2 * 32), _mm256_extractf128_si256(tr2_2,1));
+          _mm_storeu_si128((__m128i *)(output_nextStep + 3 * 32), _mm256_extractf128_si256(tr2_3,1));
+          _mm_storeu_si128((__m128i *)(output_nextStep + 4 * 32), _mm256_extractf128_si256(tr2_4,1));
+          _mm_storeu_si128((__m128i *)(output_nextStep + 5 * 32), _mm256_extractf128_si256(tr2_5,1));
+          _mm_storeu_si128((__m128i *)(output_nextStep + 6 * 32), _mm256_extractf128_si256(tr2_6,1));
+          _mm_storeu_si128((__m128i *)(output_nextStep + 7 * 32), _mm256_extractf128_si256(tr2_7,1));
+          // Process next 8x8
+          output_currStep += 8;
+          output_nextStep += 8;
+        }
+      }
+    }
+  }
+}  // NOLINT
diff --git a/media/libvpx/vp9/encoder/x86/vp9_dct32x32_sse2_impl.h b/media/libvpx/vp9/encoder/x86/vp9_dct32x32_sse2_impl.h
new file mode 100644
index 000000000..003ebd13f
--- /dev/null
+++ b/media/libvpx/vp9/encoder/x86/vp9_dct32x32_sse2_impl.h
@@ -0,0 +1,3151 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <emmintrin.h>  // SSE2
+
+#include "./vp9_rtcd.h"
+#include "vp9/common/vp9_idct.h"  // for cospi constants
+#include "vp9/encoder/x86/vp9_dct_sse2.h"
+#include "vp9/encoder/vp9_dct.h"
+#include "vpx_ports/mem.h"
+
+#if DCT_HIGH_BIT_DEPTH
+#define ADD_EPI16 _mm_adds_epi16
+#define SUB_EPI16 _mm_subs_epi16
+#if FDCT32x32_HIGH_PRECISION
+void vp9_fdct32x32_rows_c(const int16_t *intermediate, tran_low_t *out) {
+    int i, j;
+    for (i = 0; i < 32; ++i) {
+      tran_high_t temp_in[32], temp_out[32];
+      for (j = 0; j < 32; ++j)
+        temp_in[j] = intermediate[j * 32 + i];
+      vp9_fdct32(temp_in, temp_out, 0);
+      for (j = 0; j < 32; ++j)
+        out[j + i * 32] = (temp_out[j] + 1 + (temp_out[j] < 0)) >> 2;
+    }
+}
+  #define HIGH_FDCT32x32_2D_C vp9_highbd_fdct32x32_c
+  #define HIGH_FDCT32x32_2D_ROWS_C vp9_fdct32x32_rows_c
+#else
+void vp9_fdct32x32_rd_rows_c(const int16_t *intermediate, tran_low_t *out) {
+    int i, j;
+    for (i = 0; i < 32; ++i) {
+      tran_high_t temp_in[32], temp_out[32];
+      for (j = 0; j < 32; ++j)
+        temp_in[j] = intermediate[j * 32 + i];
+      vp9_fdct32(temp_in, temp_out, 1);
+      for (j = 0; j < 32; ++j)
+        out[j + i * 32] = temp_out[j];
+    }
+}
+  #define HIGH_FDCT32x32_2D_C vp9_highbd_fdct32x32_rd_c
+  #define HIGH_FDCT32x32_2D_ROWS_C vp9_fdct32x32_rd_rows_c
+#endif  // FDCT32x32_HIGH_PRECISION
+#else
+#define ADD_EPI16 _mm_add_epi16
+#define SUB_EPI16 _mm_sub_epi16
+#endif  // DCT_HIGH_BIT_DEPTH
+
+
+void FDCT32x32_2D(const int16_t *input,
+                  tran_low_t *output_org, int stride) {
+  // Calculate pre-multiplied strides
+  const int str1 = stride;
+  const int str2 = 2 * stride;
+  const int str3 = 2 * stride + str1;
+  // We need an intermediate buffer between passes.
+  DECLARE_ALIGNED(16, int16_t, intermediate[32 * 32]);
+  // Constants
+  //    When we use them, in one case, they are all the same. In all others
+  //    it's a pair of them that we need to repeat four times. This is done
+  //    by constructing the 32 bit constant corresponding to that pair.
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(+cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64,   cospi_24_64);
+  const __m128i k__cospi_m24_m08 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
+  const __m128i k__cospi_p24_p08 = pair_set_epi16(+cospi_24_64,  cospi_8_64);
+  const __m128i k__cospi_p12_p20 = pair_set_epi16(+cospi_12_64,  cospi_20_64);
+  const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64,  cospi_12_64);
+  const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64,   cospi_28_64);
+  const __m128i k__cospi_p28_p04 = pair_set_epi16(+cospi_28_64,  cospi_4_64);
+  const __m128i k__cospi_m28_m04 = pair_set_epi16(-cospi_28_64, -cospi_4_64);
+  const __m128i k__cospi_m12_m20 = pair_set_epi16(-cospi_12_64, -cospi_20_64);
+  const __m128i k__cospi_p30_p02 = pair_set_epi16(+cospi_30_64,  cospi_2_64);
+  const __m128i k__cospi_p14_p18 = pair_set_epi16(+cospi_14_64,  cospi_18_64);
+  const __m128i k__cospi_p22_p10 = pair_set_epi16(+cospi_22_64,  cospi_10_64);
+  const __m128i k__cospi_p06_p26 = pair_set_epi16(+cospi_6_64,   cospi_26_64);
+  const __m128i k__cospi_m26_p06 = pair_set_epi16(-cospi_26_64,  cospi_6_64);
+  const __m128i k__cospi_m10_p22 = pair_set_epi16(-cospi_10_64,  cospi_22_64);
+  const __m128i k__cospi_m18_p14 = pair_set_epi16(-cospi_18_64,  cospi_14_64);
+  const __m128i k__cospi_m02_p30 = pair_set_epi16(-cospi_2_64,   cospi_30_64);
+  const __m128i k__cospi_p31_p01 = pair_set_epi16(+cospi_31_64,  cospi_1_64);
+  const __m128i k__cospi_p15_p17 = pair_set_epi16(+cospi_15_64,  cospi_17_64);
+  const __m128i k__cospi_p23_p09 = pair_set_epi16(+cospi_23_64,  cospi_9_64);
+  const __m128i k__cospi_p07_p25 = pair_set_epi16(+cospi_7_64,   cospi_25_64);
+  const __m128i k__cospi_m25_p07 = pair_set_epi16(-cospi_25_64,  cospi_7_64);
+  const __m128i k__cospi_m09_p23 = pair_set_epi16(-cospi_9_64,   cospi_23_64);
+  const __m128i k__cospi_m17_p15 = pair_set_epi16(-cospi_17_64,  cospi_15_64);
+  const __m128i k__cospi_m01_p31 = pair_set_epi16(-cospi_1_64,   cospi_31_64);
+  const __m128i k__cospi_p27_p05 = pair_set_epi16(+cospi_27_64,  cospi_5_64);
+  const __m128i k__cospi_p11_p21 = pair_set_epi16(+cospi_11_64,  cospi_21_64);
+  const __m128i k__cospi_p19_p13 = pair_set_epi16(+cospi_19_64,  cospi_13_64);
+  const __m128i k__cospi_p03_p29 = pair_set_epi16(+cospi_3_64,   cospi_29_64);
+  const __m128i k__cospi_m29_p03 = pair_set_epi16(-cospi_29_64,  cospi_3_64);
+  const __m128i k__cospi_m13_p19 = pair_set_epi16(-cospi_13_64,  cospi_19_64);
+  const __m128i k__cospi_m21_p11 = pair_set_epi16(-cospi_21_64,  cospi_11_64);
+  const __m128i k__cospi_m05_p27 = pair_set_epi16(-cospi_5_64,   cospi_27_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i kZero = _mm_set1_epi16(0);
+  const __m128i kOne  = _mm_set1_epi16(1);
+  // Do the two transform/transpose passes
+  int pass;
+#if DCT_HIGH_BIT_DEPTH
+  int overflow;
+#endif
+  for (pass = 0; pass < 2; ++pass) {
+    // We process eight columns (transposed rows in second pass) at a time.
+    int column_start;
+    for (column_start = 0; column_start < 32; column_start += 8) {
+      __m128i step1[32];
+      __m128i step2[32];
+      __m128i step3[32];
+      __m128i out[32];
+      // Stage 1
+      // Note: even though all the loads below are aligned, using the aligned
+      //       intrinsic make the code slightly slower.
+      if (0 == pass) {
+        const int16_t *in  = &input[column_start];
+        // step1[i] =  (in[ 0 * stride] + in[(32 -  1) * stride]) << 2;
+        // Note: the next four blocks could be in a loop. That would help the
+        //       instruction cache but is actually slower.
+        {
+          const int16_t *ina =  in +  0 * str1;
+          const int16_t *inb =  in + 31 * str1;
+          __m128i *step1a = &step1[ 0];
+          __m128i *step1b = &step1[31];
+          const __m128i ina0  = _mm_loadu_si128((const __m128i *)(ina));
+          const __m128i ina1  = _mm_loadu_si128((const __m128i *)(ina + str1));
+          const __m128i ina2  = _mm_loadu_si128((const __m128i *)(ina + str2));
+          const __m128i ina3  = _mm_loadu_si128((const __m128i *)(ina + str3));
+          const __m128i inb3  = _mm_loadu_si128((const __m128i *)(inb - str3));
+          const __m128i inb2  = _mm_loadu_si128((const __m128i *)(inb - str2));
+          const __m128i inb1  = _mm_loadu_si128((const __m128i *)(inb - str1));
+          const __m128i inb0  = _mm_loadu_si128((const __m128i *)(inb));
+          step1a[ 0] = _mm_add_epi16(ina0, inb0);
+          step1a[ 1] = _mm_add_epi16(ina1, inb1);
+          step1a[ 2] = _mm_add_epi16(ina2, inb2);
+          step1a[ 3] = _mm_add_epi16(ina3, inb3);
+          step1b[-3] = _mm_sub_epi16(ina3, inb3);
+          step1b[-2] = _mm_sub_epi16(ina2, inb2);
+          step1b[-1] = _mm_sub_epi16(ina1, inb1);
+          step1b[-0] = _mm_sub_epi16(ina0, inb0);
+          step1a[ 0] = _mm_slli_epi16(step1a[ 0], 2);
+          step1a[ 1] = _mm_slli_epi16(step1a[ 1], 2);
+          step1a[ 2] = _mm_slli_epi16(step1a[ 2], 2);
+          step1a[ 3] = _mm_slli_epi16(step1a[ 3], 2);
+          step1b[-3] = _mm_slli_epi16(step1b[-3], 2);
+          step1b[-2] = _mm_slli_epi16(step1b[-2], 2);
+          step1b[-1] = _mm_slli_epi16(step1b[-1], 2);
+          step1b[-0] = _mm_slli_epi16(step1b[-0], 2);
+        }
+        {
+          const int16_t *ina =  in +  4 * str1;
+          const int16_t *inb =  in + 27 * str1;
+          __m128i *step1a = &step1[ 4];
+          __m128i *step1b = &step1[27];
+          const __m128i ina0  = _mm_loadu_si128((const __m128i *)(ina));
+          const __m128i ina1  = _mm_loadu_si128((const __m128i *)(ina + str1));
+          const __m128i ina2  = _mm_loadu_si128((const __m128i *)(ina + str2));
+          const __m128i ina3  = _mm_loadu_si128((const __m128i *)(ina + str3));
+          const __m128i inb3  = _mm_loadu_si128((const __m128i *)(inb - str3));
+          const __m128i inb2  = _mm_loadu_si128((const __m128i *)(inb - str2));
+          const __m128i inb1  = _mm_loadu_si128((const __m128i *)(inb - str1));
+          const __m128i inb0  = _mm_loadu_si128((const __m128i *)(inb));
+          step1a[ 0] = _mm_add_epi16(ina0, inb0);
+          step1a[ 1] = _mm_add_epi16(ina1, inb1);
+          step1a[ 2] = _mm_add_epi16(ina2, inb2);
+          step1a[ 3] = _mm_add_epi16(ina3, inb3);
+          step1b[-3] = _mm_sub_epi16(ina3, inb3);
+          step1b[-2] = _mm_sub_epi16(ina2, inb2);
+          step1b[-1] = _mm_sub_epi16(ina1, inb1);
+          step1b[-0] = _mm_sub_epi16(ina0, inb0);
+          step1a[ 0] = _mm_slli_epi16(step1a[ 0], 2);
+          step1a[ 1] = _mm_slli_epi16(step1a[ 1], 2);
+          step1a[ 2] = _mm_slli_epi16(step1a[ 2], 2);
+          step1a[ 3] = _mm_slli_epi16(step1a[ 3], 2);
+          step1b[-3] = _mm_slli_epi16(step1b[-3], 2);
+          step1b[-2] = _mm_slli_epi16(step1b[-2], 2);
+          step1b[-1] = _mm_slli_epi16(step1b[-1], 2);
+          step1b[-0] = _mm_slli_epi16(step1b[-0], 2);
+        }
+        {
+          const int16_t *ina =  in +  8 * str1;
+          const int16_t *inb =  in + 23 * str1;
+          __m128i *step1a = &step1[ 8];
+          __m128i *step1b = &step1[23];
+          const __m128i ina0  = _mm_loadu_si128((const __m128i *)(ina));
+          const __m128i ina1  = _mm_loadu_si128((const __m128i *)(ina + str1));
+          const __m128i ina2  = _mm_loadu_si128((const __m128i *)(ina + str2));
+          const __m128i ina3  = _mm_loadu_si128((const __m128i *)(ina + str3));
+          const __m128i inb3  = _mm_loadu_si128((const __m128i *)(inb - str3));
+          const __m128i inb2  = _mm_loadu_si128((const __m128i *)(inb - str2));
+          const __m128i inb1  = _mm_loadu_si128((const __m128i *)(inb - str1));
+          const __m128i inb0  = _mm_loadu_si128((const __m128i *)(inb));
+          step1a[ 0] = _mm_add_epi16(ina0, inb0);
+          step1a[ 1] = _mm_add_epi16(ina1, inb1);
+          step1a[ 2] = _mm_add_epi16(ina2, inb2);
+          step1a[ 3] = _mm_add_epi16(ina3, inb3);
+          step1b[-3] = _mm_sub_epi16(ina3, inb3);
+          step1b[-2] = _mm_sub_epi16(ina2, inb2);
+          step1b[-1] = _mm_sub_epi16(ina1, inb1);
+          step1b[-0] = _mm_sub_epi16(ina0, inb0);
+          step1a[ 0] = _mm_slli_epi16(step1a[ 0], 2);
+          step1a[ 1] = _mm_slli_epi16(step1a[ 1], 2);
+          step1a[ 2] = _mm_slli_epi16(step1a[ 2], 2);
+          step1a[ 3] = _mm_slli_epi16(step1a[ 3], 2);
+          step1b[-3] = _mm_slli_epi16(step1b[-3], 2);
+          step1b[-2] = _mm_slli_epi16(step1b[-2], 2);
+          step1b[-1] = _mm_slli_epi16(step1b[-1], 2);
+          step1b[-0] = _mm_slli_epi16(step1b[-0], 2);
+        }
+        {
+          const int16_t *ina =  in + 12 * str1;
+          const int16_t *inb =  in + 19 * str1;
+          __m128i *step1a = &step1[12];
+          __m128i *step1b = &step1[19];
+          const __m128i ina0  = _mm_loadu_si128((const __m128i *)(ina));
+          const __m128i ina1  = _mm_loadu_si128((const __m128i *)(ina + str1));
+          const __m128i ina2  = _mm_loadu_si128((const __m128i *)(ina + str2));
+          const __m128i ina3  = _mm_loadu_si128((const __m128i *)(ina + str3));
+          const __m128i inb3  = _mm_loadu_si128((const __m128i *)(inb - str3));
+          const __m128i inb2  = _mm_loadu_si128((const __m128i *)(inb - str2));
+          const __m128i inb1  = _mm_loadu_si128((const __m128i *)(inb - str1));
+          const __m128i inb0  = _mm_loadu_si128((const __m128i *)(inb));
+          step1a[ 0] = _mm_add_epi16(ina0, inb0);
+          step1a[ 1] = _mm_add_epi16(ina1, inb1);
+          step1a[ 2] = _mm_add_epi16(ina2, inb2);
+          step1a[ 3] = _mm_add_epi16(ina3, inb3);
+          step1b[-3] = _mm_sub_epi16(ina3, inb3);
+          step1b[-2] = _mm_sub_epi16(ina2, inb2);
+          step1b[-1] = _mm_sub_epi16(ina1, inb1);
+          step1b[-0] = _mm_sub_epi16(ina0, inb0);
+          step1a[ 0] = _mm_slli_epi16(step1a[ 0], 2);
+          step1a[ 1] = _mm_slli_epi16(step1a[ 1], 2);
+          step1a[ 2] = _mm_slli_epi16(step1a[ 2], 2);
+          step1a[ 3] = _mm_slli_epi16(step1a[ 3], 2);
+          step1b[-3] = _mm_slli_epi16(step1b[-3], 2);
+          step1b[-2] = _mm_slli_epi16(step1b[-2], 2);
+          step1b[-1] = _mm_slli_epi16(step1b[-1], 2);
+          step1b[-0] = _mm_slli_epi16(step1b[-0], 2);
+        }
+      } else {
+        int16_t *in = &intermediate[column_start];
+        // step1[i] =  in[ 0 * 32] + in[(32 -  1) * 32];
+        // Note: using the same approach as above to have common offset is
+        //       counter-productive as all offsets can be calculated at compile
+        //       time.
+        // Note: the next four blocks could be in a loop. That would help the
+        //       instruction cache but is actually slower.
+        {
+          __m128i in00  = _mm_loadu_si128((const __m128i *)(in +  0 * 32));
+          __m128i in01  = _mm_loadu_si128((const __m128i *)(in +  1 * 32));
+          __m128i in02  = _mm_loadu_si128((const __m128i *)(in +  2 * 32));
+          __m128i in03  = _mm_loadu_si128((const __m128i *)(in +  3 * 32));
+          __m128i in28  = _mm_loadu_si128((const __m128i *)(in + 28 * 32));
+          __m128i in29  = _mm_loadu_si128((const __m128i *)(in + 29 * 32));
+          __m128i in30  = _mm_loadu_si128((const __m128i *)(in + 30 * 32));
+          __m128i in31  = _mm_loadu_si128((const __m128i *)(in + 31 * 32));
+          step1[0] = ADD_EPI16(in00, in31);
+          step1[1] = ADD_EPI16(in01, in30);
+          step1[2] = ADD_EPI16(in02, in29);
+          step1[3] = ADD_EPI16(in03, in28);
+          step1[28] = SUB_EPI16(in03, in28);
+          step1[29] = SUB_EPI16(in02, in29);
+          step1[30] = SUB_EPI16(in01, in30);
+          step1[31] = SUB_EPI16(in00, in31);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x8(&step1[0], &step1[1], &step1[2],
+                                             &step1[3], &step1[28], &step1[29],
+                                             &step1[30], &step1[31]);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        {
+          __m128i in04  = _mm_loadu_si128((const __m128i *)(in +  4 * 32));
+          __m128i in05  = _mm_loadu_si128((const __m128i *)(in +  5 * 32));
+          __m128i in06  = _mm_loadu_si128((const __m128i *)(in +  6 * 32));
+          __m128i in07  = _mm_loadu_si128((const __m128i *)(in +  7 * 32));
+          __m128i in24  = _mm_loadu_si128((const __m128i *)(in + 24 * 32));
+          __m128i in25  = _mm_loadu_si128((const __m128i *)(in + 25 * 32));
+          __m128i in26  = _mm_loadu_si128((const __m128i *)(in + 26 * 32));
+          __m128i in27  = _mm_loadu_si128((const __m128i *)(in + 27 * 32));
+          step1[4] = ADD_EPI16(in04, in27);
+          step1[5] = ADD_EPI16(in05, in26);
+          step1[6] = ADD_EPI16(in06, in25);
+          step1[7] = ADD_EPI16(in07, in24);
+          step1[24] = SUB_EPI16(in07, in24);
+          step1[25] = SUB_EPI16(in06, in25);
+          step1[26] = SUB_EPI16(in05, in26);
+          step1[27] = SUB_EPI16(in04, in27);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x8(&step1[4], &step1[5], &step1[6],
+                                             &step1[7], &step1[24], &step1[25],
+                                             &step1[26], &step1[27]);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        {
+          __m128i in08  = _mm_loadu_si128((const __m128i *)(in +  8 * 32));
+          __m128i in09  = _mm_loadu_si128((const __m128i *)(in +  9 * 32));
+          __m128i in10  = _mm_loadu_si128((const __m128i *)(in + 10 * 32));
+          __m128i in11  = _mm_loadu_si128((const __m128i *)(in + 11 * 32));
+          __m128i in20  = _mm_loadu_si128((const __m128i *)(in + 20 * 32));
+          __m128i in21  = _mm_loadu_si128((const __m128i *)(in + 21 * 32));
+          __m128i in22  = _mm_loadu_si128((const __m128i *)(in + 22 * 32));
+          __m128i in23  = _mm_loadu_si128((const __m128i *)(in + 23 * 32));
+          step1[8] = ADD_EPI16(in08, in23);
+          step1[9] = ADD_EPI16(in09, in22);
+          step1[10] = ADD_EPI16(in10, in21);
+          step1[11] = ADD_EPI16(in11, in20);
+          step1[20] = SUB_EPI16(in11, in20);
+          step1[21] = SUB_EPI16(in10, in21);
+          step1[22] = SUB_EPI16(in09, in22);
+          step1[23] = SUB_EPI16(in08, in23);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x8(&step1[8], &step1[9], &step1[10],
+                                             &step1[11], &step1[20], &step1[21],
+                                             &step1[22], &step1[23]);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        {
+          __m128i in12  = _mm_loadu_si128((const __m128i *)(in + 12 * 32));
+          __m128i in13  = _mm_loadu_si128((const __m128i *)(in + 13 * 32));
+          __m128i in14  = _mm_loadu_si128((const __m128i *)(in + 14 * 32));
+          __m128i in15  = _mm_loadu_si128((const __m128i *)(in + 15 * 32));
+          __m128i in16  = _mm_loadu_si128((const __m128i *)(in + 16 * 32));
+          __m128i in17  = _mm_loadu_si128((const __m128i *)(in + 17 * 32));
+          __m128i in18  = _mm_loadu_si128((const __m128i *)(in + 18 * 32));
+          __m128i in19  = _mm_loadu_si128((const __m128i *)(in + 19 * 32));
+          step1[12] = ADD_EPI16(in12, in19);
+          step1[13] = ADD_EPI16(in13, in18);
+          step1[14] = ADD_EPI16(in14, in17);
+          step1[15] = ADD_EPI16(in15, in16);
+          step1[16] = SUB_EPI16(in15, in16);
+          step1[17] = SUB_EPI16(in14, in17);
+          step1[18] = SUB_EPI16(in13, in18);
+          step1[19] = SUB_EPI16(in12, in19);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x8(&step1[12], &step1[13], &step1[14],
+                                             &step1[15], &step1[16], &step1[17],
+                                             &step1[18], &step1[19]);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+      }
+      // Stage 2
+      {
+        step2[0] = ADD_EPI16(step1[0], step1[15]);
+        step2[1] = ADD_EPI16(step1[1], step1[14]);
+        step2[2] = ADD_EPI16(step1[2], step1[13]);
+        step2[3] = ADD_EPI16(step1[3], step1[12]);
+        step2[4] = ADD_EPI16(step1[4], step1[11]);
+        step2[5] = ADD_EPI16(step1[5], step1[10]);
+        step2[6] = ADD_EPI16(step1[6], step1[ 9]);
+        step2[7] = ADD_EPI16(step1[7], step1[ 8]);
+        step2[8] = SUB_EPI16(step1[7], step1[ 8]);
+        step2[9] = SUB_EPI16(step1[6], step1[ 9]);
+        step2[10] = SUB_EPI16(step1[5], step1[10]);
+        step2[11] = SUB_EPI16(step1[4], step1[11]);
+        step2[12] = SUB_EPI16(step1[3], step1[12]);
+        step2[13] = SUB_EPI16(step1[2], step1[13]);
+        step2[14] = SUB_EPI16(step1[1], step1[14]);
+        step2[15] = SUB_EPI16(step1[0], step1[15]);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x16(
+            &step2[0], &step2[1], &step2[2], &step2[3],
+            &step2[4], &step2[5], &step2[6], &step2[7],
+            &step2[8], &step2[9], &step2[10], &step2[11],
+            &step2[12], &step2[13], &step2[14], &step2[15]);
+        if (overflow) {
+          if (pass == 0)
+            HIGH_FDCT32x32_2D_C(input, output_org, stride);
+          else
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+      {
+        const __m128i s2_20_0 = _mm_unpacklo_epi16(step1[27], step1[20]);
+        const __m128i s2_20_1 = _mm_unpackhi_epi16(step1[27], step1[20]);
+        const __m128i s2_21_0 = _mm_unpacklo_epi16(step1[26], step1[21]);
+        const __m128i s2_21_1 = _mm_unpackhi_epi16(step1[26], step1[21]);
+        const __m128i s2_22_0 = _mm_unpacklo_epi16(step1[25], step1[22]);
+        const __m128i s2_22_1 = _mm_unpackhi_epi16(step1[25], step1[22]);
+        const __m128i s2_23_0 = _mm_unpacklo_epi16(step1[24], step1[23]);
+        const __m128i s2_23_1 = _mm_unpackhi_epi16(step1[24], step1[23]);
+        const __m128i s2_20_2 = _mm_madd_epi16(s2_20_0, k__cospi_p16_m16);
+        const __m128i s2_20_3 = _mm_madd_epi16(s2_20_1, k__cospi_p16_m16);
+        const __m128i s2_21_2 = _mm_madd_epi16(s2_21_0, k__cospi_p16_m16);
+        const __m128i s2_21_3 = _mm_madd_epi16(s2_21_1, k__cospi_p16_m16);
+        const __m128i s2_22_2 = _mm_madd_epi16(s2_22_0, k__cospi_p16_m16);
+        const __m128i s2_22_3 = _mm_madd_epi16(s2_22_1, k__cospi_p16_m16);
+        const __m128i s2_23_2 = _mm_madd_epi16(s2_23_0, k__cospi_p16_m16);
+        const __m128i s2_23_3 = _mm_madd_epi16(s2_23_1, k__cospi_p16_m16);
+        const __m128i s2_24_2 = _mm_madd_epi16(s2_23_0, k__cospi_p16_p16);
+        const __m128i s2_24_3 = _mm_madd_epi16(s2_23_1, k__cospi_p16_p16);
+        const __m128i s2_25_2 = _mm_madd_epi16(s2_22_0, k__cospi_p16_p16);
+        const __m128i s2_25_3 = _mm_madd_epi16(s2_22_1, k__cospi_p16_p16);
+        const __m128i s2_26_2 = _mm_madd_epi16(s2_21_0, k__cospi_p16_p16);
+        const __m128i s2_26_3 = _mm_madd_epi16(s2_21_1, k__cospi_p16_p16);
+        const __m128i s2_27_2 = _mm_madd_epi16(s2_20_0, k__cospi_p16_p16);
+        const __m128i s2_27_3 = _mm_madd_epi16(s2_20_1, k__cospi_p16_p16);
+        // dct_const_round_shift
+        const __m128i s2_20_4 = _mm_add_epi32(s2_20_2, k__DCT_CONST_ROUNDING);
+        const __m128i s2_20_5 = _mm_add_epi32(s2_20_3, k__DCT_CONST_ROUNDING);
+        const __m128i s2_21_4 = _mm_add_epi32(s2_21_2, k__DCT_CONST_ROUNDING);
+        const __m128i s2_21_5 = _mm_add_epi32(s2_21_3, k__DCT_CONST_ROUNDING);
+        const __m128i s2_22_4 = _mm_add_epi32(s2_22_2, k__DCT_CONST_ROUNDING);
+        const __m128i s2_22_5 = _mm_add_epi32(s2_22_3, k__DCT_CONST_ROUNDING);
+        const __m128i s2_23_4 = _mm_add_epi32(s2_23_2, k__DCT_CONST_ROUNDING);
+        const __m128i s2_23_5 = _mm_add_epi32(s2_23_3, k__DCT_CONST_ROUNDING);
+        const __m128i s2_24_4 = _mm_add_epi32(s2_24_2, k__DCT_CONST_ROUNDING);
+        const __m128i s2_24_5 = _mm_add_epi32(s2_24_3, k__DCT_CONST_ROUNDING);
+        const __m128i s2_25_4 = _mm_add_epi32(s2_25_2, k__DCT_CONST_ROUNDING);
+        const __m128i s2_25_5 = _mm_add_epi32(s2_25_3, k__DCT_CONST_ROUNDING);
+        const __m128i s2_26_4 = _mm_add_epi32(s2_26_2, k__DCT_CONST_ROUNDING);
+        const __m128i s2_26_5 = _mm_add_epi32(s2_26_3, k__DCT_CONST_ROUNDING);
+        const __m128i s2_27_4 = _mm_add_epi32(s2_27_2, k__DCT_CONST_ROUNDING);
+        const __m128i s2_27_5 = _mm_add_epi32(s2_27_3, k__DCT_CONST_ROUNDING);
+        const __m128i s2_20_6 = _mm_srai_epi32(s2_20_4, DCT_CONST_BITS);
+        const __m128i s2_20_7 = _mm_srai_epi32(s2_20_5, DCT_CONST_BITS);
+        const __m128i s2_21_6 = _mm_srai_epi32(s2_21_4, DCT_CONST_BITS);
+        const __m128i s2_21_7 = _mm_srai_epi32(s2_21_5, DCT_CONST_BITS);
+        const __m128i s2_22_6 = _mm_srai_epi32(s2_22_4, DCT_CONST_BITS);
+        const __m128i s2_22_7 = _mm_srai_epi32(s2_22_5, DCT_CONST_BITS);
+        const __m128i s2_23_6 = _mm_srai_epi32(s2_23_4, DCT_CONST_BITS);
+        const __m128i s2_23_7 = _mm_srai_epi32(s2_23_5, DCT_CONST_BITS);
+        const __m128i s2_24_6 = _mm_srai_epi32(s2_24_4, DCT_CONST_BITS);
+        const __m128i s2_24_7 = _mm_srai_epi32(s2_24_5, DCT_CONST_BITS);
+        const __m128i s2_25_6 = _mm_srai_epi32(s2_25_4, DCT_CONST_BITS);
+        const __m128i s2_25_7 = _mm_srai_epi32(s2_25_5, DCT_CONST_BITS);
+        const __m128i s2_26_6 = _mm_srai_epi32(s2_26_4, DCT_CONST_BITS);
+        const __m128i s2_26_7 = _mm_srai_epi32(s2_26_5, DCT_CONST_BITS);
+        const __m128i s2_27_6 = _mm_srai_epi32(s2_27_4, DCT_CONST_BITS);
+        const __m128i s2_27_7 = _mm_srai_epi32(s2_27_5, DCT_CONST_BITS);
+        // Combine
+        step2[20] = _mm_packs_epi32(s2_20_6, s2_20_7);
+        step2[21] = _mm_packs_epi32(s2_21_6, s2_21_7);
+        step2[22] = _mm_packs_epi32(s2_22_6, s2_22_7);
+        step2[23] = _mm_packs_epi32(s2_23_6, s2_23_7);
+        step2[24] = _mm_packs_epi32(s2_24_6, s2_24_7);
+        step2[25] = _mm_packs_epi32(s2_25_6, s2_25_7);
+        step2[26] = _mm_packs_epi32(s2_26_6, s2_26_7);
+        step2[27] = _mm_packs_epi32(s2_27_6, s2_27_7);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x8(&step2[20], &step2[21], &step2[22],
+                                           &step2[23], &step2[24], &step2[25],
+                                           &step2[26], &step2[27]);
+        if (overflow) {
+          if (pass == 0)
+            HIGH_FDCT32x32_2D_C(input, output_org, stride);
+          else
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+
+#if !FDCT32x32_HIGH_PRECISION
+      // dump the magnitude by half, hence the intermediate values are within
+      // the range of 16 bits.
+      if (1 == pass) {
+        __m128i s3_00_0 = _mm_cmplt_epi16(step2[ 0], kZero);
+        __m128i s3_01_0 = _mm_cmplt_epi16(step2[ 1], kZero);
+        __m128i s3_02_0 = _mm_cmplt_epi16(step2[ 2], kZero);
+        __m128i s3_03_0 = _mm_cmplt_epi16(step2[ 3], kZero);
+        __m128i s3_04_0 = _mm_cmplt_epi16(step2[ 4], kZero);
+        __m128i s3_05_0 = _mm_cmplt_epi16(step2[ 5], kZero);
+        __m128i s3_06_0 = _mm_cmplt_epi16(step2[ 6], kZero);
+        __m128i s3_07_0 = _mm_cmplt_epi16(step2[ 7], kZero);
+        __m128i s2_08_0 = _mm_cmplt_epi16(step2[ 8], kZero);
+        __m128i s2_09_0 = _mm_cmplt_epi16(step2[ 9], kZero);
+        __m128i s3_10_0 = _mm_cmplt_epi16(step2[10], kZero);
+        __m128i s3_11_0 = _mm_cmplt_epi16(step2[11], kZero);
+        __m128i s3_12_0 = _mm_cmplt_epi16(step2[12], kZero);
+        __m128i s3_13_0 = _mm_cmplt_epi16(step2[13], kZero);
+        __m128i s2_14_0 = _mm_cmplt_epi16(step2[14], kZero);
+        __m128i s2_15_0 = _mm_cmplt_epi16(step2[15], kZero);
+        __m128i s3_16_0 = _mm_cmplt_epi16(step1[16], kZero);
+        __m128i s3_17_0 = _mm_cmplt_epi16(step1[17], kZero);
+        __m128i s3_18_0 = _mm_cmplt_epi16(step1[18], kZero);
+        __m128i s3_19_0 = _mm_cmplt_epi16(step1[19], kZero);
+        __m128i s3_20_0 = _mm_cmplt_epi16(step2[20], kZero);
+        __m128i s3_21_0 = _mm_cmplt_epi16(step2[21], kZero);
+        __m128i s3_22_0 = _mm_cmplt_epi16(step2[22], kZero);
+        __m128i s3_23_0 = _mm_cmplt_epi16(step2[23], kZero);
+        __m128i s3_24_0 = _mm_cmplt_epi16(step2[24], kZero);
+        __m128i s3_25_0 = _mm_cmplt_epi16(step2[25], kZero);
+        __m128i s3_26_0 = _mm_cmplt_epi16(step2[26], kZero);
+        __m128i s3_27_0 = _mm_cmplt_epi16(step2[27], kZero);
+        __m128i s3_28_0 = _mm_cmplt_epi16(step1[28], kZero);
+        __m128i s3_29_0 = _mm_cmplt_epi16(step1[29], kZero);
+        __m128i s3_30_0 = _mm_cmplt_epi16(step1[30], kZero);
+        __m128i s3_31_0 = _mm_cmplt_epi16(step1[31], kZero);
+
+        step2[0] = SUB_EPI16(step2[ 0], s3_00_0);
+        step2[1] = SUB_EPI16(step2[ 1], s3_01_0);
+        step2[2] = SUB_EPI16(step2[ 2], s3_02_0);
+        step2[3] = SUB_EPI16(step2[ 3], s3_03_0);
+        step2[4] = SUB_EPI16(step2[ 4], s3_04_0);
+        step2[5] = SUB_EPI16(step2[ 5], s3_05_0);
+        step2[6] = SUB_EPI16(step2[ 6], s3_06_0);
+        step2[7] = SUB_EPI16(step2[ 7], s3_07_0);
+        step2[8] = SUB_EPI16(step2[ 8], s2_08_0);
+        step2[9] = SUB_EPI16(step2[ 9], s2_09_0);
+        step2[10] = SUB_EPI16(step2[10], s3_10_0);
+        step2[11] = SUB_EPI16(step2[11], s3_11_0);
+        step2[12] = SUB_EPI16(step2[12], s3_12_0);
+        step2[13] = SUB_EPI16(step2[13], s3_13_0);
+        step2[14] = SUB_EPI16(step2[14], s2_14_0);
+        step2[15] = SUB_EPI16(step2[15], s2_15_0);
+        step1[16] = SUB_EPI16(step1[16], s3_16_0);
+        step1[17] = SUB_EPI16(step1[17], s3_17_0);
+        step1[18] = SUB_EPI16(step1[18], s3_18_0);
+        step1[19] = SUB_EPI16(step1[19], s3_19_0);
+        step2[20] = SUB_EPI16(step2[20], s3_20_0);
+        step2[21] = SUB_EPI16(step2[21], s3_21_0);
+        step2[22] = SUB_EPI16(step2[22], s3_22_0);
+        step2[23] = SUB_EPI16(step2[23], s3_23_0);
+        step2[24] = SUB_EPI16(step2[24], s3_24_0);
+        step2[25] = SUB_EPI16(step2[25], s3_25_0);
+        step2[26] = SUB_EPI16(step2[26], s3_26_0);
+        step2[27] = SUB_EPI16(step2[27], s3_27_0);
+        step1[28] = SUB_EPI16(step1[28], s3_28_0);
+        step1[29] = SUB_EPI16(step1[29], s3_29_0);
+        step1[30] = SUB_EPI16(step1[30], s3_30_0);
+        step1[31] = SUB_EPI16(step1[31], s3_31_0);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x32(
+            &step2[0], &step2[1], &step2[2], &step2[3],
+            &step2[4], &step2[5], &step2[6], &step2[7],
+            &step2[8], &step2[9], &step2[10], &step2[11],
+            &step2[12], &step2[13], &step2[14], &step2[15],
+            &step1[16], &step1[17], &step1[18], &step1[19],
+            &step2[20], &step2[21], &step2[22], &step2[23],
+            &step2[24], &step2[25], &step2[26], &step2[27],
+            &step1[28], &step1[29], &step1[30], &step1[31]);
+        if (overflow) {
+          HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+        step2[0] = _mm_add_epi16(step2[ 0], kOne);
+        step2[1] = _mm_add_epi16(step2[ 1], kOne);
+        step2[2] = _mm_add_epi16(step2[ 2], kOne);
+        step2[3] = _mm_add_epi16(step2[ 3], kOne);
+        step2[4] = _mm_add_epi16(step2[ 4], kOne);
+        step2[5] = _mm_add_epi16(step2[ 5], kOne);
+        step2[6] = _mm_add_epi16(step2[ 6], kOne);
+        step2[7] = _mm_add_epi16(step2[ 7], kOne);
+        step2[8] = _mm_add_epi16(step2[ 8], kOne);
+        step2[9] = _mm_add_epi16(step2[ 9], kOne);
+        step2[10] = _mm_add_epi16(step2[10], kOne);
+        step2[11] = _mm_add_epi16(step2[11], kOne);
+        step2[12] = _mm_add_epi16(step2[12], kOne);
+        step2[13] = _mm_add_epi16(step2[13], kOne);
+        step2[14] = _mm_add_epi16(step2[14], kOne);
+        step2[15] = _mm_add_epi16(step2[15], kOne);
+        step1[16] = _mm_add_epi16(step1[16], kOne);
+        step1[17] = _mm_add_epi16(step1[17], kOne);
+        step1[18] = _mm_add_epi16(step1[18], kOne);
+        step1[19] = _mm_add_epi16(step1[19], kOne);
+        step2[20] = _mm_add_epi16(step2[20], kOne);
+        step2[21] = _mm_add_epi16(step2[21], kOne);
+        step2[22] = _mm_add_epi16(step2[22], kOne);
+        step2[23] = _mm_add_epi16(step2[23], kOne);
+        step2[24] = _mm_add_epi16(step2[24], kOne);
+        step2[25] = _mm_add_epi16(step2[25], kOne);
+        step2[26] = _mm_add_epi16(step2[26], kOne);
+        step2[27] = _mm_add_epi16(step2[27], kOne);
+        step1[28] = _mm_add_epi16(step1[28], kOne);
+        step1[29] = _mm_add_epi16(step1[29], kOne);
+        step1[30] = _mm_add_epi16(step1[30], kOne);
+        step1[31] = _mm_add_epi16(step1[31], kOne);
+
+        step2[0] = _mm_srai_epi16(step2[ 0], 2);
+        step2[1] = _mm_srai_epi16(step2[ 1], 2);
+        step2[2] = _mm_srai_epi16(step2[ 2], 2);
+        step2[3] = _mm_srai_epi16(step2[ 3], 2);
+        step2[4] = _mm_srai_epi16(step2[ 4], 2);
+        step2[5] = _mm_srai_epi16(step2[ 5], 2);
+        step2[6] = _mm_srai_epi16(step2[ 6], 2);
+        step2[7] = _mm_srai_epi16(step2[ 7], 2);
+        step2[8] = _mm_srai_epi16(step2[ 8], 2);
+        step2[9] = _mm_srai_epi16(step2[ 9], 2);
+        step2[10] = _mm_srai_epi16(step2[10], 2);
+        step2[11] = _mm_srai_epi16(step2[11], 2);
+        step2[12] = _mm_srai_epi16(step2[12], 2);
+        step2[13] = _mm_srai_epi16(step2[13], 2);
+        step2[14] = _mm_srai_epi16(step2[14], 2);
+        step2[15] = _mm_srai_epi16(step2[15], 2);
+        step1[16] = _mm_srai_epi16(step1[16], 2);
+        step1[17] = _mm_srai_epi16(step1[17], 2);
+        step1[18] = _mm_srai_epi16(step1[18], 2);
+        step1[19] = _mm_srai_epi16(step1[19], 2);
+        step2[20] = _mm_srai_epi16(step2[20], 2);
+        step2[21] = _mm_srai_epi16(step2[21], 2);
+        step2[22] = _mm_srai_epi16(step2[22], 2);
+        step2[23] = _mm_srai_epi16(step2[23], 2);
+        step2[24] = _mm_srai_epi16(step2[24], 2);
+        step2[25] = _mm_srai_epi16(step2[25], 2);
+        step2[26] = _mm_srai_epi16(step2[26], 2);
+        step2[27] = _mm_srai_epi16(step2[27], 2);
+        step1[28] = _mm_srai_epi16(step1[28], 2);
+        step1[29] = _mm_srai_epi16(step1[29], 2);
+        step1[30] = _mm_srai_epi16(step1[30], 2);
+        step1[31] = _mm_srai_epi16(step1[31], 2);
+      }
+#endif  // !FDCT32x32_HIGH_PRECISION
+
+#if FDCT32x32_HIGH_PRECISION
+      if (pass == 0) {
+#endif
+      // Stage 3
+      {
+        step3[0] = ADD_EPI16(step2[(8 - 1)], step2[0]);
+        step3[1] = ADD_EPI16(step2[(8 - 2)], step2[1]);
+        step3[2] = ADD_EPI16(step2[(8 - 3)], step2[2]);
+        step3[3] = ADD_EPI16(step2[(8 - 4)], step2[3]);
+        step3[4] = SUB_EPI16(step2[(8 - 5)], step2[4]);
+        step3[5] = SUB_EPI16(step2[(8 - 6)], step2[5]);
+        step3[6] = SUB_EPI16(step2[(8 - 7)], step2[6]);
+        step3[7] = SUB_EPI16(step2[(8 - 8)], step2[7]);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x8(&step3[0], &step3[1], &step3[2],
+                                           &step3[3], &step3[4], &step3[5],
+                                           &step3[6], &step3[7]);
+        if (overflow) {
+          if (pass == 0)
+            HIGH_FDCT32x32_2D_C(input, output_org, stride);
+          else
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+      {
+        const __m128i s3_10_0 = _mm_unpacklo_epi16(step2[13], step2[10]);
+        const __m128i s3_10_1 = _mm_unpackhi_epi16(step2[13], step2[10]);
+        const __m128i s3_11_0 = _mm_unpacklo_epi16(step2[12], step2[11]);
+        const __m128i s3_11_1 = _mm_unpackhi_epi16(step2[12], step2[11]);
+        const __m128i s3_10_2 = _mm_madd_epi16(s3_10_0, k__cospi_p16_m16);
+        const __m128i s3_10_3 = _mm_madd_epi16(s3_10_1, k__cospi_p16_m16);
+        const __m128i s3_11_2 = _mm_madd_epi16(s3_11_0, k__cospi_p16_m16);
+        const __m128i s3_11_3 = _mm_madd_epi16(s3_11_1, k__cospi_p16_m16);
+        const __m128i s3_12_2 = _mm_madd_epi16(s3_11_0, k__cospi_p16_p16);
+        const __m128i s3_12_3 = _mm_madd_epi16(s3_11_1, k__cospi_p16_p16);
+        const __m128i s3_13_2 = _mm_madd_epi16(s3_10_0, k__cospi_p16_p16);
+        const __m128i s3_13_3 = _mm_madd_epi16(s3_10_1, k__cospi_p16_p16);
+        // dct_const_round_shift
+        const __m128i s3_10_4 = _mm_add_epi32(s3_10_2, k__DCT_CONST_ROUNDING);
+        const __m128i s3_10_5 = _mm_add_epi32(s3_10_3, k__DCT_CONST_ROUNDING);
+        const __m128i s3_11_4 = _mm_add_epi32(s3_11_2, k__DCT_CONST_ROUNDING);
+        const __m128i s3_11_5 = _mm_add_epi32(s3_11_3, k__DCT_CONST_ROUNDING);
+        const __m128i s3_12_4 = _mm_add_epi32(s3_12_2, k__DCT_CONST_ROUNDING);
+        const __m128i s3_12_5 = _mm_add_epi32(s3_12_3, k__DCT_CONST_ROUNDING);
+        const __m128i s3_13_4 = _mm_add_epi32(s3_13_2, k__DCT_CONST_ROUNDING);
+        const __m128i s3_13_5 = _mm_add_epi32(s3_13_3, k__DCT_CONST_ROUNDING);
+        const __m128i s3_10_6 = _mm_srai_epi32(s3_10_4, DCT_CONST_BITS);
+        const __m128i s3_10_7 = _mm_srai_epi32(s3_10_5, DCT_CONST_BITS);
+        const __m128i s3_11_6 = _mm_srai_epi32(s3_11_4, DCT_CONST_BITS);
+        const __m128i s3_11_7 = _mm_srai_epi32(s3_11_5, DCT_CONST_BITS);
+        const __m128i s3_12_6 = _mm_srai_epi32(s3_12_4, DCT_CONST_BITS);
+        const __m128i s3_12_7 = _mm_srai_epi32(s3_12_5, DCT_CONST_BITS);
+        const __m128i s3_13_6 = _mm_srai_epi32(s3_13_4, DCT_CONST_BITS);
+        const __m128i s3_13_7 = _mm_srai_epi32(s3_13_5, DCT_CONST_BITS);
+        // Combine
+        step3[10] = _mm_packs_epi32(s3_10_6, s3_10_7);
+        step3[11] = _mm_packs_epi32(s3_11_6, s3_11_7);
+        step3[12] = _mm_packs_epi32(s3_12_6, s3_12_7);
+        step3[13] = _mm_packs_epi32(s3_13_6, s3_13_7);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x4(&step3[10], &step3[11],
+                                           &step3[12], &step3[13]);
+        if (overflow) {
+          if (pass == 0)
+            HIGH_FDCT32x32_2D_C(input, output_org, stride);
+          else
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+      {
+        step3[16] = ADD_EPI16(step2[23], step1[16]);
+        step3[17] = ADD_EPI16(step2[22], step1[17]);
+        step3[18] = ADD_EPI16(step2[21], step1[18]);
+        step3[19] = ADD_EPI16(step2[20], step1[19]);
+        step3[20] = SUB_EPI16(step1[19], step2[20]);
+        step3[21] = SUB_EPI16(step1[18], step2[21]);
+        step3[22] = SUB_EPI16(step1[17], step2[22]);
+        step3[23] = SUB_EPI16(step1[16], step2[23]);
+        step3[24] = SUB_EPI16(step1[31], step2[24]);
+        step3[25] = SUB_EPI16(step1[30], step2[25]);
+        step3[26] = SUB_EPI16(step1[29], step2[26]);
+        step3[27] = SUB_EPI16(step1[28], step2[27]);
+        step3[28] = ADD_EPI16(step2[27], step1[28]);
+        step3[29] = ADD_EPI16(step2[26], step1[29]);
+        step3[30] = ADD_EPI16(step2[25], step1[30]);
+        step3[31] = ADD_EPI16(step2[24], step1[31]);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x16(
+            &step3[16], &step3[17], &step3[18], &step3[19],
+            &step3[20], &step3[21], &step3[22], &step3[23],
+            &step3[24], &step3[25], &step3[26], &step3[27],
+            &step3[28], &step3[29], &step3[30], &step3[31]);
+        if (overflow) {
+          if (pass == 0)
+            HIGH_FDCT32x32_2D_C(input, output_org, stride);
+          else
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+
+      // Stage 4
+      {
+        step1[0] = ADD_EPI16(step3[ 3], step3[ 0]);
+        step1[1] = ADD_EPI16(step3[ 2], step3[ 1]);
+        step1[2] = SUB_EPI16(step3[ 1], step3[ 2]);
+        step1[3] = SUB_EPI16(step3[ 0], step3[ 3]);
+        step1[8] = ADD_EPI16(step3[11], step2[ 8]);
+        step1[9] = ADD_EPI16(step3[10], step2[ 9]);
+        step1[10] = SUB_EPI16(step2[ 9], step3[10]);
+        step1[11] = SUB_EPI16(step2[ 8], step3[11]);
+        step1[12] = SUB_EPI16(step2[15], step3[12]);
+        step1[13] = SUB_EPI16(step2[14], step3[13]);
+        step1[14] = ADD_EPI16(step3[13], step2[14]);
+        step1[15] = ADD_EPI16(step3[12], step2[15]);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x16(
+            &step1[0], &step1[1], &step1[2], &step1[3],
+            &step1[4], &step1[5], &step1[6], &step1[7],
+            &step1[8], &step1[9], &step1[10], &step1[11],
+            &step1[12], &step1[13], &step1[14], &step1[15]);
+        if (overflow) {
+          if (pass == 0)
+            HIGH_FDCT32x32_2D_C(input, output_org, stride);
+          else
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+      {
+        const __m128i s1_05_0 = _mm_unpacklo_epi16(step3[6], step3[5]);
+        const __m128i s1_05_1 = _mm_unpackhi_epi16(step3[6], step3[5]);
+        const __m128i s1_05_2 = _mm_madd_epi16(s1_05_0, k__cospi_p16_m16);
+        const __m128i s1_05_3 = _mm_madd_epi16(s1_05_1, k__cospi_p16_m16);
+        const __m128i s1_06_2 = _mm_madd_epi16(s1_05_0, k__cospi_p16_p16);
+        const __m128i s1_06_3 = _mm_madd_epi16(s1_05_1, k__cospi_p16_p16);
+        // dct_const_round_shift
+        const __m128i s1_05_4 = _mm_add_epi32(s1_05_2, k__DCT_CONST_ROUNDING);
+        const __m128i s1_05_5 = _mm_add_epi32(s1_05_3, k__DCT_CONST_ROUNDING);
+        const __m128i s1_06_4 = _mm_add_epi32(s1_06_2, k__DCT_CONST_ROUNDING);
+        const __m128i s1_06_5 = _mm_add_epi32(s1_06_3, k__DCT_CONST_ROUNDING);
+        const __m128i s1_05_6 = _mm_srai_epi32(s1_05_4, DCT_CONST_BITS);
+        const __m128i s1_05_7 = _mm_srai_epi32(s1_05_5, DCT_CONST_BITS);
+        const __m128i s1_06_6 = _mm_srai_epi32(s1_06_4, DCT_CONST_BITS);
+        const __m128i s1_06_7 = _mm_srai_epi32(s1_06_5, DCT_CONST_BITS);
+        // Combine
+        step1[5] = _mm_packs_epi32(s1_05_6, s1_05_7);
+        step1[6] = _mm_packs_epi32(s1_06_6, s1_06_7);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x2(&step1[5], &step1[6]);
+        if (overflow) {
+          if (pass == 0)
+            HIGH_FDCT32x32_2D_C(input, output_org, stride);
+          else
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+      {
+        const __m128i s1_18_0 = _mm_unpacklo_epi16(step3[18], step3[29]);
+        const __m128i s1_18_1 = _mm_unpackhi_epi16(step3[18], step3[29]);
+        const __m128i s1_19_0 = _mm_unpacklo_epi16(step3[19], step3[28]);
+        const __m128i s1_19_1 = _mm_unpackhi_epi16(step3[19], step3[28]);
+        const __m128i s1_20_0 = _mm_unpacklo_epi16(step3[20], step3[27]);
+        const __m128i s1_20_1 = _mm_unpackhi_epi16(step3[20], step3[27]);
+        const __m128i s1_21_0 = _mm_unpacklo_epi16(step3[21], step3[26]);
+        const __m128i s1_21_1 = _mm_unpackhi_epi16(step3[21], step3[26]);
+        const __m128i s1_18_2 = _mm_madd_epi16(s1_18_0, k__cospi_m08_p24);
+        const __m128i s1_18_3 = _mm_madd_epi16(s1_18_1, k__cospi_m08_p24);
+        const __m128i s1_19_2 = _mm_madd_epi16(s1_19_0, k__cospi_m08_p24);
+        const __m128i s1_19_3 = _mm_madd_epi16(s1_19_1, k__cospi_m08_p24);
+        const __m128i s1_20_2 = _mm_madd_epi16(s1_20_0, k__cospi_m24_m08);
+        const __m128i s1_20_3 = _mm_madd_epi16(s1_20_1, k__cospi_m24_m08);
+        const __m128i s1_21_2 = _mm_madd_epi16(s1_21_0, k__cospi_m24_m08);
+        const __m128i s1_21_3 = _mm_madd_epi16(s1_21_1, k__cospi_m24_m08);
+        const __m128i s1_26_2 = _mm_madd_epi16(s1_21_0, k__cospi_m08_p24);
+        const __m128i s1_26_3 = _mm_madd_epi16(s1_21_1, k__cospi_m08_p24);
+        const __m128i s1_27_2 = _mm_madd_epi16(s1_20_0, k__cospi_m08_p24);
+        const __m128i s1_27_3 = _mm_madd_epi16(s1_20_1, k__cospi_m08_p24);
+        const __m128i s1_28_2 = _mm_madd_epi16(s1_19_0, k__cospi_p24_p08);
+        const __m128i s1_28_3 = _mm_madd_epi16(s1_19_1, k__cospi_p24_p08);
+        const __m128i s1_29_2 = _mm_madd_epi16(s1_18_0, k__cospi_p24_p08);
+        const __m128i s1_29_3 = _mm_madd_epi16(s1_18_1, k__cospi_p24_p08);
+        // dct_const_round_shift
+        const __m128i s1_18_4 = _mm_add_epi32(s1_18_2, k__DCT_CONST_ROUNDING);
+        const __m128i s1_18_5 = _mm_add_epi32(s1_18_3, k__DCT_CONST_ROUNDING);
+        const __m128i s1_19_4 = _mm_add_epi32(s1_19_2, k__DCT_CONST_ROUNDING);
+        const __m128i s1_19_5 = _mm_add_epi32(s1_19_3, k__DCT_CONST_ROUNDING);
+        const __m128i s1_20_4 = _mm_add_epi32(s1_20_2, k__DCT_CONST_ROUNDING);
+        const __m128i s1_20_5 = _mm_add_epi32(s1_20_3, k__DCT_CONST_ROUNDING);
+        const __m128i s1_21_4 = _mm_add_epi32(s1_21_2, k__DCT_CONST_ROUNDING);
+        const __m128i s1_21_5 = _mm_add_epi32(s1_21_3, k__DCT_CONST_ROUNDING);
+        const __m128i s1_26_4 = _mm_add_epi32(s1_26_2, k__DCT_CONST_ROUNDING);
+        const __m128i s1_26_5 = _mm_add_epi32(s1_26_3, k__DCT_CONST_ROUNDING);
+        const __m128i s1_27_4 = _mm_add_epi32(s1_27_2, k__DCT_CONST_ROUNDING);
+        const __m128i s1_27_5 = _mm_add_epi32(s1_27_3, k__DCT_CONST_ROUNDING);
+        const __m128i s1_28_4 = _mm_add_epi32(s1_28_2, k__DCT_CONST_ROUNDING);
+        const __m128i s1_28_5 = _mm_add_epi32(s1_28_3, k__DCT_CONST_ROUNDING);
+        const __m128i s1_29_4 = _mm_add_epi32(s1_29_2, k__DCT_CONST_ROUNDING);
+        const __m128i s1_29_5 = _mm_add_epi32(s1_29_3, k__DCT_CONST_ROUNDING);
+        const __m128i s1_18_6 = _mm_srai_epi32(s1_18_4, DCT_CONST_BITS);
+        const __m128i s1_18_7 = _mm_srai_epi32(s1_18_5, DCT_CONST_BITS);
+        const __m128i s1_19_6 = _mm_srai_epi32(s1_19_4, DCT_CONST_BITS);
+        const __m128i s1_19_7 = _mm_srai_epi32(s1_19_5, DCT_CONST_BITS);
+        const __m128i s1_20_6 = _mm_srai_epi32(s1_20_4, DCT_CONST_BITS);
+        const __m128i s1_20_7 = _mm_srai_epi32(s1_20_5, DCT_CONST_BITS);
+        const __m128i s1_21_6 = _mm_srai_epi32(s1_21_4, DCT_CONST_BITS);
+        const __m128i s1_21_7 = _mm_srai_epi32(s1_21_5, DCT_CONST_BITS);
+        const __m128i s1_26_6 = _mm_srai_epi32(s1_26_4, DCT_CONST_BITS);
+        const __m128i s1_26_7 = _mm_srai_epi32(s1_26_5, DCT_CONST_BITS);
+        const __m128i s1_27_6 = _mm_srai_epi32(s1_27_4, DCT_CONST_BITS);
+        const __m128i s1_27_7 = _mm_srai_epi32(s1_27_5, DCT_CONST_BITS);
+        const __m128i s1_28_6 = _mm_srai_epi32(s1_28_4, DCT_CONST_BITS);
+        const __m128i s1_28_7 = _mm_srai_epi32(s1_28_5, DCT_CONST_BITS);
+        const __m128i s1_29_6 = _mm_srai_epi32(s1_29_4, DCT_CONST_BITS);
+        const __m128i s1_29_7 = _mm_srai_epi32(s1_29_5, DCT_CONST_BITS);
+        // Combine
+        step1[18] = _mm_packs_epi32(s1_18_6, s1_18_7);
+        step1[19] = _mm_packs_epi32(s1_19_6, s1_19_7);
+        step1[20] = _mm_packs_epi32(s1_20_6, s1_20_7);
+        step1[21] = _mm_packs_epi32(s1_21_6, s1_21_7);
+        step1[26] = _mm_packs_epi32(s1_26_6, s1_26_7);
+        step1[27] = _mm_packs_epi32(s1_27_6, s1_27_7);
+        step1[28] = _mm_packs_epi32(s1_28_6, s1_28_7);
+        step1[29] = _mm_packs_epi32(s1_29_6, s1_29_7);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x8(&step1[18], &step1[19], &step1[20],
+                                           &step1[21], &step1[26], &step1[27],
+                                           &step1[28], &step1[29]);
+        if (overflow) {
+          if (pass == 0)
+            HIGH_FDCT32x32_2D_C(input, output_org, stride);
+          else
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+      // Stage 5
+      {
+        step2[4] = ADD_EPI16(step1[5], step3[4]);
+        step2[5] = SUB_EPI16(step3[4], step1[5]);
+        step2[6] = SUB_EPI16(step3[7], step1[6]);
+        step2[7] = ADD_EPI16(step1[6], step3[7]);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x4(&step2[4], &step2[5],
+                                           &step2[6], &step2[7]);
+        if (overflow) {
+          if (pass == 0)
+            HIGH_FDCT32x32_2D_C(input, output_org, stride);
+          else
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+      {
+        const __m128i out_00_0 = _mm_unpacklo_epi16(step1[0], step1[1]);
+        const __m128i out_00_1 = _mm_unpackhi_epi16(step1[0], step1[1]);
+        const __m128i out_08_0 = _mm_unpacklo_epi16(step1[2], step1[3]);
+        const __m128i out_08_1 = _mm_unpackhi_epi16(step1[2], step1[3]);
+        const __m128i out_00_2 = _mm_madd_epi16(out_00_0, k__cospi_p16_p16);
+        const __m128i out_00_3 = _mm_madd_epi16(out_00_1, k__cospi_p16_p16);
+        const __m128i out_16_2 = _mm_madd_epi16(out_00_0, k__cospi_p16_m16);
+        const __m128i out_16_3 = _mm_madd_epi16(out_00_1, k__cospi_p16_m16);
+        const __m128i out_08_2 = _mm_madd_epi16(out_08_0, k__cospi_p24_p08);
+        const __m128i out_08_3 = _mm_madd_epi16(out_08_1, k__cospi_p24_p08);
+        const __m128i out_24_2 = _mm_madd_epi16(out_08_0, k__cospi_m08_p24);
+        const __m128i out_24_3 = _mm_madd_epi16(out_08_1, k__cospi_m08_p24);
+        // dct_const_round_shift
+        const __m128i out_00_4 = _mm_add_epi32(out_00_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_00_5 = _mm_add_epi32(out_00_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_16_4 = _mm_add_epi32(out_16_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_16_5 = _mm_add_epi32(out_16_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_08_4 = _mm_add_epi32(out_08_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_08_5 = _mm_add_epi32(out_08_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_24_4 = _mm_add_epi32(out_24_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_24_5 = _mm_add_epi32(out_24_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_00_6 = _mm_srai_epi32(out_00_4, DCT_CONST_BITS);
+        const __m128i out_00_7 = _mm_srai_epi32(out_00_5, DCT_CONST_BITS);
+        const __m128i out_16_6 = _mm_srai_epi32(out_16_4, DCT_CONST_BITS);
+        const __m128i out_16_7 = _mm_srai_epi32(out_16_5, DCT_CONST_BITS);
+        const __m128i out_08_6 = _mm_srai_epi32(out_08_4, DCT_CONST_BITS);
+        const __m128i out_08_7 = _mm_srai_epi32(out_08_5, DCT_CONST_BITS);
+        const __m128i out_24_6 = _mm_srai_epi32(out_24_4, DCT_CONST_BITS);
+        const __m128i out_24_7 = _mm_srai_epi32(out_24_5, DCT_CONST_BITS);
+        // Combine
+        out[ 0] = _mm_packs_epi32(out_00_6, out_00_7);
+        out[16] = _mm_packs_epi32(out_16_6, out_16_7);
+        out[ 8] = _mm_packs_epi32(out_08_6, out_08_7);
+        out[24] = _mm_packs_epi32(out_24_6, out_24_7);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x4(&out[0], &out[16],
+                                           &out[8], &out[24]);
+        if (overflow) {
+          if (pass == 0)
+            HIGH_FDCT32x32_2D_C(input, output_org, stride);
+          else
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+      {
+        const __m128i s2_09_0 = _mm_unpacklo_epi16(step1[ 9], step1[14]);
+        const __m128i s2_09_1 = _mm_unpackhi_epi16(step1[ 9], step1[14]);
+        const __m128i s2_10_0 = _mm_unpacklo_epi16(step1[10], step1[13]);
+        const __m128i s2_10_1 = _mm_unpackhi_epi16(step1[10], step1[13]);
+        const __m128i s2_09_2 = _mm_madd_epi16(s2_09_0, k__cospi_m08_p24);
+        const __m128i s2_09_3 = _mm_madd_epi16(s2_09_1, k__cospi_m08_p24);
+        const __m128i s2_10_2 = _mm_madd_epi16(s2_10_0, k__cospi_m24_m08);
+        const __m128i s2_10_3 = _mm_madd_epi16(s2_10_1, k__cospi_m24_m08);
+        const __m128i s2_13_2 = _mm_madd_epi16(s2_10_0, k__cospi_m08_p24);
+        const __m128i s2_13_3 = _mm_madd_epi16(s2_10_1, k__cospi_m08_p24);
+        const __m128i s2_14_2 = _mm_madd_epi16(s2_09_0, k__cospi_p24_p08);
+        const __m128i s2_14_3 = _mm_madd_epi16(s2_09_1, k__cospi_p24_p08);
+        // dct_const_round_shift
+        const __m128i s2_09_4 = _mm_add_epi32(s2_09_2, k__DCT_CONST_ROUNDING);
+        const __m128i s2_09_5 = _mm_add_epi32(s2_09_3, k__DCT_CONST_ROUNDING);
+        const __m128i s2_10_4 = _mm_add_epi32(s2_10_2, k__DCT_CONST_ROUNDING);
+        const __m128i s2_10_5 = _mm_add_epi32(s2_10_3, k__DCT_CONST_ROUNDING);
+        const __m128i s2_13_4 = _mm_add_epi32(s2_13_2, k__DCT_CONST_ROUNDING);
+        const __m128i s2_13_5 = _mm_add_epi32(s2_13_3, k__DCT_CONST_ROUNDING);
+        const __m128i s2_14_4 = _mm_add_epi32(s2_14_2, k__DCT_CONST_ROUNDING);
+        const __m128i s2_14_5 = _mm_add_epi32(s2_14_3, k__DCT_CONST_ROUNDING);
+        const __m128i s2_09_6 = _mm_srai_epi32(s2_09_4, DCT_CONST_BITS);
+        const __m128i s2_09_7 = _mm_srai_epi32(s2_09_5, DCT_CONST_BITS);
+        const __m128i s2_10_6 = _mm_srai_epi32(s2_10_4, DCT_CONST_BITS);
+        const __m128i s2_10_7 = _mm_srai_epi32(s2_10_5, DCT_CONST_BITS);
+        const __m128i s2_13_6 = _mm_srai_epi32(s2_13_4, DCT_CONST_BITS);
+        const __m128i s2_13_7 = _mm_srai_epi32(s2_13_5, DCT_CONST_BITS);
+        const __m128i s2_14_6 = _mm_srai_epi32(s2_14_4, DCT_CONST_BITS);
+        const __m128i s2_14_7 = _mm_srai_epi32(s2_14_5, DCT_CONST_BITS);
+        // Combine
+        step2[ 9] = _mm_packs_epi32(s2_09_6, s2_09_7);
+        step2[10] = _mm_packs_epi32(s2_10_6, s2_10_7);
+        step2[13] = _mm_packs_epi32(s2_13_6, s2_13_7);
+        step2[14] = _mm_packs_epi32(s2_14_6, s2_14_7);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x4(&step2[9], &step2[10],
+                                           &step2[13], &step2[14]);
+        if (overflow) {
+          if (pass == 0)
+            HIGH_FDCT32x32_2D_C(input, output_org, stride);
+          else
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+      {
+        step2[16] = ADD_EPI16(step1[19], step3[16]);
+        step2[17] = ADD_EPI16(step1[18], step3[17]);
+        step2[18] = SUB_EPI16(step3[17], step1[18]);
+        step2[19] = SUB_EPI16(step3[16], step1[19]);
+        step2[20] = SUB_EPI16(step3[23], step1[20]);
+        step2[21] = SUB_EPI16(step3[22], step1[21]);
+        step2[22] = ADD_EPI16(step1[21], step3[22]);
+        step2[23] = ADD_EPI16(step1[20], step3[23]);
+        step2[24] = ADD_EPI16(step1[27], step3[24]);
+        step2[25] = ADD_EPI16(step1[26], step3[25]);
+        step2[26] = SUB_EPI16(step3[25], step1[26]);
+        step2[27] = SUB_EPI16(step3[24], step1[27]);
+        step2[28] = SUB_EPI16(step3[31], step1[28]);
+        step2[29] = SUB_EPI16(step3[30], step1[29]);
+        step2[30] = ADD_EPI16(step1[29], step3[30]);
+        step2[31] = ADD_EPI16(step1[28], step3[31]);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x16(
+            &step2[16], &step2[17], &step2[18], &step2[19],
+            &step2[20], &step2[21], &step2[22], &step2[23],
+            &step2[24], &step2[25], &step2[26], &step2[27],
+            &step2[28], &step2[29], &step2[30], &step2[31]);
+        if (overflow) {
+          if (pass == 0)
+            HIGH_FDCT32x32_2D_C(input, output_org, stride);
+          else
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+      // Stage 6
+      {
+        const __m128i out_04_0 = _mm_unpacklo_epi16(step2[4], step2[7]);
+        const __m128i out_04_1 = _mm_unpackhi_epi16(step2[4], step2[7]);
+        const __m128i out_20_0 = _mm_unpacklo_epi16(step2[5], step2[6]);
+        const __m128i out_20_1 = _mm_unpackhi_epi16(step2[5], step2[6]);
+        const __m128i out_12_0 = _mm_unpacklo_epi16(step2[5], step2[6]);
+        const __m128i out_12_1 = _mm_unpackhi_epi16(step2[5], step2[6]);
+        const __m128i out_28_0 = _mm_unpacklo_epi16(step2[4], step2[7]);
+        const __m128i out_28_1 = _mm_unpackhi_epi16(step2[4], step2[7]);
+        const __m128i out_04_2 = _mm_madd_epi16(out_04_0, k__cospi_p28_p04);
+        const __m128i out_04_3 = _mm_madd_epi16(out_04_1, k__cospi_p28_p04);
+        const __m128i out_20_2 = _mm_madd_epi16(out_20_0, k__cospi_p12_p20);
+        const __m128i out_20_3 = _mm_madd_epi16(out_20_1, k__cospi_p12_p20);
+        const __m128i out_12_2 = _mm_madd_epi16(out_12_0, k__cospi_m20_p12);
+        const __m128i out_12_3 = _mm_madd_epi16(out_12_1, k__cospi_m20_p12);
+        const __m128i out_28_2 = _mm_madd_epi16(out_28_0, k__cospi_m04_p28);
+        const __m128i out_28_3 = _mm_madd_epi16(out_28_1, k__cospi_m04_p28);
+        // dct_const_round_shift
+        const __m128i out_04_4 = _mm_add_epi32(out_04_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_04_5 = _mm_add_epi32(out_04_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_20_4 = _mm_add_epi32(out_20_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_20_5 = _mm_add_epi32(out_20_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_12_4 = _mm_add_epi32(out_12_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_12_5 = _mm_add_epi32(out_12_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_28_4 = _mm_add_epi32(out_28_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_28_5 = _mm_add_epi32(out_28_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_04_6 = _mm_srai_epi32(out_04_4, DCT_CONST_BITS);
+        const __m128i out_04_7 = _mm_srai_epi32(out_04_5, DCT_CONST_BITS);
+        const __m128i out_20_6 = _mm_srai_epi32(out_20_4, DCT_CONST_BITS);
+        const __m128i out_20_7 = _mm_srai_epi32(out_20_5, DCT_CONST_BITS);
+        const __m128i out_12_6 = _mm_srai_epi32(out_12_4, DCT_CONST_BITS);
+        const __m128i out_12_7 = _mm_srai_epi32(out_12_5, DCT_CONST_BITS);
+        const __m128i out_28_6 = _mm_srai_epi32(out_28_4, DCT_CONST_BITS);
+        const __m128i out_28_7 = _mm_srai_epi32(out_28_5, DCT_CONST_BITS);
+        // Combine
+        out[4] = _mm_packs_epi32(out_04_6, out_04_7);
+        out[20] = _mm_packs_epi32(out_20_6, out_20_7);
+        out[12] = _mm_packs_epi32(out_12_6, out_12_7);
+        out[28] = _mm_packs_epi32(out_28_6, out_28_7);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x4(&out[4], &out[20],
+                                           &out[12], &out[28]);
+        if (overflow) {
+          if (pass == 0)
+            HIGH_FDCT32x32_2D_C(input, output_org, stride);
+          else
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+      {
+        step3[8] = ADD_EPI16(step2[ 9], step1[ 8]);
+        step3[9] = SUB_EPI16(step1[ 8], step2[ 9]);
+        step3[10] = SUB_EPI16(step1[11], step2[10]);
+        step3[11] = ADD_EPI16(step2[10], step1[11]);
+        step3[12] = ADD_EPI16(step2[13], step1[12]);
+        step3[13] = SUB_EPI16(step1[12], step2[13]);
+        step3[14] = SUB_EPI16(step1[15], step2[14]);
+        step3[15] = ADD_EPI16(step2[14], step1[15]);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x8(&step3[8], &step3[9], &step3[10],
+                                           &step3[11], &step3[12], &step3[13],
+                                           &step3[14], &step3[15]);
+        if (overflow) {
+          if (pass == 0)
+            HIGH_FDCT32x32_2D_C(input, output_org, stride);
+          else
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+      {
+        const __m128i s3_17_0 = _mm_unpacklo_epi16(step2[17], step2[30]);
+        const __m128i s3_17_1 = _mm_unpackhi_epi16(step2[17], step2[30]);
+        const __m128i s3_18_0 = _mm_unpacklo_epi16(step2[18], step2[29]);
+        const __m128i s3_18_1 = _mm_unpackhi_epi16(step2[18], step2[29]);
+        const __m128i s3_21_0 = _mm_unpacklo_epi16(step2[21], step2[26]);
+        const __m128i s3_21_1 = _mm_unpackhi_epi16(step2[21], step2[26]);
+        const __m128i s3_22_0 = _mm_unpacklo_epi16(step2[22], step2[25]);
+        const __m128i s3_22_1 = _mm_unpackhi_epi16(step2[22], step2[25]);
+        const __m128i s3_17_2 = _mm_madd_epi16(s3_17_0, k__cospi_m04_p28);
+        const __m128i s3_17_3 = _mm_madd_epi16(s3_17_1, k__cospi_m04_p28);
+        const __m128i s3_18_2 = _mm_madd_epi16(s3_18_0, k__cospi_m28_m04);
+        const __m128i s3_18_3 = _mm_madd_epi16(s3_18_1, k__cospi_m28_m04);
+        const __m128i s3_21_2 = _mm_madd_epi16(s3_21_0, k__cospi_m20_p12);
+        const __m128i s3_21_3 = _mm_madd_epi16(s3_21_1, k__cospi_m20_p12);
+        const __m128i s3_22_2 = _mm_madd_epi16(s3_22_0, k__cospi_m12_m20);
+        const __m128i s3_22_3 = _mm_madd_epi16(s3_22_1, k__cospi_m12_m20);
+        const __m128i s3_25_2 = _mm_madd_epi16(s3_22_0, k__cospi_m20_p12);
+        const __m128i s3_25_3 = _mm_madd_epi16(s3_22_1, k__cospi_m20_p12);
+        const __m128i s3_26_2 = _mm_madd_epi16(s3_21_0, k__cospi_p12_p20);
+        const __m128i s3_26_3 = _mm_madd_epi16(s3_21_1, k__cospi_p12_p20);
+        const __m128i s3_29_2 = _mm_madd_epi16(s3_18_0, k__cospi_m04_p28);
+        const __m128i s3_29_3 = _mm_madd_epi16(s3_18_1, k__cospi_m04_p28);
+        const __m128i s3_30_2 = _mm_madd_epi16(s3_17_0, k__cospi_p28_p04);
+        const __m128i s3_30_3 = _mm_madd_epi16(s3_17_1, k__cospi_p28_p04);
+        // dct_const_round_shift
+        const __m128i s3_17_4 = _mm_add_epi32(s3_17_2, k__DCT_CONST_ROUNDING);
+        const __m128i s3_17_5 = _mm_add_epi32(s3_17_3, k__DCT_CONST_ROUNDING);
+        const __m128i s3_18_4 = _mm_add_epi32(s3_18_2, k__DCT_CONST_ROUNDING);
+        const __m128i s3_18_5 = _mm_add_epi32(s3_18_3, k__DCT_CONST_ROUNDING);
+        const __m128i s3_21_4 = _mm_add_epi32(s3_21_2, k__DCT_CONST_ROUNDING);
+        const __m128i s3_21_5 = _mm_add_epi32(s3_21_3, k__DCT_CONST_ROUNDING);
+        const __m128i s3_22_4 = _mm_add_epi32(s3_22_2, k__DCT_CONST_ROUNDING);
+        const __m128i s3_22_5 = _mm_add_epi32(s3_22_3, k__DCT_CONST_ROUNDING);
+        const __m128i s3_17_6 = _mm_srai_epi32(s3_17_4, DCT_CONST_BITS);
+        const __m128i s3_17_7 = _mm_srai_epi32(s3_17_5, DCT_CONST_BITS);
+        const __m128i s3_18_6 = _mm_srai_epi32(s3_18_4, DCT_CONST_BITS);
+        const __m128i s3_18_7 = _mm_srai_epi32(s3_18_5, DCT_CONST_BITS);
+        const __m128i s3_21_6 = _mm_srai_epi32(s3_21_4, DCT_CONST_BITS);
+        const __m128i s3_21_7 = _mm_srai_epi32(s3_21_5, DCT_CONST_BITS);
+        const __m128i s3_22_6 = _mm_srai_epi32(s3_22_4, DCT_CONST_BITS);
+        const __m128i s3_22_7 = _mm_srai_epi32(s3_22_5, DCT_CONST_BITS);
+        const __m128i s3_25_4 = _mm_add_epi32(s3_25_2, k__DCT_CONST_ROUNDING);
+        const __m128i s3_25_5 = _mm_add_epi32(s3_25_3, k__DCT_CONST_ROUNDING);
+        const __m128i s3_26_4 = _mm_add_epi32(s3_26_2, k__DCT_CONST_ROUNDING);
+        const __m128i s3_26_5 = _mm_add_epi32(s3_26_3, k__DCT_CONST_ROUNDING);
+        const __m128i s3_29_4 = _mm_add_epi32(s3_29_2, k__DCT_CONST_ROUNDING);
+        const __m128i s3_29_5 = _mm_add_epi32(s3_29_3, k__DCT_CONST_ROUNDING);
+        const __m128i s3_30_4 = _mm_add_epi32(s3_30_2, k__DCT_CONST_ROUNDING);
+        const __m128i s3_30_5 = _mm_add_epi32(s3_30_3, k__DCT_CONST_ROUNDING);
+        const __m128i s3_25_6 = _mm_srai_epi32(s3_25_4, DCT_CONST_BITS);
+        const __m128i s3_25_7 = _mm_srai_epi32(s3_25_5, DCT_CONST_BITS);
+        const __m128i s3_26_6 = _mm_srai_epi32(s3_26_4, DCT_CONST_BITS);
+        const __m128i s3_26_7 = _mm_srai_epi32(s3_26_5, DCT_CONST_BITS);
+        const __m128i s3_29_6 = _mm_srai_epi32(s3_29_4, DCT_CONST_BITS);
+        const __m128i s3_29_7 = _mm_srai_epi32(s3_29_5, DCT_CONST_BITS);
+        const __m128i s3_30_6 = _mm_srai_epi32(s3_30_4, DCT_CONST_BITS);
+        const __m128i s3_30_7 = _mm_srai_epi32(s3_30_5, DCT_CONST_BITS);
+        // Combine
+        step3[17] = _mm_packs_epi32(s3_17_6, s3_17_7);
+        step3[18] = _mm_packs_epi32(s3_18_6, s3_18_7);
+        step3[21] = _mm_packs_epi32(s3_21_6, s3_21_7);
+        step3[22] = _mm_packs_epi32(s3_22_6, s3_22_7);
+        // Combine
+        step3[25] = _mm_packs_epi32(s3_25_6, s3_25_7);
+        step3[26] = _mm_packs_epi32(s3_26_6, s3_26_7);
+        step3[29] = _mm_packs_epi32(s3_29_6, s3_29_7);
+        step3[30] = _mm_packs_epi32(s3_30_6, s3_30_7);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x8(&step3[17], &step3[18], &step3[21],
+                                           &step3[22], &step3[25], &step3[26],
+                                           &step3[29], &step3[30]);
+        if (overflow) {
+          if (pass == 0)
+            HIGH_FDCT32x32_2D_C(input, output_org, stride);
+          else
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+      // Stage 7
+      {
+        const __m128i out_02_0 = _mm_unpacklo_epi16(step3[ 8], step3[15]);
+        const __m128i out_02_1 = _mm_unpackhi_epi16(step3[ 8], step3[15]);
+        const __m128i out_18_0 = _mm_unpacklo_epi16(step3[ 9], step3[14]);
+        const __m128i out_18_1 = _mm_unpackhi_epi16(step3[ 9], step3[14]);
+        const __m128i out_10_0 = _mm_unpacklo_epi16(step3[10], step3[13]);
+        const __m128i out_10_1 = _mm_unpackhi_epi16(step3[10], step3[13]);
+        const __m128i out_26_0 = _mm_unpacklo_epi16(step3[11], step3[12]);
+        const __m128i out_26_1 = _mm_unpackhi_epi16(step3[11], step3[12]);
+        const __m128i out_02_2 = _mm_madd_epi16(out_02_0, k__cospi_p30_p02);
+        const __m128i out_02_3 = _mm_madd_epi16(out_02_1, k__cospi_p30_p02);
+        const __m128i out_18_2 = _mm_madd_epi16(out_18_0, k__cospi_p14_p18);
+        const __m128i out_18_3 = _mm_madd_epi16(out_18_1, k__cospi_p14_p18);
+        const __m128i out_10_2 = _mm_madd_epi16(out_10_0, k__cospi_p22_p10);
+        const __m128i out_10_3 = _mm_madd_epi16(out_10_1, k__cospi_p22_p10);
+        const __m128i out_26_2 = _mm_madd_epi16(out_26_0, k__cospi_p06_p26);
+        const __m128i out_26_3 = _mm_madd_epi16(out_26_1, k__cospi_p06_p26);
+        const __m128i out_06_2 = _mm_madd_epi16(out_26_0, k__cospi_m26_p06);
+        const __m128i out_06_3 = _mm_madd_epi16(out_26_1, k__cospi_m26_p06);
+        const __m128i out_22_2 = _mm_madd_epi16(out_10_0, k__cospi_m10_p22);
+        const __m128i out_22_3 = _mm_madd_epi16(out_10_1, k__cospi_m10_p22);
+        const __m128i out_14_2 = _mm_madd_epi16(out_18_0, k__cospi_m18_p14);
+        const __m128i out_14_3 = _mm_madd_epi16(out_18_1, k__cospi_m18_p14);
+        const __m128i out_30_2 = _mm_madd_epi16(out_02_0, k__cospi_m02_p30);
+        const __m128i out_30_3 = _mm_madd_epi16(out_02_1, k__cospi_m02_p30);
+        // dct_const_round_shift
+        const __m128i out_02_4 = _mm_add_epi32(out_02_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_02_5 = _mm_add_epi32(out_02_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_18_4 = _mm_add_epi32(out_18_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_18_5 = _mm_add_epi32(out_18_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_10_4 = _mm_add_epi32(out_10_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_10_5 = _mm_add_epi32(out_10_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_26_4 = _mm_add_epi32(out_26_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_26_5 = _mm_add_epi32(out_26_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_06_4 = _mm_add_epi32(out_06_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_06_5 = _mm_add_epi32(out_06_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_22_4 = _mm_add_epi32(out_22_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_22_5 = _mm_add_epi32(out_22_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_14_4 = _mm_add_epi32(out_14_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_14_5 = _mm_add_epi32(out_14_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_30_4 = _mm_add_epi32(out_30_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_30_5 = _mm_add_epi32(out_30_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_02_6 = _mm_srai_epi32(out_02_4, DCT_CONST_BITS);
+        const __m128i out_02_7 = _mm_srai_epi32(out_02_5, DCT_CONST_BITS);
+        const __m128i out_18_6 = _mm_srai_epi32(out_18_4, DCT_CONST_BITS);
+        const __m128i out_18_7 = _mm_srai_epi32(out_18_5, DCT_CONST_BITS);
+        const __m128i out_10_6 = _mm_srai_epi32(out_10_4, DCT_CONST_BITS);
+        const __m128i out_10_7 = _mm_srai_epi32(out_10_5, DCT_CONST_BITS);
+        const __m128i out_26_6 = _mm_srai_epi32(out_26_4, DCT_CONST_BITS);
+        const __m128i out_26_7 = _mm_srai_epi32(out_26_5, DCT_CONST_BITS);
+        const __m128i out_06_6 = _mm_srai_epi32(out_06_4, DCT_CONST_BITS);
+        const __m128i out_06_7 = _mm_srai_epi32(out_06_5, DCT_CONST_BITS);
+        const __m128i out_22_6 = _mm_srai_epi32(out_22_4, DCT_CONST_BITS);
+        const __m128i out_22_7 = _mm_srai_epi32(out_22_5, DCT_CONST_BITS);
+        const __m128i out_14_6 = _mm_srai_epi32(out_14_4, DCT_CONST_BITS);
+        const __m128i out_14_7 = _mm_srai_epi32(out_14_5, DCT_CONST_BITS);
+        const __m128i out_30_6 = _mm_srai_epi32(out_30_4, DCT_CONST_BITS);
+        const __m128i out_30_7 = _mm_srai_epi32(out_30_5, DCT_CONST_BITS);
+        // Combine
+        out[ 2] = _mm_packs_epi32(out_02_6, out_02_7);
+        out[18] = _mm_packs_epi32(out_18_6, out_18_7);
+        out[10] = _mm_packs_epi32(out_10_6, out_10_7);
+        out[26] = _mm_packs_epi32(out_26_6, out_26_7);
+        out[ 6] = _mm_packs_epi32(out_06_6, out_06_7);
+        out[22] = _mm_packs_epi32(out_22_6, out_22_7);
+        out[14] = _mm_packs_epi32(out_14_6, out_14_7);
+        out[30] = _mm_packs_epi32(out_30_6, out_30_7);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x8(&out[2], &out[18], &out[10],
+                                           &out[26], &out[6], &out[22],
+                                           &out[14], &out[30]);
+        if (overflow) {
+          if (pass == 0)
+            HIGH_FDCT32x32_2D_C(input, output_org, stride);
+          else
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+      {
+        step1[16] = ADD_EPI16(step3[17], step2[16]);
+        step1[17] = SUB_EPI16(step2[16], step3[17]);
+        step1[18] = SUB_EPI16(step2[19], step3[18]);
+        step1[19] = ADD_EPI16(step3[18], step2[19]);
+        step1[20] = ADD_EPI16(step3[21], step2[20]);
+        step1[21] = SUB_EPI16(step2[20], step3[21]);
+        step1[22] = SUB_EPI16(step2[23], step3[22]);
+        step1[23] = ADD_EPI16(step3[22], step2[23]);
+        step1[24] = ADD_EPI16(step3[25], step2[24]);
+        step1[25] = SUB_EPI16(step2[24], step3[25]);
+        step1[26] = SUB_EPI16(step2[27], step3[26]);
+        step1[27] = ADD_EPI16(step3[26], step2[27]);
+        step1[28] = ADD_EPI16(step3[29], step2[28]);
+        step1[29] = SUB_EPI16(step2[28], step3[29]);
+        step1[30] = SUB_EPI16(step2[31], step3[30]);
+        step1[31] = ADD_EPI16(step3[30], step2[31]);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x16(
+            &step1[16], &step1[17], &step1[18], &step1[19],
+            &step1[20], &step1[21], &step1[22], &step1[23],
+            &step1[24], &step1[25], &step1[26], &step1[27],
+            &step1[28], &step1[29], &step1[30], &step1[31]);
+        if (overflow) {
+          if (pass == 0)
+            HIGH_FDCT32x32_2D_C(input, output_org, stride);
+          else
+             HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+      // Final stage --- outputs indices are bit-reversed.
+      {
+        const __m128i out_01_0 = _mm_unpacklo_epi16(step1[16], step1[31]);
+        const __m128i out_01_1 = _mm_unpackhi_epi16(step1[16], step1[31]);
+        const __m128i out_17_0 = _mm_unpacklo_epi16(step1[17], step1[30]);
+        const __m128i out_17_1 = _mm_unpackhi_epi16(step1[17], step1[30]);
+        const __m128i out_09_0 = _mm_unpacklo_epi16(step1[18], step1[29]);
+        const __m128i out_09_1 = _mm_unpackhi_epi16(step1[18], step1[29]);
+        const __m128i out_25_0 = _mm_unpacklo_epi16(step1[19], step1[28]);
+        const __m128i out_25_1 = _mm_unpackhi_epi16(step1[19], step1[28]);
+        const __m128i out_01_2 = _mm_madd_epi16(out_01_0, k__cospi_p31_p01);
+        const __m128i out_01_3 = _mm_madd_epi16(out_01_1, k__cospi_p31_p01);
+        const __m128i out_17_2 = _mm_madd_epi16(out_17_0, k__cospi_p15_p17);
+        const __m128i out_17_3 = _mm_madd_epi16(out_17_1, k__cospi_p15_p17);
+        const __m128i out_09_2 = _mm_madd_epi16(out_09_0, k__cospi_p23_p09);
+        const __m128i out_09_3 = _mm_madd_epi16(out_09_1, k__cospi_p23_p09);
+        const __m128i out_25_2 = _mm_madd_epi16(out_25_0, k__cospi_p07_p25);
+        const __m128i out_25_3 = _mm_madd_epi16(out_25_1, k__cospi_p07_p25);
+        const __m128i out_07_2 = _mm_madd_epi16(out_25_0, k__cospi_m25_p07);
+        const __m128i out_07_3 = _mm_madd_epi16(out_25_1, k__cospi_m25_p07);
+        const __m128i out_23_2 = _mm_madd_epi16(out_09_0, k__cospi_m09_p23);
+        const __m128i out_23_3 = _mm_madd_epi16(out_09_1, k__cospi_m09_p23);
+        const __m128i out_15_2 = _mm_madd_epi16(out_17_0, k__cospi_m17_p15);
+        const __m128i out_15_3 = _mm_madd_epi16(out_17_1, k__cospi_m17_p15);
+        const __m128i out_31_2 = _mm_madd_epi16(out_01_0, k__cospi_m01_p31);
+        const __m128i out_31_3 = _mm_madd_epi16(out_01_1, k__cospi_m01_p31);
+        // dct_const_round_shift
+        const __m128i out_01_4 = _mm_add_epi32(out_01_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_01_5 = _mm_add_epi32(out_01_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_17_4 = _mm_add_epi32(out_17_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_17_5 = _mm_add_epi32(out_17_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_09_4 = _mm_add_epi32(out_09_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_09_5 = _mm_add_epi32(out_09_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_25_4 = _mm_add_epi32(out_25_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_25_5 = _mm_add_epi32(out_25_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_07_4 = _mm_add_epi32(out_07_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_07_5 = _mm_add_epi32(out_07_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_23_4 = _mm_add_epi32(out_23_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_23_5 = _mm_add_epi32(out_23_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_15_4 = _mm_add_epi32(out_15_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_15_5 = _mm_add_epi32(out_15_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_31_4 = _mm_add_epi32(out_31_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_31_5 = _mm_add_epi32(out_31_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_01_6 = _mm_srai_epi32(out_01_4, DCT_CONST_BITS);
+        const __m128i out_01_7 = _mm_srai_epi32(out_01_5, DCT_CONST_BITS);
+        const __m128i out_17_6 = _mm_srai_epi32(out_17_4, DCT_CONST_BITS);
+        const __m128i out_17_7 = _mm_srai_epi32(out_17_5, DCT_CONST_BITS);
+        const __m128i out_09_6 = _mm_srai_epi32(out_09_4, DCT_CONST_BITS);
+        const __m128i out_09_7 = _mm_srai_epi32(out_09_5, DCT_CONST_BITS);
+        const __m128i out_25_6 = _mm_srai_epi32(out_25_4, DCT_CONST_BITS);
+        const __m128i out_25_7 = _mm_srai_epi32(out_25_5, DCT_CONST_BITS);
+        const __m128i out_07_6 = _mm_srai_epi32(out_07_4, DCT_CONST_BITS);
+        const __m128i out_07_7 = _mm_srai_epi32(out_07_5, DCT_CONST_BITS);
+        const __m128i out_23_6 = _mm_srai_epi32(out_23_4, DCT_CONST_BITS);
+        const __m128i out_23_7 = _mm_srai_epi32(out_23_5, DCT_CONST_BITS);
+        const __m128i out_15_6 = _mm_srai_epi32(out_15_4, DCT_CONST_BITS);
+        const __m128i out_15_7 = _mm_srai_epi32(out_15_5, DCT_CONST_BITS);
+        const __m128i out_31_6 = _mm_srai_epi32(out_31_4, DCT_CONST_BITS);
+        const __m128i out_31_7 = _mm_srai_epi32(out_31_5, DCT_CONST_BITS);
+        // Combine
+        out[ 1] = _mm_packs_epi32(out_01_6, out_01_7);
+        out[17] = _mm_packs_epi32(out_17_6, out_17_7);
+        out[ 9] = _mm_packs_epi32(out_09_6, out_09_7);
+        out[25] = _mm_packs_epi32(out_25_6, out_25_7);
+        out[ 7] = _mm_packs_epi32(out_07_6, out_07_7);
+        out[23] = _mm_packs_epi32(out_23_6, out_23_7);
+        out[15] = _mm_packs_epi32(out_15_6, out_15_7);
+        out[31] = _mm_packs_epi32(out_31_6, out_31_7);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x8(&out[1], &out[17], &out[9],
+                                           &out[25], &out[7], &out[23],
+                                           &out[15], &out[31]);
+        if (overflow) {
+          if (pass == 0)
+            HIGH_FDCT32x32_2D_C(input, output_org, stride);
+          else
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+      {
+        const __m128i out_05_0 = _mm_unpacklo_epi16(step1[20], step1[27]);
+        const __m128i out_05_1 = _mm_unpackhi_epi16(step1[20], step1[27]);
+        const __m128i out_21_0 = _mm_unpacklo_epi16(step1[21], step1[26]);
+        const __m128i out_21_1 = _mm_unpackhi_epi16(step1[21], step1[26]);
+        const __m128i out_13_0 = _mm_unpacklo_epi16(step1[22], step1[25]);
+        const __m128i out_13_1 = _mm_unpackhi_epi16(step1[22], step1[25]);
+        const __m128i out_29_0 = _mm_unpacklo_epi16(step1[23], step1[24]);
+        const __m128i out_29_1 = _mm_unpackhi_epi16(step1[23], step1[24]);
+        const __m128i out_05_2 = _mm_madd_epi16(out_05_0, k__cospi_p27_p05);
+        const __m128i out_05_3 = _mm_madd_epi16(out_05_1, k__cospi_p27_p05);
+        const __m128i out_21_2 = _mm_madd_epi16(out_21_0, k__cospi_p11_p21);
+        const __m128i out_21_3 = _mm_madd_epi16(out_21_1, k__cospi_p11_p21);
+        const __m128i out_13_2 = _mm_madd_epi16(out_13_0, k__cospi_p19_p13);
+        const __m128i out_13_3 = _mm_madd_epi16(out_13_1, k__cospi_p19_p13);
+        const __m128i out_29_2 = _mm_madd_epi16(out_29_0, k__cospi_p03_p29);
+        const __m128i out_29_3 = _mm_madd_epi16(out_29_1, k__cospi_p03_p29);
+        const __m128i out_03_2 = _mm_madd_epi16(out_29_0, k__cospi_m29_p03);
+        const __m128i out_03_3 = _mm_madd_epi16(out_29_1, k__cospi_m29_p03);
+        const __m128i out_19_2 = _mm_madd_epi16(out_13_0, k__cospi_m13_p19);
+        const __m128i out_19_3 = _mm_madd_epi16(out_13_1, k__cospi_m13_p19);
+        const __m128i out_11_2 = _mm_madd_epi16(out_21_0, k__cospi_m21_p11);
+        const __m128i out_11_3 = _mm_madd_epi16(out_21_1, k__cospi_m21_p11);
+        const __m128i out_27_2 = _mm_madd_epi16(out_05_0, k__cospi_m05_p27);
+        const __m128i out_27_3 = _mm_madd_epi16(out_05_1, k__cospi_m05_p27);
+        // dct_const_round_shift
+        const __m128i out_05_4 = _mm_add_epi32(out_05_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_05_5 = _mm_add_epi32(out_05_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_21_4 = _mm_add_epi32(out_21_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_21_5 = _mm_add_epi32(out_21_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_13_4 = _mm_add_epi32(out_13_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_13_5 = _mm_add_epi32(out_13_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_29_4 = _mm_add_epi32(out_29_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_29_5 = _mm_add_epi32(out_29_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_03_4 = _mm_add_epi32(out_03_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_03_5 = _mm_add_epi32(out_03_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_19_4 = _mm_add_epi32(out_19_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_19_5 = _mm_add_epi32(out_19_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_11_4 = _mm_add_epi32(out_11_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_11_5 = _mm_add_epi32(out_11_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_27_4 = _mm_add_epi32(out_27_2, k__DCT_CONST_ROUNDING);
+        const __m128i out_27_5 = _mm_add_epi32(out_27_3, k__DCT_CONST_ROUNDING);
+        const __m128i out_05_6 = _mm_srai_epi32(out_05_4, DCT_CONST_BITS);
+        const __m128i out_05_7 = _mm_srai_epi32(out_05_5, DCT_CONST_BITS);
+        const __m128i out_21_6 = _mm_srai_epi32(out_21_4, DCT_CONST_BITS);
+        const __m128i out_21_7 = _mm_srai_epi32(out_21_5, DCT_CONST_BITS);
+        const __m128i out_13_6 = _mm_srai_epi32(out_13_4, DCT_CONST_BITS);
+        const __m128i out_13_7 = _mm_srai_epi32(out_13_5, DCT_CONST_BITS);
+        const __m128i out_29_6 = _mm_srai_epi32(out_29_4, DCT_CONST_BITS);
+        const __m128i out_29_7 = _mm_srai_epi32(out_29_5, DCT_CONST_BITS);
+        const __m128i out_03_6 = _mm_srai_epi32(out_03_4, DCT_CONST_BITS);
+        const __m128i out_03_7 = _mm_srai_epi32(out_03_5, DCT_CONST_BITS);
+        const __m128i out_19_6 = _mm_srai_epi32(out_19_4, DCT_CONST_BITS);
+        const __m128i out_19_7 = _mm_srai_epi32(out_19_5, DCT_CONST_BITS);
+        const __m128i out_11_6 = _mm_srai_epi32(out_11_4, DCT_CONST_BITS);
+        const __m128i out_11_7 = _mm_srai_epi32(out_11_5, DCT_CONST_BITS);
+        const __m128i out_27_6 = _mm_srai_epi32(out_27_4, DCT_CONST_BITS);
+        const __m128i out_27_7 = _mm_srai_epi32(out_27_5, DCT_CONST_BITS);
+        // Combine
+        out[ 5] = _mm_packs_epi32(out_05_6, out_05_7);
+        out[21] = _mm_packs_epi32(out_21_6, out_21_7);
+        out[13] = _mm_packs_epi32(out_13_6, out_13_7);
+        out[29] = _mm_packs_epi32(out_29_6, out_29_7);
+        out[ 3] = _mm_packs_epi32(out_03_6, out_03_7);
+        out[19] = _mm_packs_epi32(out_19_6, out_19_7);
+        out[11] = _mm_packs_epi32(out_11_6, out_11_7);
+        out[27] = _mm_packs_epi32(out_27_6, out_27_7);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x8(&out[5], &out[21], &out[13],
+                                           &out[29], &out[3], &out[19],
+                                           &out[11], &out[27]);
+        if (overflow) {
+          if (pass == 0)
+            HIGH_FDCT32x32_2D_C(input, output_org, stride);
+          else
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+#if FDCT32x32_HIGH_PRECISION
+      } else {
+        __m128i lstep1[64], lstep2[64], lstep3[64];
+        __m128i u[32], v[32], sign[16];
+        const __m128i K32One = _mm_set_epi32(1, 1, 1, 1);
+        // start using 32-bit operations
+        // stage 3
+        {
+          // expanding to 32-bit length priori to addition operations
+          lstep2[ 0] = _mm_unpacklo_epi16(step2[ 0], kZero);
+          lstep2[ 1] = _mm_unpackhi_epi16(step2[ 0], kZero);
+          lstep2[ 2] = _mm_unpacklo_epi16(step2[ 1], kZero);
+          lstep2[ 3] = _mm_unpackhi_epi16(step2[ 1], kZero);
+          lstep2[ 4] = _mm_unpacklo_epi16(step2[ 2], kZero);
+          lstep2[ 5] = _mm_unpackhi_epi16(step2[ 2], kZero);
+          lstep2[ 6] = _mm_unpacklo_epi16(step2[ 3], kZero);
+          lstep2[ 7] = _mm_unpackhi_epi16(step2[ 3], kZero);
+          lstep2[ 8] = _mm_unpacklo_epi16(step2[ 4], kZero);
+          lstep2[ 9] = _mm_unpackhi_epi16(step2[ 4], kZero);
+          lstep2[10] = _mm_unpacklo_epi16(step2[ 5], kZero);
+          lstep2[11] = _mm_unpackhi_epi16(step2[ 5], kZero);
+          lstep2[12] = _mm_unpacklo_epi16(step2[ 6], kZero);
+          lstep2[13] = _mm_unpackhi_epi16(step2[ 6], kZero);
+          lstep2[14] = _mm_unpacklo_epi16(step2[ 7], kZero);
+          lstep2[15] = _mm_unpackhi_epi16(step2[ 7], kZero);
+          lstep2[ 0] = _mm_madd_epi16(lstep2[ 0], kOne);
+          lstep2[ 1] = _mm_madd_epi16(lstep2[ 1], kOne);
+          lstep2[ 2] = _mm_madd_epi16(lstep2[ 2], kOne);
+          lstep2[ 3] = _mm_madd_epi16(lstep2[ 3], kOne);
+          lstep2[ 4] = _mm_madd_epi16(lstep2[ 4], kOne);
+          lstep2[ 5] = _mm_madd_epi16(lstep2[ 5], kOne);
+          lstep2[ 6] = _mm_madd_epi16(lstep2[ 6], kOne);
+          lstep2[ 7] = _mm_madd_epi16(lstep2[ 7], kOne);
+          lstep2[ 8] = _mm_madd_epi16(lstep2[ 8], kOne);
+          lstep2[ 9] = _mm_madd_epi16(lstep2[ 9], kOne);
+          lstep2[10] = _mm_madd_epi16(lstep2[10], kOne);
+          lstep2[11] = _mm_madd_epi16(lstep2[11], kOne);
+          lstep2[12] = _mm_madd_epi16(lstep2[12], kOne);
+          lstep2[13] = _mm_madd_epi16(lstep2[13], kOne);
+          lstep2[14] = _mm_madd_epi16(lstep2[14], kOne);
+          lstep2[15] = _mm_madd_epi16(lstep2[15], kOne);
+
+          lstep3[ 0] = _mm_add_epi32(lstep2[14], lstep2[ 0]);
+          lstep3[ 1] = _mm_add_epi32(lstep2[15], lstep2[ 1]);
+          lstep3[ 2] = _mm_add_epi32(lstep2[12], lstep2[ 2]);
+          lstep3[ 3] = _mm_add_epi32(lstep2[13], lstep2[ 3]);
+          lstep3[ 4] = _mm_add_epi32(lstep2[10], lstep2[ 4]);
+          lstep3[ 5] = _mm_add_epi32(lstep2[11], lstep2[ 5]);
+          lstep3[ 6] = _mm_add_epi32(lstep2[ 8], lstep2[ 6]);
+          lstep3[ 7] = _mm_add_epi32(lstep2[ 9], lstep2[ 7]);
+          lstep3[ 8] = _mm_sub_epi32(lstep2[ 6], lstep2[ 8]);
+          lstep3[ 9] = _mm_sub_epi32(lstep2[ 7], lstep2[ 9]);
+          lstep3[10] = _mm_sub_epi32(lstep2[ 4], lstep2[10]);
+          lstep3[11] = _mm_sub_epi32(lstep2[ 5], lstep2[11]);
+          lstep3[12] = _mm_sub_epi32(lstep2[ 2], lstep2[12]);
+          lstep3[13] = _mm_sub_epi32(lstep2[ 3], lstep2[13]);
+          lstep3[14] = _mm_sub_epi32(lstep2[ 0], lstep2[14]);
+          lstep3[15] = _mm_sub_epi32(lstep2[ 1], lstep2[15]);
+        }
+        {
+          const __m128i s3_10_0 = _mm_unpacklo_epi16(step2[13], step2[10]);
+          const __m128i s3_10_1 = _mm_unpackhi_epi16(step2[13], step2[10]);
+          const __m128i s3_11_0 = _mm_unpacklo_epi16(step2[12], step2[11]);
+          const __m128i s3_11_1 = _mm_unpackhi_epi16(step2[12], step2[11]);
+          const __m128i s3_10_2 = _mm_madd_epi16(s3_10_0, k__cospi_p16_m16);
+          const __m128i s3_10_3 = _mm_madd_epi16(s3_10_1, k__cospi_p16_m16);
+          const __m128i s3_11_2 = _mm_madd_epi16(s3_11_0, k__cospi_p16_m16);
+          const __m128i s3_11_3 = _mm_madd_epi16(s3_11_1, k__cospi_p16_m16);
+          const __m128i s3_12_2 = _mm_madd_epi16(s3_11_0, k__cospi_p16_p16);
+          const __m128i s3_12_3 = _mm_madd_epi16(s3_11_1, k__cospi_p16_p16);
+          const __m128i s3_13_2 = _mm_madd_epi16(s3_10_0, k__cospi_p16_p16);
+          const __m128i s3_13_3 = _mm_madd_epi16(s3_10_1, k__cospi_p16_p16);
+          // dct_const_round_shift
+          const __m128i s3_10_4 = _mm_add_epi32(s3_10_2, k__DCT_CONST_ROUNDING);
+          const __m128i s3_10_5 = _mm_add_epi32(s3_10_3, k__DCT_CONST_ROUNDING);
+          const __m128i s3_11_4 = _mm_add_epi32(s3_11_2, k__DCT_CONST_ROUNDING);
+          const __m128i s3_11_5 = _mm_add_epi32(s3_11_3, k__DCT_CONST_ROUNDING);
+          const __m128i s3_12_4 = _mm_add_epi32(s3_12_2, k__DCT_CONST_ROUNDING);
+          const __m128i s3_12_5 = _mm_add_epi32(s3_12_3, k__DCT_CONST_ROUNDING);
+          const __m128i s3_13_4 = _mm_add_epi32(s3_13_2, k__DCT_CONST_ROUNDING);
+          const __m128i s3_13_5 = _mm_add_epi32(s3_13_3, k__DCT_CONST_ROUNDING);
+          lstep3[20] = _mm_srai_epi32(s3_10_4, DCT_CONST_BITS);
+          lstep3[21] = _mm_srai_epi32(s3_10_5, DCT_CONST_BITS);
+          lstep3[22] = _mm_srai_epi32(s3_11_4, DCT_CONST_BITS);
+          lstep3[23] = _mm_srai_epi32(s3_11_5, DCT_CONST_BITS);
+          lstep3[24] = _mm_srai_epi32(s3_12_4, DCT_CONST_BITS);
+          lstep3[25] = _mm_srai_epi32(s3_12_5, DCT_CONST_BITS);
+          lstep3[26] = _mm_srai_epi32(s3_13_4, DCT_CONST_BITS);
+          lstep3[27] = _mm_srai_epi32(s3_13_5, DCT_CONST_BITS);
+        }
+        {
+          lstep2[40] = _mm_unpacklo_epi16(step2[20], kZero);
+          lstep2[41] = _mm_unpackhi_epi16(step2[20], kZero);
+          lstep2[42] = _mm_unpacklo_epi16(step2[21], kZero);
+          lstep2[43] = _mm_unpackhi_epi16(step2[21], kZero);
+          lstep2[44] = _mm_unpacklo_epi16(step2[22], kZero);
+          lstep2[45] = _mm_unpackhi_epi16(step2[22], kZero);
+          lstep2[46] = _mm_unpacklo_epi16(step2[23], kZero);
+          lstep2[47] = _mm_unpackhi_epi16(step2[23], kZero);
+          lstep2[48] = _mm_unpacklo_epi16(step2[24], kZero);
+          lstep2[49] = _mm_unpackhi_epi16(step2[24], kZero);
+          lstep2[50] = _mm_unpacklo_epi16(step2[25], kZero);
+          lstep2[51] = _mm_unpackhi_epi16(step2[25], kZero);
+          lstep2[52] = _mm_unpacklo_epi16(step2[26], kZero);
+          lstep2[53] = _mm_unpackhi_epi16(step2[26], kZero);
+          lstep2[54] = _mm_unpacklo_epi16(step2[27], kZero);
+          lstep2[55] = _mm_unpackhi_epi16(step2[27], kZero);
+          lstep2[40] = _mm_madd_epi16(lstep2[40], kOne);
+          lstep2[41] = _mm_madd_epi16(lstep2[41], kOne);
+          lstep2[42] = _mm_madd_epi16(lstep2[42], kOne);
+          lstep2[43] = _mm_madd_epi16(lstep2[43], kOne);
+          lstep2[44] = _mm_madd_epi16(lstep2[44], kOne);
+          lstep2[45] = _mm_madd_epi16(lstep2[45], kOne);
+          lstep2[46] = _mm_madd_epi16(lstep2[46], kOne);
+          lstep2[47] = _mm_madd_epi16(lstep2[47], kOne);
+          lstep2[48] = _mm_madd_epi16(lstep2[48], kOne);
+          lstep2[49] = _mm_madd_epi16(lstep2[49], kOne);
+          lstep2[50] = _mm_madd_epi16(lstep2[50], kOne);
+          lstep2[51] = _mm_madd_epi16(lstep2[51], kOne);
+          lstep2[52] = _mm_madd_epi16(lstep2[52], kOne);
+          lstep2[53] = _mm_madd_epi16(lstep2[53], kOne);
+          lstep2[54] = _mm_madd_epi16(lstep2[54], kOne);
+          lstep2[55] = _mm_madd_epi16(lstep2[55], kOne);
+
+          lstep1[32] = _mm_unpacklo_epi16(step1[16], kZero);
+          lstep1[33] = _mm_unpackhi_epi16(step1[16], kZero);
+          lstep1[34] = _mm_unpacklo_epi16(step1[17], kZero);
+          lstep1[35] = _mm_unpackhi_epi16(step1[17], kZero);
+          lstep1[36] = _mm_unpacklo_epi16(step1[18], kZero);
+          lstep1[37] = _mm_unpackhi_epi16(step1[18], kZero);
+          lstep1[38] = _mm_unpacklo_epi16(step1[19], kZero);
+          lstep1[39] = _mm_unpackhi_epi16(step1[19], kZero);
+          lstep1[56] = _mm_unpacklo_epi16(step1[28], kZero);
+          lstep1[57] = _mm_unpackhi_epi16(step1[28], kZero);
+          lstep1[58] = _mm_unpacklo_epi16(step1[29], kZero);
+          lstep1[59] = _mm_unpackhi_epi16(step1[29], kZero);
+          lstep1[60] = _mm_unpacklo_epi16(step1[30], kZero);
+          lstep1[61] = _mm_unpackhi_epi16(step1[30], kZero);
+          lstep1[62] = _mm_unpacklo_epi16(step1[31], kZero);
+          lstep1[63] = _mm_unpackhi_epi16(step1[31], kZero);
+          lstep1[32] = _mm_madd_epi16(lstep1[32], kOne);
+          lstep1[33] = _mm_madd_epi16(lstep1[33], kOne);
+          lstep1[34] = _mm_madd_epi16(lstep1[34], kOne);
+          lstep1[35] = _mm_madd_epi16(lstep1[35], kOne);
+          lstep1[36] = _mm_madd_epi16(lstep1[36], kOne);
+          lstep1[37] = _mm_madd_epi16(lstep1[37], kOne);
+          lstep1[38] = _mm_madd_epi16(lstep1[38], kOne);
+          lstep1[39] = _mm_madd_epi16(lstep1[39], kOne);
+          lstep1[56] = _mm_madd_epi16(lstep1[56], kOne);
+          lstep1[57] = _mm_madd_epi16(lstep1[57], kOne);
+          lstep1[58] = _mm_madd_epi16(lstep1[58], kOne);
+          lstep1[59] = _mm_madd_epi16(lstep1[59], kOne);
+          lstep1[60] = _mm_madd_epi16(lstep1[60], kOne);
+          lstep1[61] = _mm_madd_epi16(lstep1[61], kOne);
+          lstep1[62] = _mm_madd_epi16(lstep1[62], kOne);
+          lstep1[63] = _mm_madd_epi16(lstep1[63], kOne);
+
+          lstep3[32] = _mm_add_epi32(lstep2[46], lstep1[32]);
+          lstep3[33] = _mm_add_epi32(lstep2[47], lstep1[33]);
+
+          lstep3[34] = _mm_add_epi32(lstep2[44], lstep1[34]);
+          lstep3[35] = _mm_add_epi32(lstep2[45], lstep1[35]);
+          lstep3[36] = _mm_add_epi32(lstep2[42], lstep1[36]);
+          lstep3[37] = _mm_add_epi32(lstep2[43], lstep1[37]);
+          lstep3[38] = _mm_add_epi32(lstep2[40], lstep1[38]);
+          lstep3[39] = _mm_add_epi32(lstep2[41], lstep1[39]);
+          lstep3[40] = _mm_sub_epi32(lstep1[38], lstep2[40]);
+          lstep3[41] = _mm_sub_epi32(lstep1[39], lstep2[41]);
+          lstep3[42] = _mm_sub_epi32(lstep1[36], lstep2[42]);
+          lstep3[43] = _mm_sub_epi32(lstep1[37], lstep2[43]);
+          lstep3[44] = _mm_sub_epi32(lstep1[34], lstep2[44]);
+          lstep3[45] = _mm_sub_epi32(lstep1[35], lstep2[45]);
+          lstep3[46] = _mm_sub_epi32(lstep1[32], lstep2[46]);
+          lstep3[47] = _mm_sub_epi32(lstep1[33], lstep2[47]);
+          lstep3[48] = _mm_sub_epi32(lstep1[62], lstep2[48]);
+          lstep3[49] = _mm_sub_epi32(lstep1[63], lstep2[49]);
+          lstep3[50] = _mm_sub_epi32(lstep1[60], lstep2[50]);
+          lstep3[51] = _mm_sub_epi32(lstep1[61], lstep2[51]);
+          lstep3[52] = _mm_sub_epi32(lstep1[58], lstep2[52]);
+          lstep3[53] = _mm_sub_epi32(lstep1[59], lstep2[53]);
+          lstep3[54] = _mm_sub_epi32(lstep1[56], lstep2[54]);
+          lstep3[55] = _mm_sub_epi32(lstep1[57], lstep2[55]);
+          lstep3[56] = _mm_add_epi32(lstep2[54], lstep1[56]);
+          lstep3[57] = _mm_add_epi32(lstep2[55], lstep1[57]);
+          lstep3[58] = _mm_add_epi32(lstep2[52], lstep1[58]);
+          lstep3[59] = _mm_add_epi32(lstep2[53], lstep1[59]);
+          lstep3[60] = _mm_add_epi32(lstep2[50], lstep1[60]);
+          lstep3[61] = _mm_add_epi32(lstep2[51], lstep1[61]);
+          lstep3[62] = _mm_add_epi32(lstep2[48], lstep1[62]);
+          lstep3[63] = _mm_add_epi32(lstep2[49], lstep1[63]);
+        }
+
+        // stage 4
+        {
+          // expanding to 32-bit length priori to addition operations
+          lstep2[16] = _mm_unpacklo_epi16(step2[ 8], kZero);
+          lstep2[17] = _mm_unpackhi_epi16(step2[ 8], kZero);
+          lstep2[18] = _mm_unpacklo_epi16(step2[ 9], kZero);
+          lstep2[19] = _mm_unpackhi_epi16(step2[ 9], kZero);
+          lstep2[28] = _mm_unpacklo_epi16(step2[14], kZero);
+          lstep2[29] = _mm_unpackhi_epi16(step2[14], kZero);
+          lstep2[30] = _mm_unpacklo_epi16(step2[15], kZero);
+          lstep2[31] = _mm_unpackhi_epi16(step2[15], kZero);
+          lstep2[16] = _mm_madd_epi16(lstep2[16], kOne);
+          lstep2[17] = _mm_madd_epi16(lstep2[17], kOne);
+          lstep2[18] = _mm_madd_epi16(lstep2[18], kOne);
+          lstep2[19] = _mm_madd_epi16(lstep2[19], kOne);
+          lstep2[28] = _mm_madd_epi16(lstep2[28], kOne);
+          lstep2[29] = _mm_madd_epi16(lstep2[29], kOne);
+          lstep2[30] = _mm_madd_epi16(lstep2[30], kOne);
+          lstep2[31] = _mm_madd_epi16(lstep2[31], kOne);
+
+          lstep1[ 0] = _mm_add_epi32(lstep3[ 6], lstep3[ 0]);
+          lstep1[ 1] = _mm_add_epi32(lstep3[ 7], lstep3[ 1]);
+          lstep1[ 2] = _mm_add_epi32(lstep3[ 4], lstep3[ 2]);
+          lstep1[ 3] = _mm_add_epi32(lstep3[ 5], lstep3[ 3]);
+          lstep1[ 4] = _mm_sub_epi32(lstep3[ 2], lstep3[ 4]);
+          lstep1[ 5] = _mm_sub_epi32(lstep3[ 3], lstep3[ 5]);
+          lstep1[ 6] = _mm_sub_epi32(lstep3[ 0], lstep3[ 6]);
+          lstep1[ 7] = _mm_sub_epi32(lstep3[ 1], lstep3[ 7]);
+          lstep1[16] = _mm_add_epi32(lstep3[22], lstep2[16]);
+          lstep1[17] = _mm_add_epi32(lstep3[23], lstep2[17]);
+          lstep1[18] = _mm_add_epi32(lstep3[20], lstep2[18]);
+          lstep1[19] = _mm_add_epi32(lstep3[21], lstep2[19]);
+          lstep1[20] = _mm_sub_epi32(lstep2[18], lstep3[20]);
+          lstep1[21] = _mm_sub_epi32(lstep2[19], lstep3[21]);
+          lstep1[22] = _mm_sub_epi32(lstep2[16], lstep3[22]);
+          lstep1[23] = _mm_sub_epi32(lstep2[17], lstep3[23]);
+          lstep1[24] = _mm_sub_epi32(lstep2[30], lstep3[24]);
+          lstep1[25] = _mm_sub_epi32(lstep2[31], lstep3[25]);
+          lstep1[26] = _mm_sub_epi32(lstep2[28], lstep3[26]);
+          lstep1[27] = _mm_sub_epi32(lstep2[29], lstep3[27]);
+          lstep1[28] = _mm_add_epi32(lstep3[26], lstep2[28]);
+          lstep1[29] = _mm_add_epi32(lstep3[27], lstep2[29]);
+          lstep1[30] = _mm_add_epi32(lstep3[24], lstep2[30]);
+          lstep1[31] = _mm_add_epi32(lstep3[25], lstep2[31]);
+        }
+        {
+        // to be continued...
+        //
+        const __m128i k32_p16_p16 = pair_set_epi32(cospi_16_64, cospi_16_64);
+        const __m128i k32_p16_m16 = pair_set_epi32(cospi_16_64, -cospi_16_64);
+
+        u[0] = _mm_unpacklo_epi32(lstep3[12], lstep3[10]);
+        u[1] = _mm_unpackhi_epi32(lstep3[12], lstep3[10]);
+        u[2] = _mm_unpacklo_epi32(lstep3[13], lstep3[11]);
+        u[3] = _mm_unpackhi_epi32(lstep3[13], lstep3[11]);
+
+        // TODO(jingning): manually inline k_madd_epi32_ to further hide
+        // instruction latency.
+        v[0] = k_madd_epi32(u[0], k32_p16_m16);
+        v[1] = k_madd_epi32(u[1], k32_p16_m16);
+        v[2] = k_madd_epi32(u[2], k32_p16_m16);
+        v[3] = k_madd_epi32(u[3], k32_p16_m16);
+        v[4] = k_madd_epi32(u[0], k32_p16_p16);
+        v[5] = k_madd_epi32(u[1], k32_p16_p16);
+        v[6] = k_madd_epi32(u[2], k32_p16_p16);
+        v[7] = k_madd_epi32(u[3], k32_p16_p16);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = k_check_epi32_overflow_8(&v[0], &v[1], &v[2], &v[3],
+                                            &v[4], &v[5], &v[6], &v[7], &kZero);
+        if (overflow) {
+          HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+        u[0] = k_packs_epi64(v[0], v[1]);
+        u[1] = k_packs_epi64(v[2], v[3]);
+        u[2] = k_packs_epi64(v[4], v[5]);
+        u[3] = k_packs_epi64(v[6], v[7]);
+
+        v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+        v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+        v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+        v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+
+        lstep1[10] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
+        lstep1[11] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
+        lstep1[12] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
+        lstep1[13] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
+        }
+        {
+          const __m128i k32_m08_p24 = pair_set_epi32(-cospi_8_64, cospi_24_64);
+          const __m128i k32_m24_m08 = pair_set_epi32(-cospi_24_64, -cospi_8_64);
+          const __m128i k32_p24_p08 = pair_set_epi32(cospi_24_64, cospi_8_64);
+
+          u[ 0] = _mm_unpacklo_epi32(lstep3[36], lstep3[58]);
+          u[ 1] = _mm_unpackhi_epi32(lstep3[36], lstep3[58]);
+          u[ 2] = _mm_unpacklo_epi32(lstep3[37], lstep3[59]);
+          u[ 3] = _mm_unpackhi_epi32(lstep3[37], lstep3[59]);
+          u[ 4] = _mm_unpacklo_epi32(lstep3[38], lstep3[56]);
+          u[ 5] = _mm_unpackhi_epi32(lstep3[38], lstep3[56]);
+          u[ 6] = _mm_unpacklo_epi32(lstep3[39], lstep3[57]);
+          u[ 7] = _mm_unpackhi_epi32(lstep3[39], lstep3[57]);
+          u[ 8] = _mm_unpacklo_epi32(lstep3[40], lstep3[54]);
+          u[ 9] = _mm_unpackhi_epi32(lstep3[40], lstep3[54]);
+          u[10] = _mm_unpacklo_epi32(lstep3[41], lstep3[55]);
+          u[11] = _mm_unpackhi_epi32(lstep3[41], lstep3[55]);
+          u[12] = _mm_unpacklo_epi32(lstep3[42], lstep3[52]);
+          u[13] = _mm_unpackhi_epi32(lstep3[42], lstep3[52]);
+          u[14] = _mm_unpacklo_epi32(lstep3[43], lstep3[53]);
+          u[15] = _mm_unpackhi_epi32(lstep3[43], lstep3[53]);
+
+          v[ 0] = k_madd_epi32(u[ 0], k32_m08_p24);
+          v[ 1] = k_madd_epi32(u[ 1], k32_m08_p24);
+          v[ 2] = k_madd_epi32(u[ 2], k32_m08_p24);
+          v[ 3] = k_madd_epi32(u[ 3], k32_m08_p24);
+          v[ 4] = k_madd_epi32(u[ 4], k32_m08_p24);
+          v[ 5] = k_madd_epi32(u[ 5], k32_m08_p24);
+          v[ 6] = k_madd_epi32(u[ 6], k32_m08_p24);
+          v[ 7] = k_madd_epi32(u[ 7], k32_m08_p24);
+          v[ 8] = k_madd_epi32(u[ 8], k32_m24_m08);
+          v[ 9] = k_madd_epi32(u[ 9], k32_m24_m08);
+          v[10] = k_madd_epi32(u[10], k32_m24_m08);
+          v[11] = k_madd_epi32(u[11], k32_m24_m08);
+          v[12] = k_madd_epi32(u[12], k32_m24_m08);
+          v[13] = k_madd_epi32(u[13], k32_m24_m08);
+          v[14] = k_madd_epi32(u[14], k32_m24_m08);
+          v[15] = k_madd_epi32(u[15], k32_m24_m08);
+          v[16] = k_madd_epi32(u[12], k32_m08_p24);
+          v[17] = k_madd_epi32(u[13], k32_m08_p24);
+          v[18] = k_madd_epi32(u[14], k32_m08_p24);
+          v[19] = k_madd_epi32(u[15], k32_m08_p24);
+          v[20] = k_madd_epi32(u[ 8], k32_m08_p24);
+          v[21] = k_madd_epi32(u[ 9], k32_m08_p24);
+          v[22] = k_madd_epi32(u[10], k32_m08_p24);
+          v[23] = k_madd_epi32(u[11], k32_m08_p24);
+          v[24] = k_madd_epi32(u[ 4], k32_p24_p08);
+          v[25] = k_madd_epi32(u[ 5], k32_p24_p08);
+          v[26] = k_madd_epi32(u[ 6], k32_p24_p08);
+          v[27] = k_madd_epi32(u[ 7], k32_p24_p08);
+          v[28] = k_madd_epi32(u[ 0], k32_p24_p08);
+          v[29] = k_madd_epi32(u[ 1], k32_p24_p08);
+          v[30] = k_madd_epi32(u[ 2], k32_p24_p08);
+          v[31] = k_madd_epi32(u[ 3], k32_p24_p08);
+
+#if DCT_HIGH_BIT_DEPTH
+          overflow = k_check_epi32_overflow_32(
+              &v[0], &v[1], &v[2], &v[3], &v[4], &v[5], &v[6], &v[7],
+              &v[8], &v[9], &v[10], &v[11], &v[12], &v[13], &v[14], &v[15],
+              &v[16], &v[17], &v[18], &v[19], &v[20], &v[21], &v[22], &v[23],
+              &v[24], &v[25], &v[26], &v[27], &v[28], &v[29], &v[30], &v[31],
+              &kZero);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+          u[ 0] = k_packs_epi64(v[ 0], v[ 1]);
+          u[ 1] = k_packs_epi64(v[ 2], v[ 3]);
+          u[ 2] = k_packs_epi64(v[ 4], v[ 5]);
+          u[ 3] = k_packs_epi64(v[ 6], v[ 7]);
+          u[ 4] = k_packs_epi64(v[ 8], v[ 9]);
+          u[ 5] = k_packs_epi64(v[10], v[11]);
+          u[ 6] = k_packs_epi64(v[12], v[13]);
+          u[ 7] = k_packs_epi64(v[14], v[15]);
+          u[ 8] = k_packs_epi64(v[16], v[17]);
+          u[ 9] = k_packs_epi64(v[18], v[19]);
+          u[10] = k_packs_epi64(v[20], v[21]);
+          u[11] = k_packs_epi64(v[22], v[23]);
+          u[12] = k_packs_epi64(v[24], v[25]);
+          u[13] = k_packs_epi64(v[26], v[27]);
+          u[14] = k_packs_epi64(v[28], v[29]);
+          u[15] = k_packs_epi64(v[30], v[31]);
+
+          v[ 0] = _mm_add_epi32(u[ 0], k__DCT_CONST_ROUNDING);
+          v[ 1] = _mm_add_epi32(u[ 1], k__DCT_CONST_ROUNDING);
+          v[ 2] = _mm_add_epi32(u[ 2], k__DCT_CONST_ROUNDING);
+          v[ 3] = _mm_add_epi32(u[ 3], k__DCT_CONST_ROUNDING);
+          v[ 4] = _mm_add_epi32(u[ 4], k__DCT_CONST_ROUNDING);
+          v[ 5] = _mm_add_epi32(u[ 5], k__DCT_CONST_ROUNDING);
+          v[ 6] = _mm_add_epi32(u[ 6], k__DCT_CONST_ROUNDING);
+          v[ 7] = _mm_add_epi32(u[ 7], k__DCT_CONST_ROUNDING);
+          v[ 8] = _mm_add_epi32(u[ 8], k__DCT_CONST_ROUNDING);
+          v[ 9] = _mm_add_epi32(u[ 9], k__DCT_CONST_ROUNDING);
+          v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+          v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+          v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+          v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+          v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+          v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+          lstep1[36] = _mm_srai_epi32(v[ 0], DCT_CONST_BITS);
+          lstep1[37] = _mm_srai_epi32(v[ 1], DCT_CONST_BITS);
+          lstep1[38] = _mm_srai_epi32(v[ 2], DCT_CONST_BITS);
+          lstep1[39] = _mm_srai_epi32(v[ 3], DCT_CONST_BITS);
+          lstep1[40] = _mm_srai_epi32(v[ 4], DCT_CONST_BITS);
+          lstep1[41] = _mm_srai_epi32(v[ 5], DCT_CONST_BITS);
+          lstep1[42] = _mm_srai_epi32(v[ 6], DCT_CONST_BITS);
+          lstep1[43] = _mm_srai_epi32(v[ 7], DCT_CONST_BITS);
+          lstep1[52] = _mm_srai_epi32(v[ 8], DCT_CONST_BITS);
+          lstep1[53] = _mm_srai_epi32(v[ 9], DCT_CONST_BITS);
+          lstep1[54] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
+          lstep1[55] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
+          lstep1[56] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
+          lstep1[57] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
+          lstep1[58] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
+          lstep1[59] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
+        }
+        // stage 5
+        {
+          lstep2[ 8] = _mm_add_epi32(lstep1[10], lstep3[ 8]);
+          lstep2[ 9] = _mm_add_epi32(lstep1[11], lstep3[ 9]);
+          lstep2[10] = _mm_sub_epi32(lstep3[ 8], lstep1[10]);
+          lstep2[11] = _mm_sub_epi32(lstep3[ 9], lstep1[11]);
+          lstep2[12] = _mm_sub_epi32(lstep3[14], lstep1[12]);
+          lstep2[13] = _mm_sub_epi32(lstep3[15], lstep1[13]);
+          lstep2[14] = _mm_add_epi32(lstep1[12], lstep3[14]);
+          lstep2[15] = _mm_add_epi32(lstep1[13], lstep3[15]);
+        }
+        {
+          const __m128i k32_p16_p16 = pair_set_epi32(cospi_16_64, cospi_16_64);
+          const __m128i k32_p16_m16 = pair_set_epi32(cospi_16_64, -cospi_16_64);
+          const __m128i k32_p24_p08 = pair_set_epi32(cospi_24_64, cospi_8_64);
+          const __m128i k32_m08_p24 = pair_set_epi32(-cospi_8_64, cospi_24_64);
+
+          u[0] = _mm_unpacklo_epi32(lstep1[0], lstep1[2]);
+          u[1] = _mm_unpackhi_epi32(lstep1[0], lstep1[2]);
+          u[2] = _mm_unpacklo_epi32(lstep1[1], lstep1[3]);
+          u[3] = _mm_unpackhi_epi32(lstep1[1], lstep1[3]);
+          u[4] = _mm_unpacklo_epi32(lstep1[4], lstep1[6]);
+          u[5] = _mm_unpackhi_epi32(lstep1[4], lstep1[6]);
+          u[6] = _mm_unpacklo_epi32(lstep1[5], lstep1[7]);
+          u[7] = _mm_unpackhi_epi32(lstep1[5], lstep1[7]);
+
+          // TODO(jingning): manually inline k_madd_epi32_ to further hide
+          // instruction latency.
+          v[ 0] = k_madd_epi32(u[0], k32_p16_p16);
+          v[ 1] = k_madd_epi32(u[1], k32_p16_p16);
+          v[ 2] = k_madd_epi32(u[2], k32_p16_p16);
+          v[ 3] = k_madd_epi32(u[3], k32_p16_p16);
+          v[ 4] = k_madd_epi32(u[0], k32_p16_m16);
+          v[ 5] = k_madd_epi32(u[1], k32_p16_m16);
+          v[ 6] = k_madd_epi32(u[2], k32_p16_m16);
+          v[ 7] = k_madd_epi32(u[3], k32_p16_m16);
+          v[ 8] = k_madd_epi32(u[4], k32_p24_p08);
+          v[ 9] = k_madd_epi32(u[5], k32_p24_p08);
+          v[10] = k_madd_epi32(u[6], k32_p24_p08);
+          v[11] = k_madd_epi32(u[7], k32_p24_p08);
+          v[12] = k_madd_epi32(u[4], k32_m08_p24);
+          v[13] = k_madd_epi32(u[5], k32_m08_p24);
+          v[14] = k_madd_epi32(u[6], k32_m08_p24);
+          v[15] = k_madd_epi32(u[7], k32_m08_p24);
+
+#if DCT_HIGH_BIT_DEPTH
+          overflow = k_check_epi32_overflow_16(
+              &v[0], &v[1], &v[2], &v[3], &v[4], &v[5], &v[6], &v[7],
+              &v[8], &v[9], &v[10], &v[11], &v[12], &v[13], &v[14], &v[15],
+              &kZero);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+          u[0] = k_packs_epi64(v[0], v[1]);
+          u[1] = k_packs_epi64(v[2], v[3]);
+          u[2] = k_packs_epi64(v[4], v[5]);
+          u[3] = k_packs_epi64(v[6], v[7]);
+          u[4] = k_packs_epi64(v[8], v[9]);
+          u[5] = k_packs_epi64(v[10], v[11]);
+          u[6] = k_packs_epi64(v[12], v[13]);
+          u[7] = k_packs_epi64(v[14], v[15]);
+
+          v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+          v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+          v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+          v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+          v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+          v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+          v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+          v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+
+          u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
+          u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
+          u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
+          u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
+          u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
+          u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
+          u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
+          u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
+
+          sign[0] = _mm_cmplt_epi32(u[0], kZero);
+          sign[1] = _mm_cmplt_epi32(u[1], kZero);
+          sign[2] = _mm_cmplt_epi32(u[2], kZero);
+          sign[3] = _mm_cmplt_epi32(u[3], kZero);
+          sign[4] = _mm_cmplt_epi32(u[4], kZero);
+          sign[5] = _mm_cmplt_epi32(u[5], kZero);
+          sign[6] = _mm_cmplt_epi32(u[6], kZero);
+          sign[7] = _mm_cmplt_epi32(u[7], kZero);
+
+          u[0] = _mm_sub_epi32(u[0], sign[0]);
+          u[1] = _mm_sub_epi32(u[1], sign[1]);
+          u[2] = _mm_sub_epi32(u[2], sign[2]);
+          u[3] = _mm_sub_epi32(u[3], sign[3]);
+          u[4] = _mm_sub_epi32(u[4], sign[4]);
+          u[5] = _mm_sub_epi32(u[5], sign[5]);
+          u[6] = _mm_sub_epi32(u[6], sign[6]);
+          u[7] = _mm_sub_epi32(u[7], sign[7]);
+
+          u[0] = _mm_add_epi32(u[0], K32One);
+          u[1] = _mm_add_epi32(u[1], K32One);
+          u[2] = _mm_add_epi32(u[2], K32One);
+          u[3] = _mm_add_epi32(u[3], K32One);
+          u[4] = _mm_add_epi32(u[4], K32One);
+          u[5] = _mm_add_epi32(u[5], K32One);
+          u[6] = _mm_add_epi32(u[6], K32One);
+          u[7] = _mm_add_epi32(u[7], K32One);
+
+          u[0] = _mm_srai_epi32(u[0], 2);
+          u[1] = _mm_srai_epi32(u[1], 2);
+          u[2] = _mm_srai_epi32(u[2], 2);
+          u[3] = _mm_srai_epi32(u[3], 2);
+          u[4] = _mm_srai_epi32(u[4], 2);
+          u[5] = _mm_srai_epi32(u[5], 2);
+          u[6] = _mm_srai_epi32(u[6], 2);
+          u[7] = _mm_srai_epi32(u[7], 2);
+
+          // Combine
+          out[ 0] = _mm_packs_epi32(u[0], u[1]);
+          out[16] = _mm_packs_epi32(u[2], u[3]);
+          out[ 8] = _mm_packs_epi32(u[4], u[5]);
+          out[24] = _mm_packs_epi32(u[6], u[7]);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x4(&out[0], &out[16],
+                                             &out[8], &out[24]);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        {
+          const __m128i k32_m08_p24 = pair_set_epi32(-cospi_8_64, cospi_24_64);
+          const __m128i k32_m24_m08 = pair_set_epi32(-cospi_24_64, -cospi_8_64);
+          const __m128i k32_p24_p08 = pair_set_epi32(cospi_24_64, cospi_8_64);
+
+          u[0] = _mm_unpacklo_epi32(lstep1[18], lstep1[28]);
+          u[1] = _mm_unpackhi_epi32(lstep1[18], lstep1[28]);
+          u[2] = _mm_unpacklo_epi32(lstep1[19], lstep1[29]);
+          u[3] = _mm_unpackhi_epi32(lstep1[19], lstep1[29]);
+          u[4] = _mm_unpacklo_epi32(lstep1[20], lstep1[26]);
+          u[5] = _mm_unpackhi_epi32(lstep1[20], lstep1[26]);
+          u[6] = _mm_unpacklo_epi32(lstep1[21], lstep1[27]);
+          u[7] = _mm_unpackhi_epi32(lstep1[21], lstep1[27]);
+
+          v[0] = k_madd_epi32(u[0], k32_m08_p24);
+          v[1] = k_madd_epi32(u[1], k32_m08_p24);
+          v[2] = k_madd_epi32(u[2], k32_m08_p24);
+          v[3] = k_madd_epi32(u[3], k32_m08_p24);
+          v[4] = k_madd_epi32(u[4], k32_m24_m08);
+          v[5] = k_madd_epi32(u[5], k32_m24_m08);
+          v[6] = k_madd_epi32(u[6], k32_m24_m08);
+          v[7] = k_madd_epi32(u[7], k32_m24_m08);
+          v[ 8] = k_madd_epi32(u[4], k32_m08_p24);
+          v[ 9] = k_madd_epi32(u[5], k32_m08_p24);
+          v[10] = k_madd_epi32(u[6], k32_m08_p24);
+          v[11] = k_madd_epi32(u[7], k32_m08_p24);
+          v[12] = k_madd_epi32(u[0], k32_p24_p08);
+          v[13] = k_madd_epi32(u[1], k32_p24_p08);
+          v[14] = k_madd_epi32(u[2], k32_p24_p08);
+          v[15] = k_madd_epi32(u[3], k32_p24_p08);
+
+#if DCT_HIGH_BIT_DEPTH
+          overflow = k_check_epi32_overflow_16(
+              &v[0], &v[1], &v[2], &v[3], &v[4], &v[5], &v[6], &v[7],
+              &v[8], &v[9], &v[10], &v[11], &v[12], &v[13], &v[14], &v[15],
+              &kZero);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+          u[0] = k_packs_epi64(v[0], v[1]);
+          u[1] = k_packs_epi64(v[2], v[3]);
+          u[2] = k_packs_epi64(v[4], v[5]);
+          u[3] = k_packs_epi64(v[6], v[7]);
+          u[4] = k_packs_epi64(v[8], v[9]);
+          u[5] = k_packs_epi64(v[10], v[11]);
+          u[6] = k_packs_epi64(v[12], v[13]);
+          u[7] = k_packs_epi64(v[14], v[15]);
+
+          u[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+          u[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+          u[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+          u[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+          u[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+          u[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+          u[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+          u[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+
+          lstep2[18] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+          lstep2[19] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+          lstep2[20] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+          lstep2[21] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+          lstep2[26] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+          lstep2[27] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+          lstep2[28] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+          lstep2[29] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+        }
+        {
+          lstep2[32] = _mm_add_epi32(lstep1[38], lstep3[32]);
+          lstep2[33] = _mm_add_epi32(lstep1[39], lstep3[33]);
+          lstep2[34] = _mm_add_epi32(lstep1[36], lstep3[34]);
+          lstep2[35] = _mm_add_epi32(lstep1[37], lstep3[35]);
+          lstep2[36] = _mm_sub_epi32(lstep3[34], lstep1[36]);
+          lstep2[37] = _mm_sub_epi32(lstep3[35], lstep1[37]);
+          lstep2[38] = _mm_sub_epi32(lstep3[32], lstep1[38]);
+          lstep2[39] = _mm_sub_epi32(lstep3[33], lstep1[39]);
+          lstep2[40] = _mm_sub_epi32(lstep3[46], lstep1[40]);
+          lstep2[41] = _mm_sub_epi32(lstep3[47], lstep1[41]);
+          lstep2[42] = _mm_sub_epi32(lstep3[44], lstep1[42]);
+          lstep2[43] = _mm_sub_epi32(lstep3[45], lstep1[43]);
+          lstep2[44] = _mm_add_epi32(lstep1[42], lstep3[44]);
+          lstep2[45] = _mm_add_epi32(lstep1[43], lstep3[45]);
+          lstep2[46] = _mm_add_epi32(lstep1[40], lstep3[46]);
+          lstep2[47] = _mm_add_epi32(lstep1[41], lstep3[47]);
+          lstep2[48] = _mm_add_epi32(lstep1[54], lstep3[48]);
+          lstep2[49] = _mm_add_epi32(lstep1[55], lstep3[49]);
+          lstep2[50] = _mm_add_epi32(lstep1[52], lstep3[50]);
+          lstep2[51] = _mm_add_epi32(lstep1[53], lstep3[51]);
+          lstep2[52] = _mm_sub_epi32(lstep3[50], lstep1[52]);
+          lstep2[53] = _mm_sub_epi32(lstep3[51], lstep1[53]);
+          lstep2[54] = _mm_sub_epi32(lstep3[48], lstep1[54]);
+          lstep2[55] = _mm_sub_epi32(lstep3[49], lstep1[55]);
+          lstep2[56] = _mm_sub_epi32(lstep3[62], lstep1[56]);
+          lstep2[57] = _mm_sub_epi32(lstep3[63], lstep1[57]);
+          lstep2[58] = _mm_sub_epi32(lstep3[60], lstep1[58]);
+          lstep2[59] = _mm_sub_epi32(lstep3[61], lstep1[59]);
+          lstep2[60] = _mm_add_epi32(lstep1[58], lstep3[60]);
+          lstep2[61] = _mm_add_epi32(lstep1[59], lstep3[61]);
+          lstep2[62] = _mm_add_epi32(lstep1[56], lstep3[62]);
+          lstep2[63] = _mm_add_epi32(lstep1[57], lstep3[63]);
+        }
+        // stage 6
+        {
+          const __m128i k32_p28_p04 = pair_set_epi32(cospi_28_64, cospi_4_64);
+          const __m128i k32_p12_p20 = pair_set_epi32(cospi_12_64, cospi_20_64);
+          const __m128i k32_m20_p12 = pair_set_epi32(-cospi_20_64, cospi_12_64);
+          const __m128i k32_m04_p28 = pair_set_epi32(-cospi_4_64, cospi_28_64);
+
+          u[0] = _mm_unpacklo_epi32(lstep2[ 8], lstep2[14]);
+          u[1] = _mm_unpackhi_epi32(lstep2[ 8], lstep2[14]);
+          u[2] = _mm_unpacklo_epi32(lstep2[ 9], lstep2[15]);
+          u[3] = _mm_unpackhi_epi32(lstep2[ 9], lstep2[15]);
+          u[4] = _mm_unpacklo_epi32(lstep2[10], lstep2[12]);
+          u[5] = _mm_unpackhi_epi32(lstep2[10], lstep2[12]);
+          u[6] = _mm_unpacklo_epi32(lstep2[11], lstep2[13]);
+          u[7] = _mm_unpackhi_epi32(lstep2[11], lstep2[13]);
+          u[8] = _mm_unpacklo_epi32(lstep2[10], lstep2[12]);
+          u[9] = _mm_unpackhi_epi32(lstep2[10], lstep2[12]);
+          u[10] = _mm_unpacklo_epi32(lstep2[11], lstep2[13]);
+          u[11] = _mm_unpackhi_epi32(lstep2[11], lstep2[13]);
+          u[12] = _mm_unpacklo_epi32(lstep2[ 8], lstep2[14]);
+          u[13] = _mm_unpackhi_epi32(lstep2[ 8], lstep2[14]);
+          u[14] = _mm_unpacklo_epi32(lstep2[ 9], lstep2[15]);
+          u[15] = _mm_unpackhi_epi32(lstep2[ 9], lstep2[15]);
+
+          v[0] = k_madd_epi32(u[0], k32_p28_p04);
+          v[1] = k_madd_epi32(u[1], k32_p28_p04);
+          v[2] = k_madd_epi32(u[2], k32_p28_p04);
+          v[3] = k_madd_epi32(u[3], k32_p28_p04);
+          v[4] = k_madd_epi32(u[4], k32_p12_p20);
+          v[5] = k_madd_epi32(u[5], k32_p12_p20);
+          v[6] = k_madd_epi32(u[6], k32_p12_p20);
+          v[7] = k_madd_epi32(u[7], k32_p12_p20);
+          v[ 8] = k_madd_epi32(u[ 8], k32_m20_p12);
+          v[ 9] = k_madd_epi32(u[ 9], k32_m20_p12);
+          v[10] = k_madd_epi32(u[10], k32_m20_p12);
+          v[11] = k_madd_epi32(u[11], k32_m20_p12);
+          v[12] = k_madd_epi32(u[12], k32_m04_p28);
+          v[13] = k_madd_epi32(u[13], k32_m04_p28);
+          v[14] = k_madd_epi32(u[14], k32_m04_p28);
+          v[15] = k_madd_epi32(u[15], k32_m04_p28);
+
+#if DCT_HIGH_BIT_DEPTH
+          overflow = k_check_epi32_overflow_16(
+              &v[0], &v[1], &v[2], &v[3], &v[4], &v[5], &v[6], &v[7],
+              &v[8], &v[9], &v[10], &v[11], &v[12], &v[13], &v[14], &v[15],
+              &kZero);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+          u[0] = k_packs_epi64(v[0], v[1]);
+          u[1] = k_packs_epi64(v[2], v[3]);
+          u[2] = k_packs_epi64(v[4], v[5]);
+          u[3] = k_packs_epi64(v[6], v[7]);
+          u[4] = k_packs_epi64(v[8], v[9]);
+          u[5] = k_packs_epi64(v[10], v[11]);
+          u[6] = k_packs_epi64(v[12], v[13]);
+          u[7] = k_packs_epi64(v[14], v[15]);
+
+          v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+          v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+          v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+          v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+          v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+          v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+          v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+          v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+
+          u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
+          u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
+          u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
+          u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
+          u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
+          u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
+          u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
+          u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
+
+          sign[0] = _mm_cmplt_epi32(u[0], kZero);
+          sign[1] = _mm_cmplt_epi32(u[1], kZero);
+          sign[2] = _mm_cmplt_epi32(u[2], kZero);
+          sign[3] = _mm_cmplt_epi32(u[3], kZero);
+          sign[4] = _mm_cmplt_epi32(u[4], kZero);
+          sign[5] = _mm_cmplt_epi32(u[5], kZero);
+          sign[6] = _mm_cmplt_epi32(u[6], kZero);
+          sign[7] = _mm_cmplt_epi32(u[7], kZero);
+
+          u[0] = _mm_sub_epi32(u[0], sign[0]);
+          u[1] = _mm_sub_epi32(u[1], sign[1]);
+          u[2] = _mm_sub_epi32(u[2], sign[2]);
+          u[3] = _mm_sub_epi32(u[3], sign[3]);
+          u[4] = _mm_sub_epi32(u[4], sign[4]);
+          u[5] = _mm_sub_epi32(u[5], sign[5]);
+          u[6] = _mm_sub_epi32(u[6], sign[6]);
+          u[7] = _mm_sub_epi32(u[7], sign[7]);
+
+          u[0] = _mm_add_epi32(u[0], K32One);
+          u[1] = _mm_add_epi32(u[1], K32One);
+          u[2] = _mm_add_epi32(u[2], K32One);
+          u[3] = _mm_add_epi32(u[3], K32One);
+          u[4] = _mm_add_epi32(u[4], K32One);
+          u[5] = _mm_add_epi32(u[5], K32One);
+          u[6] = _mm_add_epi32(u[6], K32One);
+          u[7] = _mm_add_epi32(u[7], K32One);
+
+          u[0] = _mm_srai_epi32(u[0], 2);
+          u[1] = _mm_srai_epi32(u[1], 2);
+          u[2] = _mm_srai_epi32(u[2], 2);
+          u[3] = _mm_srai_epi32(u[3], 2);
+          u[4] = _mm_srai_epi32(u[4], 2);
+          u[5] = _mm_srai_epi32(u[5], 2);
+          u[6] = _mm_srai_epi32(u[6], 2);
+          u[7] = _mm_srai_epi32(u[7], 2);
+
+          out[ 4] = _mm_packs_epi32(u[0], u[1]);
+          out[20] = _mm_packs_epi32(u[2], u[3]);
+          out[12] = _mm_packs_epi32(u[4], u[5]);
+          out[28] = _mm_packs_epi32(u[6], u[7]);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x4(&out[4], &out[20],
+                                             &out[12], &out[28]);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        {
+          lstep3[16] = _mm_add_epi32(lstep2[18], lstep1[16]);
+          lstep3[17] = _mm_add_epi32(lstep2[19], lstep1[17]);
+          lstep3[18] = _mm_sub_epi32(lstep1[16], lstep2[18]);
+          lstep3[19] = _mm_sub_epi32(lstep1[17], lstep2[19]);
+          lstep3[20] = _mm_sub_epi32(lstep1[22], lstep2[20]);
+          lstep3[21] = _mm_sub_epi32(lstep1[23], lstep2[21]);
+          lstep3[22] = _mm_add_epi32(lstep2[20], lstep1[22]);
+          lstep3[23] = _mm_add_epi32(lstep2[21], lstep1[23]);
+          lstep3[24] = _mm_add_epi32(lstep2[26], lstep1[24]);
+          lstep3[25] = _mm_add_epi32(lstep2[27], lstep1[25]);
+          lstep3[26] = _mm_sub_epi32(lstep1[24], lstep2[26]);
+          lstep3[27] = _mm_sub_epi32(lstep1[25], lstep2[27]);
+          lstep3[28] = _mm_sub_epi32(lstep1[30], lstep2[28]);
+          lstep3[29] = _mm_sub_epi32(lstep1[31], lstep2[29]);
+          lstep3[30] = _mm_add_epi32(lstep2[28], lstep1[30]);
+          lstep3[31] = _mm_add_epi32(lstep2[29], lstep1[31]);
+        }
+        {
+          const __m128i k32_m04_p28 = pair_set_epi32(-cospi_4_64, cospi_28_64);
+          const __m128i k32_m28_m04 = pair_set_epi32(-cospi_28_64, -cospi_4_64);
+          const __m128i k32_m20_p12 = pair_set_epi32(-cospi_20_64, cospi_12_64);
+          const __m128i k32_m12_m20 = pair_set_epi32(-cospi_12_64,
+                                                     -cospi_20_64);
+          const __m128i k32_p12_p20 = pair_set_epi32(cospi_12_64, cospi_20_64);
+          const __m128i k32_p28_p04 = pair_set_epi32(cospi_28_64, cospi_4_64);
+
+          u[ 0] = _mm_unpacklo_epi32(lstep2[34], lstep2[60]);
+          u[ 1] = _mm_unpackhi_epi32(lstep2[34], lstep2[60]);
+          u[ 2] = _mm_unpacklo_epi32(lstep2[35], lstep2[61]);
+          u[ 3] = _mm_unpackhi_epi32(lstep2[35], lstep2[61]);
+          u[ 4] = _mm_unpacklo_epi32(lstep2[36], lstep2[58]);
+          u[ 5] = _mm_unpackhi_epi32(lstep2[36], lstep2[58]);
+          u[ 6] = _mm_unpacklo_epi32(lstep2[37], lstep2[59]);
+          u[ 7] = _mm_unpackhi_epi32(lstep2[37], lstep2[59]);
+          u[ 8] = _mm_unpacklo_epi32(lstep2[42], lstep2[52]);
+          u[ 9] = _mm_unpackhi_epi32(lstep2[42], lstep2[52]);
+          u[10] = _mm_unpacklo_epi32(lstep2[43], lstep2[53]);
+          u[11] = _mm_unpackhi_epi32(lstep2[43], lstep2[53]);
+          u[12] = _mm_unpacklo_epi32(lstep2[44], lstep2[50]);
+          u[13] = _mm_unpackhi_epi32(lstep2[44], lstep2[50]);
+          u[14] = _mm_unpacklo_epi32(lstep2[45], lstep2[51]);
+          u[15] = _mm_unpackhi_epi32(lstep2[45], lstep2[51]);
+
+          v[ 0] = k_madd_epi32(u[ 0], k32_m04_p28);
+          v[ 1] = k_madd_epi32(u[ 1], k32_m04_p28);
+          v[ 2] = k_madd_epi32(u[ 2], k32_m04_p28);
+          v[ 3] = k_madd_epi32(u[ 3], k32_m04_p28);
+          v[ 4] = k_madd_epi32(u[ 4], k32_m28_m04);
+          v[ 5] = k_madd_epi32(u[ 5], k32_m28_m04);
+          v[ 6] = k_madd_epi32(u[ 6], k32_m28_m04);
+          v[ 7] = k_madd_epi32(u[ 7], k32_m28_m04);
+          v[ 8] = k_madd_epi32(u[ 8], k32_m20_p12);
+          v[ 9] = k_madd_epi32(u[ 9], k32_m20_p12);
+          v[10] = k_madd_epi32(u[10], k32_m20_p12);
+          v[11] = k_madd_epi32(u[11], k32_m20_p12);
+          v[12] = k_madd_epi32(u[12], k32_m12_m20);
+          v[13] = k_madd_epi32(u[13], k32_m12_m20);
+          v[14] = k_madd_epi32(u[14], k32_m12_m20);
+          v[15] = k_madd_epi32(u[15], k32_m12_m20);
+          v[16] = k_madd_epi32(u[12], k32_m20_p12);
+          v[17] = k_madd_epi32(u[13], k32_m20_p12);
+          v[18] = k_madd_epi32(u[14], k32_m20_p12);
+          v[19] = k_madd_epi32(u[15], k32_m20_p12);
+          v[20] = k_madd_epi32(u[ 8], k32_p12_p20);
+          v[21] = k_madd_epi32(u[ 9], k32_p12_p20);
+          v[22] = k_madd_epi32(u[10], k32_p12_p20);
+          v[23] = k_madd_epi32(u[11], k32_p12_p20);
+          v[24] = k_madd_epi32(u[ 4], k32_m04_p28);
+          v[25] = k_madd_epi32(u[ 5], k32_m04_p28);
+          v[26] = k_madd_epi32(u[ 6], k32_m04_p28);
+          v[27] = k_madd_epi32(u[ 7], k32_m04_p28);
+          v[28] = k_madd_epi32(u[ 0], k32_p28_p04);
+          v[29] = k_madd_epi32(u[ 1], k32_p28_p04);
+          v[30] = k_madd_epi32(u[ 2], k32_p28_p04);
+          v[31] = k_madd_epi32(u[ 3], k32_p28_p04);
+
+#if DCT_HIGH_BIT_DEPTH
+          overflow = k_check_epi32_overflow_32(
+              &v[0], &v[1], &v[2], &v[3], &v[4], &v[5], &v[6], &v[7],
+              &v[8], &v[9], &v[10], &v[11], &v[12], &v[13], &v[14], &v[15],
+              &v[16], &v[17], &v[18], &v[19], &v[20], &v[21], &v[22], &v[23],
+              &v[24], &v[25], &v[26], &v[27], &v[28], &v[29], &v[30], &v[31],
+              &kZero);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+          u[ 0] = k_packs_epi64(v[ 0], v[ 1]);
+          u[ 1] = k_packs_epi64(v[ 2], v[ 3]);
+          u[ 2] = k_packs_epi64(v[ 4], v[ 5]);
+          u[ 3] = k_packs_epi64(v[ 6], v[ 7]);
+          u[ 4] = k_packs_epi64(v[ 8], v[ 9]);
+          u[ 5] = k_packs_epi64(v[10], v[11]);
+          u[ 6] = k_packs_epi64(v[12], v[13]);
+          u[ 7] = k_packs_epi64(v[14], v[15]);
+          u[ 8] = k_packs_epi64(v[16], v[17]);
+          u[ 9] = k_packs_epi64(v[18], v[19]);
+          u[10] = k_packs_epi64(v[20], v[21]);
+          u[11] = k_packs_epi64(v[22], v[23]);
+          u[12] = k_packs_epi64(v[24], v[25]);
+          u[13] = k_packs_epi64(v[26], v[27]);
+          u[14] = k_packs_epi64(v[28], v[29]);
+          u[15] = k_packs_epi64(v[30], v[31]);
+
+          v[ 0] = _mm_add_epi32(u[ 0], k__DCT_CONST_ROUNDING);
+          v[ 1] = _mm_add_epi32(u[ 1], k__DCT_CONST_ROUNDING);
+          v[ 2] = _mm_add_epi32(u[ 2], k__DCT_CONST_ROUNDING);
+          v[ 3] = _mm_add_epi32(u[ 3], k__DCT_CONST_ROUNDING);
+          v[ 4] = _mm_add_epi32(u[ 4], k__DCT_CONST_ROUNDING);
+          v[ 5] = _mm_add_epi32(u[ 5], k__DCT_CONST_ROUNDING);
+          v[ 6] = _mm_add_epi32(u[ 6], k__DCT_CONST_ROUNDING);
+          v[ 7] = _mm_add_epi32(u[ 7], k__DCT_CONST_ROUNDING);
+          v[ 8] = _mm_add_epi32(u[ 8], k__DCT_CONST_ROUNDING);
+          v[ 9] = _mm_add_epi32(u[ 9], k__DCT_CONST_ROUNDING);
+          v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+          v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+          v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+          v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+          v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+          v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+          lstep3[34] = _mm_srai_epi32(v[ 0], DCT_CONST_BITS);
+          lstep3[35] = _mm_srai_epi32(v[ 1], DCT_CONST_BITS);
+          lstep3[36] = _mm_srai_epi32(v[ 2], DCT_CONST_BITS);
+          lstep3[37] = _mm_srai_epi32(v[ 3], DCT_CONST_BITS);
+          lstep3[42] = _mm_srai_epi32(v[ 4], DCT_CONST_BITS);
+          lstep3[43] = _mm_srai_epi32(v[ 5], DCT_CONST_BITS);
+          lstep3[44] = _mm_srai_epi32(v[ 6], DCT_CONST_BITS);
+          lstep3[45] = _mm_srai_epi32(v[ 7], DCT_CONST_BITS);
+          lstep3[50] = _mm_srai_epi32(v[ 8], DCT_CONST_BITS);
+          lstep3[51] = _mm_srai_epi32(v[ 9], DCT_CONST_BITS);
+          lstep3[52] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
+          lstep3[53] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
+          lstep3[58] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
+          lstep3[59] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
+          lstep3[60] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
+          lstep3[61] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
+        }
+        // stage 7
+        {
+          const __m128i k32_p30_p02 = pair_set_epi32(cospi_30_64, cospi_2_64);
+          const __m128i k32_p14_p18 = pair_set_epi32(cospi_14_64, cospi_18_64);
+          const __m128i k32_p22_p10 = pair_set_epi32(cospi_22_64, cospi_10_64);
+          const __m128i k32_p06_p26 = pair_set_epi32(cospi_6_64,  cospi_26_64);
+          const __m128i k32_m26_p06 = pair_set_epi32(-cospi_26_64, cospi_6_64);
+          const __m128i k32_m10_p22 = pair_set_epi32(-cospi_10_64, cospi_22_64);
+          const __m128i k32_m18_p14 = pair_set_epi32(-cospi_18_64, cospi_14_64);
+          const __m128i k32_m02_p30 = pair_set_epi32(-cospi_2_64, cospi_30_64);
+
+          u[ 0] = _mm_unpacklo_epi32(lstep3[16], lstep3[30]);
+          u[ 1] = _mm_unpackhi_epi32(lstep3[16], lstep3[30]);
+          u[ 2] = _mm_unpacklo_epi32(lstep3[17], lstep3[31]);
+          u[ 3] = _mm_unpackhi_epi32(lstep3[17], lstep3[31]);
+          u[ 4] = _mm_unpacklo_epi32(lstep3[18], lstep3[28]);
+          u[ 5] = _mm_unpackhi_epi32(lstep3[18], lstep3[28]);
+          u[ 6] = _mm_unpacklo_epi32(lstep3[19], lstep3[29]);
+          u[ 7] = _mm_unpackhi_epi32(lstep3[19], lstep3[29]);
+          u[ 8] = _mm_unpacklo_epi32(lstep3[20], lstep3[26]);
+          u[ 9] = _mm_unpackhi_epi32(lstep3[20], lstep3[26]);
+          u[10] = _mm_unpacklo_epi32(lstep3[21], lstep3[27]);
+          u[11] = _mm_unpackhi_epi32(lstep3[21], lstep3[27]);
+          u[12] = _mm_unpacklo_epi32(lstep3[22], lstep3[24]);
+          u[13] = _mm_unpackhi_epi32(lstep3[22], lstep3[24]);
+          u[14] = _mm_unpacklo_epi32(lstep3[23], lstep3[25]);
+          u[15] = _mm_unpackhi_epi32(lstep3[23], lstep3[25]);
+
+          v[ 0] = k_madd_epi32(u[ 0], k32_p30_p02);
+          v[ 1] = k_madd_epi32(u[ 1], k32_p30_p02);
+          v[ 2] = k_madd_epi32(u[ 2], k32_p30_p02);
+          v[ 3] = k_madd_epi32(u[ 3], k32_p30_p02);
+          v[ 4] = k_madd_epi32(u[ 4], k32_p14_p18);
+          v[ 5] = k_madd_epi32(u[ 5], k32_p14_p18);
+          v[ 6] = k_madd_epi32(u[ 6], k32_p14_p18);
+          v[ 7] = k_madd_epi32(u[ 7], k32_p14_p18);
+          v[ 8] = k_madd_epi32(u[ 8], k32_p22_p10);
+          v[ 9] = k_madd_epi32(u[ 9], k32_p22_p10);
+          v[10] = k_madd_epi32(u[10], k32_p22_p10);
+          v[11] = k_madd_epi32(u[11], k32_p22_p10);
+          v[12] = k_madd_epi32(u[12], k32_p06_p26);
+          v[13] = k_madd_epi32(u[13], k32_p06_p26);
+          v[14] = k_madd_epi32(u[14], k32_p06_p26);
+          v[15] = k_madd_epi32(u[15], k32_p06_p26);
+          v[16] = k_madd_epi32(u[12], k32_m26_p06);
+          v[17] = k_madd_epi32(u[13], k32_m26_p06);
+          v[18] = k_madd_epi32(u[14], k32_m26_p06);
+          v[19] = k_madd_epi32(u[15], k32_m26_p06);
+          v[20] = k_madd_epi32(u[ 8], k32_m10_p22);
+          v[21] = k_madd_epi32(u[ 9], k32_m10_p22);
+          v[22] = k_madd_epi32(u[10], k32_m10_p22);
+          v[23] = k_madd_epi32(u[11], k32_m10_p22);
+          v[24] = k_madd_epi32(u[ 4], k32_m18_p14);
+          v[25] = k_madd_epi32(u[ 5], k32_m18_p14);
+          v[26] = k_madd_epi32(u[ 6], k32_m18_p14);
+          v[27] = k_madd_epi32(u[ 7], k32_m18_p14);
+          v[28] = k_madd_epi32(u[ 0], k32_m02_p30);
+          v[29] = k_madd_epi32(u[ 1], k32_m02_p30);
+          v[30] = k_madd_epi32(u[ 2], k32_m02_p30);
+          v[31] = k_madd_epi32(u[ 3], k32_m02_p30);
+
+#if DCT_HIGH_BIT_DEPTH
+          overflow = k_check_epi32_overflow_32(
+              &v[0], &v[1], &v[2], &v[3], &v[4], &v[5], &v[6], &v[7],
+              &v[8], &v[9], &v[10], &v[11], &v[12], &v[13], &v[14], &v[15],
+              &v[16], &v[17], &v[18], &v[19], &v[20], &v[21], &v[22], &v[23],
+              &v[24], &v[25], &v[26], &v[27], &v[28], &v[29], &v[30], &v[31],
+              &kZero);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+          u[ 0] = k_packs_epi64(v[ 0], v[ 1]);
+          u[ 1] = k_packs_epi64(v[ 2], v[ 3]);
+          u[ 2] = k_packs_epi64(v[ 4], v[ 5]);
+          u[ 3] = k_packs_epi64(v[ 6], v[ 7]);
+          u[ 4] = k_packs_epi64(v[ 8], v[ 9]);
+          u[ 5] = k_packs_epi64(v[10], v[11]);
+          u[ 6] = k_packs_epi64(v[12], v[13]);
+          u[ 7] = k_packs_epi64(v[14], v[15]);
+          u[ 8] = k_packs_epi64(v[16], v[17]);
+          u[ 9] = k_packs_epi64(v[18], v[19]);
+          u[10] = k_packs_epi64(v[20], v[21]);
+          u[11] = k_packs_epi64(v[22], v[23]);
+          u[12] = k_packs_epi64(v[24], v[25]);
+          u[13] = k_packs_epi64(v[26], v[27]);
+          u[14] = k_packs_epi64(v[28], v[29]);
+          u[15] = k_packs_epi64(v[30], v[31]);
+
+          v[ 0] = _mm_add_epi32(u[ 0], k__DCT_CONST_ROUNDING);
+          v[ 1] = _mm_add_epi32(u[ 1], k__DCT_CONST_ROUNDING);
+          v[ 2] = _mm_add_epi32(u[ 2], k__DCT_CONST_ROUNDING);
+          v[ 3] = _mm_add_epi32(u[ 3], k__DCT_CONST_ROUNDING);
+          v[ 4] = _mm_add_epi32(u[ 4], k__DCT_CONST_ROUNDING);
+          v[ 5] = _mm_add_epi32(u[ 5], k__DCT_CONST_ROUNDING);
+          v[ 6] = _mm_add_epi32(u[ 6], k__DCT_CONST_ROUNDING);
+          v[ 7] = _mm_add_epi32(u[ 7], k__DCT_CONST_ROUNDING);
+          v[ 8] = _mm_add_epi32(u[ 8], k__DCT_CONST_ROUNDING);
+          v[ 9] = _mm_add_epi32(u[ 9], k__DCT_CONST_ROUNDING);
+          v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+          v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+          v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+          v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+          v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+          v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+          u[ 0] = _mm_srai_epi32(v[ 0], DCT_CONST_BITS);
+          u[ 1] = _mm_srai_epi32(v[ 1], DCT_CONST_BITS);
+          u[ 2] = _mm_srai_epi32(v[ 2], DCT_CONST_BITS);
+          u[ 3] = _mm_srai_epi32(v[ 3], DCT_CONST_BITS);
+          u[ 4] = _mm_srai_epi32(v[ 4], DCT_CONST_BITS);
+          u[ 5] = _mm_srai_epi32(v[ 5], DCT_CONST_BITS);
+          u[ 6] = _mm_srai_epi32(v[ 6], DCT_CONST_BITS);
+          u[ 7] = _mm_srai_epi32(v[ 7], DCT_CONST_BITS);
+          u[ 8] = _mm_srai_epi32(v[ 8], DCT_CONST_BITS);
+          u[ 9] = _mm_srai_epi32(v[ 9], DCT_CONST_BITS);
+          u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
+          u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
+          u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
+          u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
+          u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
+          u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
+
+          v[ 0] = _mm_cmplt_epi32(u[ 0], kZero);
+          v[ 1] = _mm_cmplt_epi32(u[ 1], kZero);
+          v[ 2] = _mm_cmplt_epi32(u[ 2], kZero);
+          v[ 3] = _mm_cmplt_epi32(u[ 3], kZero);
+          v[ 4] = _mm_cmplt_epi32(u[ 4], kZero);
+          v[ 5] = _mm_cmplt_epi32(u[ 5], kZero);
+          v[ 6] = _mm_cmplt_epi32(u[ 6], kZero);
+          v[ 7] = _mm_cmplt_epi32(u[ 7], kZero);
+          v[ 8] = _mm_cmplt_epi32(u[ 8], kZero);
+          v[ 9] = _mm_cmplt_epi32(u[ 9], kZero);
+          v[10] = _mm_cmplt_epi32(u[10], kZero);
+          v[11] = _mm_cmplt_epi32(u[11], kZero);
+          v[12] = _mm_cmplt_epi32(u[12], kZero);
+          v[13] = _mm_cmplt_epi32(u[13], kZero);
+          v[14] = _mm_cmplt_epi32(u[14], kZero);
+          v[15] = _mm_cmplt_epi32(u[15], kZero);
+
+          u[ 0] = _mm_sub_epi32(u[ 0], v[ 0]);
+          u[ 1] = _mm_sub_epi32(u[ 1], v[ 1]);
+          u[ 2] = _mm_sub_epi32(u[ 2], v[ 2]);
+          u[ 3] = _mm_sub_epi32(u[ 3], v[ 3]);
+          u[ 4] = _mm_sub_epi32(u[ 4], v[ 4]);
+          u[ 5] = _mm_sub_epi32(u[ 5], v[ 5]);
+          u[ 6] = _mm_sub_epi32(u[ 6], v[ 6]);
+          u[ 7] = _mm_sub_epi32(u[ 7], v[ 7]);
+          u[ 8] = _mm_sub_epi32(u[ 8], v[ 8]);
+          u[ 9] = _mm_sub_epi32(u[ 9], v[ 9]);
+          u[10] = _mm_sub_epi32(u[10], v[10]);
+          u[11] = _mm_sub_epi32(u[11], v[11]);
+          u[12] = _mm_sub_epi32(u[12], v[12]);
+          u[13] = _mm_sub_epi32(u[13], v[13]);
+          u[14] = _mm_sub_epi32(u[14], v[14]);
+          u[15] = _mm_sub_epi32(u[15], v[15]);
+
+          v[ 0] = _mm_add_epi32(u[ 0], K32One);
+          v[ 1] = _mm_add_epi32(u[ 1], K32One);
+          v[ 2] = _mm_add_epi32(u[ 2], K32One);
+          v[ 3] = _mm_add_epi32(u[ 3], K32One);
+          v[ 4] = _mm_add_epi32(u[ 4], K32One);
+          v[ 5] = _mm_add_epi32(u[ 5], K32One);
+          v[ 6] = _mm_add_epi32(u[ 6], K32One);
+          v[ 7] = _mm_add_epi32(u[ 7], K32One);
+          v[ 8] = _mm_add_epi32(u[ 8], K32One);
+          v[ 9] = _mm_add_epi32(u[ 9], K32One);
+          v[10] = _mm_add_epi32(u[10], K32One);
+          v[11] = _mm_add_epi32(u[11], K32One);
+          v[12] = _mm_add_epi32(u[12], K32One);
+          v[13] = _mm_add_epi32(u[13], K32One);
+          v[14] = _mm_add_epi32(u[14], K32One);
+          v[15] = _mm_add_epi32(u[15], K32One);
+
+          u[ 0] = _mm_srai_epi32(v[ 0], 2);
+          u[ 1] = _mm_srai_epi32(v[ 1], 2);
+          u[ 2] = _mm_srai_epi32(v[ 2], 2);
+          u[ 3] = _mm_srai_epi32(v[ 3], 2);
+          u[ 4] = _mm_srai_epi32(v[ 4], 2);
+          u[ 5] = _mm_srai_epi32(v[ 5], 2);
+          u[ 6] = _mm_srai_epi32(v[ 6], 2);
+          u[ 7] = _mm_srai_epi32(v[ 7], 2);
+          u[ 8] = _mm_srai_epi32(v[ 8], 2);
+          u[ 9] = _mm_srai_epi32(v[ 9], 2);
+          u[10] = _mm_srai_epi32(v[10], 2);
+          u[11] = _mm_srai_epi32(v[11], 2);
+          u[12] = _mm_srai_epi32(v[12], 2);
+          u[13] = _mm_srai_epi32(v[13], 2);
+          u[14] = _mm_srai_epi32(v[14], 2);
+          u[15] = _mm_srai_epi32(v[15], 2);
+
+          out[ 2] = _mm_packs_epi32(u[0], u[1]);
+          out[18] = _mm_packs_epi32(u[2], u[3]);
+          out[10] = _mm_packs_epi32(u[4], u[5]);
+          out[26] = _mm_packs_epi32(u[6], u[7]);
+          out[ 6] = _mm_packs_epi32(u[8], u[9]);
+          out[22] = _mm_packs_epi32(u[10], u[11]);
+          out[14] = _mm_packs_epi32(u[12], u[13]);
+          out[30] = _mm_packs_epi32(u[14], u[15]);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x8(&out[2], &out[18], &out[10],
+                                             &out[26], &out[6], &out[22],
+                                             &out[14], &out[30]);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        {
+          lstep1[32] = _mm_add_epi32(lstep3[34], lstep2[32]);
+          lstep1[33] = _mm_add_epi32(lstep3[35], lstep2[33]);
+          lstep1[34] = _mm_sub_epi32(lstep2[32], lstep3[34]);
+          lstep1[35] = _mm_sub_epi32(lstep2[33], lstep3[35]);
+          lstep1[36] = _mm_sub_epi32(lstep2[38], lstep3[36]);
+          lstep1[37] = _mm_sub_epi32(lstep2[39], lstep3[37]);
+          lstep1[38] = _mm_add_epi32(lstep3[36], lstep2[38]);
+          lstep1[39] = _mm_add_epi32(lstep3[37], lstep2[39]);
+          lstep1[40] = _mm_add_epi32(lstep3[42], lstep2[40]);
+          lstep1[41] = _mm_add_epi32(lstep3[43], lstep2[41]);
+          lstep1[42] = _mm_sub_epi32(lstep2[40], lstep3[42]);
+          lstep1[43] = _mm_sub_epi32(lstep2[41], lstep3[43]);
+          lstep1[44] = _mm_sub_epi32(lstep2[46], lstep3[44]);
+          lstep1[45] = _mm_sub_epi32(lstep2[47], lstep3[45]);
+          lstep1[46] = _mm_add_epi32(lstep3[44], lstep2[46]);
+          lstep1[47] = _mm_add_epi32(lstep3[45], lstep2[47]);
+          lstep1[48] = _mm_add_epi32(lstep3[50], lstep2[48]);
+          lstep1[49] = _mm_add_epi32(lstep3[51], lstep2[49]);
+          lstep1[50] = _mm_sub_epi32(lstep2[48], lstep3[50]);
+          lstep1[51] = _mm_sub_epi32(lstep2[49], lstep3[51]);
+          lstep1[52] = _mm_sub_epi32(lstep2[54], lstep3[52]);
+          lstep1[53] = _mm_sub_epi32(lstep2[55], lstep3[53]);
+          lstep1[54] = _mm_add_epi32(lstep3[52], lstep2[54]);
+          lstep1[55] = _mm_add_epi32(lstep3[53], lstep2[55]);
+          lstep1[56] = _mm_add_epi32(lstep3[58], lstep2[56]);
+          lstep1[57] = _mm_add_epi32(lstep3[59], lstep2[57]);
+          lstep1[58] = _mm_sub_epi32(lstep2[56], lstep3[58]);
+          lstep1[59] = _mm_sub_epi32(lstep2[57], lstep3[59]);
+          lstep1[60] = _mm_sub_epi32(lstep2[62], lstep3[60]);
+          lstep1[61] = _mm_sub_epi32(lstep2[63], lstep3[61]);
+          lstep1[62] = _mm_add_epi32(lstep3[60], lstep2[62]);
+          lstep1[63] = _mm_add_epi32(lstep3[61], lstep2[63]);
+        }
+        // stage 8
+        {
+          const __m128i k32_p31_p01 = pair_set_epi32(cospi_31_64, cospi_1_64);
+          const __m128i k32_p15_p17 = pair_set_epi32(cospi_15_64, cospi_17_64);
+          const __m128i k32_p23_p09 = pair_set_epi32(cospi_23_64, cospi_9_64);
+          const __m128i k32_p07_p25 = pair_set_epi32(cospi_7_64, cospi_25_64);
+          const __m128i k32_m25_p07 = pair_set_epi32(-cospi_25_64, cospi_7_64);
+          const __m128i k32_m09_p23 = pair_set_epi32(-cospi_9_64, cospi_23_64);
+          const __m128i k32_m17_p15 = pair_set_epi32(-cospi_17_64, cospi_15_64);
+          const __m128i k32_m01_p31 = pair_set_epi32(-cospi_1_64, cospi_31_64);
+
+          u[ 0] = _mm_unpacklo_epi32(lstep1[32], lstep1[62]);
+          u[ 1] = _mm_unpackhi_epi32(lstep1[32], lstep1[62]);
+          u[ 2] = _mm_unpacklo_epi32(lstep1[33], lstep1[63]);
+          u[ 3] = _mm_unpackhi_epi32(lstep1[33], lstep1[63]);
+          u[ 4] = _mm_unpacklo_epi32(lstep1[34], lstep1[60]);
+          u[ 5] = _mm_unpackhi_epi32(lstep1[34], lstep1[60]);
+          u[ 6] = _mm_unpacklo_epi32(lstep1[35], lstep1[61]);
+          u[ 7] = _mm_unpackhi_epi32(lstep1[35], lstep1[61]);
+          u[ 8] = _mm_unpacklo_epi32(lstep1[36], lstep1[58]);
+          u[ 9] = _mm_unpackhi_epi32(lstep1[36], lstep1[58]);
+          u[10] = _mm_unpacklo_epi32(lstep1[37], lstep1[59]);
+          u[11] = _mm_unpackhi_epi32(lstep1[37], lstep1[59]);
+          u[12] = _mm_unpacklo_epi32(lstep1[38], lstep1[56]);
+          u[13] = _mm_unpackhi_epi32(lstep1[38], lstep1[56]);
+          u[14] = _mm_unpacklo_epi32(lstep1[39], lstep1[57]);
+          u[15] = _mm_unpackhi_epi32(lstep1[39], lstep1[57]);
+
+          v[ 0] = k_madd_epi32(u[ 0], k32_p31_p01);
+          v[ 1] = k_madd_epi32(u[ 1], k32_p31_p01);
+          v[ 2] = k_madd_epi32(u[ 2], k32_p31_p01);
+          v[ 3] = k_madd_epi32(u[ 3], k32_p31_p01);
+          v[ 4] = k_madd_epi32(u[ 4], k32_p15_p17);
+          v[ 5] = k_madd_epi32(u[ 5], k32_p15_p17);
+          v[ 6] = k_madd_epi32(u[ 6], k32_p15_p17);
+          v[ 7] = k_madd_epi32(u[ 7], k32_p15_p17);
+          v[ 8] = k_madd_epi32(u[ 8], k32_p23_p09);
+          v[ 9] = k_madd_epi32(u[ 9], k32_p23_p09);
+          v[10] = k_madd_epi32(u[10], k32_p23_p09);
+          v[11] = k_madd_epi32(u[11], k32_p23_p09);
+          v[12] = k_madd_epi32(u[12], k32_p07_p25);
+          v[13] = k_madd_epi32(u[13], k32_p07_p25);
+          v[14] = k_madd_epi32(u[14], k32_p07_p25);
+          v[15] = k_madd_epi32(u[15], k32_p07_p25);
+          v[16] = k_madd_epi32(u[12], k32_m25_p07);
+          v[17] = k_madd_epi32(u[13], k32_m25_p07);
+          v[18] = k_madd_epi32(u[14], k32_m25_p07);
+          v[19] = k_madd_epi32(u[15], k32_m25_p07);
+          v[20] = k_madd_epi32(u[ 8], k32_m09_p23);
+          v[21] = k_madd_epi32(u[ 9], k32_m09_p23);
+          v[22] = k_madd_epi32(u[10], k32_m09_p23);
+          v[23] = k_madd_epi32(u[11], k32_m09_p23);
+          v[24] = k_madd_epi32(u[ 4], k32_m17_p15);
+          v[25] = k_madd_epi32(u[ 5], k32_m17_p15);
+          v[26] = k_madd_epi32(u[ 6], k32_m17_p15);
+          v[27] = k_madd_epi32(u[ 7], k32_m17_p15);
+          v[28] = k_madd_epi32(u[ 0], k32_m01_p31);
+          v[29] = k_madd_epi32(u[ 1], k32_m01_p31);
+          v[30] = k_madd_epi32(u[ 2], k32_m01_p31);
+          v[31] = k_madd_epi32(u[ 3], k32_m01_p31);
+
+#if DCT_HIGH_BIT_DEPTH
+          overflow = k_check_epi32_overflow_32(
+              &v[0], &v[1], &v[2], &v[3], &v[4], &v[5], &v[6], &v[7],
+              &v[8], &v[9], &v[10], &v[11], &v[12], &v[13], &v[14], &v[15],
+              &v[16], &v[17], &v[18], &v[19], &v[20], &v[21], &v[22], &v[23],
+              &v[24], &v[25], &v[26], &v[27], &v[28], &v[29], &v[30], &v[31],
+              &kZero);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+          u[ 0] = k_packs_epi64(v[ 0], v[ 1]);
+          u[ 1] = k_packs_epi64(v[ 2], v[ 3]);
+          u[ 2] = k_packs_epi64(v[ 4], v[ 5]);
+          u[ 3] = k_packs_epi64(v[ 6], v[ 7]);
+          u[ 4] = k_packs_epi64(v[ 8], v[ 9]);
+          u[ 5] = k_packs_epi64(v[10], v[11]);
+          u[ 6] = k_packs_epi64(v[12], v[13]);
+          u[ 7] = k_packs_epi64(v[14], v[15]);
+          u[ 8] = k_packs_epi64(v[16], v[17]);
+          u[ 9] = k_packs_epi64(v[18], v[19]);
+          u[10] = k_packs_epi64(v[20], v[21]);
+          u[11] = k_packs_epi64(v[22], v[23]);
+          u[12] = k_packs_epi64(v[24], v[25]);
+          u[13] = k_packs_epi64(v[26], v[27]);
+          u[14] = k_packs_epi64(v[28], v[29]);
+          u[15] = k_packs_epi64(v[30], v[31]);
+
+          v[ 0] = _mm_add_epi32(u[ 0], k__DCT_CONST_ROUNDING);
+          v[ 1] = _mm_add_epi32(u[ 1], k__DCT_CONST_ROUNDING);
+          v[ 2] = _mm_add_epi32(u[ 2], k__DCT_CONST_ROUNDING);
+          v[ 3] = _mm_add_epi32(u[ 3], k__DCT_CONST_ROUNDING);
+          v[ 4] = _mm_add_epi32(u[ 4], k__DCT_CONST_ROUNDING);
+          v[ 5] = _mm_add_epi32(u[ 5], k__DCT_CONST_ROUNDING);
+          v[ 6] = _mm_add_epi32(u[ 6], k__DCT_CONST_ROUNDING);
+          v[ 7] = _mm_add_epi32(u[ 7], k__DCT_CONST_ROUNDING);
+          v[ 8] = _mm_add_epi32(u[ 8], k__DCT_CONST_ROUNDING);
+          v[ 9] = _mm_add_epi32(u[ 9], k__DCT_CONST_ROUNDING);
+          v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+          v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+          v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+          v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+          v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+          v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+          u[ 0] = _mm_srai_epi32(v[ 0], DCT_CONST_BITS);
+          u[ 1] = _mm_srai_epi32(v[ 1], DCT_CONST_BITS);
+          u[ 2] = _mm_srai_epi32(v[ 2], DCT_CONST_BITS);
+          u[ 3] = _mm_srai_epi32(v[ 3], DCT_CONST_BITS);
+          u[ 4] = _mm_srai_epi32(v[ 4], DCT_CONST_BITS);
+          u[ 5] = _mm_srai_epi32(v[ 5], DCT_CONST_BITS);
+          u[ 6] = _mm_srai_epi32(v[ 6], DCT_CONST_BITS);
+          u[ 7] = _mm_srai_epi32(v[ 7], DCT_CONST_BITS);
+          u[ 8] = _mm_srai_epi32(v[ 8], DCT_CONST_BITS);
+          u[ 9] = _mm_srai_epi32(v[ 9], DCT_CONST_BITS);
+          u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
+          u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
+          u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
+          u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
+          u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
+          u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
+
+          v[ 0] = _mm_cmplt_epi32(u[ 0], kZero);
+          v[ 1] = _mm_cmplt_epi32(u[ 1], kZero);
+          v[ 2] = _mm_cmplt_epi32(u[ 2], kZero);
+          v[ 3] = _mm_cmplt_epi32(u[ 3], kZero);
+          v[ 4] = _mm_cmplt_epi32(u[ 4], kZero);
+          v[ 5] = _mm_cmplt_epi32(u[ 5], kZero);
+          v[ 6] = _mm_cmplt_epi32(u[ 6], kZero);
+          v[ 7] = _mm_cmplt_epi32(u[ 7], kZero);
+          v[ 8] = _mm_cmplt_epi32(u[ 8], kZero);
+          v[ 9] = _mm_cmplt_epi32(u[ 9], kZero);
+          v[10] = _mm_cmplt_epi32(u[10], kZero);
+          v[11] = _mm_cmplt_epi32(u[11], kZero);
+          v[12] = _mm_cmplt_epi32(u[12], kZero);
+          v[13] = _mm_cmplt_epi32(u[13], kZero);
+          v[14] = _mm_cmplt_epi32(u[14], kZero);
+          v[15] = _mm_cmplt_epi32(u[15], kZero);
+
+          u[ 0] = _mm_sub_epi32(u[ 0], v[ 0]);
+          u[ 1] = _mm_sub_epi32(u[ 1], v[ 1]);
+          u[ 2] = _mm_sub_epi32(u[ 2], v[ 2]);
+          u[ 3] = _mm_sub_epi32(u[ 3], v[ 3]);
+          u[ 4] = _mm_sub_epi32(u[ 4], v[ 4]);
+          u[ 5] = _mm_sub_epi32(u[ 5], v[ 5]);
+          u[ 6] = _mm_sub_epi32(u[ 6], v[ 6]);
+          u[ 7] = _mm_sub_epi32(u[ 7], v[ 7]);
+          u[ 8] = _mm_sub_epi32(u[ 8], v[ 8]);
+          u[ 9] = _mm_sub_epi32(u[ 9], v[ 9]);
+          u[10] = _mm_sub_epi32(u[10], v[10]);
+          u[11] = _mm_sub_epi32(u[11], v[11]);
+          u[12] = _mm_sub_epi32(u[12], v[12]);
+          u[13] = _mm_sub_epi32(u[13], v[13]);
+          u[14] = _mm_sub_epi32(u[14], v[14]);
+          u[15] = _mm_sub_epi32(u[15], v[15]);
+
+          v[0] = _mm_add_epi32(u[0], K32One);
+          v[1] = _mm_add_epi32(u[1], K32One);
+          v[2] = _mm_add_epi32(u[2], K32One);
+          v[3] = _mm_add_epi32(u[3], K32One);
+          v[4] = _mm_add_epi32(u[4], K32One);
+          v[5] = _mm_add_epi32(u[5], K32One);
+          v[6] = _mm_add_epi32(u[6], K32One);
+          v[7] = _mm_add_epi32(u[7], K32One);
+          v[8] = _mm_add_epi32(u[8], K32One);
+          v[9] = _mm_add_epi32(u[9], K32One);
+          v[10] = _mm_add_epi32(u[10], K32One);
+          v[11] = _mm_add_epi32(u[11], K32One);
+          v[12] = _mm_add_epi32(u[12], K32One);
+          v[13] = _mm_add_epi32(u[13], K32One);
+          v[14] = _mm_add_epi32(u[14], K32One);
+          v[15] = _mm_add_epi32(u[15], K32One);
+
+          u[0] = _mm_srai_epi32(v[0], 2);
+          u[1] = _mm_srai_epi32(v[1], 2);
+          u[2] = _mm_srai_epi32(v[2], 2);
+          u[3] = _mm_srai_epi32(v[3], 2);
+          u[4] = _mm_srai_epi32(v[4], 2);
+          u[5] = _mm_srai_epi32(v[5], 2);
+          u[6] = _mm_srai_epi32(v[6], 2);
+          u[7] = _mm_srai_epi32(v[7], 2);
+          u[8] = _mm_srai_epi32(v[8], 2);
+          u[9] = _mm_srai_epi32(v[9], 2);
+          u[10] = _mm_srai_epi32(v[10], 2);
+          u[11] = _mm_srai_epi32(v[11], 2);
+          u[12] = _mm_srai_epi32(v[12], 2);
+          u[13] = _mm_srai_epi32(v[13], 2);
+          u[14] = _mm_srai_epi32(v[14], 2);
+          u[15] = _mm_srai_epi32(v[15], 2);
+
+          out[ 1] = _mm_packs_epi32(u[0], u[1]);
+          out[17] = _mm_packs_epi32(u[2], u[3]);
+          out[ 9] = _mm_packs_epi32(u[4], u[5]);
+          out[25] = _mm_packs_epi32(u[6], u[7]);
+          out[ 7] = _mm_packs_epi32(u[8], u[9]);
+          out[23] = _mm_packs_epi32(u[10], u[11]);
+          out[15] = _mm_packs_epi32(u[12], u[13]);
+          out[31] = _mm_packs_epi32(u[14], u[15]);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x8(&out[1], &out[17], &out[9],
+                                             &out[25], &out[7], &out[23],
+                                             &out[15], &out[31]);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        {
+          const __m128i k32_p27_p05 = pair_set_epi32(cospi_27_64, cospi_5_64);
+          const __m128i k32_p11_p21 = pair_set_epi32(cospi_11_64, cospi_21_64);
+          const __m128i k32_p19_p13 = pair_set_epi32(cospi_19_64, cospi_13_64);
+          const __m128i k32_p03_p29 = pair_set_epi32(cospi_3_64, cospi_29_64);
+          const __m128i k32_m29_p03 = pair_set_epi32(-cospi_29_64, cospi_3_64);
+          const __m128i k32_m13_p19 = pair_set_epi32(-cospi_13_64, cospi_19_64);
+          const __m128i k32_m21_p11 = pair_set_epi32(-cospi_21_64, cospi_11_64);
+          const __m128i k32_m05_p27 = pair_set_epi32(-cospi_5_64, cospi_27_64);
+
+          u[ 0] = _mm_unpacklo_epi32(lstep1[40], lstep1[54]);
+          u[ 1] = _mm_unpackhi_epi32(lstep1[40], lstep1[54]);
+          u[ 2] = _mm_unpacklo_epi32(lstep1[41], lstep1[55]);
+          u[ 3] = _mm_unpackhi_epi32(lstep1[41], lstep1[55]);
+          u[ 4] = _mm_unpacklo_epi32(lstep1[42], lstep1[52]);
+          u[ 5] = _mm_unpackhi_epi32(lstep1[42], lstep1[52]);
+          u[ 6] = _mm_unpacklo_epi32(lstep1[43], lstep1[53]);
+          u[ 7] = _mm_unpackhi_epi32(lstep1[43], lstep1[53]);
+          u[ 8] = _mm_unpacklo_epi32(lstep1[44], lstep1[50]);
+          u[ 9] = _mm_unpackhi_epi32(lstep1[44], lstep1[50]);
+          u[10] = _mm_unpacklo_epi32(lstep1[45], lstep1[51]);
+          u[11] = _mm_unpackhi_epi32(lstep1[45], lstep1[51]);
+          u[12] = _mm_unpacklo_epi32(lstep1[46], lstep1[48]);
+          u[13] = _mm_unpackhi_epi32(lstep1[46], lstep1[48]);
+          u[14] = _mm_unpacklo_epi32(lstep1[47], lstep1[49]);
+          u[15] = _mm_unpackhi_epi32(lstep1[47], lstep1[49]);
+
+          v[ 0] = k_madd_epi32(u[ 0], k32_p27_p05);
+          v[ 1] = k_madd_epi32(u[ 1], k32_p27_p05);
+          v[ 2] = k_madd_epi32(u[ 2], k32_p27_p05);
+          v[ 3] = k_madd_epi32(u[ 3], k32_p27_p05);
+          v[ 4] = k_madd_epi32(u[ 4], k32_p11_p21);
+          v[ 5] = k_madd_epi32(u[ 5], k32_p11_p21);
+          v[ 6] = k_madd_epi32(u[ 6], k32_p11_p21);
+          v[ 7] = k_madd_epi32(u[ 7], k32_p11_p21);
+          v[ 8] = k_madd_epi32(u[ 8], k32_p19_p13);
+          v[ 9] = k_madd_epi32(u[ 9], k32_p19_p13);
+          v[10] = k_madd_epi32(u[10], k32_p19_p13);
+          v[11] = k_madd_epi32(u[11], k32_p19_p13);
+          v[12] = k_madd_epi32(u[12], k32_p03_p29);
+          v[13] = k_madd_epi32(u[13], k32_p03_p29);
+          v[14] = k_madd_epi32(u[14], k32_p03_p29);
+          v[15] = k_madd_epi32(u[15], k32_p03_p29);
+          v[16] = k_madd_epi32(u[12], k32_m29_p03);
+          v[17] = k_madd_epi32(u[13], k32_m29_p03);
+          v[18] = k_madd_epi32(u[14], k32_m29_p03);
+          v[19] = k_madd_epi32(u[15], k32_m29_p03);
+          v[20] = k_madd_epi32(u[ 8], k32_m13_p19);
+          v[21] = k_madd_epi32(u[ 9], k32_m13_p19);
+          v[22] = k_madd_epi32(u[10], k32_m13_p19);
+          v[23] = k_madd_epi32(u[11], k32_m13_p19);
+          v[24] = k_madd_epi32(u[ 4], k32_m21_p11);
+          v[25] = k_madd_epi32(u[ 5], k32_m21_p11);
+          v[26] = k_madd_epi32(u[ 6], k32_m21_p11);
+          v[27] = k_madd_epi32(u[ 7], k32_m21_p11);
+          v[28] = k_madd_epi32(u[ 0], k32_m05_p27);
+          v[29] = k_madd_epi32(u[ 1], k32_m05_p27);
+          v[30] = k_madd_epi32(u[ 2], k32_m05_p27);
+          v[31] = k_madd_epi32(u[ 3], k32_m05_p27);
+
+#if DCT_HIGH_BIT_DEPTH
+          overflow = k_check_epi32_overflow_32(
+              &v[0], &v[1], &v[2], &v[3], &v[4], &v[5], &v[6], &v[7],
+              &v[8], &v[9], &v[10], &v[11], &v[12], &v[13], &v[14], &v[15],
+              &v[16], &v[17], &v[18], &v[19], &v[20], &v[21], &v[22], &v[23],
+              &v[24], &v[25], &v[26], &v[27], &v[28], &v[29], &v[30], &v[31],
+              &kZero);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+          u[ 0] = k_packs_epi64(v[ 0], v[ 1]);
+          u[ 1] = k_packs_epi64(v[ 2], v[ 3]);
+          u[ 2] = k_packs_epi64(v[ 4], v[ 5]);
+          u[ 3] = k_packs_epi64(v[ 6], v[ 7]);
+          u[ 4] = k_packs_epi64(v[ 8], v[ 9]);
+          u[ 5] = k_packs_epi64(v[10], v[11]);
+          u[ 6] = k_packs_epi64(v[12], v[13]);
+          u[ 7] = k_packs_epi64(v[14], v[15]);
+          u[ 8] = k_packs_epi64(v[16], v[17]);
+          u[ 9] = k_packs_epi64(v[18], v[19]);
+          u[10] = k_packs_epi64(v[20], v[21]);
+          u[11] = k_packs_epi64(v[22], v[23]);
+          u[12] = k_packs_epi64(v[24], v[25]);
+          u[13] = k_packs_epi64(v[26], v[27]);
+          u[14] = k_packs_epi64(v[28], v[29]);
+          u[15] = k_packs_epi64(v[30], v[31]);
+
+          v[ 0] = _mm_add_epi32(u[ 0], k__DCT_CONST_ROUNDING);
+          v[ 1] = _mm_add_epi32(u[ 1], k__DCT_CONST_ROUNDING);
+          v[ 2] = _mm_add_epi32(u[ 2], k__DCT_CONST_ROUNDING);
+          v[ 3] = _mm_add_epi32(u[ 3], k__DCT_CONST_ROUNDING);
+          v[ 4] = _mm_add_epi32(u[ 4], k__DCT_CONST_ROUNDING);
+          v[ 5] = _mm_add_epi32(u[ 5], k__DCT_CONST_ROUNDING);
+          v[ 6] = _mm_add_epi32(u[ 6], k__DCT_CONST_ROUNDING);
+          v[ 7] = _mm_add_epi32(u[ 7], k__DCT_CONST_ROUNDING);
+          v[ 8] = _mm_add_epi32(u[ 8], k__DCT_CONST_ROUNDING);
+          v[ 9] = _mm_add_epi32(u[ 9], k__DCT_CONST_ROUNDING);
+          v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+          v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+          v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+          v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+          v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+          v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+          u[ 0] = _mm_srai_epi32(v[ 0], DCT_CONST_BITS);
+          u[ 1] = _mm_srai_epi32(v[ 1], DCT_CONST_BITS);
+          u[ 2] = _mm_srai_epi32(v[ 2], DCT_CONST_BITS);
+          u[ 3] = _mm_srai_epi32(v[ 3], DCT_CONST_BITS);
+          u[ 4] = _mm_srai_epi32(v[ 4], DCT_CONST_BITS);
+          u[ 5] = _mm_srai_epi32(v[ 5], DCT_CONST_BITS);
+          u[ 6] = _mm_srai_epi32(v[ 6], DCT_CONST_BITS);
+          u[ 7] = _mm_srai_epi32(v[ 7], DCT_CONST_BITS);
+          u[ 8] = _mm_srai_epi32(v[ 8], DCT_CONST_BITS);
+          u[ 9] = _mm_srai_epi32(v[ 9], DCT_CONST_BITS);
+          u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
+          u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
+          u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
+          u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
+          u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
+          u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
+
+          v[ 0] = _mm_cmplt_epi32(u[ 0], kZero);
+          v[ 1] = _mm_cmplt_epi32(u[ 1], kZero);
+          v[ 2] = _mm_cmplt_epi32(u[ 2], kZero);
+          v[ 3] = _mm_cmplt_epi32(u[ 3], kZero);
+          v[ 4] = _mm_cmplt_epi32(u[ 4], kZero);
+          v[ 5] = _mm_cmplt_epi32(u[ 5], kZero);
+          v[ 6] = _mm_cmplt_epi32(u[ 6], kZero);
+          v[ 7] = _mm_cmplt_epi32(u[ 7], kZero);
+          v[ 8] = _mm_cmplt_epi32(u[ 8], kZero);
+          v[ 9] = _mm_cmplt_epi32(u[ 9], kZero);
+          v[10] = _mm_cmplt_epi32(u[10], kZero);
+          v[11] = _mm_cmplt_epi32(u[11], kZero);
+          v[12] = _mm_cmplt_epi32(u[12], kZero);
+          v[13] = _mm_cmplt_epi32(u[13], kZero);
+          v[14] = _mm_cmplt_epi32(u[14], kZero);
+          v[15] = _mm_cmplt_epi32(u[15], kZero);
+
+          u[ 0] = _mm_sub_epi32(u[ 0], v[ 0]);
+          u[ 1] = _mm_sub_epi32(u[ 1], v[ 1]);
+          u[ 2] = _mm_sub_epi32(u[ 2], v[ 2]);
+          u[ 3] = _mm_sub_epi32(u[ 3], v[ 3]);
+          u[ 4] = _mm_sub_epi32(u[ 4], v[ 4]);
+          u[ 5] = _mm_sub_epi32(u[ 5], v[ 5]);
+          u[ 6] = _mm_sub_epi32(u[ 6], v[ 6]);
+          u[ 7] = _mm_sub_epi32(u[ 7], v[ 7]);
+          u[ 8] = _mm_sub_epi32(u[ 8], v[ 8]);
+          u[ 9] = _mm_sub_epi32(u[ 9], v[ 9]);
+          u[10] = _mm_sub_epi32(u[10], v[10]);
+          u[11] = _mm_sub_epi32(u[11], v[11]);
+          u[12] = _mm_sub_epi32(u[12], v[12]);
+          u[13] = _mm_sub_epi32(u[13], v[13]);
+          u[14] = _mm_sub_epi32(u[14], v[14]);
+          u[15] = _mm_sub_epi32(u[15], v[15]);
+
+          v[0] = _mm_add_epi32(u[0], K32One);
+          v[1] = _mm_add_epi32(u[1], K32One);
+          v[2] = _mm_add_epi32(u[2], K32One);
+          v[3] = _mm_add_epi32(u[3], K32One);
+          v[4] = _mm_add_epi32(u[4], K32One);
+          v[5] = _mm_add_epi32(u[5], K32One);
+          v[6] = _mm_add_epi32(u[6], K32One);
+          v[7] = _mm_add_epi32(u[7], K32One);
+          v[8] = _mm_add_epi32(u[8], K32One);
+          v[9] = _mm_add_epi32(u[9], K32One);
+          v[10] = _mm_add_epi32(u[10], K32One);
+          v[11] = _mm_add_epi32(u[11], K32One);
+          v[12] = _mm_add_epi32(u[12], K32One);
+          v[13] = _mm_add_epi32(u[13], K32One);
+          v[14] = _mm_add_epi32(u[14], K32One);
+          v[15] = _mm_add_epi32(u[15], K32One);
+
+          u[0] = _mm_srai_epi32(v[0], 2);
+          u[1] = _mm_srai_epi32(v[1], 2);
+          u[2] = _mm_srai_epi32(v[2], 2);
+          u[3] = _mm_srai_epi32(v[3], 2);
+          u[4] = _mm_srai_epi32(v[4], 2);
+          u[5] = _mm_srai_epi32(v[5], 2);
+          u[6] = _mm_srai_epi32(v[6], 2);
+          u[7] = _mm_srai_epi32(v[7], 2);
+          u[8] = _mm_srai_epi32(v[8], 2);
+          u[9] = _mm_srai_epi32(v[9], 2);
+          u[10] = _mm_srai_epi32(v[10], 2);
+          u[11] = _mm_srai_epi32(v[11], 2);
+          u[12] = _mm_srai_epi32(v[12], 2);
+          u[13] = _mm_srai_epi32(v[13], 2);
+          u[14] = _mm_srai_epi32(v[14], 2);
+          u[15] = _mm_srai_epi32(v[15], 2);
+
+          out[ 5] = _mm_packs_epi32(u[0], u[1]);
+          out[21] = _mm_packs_epi32(u[2], u[3]);
+          out[13] = _mm_packs_epi32(u[4], u[5]);
+          out[29] = _mm_packs_epi32(u[6], u[7]);
+          out[ 3] = _mm_packs_epi32(u[8], u[9]);
+          out[19] = _mm_packs_epi32(u[10], u[11]);
+          out[11] = _mm_packs_epi32(u[12], u[13]);
+          out[27] = _mm_packs_epi32(u[14], u[15]);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x8(&out[5], &out[21], &out[13],
+                                             &out[29], &out[3], &out[19],
+                                             &out[11], &out[27]);
+          if (overflow) {
+            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+      }
+#endif  // FDCT32x32_HIGH_PRECISION
+      // Transpose the results, do it as four 8x8 transposes.
+      {
+        int transpose_block;
+        int16_t *output0 = &intermediate[column_start * 32];
+        tran_low_t *output1 = &output_org[column_start * 32];
+        for (transpose_block = 0; transpose_block < 4; ++transpose_block) {
+          __m128i *this_out = &out[8 * transpose_block];
+          // 00 01 02 03 04 05 06 07
+          // 10 11 12 13 14 15 16 17
+          // 20 21 22 23 24 25 26 27
+          // 30 31 32 33 34 35 36 37
+          // 40 41 42 43 44 45 46 47
+          // 50 51 52 53 54 55 56 57
+          // 60 61 62 63 64 65 66 67
+          // 70 71 72 73 74 75 76 77
+          const __m128i tr0_0 = _mm_unpacklo_epi16(this_out[0], this_out[1]);
+          const __m128i tr0_1 = _mm_unpacklo_epi16(this_out[2], this_out[3]);
+          const __m128i tr0_2 = _mm_unpackhi_epi16(this_out[0], this_out[1]);
+          const __m128i tr0_3 = _mm_unpackhi_epi16(this_out[2], this_out[3]);
+          const __m128i tr0_4 = _mm_unpacklo_epi16(this_out[4], this_out[5]);
+          const __m128i tr0_5 = _mm_unpacklo_epi16(this_out[6], this_out[7]);
+          const __m128i tr0_6 = _mm_unpackhi_epi16(this_out[4], this_out[5]);
+          const __m128i tr0_7 = _mm_unpackhi_epi16(this_out[6], this_out[7]);
+          // 00 10 01 11 02 12 03 13
+          // 20 30 21 31 22 32 23 33
+          // 04 14 05 15 06 16 07 17
+          // 24 34 25 35 26 36 27 37
+          // 40 50 41 51 42 52 43 53
+          // 60 70 61 71 62 72 63 73
+          // 54 54 55 55 56 56 57 57
+          // 64 74 65 75 66 76 67 77
+          const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
+          const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3);
+          const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
+          const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3);
+          const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5);
+          const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
+          const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5);
+          const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
+          // 00 10 20 30 01 11 21 31
+          // 40 50 60 70 41 51 61 71
+          // 02 12 22 32 03 13 23 33
+          // 42 52 62 72 43 53 63 73
+          // 04 14 24 34 05 15 21 36
+          // 44 54 64 74 45 55 61 76
+          // 06 16 26 36 07 17 27 37
+          // 46 56 66 76 47 57 67 77
+          __m128i tr2_0 = _mm_unpacklo_epi64(tr1_0, tr1_4);
+          __m128i tr2_1 = _mm_unpackhi_epi64(tr1_0, tr1_4);
+          __m128i tr2_2 = _mm_unpacklo_epi64(tr1_2, tr1_6);
+          __m128i tr2_3 = _mm_unpackhi_epi64(tr1_2, tr1_6);
+          __m128i tr2_4 = _mm_unpacklo_epi64(tr1_1, tr1_5);
+          __m128i tr2_5 = _mm_unpackhi_epi64(tr1_1, tr1_5);
+          __m128i tr2_6 = _mm_unpacklo_epi64(tr1_3, tr1_7);
+          __m128i tr2_7 = _mm_unpackhi_epi64(tr1_3, tr1_7);
+          // 00 10 20 30 40 50 60 70
+          // 01 11 21 31 41 51 61 71
+          // 02 12 22 32 42 52 62 72
+          // 03 13 23 33 43 53 63 73
+          // 04 14 24 34 44 54 64 74
+          // 05 15 25 35 45 55 65 75
+          // 06 16 26 36 46 56 66 76
+          // 07 17 27 37 47 57 67 77
+          if (0 == pass) {
+            // output[j] = (output[j] + 1 + (output[j] > 0)) >> 2;
+            // TODO(cd): see quality impact of only doing
+            //           output[j] = (output[j] + 1) >> 2;
+            //           which would remove the code between here ...
+            __m128i tr2_0_0 = _mm_cmpgt_epi16(tr2_0, kZero);
+            __m128i tr2_1_0 = _mm_cmpgt_epi16(tr2_1, kZero);
+            __m128i tr2_2_0 = _mm_cmpgt_epi16(tr2_2, kZero);
+            __m128i tr2_3_0 = _mm_cmpgt_epi16(tr2_3, kZero);
+            __m128i tr2_4_0 = _mm_cmpgt_epi16(tr2_4, kZero);
+            __m128i tr2_5_0 = _mm_cmpgt_epi16(tr2_5, kZero);
+            __m128i tr2_6_0 = _mm_cmpgt_epi16(tr2_6, kZero);
+            __m128i tr2_7_0 = _mm_cmpgt_epi16(tr2_7, kZero);
+            tr2_0 = _mm_sub_epi16(tr2_0, tr2_0_0);
+            tr2_1 = _mm_sub_epi16(tr2_1, tr2_1_0);
+            tr2_2 = _mm_sub_epi16(tr2_2, tr2_2_0);
+            tr2_3 = _mm_sub_epi16(tr2_3, tr2_3_0);
+            tr2_4 = _mm_sub_epi16(tr2_4, tr2_4_0);
+            tr2_5 = _mm_sub_epi16(tr2_5, tr2_5_0);
+            tr2_6 = _mm_sub_epi16(tr2_6, tr2_6_0);
+            tr2_7 = _mm_sub_epi16(tr2_7, tr2_7_0);
+            //           ... and here.
+            //           PS: also change code in vp9/encoder/vp9_dct.c
+            tr2_0 = _mm_add_epi16(tr2_0, kOne);
+            tr2_1 = _mm_add_epi16(tr2_1, kOne);
+            tr2_2 = _mm_add_epi16(tr2_2, kOne);
+            tr2_3 = _mm_add_epi16(tr2_3, kOne);
+            tr2_4 = _mm_add_epi16(tr2_4, kOne);
+            tr2_5 = _mm_add_epi16(tr2_5, kOne);
+            tr2_6 = _mm_add_epi16(tr2_6, kOne);
+            tr2_7 = _mm_add_epi16(tr2_7, kOne);
+            tr2_0 = _mm_srai_epi16(tr2_0, 2);
+            tr2_1 = _mm_srai_epi16(tr2_1, 2);
+            tr2_2 = _mm_srai_epi16(tr2_2, 2);
+            tr2_3 = _mm_srai_epi16(tr2_3, 2);
+            tr2_4 = _mm_srai_epi16(tr2_4, 2);
+            tr2_5 = _mm_srai_epi16(tr2_5, 2);
+            tr2_6 = _mm_srai_epi16(tr2_6, 2);
+            tr2_7 = _mm_srai_epi16(tr2_7, 2);
+          }
+          // Note: even though all these stores are aligned, using the aligned
+          //       intrinsic make the code slightly slower.
+          if (pass == 0) {
+            _mm_storeu_si128((__m128i *)(output0 + 0 * 32), tr2_0);
+            _mm_storeu_si128((__m128i *)(output0 + 1 * 32), tr2_1);
+            _mm_storeu_si128((__m128i *)(output0 + 2 * 32), tr2_2);
+            _mm_storeu_si128((__m128i *)(output0 + 3 * 32), tr2_3);
+            _mm_storeu_si128((__m128i *)(output0 + 4 * 32), tr2_4);
+            _mm_storeu_si128((__m128i *)(output0 + 5 * 32), tr2_5);
+            _mm_storeu_si128((__m128i *)(output0 + 6 * 32), tr2_6);
+            _mm_storeu_si128((__m128i *)(output0 + 7 * 32), tr2_7);
+            // Process next 8x8
+            output0 += 8;
+          } else {
+            storeu_output(&tr2_0, (output1 + 0 * 32));
+            storeu_output(&tr2_1, (output1 + 1 * 32));
+            storeu_output(&tr2_2, (output1 + 2 * 32));
+            storeu_output(&tr2_3, (output1 + 3 * 32));
+            storeu_output(&tr2_4, (output1 + 4 * 32));
+            storeu_output(&tr2_5, (output1 + 5 * 32));
+            storeu_output(&tr2_6, (output1 + 6 * 32));
+            storeu_output(&tr2_7, (output1 + 7 * 32));
+            // Process next 8x8
+            output1 += 8;
+          }
+        }
+      }
+    }
+  }
+}  // NOLINT
+
+#undef ADD_EPI16
+#undef SUB_EPI16
+#undef HIGH_FDCT32x32_2D_C
+#undef HIGH_FDCT32x32_2D_ROWS_C
diff --git a/media/libvpx/vp9/encoder/x86/vp9_dct_avx2.c b/media/libvpx/vp9/encoder/x86/vp9_dct_avx2.c
new file mode 100644
index 000000000..8f3b61ad8
--- /dev/null
+++ b/media/libvpx/vp9/encoder/x86/vp9_dct_avx2.c
@@ -0,0 +1,26 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <immintrin.h>  // AVX2
+#include "vp9/common/vp9_idct.h"  // for cospi constants
+#include "vpx_ports/mem.h"
+
+
+#define FDCT32x32_2D_AVX2 vp9_fdct32x32_rd_avx2
+#define FDCT32x32_HIGH_PRECISION 0
+#include "vp9/encoder/x86/vp9_dct32x32_avx2_impl.h"
+#undef  FDCT32x32_2D_AVX2
+#undef  FDCT32x32_HIGH_PRECISION
+
+#define FDCT32x32_2D_AVX2 vp9_fdct32x32_avx2
+#define FDCT32x32_HIGH_PRECISION 1
+#include "vp9/encoder/x86/vp9_dct32x32_avx2_impl.h" // NOLINT
+#undef  FDCT32x32_2D_AVX2
+#undef  FDCT32x32_HIGH_PRECISION
diff --git a/media/libvpx/vp9/encoder/x86/vp9_dct_mmx.asm b/media/libvpx/vp9/encoder/x86/vp9_dct_mmx.asm
new file mode 100644
index 000000000..b41fbc8b3
--- /dev/null
+++ b/media/libvpx/vp9/encoder/x86/vp9_dct_mmx.asm
@@ -0,0 +1,101 @@
+;
+;  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION .text
+
+%macro TRANSFORM_COLS 0
+  paddw           m0,        m1
+  movq            m4,        m0
+  psubw           m3,        m2
+  psubw           m4,        m3
+  psraw           m4,        1
+  movq            m5,        m4
+  psubw           m5,        m1 ;b1
+  psubw           m4,        m2 ;c1
+  psubw           m0,        m4
+  paddw           m3,        m5
+                                ; m0 a0
+  SWAP            1,         4  ; m1 c1
+  SWAP            2,         3  ; m2 d1
+  SWAP            3,         5  ; m3 b1
+%endmacro
+
+%macro TRANSPOSE_4X4 0
+  movq            m4,        m0
+  movq            m5,        m2
+  punpcklwd       m4,        m1
+  punpckhwd       m0,        m1
+  punpcklwd       m5,        m3
+  punpckhwd       m2,        m3
+  movq            m1,        m4
+  movq            m3,        m0
+  punpckldq       m1,        m5
+  punpckhdq       m4,        m5
+  punpckldq       m3,        m2
+  punpckhdq       m0,        m2
+  SWAP            2, 3, 0, 1, 4
+%endmacro
+
+INIT_MMX mmx
+cglobal fwht4x4, 3, 4, 8, input, output, stride
+  lea             r3q,       [inputq + strideq*4]
+  movq            m0,        [inputq] ;a1
+  movq            m1,        [inputq + strideq*2] ;b1
+  movq            m2,        [r3q] ;c1
+  movq            m3,        [r3q + strideq*2] ;d1
+
+  TRANSFORM_COLS
+  TRANSPOSE_4X4
+  TRANSFORM_COLS
+  TRANSPOSE_4X4
+
+  psllw           m0,        2
+  psllw           m1,        2
+  psllw           m2,        2
+  psllw           m3,        2
+
+%if CONFIG_VP9_HIGHBITDEPTH
+  pxor            m4,             m4
+  pxor            m5,             m5
+  pcmpgtw         m4,             m0
+  pcmpgtw         m5,             m1
+  movq            m6,             m0
+  movq            m7,             m1
+  punpcklwd       m0,             m4
+  punpcklwd       m1,             m5
+  punpckhwd       m6,             m4
+  punpckhwd       m7,             m5
+  movq            [outputq],      m0
+  movq            [outputq + 8],  m6
+  movq            [outputq + 16], m1
+  movq            [outputq + 24], m7
+  pxor            m4,             m4
+  pxor            m5,             m5
+  pcmpgtw         m4,             m2
+  pcmpgtw         m5,             m3
+  movq            m6,             m2
+  movq            m7,             m3
+  punpcklwd       m2,             m4
+  punpcklwd       m3,             m5
+  punpckhwd       m6,             m4
+  punpckhwd       m7,             m5
+  movq            [outputq + 32], m2
+  movq            [outputq + 40], m6
+  movq            [outputq + 48], m3
+  movq            [outputq + 56], m7
+%else
+  movq            [outputq],      m0
+  movq            [outputq + 8],  m1
+  movq            [outputq + 16], m2
+  movq            [outputq + 24], m3
+%endif
+
+  RET
diff --git a/media/libvpx/vp9/encoder/x86/vp9_dct_sse2.c b/media/libvpx/vp9/encoder/x86/vp9_dct_sse2.c
new file mode 100644
index 000000000..cff4fcbdc
--- /dev/null
+++ b/media/libvpx/vp9/encoder/x86/vp9_dct_sse2.c
@@ -0,0 +1,2429 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#include <emmintrin.h>  // SSE2
+
+#include "./vp9_rtcd.h"
+#include "vp9/common/vp9_idct.h"  // for cospi constants
+#include "vp9/encoder/vp9_dct.h"
+#include "vp9/encoder/x86/vp9_dct_sse2.h"
+#include "vpx_ports/mem.h"
+
+void vp9_fdct4x4_1_sse2(const int16_t *input, tran_low_t *output, int stride) {
+  __m128i in0, in1;
+  __m128i tmp;
+  const __m128i zero = _mm_setzero_si128();
+  in0  = _mm_loadl_epi64((const __m128i *)(input +  0 * stride));
+  in1  = _mm_loadl_epi64((const __m128i *)(input +  1 * stride));
+  in1  = _mm_unpacklo_epi64(in1, _mm_loadl_epi64((const __m128i *)
+         (input +  2 * stride)));
+  in0  = _mm_unpacklo_epi64(in0, _mm_loadl_epi64((const __m128i *)
+         (input +  3 * stride)));
+
+  tmp = _mm_add_epi16(in0, in1);
+  in0 = _mm_unpacklo_epi16(zero, tmp);
+  in1 = _mm_unpackhi_epi16(zero, tmp);
+  in0 = _mm_srai_epi32(in0, 16);
+  in1 = _mm_srai_epi32(in1, 16);
+
+  tmp = _mm_add_epi32(in0, in1);
+  in0 = _mm_unpacklo_epi32(tmp, zero);
+  in1 = _mm_unpackhi_epi32(tmp, zero);
+
+  tmp = _mm_add_epi32(in0, in1);
+  in0 = _mm_srli_si128(tmp, 8);
+
+  in1 = _mm_add_epi32(tmp, in0);
+  in0 = _mm_slli_epi32(in1, 1);
+  store_output(&in0, output);
+}
+
+static INLINE void load_buffer_4x4(const int16_t *input, __m128i *in,
+                                   int stride) {
+  const __m128i k__nonzero_bias_a = _mm_setr_epi16(0, 1, 1, 1, 1, 1, 1, 1);
+  const __m128i k__nonzero_bias_b = _mm_setr_epi16(1, 0, 0, 0, 0, 0, 0, 0);
+  __m128i mask;
+
+  in[0] = _mm_loadl_epi64((const __m128i *)(input + 0 * stride));
+  in[1] = _mm_loadl_epi64((const __m128i *)(input + 1 * stride));
+  in[2] = _mm_loadl_epi64((const __m128i *)(input + 2 * stride));
+  in[3] = _mm_loadl_epi64((const __m128i *)(input + 3 * stride));
+
+  in[0] = _mm_slli_epi16(in[0], 4);
+  in[1] = _mm_slli_epi16(in[1], 4);
+  in[2] = _mm_slli_epi16(in[2], 4);
+  in[3] = _mm_slli_epi16(in[3], 4);
+
+  mask = _mm_cmpeq_epi16(in[0], k__nonzero_bias_a);
+  in[0] = _mm_add_epi16(in[0], mask);
+  in[0] = _mm_add_epi16(in[0], k__nonzero_bias_b);
+}
+
+static INLINE void write_buffer_4x4(tran_low_t *output, __m128i *res) {
+  const __m128i kOne = _mm_set1_epi16(1);
+  __m128i in01 = _mm_unpacklo_epi64(res[0], res[1]);
+  __m128i in23 = _mm_unpacklo_epi64(res[2], res[3]);
+  __m128i out01 = _mm_add_epi16(in01, kOne);
+  __m128i out23 = _mm_add_epi16(in23, kOne);
+  out01 = _mm_srai_epi16(out01, 2);
+  out23 = _mm_srai_epi16(out23, 2);
+  store_output(&out01, (output + 0 * 8));
+  store_output(&out23, (output + 1 * 8));
+}
+
+static INLINE void transpose_4x4(__m128i *res) {
+  // Combine and transpose
+  // 00 01 02 03 20 21 22 23
+  // 10 11 12 13 30 31 32 33
+  const __m128i tr0_0 = _mm_unpacklo_epi16(res[0], res[1]);
+  const __m128i tr0_1 = _mm_unpackhi_epi16(res[0], res[1]);
+
+  // 00 10 01 11 02 12 03 13
+  // 20 30 21 31 22 32 23 33
+  res[0] = _mm_unpacklo_epi32(tr0_0, tr0_1);
+  res[2] = _mm_unpackhi_epi32(tr0_0, tr0_1);
+
+  // 00 10 20 30 01 11 21 31
+  // 02 12 22 32 03 13 23 33
+  // only use the first 4 16-bit integers
+  res[1] = _mm_unpackhi_epi64(res[0], res[0]);
+  res[3] = _mm_unpackhi_epi64(res[2], res[2]);
+}
+
+static void fdct4_sse2(__m128i *in) {
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+
+  __m128i u[4], v[4];
+  u[0]=_mm_unpacklo_epi16(in[0], in[1]);
+  u[1]=_mm_unpacklo_epi16(in[3], in[2]);
+
+  v[0] = _mm_add_epi16(u[0], u[1]);
+  v[1] = _mm_sub_epi16(u[0], u[1]);
+
+  u[0] = _mm_madd_epi16(v[0], k__cospi_p16_p16);  // 0
+  u[1] = _mm_madd_epi16(v[0], k__cospi_p16_m16);  // 2
+  u[2] = _mm_madd_epi16(v[1], k__cospi_p08_p24);  // 1
+  u[3] = _mm_madd_epi16(v[1], k__cospi_p24_m08);  // 3
+
+  v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+  v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+  v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+  v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+  u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
+
+  in[0] = _mm_packs_epi32(u[0], u[1]);
+  in[1] = _mm_packs_epi32(u[2], u[3]);
+  transpose_4x4(in);
+}
+
+static void fadst4_sse2(__m128i *in) {
+  const __m128i k__sinpi_p01_p02 = pair_set_epi16(sinpi_1_9, sinpi_2_9);
+  const __m128i k__sinpi_p04_m01 = pair_set_epi16(sinpi_4_9, -sinpi_1_9);
+  const __m128i k__sinpi_p03_p04 = pair_set_epi16(sinpi_3_9, sinpi_4_9);
+  const __m128i k__sinpi_m03_p02 = pair_set_epi16(-sinpi_3_9, sinpi_2_9);
+  const __m128i k__sinpi_p03_p03 = _mm_set1_epi16((int16_t)sinpi_3_9);
+  const __m128i kZero = _mm_set1_epi16(0);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  __m128i u[8], v[8];
+  __m128i in7 = _mm_add_epi16(in[0], in[1]);
+
+  u[0] = _mm_unpacklo_epi16(in[0], in[1]);
+  u[1] = _mm_unpacklo_epi16(in[2], in[3]);
+  u[2] = _mm_unpacklo_epi16(in7, kZero);
+  u[3] = _mm_unpacklo_epi16(in[2], kZero);
+  u[4] = _mm_unpacklo_epi16(in[3], kZero);
+
+  v[0] = _mm_madd_epi16(u[0], k__sinpi_p01_p02);  // s0 + s2
+  v[1] = _mm_madd_epi16(u[1], k__sinpi_p03_p04);  // s4 + s5
+  v[2] = _mm_madd_epi16(u[2], k__sinpi_p03_p03);  // x1
+  v[3] = _mm_madd_epi16(u[0], k__sinpi_p04_m01);  // s1 - s3
+  v[4] = _mm_madd_epi16(u[1], k__sinpi_m03_p02);  // -s4 + s6
+  v[5] = _mm_madd_epi16(u[3], k__sinpi_p03_p03);  // s4
+  v[6] = _mm_madd_epi16(u[4], k__sinpi_p03_p03);
+
+  u[0] = _mm_add_epi32(v[0], v[1]);
+  u[1] = _mm_sub_epi32(v[2], v[6]);
+  u[2] = _mm_add_epi32(v[3], v[4]);
+  u[3] = _mm_sub_epi32(u[2], u[0]);
+  u[4] = _mm_slli_epi32(v[5], 2);
+  u[5] = _mm_sub_epi32(u[4], v[5]);
+  u[6] = _mm_add_epi32(u[3], u[5]);
+
+  v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+  v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+  v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+  v[3] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+
+  u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
+
+  in[0] = _mm_packs_epi32(u[0], u[2]);
+  in[1] = _mm_packs_epi32(u[1], u[3]);
+  transpose_4x4(in);
+}
+
+void vp9_fht4x4_sse2(const int16_t *input, tran_low_t *output,
+                     int stride, int tx_type) {
+  __m128i in[4];
+
+  switch (tx_type) {
+    case DCT_DCT:
+      vp9_fdct4x4_sse2(input, output, stride);
+      break;
+    case ADST_DCT:
+      load_buffer_4x4(input, in, stride);
+      fadst4_sse2(in);
+      fdct4_sse2(in);
+      write_buffer_4x4(output, in);
+      break;
+    case DCT_ADST:
+      load_buffer_4x4(input, in, stride);
+      fdct4_sse2(in);
+      fadst4_sse2(in);
+      write_buffer_4x4(output, in);
+      break;
+    case ADST_ADST:
+      load_buffer_4x4(input, in, stride);
+      fadst4_sse2(in);
+      fadst4_sse2(in);
+      write_buffer_4x4(output, in);
+      break;
+   default:
+     assert(0);
+     break;
+  }
+}
+
+void vp9_fdct8x8_1_sse2(const int16_t *input, tran_low_t *output, int stride) {
+  __m128i in0  = _mm_load_si128((const __m128i *)(input + 0 * stride));
+  __m128i in1  = _mm_load_si128((const __m128i *)(input + 1 * stride));
+  __m128i in2  = _mm_load_si128((const __m128i *)(input + 2 * stride));
+  __m128i in3  = _mm_load_si128((const __m128i *)(input + 3 * stride));
+  __m128i u0, u1, sum;
+
+  u0 = _mm_add_epi16(in0, in1);
+  u1 = _mm_add_epi16(in2, in3);
+
+  in0  = _mm_load_si128((const __m128i *)(input + 4 * stride));
+  in1  = _mm_load_si128((const __m128i *)(input + 5 * stride));
+  in2  = _mm_load_si128((const __m128i *)(input + 6 * stride));
+  in3  = _mm_load_si128((const __m128i *)(input + 7 * stride));
+
+  sum = _mm_add_epi16(u0, u1);
+
+  in0 = _mm_add_epi16(in0, in1);
+  in2 = _mm_add_epi16(in2, in3);
+  sum = _mm_add_epi16(sum, in0);
+
+  u0  = _mm_setzero_si128();
+  sum = _mm_add_epi16(sum, in2);
+
+  in0 = _mm_unpacklo_epi16(u0, sum);
+  in1 = _mm_unpackhi_epi16(u0, sum);
+  in0 = _mm_srai_epi32(in0, 16);
+  in1 = _mm_srai_epi32(in1, 16);
+
+  sum = _mm_add_epi32(in0, in1);
+  in0 = _mm_unpacklo_epi32(sum, u0);
+  in1 = _mm_unpackhi_epi32(sum, u0);
+
+  sum = _mm_add_epi32(in0, in1);
+  in0 = _mm_srli_si128(sum, 8);
+
+  in1 = _mm_add_epi32(sum, in0);
+  store_output(&in1, output);
+}
+
+void vp9_fdct8x8_quant_sse2(const int16_t *input, int stride,
+                            int16_t* coeff_ptr, intptr_t n_coeffs,
+                            int skip_block, const int16_t* zbin_ptr,
+                            const int16_t* round_ptr, const int16_t* quant_ptr,
+                            const int16_t* quant_shift_ptr, int16_t* qcoeff_ptr,
+                            int16_t* dqcoeff_ptr, const int16_t* dequant_ptr,
+                            uint16_t* eob_ptr,
+                            const int16_t* scan_ptr,
+                            const int16_t* iscan_ptr) {
+  __m128i zero;
+  int pass;
+  // Constants
+  //    When we use them, in one case, they are all the same. In all others
+  //    it's a pair of them that we need to repeat four times. This is done
+  //    by constructing the 32 bit constant corresponding to that pair.
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64);
+  const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64);
+  const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  // Load input
+  __m128i in0  = _mm_load_si128((const __m128i *)(input + 0 * stride));
+  __m128i in1  = _mm_load_si128((const __m128i *)(input + 1 * stride));
+  __m128i in2  = _mm_load_si128((const __m128i *)(input + 2 * stride));
+  __m128i in3  = _mm_load_si128((const __m128i *)(input + 3 * stride));
+  __m128i in4  = _mm_load_si128((const __m128i *)(input + 4 * stride));
+  __m128i in5  = _mm_load_si128((const __m128i *)(input + 5 * stride));
+  __m128i in6  = _mm_load_si128((const __m128i *)(input + 6 * stride));
+  __m128i in7  = _mm_load_si128((const __m128i *)(input + 7 * stride));
+  __m128i *in[8];
+  int index = 0;
+
+  (void)scan_ptr;
+  (void)zbin_ptr;
+  (void)quant_shift_ptr;
+  (void)coeff_ptr;
+
+  // Pre-condition input (shift by two)
+  in0 = _mm_slli_epi16(in0, 2);
+  in1 = _mm_slli_epi16(in1, 2);
+  in2 = _mm_slli_epi16(in2, 2);
+  in3 = _mm_slli_epi16(in3, 2);
+  in4 = _mm_slli_epi16(in4, 2);
+  in5 = _mm_slli_epi16(in5, 2);
+  in6 = _mm_slli_epi16(in6, 2);
+  in7 = _mm_slli_epi16(in7, 2);
+
+  in[0] = &in0;
+  in[1] = &in1;
+  in[2] = &in2;
+  in[3] = &in3;
+  in[4] = &in4;
+  in[5] = &in5;
+  in[6] = &in6;
+  in[7] = &in7;
+
+  // We do two passes, first the columns, then the rows. The results of the
+  // first pass are transposed so that the same column code can be reused. The
+  // results of the second pass are also transposed so that the rows (processed
+  // as columns) are put back in row positions.
+  for (pass = 0; pass < 2; pass++) {
+    // To store results of each pass before the transpose.
+    __m128i res0, res1, res2, res3, res4, res5, res6, res7;
+    // Add/subtract
+    const __m128i q0 = _mm_add_epi16(in0, in7);
+    const __m128i q1 = _mm_add_epi16(in1, in6);
+    const __m128i q2 = _mm_add_epi16(in2, in5);
+    const __m128i q3 = _mm_add_epi16(in3, in4);
+    const __m128i q4 = _mm_sub_epi16(in3, in4);
+    const __m128i q5 = _mm_sub_epi16(in2, in5);
+    const __m128i q6 = _mm_sub_epi16(in1, in6);
+    const __m128i q7 = _mm_sub_epi16(in0, in7);
+    // Work on first four results
+    {
+      // Add/subtract
+      const __m128i r0 = _mm_add_epi16(q0, q3);
+      const __m128i r1 = _mm_add_epi16(q1, q2);
+      const __m128i r2 = _mm_sub_epi16(q1, q2);
+      const __m128i r3 = _mm_sub_epi16(q0, q3);
+      // Interleave to do the multiply by constants which gets us into 32bits
+      const __m128i t0 = _mm_unpacklo_epi16(r0, r1);
+      const __m128i t1 = _mm_unpackhi_epi16(r0, r1);
+      const __m128i t2 = _mm_unpacklo_epi16(r2, r3);
+      const __m128i t3 = _mm_unpackhi_epi16(r2, r3);
+      const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16);
+      const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_p16);
+      const __m128i u2 = _mm_madd_epi16(t0, k__cospi_p16_m16);
+      const __m128i u3 = _mm_madd_epi16(t1, k__cospi_p16_m16);
+      const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p24_p08);
+      const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p24_p08);
+      const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m08_p24);
+      const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m08_p24);
+      // dct_const_round_shift
+      const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
+      const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
+      const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
+      const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
+      const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
+      const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
+      const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
+      const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
+      const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+      const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+      const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+      const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+      const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
+      const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
+      const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
+      const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
+      // Combine
+      res0 = _mm_packs_epi32(w0, w1);
+      res4 = _mm_packs_epi32(w2, w3);
+      res2 = _mm_packs_epi32(w4, w5);
+      res6 = _mm_packs_epi32(w6, w7);
+    }
+    // Work on next four results
+    {
+      // Interleave to do the multiply by constants which gets us into 32bits
+      const __m128i d0 = _mm_unpacklo_epi16(q6, q5);
+      const __m128i d1 = _mm_unpackhi_epi16(q6, q5);
+      const __m128i e0 = _mm_madd_epi16(d0, k__cospi_p16_m16);
+      const __m128i e1 = _mm_madd_epi16(d1, k__cospi_p16_m16);
+      const __m128i e2 = _mm_madd_epi16(d0, k__cospi_p16_p16);
+      const __m128i e3 = _mm_madd_epi16(d1, k__cospi_p16_p16);
+      // dct_const_round_shift
+      const __m128i f0 = _mm_add_epi32(e0, k__DCT_CONST_ROUNDING);
+      const __m128i f1 = _mm_add_epi32(e1, k__DCT_CONST_ROUNDING);
+      const __m128i f2 = _mm_add_epi32(e2, k__DCT_CONST_ROUNDING);
+      const __m128i f3 = _mm_add_epi32(e3, k__DCT_CONST_ROUNDING);
+      const __m128i s0 = _mm_srai_epi32(f0, DCT_CONST_BITS);
+      const __m128i s1 = _mm_srai_epi32(f1, DCT_CONST_BITS);
+      const __m128i s2 = _mm_srai_epi32(f2, DCT_CONST_BITS);
+      const __m128i s3 = _mm_srai_epi32(f3, DCT_CONST_BITS);
+      // Combine
+      const __m128i r0 = _mm_packs_epi32(s0, s1);
+      const __m128i r1 = _mm_packs_epi32(s2, s3);
+      // Add/subtract
+      const __m128i x0 = _mm_add_epi16(q4, r0);
+      const __m128i x1 = _mm_sub_epi16(q4, r0);
+      const __m128i x2 = _mm_sub_epi16(q7, r1);
+      const __m128i x3 = _mm_add_epi16(q7, r1);
+      // Interleave to do the multiply by constants which gets us into 32bits
+      const __m128i t0 = _mm_unpacklo_epi16(x0, x3);
+      const __m128i t1 = _mm_unpackhi_epi16(x0, x3);
+      const __m128i t2 = _mm_unpacklo_epi16(x1, x2);
+      const __m128i t3 = _mm_unpackhi_epi16(x1, x2);
+      const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p28_p04);
+      const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p28_p04);
+      const __m128i u2 = _mm_madd_epi16(t0, k__cospi_m04_p28);
+      const __m128i u3 = _mm_madd_epi16(t1, k__cospi_m04_p28);
+      const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p12_p20);
+      const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p12_p20);
+      const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m20_p12);
+      const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m20_p12);
+      // dct_const_round_shift
+      const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
+      const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
+      const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
+      const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
+      const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
+      const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
+      const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
+      const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
+      const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+      const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+      const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+      const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+      const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
+      const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
+      const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
+      const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
+      // Combine
+      res1 = _mm_packs_epi32(w0, w1);
+      res7 = _mm_packs_epi32(w2, w3);
+      res5 = _mm_packs_epi32(w4, w5);
+      res3 = _mm_packs_epi32(w6, w7);
+    }
+    // Transpose the 8x8.
+    {
+      // 00 01 02 03 04 05 06 07
+      // 10 11 12 13 14 15 16 17
+      // 20 21 22 23 24 25 26 27
+      // 30 31 32 33 34 35 36 37
+      // 40 41 42 43 44 45 46 47
+      // 50 51 52 53 54 55 56 57
+      // 60 61 62 63 64 65 66 67
+      // 70 71 72 73 74 75 76 77
+      const __m128i tr0_0 = _mm_unpacklo_epi16(res0, res1);
+      const __m128i tr0_1 = _mm_unpacklo_epi16(res2, res3);
+      const __m128i tr0_2 = _mm_unpackhi_epi16(res0, res1);
+      const __m128i tr0_3 = _mm_unpackhi_epi16(res2, res3);
+      const __m128i tr0_4 = _mm_unpacklo_epi16(res4, res5);
+      const __m128i tr0_5 = _mm_unpacklo_epi16(res6, res7);
+      const __m128i tr0_6 = _mm_unpackhi_epi16(res4, res5);
+      const __m128i tr0_7 = _mm_unpackhi_epi16(res6, res7);
+      // 00 10 01 11 02 12 03 13
+      // 20 30 21 31 22 32 23 33
+      // 04 14 05 15 06 16 07 17
+      // 24 34 25 35 26 36 27 37
+      // 40 50 41 51 42 52 43 53
+      // 60 70 61 71 62 72 63 73
+      // 54 54 55 55 56 56 57 57
+      // 64 74 65 75 66 76 67 77
+      const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
+      const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3);
+      const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
+      const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3);
+      const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5);
+      const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
+      const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5);
+      const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
+      // 00 10 20 30 01 11 21 31
+      // 40 50 60 70 41 51 61 71
+      // 02 12 22 32 03 13 23 33
+      // 42 52 62 72 43 53 63 73
+      // 04 14 24 34 05 15 21 36
+      // 44 54 64 74 45 55 61 76
+      // 06 16 26 36 07 17 27 37
+      // 46 56 66 76 47 57 67 77
+      in0 = _mm_unpacklo_epi64(tr1_0, tr1_4);
+      in1 = _mm_unpackhi_epi64(tr1_0, tr1_4);
+      in2 = _mm_unpacklo_epi64(tr1_2, tr1_6);
+      in3 = _mm_unpackhi_epi64(tr1_2, tr1_6);
+      in4 = _mm_unpacklo_epi64(tr1_1, tr1_5);
+      in5 = _mm_unpackhi_epi64(tr1_1, tr1_5);
+      in6 = _mm_unpacklo_epi64(tr1_3, tr1_7);
+      in7 = _mm_unpackhi_epi64(tr1_3, tr1_7);
+      // 00 10 20 30 40 50 60 70
+      // 01 11 21 31 41 51 61 71
+      // 02 12 22 32 42 52 62 72
+      // 03 13 23 33 43 53 63 73
+      // 04 14 24 34 44 54 64 74
+      // 05 15 25 35 45 55 65 75
+      // 06 16 26 36 46 56 66 76
+      // 07 17 27 37 47 57 67 77
+    }
+  }
+  // Post-condition output and store it
+  {
+    // Post-condition (division by two)
+    //    division of two 16 bits signed numbers using shifts
+    //    n / 2 = (n - (n >> 15)) >> 1
+    const __m128i sign_in0 = _mm_srai_epi16(in0, 15);
+    const __m128i sign_in1 = _mm_srai_epi16(in1, 15);
+    const __m128i sign_in2 = _mm_srai_epi16(in2, 15);
+    const __m128i sign_in3 = _mm_srai_epi16(in3, 15);
+    const __m128i sign_in4 = _mm_srai_epi16(in4, 15);
+    const __m128i sign_in5 = _mm_srai_epi16(in5, 15);
+    const __m128i sign_in6 = _mm_srai_epi16(in6, 15);
+    const __m128i sign_in7 = _mm_srai_epi16(in7, 15);
+    in0 = _mm_sub_epi16(in0, sign_in0);
+    in1 = _mm_sub_epi16(in1, sign_in1);
+    in2 = _mm_sub_epi16(in2, sign_in2);
+    in3 = _mm_sub_epi16(in3, sign_in3);
+    in4 = _mm_sub_epi16(in4, sign_in4);
+    in5 = _mm_sub_epi16(in5, sign_in5);
+    in6 = _mm_sub_epi16(in6, sign_in6);
+    in7 = _mm_sub_epi16(in7, sign_in7);
+    in0 = _mm_srai_epi16(in0, 1);
+    in1 = _mm_srai_epi16(in1, 1);
+    in2 = _mm_srai_epi16(in2, 1);
+    in3 = _mm_srai_epi16(in3, 1);
+    in4 = _mm_srai_epi16(in4, 1);
+    in5 = _mm_srai_epi16(in5, 1);
+    in6 = _mm_srai_epi16(in6, 1);
+    in7 = _mm_srai_epi16(in7, 1);
+  }
+
+  iscan_ptr += n_coeffs;
+  qcoeff_ptr += n_coeffs;
+  dqcoeff_ptr += n_coeffs;
+  n_coeffs = -n_coeffs;
+  zero = _mm_setzero_si128();
+
+  if (!skip_block) {
+    __m128i eob;
+    __m128i round, quant, dequant;
+    {
+      __m128i coeff0, coeff1;
+
+      // Setup global values
+      {
+        round = _mm_load_si128((const __m128i*)round_ptr);
+        quant = _mm_load_si128((const __m128i*)quant_ptr);
+        dequant = _mm_load_si128((const __m128i*)dequant_ptr);
+      }
+
+      {
+        __m128i coeff0_sign, coeff1_sign;
+        __m128i qcoeff0, qcoeff1;
+        __m128i qtmp0, qtmp1;
+        // Do DC and first 15 AC
+        coeff0 = *in[0];
+        coeff1 = *in[1];
+
+        // Poor man's sign extract
+        coeff0_sign = _mm_srai_epi16(coeff0, 15);
+        coeff1_sign = _mm_srai_epi16(coeff1, 15);
+        qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        qcoeff0 = _mm_adds_epi16(qcoeff0, round);
+        round = _mm_unpackhi_epi64(round, round);
+        qcoeff1 = _mm_adds_epi16(qcoeff1, round);
+        qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
+        quant = _mm_unpackhi_epi64(quant, quant);
+        qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
+
+        // Reinsert signs
+        qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), qcoeff0);
+        _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
+
+        coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
+        dequant = _mm_unpackhi_epi64(dequant, dequant);
+        coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+
+        _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), coeff0);
+        _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, coeff1);
+      }
+
+      {
+        // Scan for eob
+        __m128i zero_coeff0, zero_coeff1;
+        __m128i nzero_coeff0, nzero_coeff1;
+        __m128i iscan0, iscan1;
+        __m128i eob1;
+        zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
+        zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
+        nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
+        nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
+        iscan0 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs));
+        iscan1 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs) + 1);
+        // Add one to convert from indices to counts
+        iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
+        iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
+        eob = _mm_and_si128(iscan0, nzero_coeff0);
+        eob1 = _mm_and_si128(iscan1, nzero_coeff1);
+        eob = _mm_max_epi16(eob, eob1);
+      }
+      n_coeffs += 8 * 2;
+    }
+
+    // AC only loop
+    index = 2;
+    while (n_coeffs < 0) {
+      __m128i coeff0, coeff1;
+      {
+        __m128i coeff0_sign, coeff1_sign;
+        __m128i qcoeff0, qcoeff1;
+        __m128i qtmp0, qtmp1;
+
+        assert(index < (int)(sizeof(in) / sizeof(in[0])) - 1);
+        coeff0 = *in[index];
+        coeff1 = *in[index + 1];
+
+        // Poor man's sign extract
+        coeff0_sign = _mm_srai_epi16(coeff0, 15);
+        coeff1_sign = _mm_srai_epi16(coeff1, 15);
+        qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        qcoeff0 = _mm_adds_epi16(qcoeff0, round);
+        qcoeff1 = _mm_adds_epi16(qcoeff1, round);
+        qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
+        qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
+
+        // Reinsert signs
+        qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), qcoeff0);
+        _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
+
+        coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
+        coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+
+        _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), coeff0);
+        _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, coeff1);
+      }
+
+      {
+        // Scan for eob
+        __m128i zero_coeff0, zero_coeff1;
+        __m128i nzero_coeff0, nzero_coeff1;
+        __m128i iscan0, iscan1;
+        __m128i eob0, eob1;
+        zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
+        zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
+        nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
+        nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
+        iscan0 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs));
+        iscan1 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs) + 1);
+        // Add one to convert from indices to counts
+        iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
+        iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
+        eob0 = _mm_and_si128(iscan0, nzero_coeff0);
+        eob1 = _mm_and_si128(iscan1, nzero_coeff1);
+        eob0 = _mm_max_epi16(eob0, eob1);
+        eob = _mm_max_epi16(eob, eob0);
+      }
+      n_coeffs += 8 * 2;
+      index += 2;
+    }
+
+    // Accumulate EOB
+    {
+      __m128i eob_shuffled;
+      eob_shuffled = _mm_shuffle_epi32(eob, 0xe);
+      eob = _mm_max_epi16(eob, eob_shuffled);
+      eob_shuffled = _mm_shufflelo_epi16(eob, 0xe);
+      eob = _mm_max_epi16(eob, eob_shuffled);
+      eob_shuffled = _mm_shufflelo_epi16(eob, 0x1);
+      eob = _mm_max_epi16(eob, eob_shuffled);
+      *eob_ptr = _mm_extract_epi16(eob, 1);
+    }
+  } else {
+    do {
+      _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), zero);
+      _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, zero);
+      _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), zero);
+      _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, zero);
+      n_coeffs += 8 * 2;
+    } while (n_coeffs < 0);
+    *eob_ptr = 0;
+  }
+}
+
+// load 8x8 array
+static INLINE void load_buffer_8x8(const int16_t *input, __m128i *in,
+                                   int stride) {
+  in[0]  = _mm_load_si128((const __m128i *)(input + 0 * stride));
+  in[1]  = _mm_load_si128((const __m128i *)(input + 1 * stride));
+  in[2]  = _mm_load_si128((const __m128i *)(input + 2 * stride));
+  in[3]  = _mm_load_si128((const __m128i *)(input + 3 * stride));
+  in[4]  = _mm_load_si128((const __m128i *)(input + 4 * stride));
+  in[5]  = _mm_load_si128((const __m128i *)(input + 5 * stride));
+  in[6]  = _mm_load_si128((const __m128i *)(input + 6 * stride));
+  in[7]  = _mm_load_si128((const __m128i *)(input + 7 * stride));
+
+  in[0] = _mm_slli_epi16(in[0], 2);
+  in[1] = _mm_slli_epi16(in[1], 2);
+  in[2] = _mm_slli_epi16(in[2], 2);
+  in[3] = _mm_slli_epi16(in[3], 2);
+  in[4] = _mm_slli_epi16(in[4], 2);
+  in[5] = _mm_slli_epi16(in[5], 2);
+  in[6] = _mm_slli_epi16(in[6], 2);
+  in[7] = _mm_slli_epi16(in[7], 2);
+}
+
+// right shift and rounding
+static INLINE void right_shift_8x8(__m128i *res, const int bit) {
+  __m128i sign0 = _mm_srai_epi16(res[0], 15);
+  __m128i sign1 = _mm_srai_epi16(res[1], 15);
+  __m128i sign2 = _mm_srai_epi16(res[2], 15);
+  __m128i sign3 = _mm_srai_epi16(res[3], 15);
+  __m128i sign4 = _mm_srai_epi16(res[4], 15);
+  __m128i sign5 = _mm_srai_epi16(res[5], 15);
+  __m128i sign6 = _mm_srai_epi16(res[6], 15);
+  __m128i sign7 = _mm_srai_epi16(res[7], 15);
+
+  if (bit == 2) {
+    const __m128i const_rounding = _mm_set1_epi16(1);
+    res[0] = _mm_add_epi16(res[0], const_rounding);
+    res[1] = _mm_add_epi16(res[1], const_rounding);
+    res[2] = _mm_add_epi16(res[2], const_rounding);
+    res[3] = _mm_add_epi16(res[3], const_rounding);
+    res[4] = _mm_add_epi16(res[4], const_rounding);
+    res[5] = _mm_add_epi16(res[5], const_rounding);
+    res[6] = _mm_add_epi16(res[6], const_rounding);
+    res[7] = _mm_add_epi16(res[7], const_rounding);
+  }
+
+  res[0] = _mm_sub_epi16(res[0], sign0);
+  res[1] = _mm_sub_epi16(res[1], sign1);
+  res[2] = _mm_sub_epi16(res[2], sign2);
+  res[3] = _mm_sub_epi16(res[3], sign3);
+  res[4] = _mm_sub_epi16(res[4], sign4);
+  res[5] = _mm_sub_epi16(res[5], sign5);
+  res[6] = _mm_sub_epi16(res[6], sign6);
+  res[7] = _mm_sub_epi16(res[7], sign7);
+
+  if (bit == 1) {
+    res[0] = _mm_srai_epi16(res[0], 1);
+    res[1] = _mm_srai_epi16(res[1], 1);
+    res[2] = _mm_srai_epi16(res[2], 1);
+    res[3] = _mm_srai_epi16(res[3], 1);
+    res[4] = _mm_srai_epi16(res[4], 1);
+    res[5] = _mm_srai_epi16(res[5], 1);
+    res[6] = _mm_srai_epi16(res[6], 1);
+    res[7] = _mm_srai_epi16(res[7], 1);
+  } else {
+    res[0] = _mm_srai_epi16(res[0], 2);
+    res[1] = _mm_srai_epi16(res[1], 2);
+    res[2] = _mm_srai_epi16(res[2], 2);
+    res[3] = _mm_srai_epi16(res[3], 2);
+    res[4] = _mm_srai_epi16(res[4], 2);
+    res[5] = _mm_srai_epi16(res[5], 2);
+    res[6] = _mm_srai_epi16(res[6], 2);
+    res[7] = _mm_srai_epi16(res[7], 2);
+  }
+}
+
+// write 8x8 array
+static INLINE void write_buffer_8x8(tran_low_t *output, __m128i *res,
+                                    int stride) {
+  store_output(&res[0], (output + 0 * stride));
+  store_output(&res[1], (output + 1 * stride));
+  store_output(&res[2], (output + 2 * stride));
+  store_output(&res[3], (output + 3 * stride));
+  store_output(&res[4], (output + 4 * stride));
+  store_output(&res[5], (output + 5 * stride));
+  store_output(&res[6], (output + 6 * stride));
+  store_output(&res[7], (output + 7 * stride));
+}
+
+// perform in-place transpose
+static INLINE void array_transpose_8x8(__m128i *in, __m128i *res) {
+  const __m128i tr0_0 = _mm_unpacklo_epi16(in[0], in[1]);
+  const __m128i tr0_1 = _mm_unpacklo_epi16(in[2], in[3]);
+  const __m128i tr0_2 = _mm_unpackhi_epi16(in[0], in[1]);
+  const __m128i tr0_3 = _mm_unpackhi_epi16(in[2], in[3]);
+  const __m128i tr0_4 = _mm_unpacklo_epi16(in[4], in[5]);
+  const __m128i tr0_5 = _mm_unpacklo_epi16(in[6], in[7]);
+  const __m128i tr0_6 = _mm_unpackhi_epi16(in[4], in[5]);
+  const __m128i tr0_7 = _mm_unpackhi_epi16(in[6], in[7]);
+  // 00 10 01 11 02 12 03 13
+  // 20 30 21 31 22 32 23 33
+  // 04 14 05 15 06 16 07 17
+  // 24 34 25 35 26 36 27 37
+  // 40 50 41 51 42 52 43 53
+  // 60 70 61 71 62 72 63 73
+  // 44 54 45 55 46 56 47 57
+  // 64 74 65 75 66 76 67 77
+  const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
+  const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_4, tr0_5);
+  const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
+  const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_4, tr0_5);
+  const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_2, tr0_3);
+  const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
+  const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_2, tr0_3);
+  const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
+  // 00 10 20 30 01 11 21 31
+  // 40 50 60 70 41 51 61 71
+  // 02 12 22 32 03 13 23 33
+  // 42 52 62 72 43 53 63 73
+  // 04 14 24 34 05 15 25 35
+  // 44 54 64 74 45 55 65 75
+  // 06 16 26 36 07 17 27 37
+  // 46 56 66 76 47 57 67 77
+  res[0] = _mm_unpacklo_epi64(tr1_0, tr1_1);
+  res[1] = _mm_unpackhi_epi64(tr1_0, tr1_1);
+  res[2] = _mm_unpacklo_epi64(tr1_2, tr1_3);
+  res[3] = _mm_unpackhi_epi64(tr1_2, tr1_3);
+  res[4] = _mm_unpacklo_epi64(tr1_4, tr1_5);
+  res[5] = _mm_unpackhi_epi64(tr1_4, tr1_5);
+  res[6] = _mm_unpacklo_epi64(tr1_6, tr1_7);
+  res[7] = _mm_unpackhi_epi64(tr1_6, tr1_7);
+  // 00 10 20 30 40 50 60 70
+  // 01 11 21 31 41 51 61 71
+  // 02 12 22 32 42 52 62 72
+  // 03 13 23 33 43 53 63 73
+  // 04 14 24 34 44 54 64 74
+  // 05 15 25 35 45 55 65 75
+  // 06 16 26 36 46 56 66 76
+  // 07 17 27 37 47 57 67 77
+}
+
+static void fdct8_sse2(__m128i *in) {
+  // constants
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64);
+  const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64);
+  const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  __m128i u0, u1, u2, u3, u4, u5, u6, u7;
+  __m128i v0, v1, v2, v3, v4, v5, v6, v7;
+  __m128i s0, s1, s2, s3, s4, s5, s6, s7;
+
+  // stage 1
+  s0 = _mm_add_epi16(in[0], in[7]);
+  s1 = _mm_add_epi16(in[1], in[6]);
+  s2 = _mm_add_epi16(in[2], in[5]);
+  s3 = _mm_add_epi16(in[3], in[4]);
+  s4 = _mm_sub_epi16(in[3], in[4]);
+  s5 = _mm_sub_epi16(in[2], in[5]);
+  s6 = _mm_sub_epi16(in[1], in[6]);
+  s7 = _mm_sub_epi16(in[0], in[7]);
+
+  u0 = _mm_add_epi16(s0, s3);
+  u1 = _mm_add_epi16(s1, s2);
+  u2 = _mm_sub_epi16(s1, s2);
+  u3 = _mm_sub_epi16(s0, s3);
+  // interleave and perform butterfly multiplication/addition
+  v0 = _mm_unpacklo_epi16(u0, u1);
+  v1 = _mm_unpackhi_epi16(u0, u1);
+  v2 = _mm_unpacklo_epi16(u2, u3);
+  v3 = _mm_unpackhi_epi16(u2, u3);
+
+  u0 = _mm_madd_epi16(v0, k__cospi_p16_p16);
+  u1 = _mm_madd_epi16(v1, k__cospi_p16_p16);
+  u2 = _mm_madd_epi16(v0, k__cospi_p16_m16);
+  u3 = _mm_madd_epi16(v1, k__cospi_p16_m16);
+  u4 = _mm_madd_epi16(v2, k__cospi_p24_p08);
+  u5 = _mm_madd_epi16(v3, k__cospi_p24_p08);
+  u6 = _mm_madd_epi16(v2, k__cospi_m08_p24);
+  u7 = _mm_madd_epi16(v3, k__cospi_m08_p24);
+
+  // shift and rounding
+  v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
+  v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
+  v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
+  v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
+  v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
+  v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
+  v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
+  v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
+
+  u0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+  u1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+  u2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+  u3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+  u4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
+  u5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
+  u6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
+  u7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
+
+  in[0] = _mm_packs_epi32(u0, u1);
+  in[2] = _mm_packs_epi32(u4, u5);
+  in[4] = _mm_packs_epi32(u2, u3);
+  in[6] = _mm_packs_epi32(u6, u7);
+
+  // stage 2
+  // interleave and perform butterfly multiplication/addition
+  u0 = _mm_unpacklo_epi16(s6, s5);
+  u1 = _mm_unpackhi_epi16(s6, s5);
+  v0 = _mm_madd_epi16(u0, k__cospi_p16_m16);
+  v1 = _mm_madd_epi16(u1, k__cospi_p16_m16);
+  v2 = _mm_madd_epi16(u0, k__cospi_p16_p16);
+  v3 = _mm_madd_epi16(u1, k__cospi_p16_p16);
+
+  // shift and rounding
+  u0 = _mm_add_epi32(v0, k__DCT_CONST_ROUNDING);
+  u1 = _mm_add_epi32(v1, k__DCT_CONST_ROUNDING);
+  u2 = _mm_add_epi32(v2, k__DCT_CONST_ROUNDING);
+  u3 = _mm_add_epi32(v3, k__DCT_CONST_ROUNDING);
+
+  v0 = _mm_srai_epi32(u0, DCT_CONST_BITS);
+  v1 = _mm_srai_epi32(u1, DCT_CONST_BITS);
+  v2 = _mm_srai_epi32(u2, DCT_CONST_BITS);
+  v3 = _mm_srai_epi32(u3, DCT_CONST_BITS);
+
+  u0 = _mm_packs_epi32(v0, v1);
+  u1 = _mm_packs_epi32(v2, v3);
+
+  // stage 3
+  s0 = _mm_add_epi16(s4, u0);
+  s1 = _mm_sub_epi16(s4, u0);
+  s2 = _mm_sub_epi16(s7, u1);
+  s3 = _mm_add_epi16(s7, u1);
+
+  // stage 4
+  u0 = _mm_unpacklo_epi16(s0, s3);
+  u1 = _mm_unpackhi_epi16(s0, s3);
+  u2 = _mm_unpacklo_epi16(s1, s2);
+  u3 = _mm_unpackhi_epi16(s1, s2);
+
+  v0 = _mm_madd_epi16(u0, k__cospi_p28_p04);
+  v1 = _mm_madd_epi16(u1, k__cospi_p28_p04);
+  v2 = _mm_madd_epi16(u2, k__cospi_p12_p20);
+  v3 = _mm_madd_epi16(u3, k__cospi_p12_p20);
+  v4 = _mm_madd_epi16(u2, k__cospi_m20_p12);
+  v5 = _mm_madd_epi16(u3, k__cospi_m20_p12);
+  v6 = _mm_madd_epi16(u0, k__cospi_m04_p28);
+  v7 = _mm_madd_epi16(u1, k__cospi_m04_p28);
+
+  // shift and rounding
+  u0 = _mm_add_epi32(v0, k__DCT_CONST_ROUNDING);
+  u1 = _mm_add_epi32(v1, k__DCT_CONST_ROUNDING);
+  u2 = _mm_add_epi32(v2, k__DCT_CONST_ROUNDING);
+  u3 = _mm_add_epi32(v3, k__DCT_CONST_ROUNDING);
+  u4 = _mm_add_epi32(v4, k__DCT_CONST_ROUNDING);
+  u5 = _mm_add_epi32(v5, k__DCT_CONST_ROUNDING);
+  u6 = _mm_add_epi32(v6, k__DCT_CONST_ROUNDING);
+  u7 = _mm_add_epi32(v7, k__DCT_CONST_ROUNDING);
+
+  v0 = _mm_srai_epi32(u0, DCT_CONST_BITS);
+  v1 = _mm_srai_epi32(u1, DCT_CONST_BITS);
+  v2 = _mm_srai_epi32(u2, DCT_CONST_BITS);
+  v3 = _mm_srai_epi32(u3, DCT_CONST_BITS);
+  v4 = _mm_srai_epi32(u4, DCT_CONST_BITS);
+  v5 = _mm_srai_epi32(u5, DCT_CONST_BITS);
+  v6 = _mm_srai_epi32(u6, DCT_CONST_BITS);
+  v7 = _mm_srai_epi32(u7, DCT_CONST_BITS);
+
+  in[1] = _mm_packs_epi32(v0, v1);
+  in[3] = _mm_packs_epi32(v4, v5);
+  in[5] = _mm_packs_epi32(v2, v3);
+  in[7] = _mm_packs_epi32(v6, v7);
+
+  // transpose
+  array_transpose_8x8(in, in);
+}
+
+static void fadst8_sse2(__m128i *in) {
+  // Constants
+  const __m128i k__cospi_p02_p30 = pair_set_epi16(cospi_2_64, cospi_30_64);
+  const __m128i k__cospi_p30_m02 = pair_set_epi16(cospi_30_64, -cospi_2_64);
+  const __m128i k__cospi_p10_p22 = pair_set_epi16(cospi_10_64, cospi_22_64);
+  const __m128i k__cospi_p22_m10 = pair_set_epi16(cospi_22_64, -cospi_10_64);
+  const __m128i k__cospi_p18_p14 = pair_set_epi16(cospi_18_64, cospi_14_64);
+  const __m128i k__cospi_p14_m18 = pair_set_epi16(cospi_14_64, -cospi_18_64);
+  const __m128i k__cospi_p26_p06 = pair_set_epi16(cospi_26_64, cospi_6_64);
+  const __m128i k__cospi_p06_m26 = pair_set_epi16(cospi_6_64, -cospi_26_64);
+  const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i k__cospi_m24_p08 = pair_set_epi16(-cospi_24_64, cospi_8_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__const_0 = _mm_set1_epi16(0);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+
+  __m128i u0, u1, u2, u3, u4, u5, u6, u7, u8, u9, u10, u11, u12, u13, u14, u15;
+  __m128i v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15;
+  __m128i w0, w1, w2, w3, w4, w5, w6, w7, w8, w9, w10, w11, w12, w13, w14, w15;
+  __m128i s0, s1, s2, s3, s4, s5, s6, s7;
+  __m128i in0, in1, in2, in3, in4, in5, in6, in7;
+
+  // properly aligned for butterfly input
+  in0  = in[7];
+  in1  = in[0];
+  in2  = in[5];
+  in3  = in[2];
+  in4  = in[3];
+  in5  = in[4];
+  in6  = in[1];
+  in7  = in[6];
+
+  // column transformation
+  // stage 1
+  // interleave and multiply/add into 32-bit integer
+  s0 = _mm_unpacklo_epi16(in0, in1);
+  s1 = _mm_unpackhi_epi16(in0, in1);
+  s2 = _mm_unpacklo_epi16(in2, in3);
+  s3 = _mm_unpackhi_epi16(in2, in3);
+  s4 = _mm_unpacklo_epi16(in4, in5);
+  s5 = _mm_unpackhi_epi16(in4, in5);
+  s6 = _mm_unpacklo_epi16(in6, in7);
+  s7 = _mm_unpackhi_epi16(in6, in7);
+
+  u0 = _mm_madd_epi16(s0, k__cospi_p02_p30);
+  u1 = _mm_madd_epi16(s1, k__cospi_p02_p30);
+  u2 = _mm_madd_epi16(s0, k__cospi_p30_m02);
+  u3 = _mm_madd_epi16(s1, k__cospi_p30_m02);
+  u4 = _mm_madd_epi16(s2, k__cospi_p10_p22);
+  u5 = _mm_madd_epi16(s3, k__cospi_p10_p22);
+  u6 = _mm_madd_epi16(s2, k__cospi_p22_m10);
+  u7 = _mm_madd_epi16(s3, k__cospi_p22_m10);
+  u8 = _mm_madd_epi16(s4, k__cospi_p18_p14);
+  u9 = _mm_madd_epi16(s5, k__cospi_p18_p14);
+  u10 = _mm_madd_epi16(s4, k__cospi_p14_m18);
+  u11 = _mm_madd_epi16(s5, k__cospi_p14_m18);
+  u12 = _mm_madd_epi16(s6, k__cospi_p26_p06);
+  u13 = _mm_madd_epi16(s7, k__cospi_p26_p06);
+  u14 = _mm_madd_epi16(s6, k__cospi_p06_m26);
+  u15 = _mm_madd_epi16(s7, k__cospi_p06_m26);
+
+  // addition
+  w0 = _mm_add_epi32(u0, u8);
+  w1 = _mm_add_epi32(u1, u9);
+  w2 = _mm_add_epi32(u2, u10);
+  w3 = _mm_add_epi32(u3, u11);
+  w4 = _mm_add_epi32(u4, u12);
+  w5 = _mm_add_epi32(u5, u13);
+  w6 = _mm_add_epi32(u6, u14);
+  w7 = _mm_add_epi32(u7, u15);
+  w8 = _mm_sub_epi32(u0, u8);
+  w9 = _mm_sub_epi32(u1, u9);
+  w10 = _mm_sub_epi32(u2, u10);
+  w11 = _mm_sub_epi32(u3, u11);
+  w12 = _mm_sub_epi32(u4, u12);
+  w13 = _mm_sub_epi32(u5, u13);
+  w14 = _mm_sub_epi32(u6, u14);
+  w15 = _mm_sub_epi32(u7, u15);
+
+  // shift and rounding
+  v0 = _mm_add_epi32(w0, k__DCT_CONST_ROUNDING);
+  v1 = _mm_add_epi32(w1, k__DCT_CONST_ROUNDING);
+  v2 = _mm_add_epi32(w2, k__DCT_CONST_ROUNDING);
+  v3 = _mm_add_epi32(w3, k__DCT_CONST_ROUNDING);
+  v4 = _mm_add_epi32(w4, k__DCT_CONST_ROUNDING);
+  v5 = _mm_add_epi32(w5, k__DCT_CONST_ROUNDING);
+  v6 = _mm_add_epi32(w6, k__DCT_CONST_ROUNDING);
+  v7 = _mm_add_epi32(w7, k__DCT_CONST_ROUNDING);
+  v8 = _mm_add_epi32(w8, k__DCT_CONST_ROUNDING);
+  v9 = _mm_add_epi32(w9, k__DCT_CONST_ROUNDING);
+  v10 = _mm_add_epi32(w10, k__DCT_CONST_ROUNDING);
+  v11 = _mm_add_epi32(w11, k__DCT_CONST_ROUNDING);
+  v12 = _mm_add_epi32(w12, k__DCT_CONST_ROUNDING);
+  v13 = _mm_add_epi32(w13, k__DCT_CONST_ROUNDING);
+  v14 = _mm_add_epi32(w14, k__DCT_CONST_ROUNDING);
+  v15 = _mm_add_epi32(w15, k__DCT_CONST_ROUNDING);
+
+  u0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+  u1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+  u2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+  u3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+  u4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
+  u5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
+  u6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
+  u7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
+  u8 = _mm_srai_epi32(v8, DCT_CONST_BITS);
+  u9 = _mm_srai_epi32(v9, DCT_CONST_BITS);
+  u10 = _mm_srai_epi32(v10, DCT_CONST_BITS);
+  u11 = _mm_srai_epi32(v11, DCT_CONST_BITS);
+  u12 = _mm_srai_epi32(v12, DCT_CONST_BITS);
+  u13 = _mm_srai_epi32(v13, DCT_CONST_BITS);
+  u14 = _mm_srai_epi32(v14, DCT_CONST_BITS);
+  u15 = _mm_srai_epi32(v15, DCT_CONST_BITS);
+
+  // back to 16-bit and pack 8 integers into __m128i
+  in[0] = _mm_packs_epi32(u0, u1);
+  in[1] = _mm_packs_epi32(u2, u3);
+  in[2] = _mm_packs_epi32(u4, u5);
+  in[3] = _mm_packs_epi32(u6, u7);
+  in[4] = _mm_packs_epi32(u8, u9);
+  in[5] = _mm_packs_epi32(u10, u11);
+  in[6] = _mm_packs_epi32(u12, u13);
+  in[7] = _mm_packs_epi32(u14, u15);
+
+  // stage 2
+  s0 = _mm_add_epi16(in[0], in[2]);
+  s1 = _mm_add_epi16(in[1], in[3]);
+  s2 = _mm_sub_epi16(in[0], in[2]);
+  s3 = _mm_sub_epi16(in[1], in[3]);
+  u0 = _mm_unpacklo_epi16(in[4], in[5]);
+  u1 = _mm_unpackhi_epi16(in[4], in[5]);
+  u2 = _mm_unpacklo_epi16(in[6], in[7]);
+  u3 = _mm_unpackhi_epi16(in[6], in[7]);
+
+  v0 = _mm_madd_epi16(u0, k__cospi_p08_p24);
+  v1 = _mm_madd_epi16(u1, k__cospi_p08_p24);
+  v2 = _mm_madd_epi16(u0, k__cospi_p24_m08);
+  v3 = _mm_madd_epi16(u1, k__cospi_p24_m08);
+  v4 = _mm_madd_epi16(u2, k__cospi_m24_p08);
+  v5 = _mm_madd_epi16(u3, k__cospi_m24_p08);
+  v6 = _mm_madd_epi16(u2, k__cospi_p08_p24);
+  v7 = _mm_madd_epi16(u3, k__cospi_p08_p24);
+
+  w0 = _mm_add_epi32(v0, v4);
+  w1 = _mm_add_epi32(v1, v5);
+  w2 = _mm_add_epi32(v2, v6);
+  w3 = _mm_add_epi32(v3, v7);
+  w4 = _mm_sub_epi32(v0, v4);
+  w5 = _mm_sub_epi32(v1, v5);
+  w6 = _mm_sub_epi32(v2, v6);
+  w7 = _mm_sub_epi32(v3, v7);
+
+  v0 = _mm_add_epi32(w0, k__DCT_CONST_ROUNDING);
+  v1 = _mm_add_epi32(w1, k__DCT_CONST_ROUNDING);
+  v2 = _mm_add_epi32(w2, k__DCT_CONST_ROUNDING);
+  v3 = _mm_add_epi32(w3, k__DCT_CONST_ROUNDING);
+  v4 = _mm_add_epi32(w4, k__DCT_CONST_ROUNDING);
+  v5 = _mm_add_epi32(w5, k__DCT_CONST_ROUNDING);
+  v6 = _mm_add_epi32(w6, k__DCT_CONST_ROUNDING);
+  v7 = _mm_add_epi32(w7, k__DCT_CONST_ROUNDING);
+
+  u0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+  u1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+  u2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+  u3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+  u4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
+  u5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
+  u6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
+  u7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
+
+  // back to 16-bit intergers
+  s4 = _mm_packs_epi32(u0, u1);
+  s5 = _mm_packs_epi32(u2, u3);
+  s6 = _mm_packs_epi32(u4, u5);
+  s7 = _mm_packs_epi32(u6, u7);
+
+  // stage 3
+  u0 = _mm_unpacklo_epi16(s2, s3);
+  u1 = _mm_unpackhi_epi16(s2, s3);
+  u2 = _mm_unpacklo_epi16(s6, s7);
+  u3 = _mm_unpackhi_epi16(s6, s7);
+
+  v0 = _mm_madd_epi16(u0, k__cospi_p16_p16);
+  v1 = _mm_madd_epi16(u1, k__cospi_p16_p16);
+  v2 = _mm_madd_epi16(u0, k__cospi_p16_m16);
+  v3 = _mm_madd_epi16(u1, k__cospi_p16_m16);
+  v4 = _mm_madd_epi16(u2, k__cospi_p16_p16);
+  v5 = _mm_madd_epi16(u3, k__cospi_p16_p16);
+  v6 = _mm_madd_epi16(u2, k__cospi_p16_m16);
+  v7 = _mm_madd_epi16(u3, k__cospi_p16_m16);
+
+  u0 = _mm_add_epi32(v0, k__DCT_CONST_ROUNDING);
+  u1 = _mm_add_epi32(v1, k__DCT_CONST_ROUNDING);
+  u2 = _mm_add_epi32(v2, k__DCT_CONST_ROUNDING);
+  u3 = _mm_add_epi32(v3, k__DCT_CONST_ROUNDING);
+  u4 = _mm_add_epi32(v4, k__DCT_CONST_ROUNDING);
+  u5 = _mm_add_epi32(v5, k__DCT_CONST_ROUNDING);
+  u6 = _mm_add_epi32(v6, k__DCT_CONST_ROUNDING);
+  u7 = _mm_add_epi32(v7, k__DCT_CONST_ROUNDING);
+
+  v0 = _mm_srai_epi32(u0, DCT_CONST_BITS);
+  v1 = _mm_srai_epi32(u1, DCT_CONST_BITS);
+  v2 = _mm_srai_epi32(u2, DCT_CONST_BITS);
+  v3 = _mm_srai_epi32(u3, DCT_CONST_BITS);
+  v4 = _mm_srai_epi32(u4, DCT_CONST_BITS);
+  v5 = _mm_srai_epi32(u5, DCT_CONST_BITS);
+  v6 = _mm_srai_epi32(u6, DCT_CONST_BITS);
+  v7 = _mm_srai_epi32(u7, DCT_CONST_BITS);
+
+  s2 = _mm_packs_epi32(v0, v1);
+  s3 = _mm_packs_epi32(v2, v3);
+  s6 = _mm_packs_epi32(v4, v5);
+  s7 = _mm_packs_epi32(v6, v7);
+
+  // FIXME(jingning): do subtract using bit inversion?
+  in[0] = s0;
+  in[1] = _mm_sub_epi16(k__const_0, s4);
+  in[2] = s6;
+  in[3] = _mm_sub_epi16(k__const_0, s2);
+  in[4] = s3;
+  in[5] = _mm_sub_epi16(k__const_0, s7);
+  in[6] = s5;
+  in[7] = _mm_sub_epi16(k__const_0, s1);
+
+  // transpose
+  array_transpose_8x8(in, in);
+}
+
+void vp9_fht8x8_sse2(const int16_t *input, tran_low_t *output,
+                     int stride, int tx_type) {
+  __m128i in[8];
+
+  switch (tx_type) {
+    case DCT_DCT:
+      vp9_fdct8x8_sse2(input, output, stride);
+      break;
+    case ADST_DCT:
+      load_buffer_8x8(input, in, stride);
+      fadst8_sse2(in);
+      fdct8_sse2(in);
+      right_shift_8x8(in, 1);
+      write_buffer_8x8(output, in, 8);
+      break;
+    case DCT_ADST:
+      load_buffer_8x8(input, in, stride);
+      fdct8_sse2(in);
+      fadst8_sse2(in);
+      right_shift_8x8(in, 1);
+      write_buffer_8x8(output, in, 8);
+      break;
+    case ADST_ADST:
+      load_buffer_8x8(input, in, stride);
+      fadst8_sse2(in);
+      fadst8_sse2(in);
+      right_shift_8x8(in, 1);
+      write_buffer_8x8(output, in, 8);
+      break;
+    default:
+      assert(0);
+      break;
+  }
+}
+
+void vp9_fdct16x16_1_sse2(const int16_t *input, tran_low_t *output,
+                          int stride) {
+  __m128i in0, in1, in2, in3;
+  __m128i u0, u1;
+  __m128i sum = _mm_setzero_si128();
+  int i;
+
+  for (i = 0; i < 2; ++i) {
+    input += 8 * i;
+    in0  = _mm_load_si128((const __m128i *)(input +  0 * stride));
+    in1  = _mm_load_si128((const __m128i *)(input +  1 * stride));
+    in2  = _mm_load_si128((const __m128i *)(input +  2 * stride));
+    in3  = _mm_load_si128((const __m128i *)(input +  3 * stride));
+
+    u0 = _mm_add_epi16(in0, in1);
+    u1 = _mm_add_epi16(in2, in3);
+    sum = _mm_add_epi16(sum, u0);
+
+    in0  = _mm_load_si128((const __m128i *)(input +  4 * stride));
+    in1  = _mm_load_si128((const __m128i *)(input +  5 * stride));
+    in2  = _mm_load_si128((const __m128i *)(input +  6 * stride));
+    in3  = _mm_load_si128((const __m128i *)(input +  7 * stride));
+
+    sum = _mm_add_epi16(sum, u1);
+    u0  = _mm_add_epi16(in0, in1);
+    u1  = _mm_add_epi16(in2, in3);
+    sum = _mm_add_epi16(sum, u0);
+
+    in0  = _mm_load_si128((const __m128i *)(input +  8 * stride));
+    in1  = _mm_load_si128((const __m128i *)(input +  9 * stride));
+    in2  = _mm_load_si128((const __m128i *)(input + 10 * stride));
+    in3  = _mm_load_si128((const __m128i *)(input + 11 * stride));
+
+    sum = _mm_add_epi16(sum, u1);
+    u0  = _mm_add_epi16(in0, in1);
+    u1  = _mm_add_epi16(in2, in3);
+    sum = _mm_add_epi16(sum, u0);
+
+    in0  = _mm_load_si128((const __m128i *)(input + 12 * stride));
+    in1  = _mm_load_si128((const __m128i *)(input + 13 * stride));
+    in2  = _mm_load_si128((const __m128i *)(input + 14 * stride));
+    in3  = _mm_load_si128((const __m128i *)(input + 15 * stride));
+
+    sum = _mm_add_epi16(sum, u1);
+    u0  = _mm_add_epi16(in0, in1);
+    u1  = _mm_add_epi16(in2, in3);
+    sum = _mm_add_epi16(sum, u0);
+
+    sum = _mm_add_epi16(sum, u1);
+  }
+
+  u0  = _mm_setzero_si128();
+  in0 = _mm_unpacklo_epi16(u0, sum);
+  in1 = _mm_unpackhi_epi16(u0, sum);
+  in0 = _mm_srai_epi32(in0, 16);
+  in1 = _mm_srai_epi32(in1, 16);
+
+  sum = _mm_add_epi32(in0, in1);
+  in0 = _mm_unpacklo_epi32(sum, u0);
+  in1 = _mm_unpackhi_epi32(sum, u0);
+
+  sum = _mm_add_epi32(in0, in1);
+  in0 = _mm_srli_si128(sum, 8);
+
+  in1 = _mm_add_epi32(sum, in0);
+  in1 = _mm_srai_epi32(in1, 1);
+  store_output(&in1, output);
+}
+
+static INLINE void load_buffer_16x16(const int16_t* input, __m128i *in0,
+                                     __m128i *in1, int stride) {
+  // load first 8 columns
+  load_buffer_8x8(input, in0, stride);
+  load_buffer_8x8(input + 8 * stride, in0 + 8, stride);
+
+  input += 8;
+  // load second 8 columns
+  load_buffer_8x8(input, in1, stride);
+  load_buffer_8x8(input + 8 * stride, in1 + 8, stride);
+}
+
+static INLINE void write_buffer_16x16(tran_low_t *output, __m128i *in0,
+                                      __m128i *in1, int stride) {
+  // write first 8 columns
+  write_buffer_8x8(output, in0, stride);
+  write_buffer_8x8(output + 8 * stride, in0 + 8, stride);
+  // write second 8 columns
+  output += 8;
+  write_buffer_8x8(output, in1, stride);
+  write_buffer_8x8(output + 8 * stride, in1 + 8, stride);
+}
+
+static INLINE void array_transpose_16x16(__m128i *res0, __m128i *res1) {
+  __m128i tbuf[8];
+  array_transpose_8x8(res0, res0);
+  array_transpose_8x8(res1, tbuf);
+  array_transpose_8x8(res0 + 8, res1);
+  array_transpose_8x8(res1 + 8, res1 + 8);
+
+  res0[8] = tbuf[0];
+  res0[9] = tbuf[1];
+  res0[10] = tbuf[2];
+  res0[11] = tbuf[3];
+  res0[12] = tbuf[4];
+  res0[13] = tbuf[5];
+  res0[14] = tbuf[6];
+  res0[15] = tbuf[7];
+}
+
+static INLINE void right_shift_16x16(__m128i *res0, __m128i *res1) {
+  // perform rounding operations
+  right_shift_8x8(res0, 2);
+  right_shift_8x8(res0 + 8, 2);
+  right_shift_8x8(res1, 2);
+  right_shift_8x8(res1 + 8, 2);
+}
+
+static void fdct16_8col(__m128i *in) {
+  // perform 16x16 1-D DCT for 8 columns
+  __m128i i[8], s[8], p[8], t[8], u[16], v[16];
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_m16_p16 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+  const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i k__cospi_p08_m24 = pair_set_epi16(cospi_8_64, -cospi_24_64);
+  const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64);
+  const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64);
+  const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i k__cospi_p30_p02 = pair_set_epi16(cospi_30_64, cospi_2_64);
+  const __m128i k__cospi_p14_p18 = pair_set_epi16(cospi_14_64, cospi_18_64);
+  const __m128i k__cospi_m02_p30 = pair_set_epi16(-cospi_2_64, cospi_30_64);
+  const __m128i k__cospi_m18_p14 = pair_set_epi16(-cospi_18_64, cospi_14_64);
+  const __m128i k__cospi_p22_p10 = pair_set_epi16(cospi_22_64, cospi_10_64);
+  const __m128i k__cospi_p06_p26 = pair_set_epi16(cospi_6_64, cospi_26_64);
+  const __m128i k__cospi_m10_p22 = pair_set_epi16(-cospi_10_64, cospi_22_64);
+  const __m128i k__cospi_m26_p06 = pair_set_epi16(-cospi_26_64, cospi_6_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+
+  // stage 1
+  i[0] = _mm_add_epi16(in[0], in[15]);
+  i[1] = _mm_add_epi16(in[1], in[14]);
+  i[2] = _mm_add_epi16(in[2], in[13]);
+  i[3] = _mm_add_epi16(in[3], in[12]);
+  i[4] = _mm_add_epi16(in[4], in[11]);
+  i[5] = _mm_add_epi16(in[5], in[10]);
+  i[6] = _mm_add_epi16(in[6], in[9]);
+  i[7] = _mm_add_epi16(in[7], in[8]);
+
+  s[0] = _mm_sub_epi16(in[7], in[8]);
+  s[1] = _mm_sub_epi16(in[6], in[9]);
+  s[2] = _mm_sub_epi16(in[5], in[10]);
+  s[3] = _mm_sub_epi16(in[4], in[11]);
+  s[4] = _mm_sub_epi16(in[3], in[12]);
+  s[5] = _mm_sub_epi16(in[2], in[13]);
+  s[6] = _mm_sub_epi16(in[1], in[14]);
+  s[7] = _mm_sub_epi16(in[0], in[15]);
+
+  p[0] = _mm_add_epi16(i[0], i[7]);
+  p[1] = _mm_add_epi16(i[1], i[6]);
+  p[2] = _mm_add_epi16(i[2], i[5]);
+  p[3] = _mm_add_epi16(i[3], i[4]);
+  p[4] = _mm_sub_epi16(i[3], i[4]);
+  p[5] = _mm_sub_epi16(i[2], i[5]);
+  p[6] = _mm_sub_epi16(i[1], i[6]);
+  p[7] = _mm_sub_epi16(i[0], i[7]);
+
+  u[0] = _mm_add_epi16(p[0], p[3]);
+  u[1] = _mm_add_epi16(p[1], p[2]);
+  u[2] = _mm_sub_epi16(p[1], p[2]);
+  u[3] = _mm_sub_epi16(p[0], p[3]);
+
+  v[0] = _mm_unpacklo_epi16(u[0], u[1]);
+  v[1] = _mm_unpackhi_epi16(u[0], u[1]);
+  v[2] = _mm_unpacklo_epi16(u[2], u[3]);
+  v[3] = _mm_unpackhi_epi16(u[2], u[3]);
+
+  u[0] = _mm_madd_epi16(v[0], k__cospi_p16_p16);
+  u[1] = _mm_madd_epi16(v[1], k__cospi_p16_p16);
+  u[2] = _mm_madd_epi16(v[0], k__cospi_p16_m16);
+  u[3] = _mm_madd_epi16(v[1], k__cospi_p16_m16);
+  u[4] = _mm_madd_epi16(v[2], k__cospi_p24_p08);
+  u[5] = _mm_madd_epi16(v[3], k__cospi_p24_p08);
+  u[6] = _mm_madd_epi16(v[2], k__cospi_m08_p24);
+  u[7] = _mm_madd_epi16(v[3], k__cospi_m08_p24);
+
+  v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+  v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+  v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+  v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+  v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+  v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+  v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+  v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+
+  u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
+  u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
+  u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
+  u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
+  u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
+
+  in[0] = _mm_packs_epi32(u[0], u[1]);
+  in[4] = _mm_packs_epi32(u[4], u[5]);
+  in[8] = _mm_packs_epi32(u[2], u[3]);
+  in[12] = _mm_packs_epi32(u[6], u[7]);
+
+  u[0] = _mm_unpacklo_epi16(p[5], p[6]);
+  u[1] = _mm_unpackhi_epi16(p[5], p[6]);
+  v[0] = _mm_madd_epi16(u[0], k__cospi_m16_p16);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_m16_p16);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p16_p16);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p16_p16);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+
+  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+
+  u[0] = _mm_packs_epi32(v[0], v[1]);
+  u[1] = _mm_packs_epi32(v[2], v[3]);
+
+  t[0] = _mm_add_epi16(p[4], u[0]);
+  t[1] = _mm_sub_epi16(p[4], u[0]);
+  t[2] = _mm_sub_epi16(p[7], u[1]);
+  t[3] = _mm_add_epi16(p[7], u[1]);
+
+  u[0] = _mm_unpacklo_epi16(t[0], t[3]);
+  u[1] = _mm_unpackhi_epi16(t[0], t[3]);
+  u[2] = _mm_unpacklo_epi16(t[1], t[2]);
+  u[3] = _mm_unpackhi_epi16(t[1], t[2]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p28_p04);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_p28_p04);
+  v[2] = _mm_madd_epi16(u[2], k__cospi_p12_p20);
+  v[3] = _mm_madd_epi16(u[3], k__cospi_p12_p20);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_m20_p12);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_m20_p12);
+  v[6] = _mm_madd_epi16(u[0], k__cospi_m04_p28);
+  v[7] = _mm_madd_epi16(u[1], k__cospi_m04_p28);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
+
+  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+
+  in[2] = _mm_packs_epi32(v[0], v[1]);
+  in[6] = _mm_packs_epi32(v[4], v[5]);
+  in[10] = _mm_packs_epi32(v[2], v[3]);
+  in[14] = _mm_packs_epi32(v[6], v[7]);
+
+  // stage 2
+  u[0] = _mm_unpacklo_epi16(s[2], s[5]);
+  u[1] = _mm_unpackhi_epi16(s[2], s[5]);
+  u[2] = _mm_unpacklo_epi16(s[3], s[4]);
+  u[3] = _mm_unpackhi_epi16(s[3], s[4]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_m16_p16);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_m16_p16);
+  v[2] = _mm_madd_epi16(u[2], k__cospi_m16_p16);
+  v[3] = _mm_madd_epi16(u[3], k__cospi_m16_p16);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p16_p16);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p16_p16);
+  v[6] = _mm_madd_epi16(u[0], k__cospi_p16_p16);
+  v[7] = _mm_madd_epi16(u[1], k__cospi_p16_p16);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
+
+  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+
+  t[2] = _mm_packs_epi32(v[0], v[1]);
+  t[3] = _mm_packs_epi32(v[2], v[3]);
+  t[4] = _mm_packs_epi32(v[4], v[5]);
+  t[5] = _mm_packs_epi32(v[6], v[7]);
+
+  // stage 3
+  p[0] = _mm_add_epi16(s[0], t[3]);
+  p[1] = _mm_add_epi16(s[1], t[2]);
+  p[2] = _mm_sub_epi16(s[1], t[2]);
+  p[3] = _mm_sub_epi16(s[0], t[3]);
+  p[4] = _mm_sub_epi16(s[7], t[4]);
+  p[5] = _mm_sub_epi16(s[6], t[5]);
+  p[6] = _mm_add_epi16(s[6], t[5]);
+  p[7] = _mm_add_epi16(s[7], t[4]);
+
+  // stage 4
+  u[0] = _mm_unpacklo_epi16(p[1], p[6]);
+  u[1] = _mm_unpackhi_epi16(p[1], p[6]);
+  u[2] = _mm_unpacklo_epi16(p[2], p[5]);
+  u[3] = _mm_unpackhi_epi16(p[2], p[5]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_m08_p24);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_m08_p24);
+  v[2] = _mm_madd_epi16(u[2], k__cospi_p24_p08);
+  v[3] = _mm_madd_epi16(u[3], k__cospi_p24_p08);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p08_m24);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p08_m24);
+  v[6] = _mm_madd_epi16(u[0], k__cospi_p24_p08);
+  v[7] = _mm_madd_epi16(u[1], k__cospi_p24_p08);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
+
+  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+
+  t[1] = _mm_packs_epi32(v[0], v[1]);
+  t[2] = _mm_packs_epi32(v[2], v[3]);
+  t[5] = _mm_packs_epi32(v[4], v[5]);
+  t[6] = _mm_packs_epi32(v[6], v[7]);
+
+  // stage 5
+  s[0] = _mm_add_epi16(p[0], t[1]);
+  s[1] = _mm_sub_epi16(p[0], t[1]);
+  s[2] = _mm_add_epi16(p[3], t[2]);
+  s[3] = _mm_sub_epi16(p[3], t[2]);
+  s[4] = _mm_sub_epi16(p[4], t[5]);
+  s[5] = _mm_add_epi16(p[4], t[5]);
+  s[6] = _mm_sub_epi16(p[7], t[6]);
+  s[7] = _mm_add_epi16(p[7], t[6]);
+
+  // stage 6
+  u[0] = _mm_unpacklo_epi16(s[0], s[7]);
+  u[1] = _mm_unpackhi_epi16(s[0], s[7]);
+  u[2] = _mm_unpacklo_epi16(s[1], s[6]);
+  u[3] = _mm_unpackhi_epi16(s[1], s[6]);
+  u[4] = _mm_unpacklo_epi16(s[2], s[5]);
+  u[5] = _mm_unpackhi_epi16(s[2], s[5]);
+  u[6] = _mm_unpacklo_epi16(s[3], s[4]);
+  u[7] = _mm_unpackhi_epi16(s[3], s[4]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p30_p02);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_p30_p02);
+  v[2] = _mm_madd_epi16(u[2], k__cospi_p14_p18);
+  v[3] = _mm_madd_epi16(u[3], k__cospi_p14_p18);
+  v[4] = _mm_madd_epi16(u[4], k__cospi_p22_p10);
+  v[5] = _mm_madd_epi16(u[5], k__cospi_p22_p10);
+  v[6] = _mm_madd_epi16(u[6], k__cospi_p06_p26);
+  v[7] = _mm_madd_epi16(u[7], k__cospi_p06_p26);
+  v[8] = _mm_madd_epi16(u[6], k__cospi_m26_p06);
+  v[9] = _mm_madd_epi16(u[7], k__cospi_m26_p06);
+  v[10] = _mm_madd_epi16(u[4], k__cospi_m10_p22);
+  v[11] = _mm_madd_epi16(u[5], k__cospi_m10_p22);
+  v[12] = _mm_madd_epi16(u[2], k__cospi_m18_p14);
+  v[13] = _mm_madd_epi16(u[3], k__cospi_m18_p14);
+  v[14] = _mm_madd_epi16(u[0], k__cospi_m02_p30);
+  v[15] = _mm_madd_epi16(u[1], k__cospi_m02_p30);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
+  u[8] = _mm_add_epi32(v[8], k__DCT_CONST_ROUNDING);
+  u[9] = _mm_add_epi32(v[9], k__DCT_CONST_ROUNDING);
+  u[10] = _mm_add_epi32(v[10], k__DCT_CONST_ROUNDING);
+  u[11] = _mm_add_epi32(v[11], k__DCT_CONST_ROUNDING);
+  u[12] = _mm_add_epi32(v[12], k__DCT_CONST_ROUNDING);
+  u[13] = _mm_add_epi32(v[13], k__DCT_CONST_ROUNDING);
+  u[14] = _mm_add_epi32(v[14], k__DCT_CONST_ROUNDING);
+  u[15] = _mm_add_epi32(v[15], k__DCT_CONST_ROUNDING);
+
+  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+  v[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS);
+  v[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS);
+  v[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS);
+  v[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS);
+  v[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS);
+  v[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS);
+  v[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS);
+  v[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS);
+
+  in[1]  = _mm_packs_epi32(v[0], v[1]);
+  in[9]  = _mm_packs_epi32(v[2], v[3]);
+  in[5]  = _mm_packs_epi32(v[4], v[5]);
+  in[13] = _mm_packs_epi32(v[6], v[7]);
+  in[3]  = _mm_packs_epi32(v[8], v[9]);
+  in[11] = _mm_packs_epi32(v[10], v[11]);
+  in[7]  = _mm_packs_epi32(v[12], v[13]);
+  in[15] = _mm_packs_epi32(v[14], v[15]);
+}
+
+static void fadst16_8col(__m128i *in) {
+  // perform 16x16 1-D ADST for 8 columns
+  __m128i s[16], x[16], u[32], v[32];
+  const __m128i k__cospi_p01_p31 = pair_set_epi16(cospi_1_64, cospi_31_64);
+  const __m128i k__cospi_p31_m01 = pair_set_epi16(cospi_31_64, -cospi_1_64);
+  const __m128i k__cospi_p05_p27 = pair_set_epi16(cospi_5_64, cospi_27_64);
+  const __m128i k__cospi_p27_m05 = pair_set_epi16(cospi_27_64, -cospi_5_64);
+  const __m128i k__cospi_p09_p23 = pair_set_epi16(cospi_9_64, cospi_23_64);
+  const __m128i k__cospi_p23_m09 = pair_set_epi16(cospi_23_64, -cospi_9_64);
+  const __m128i k__cospi_p13_p19 = pair_set_epi16(cospi_13_64, cospi_19_64);
+  const __m128i k__cospi_p19_m13 = pair_set_epi16(cospi_19_64, -cospi_13_64);
+  const __m128i k__cospi_p17_p15 = pair_set_epi16(cospi_17_64, cospi_15_64);
+  const __m128i k__cospi_p15_m17 = pair_set_epi16(cospi_15_64, -cospi_17_64);
+  const __m128i k__cospi_p21_p11 = pair_set_epi16(cospi_21_64, cospi_11_64);
+  const __m128i k__cospi_p11_m21 = pair_set_epi16(cospi_11_64, -cospi_21_64);
+  const __m128i k__cospi_p25_p07 = pair_set_epi16(cospi_25_64, cospi_7_64);
+  const __m128i k__cospi_p07_m25 = pair_set_epi16(cospi_7_64, -cospi_25_64);
+  const __m128i k__cospi_p29_p03 = pair_set_epi16(cospi_29_64, cospi_3_64);
+  const __m128i k__cospi_p03_m29 = pair_set_epi16(cospi_3_64, -cospi_29_64);
+  const __m128i k__cospi_p04_p28 = pair_set_epi16(cospi_4_64, cospi_28_64);
+  const __m128i k__cospi_p28_m04 = pair_set_epi16(cospi_28_64, -cospi_4_64);
+  const __m128i k__cospi_p20_p12 = pair_set_epi16(cospi_20_64, cospi_12_64);
+  const __m128i k__cospi_p12_m20 = pair_set_epi16(cospi_12_64, -cospi_20_64);
+  const __m128i k__cospi_m28_p04 = pair_set_epi16(-cospi_28_64, cospi_4_64);
+  const __m128i k__cospi_m12_p20 = pair_set_epi16(-cospi_12_64, cospi_20_64);
+  const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
+  const __m128i k__cospi_m24_p08 = pair_set_epi16(-cospi_24_64, cospi_8_64);
+  const __m128i k__cospi_m16_m16 = _mm_set1_epi16((int16_t)-cospi_16_64);
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_m16_p16 = pair_set_epi16(-cospi_16_64, cospi_16_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i kZero = _mm_set1_epi16(0);
+
+  u[0] = _mm_unpacklo_epi16(in[15], in[0]);
+  u[1] = _mm_unpackhi_epi16(in[15], in[0]);
+  u[2] = _mm_unpacklo_epi16(in[13], in[2]);
+  u[3] = _mm_unpackhi_epi16(in[13], in[2]);
+  u[4] = _mm_unpacklo_epi16(in[11], in[4]);
+  u[5] = _mm_unpackhi_epi16(in[11], in[4]);
+  u[6] = _mm_unpacklo_epi16(in[9], in[6]);
+  u[7] = _mm_unpackhi_epi16(in[9], in[6]);
+  u[8] = _mm_unpacklo_epi16(in[7], in[8]);
+  u[9] = _mm_unpackhi_epi16(in[7], in[8]);
+  u[10] = _mm_unpacklo_epi16(in[5], in[10]);
+  u[11] = _mm_unpackhi_epi16(in[5], in[10]);
+  u[12] = _mm_unpacklo_epi16(in[3], in[12]);
+  u[13] = _mm_unpackhi_epi16(in[3], in[12]);
+  u[14] = _mm_unpacklo_epi16(in[1], in[14]);
+  u[15] = _mm_unpackhi_epi16(in[1], in[14]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p01_p31);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_p01_p31);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p31_m01);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p31_m01);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p05_p27);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p05_p27);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_p27_m05);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_p27_m05);
+  v[8] = _mm_madd_epi16(u[4], k__cospi_p09_p23);
+  v[9] = _mm_madd_epi16(u[5], k__cospi_p09_p23);
+  v[10] = _mm_madd_epi16(u[4], k__cospi_p23_m09);
+  v[11] = _mm_madd_epi16(u[5], k__cospi_p23_m09);
+  v[12] = _mm_madd_epi16(u[6], k__cospi_p13_p19);
+  v[13] = _mm_madd_epi16(u[7], k__cospi_p13_p19);
+  v[14] = _mm_madd_epi16(u[6], k__cospi_p19_m13);
+  v[15] = _mm_madd_epi16(u[7], k__cospi_p19_m13);
+  v[16] = _mm_madd_epi16(u[8], k__cospi_p17_p15);
+  v[17] = _mm_madd_epi16(u[9], k__cospi_p17_p15);
+  v[18] = _mm_madd_epi16(u[8], k__cospi_p15_m17);
+  v[19] = _mm_madd_epi16(u[9], k__cospi_p15_m17);
+  v[20] = _mm_madd_epi16(u[10], k__cospi_p21_p11);
+  v[21] = _mm_madd_epi16(u[11], k__cospi_p21_p11);
+  v[22] = _mm_madd_epi16(u[10], k__cospi_p11_m21);
+  v[23] = _mm_madd_epi16(u[11], k__cospi_p11_m21);
+  v[24] = _mm_madd_epi16(u[12], k__cospi_p25_p07);
+  v[25] = _mm_madd_epi16(u[13], k__cospi_p25_p07);
+  v[26] = _mm_madd_epi16(u[12], k__cospi_p07_m25);
+  v[27] = _mm_madd_epi16(u[13], k__cospi_p07_m25);
+  v[28] = _mm_madd_epi16(u[14], k__cospi_p29_p03);
+  v[29] = _mm_madd_epi16(u[15], k__cospi_p29_p03);
+  v[30] = _mm_madd_epi16(u[14], k__cospi_p03_m29);
+  v[31] = _mm_madd_epi16(u[15], k__cospi_p03_m29);
+
+  u[0] = _mm_add_epi32(v[0], v[16]);
+  u[1] = _mm_add_epi32(v[1], v[17]);
+  u[2] = _mm_add_epi32(v[2], v[18]);
+  u[3] = _mm_add_epi32(v[3], v[19]);
+  u[4] = _mm_add_epi32(v[4], v[20]);
+  u[5] = _mm_add_epi32(v[5], v[21]);
+  u[6] = _mm_add_epi32(v[6], v[22]);
+  u[7] = _mm_add_epi32(v[7], v[23]);
+  u[8] = _mm_add_epi32(v[8], v[24]);
+  u[9] = _mm_add_epi32(v[9], v[25]);
+  u[10] = _mm_add_epi32(v[10], v[26]);
+  u[11] = _mm_add_epi32(v[11], v[27]);
+  u[12] = _mm_add_epi32(v[12], v[28]);
+  u[13] = _mm_add_epi32(v[13], v[29]);
+  u[14] = _mm_add_epi32(v[14], v[30]);
+  u[15] = _mm_add_epi32(v[15], v[31]);
+  u[16] = _mm_sub_epi32(v[0], v[16]);
+  u[17] = _mm_sub_epi32(v[1], v[17]);
+  u[18] = _mm_sub_epi32(v[2], v[18]);
+  u[19] = _mm_sub_epi32(v[3], v[19]);
+  u[20] = _mm_sub_epi32(v[4], v[20]);
+  u[21] = _mm_sub_epi32(v[5], v[21]);
+  u[22] = _mm_sub_epi32(v[6], v[22]);
+  u[23] = _mm_sub_epi32(v[7], v[23]);
+  u[24] = _mm_sub_epi32(v[8], v[24]);
+  u[25] = _mm_sub_epi32(v[9], v[25]);
+  u[26] = _mm_sub_epi32(v[10], v[26]);
+  u[27] = _mm_sub_epi32(v[11], v[27]);
+  u[28] = _mm_sub_epi32(v[12], v[28]);
+  u[29] = _mm_sub_epi32(v[13], v[29]);
+  u[30] = _mm_sub_epi32(v[14], v[30]);
+  u[31] = _mm_sub_epi32(v[15], v[31]);
+
+  v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+  v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+  v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+  v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+  v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+  v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+  v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+  v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+  v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
+  v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
+  v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+  v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+  v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+  v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+  v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+  v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+  v[16] = _mm_add_epi32(u[16], k__DCT_CONST_ROUNDING);
+  v[17] = _mm_add_epi32(u[17], k__DCT_CONST_ROUNDING);
+  v[18] = _mm_add_epi32(u[18], k__DCT_CONST_ROUNDING);
+  v[19] = _mm_add_epi32(u[19], k__DCT_CONST_ROUNDING);
+  v[20] = _mm_add_epi32(u[20], k__DCT_CONST_ROUNDING);
+  v[21] = _mm_add_epi32(u[21], k__DCT_CONST_ROUNDING);
+  v[22] = _mm_add_epi32(u[22], k__DCT_CONST_ROUNDING);
+  v[23] = _mm_add_epi32(u[23], k__DCT_CONST_ROUNDING);
+  v[24] = _mm_add_epi32(u[24], k__DCT_CONST_ROUNDING);
+  v[25] = _mm_add_epi32(u[25], k__DCT_CONST_ROUNDING);
+  v[26] = _mm_add_epi32(u[26], k__DCT_CONST_ROUNDING);
+  v[27] = _mm_add_epi32(u[27], k__DCT_CONST_ROUNDING);
+  v[28] = _mm_add_epi32(u[28], k__DCT_CONST_ROUNDING);
+  v[29] = _mm_add_epi32(u[29], k__DCT_CONST_ROUNDING);
+  v[30] = _mm_add_epi32(u[30], k__DCT_CONST_ROUNDING);
+  v[31] = _mm_add_epi32(u[31], k__DCT_CONST_ROUNDING);
+
+  u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
+  u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
+  u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
+  u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
+  u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
+  u[8] = _mm_srai_epi32(v[8], DCT_CONST_BITS);
+  u[9] = _mm_srai_epi32(v[9], DCT_CONST_BITS);
+  u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
+  u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
+  u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
+  u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
+  u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
+  u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
+  u[16] = _mm_srai_epi32(v[16], DCT_CONST_BITS);
+  u[17] = _mm_srai_epi32(v[17], DCT_CONST_BITS);
+  u[18] = _mm_srai_epi32(v[18], DCT_CONST_BITS);
+  u[19] = _mm_srai_epi32(v[19], DCT_CONST_BITS);
+  u[20] = _mm_srai_epi32(v[20], DCT_CONST_BITS);
+  u[21] = _mm_srai_epi32(v[21], DCT_CONST_BITS);
+  u[22] = _mm_srai_epi32(v[22], DCT_CONST_BITS);
+  u[23] = _mm_srai_epi32(v[23], DCT_CONST_BITS);
+  u[24] = _mm_srai_epi32(v[24], DCT_CONST_BITS);
+  u[25] = _mm_srai_epi32(v[25], DCT_CONST_BITS);
+  u[26] = _mm_srai_epi32(v[26], DCT_CONST_BITS);
+  u[27] = _mm_srai_epi32(v[27], DCT_CONST_BITS);
+  u[28] = _mm_srai_epi32(v[28], DCT_CONST_BITS);
+  u[29] = _mm_srai_epi32(v[29], DCT_CONST_BITS);
+  u[30] = _mm_srai_epi32(v[30], DCT_CONST_BITS);
+  u[31] = _mm_srai_epi32(v[31], DCT_CONST_BITS);
+
+  s[0] = _mm_packs_epi32(u[0], u[1]);
+  s[1] = _mm_packs_epi32(u[2], u[3]);
+  s[2] = _mm_packs_epi32(u[4], u[5]);
+  s[3] = _mm_packs_epi32(u[6], u[7]);
+  s[4] = _mm_packs_epi32(u[8], u[9]);
+  s[5] = _mm_packs_epi32(u[10], u[11]);
+  s[6] = _mm_packs_epi32(u[12], u[13]);
+  s[7] = _mm_packs_epi32(u[14], u[15]);
+  s[8] = _mm_packs_epi32(u[16], u[17]);
+  s[9] = _mm_packs_epi32(u[18], u[19]);
+  s[10] = _mm_packs_epi32(u[20], u[21]);
+  s[11] = _mm_packs_epi32(u[22], u[23]);
+  s[12] = _mm_packs_epi32(u[24], u[25]);
+  s[13] = _mm_packs_epi32(u[26], u[27]);
+  s[14] = _mm_packs_epi32(u[28], u[29]);
+  s[15] = _mm_packs_epi32(u[30], u[31]);
+
+  // stage 2
+  u[0] = _mm_unpacklo_epi16(s[8], s[9]);
+  u[1] = _mm_unpackhi_epi16(s[8], s[9]);
+  u[2] = _mm_unpacklo_epi16(s[10], s[11]);
+  u[3] = _mm_unpackhi_epi16(s[10], s[11]);
+  u[4] = _mm_unpacklo_epi16(s[12], s[13]);
+  u[5] = _mm_unpackhi_epi16(s[12], s[13]);
+  u[6] = _mm_unpacklo_epi16(s[14], s[15]);
+  u[7] = _mm_unpackhi_epi16(s[14], s[15]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p04_p28);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_p04_p28);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p28_m04);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p28_m04);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p20_p12);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p20_p12);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_p12_m20);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_p12_m20);
+  v[8] = _mm_madd_epi16(u[4], k__cospi_m28_p04);
+  v[9] = _mm_madd_epi16(u[5], k__cospi_m28_p04);
+  v[10] = _mm_madd_epi16(u[4], k__cospi_p04_p28);
+  v[11] = _mm_madd_epi16(u[5], k__cospi_p04_p28);
+  v[12] = _mm_madd_epi16(u[6], k__cospi_m12_p20);
+  v[13] = _mm_madd_epi16(u[7], k__cospi_m12_p20);
+  v[14] = _mm_madd_epi16(u[6], k__cospi_p20_p12);
+  v[15] = _mm_madd_epi16(u[7], k__cospi_p20_p12);
+
+  u[0] = _mm_add_epi32(v[0], v[8]);
+  u[1] = _mm_add_epi32(v[1], v[9]);
+  u[2] = _mm_add_epi32(v[2], v[10]);
+  u[3] = _mm_add_epi32(v[3], v[11]);
+  u[4] = _mm_add_epi32(v[4], v[12]);
+  u[5] = _mm_add_epi32(v[5], v[13]);
+  u[6] = _mm_add_epi32(v[6], v[14]);
+  u[7] = _mm_add_epi32(v[7], v[15]);
+  u[8] = _mm_sub_epi32(v[0], v[8]);
+  u[9] = _mm_sub_epi32(v[1], v[9]);
+  u[10] = _mm_sub_epi32(v[2], v[10]);
+  u[11] = _mm_sub_epi32(v[3], v[11]);
+  u[12] = _mm_sub_epi32(v[4], v[12]);
+  u[13] = _mm_sub_epi32(v[5], v[13]);
+  u[14] = _mm_sub_epi32(v[6], v[14]);
+  u[15] = _mm_sub_epi32(v[7], v[15]);
+
+  v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+  v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+  v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+  v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+  v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+  v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+  v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+  v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+  v[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
+  v[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
+  v[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+  v[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+  v[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+  v[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+  v[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+  v[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+  u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
+  u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
+  u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
+  u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
+  u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
+  u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
+  u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
+  u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);
+  u[8] = _mm_srai_epi32(v[8], DCT_CONST_BITS);
+  u[9] = _mm_srai_epi32(v[9], DCT_CONST_BITS);
+  u[10] = _mm_srai_epi32(v[10], DCT_CONST_BITS);
+  u[11] = _mm_srai_epi32(v[11], DCT_CONST_BITS);
+  u[12] = _mm_srai_epi32(v[12], DCT_CONST_BITS);
+  u[13] = _mm_srai_epi32(v[13], DCT_CONST_BITS);
+  u[14] = _mm_srai_epi32(v[14], DCT_CONST_BITS);
+  u[15] = _mm_srai_epi32(v[15], DCT_CONST_BITS);
+
+  x[0] = _mm_add_epi16(s[0], s[4]);
+  x[1] = _mm_add_epi16(s[1], s[5]);
+  x[2] = _mm_add_epi16(s[2], s[6]);
+  x[3] = _mm_add_epi16(s[3], s[7]);
+  x[4] = _mm_sub_epi16(s[0], s[4]);
+  x[5] = _mm_sub_epi16(s[1], s[5]);
+  x[6] = _mm_sub_epi16(s[2], s[6]);
+  x[7] = _mm_sub_epi16(s[3], s[7]);
+  x[8] = _mm_packs_epi32(u[0], u[1]);
+  x[9] = _mm_packs_epi32(u[2], u[3]);
+  x[10] = _mm_packs_epi32(u[4], u[5]);
+  x[11] = _mm_packs_epi32(u[6], u[7]);
+  x[12] = _mm_packs_epi32(u[8], u[9]);
+  x[13] = _mm_packs_epi32(u[10], u[11]);
+  x[14] = _mm_packs_epi32(u[12], u[13]);
+  x[15] = _mm_packs_epi32(u[14], u[15]);
+
+  // stage 3
+  u[0] = _mm_unpacklo_epi16(x[4], x[5]);
+  u[1] = _mm_unpackhi_epi16(x[4], x[5]);
+  u[2] = _mm_unpacklo_epi16(x[6], x[7]);
+  u[3] = _mm_unpackhi_epi16(x[6], x[7]);
+  u[4] = _mm_unpacklo_epi16(x[12], x[13]);
+  u[5] = _mm_unpackhi_epi16(x[12], x[13]);
+  u[6] = _mm_unpacklo_epi16(x[14], x[15]);
+  u[7] = _mm_unpackhi_epi16(x[14], x[15]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_p08_p24);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_p08_p24);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p24_m08);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p24_m08);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_m24_p08);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_m24_p08);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_p08_p24);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_p08_p24);
+  v[8] = _mm_madd_epi16(u[4], k__cospi_p08_p24);
+  v[9] = _mm_madd_epi16(u[5], k__cospi_p08_p24);
+  v[10] = _mm_madd_epi16(u[4], k__cospi_p24_m08);
+  v[11] = _mm_madd_epi16(u[5], k__cospi_p24_m08);
+  v[12] = _mm_madd_epi16(u[6], k__cospi_m24_p08);
+  v[13] = _mm_madd_epi16(u[7], k__cospi_m24_p08);
+  v[14] = _mm_madd_epi16(u[6], k__cospi_p08_p24);
+  v[15] = _mm_madd_epi16(u[7], k__cospi_p08_p24);
+
+  u[0] = _mm_add_epi32(v[0], v[4]);
+  u[1] = _mm_add_epi32(v[1], v[5]);
+  u[2] = _mm_add_epi32(v[2], v[6]);
+  u[3] = _mm_add_epi32(v[3], v[7]);
+  u[4] = _mm_sub_epi32(v[0], v[4]);
+  u[5] = _mm_sub_epi32(v[1], v[5]);
+  u[6] = _mm_sub_epi32(v[2], v[6]);
+  u[7] = _mm_sub_epi32(v[3], v[7]);
+  u[8] = _mm_add_epi32(v[8], v[12]);
+  u[9] = _mm_add_epi32(v[9], v[13]);
+  u[10] = _mm_add_epi32(v[10], v[14]);
+  u[11] = _mm_add_epi32(v[11], v[15]);
+  u[12] = _mm_sub_epi32(v[8], v[12]);
+  u[13] = _mm_sub_epi32(v[9], v[13]);
+  u[14] = _mm_sub_epi32(v[10], v[14]);
+  u[15] = _mm_sub_epi32(v[11], v[15]);
+
+  u[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);
+  u[8] = _mm_add_epi32(u[8], k__DCT_CONST_ROUNDING);
+  u[9] = _mm_add_epi32(u[9], k__DCT_CONST_ROUNDING);
+  u[10] = _mm_add_epi32(u[10], k__DCT_CONST_ROUNDING);
+  u[11] = _mm_add_epi32(u[11], k__DCT_CONST_ROUNDING);
+  u[12] = _mm_add_epi32(u[12], k__DCT_CONST_ROUNDING);
+  u[13] = _mm_add_epi32(u[13], k__DCT_CONST_ROUNDING);
+  u[14] = _mm_add_epi32(u[14], k__DCT_CONST_ROUNDING);
+  u[15] = _mm_add_epi32(u[15], k__DCT_CONST_ROUNDING);
+
+  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+  v[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS);
+  v[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS);
+  v[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS);
+  v[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS);
+  v[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS);
+  v[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS);
+  v[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS);
+  v[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS);
+
+  s[0] = _mm_add_epi16(x[0], x[2]);
+  s[1] = _mm_add_epi16(x[1], x[3]);
+  s[2] = _mm_sub_epi16(x[0], x[2]);
+  s[3] = _mm_sub_epi16(x[1], x[3]);
+  s[4] = _mm_packs_epi32(v[0], v[1]);
+  s[5] = _mm_packs_epi32(v[2], v[3]);
+  s[6] = _mm_packs_epi32(v[4], v[5]);
+  s[7] = _mm_packs_epi32(v[6], v[7]);
+  s[8] = _mm_add_epi16(x[8], x[10]);
+  s[9] = _mm_add_epi16(x[9], x[11]);
+  s[10] = _mm_sub_epi16(x[8], x[10]);
+  s[11] = _mm_sub_epi16(x[9], x[11]);
+  s[12] = _mm_packs_epi32(v[8], v[9]);
+  s[13] = _mm_packs_epi32(v[10], v[11]);
+  s[14] = _mm_packs_epi32(v[12], v[13]);
+  s[15] = _mm_packs_epi32(v[14], v[15]);
+
+  // stage 4
+  u[0] = _mm_unpacklo_epi16(s[2], s[3]);
+  u[1] = _mm_unpackhi_epi16(s[2], s[3]);
+  u[2] = _mm_unpacklo_epi16(s[6], s[7]);
+  u[3] = _mm_unpackhi_epi16(s[6], s[7]);
+  u[4] = _mm_unpacklo_epi16(s[10], s[11]);
+  u[5] = _mm_unpackhi_epi16(s[10], s[11]);
+  u[6] = _mm_unpacklo_epi16(s[14], s[15]);
+  u[7] = _mm_unpackhi_epi16(s[14], s[15]);
+
+  v[0] = _mm_madd_epi16(u[0], k__cospi_m16_m16);
+  v[1] = _mm_madd_epi16(u[1], k__cospi_m16_m16);
+  v[2] = _mm_madd_epi16(u[0], k__cospi_p16_m16);
+  v[3] = _mm_madd_epi16(u[1], k__cospi_p16_m16);
+  v[4] = _mm_madd_epi16(u[2], k__cospi_p16_p16);
+  v[5] = _mm_madd_epi16(u[3], k__cospi_p16_p16);
+  v[6] = _mm_madd_epi16(u[2], k__cospi_m16_p16);
+  v[7] = _mm_madd_epi16(u[3], k__cospi_m16_p16);
+  v[8] = _mm_madd_epi16(u[4], k__cospi_p16_p16);
+  v[9] = _mm_madd_epi16(u[5], k__cospi_p16_p16);
+  v[10] = _mm_madd_epi16(u[4], k__cospi_m16_p16);
+  v[11] = _mm_madd_epi16(u[5], k__cospi_m16_p16);
+  v[12] = _mm_madd_epi16(u[6], k__cospi_m16_m16);
+  v[13] = _mm_madd_epi16(u[7], k__cospi_m16_m16);
+  v[14] = _mm_madd_epi16(u[6], k__cospi_p16_m16);
+  v[15] = _mm_madd_epi16(u[7], k__cospi_p16_m16);
+
+  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
+  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
+  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
+  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
+  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
+  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
+  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
+  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);
+  u[8] = _mm_add_epi32(v[8], k__DCT_CONST_ROUNDING);
+  u[9] = _mm_add_epi32(v[9], k__DCT_CONST_ROUNDING);
+  u[10] = _mm_add_epi32(v[10], k__DCT_CONST_ROUNDING);
+  u[11] = _mm_add_epi32(v[11], k__DCT_CONST_ROUNDING);
+  u[12] = _mm_add_epi32(v[12], k__DCT_CONST_ROUNDING);
+  u[13] = _mm_add_epi32(v[13], k__DCT_CONST_ROUNDING);
+  u[14] = _mm_add_epi32(v[14], k__DCT_CONST_ROUNDING);
+  u[15] = _mm_add_epi32(v[15], k__DCT_CONST_ROUNDING);
+
+  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
+  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
+  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
+  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
+  v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
+  v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
+  v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
+  v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);
+  v[8] = _mm_srai_epi32(u[8], DCT_CONST_BITS);
+  v[9] = _mm_srai_epi32(u[9], DCT_CONST_BITS);
+  v[10] = _mm_srai_epi32(u[10], DCT_CONST_BITS);
+  v[11] = _mm_srai_epi32(u[11], DCT_CONST_BITS);
+  v[12] = _mm_srai_epi32(u[12], DCT_CONST_BITS);
+  v[13] = _mm_srai_epi32(u[13], DCT_CONST_BITS);
+  v[14] = _mm_srai_epi32(u[14], DCT_CONST_BITS);
+  v[15] = _mm_srai_epi32(u[15], DCT_CONST_BITS);
+
+  in[0] = s[0];
+  in[1] = _mm_sub_epi16(kZero, s[8]);
+  in[2] = s[12];
+  in[3] = _mm_sub_epi16(kZero, s[4]);
+  in[4] = _mm_packs_epi32(v[4], v[5]);
+  in[5] = _mm_packs_epi32(v[12], v[13]);
+  in[6] = _mm_packs_epi32(v[8], v[9]);
+  in[7] = _mm_packs_epi32(v[0], v[1]);
+  in[8] = _mm_packs_epi32(v[2], v[3]);
+  in[9] = _mm_packs_epi32(v[10], v[11]);
+  in[10] = _mm_packs_epi32(v[14], v[15]);
+  in[11] = _mm_packs_epi32(v[6], v[7]);
+  in[12] = s[5];
+  in[13] = _mm_sub_epi16(kZero, s[13]);
+  in[14] = s[9];
+  in[15] = _mm_sub_epi16(kZero, s[1]);
+}
+
+static void fdct16_sse2(__m128i *in0, __m128i *in1) {
+  fdct16_8col(in0);
+  fdct16_8col(in1);
+  array_transpose_16x16(in0, in1);
+}
+
+static void fadst16_sse2(__m128i *in0, __m128i *in1) {
+  fadst16_8col(in0);
+  fadst16_8col(in1);
+  array_transpose_16x16(in0, in1);
+}
+
+void vp9_fht16x16_sse2(const int16_t *input, tran_low_t *output,
+                       int stride, int tx_type) {
+  __m128i in0[16], in1[16];
+
+  switch (tx_type) {
+    case DCT_DCT:
+      vp9_fdct16x16_sse2(input, output, stride);
+      break;
+    case ADST_DCT:
+      load_buffer_16x16(input, in0, in1, stride);
+      fadst16_sse2(in0, in1);
+      right_shift_16x16(in0, in1);
+      fdct16_sse2(in0, in1);
+      write_buffer_16x16(output, in0, in1, 16);
+      break;
+    case DCT_ADST:
+      load_buffer_16x16(input, in0, in1, stride);
+      fdct16_sse2(in0, in1);
+      right_shift_16x16(in0, in1);
+      fadst16_sse2(in0, in1);
+      write_buffer_16x16(output, in0, in1, 16);
+      break;
+    case ADST_ADST:
+      load_buffer_16x16(input, in0, in1, stride);
+      fadst16_sse2(in0, in1);
+      right_shift_16x16(in0, in1);
+      fadst16_sse2(in0, in1);
+      write_buffer_16x16(output, in0, in1, 16);
+      break;
+    default:
+      assert(0);
+      break;
+  }
+}
+
+void vp9_fdct32x32_1_sse2(const int16_t *input, tran_low_t *output,
+                          int stride) {
+  __m128i in0, in1, in2, in3;
+  __m128i u0, u1;
+  __m128i sum = _mm_setzero_si128();
+  int i;
+
+  for (i = 0; i < 8; ++i) {
+    in0  = _mm_load_si128((const __m128i *)(input +  0));
+    in1  = _mm_load_si128((const __m128i *)(input +  8));
+    in2  = _mm_load_si128((const __m128i *)(input + 16));
+    in3  = _mm_load_si128((const __m128i *)(input + 24));
+
+    input += stride;
+    u0 = _mm_add_epi16(in0, in1);
+    u1 = _mm_add_epi16(in2, in3);
+    sum = _mm_add_epi16(sum, u0);
+
+    in0  = _mm_load_si128((const __m128i *)(input +  0));
+    in1  = _mm_load_si128((const __m128i *)(input +  8));
+    in2  = _mm_load_si128((const __m128i *)(input + 16));
+    in3  = _mm_load_si128((const __m128i *)(input + 24));
+
+    input += stride;
+    sum = _mm_add_epi16(sum, u1);
+    u0  = _mm_add_epi16(in0, in1);
+    u1  = _mm_add_epi16(in2, in3);
+    sum = _mm_add_epi16(sum, u0);
+
+    in0  = _mm_load_si128((const __m128i *)(input +  0));
+    in1  = _mm_load_si128((const __m128i *)(input +  8));
+    in2  = _mm_load_si128((const __m128i *)(input + 16));
+    in3  = _mm_load_si128((const __m128i *)(input + 24));
+
+    input += stride;
+    sum = _mm_add_epi16(sum, u1);
+    u0  = _mm_add_epi16(in0, in1);
+    u1  = _mm_add_epi16(in2, in3);
+    sum = _mm_add_epi16(sum, u0);
+
+    in0  = _mm_load_si128((const __m128i *)(input +  0));
+    in1  = _mm_load_si128((const __m128i *)(input +  8));
+    in2  = _mm_load_si128((const __m128i *)(input + 16));
+    in3  = _mm_load_si128((const __m128i *)(input + 24));
+
+    input += stride;
+    sum = _mm_add_epi16(sum, u1);
+    u0  = _mm_add_epi16(in0, in1);
+    u1  = _mm_add_epi16(in2, in3);
+    sum = _mm_add_epi16(sum, u0);
+
+    sum = _mm_add_epi16(sum, u1);
+  }
+
+  u0  = _mm_setzero_si128();
+  in0 = _mm_unpacklo_epi16(u0, sum);
+  in1 = _mm_unpackhi_epi16(u0, sum);
+  in0 = _mm_srai_epi32(in0, 16);
+  in1 = _mm_srai_epi32(in1, 16);
+
+  sum = _mm_add_epi32(in0, in1);
+  in0 = _mm_unpacklo_epi32(sum, u0);
+  in1 = _mm_unpackhi_epi32(sum, u0);
+
+  sum = _mm_add_epi32(in0, in1);
+  in0 = _mm_srli_si128(sum, 8);
+
+  in1 = _mm_add_epi32(sum, in0);
+  in1 = _mm_srai_epi32(in1, 3);
+  store_output(&in1, output);
+}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+/* These SSE2 versions of the FHT functions only actually use SSE2 in the
+ * DCT_DCT case in all other cases, they revert to C code which is identical
+ * to that used by the C versions of them.
+ */
+
+void vp9_highbd_fht4x4_sse2(const int16_t *input, tran_low_t *output,
+                            int stride, int tx_type) {
+  if (tx_type == DCT_DCT) {
+    vp9_highbd_fdct4x4_sse2(input, output, stride);
+  } else {
+    tran_low_t out[4 * 4];
+    tran_low_t *outptr = &out[0];
+    int i, j;
+    tran_low_t temp_in[4], temp_out[4];
+    const transform_2d ht = FHT_4[tx_type];
+
+    // Columns
+    for (i = 0; i < 4; ++i) {
+      for (j = 0; j < 4; ++j)
+        temp_in[j] = input[j * stride + i] * 16;
+      if (i == 0 && temp_in[0])
+        temp_in[0] += 1;
+      ht.cols(temp_in, temp_out);
+      for (j = 0; j < 4; ++j)
+        outptr[j * 4 + i] = temp_out[j];
+    }
+
+    // Rows
+    for (i = 0; i < 4; ++i) {
+      for (j = 0; j < 4; ++j)
+        temp_in[j] = out[j + i * 4];
+      ht.rows(temp_in, temp_out);
+      for (j = 0; j < 4; ++j)
+        output[j + i * 4] = (temp_out[j] + 1) >> 2;
+    }
+  }
+}
+
+void vp9_highbd_fht8x8_sse2(const int16_t *input, tran_low_t *output,
+                            int stride, int tx_type) {
+  if (tx_type == DCT_DCT) {
+    vp9_highbd_fdct8x8_sse2(input, output, stride);
+  } else {
+    tran_low_t out[64];
+    tran_low_t *outptr = &out[0];
+    int i, j;
+    tran_low_t temp_in[8], temp_out[8];
+    const transform_2d ht = FHT_8[tx_type];
+
+    // Columns
+    for (i = 0; i < 8; ++i) {
+      for (j = 0; j < 8; ++j)
+        temp_in[j] = input[j * stride + i] * 4;
+      ht.cols(temp_in, temp_out);
+      for (j = 0; j < 8; ++j)
+        outptr[j * 8 + i] = temp_out[j];
+    }
+
+    // Rows
+    for (i = 0; i < 8; ++i) {
+      for (j = 0; j < 8; ++j)
+        temp_in[j] = out[j + i * 8];
+      ht.rows(temp_in, temp_out);
+      for (j = 0; j < 8; ++j)
+        output[j + i * 8] = (temp_out[j] + (temp_out[j] < 0)) >> 1;
+    }
+  }
+}
+
+void vp9_highbd_fht16x16_sse2(const int16_t *input, tran_low_t *output,
+                              int stride, int tx_type) {
+  if (tx_type == DCT_DCT) {
+    vp9_highbd_fdct16x16_sse2(input, output, stride);
+  } else {
+    tran_low_t out[256];
+    tran_low_t *outptr = &out[0];
+    int i, j;
+    tran_low_t temp_in[16], temp_out[16];
+    const transform_2d ht = FHT_16[tx_type];
+
+    // Columns
+    for (i = 0; i < 16; ++i) {
+      for (j = 0; j < 16; ++j)
+        temp_in[j] = input[j * stride + i] * 4;
+      ht.cols(temp_in, temp_out);
+      for (j = 0; j < 16; ++j)
+        outptr[j * 16 + i] = (temp_out[j] + 1 + (temp_out[j] < 0)) >> 2;
+    }
+
+    // Rows
+    for (i = 0; i < 16; ++i) {
+      for (j = 0; j < 16; ++j)
+        temp_in[j] = out[j + i * 16];
+      ht.rows(temp_in, temp_out);
+      for (j = 0; j < 16; ++j)
+        output[j + i * 16] = temp_out[j];
+    }
+  }
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
+/*
+ * The DCTnxn functions are defined using the macros below. The main code for
+ * them is in separate files (vp9/encoder/x86/vp9_dct_sse2_impl.h &
+ * vp9/encoder/x86/vp9_dct32x32_sse2_impl.h) which are used by both the 8 bit code
+ * and the high bit depth code.
+ */
+
+#define DCT_HIGH_BIT_DEPTH 0
+
+#define FDCT4x4_2D vp9_fdct4x4_sse2
+#define FDCT8x8_2D vp9_fdct8x8_sse2
+#define FDCT16x16_2D vp9_fdct16x16_sse2
+#include "vp9/encoder/x86/vp9_dct_sse2_impl.h"
+#undef  FDCT4x4_2D
+#undef  FDCT8x8_2D
+#undef  FDCT16x16_2D
+
+#define FDCT32x32_2D vp9_fdct32x32_rd_sse2
+#define FDCT32x32_HIGH_PRECISION 0
+#include "vp9/encoder/x86/vp9_dct32x32_sse2_impl.h"
+#undef  FDCT32x32_2D
+#undef  FDCT32x32_HIGH_PRECISION
+
+#define FDCT32x32_2D vp9_fdct32x32_sse2
+#define FDCT32x32_HIGH_PRECISION 1
+#include "vp9/encoder/x86/vp9_dct32x32_sse2_impl.h" // NOLINT
+#undef  FDCT32x32_2D
+#undef  FDCT32x32_HIGH_PRECISION
+
+#undef  DCT_HIGH_BIT_DEPTH
+
+
+#if CONFIG_VP9_HIGHBITDEPTH
+
+#define DCT_HIGH_BIT_DEPTH 1
+
+#define FDCT4x4_2D vp9_highbd_fdct4x4_sse2
+#define FDCT8x8_2D vp9_highbd_fdct8x8_sse2
+#define FDCT16x16_2D vp9_highbd_fdct16x16_sse2
+#include "vp9/encoder/x86/vp9_dct_sse2_impl.h" // NOLINT
+#undef  FDCT4x4_2D
+#undef  FDCT8x8_2D
+#undef  FDCT16x16_2D
+
+#define FDCT32x32_2D vp9_highbd_fdct32x32_rd_sse2
+#define FDCT32x32_HIGH_PRECISION 0
+#include "vp9/encoder/x86/vp9_dct32x32_sse2_impl.h" // NOLINT
+#undef  FDCT32x32_2D
+#undef  FDCT32x32_HIGH_PRECISION
+
+#define FDCT32x32_2D vp9_highbd_fdct32x32_sse2
+#define FDCT32x32_HIGH_PRECISION 1
+#include "vp9/encoder/x86/vp9_dct32x32_sse2_impl.h" // NOLINT
+#undef  FDCT32x32_2D
+#undef  FDCT32x32_HIGH_PRECISION
+
+#undef  DCT_HIGH_BIT_DEPTH
+
+#endif  // CONFIG_VP9_HIGHBITDEPTH
diff --git a/media/libvpx/vp9/encoder/x86/vp9_dct_sse2.h b/media/libvpx/vp9/encoder/x86/vp9_dct_sse2.h
new file mode 100644
index 000000000..b99db923e
--- /dev/null
+++ b/media/libvpx/vp9/encoder/x86/vp9_dct_sse2.h
@@ -0,0 +1,464 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#ifndef VP9_ENCODER_X86_VP9_DCT_SSE2_H_
+#define VP9_ENCODER_X86_VP9_DCT_SSE2_H_
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define pair_set_epi32(a, b) \
+  _mm_set_epi32((int)(b), (int)(a), (int)(b), (int)(a))
+
+void vp9_fdct4x4_sse2(const int16_t *input, tran_low_t *output, int stride);
+void vp9_fdct8x8_sse2(const int16_t *input, tran_low_t *output, int stride);
+void vp9_fdct16x16_sse2(const int16_t *input, tran_low_t *output, int stride);
+void vp9_highbd_fdct4x4_sse2(const int16_t *input, tran_low_t *output,
+                            int stride);
+void vp9_highbd_fdct8x8_sse2(const int16_t *input, tran_low_t *output,
+                            int stride);
+void vp9_highbd_fdct16x16_sse2(const int16_t *input, tran_low_t *output,
+                               int stride);
+
+static INLINE __m128i k_madd_epi32(__m128i a, __m128i b) {
+  __m128i buf0, buf1;
+  buf0 = _mm_mul_epu32(a, b);
+  a = _mm_srli_epi64(a, 32);
+  b = _mm_srli_epi64(b, 32);
+  buf1 = _mm_mul_epu32(a, b);
+  return _mm_add_epi64(buf0, buf1);
+}
+
+static INLINE __m128i k_packs_epi64(__m128i a, __m128i b) {
+  __m128i buf0 = _mm_shuffle_epi32(a, _MM_SHUFFLE(0, 0, 2, 0));
+  __m128i buf1 = _mm_shuffle_epi32(b, _MM_SHUFFLE(0, 0, 2, 0));
+  return _mm_unpacklo_epi64(buf0, buf1);
+}
+
+static INLINE int check_epi16_overflow_x2(const __m128i *preg0,
+                                          const __m128i *preg1) {
+  const __m128i max_overflow = _mm_set1_epi16(0x7fff);
+  const __m128i min_overflow = _mm_set1_epi16(0x8000);
+  __m128i cmp0 = _mm_or_si128(_mm_cmpeq_epi16(*preg0, max_overflow),
+                              _mm_cmpeq_epi16(*preg0, min_overflow));
+  __m128i cmp1 = _mm_or_si128(_mm_cmpeq_epi16(*preg1, max_overflow),
+                              _mm_cmpeq_epi16(*preg1, min_overflow));
+  cmp0 = _mm_or_si128(cmp0, cmp1);
+  return _mm_movemask_epi8(cmp0);
+}
+
+static INLINE int check_epi16_overflow_x4(const __m128i *preg0,
+                                          const __m128i *preg1,
+                                          const __m128i *preg2,
+                                          const __m128i *preg3) {
+  const __m128i max_overflow = _mm_set1_epi16(0x7fff);
+  const __m128i min_overflow = _mm_set1_epi16(0x8000);
+  __m128i cmp0 = _mm_or_si128(_mm_cmpeq_epi16(*preg0, max_overflow),
+                              _mm_cmpeq_epi16(*preg0, min_overflow));
+  __m128i cmp1 = _mm_or_si128(_mm_cmpeq_epi16(*preg1, max_overflow),
+                              _mm_cmpeq_epi16(*preg1, min_overflow));
+  __m128i cmp2 = _mm_or_si128(_mm_cmpeq_epi16(*preg2, max_overflow),
+                              _mm_cmpeq_epi16(*preg2, min_overflow));
+  __m128i cmp3 = _mm_or_si128(_mm_cmpeq_epi16(*preg3, max_overflow),
+                              _mm_cmpeq_epi16(*preg3, min_overflow));
+  cmp0 = _mm_or_si128(_mm_or_si128(cmp0, cmp1), _mm_or_si128(cmp2, cmp3));
+  return _mm_movemask_epi8(cmp0);
+}
+
+static INLINE int check_epi16_overflow_x8(const __m128i *preg0,
+                                          const __m128i *preg1,
+                                          const __m128i *preg2,
+                                          const __m128i *preg3,
+                                          const __m128i *preg4,
+                                          const __m128i *preg5,
+                                          const __m128i *preg6,
+                                          const __m128i *preg7) {
+  int res0, res1;
+  res0 = check_epi16_overflow_x4(preg0, preg1, preg2, preg3);
+  res1 = check_epi16_overflow_x4(preg4, preg5, preg6, preg7);
+  return res0 + res1;
+}
+
+static INLINE int check_epi16_overflow_x12(const __m128i *preg0,
+                                           const __m128i *preg1,
+                                           const __m128i *preg2,
+                                           const __m128i *preg3,
+                                           const __m128i *preg4,
+                                           const __m128i *preg5,
+                                           const __m128i *preg6,
+                                           const __m128i *preg7,
+                                           const __m128i *preg8,
+                                           const __m128i *preg9,
+                                           const __m128i *preg10,
+                                           const __m128i *preg11) {
+  int res0, res1;
+  res0 = check_epi16_overflow_x4(preg0, preg1, preg2, preg3);
+  res1 = check_epi16_overflow_x4(preg4, preg5, preg6, preg7);
+  if (!res0)
+    res0 = check_epi16_overflow_x4(preg8, preg9, preg10, preg11);
+  return res0 + res1;
+}
+
+static INLINE int check_epi16_overflow_x16(const __m128i *preg0,
+                                           const __m128i *preg1,
+                                           const __m128i *preg2,
+                                           const __m128i *preg3,
+                                           const __m128i *preg4,
+                                           const __m128i *preg5,
+                                           const __m128i *preg6,
+                                           const __m128i *preg7,
+                                           const __m128i *preg8,
+                                           const __m128i *preg9,
+                                           const __m128i *preg10,
+                                           const __m128i *preg11,
+                                           const __m128i *preg12,
+                                           const __m128i *preg13,
+                                           const __m128i *preg14,
+                                           const __m128i *preg15) {
+  int res0, res1;
+  res0 = check_epi16_overflow_x4(preg0, preg1, preg2, preg3);
+  res1 = check_epi16_overflow_x4(preg4, preg5, preg6, preg7);
+  if (!res0) {
+    res0 = check_epi16_overflow_x4(preg8, preg9, preg10, preg11);
+    if (!res1)
+      res1 = check_epi16_overflow_x4(preg12, preg13, preg14, preg15);
+  }
+  return res0 + res1;
+}
+
+static INLINE int check_epi16_overflow_x32(const __m128i *preg0,
+                                           const __m128i *preg1,
+                                           const __m128i *preg2,
+                                           const __m128i *preg3,
+                                           const __m128i *preg4,
+                                           const __m128i *preg5,
+                                           const __m128i *preg6,
+                                           const __m128i *preg7,
+                                           const __m128i *preg8,
+                                           const __m128i *preg9,
+                                           const __m128i *preg10,
+                                           const __m128i *preg11,
+                                           const __m128i *preg12,
+                                           const __m128i *preg13,
+                                           const __m128i *preg14,
+                                           const __m128i *preg15,
+                                           const __m128i *preg16,
+                                           const __m128i *preg17,
+                                           const __m128i *preg18,
+                                           const __m128i *preg19,
+                                           const __m128i *preg20,
+                                           const __m128i *preg21,
+                                           const __m128i *preg22,
+                                           const __m128i *preg23,
+                                           const __m128i *preg24,
+                                           const __m128i *preg25,
+                                           const __m128i *preg26,
+                                           const __m128i *preg27,
+                                           const __m128i *preg28,
+                                           const __m128i *preg29,
+                                           const __m128i *preg30,
+                                           const __m128i *preg31) {
+  int res0, res1;
+  res0 = check_epi16_overflow_x4(preg0, preg1, preg2, preg3);
+  res1 = check_epi16_overflow_x4(preg4, preg5, preg6, preg7);
+  if (!res0) {
+    res0 = check_epi16_overflow_x4(preg8, preg9, preg10, preg11);
+    if (!res1) {
+      res1 = check_epi16_overflow_x4(preg12, preg13, preg14, preg15);
+      if (!res0) {
+        res0 = check_epi16_overflow_x4(preg16, preg17, preg18, preg19);
+        if (!res1) {
+          res1 = check_epi16_overflow_x4(preg20, preg21, preg22, preg23);
+          if (!res0) {
+            res0 = check_epi16_overflow_x4(preg24, preg25, preg26, preg27);
+            if (!res1)
+              res1 = check_epi16_overflow_x4(preg28, preg29, preg30, preg31);
+          }
+        }
+      }
+    }
+  }
+  return res0 + res1;
+}
+
+static INLINE int k_check_epi32_overflow_4(const __m128i *preg0,
+                                           const __m128i *preg1,
+                                           const __m128i *preg2,
+                                           const __m128i *preg3,
+                                           const __m128i *zero) {
+  __m128i minus_one = _mm_set1_epi32(-1);
+  // Check for overflows
+  __m128i reg0_shifted = _mm_slli_epi64(*preg0, 1);
+  __m128i reg1_shifted = _mm_slli_epi64(*preg1, 1);
+  __m128i reg2_shifted = _mm_slli_epi64(*preg2, 1);
+  __m128i reg3_shifted = _mm_slli_epi64(*preg3, 1);
+  __m128i reg0_top_dwords = _mm_shuffle_epi32(
+      reg0_shifted, _MM_SHUFFLE(0, 0, 3, 1));
+  __m128i reg1_top_dwords = _mm_shuffle_epi32(
+      reg1_shifted, _MM_SHUFFLE(0, 0, 3, 1));
+  __m128i reg2_top_dwords = _mm_shuffle_epi32(
+      reg2_shifted, _MM_SHUFFLE(0, 0, 3, 1));
+  __m128i reg3_top_dwords = _mm_shuffle_epi32(
+      reg3_shifted, _MM_SHUFFLE(0, 0, 3, 1));
+  __m128i top_dwords_01 = _mm_unpacklo_epi64(reg0_top_dwords, reg1_top_dwords);
+  __m128i top_dwords_23 = _mm_unpacklo_epi64(reg2_top_dwords, reg3_top_dwords);
+  __m128i valid_positve_01 = _mm_cmpeq_epi32(top_dwords_01, *zero);
+  __m128i valid_positve_23 = _mm_cmpeq_epi32(top_dwords_23, *zero);
+  __m128i valid_negative_01 = _mm_cmpeq_epi32(top_dwords_01, minus_one);
+  __m128i valid_negative_23 = _mm_cmpeq_epi32(top_dwords_23, minus_one);
+  int overflow_01 = _mm_movemask_epi8(
+      _mm_cmpeq_epi32(valid_positve_01, valid_negative_01));
+  int overflow_23 = _mm_movemask_epi8(
+      _mm_cmpeq_epi32(valid_positve_23, valid_negative_23));
+  return (overflow_01 + overflow_23);
+}
+
+static INLINE int k_check_epi32_overflow_8(const __m128i *preg0,
+                                           const __m128i *preg1,
+                                           const __m128i *preg2,
+                                           const __m128i *preg3,
+                                           const __m128i *preg4,
+                                           const __m128i *preg5,
+                                           const __m128i *preg6,
+                                           const __m128i *preg7,
+                                           const __m128i *zero) {
+  int overflow = k_check_epi32_overflow_4(preg0, preg1, preg2, preg3, zero);
+  if (!overflow) {
+    overflow = k_check_epi32_overflow_4(preg4, preg5, preg6, preg7, zero);
+  }
+  return overflow;
+}
+
+static INLINE int k_check_epi32_overflow_16(const __m128i *preg0,
+                                            const __m128i *preg1,
+                                            const __m128i *preg2,
+                                            const __m128i *preg3,
+                                            const __m128i *preg4,
+                                            const __m128i *preg5,
+                                            const __m128i *preg6,
+                                            const __m128i *preg7,
+                                            const __m128i *preg8,
+                                            const __m128i *preg9,
+                                            const __m128i *preg10,
+                                            const __m128i *preg11,
+                                            const __m128i *preg12,
+                                            const __m128i *preg13,
+                                            const __m128i *preg14,
+                                            const __m128i *preg15,
+                                            const __m128i *zero) {
+  int overflow = k_check_epi32_overflow_4(preg0, preg1, preg2, preg3, zero);
+  if (!overflow) {
+    overflow = k_check_epi32_overflow_4(preg4, preg5, preg6, preg7, zero);
+    if (!overflow) {
+      overflow = k_check_epi32_overflow_4(preg8, preg9, preg10, preg11,
+                                          zero);
+      if (!overflow) {
+        overflow = k_check_epi32_overflow_4(preg12, preg13, preg14, preg15,
+                                            zero);
+      }
+    }
+  }
+  return overflow;
+}
+
+static INLINE int k_check_epi32_overflow_32(const __m128i *preg0,
+                                            const __m128i *preg1,
+                                            const __m128i *preg2,
+                                            const __m128i *preg3,
+                                            const __m128i *preg4,
+                                            const __m128i *preg5,
+                                            const __m128i *preg6,
+                                            const __m128i *preg7,
+                                            const __m128i *preg8,
+                                            const __m128i *preg9,
+                                            const __m128i *preg10,
+                                            const __m128i *preg11,
+                                            const __m128i *preg12,
+                                            const __m128i *preg13,
+                                            const __m128i *preg14,
+                                            const __m128i *preg15,
+                                            const __m128i *preg16,
+                                            const __m128i *preg17,
+                                            const __m128i *preg18,
+                                            const __m128i *preg19,
+                                            const __m128i *preg20,
+                                            const __m128i *preg21,
+                                            const __m128i *preg22,
+                                            const __m128i *preg23,
+                                            const __m128i *preg24,
+                                            const __m128i *preg25,
+                                            const __m128i *preg26,
+                                            const __m128i *preg27,
+                                            const __m128i *preg28,
+                                            const __m128i *preg29,
+                                            const __m128i *preg30,
+                                            const __m128i *preg31,
+                                            const __m128i *zero) {
+  int overflow = k_check_epi32_overflow_4(preg0, preg1, preg2, preg3, zero);
+  if (!overflow) {
+    overflow = k_check_epi32_overflow_4(preg4, preg5, preg6, preg7, zero);
+    if (!overflow) {
+      overflow = k_check_epi32_overflow_4(preg8, preg9, preg10, preg11, zero);
+      if (!overflow) {
+        overflow = k_check_epi32_overflow_4(preg12, preg13, preg14, preg15,
+                                            zero);
+        if (!overflow) {
+          overflow = k_check_epi32_overflow_4(preg16, preg17, preg18, preg19,
+                                              zero);
+          if (!overflow) {
+            overflow = k_check_epi32_overflow_4(preg20, preg21,
+                                                preg22, preg23, zero);
+            if (!overflow) {
+              overflow = k_check_epi32_overflow_4(preg24, preg25,
+                                                  preg26, preg27, zero);
+              if (!overflow) {
+                overflow = k_check_epi32_overflow_4(preg28, preg29,
+                                                    preg30, preg31, zero);
+              }
+            }
+          }
+        }
+      }
+    }
+  }
+  return overflow;
+}
+
+static INLINE void store_output(const __m128i *poutput, tran_low_t* dst_ptr) {
+#if CONFIG_VP9_HIGHBITDEPTH
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i sign_bits = _mm_cmplt_epi16(*poutput, zero);
+  __m128i out0 = _mm_unpacklo_epi16(*poutput, sign_bits);
+  __m128i out1 = _mm_unpackhi_epi16(*poutput, sign_bits);
+  _mm_store_si128((__m128i *)(dst_ptr), out0);
+  _mm_store_si128((__m128i *)(dst_ptr + 4), out1);
+#else
+  _mm_store_si128((__m128i *)(dst_ptr), *poutput);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+}
+
+static INLINE void storeu_output(const __m128i *poutput, tran_low_t* dst_ptr) {
+#if CONFIG_VP9_HIGHBITDEPTH
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i sign_bits = _mm_cmplt_epi16(*poutput, zero);
+  __m128i out0 = _mm_unpacklo_epi16(*poutput, sign_bits);
+  __m128i out1 = _mm_unpackhi_epi16(*poutput, sign_bits);
+  _mm_storeu_si128((__m128i *)(dst_ptr), out0);
+  _mm_storeu_si128((__m128i *)(dst_ptr + 4), out1);
+#else
+  _mm_storeu_si128((__m128i *)(dst_ptr), *poutput);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+}
+
+
+static INLINE __m128i mult_round_shift(const __m128i *pin0,
+                                       const __m128i *pin1,
+                                       const __m128i *pmultiplier,
+                                       const __m128i *prounding,
+                                       const int shift) {
+  const __m128i u0 = _mm_madd_epi16(*pin0, *pmultiplier);
+  const __m128i u1 = _mm_madd_epi16(*pin1, *pmultiplier);
+  const __m128i v0 = _mm_add_epi32(u0, *prounding);
+  const __m128i v1 = _mm_add_epi32(u1, *prounding);
+  const __m128i w0 = _mm_srai_epi32(v0, shift);
+  const __m128i w1 = _mm_srai_epi32(v1, shift);
+  return _mm_packs_epi32(w0, w1);
+}
+
+static INLINE void transpose_and_output8x8(
+    const __m128i *pin00, const __m128i *pin01,
+    const __m128i *pin02, const __m128i *pin03,
+    const __m128i *pin04, const __m128i *pin05,
+    const __m128i *pin06, const __m128i *pin07,
+    const int pass, int16_t* out0_ptr,
+    tran_low_t* out1_ptr) {
+  // 00 01 02 03 04 05 06 07
+  // 10 11 12 13 14 15 16 17
+  // 20 21 22 23 24 25 26 27
+  // 30 31 32 33 34 35 36 37
+  // 40 41 42 43 44 45 46 47
+  // 50 51 52 53 54 55 56 57
+  // 60 61 62 63 64 65 66 67
+  // 70 71 72 73 74 75 76 77
+  const __m128i tr0_0 = _mm_unpacklo_epi16(*pin00, *pin01);
+  const __m128i tr0_1 = _mm_unpacklo_epi16(*pin02, *pin03);
+  const __m128i tr0_2 = _mm_unpackhi_epi16(*pin00, *pin01);
+  const __m128i tr0_3 = _mm_unpackhi_epi16(*pin02, *pin03);
+  const __m128i tr0_4 = _mm_unpacklo_epi16(*pin04, *pin05);
+  const __m128i tr0_5 = _mm_unpacklo_epi16(*pin06, *pin07);
+  const __m128i tr0_6 = _mm_unpackhi_epi16(*pin04, *pin05);
+  const __m128i tr0_7 = _mm_unpackhi_epi16(*pin06, *pin07);
+  // 00 10 01 11 02 12 03 13
+  // 20 30 21 31 22 32 23 33
+  // 04 14 05 15 06 16 07 17
+  // 24 34 25 35 26 36 27 37
+  // 40 50 41 51 42 52 43 53
+  // 60 70 61 71 62 72 63 73
+  // 54 54 55 55 56 56 57 57
+  // 64 74 65 75 66 76 67 77
+  const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
+  const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3);
+  const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
+  const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3);
+  const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5);
+  const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
+  const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5);
+  const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
+  // 00 10 20 30 01 11 21 31
+  // 40 50 60 70 41 51 61 71
+  // 02 12 22 32 03 13 23 33
+  // 42 52 62 72 43 53 63 73
+  // 04 14 24 34 05 15 21 36
+  // 44 54 64 74 45 55 61 76
+  // 06 16 26 36 07 17 27 37
+  // 46 56 66 76 47 57 67 77
+  const __m128i tr2_0 = _mm_unpacklo_epi64(tr1_0, tr1_4);
+  const __m128i tr2_1 = _mm_unpackhi_epi64(tr1_0, tr1_4);
+  const __m128i tr2_2 = _mm_unpacklo_epi64(tr1_2, tr1_6);
+  const __m128i tr2_3 = _mm_unpackhi_epi64(tr1_2, tr1_6);
+  const __m128i tr2_4 = _mm_unpacklo_epi64(tr1_1, tr1_5);
+  const __m128i tr2_5 = _mm_unpackhi_epi64(tr1_1, tr1_5);
+  const __m128i tr2_6 = _mm_unpacklo_epi64(tr1_3, tr1_7);
+  const __m128i tr2_7 = _mm_unpackhi_epi64(tr1_3, tr1_7);
+  // 00 10 20 30 40 50 60 70
+  // 01 11 21 31 41 51 61 71
+  // 02 12 22 32 42 52 62 72
+  // 03 13 23 33 43 53 63 73
+  // 04 14 24 34 44 54 64 74
+  // 05 15 25 35 45 55 65 75
+  // 06 16 26 36 46 56 66 76
+  // 07 17 27 37 47 57 67 77
+  if (pass == 0) {
+    _mm_storeu_si128((__m128i*)(out0_ptr + 0 * 16), tr2_0);
+    _mm_storeu_si128((__m128i*)(out0_ptr + 1 * 16), tr2_1);
+    _mm_storeu_si128((__m128i*)(out0_ptr + 2 * 16), tr2_2);
+    _mm_storeu_si128((__m128i*)(out0_ptr + 3 * 16), tr2_3);
+    _mm_storeu_si128((__m128i*)(out0_ptr + 4 * 16), tr2_4);
+    _mm_storeu_si128((__m128i*)(out0_ptr + 5 * 16), tr2_5);
+    _mm_storeu_si128((__m128i*)(out0_ptr + 6 * 16), tr2_6);
+    _mm_storeu_si128((__m128i*)(out0_ptr + 7 * 16), tr2_7);
+  } else {
+    storeu_output(&tr2_0, (out1_ptr + 0 * 16));
+    storeu_output(&tr2_1, (out1_ptr + 1 * 16));
+    storeu_output(&tr2_2, (out1_ptr + 2 * 16));
+    storeu_output(&tr2_3, (out1_ptr + 3 * 16));
+    storeu_output(&tr2_4, (out1_ptr + 4 * 16));
+    storeu_output(&tr2_5, (out1_ptr + 5 * 16));
+    storeu_output(&tr2_6, (out1_ptr + 6 * 16));
+    storeu_output(&tr2_7, (out1_ptr + 7 * 16));
+  }
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // VP9_ENCODER_X86_VP9_DCT_SSE2_H_
diff --git a/media/libvpx/vp9/encoder/x86/vp9_dct_sse2_impl.h b/media/libvpx/vp9/encoder/x86/vp9_dct_sse2_impl.h
new file mode 100644
index 000000000..11bf5a25e
--- /dev/null
+++ b/media/libvpx/vp9/encoder/x86/vp9_dct_sse2_impl.h
@@ -0,0 +1,1024 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <emmintrin.h>  // SSE2
+
+#include "./vp9_rtcd.h"
+#include "vp9/common/vp9_idct.h"  // for cospi constants
+#include "vp9/encoder/vp9_dct.h"
+#include "vp9/encoder/x86/vp9_dct_sse2.h"
+#include "vpx_ports/mem.h"
+
+#if DCT_HIGH_BIT_DEPTH
+#define ADD_EPI16 _mm_adds_epi16
+#define SUB_EPI16 _mm_subs_epi16
+
+#else
+#define ADD_EPI16 _mm_add_epi16
+#define SUB_EPI16 _mm_sub_epi16
+#endif
+
+void FDCT4x4_2D(const int16_t *input, tran_low_t *output, int stride) {
+  // This 2D transform implements 4 vertical 1D transforms followed
+  // by 4 horizontal 1D transforms.  The multiplies and adds are as given
+  // by Chen, Smith and Fralick ('77).  The commands for moving the data
+  // around have been minimized by hand.
+  // For the purposes of the comments, the 16 inputs are referred to at i0
+  // through iF (in raster order), intermediate variables are a0, b0, c0
+  // through f, and correspond to the in-place computations mapped to input
+  // locations.  The outputs, o0 through oF are labeled according to the
+  // output locations.
+
+  // Constants
+  // These are the coefficients used for the multiplies.
+  // In the comments, pN means cos(N pi /64) and mN is -cos(N pi /64),
+  // where cospi_N_64 = cos(N pi /64)
+  const __m128i k__cospi_A = _mm_setr_epi16(cospi_16_64, cospi_16_64,
+                                            cospi_16_64, cospi_16_64,
+                                            cospi_16_64, -cospi_16_64,
+                                            cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_B = _mm_setr_epi16(cospi_16_64, -cospi_16_64,
+                                            cospi_16_64, -cospi_16_64,
+                                            cospi_16_64, cospi_16_64,
+                                            cospi_16_64, cospi_16_64);
+  const __m128i k__cospi_C = _mm_setr_epi16(cospi_8_64, cospi_24_64,
+                                            cospi_8_64, cospi_24_64,
+                                            cospi_24_64, -cospi_8_64,
+                                            cospi_24_64, -cospi_8_64);
+  const __m128i k__cospi_D = _mm_setr_epi16(cospi_24_64, -cospi_8_64,
+                                            cospi_24_64, -cospi_8_64,
+                                            cospi_8_64, cospi_24_64,
+                                            cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_E = _mm_setr_epi16(cospi_16_64, cospi_16_64,
+                                            cospi_16_64, cospi_16_64,
+                                            cospi_16_64, cospi_16_64,
+                                            cospi_16_64, cospi_16_64);
+  const __m128i k__cospi_F = _mm_setr_epi16(cospi_16_64, -cospi_16_64,
+                                            cospi_16_64, -cospi_16_64,
+                                            cospi_16_64, -cospi_16_64,
+                                            cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_G = _mm_setr_epi16(cospi_8_64, cospi_24_64,
+                                            cospi_8_64, cospi_24_64,
+                                            -cospi_8_64, -cospi_24_64,
+                                            -cospi_8_64, -cospi_24_64);
+  const __m128i k__cospi_H = _mm_setr_epi16(cospi_24_64, -cospi_8_64,
+                                            cospi_24_64, -cospi_8_64,
+                                            -cospi_24_64, cospi_8_64,
+                                            -cospi_24_64, cospi_8_64);
+
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  // This second rounding constant saves doing some extra adds at the end
+  const __m128i k__DCT_CONST_ROUNDING2 = _mm_set1_epi32(DCT_CONST_ROUNDING
+                                               +(DCT_CONST_ROUNDING << 1));
+  const int DCT_CONST_BITS2 =  DCT_CONST_BITS + 2;
+  const __m128i k__nonzero_bias_a = _mm_setr_epi16(0, 1, 1, 1, 1, 1, 1, 1);
+  const __m128i k__nonzero_bias_b = _mm_setr_epi16(1, 0, 0, 0, 0, 0, 0, 0);
+  __m128i in0, in1;
+#if DCT_HIGH_BIT_DEPTH
+  __m128i cmp0, cmp1;
+  int test, overflow;
+#endif
+
+  // Load inputs.
+  in0  = _mm_loadl_epi64((const __m128i *)(input +  0 * stride));
+  in1  = _mm_loadl_epi64((const __m128i *)(input +  1 * stride));
+  in1  = _mm_unpacklo_epi64(in1, _mm_loadl_epi64((const __m128i *)
+                                                 (input +  2 * stride)));
+  in0  = _mm_unpacklo_epi64(in0, _mm_loadl_epi64((const __m128i *)
+                                                 (input +  3 * stride)));
+  // in0 = [i0 i1 i2 i3 iC iD iE iF]
+  // in1 = [i4 i5 i6 i7 i8 i9 iA iB]
+#if DCT_HIGH_BIT_DEPTH
+  // Check inputs small enough to use optimised code
+  cmp0 = _mm_xor_si128(_mm_cmpgt_epi16(in0, _mm_set1_epi16(0x3ff)),
+                       _mm_cmplt_epi16(in0, _mm_set1_epi16(0xfc00)));
+  cmp1 = _mm_xor_si128(_mm_cmpgt_epi16(in1, _mm_set1_epi16(0x3ff)),
+                       _mm_cmplt_epi16(in1, _mm_set1_epi16(0xfc00)));
+  test = _mm_movemask_epi8(_mm_or_si128(cmp0, cmp1));
+  if (test) {
+    vp9_highbd_fdct4x4_c(input, output, stride);
+    return;
+  }
+#endif  // DCT_HIGH_BIT_DEPTH
+
+  // multiply by 16 to give some extra precision
+  in0 = _mm_slli_epi16(in0, 4);
+  in1 = _mm_slli_epi16(in1, 4);
+  // if (i == 0 && input[0]) input[0] += 1;
+  // add 1 to the upper left pixel if it is non-zero, which helps reduce
+  // the round-trip error
+  {
+    // The mask will only contain whether the first value is zero, all
+    // other comparison will fail as something shifted by 4 (above << 4)
+    // can never be equal to one. To increment in the non-zero case, we
+    // add the mask and one for the first element:
+    //   - if zero, mask = -1, v = v - 1 + 1 = v
+    //   - if non-zero, mask = 0, v = v + 0 + 1 = v + 1
+    __m128i mask = _mm_cmpeq_epi16(in0, k__nonzero_bias_a);
+    in0 = _mm_add_epi16(in0, mask);
+    in0 = _mm_add_epi16(in0, k__nonzero_bias_b);
+  }
+  // There are 4 total stages, alternating between an add/subtract stage
+  // followed by an multiply-and-add stage.
+  {
+    // Stage 1: Add/subtract
+
+    // in0 = [i0 i1 i2 i3 iC iD iE iF]
+    // in1 = [i4 i5 i6 i7 i8 i9 iA iB]
+    const __m128i r0 = _mm_unpacklo_epi16(in0, in1);
+    const __m128i r1 = _mm_unpackhi_epi16(in0, in1);
+    // r0 = [i0 i4 i1 i5 i2 i6 i3 i7]
+    // r1 = [iC i8 iD i9 iE iA iF iB]
+    const __m128i r2 = _mm_shuffle_epi32(r0, 0xB4);
+    const __m128i r3 = _mm_shuffle_epi32(r1, 0xB4);
+    // r2 = [i0 i4 i1 i5 i3 i7 i2 i6]
+    // r3 = [iC i8 iD i9 iF iB iE iA]
+
+    const __m128i t0 = _mm_add_epi16(r2, r3);
+    const __m128i t1 = _mm_sub_epi16(r2, r3);
+    // t0 = [a0 a4 a1 a5 a3 a7 a2 a6]
+    // t1 = [aC a8 aD a9 aF aB aE aA]
+
+    // Stage 2: multiply by constants (which gets us into 32 bits).
+    // The constants needed here are:
+    // k__cospi_A = [p16 p16 p16 p16 p16 m16 p16 m16]
+    // k__cospi_B = [p16 m16 p16 m16 p16 p16 p16 p16]
+    // k__cospi_C = [p08 p24 p08 p24 p24 m08 p24 m08]
+    // k__cospi_D = [p24 m08 p24 m08 p08 p24 p08 p24]
+    const __m128i u0 = _mm_madd_epi16(t0, k__cospi_A);
+    const __m128i u2 = _mm_madd_epi16(t0, k__cospi_B);
+    const __m128i u1 = _mm_madd_epi16(t1, k__cospi_C);
+    const __m128i u3 = _mm_madd_epi16(t1, k__cospi_D);
+    // Then add and right-shift to get back to 16-bit range
+    const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
+    const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
+    const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
+    const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
+    const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+    const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+    const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+    const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+    // w0 = [b0 b1 b7 b6]
+    // w1 = [b8 b9 bF bE]
+    // w2 = [b4 b5 b3 b2]
+    // w3 = [bC bD bB bA]
+    const __m128i x0 = _mm_packs_epi32(w0, w1);
+    const __m128i x1 = _mm_packs_epi32(w2, w3);
+#if DCT_HIGH_BIT_DEPTH
+    overflow = check_epi16_overflow_x2(&x0, &x1);
+    if (overflow) {
+      vp9_highbd_fdct4x4_c(input, output, stride);
+      return;
+    }
+#endif  // DCT_HIGH_BIT_DEPTH
+    // x0 = [b0 b1 b7 b6 b8 b9 bF bE]
+    // x1 = [b4 b5 b3 b2 bC bD bB bA]
+    in0 = _mm_shuffle_epi32(x0, 0xD8);
+    in1 = _mm_shuffle_epi32(x1, 0x8D);
+    // in0 = [b0 b1 b8 b9 b7 b6 bF bE]
+    // in1 = [b3 b2 bB bA b4 b5 bC bD]
+  }
+  {
+    // vertical DCTs finished. Now we do the horizontal DCTs.
+    // Stage 3: Add/subtract
+
+    const __m128i t0 = ADD_EPI16(in0, in1);
+    const __m128i t1 = SUB_EPI16(in0, in1);
+    // t0 = [c0 c1 c8 c9  c4  c5  cC  cD]
+    // t1 = [c3 c2 cB cA -c7 -c6 -cF -cE]
+#if DCT_HIGH_BIT_DEPTH
+    overflow = check_epi16_overflow_x2(&t0, &t1);
+    if (overflow) {
+      vp9_highbd_fdct4x4_c(input, output, stride);
+      return;
+    }
+#endif  // DCT_HIGH_BIT_DEPTH
+
+    // Stage 4: multiply by constants (which gets us into 32 bits).
+    {
+      // The constants needed here are:
+      // k__cospi_E = [p16 p16 p16 p16 p16 p16 p16 p16]
+      // k__cospi_F = [p16 m16 p16 m16 p16 m16 p16 m16]
+      // k__cospi_G = [p08 p24 p08 p24 m08 m24 m08 m24]
+      // k__cospi_H = [p24 m08 p24 m08 m24 p08 m24 p08]
+      const __m128i u0 = _mm_madd_epi16(t0, k__cospi_E);
+      const __m128i u1 = _mm_madd_epi16(t0, k__cospi_F);
+      const __m128i u2 = _mm_madd_epi16(t1, k__cospi_G);
+      const __m128i u3 = _mm_madd_epi16(t1, k__cospi_H);
+      // Then add and right-shift to get back to 16-bit range
+      // but this combines the final right-shift as well to save operations
+      // This unusual rounding operations is to maintain bit-accurate
+      // compatibility with the c version of this function which has two
+      // rounding steps in a row.
+      const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING2);
+      const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING2);
+      const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING2);
+      const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING2);
+      const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS2);
+      const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS2);
+      const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS2);
+      const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS2);
+      // w0 = [o0 o4 o8 oC]
+      // w1 = [o2 o6 oA oE]
+      // w2 = [o1 o5 o9 oD]
+      // w3 = [o3 o7 oB oF]
+      // remember the o's are numbered according to the correct output location
+      const __m128i x0 = _mm_packs_epi32(w0, w1);
+      const __m128i x1 = _mm_packs_epi32(w2, w3);
+#if DCT_HIGH_BIT_DEPTH
+      overflow = check_epi16_overflow_x2(&x0, &x1);
+      if (overflow) {
+        vp9_highbd_fdct4x4_c(input, output, stride);
+        return;
+      }
+#endif  // DCT_HIGH_BIT_DEPTH
+      {
+        // x0 = [o0 o4 o8 oC o2 o6 oA oE]
+        // x1 = [o1 o5 o9 oD o3 o7 oB oF]
+        const __m128i y0 = _mm_unpacklo_epi16(x0, x1);
+        const __m128i y1 = _mm_unpackhi_epi16(x0, x1);
+        // y0 = [o0 o1 o4 o5 o8 o9 oC oD]
+        // y1 = [o2 o3 o6 o7 oA oB oE oF]
+        in0 = _mm_unpacklo_epi32(y0, y1);
+        // in0 = [o0 o1 o2 o3 o4 o5 o6 o7]
+        in1 = _mm_unpackhi_epi32(y0, y1);
+        // in1 = [o8 o9 oA oB oC oD oE oF]
+      }
+    }
+  }
+  // Post-condition (v + 1) >> 2 is now incorporated into previous
+  // add and right-shift commands.  Only 2 store instructions needed
+  // because we are using the fact that 1/3 are stored just after 0/2.
+  storeu_output(&in0, output + 0 * 4);
+  storeu_output(&in1, output + 2 * 4);
+}
+
+
+void FDCT8x8_2D(const int16_t *input, tran_low_t *output, int stride) {
+  int pass;
+  // Constants
+  //    When we use them, in one case, they are all the same. In all others
+  //    it's a pair of them that we need to repeat four times. This is done
+  //    by constructing the 32 bit constant corresponding to that pair.
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16(cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64);
+  const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64);
+  const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+#if DCT_HIGH_BIT_DEPTH
+  int overflow;
+#endif
+  // Load input
+  __m128i in0  = _mm_load_si128((const __m128i *)(input + 0 * stride));
+  __m128i in1  = _mm_load_si128((const __m128i *)(input + 1 * stride));
+  __m128i in2  = _mm_load_si128((const __m128i *)(input + 2 * stride));
+  __m128i in3  = _mm_load_si128((const __m128i *)(input + 3 * stride));
+  __m128i in4  = _mm_load_si128((const __m128i *)(input + 4 * stride));
+  __m128i in5  = _mm_load_si128((const __m128i *)(input + 5 * stride));
+  __m128i in6  = _mm_load_si128((const __m128i *)(input + 6 * stride));
+  __m128i in7  = _mm_load_si128((const __m128i *)(input + 7 * stride));
+  // Pre-condition input (shift by two)
+  in0 = _mm_slli_epi16(in0, 2);
+  in1 = _mm_slli_epi16(in1, 2);
+  in2 = _mm_slli_epi16(in2, 2);
+  in3 = _mm_slli_epi16(in3, 2);
+  in4 = _mm_slli_epi16(in4, 2);
+  in5 = _mm_slli_epi16(in5, 2);
+  in6 = _mm_slli_epi16(in6, 2);
+  in7 = _mm_slli_epi16(in7, 2);
+
+  // We do two passes, first the columns, then the rows. The results of the
+  // first pass are transposed so that the same column code can be reused. The
+  // results of the second pass are also transposed so that the rows (processed
+  // as columns) are put back in row positions.
+  for (pass = 0; pass < 2; pass++) {
+    // To store results of each pass before the transpose.
+    __m128i res0, res1, res2, res3, res4, res5, res6, res7;
+    // Add/subtract
+    const __m128i q0 = ADD_EPI16(in0, in7);
+    const __m128i q1 = ADD_EPI16(in1, in6);
+    const __m128i q2 = ADD_EPI16(in2, in5);
+    const __m128i q3 = ADD_EPI16(in3, in4);
+    const __m128i q4 = SUB_EPI16(in3, in4);
+    const __m128i q5 = SUB_EPI16(in2, in5);
+    const __m128i q6 = SUB_EPI16(in1, in6);
+    const __m128i q7 = SUB_EPI16(in0, in7);
+#if DCT_HIGH_BIT_DEPTH
+    if (pass == 1) {
+      overflow = check_epi16_overflow_x8(&q0, &q1, &q2, &q3,
+                                         &q4, &q5, &q6, &q7);
+      if (overflow) {
+        vp9_highbd_fdct8x8_c(input, output, stride);
+        return;
+      }
+    }
+#endif  // DCT_HIGH_BIT_DEPTH
+    // Work on first four results
+    {
+      // Add/subtract
+      const __m128i r0 = ADD_EPI16(q0, q3);
+      const __m128i r1 = ADD_EPI16(q1, q2);
+      const __m128i r2 = SUB_EPI16(q1, q2);
+      const __m128i r3 = SUB_EPI16(q0, q3);
+#if DCT_HIGH_BIT_DEPTH
+      overflow = check_epi16_overflow_x4(&r0, &r1, &r2, &r3);
+      if (overflow) {
+        vp9_highbd_fdct8x8_c(input, output, stride);
+        return;
+      }
+#endif  // DCT_HIGH_BIT_DEPTH
+      // Interleave to do the multiply by constants which gets us into 32bits
+      {
+        const __m128i t0 = _mm_unpacklo_epi16(r0, r1);
+        const __m128i t1 = _mm_unpackhi_epi16(r0, r1);
+        const __m128i t2 = _mm_unpacklo_epi16(r2, r3);
+        const __m128i t3 = _mm_unpackhi_epi16(r2, r3);
+        const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16);
+        const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_p16);
+        const __m128i u2 = _mm_madd_epi16(t0, k__cospi_p16_m16);
+        const __m128i u3 = _mm_madd_epi16(t1, k__cospi_p16_m16);
+        const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p24_p08);
+        const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p24_p08);
+        const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m08_p24);
+        const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m08_p24);
+        // dct_const_round_shift
+        const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
+        const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
+        const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
+        const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
+        const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
+        const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
+        const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
+        const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
+        const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+        const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+        const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+        const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+        const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
+        const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
+        const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
+        const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
+        // Combine
+        res0 = _mm_packs_epi32(w0, w1);
+        res4 = _mm_packs_epi32(w2, w3);
+        res2 = _mm_packs_epi32(w4, w5);
+        res6 = _mm_packs_epi32(w6, w7);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x4(&res0, &res4, &res2, &res6);
+        if (overflow) {
+          vp9_highbd_fdct8x8_c(input, output, stride);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+    }
+    // Work on next four results
+    {
+      // Interleave to do the multiply by constants which gets us into 32bits
+      const __m128i d0 = _mm_unpacklo_epi16(q6, q5);
+      const __m128i d1 = _mm_unpackhi_epi16(q6, q5);
+      const __m128i e0 = _mm_madd_epi16(d0, k__cospi_p16_m16);
+      const __m128i e1 = _mm_madd_epi16(d1, k__cospi_p16_m16);
+      const __m128i e2 = _mm_madd_epi16(d0, k__cospi_p16_p16);
+      const __m128i e3 = _mm_madd_epi16(d1, k__cospi_p16_p16);
+      // dct_const_round_shift
+      const __m128i f0 = _mm_add_epi32(e0, k__DCT_CONST_ROUNDING);
+      const __m128i f1 = _mm_add_epi32(e1, k__DCT_CONST_ROUNDING);
+      const __m128i f2 = _mm_add_epi32(e2, k__DCT_CONST_ROUNDING);
+      const __m128i f3 = _mm_add_epi32(e3, k__DCT_CONST_ROUNDING);
+      const __m128i s0 = _mm_srai_epi32(f0, DCT_CONST_BITS);
+      const __m128i s1 = _mm_srai_epi32(f1, DCT_CONST_BITS);
+      const __m128i s2 = _mm_srai_epi32(f2, DCT_CONST_BITS);
+      const __m128i s3 = _mm_srai_epi32(f3, DCT_CONST_BITS);
+      // Combine
+      const __m128i r0 = _mm_packs_epi32(s0, s1);
+      const __m128i r1 = _mm_packs_epi32(s2, s3);
+#if DCT_HIGH_BIT_DEPTH
+      overflow = check_epi16_overflow_x2(&r0, &r1);
+      if (overflow) {
+        vp9_highbd_fdct8x8_c(input, output, stride);
+        return;
+      }
+#endif  // DCT_HIGH_BIT_DEPTH
+      {
+        // Add/subtract
+        const __m128i x0 = ADD_EPI16(q4, r0);
+        const __m128i x1 = SUB_EPI16(q4, r0);
+        const __m128i x2 = SUB_EPI16(q7, r1);
+        const __m128i x3 = ADD_EPI16(q7, r1);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x4(&x0, &x1, &x2, &x3);
+        if (overflow) {
+          vp9_highbd_fdct8x8_c(input, output, stride);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+        // Interleave to do the multiply by constants which gets us into 32bits
+        {
+          const __m128i t0 = _mm_unpacklo_epi16(x0, x3);
+          const __m128i t1 = _mm_unpackhi_epi16(x0, x3);
+          const __m128i t2 = _mm_unpacklo_epi16(x1, x2);
+          const __m128i t3 = _mm_unpackhi_epi16(x1, x2);
+          const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p28_p04);
+          const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p28_p04);
+          const __m128i u2 = _mm_madd_epi16(t0, k__cospi_m04_p28);
+          const __m128i u3 = _mm_madd_epi16(t1, k__cospi_m04_p28);
+          const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p12_p20);
+          const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p12_p20);
+          const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m20_p12);
+          const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m20_p12);
+          // dct_const_round_shift
+          const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
+          const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
+          const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
+          const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
+          const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
+          const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
+          const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
+          const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
+          const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+          const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+          const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+          const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+          const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
+          const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
+          const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
+          const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
+          // Combine
+          res1 = _mm_packs_epi32(w0, w1);
+          res7 = _mm_packs_epi32(w2, w3);
+          res5 = _mm_packs_epi32(w4, w5);
+          res3 = _mm_packs_epi32(w6, w7);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x4(&res1, &res7, &res5, &res3);
+          if (overflow) {
+            vp9_highbd_fdct8x8_c(input, output, stride);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+      }
+    }
+    // Transpose the 8x8.
+    {
+      // 00 01 02 03 04 05 06 07
+      // 10 11 12 13 14 15 16 17
+      // 20 21 22 23 24 25 26 27
+      // 30 31 32 33 34 35 36 37
+      // 40 41 42 43 44 45 46 47
+      // 50 51 52 53 54 55 56 57
+      // 60 61 62 63 64 65 66 67
+      // 70 71 72 73 74 75 76 77
+      const __m128i tr0_0 = _mm_unpacklo_epi16(res0, res1);
+      const __m128i tr0_1 = _mm_unpacklo_epi16(res2, res3);
+      const __m128i tr0_2 = _mm_unpackhi_epi16(res0, res1);
+      const __m128i tr0_3 = _mm_unpackhi_epi16(res2, res3);
+      const __m128i tr0_4 = _mm_unpacklo_epi16(res4, res5);
+      const __m128i tr0_5 = _mm_unpacklo_epi16(res6, res7);
+      const __m128i tr0_6 = _mm_unpackhi_epi16(res4, res5);
+      const __m128i tr0_7 = _mm_unpackhi_epi16(res6, res7);
+      // 00 10 01 11 02 12 03 13
+      // 20 30 21 31 22 32 23 33
+      // 04 14 05 15 06 16 07 17
+      // 24 34 25 35 26 36 27 37
+      // 40 50 41 51 42 52 43 53
+      // 60 70 61 71 62 72 63 73
+      // 54 54 55 55 56 56 57 57
+      // 64 74 65 75 66 76 67 77
+      const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
+      const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3);
+      const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
+      const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3);
+      const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5);
+      const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
+      const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5);
+      const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
+      // 00 10 20 30 01 11 21 31
+      // 40 50 60 70 41 51 61 71
+      // 02 12 22 32 03 13 23 33
+      // 42 52 62 72 43 53 63 73
+      // 04 14 24 34 05 15 21 36
+      // 44 54 64 74 45 55 61 76
+      // 06 16 26 36 07 17 27 37
+      // 46 56 66 76 47 57 67 77
+      in0 = _mm_unpacklo_epi64(tr1_0, tr1_4);
+      in1 = _mm_unpackhi_epi64(tr1_0, tr1_4);
+      in2 = _mm_unpacklo_epi64(tr1_2, tr1_6);
+      in3 = _mm_unpackhi_epi64(tr1_2, tr1_6);
+      in4 = _mm_unpacklo_epi64(tr1_1, tr1_5);
+      in5 = _mm_unpackhi_epi64(tr1_1, tr1_5);
+      in6 = _mm_unpacklo_epi64(tr1_3, tr1_7);
+      in7 = _mm_unpackhi_epi64(tr1_3, tr1_7);
+      // 00 10 20 30 40 50 60 70
+      // 01 11 21 31 41 51 61 71
+      // 02 12 22 32 42 52 62 72
+      // 03 13 23 33 43 53 63 73
+      // 04 14 24 34 44 54 64 74
+      // 05 15 25 35 45 55 65 75
+      // 06 16 26 36 46 56 66 76
+      // 07 17 27 37 47 57 67 77
+    }
+  }
+  // Post-condition output and store it
+  {
+    // Post-condition (division by two)
+    //    division of two 16 bits signed numbers using shifts
+    //    n / 2 = (n - (n >> 15)) >> 1
+    const __m128i sign_in0 = _mm_srai_epi16(in0, 15);
+    const __m128i sign_in1 = _mm_srai_epi16(in1, 15);
+    const __m128i sign_in2 = _mm_srai_epi16(in2, 15);
+    const __m128i sign_in3 = _mm_srai_epi16(in3, 15);
+    const __m128i sign_in4 = _mm_srai_epi16(in4, 15);
+    const __m128i sign_in5 = _mm_srai_epi16(in5, 15);
+    const __m128i sign_in6 = _mm_srai_epi16(in6, 15);
+    const __m128i sign_in7 = _mm_srai_epi16(in7, 15);
+    in0 = _mm_sub_epi16(in0, sign_in0);
+    in1 = _mm_sub_epi16(in1, sign_in1);
+    in2 = _mm_sub_epi16(in2, sign_in2);
+    in3 = _mm_sub_epi16(in3, sign_in3);
+    in4 = _mm_sub_epi16(in4, sign_in4);
+    in5 = _mm_sub_epi16(in5, sign_in5);
+    in6 = _mm_sub_epi16(in6, sign_in6);
+    in7 = _mm_sub_epi16(in7, sign_in7);
+    in0 = _mm_srai_epi16(in0, 1);
+    in1 = _mm_srai_epi16(in1, 1);
+    in2 = _mm_srai_epi16(in2, 1);
+    in3 = _mm_srai_epi16(in3, 1);
+    in4 = _mm_srai_epi16(in4, 1);
+    in5 = _mm_srai_epi16(in5, 1);
+    in6 = _mm_srai_epi16(in6, 1);
+    in7 = _mm_srai_epi16(in7, 1);
+    // store results
+    store_output(&in0, (output + 0 * 8));
+    store_output(&in1, (output + 1 * 8));
+    store_output(&in2, (output + 2 * 8));
+    store_output(&in3, (output + 3 * 8));
+    store_output(&in4, (output + 4 * 8));
+    store_output(&in5, (output + 5 * 8));
+    store_output(&in6, (output + 6 * 8));
+    store_output(&in7, (output + 7 * 8));
+  }
+}
+
+void FDCT16x16_2D(const int16_t *input, tran_low_t *output, int stride) {
+  // The 2D transform is done with two passes which are actually pretty
+  // similar. In the first one, we transform the columns and transpose
+  // the results. In the second one, we transform the rows. To achieve that,
+  // as the first pass results are transposed, we transpose the columns (that
+  // is the transposed rows) and transpose the results (so that it goes back
+  // in normal/row positions).
+  int pass;
+  // We need an intermediate buffer between passes.
+  DECLARE_ALIGNED(16, int16_t, intermediate[256]);
+  const int16_t *in = input;
+  int16_t *out0 = intermediate;
+  tran_low_t *out1 = output;
+  // Constants
+  //    When we use them, in one case, they are all the same. In all others
+  //    it's a pair of them that we need to repeat four times. This is done
+  //    by constructing the 32 bit constant corresponding to that pair.
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16(cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i k__cospi_p08_m24 = pair_set_epi16(cospi_8_64, -cospi_24_64);
+  const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64);
+  const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64);
+  const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i k__cospi_p30_p02 = pair_set_epi16(cospi_30_64, cospi_2_64);
+  const __m128i k__cospi_p14_p18 = pair_set_epi16(cospi_14_64, cospi_18_64);
+  const __m128i k__cospi_m02_p30 = pair_set_epi16(-cospi_2_64, cospi_30_64);
+  const __m128i k__cospi_m18_p14 = pair_set_epi16(-cospi_18_64, cospi_14_64);
+  const __m128i k__cospi_p22_p10 = pair_set_epi16(cospi_22_64, cospi_10_64);
+  const __m128i k__cospi_p06_p26 = pair_set_epi16(cospi_6_64, cospi_26_64);
+  const __m128i k__cospi_m10_p22 = pair_set_epi16(-cospi_10_64, cospi_22_64);
+  const __m128i k__cospi_m26_p06 = pair_set_epi16(-cospi_26_64, cospi_6_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  const __m128i kOne = _mm_set1_epi16(1);
+  // Do the two transform/transpose passes
+  for (pass = 0; pass < 2; ++pass) {
+    // We process eight columns (transposed rows in second pass) at a time.
+    int column_start;
+#if DCT_HIGH_BIT_DEPTH
+    int overflow;
+#endif
+    for (column_start = 0; column_start < 16; column_start += 8) {
+      __m128i in00, in01, in02, in03, in04, in05, in06, in07;
+      __m128i in08, in09, in10, in11, in12, in13, in14, in15;
+      __m128i input0, input1, input2, input3, input4, input5, input6, input7;
+      __m128i step1_0, step1_1, step1_2, step1_3;
+      __m128i step1_4, step1_5, step1_6, step1_7;
+      __m128i step2_1, step2_2, step2_3, step2_4, step2_5, step2_6;
+      __m128i step3_0, step3_1, step3_2, step3_3;
+      __m128i step3_4, step3_5, step3_6, step3_7;
+      __m128i res00, res01, res02, res03, res04, res05, res06, res07;
+      __m128i res08, res09, res10, res11, res12, res13, res14, res15;
+      // Load and pre-condition input.
+      if (0 == pass) {
+        in00  = _mm_load_si128((const __m128i *)(in +  0 * stride));
+        in01  = _mm_load_si128((const __m128i *)(in +  1 * stride));
+        in02  = _mm_load_si128((const __m128i *)(in +  2 * stride));
+        in03  = _mm_load_si128((const __m128i *)(in +  3 * stride));
+        in04  = _mm_load_si128((const __m128i *)(in +  4 * stride));
+        in05  = _mm_load_si128((const __m128i *)(in +  5 * stride));
+        in06  = _mm_load_si128((const __m128i *)(in +  6 * stride));
+        in07  = _mm_load_si128((const __m128i *)(in +  7 * stride));
+        in08  = _mm_load_si128((const __m128i *)(in +  8 * stride));
+        in09  = _mm_load_si128((const __m128i *)(in +  9 * stride));
+        in10  = _mm_load_si128((const __m128i *)(in + 10 * stride));
+        in11  = _mm_load_si128((const __m128i *)(in + 11 * stride));
+        in12  = _mm_load_si128((const __m128i *)(in + 12 * stride));
+        in13  = _mm_load_si128((const __m128i *)(in + 13 * stride));
+        in14  = _mm_load_si128((const __m128i *)(in + 14 * stride));
+        in15  = _mm_load_si128((const __m128i *)(in + 15 * stride));
+        // x = x << 2
+        in00 = _mm_slli_epi16(in00, 2);
+        in01 = _mm_slli_epi16(in01, 2);
+        in02 = _mm_slli_epi16(in02, 2);
+        in03 = _mm_slli_epi16(in03, 2);
+        in04 = _mm_slli_epi16(in04, 2);
+        in05 = _mm_slli_epi16(in05, 2);
+        in06 = _mm_slli_epi16(in06, 2);
+        in07 = _mm_slli_epi16(in07, 2);
+        in08 = _mm_slli_epi16(in08, 2);
+        in09 = _mm_slli_epi16(in09, 2);
+        in10 = _mm_slli_epi16(in10, 2);
+        in11 = _mm_slli_epi16(in11, 2);
+        in12 = _mm_slli_epi16(in12, 2);
+        in13 = _mm_slli_epi16(in13, 2);
+        in14 = _mm_slli_epi16(in14, 2);
+        in15 = _mm_slli_epi16(in15, 2);
+      } else {
+        in00  = _mm_load_si128((const __m128i *)(in +  0 * 16));
+        in01  = _mm_load_si128((const __m128i *)(in +  1 * 16));
+        in02  = _mm_load_si128((const __m128i *)(in +  2 * 16));
+        in03  = _mm_load_si128((const __m128i *)(in +  3 * 16));
+        in04  = _mm_load_si128((const __m128i *)(in +  4 * 16));
+        in05  = _mm_load_si128((const __m128i *)(in +  5 * 16));
+        in06  = _mm_load_si128((const __m128i *)(in +  6 * 16));
+        in07  = _mm_load_si128((const __m128i *)(in +  7 * 16));
+        in08  = _mm_load_si128((const __m128i *)(in +  8 * 16));
+        in09  = _mm_load_si128((const __m128i *)(in +  9 * 16));
+        in10  = _mm_load_si128((const __m128i *)(in + 10 * 16));
+        in11  = _mm_load_si128((const __m128i *)(in + 11 * 16));
+        in12  = _mm_load_si128((const __m128i *)(in + 12 * 16));
+        in13  = _mm_load_si128((const __m128i *)(in + 13 * 16));
+        in14  = _mm_load_si128((const __m128i *)(in + 14 * 16));
+        in15  = _mm_load_si128((const __m128i *)(in + 15 * 16));
+        // x = (x + 1) >> 2
+        in00 = _mm_add_epi16(in00, kOne);
+        in01 = _mm_add_epi16(in01, kOne);
+        in02 = _mm_add_epi16(in02, kOne);
+        in03 = _mm_add_epi16(in03, kOne);
+        in04 = _mm_add_epi16(in04, kOne);
+        in05 = _mm_add_epi16(in05, kOne);
+        in06 = _mm_add_epi16(in06, kOne);
+        in07 = _mm_add_epi16(in07, kOne);
+        in08 = _mm_add_epi16(in08, kOne);
+        in09 = _mm_add_epi16(in09, kOne);
+        in10 = _mm_add_epi16(in10, kOne);
+        in11 = _mm_add_epi16(in11, kOne);
+        in12 = _mm_add_epi16(in12, kOne);
+        in13 = _mm_add_epi16(in13, kOne);
+        in14 = _mm_add_epi16(in14, kOne);
+        in15 = _mm_add_epi16(in15, kOne);
+        in00 = _mm_srai_epi16(in00, 2);
+        in01 = _mm_srai_epi16(in01, 2);
+        in02 = _mm_srai_epi16(in02, 2);
+        in03 = _mm_srai_epi16(in03, 2);
+        in04 = _mm_srai_epi16(in04, 2);
+        in05 = _mm_srai_epi16(in05, 2);
+        in06 = _mm_srai_epi16(in06, 2);
+        in07 = _mm_srai_epi16(in07, 2);
+        in08 = _mm_srai_epi16(in08, 2);
+        in09 = _mm_srai_epi16(in09, 2);
+        in10 = _mm_srai_epi16(in10, 2);
+        in11 = _mm_srai_epi16(in11, 2);
+        in12 = _mm_srai_epi16(in12, 2);
+        in13 = _mm_srai_epi16(in13, 2);
+        in14 = _mm_srai_epi16(in14, 2);
+        in15 = _mm_srai_epi16(in15, 2);
+      }
+      in += 8;
+      // Calculate input for the first 8 results.
+      {
+        input0 = ADD_EPI16(in00, in15);
+        input1 = ADD_EPI16(in01, in14);
+        input2 = ADD_EPI16(in02, in13);
+        input3 = ADD_EPI16(in03, in12);
+        input4 = ADD_EPI16(in04, in11);
+        input5 = ADD_EPI16(in05, in10);
+        input6 = ADD_EPI16(in06, in09);
+        input7 = ADD_EPI16(in07, in08);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x8(&input0, &input1, &input2, &input3,
+                                           &input4, &input5, &input6, &input7);
+        if (overflow) {
+          vp9_highbd_fdct16x16_c(input, output, stride);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+      // Calculate input for the next 8 results.
+      {
+        step1_0 = SUB_EPI16(in07, in08);
+        step1_1 = SUB_EPI16(in06, in09);
+        step1_2 = SUB_EPI16(in05, in10);
+        step1_3 = SUB_EPI16(in04, in11);
+        step1_4 = SUB_EPI16(in03, in12);
+        step1_5 = SUB_EPI16(in02, in13);
+        step1_6 = SUB_EPI16(in01, in14);
+        step1_7 = SUB_EPI16(in00, in15);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x8(&step1_0, &step1_1,
+                                           &step1_2, &step1_3,
+                                           &step1_4, &step1_5,
+                                           &step1_6, &step1_7);
+        if (overflow) {
+          vp9_highbd_fdct16x16_c(input, output, stride);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+      }
+      // Work on the first eight values; fdct8(input, even_results);
+      {
+        // Add/subtract
+        const __m128i q0 = ADD_EPI16(input0, input7);
+        const __m128i q1 = ADD_EPI16(input1, input6);
+        const __m128i q2 = ADD_EPI16(input2, input5);
+        const __m128i q3 = ADD_EPI16(input3, input4);
+        const __m128i q4 = SUB_EPI16(input3, input4);
+        const __m128i q5 = SUB_EPI16(input2, input5);
+        const __m128i q6 = SUB_EPI16(input1, input6);
+        const __m128i q7 = SUB_EPI16(input0, input7);
+#if DCT_HIGH_BIT_DEPTH
+        overflow = check_epi16_overflow_x8(&q0, &q1, &q2, &q3,
+                                           &q4, &q5, &q6, &q7);
+        if (overflow) {
+          vp9_highbd_fdct16x16_c(input, output, stride);
+          return;
+        }
+#endif  // DCT_HIGH_BIT_DEPTH
+        // Work on first four results
+        {
+          // Add/subtract
+          const __m128i r0 = ADD_EPI16(q0, q3);
+          const __m128i r1 = ADD_EPI16(q1, q2);
+          const __m128i r2 = SUB_EPI16(q1, q2);
+          const __m128i r3 = SUB_EPI16(q0, q3);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x4(&r0, &r1, &r2, &r3);
+          if (overflow) {
+            vp9_highbd_fdct16x16_c(input, output, stride);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+          // Interleave to do the multiply by constants which gets us
+          // into 32 bits.
+          {
+            const __m128i t0 = _mm_unpacklo_epi16(r0, r1);
+            const __m128i t1 = _mm_unpackhi_epi16(r0, r1);
+            const __m128i t2 = _mm_unpacklo_epi16(r2, r3);
+            const __m128i t3 = _mm_unpackhi_epi16(r2, r3);
+            res00 = mult_round_shift(&t0, &t1, &k__cospi_p16_p16,
+                                     &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+            res08 = mult_round_shift(&t0, &t1, &k__cospi_p16_m16,
+                                     &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+            res04 = mult_round_shift(&t2, &t3, &k__cospi_p24_p08,
+                                     &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+            res12 = mult_round_shift(&t2, &t3, &k__cospi_m08_p24,
+                                     &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+#if DCT_HIGH_BIT_DEPTH
+            overflow = check_epi16_overflow_x4(&res00, &res08, &res04, &res12);
+            if (overflow) {
+              vp9_highbd_fdct16x16_c(input, output, stride);
+              return;
+            }
+#endif  // DCT_HIGH_BIT_DEPTH
+          }
+        }
+        // Work on next four results
+        {
+          // Interleave to do the multiply by constants which gets us
+          // into 32 bits.
+          const __m128i d0 = _mm_unpacklo_epi16(q6, q5);
+          const __m128i d1 = _mm_unpackhi_epi16(q6, q5);
+          const __m128i r0 = mult_round_shift(&d0, &d1, &k__cospi_p16_m16,
+                                              &k__DCT_CONST_ROUNDING,
+                                              DCT_CONST_BITS);
+          const __m128i r1 = mult_round_shift(&d0, &d1, &k__cospi_p16_p16,
+                                              &k__DCT_CONST_ROUNDING,
+                                              DCT_CONST_BITS);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x2(&r0, &r1);
+          if (overflow) {
+            vp9_highbd_fdct16x16_c(input, output, stride);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+          {
+            // Add/subtract
+            const __m128i x0 = ADD_EPI16(q4, r0);
+            const __m128i x1 = SUB_EPI16(q4, r0);
+            const __m128i x2 = SUB_EPI16(q7, r1);
+            const __m128i x3 = ADD_EPI16(q7, r1);
+#if DCT_HIGH_BIT_DEPTH
+            overflow = check_epi16_overflow_x4(&x0, &x1, &x2, &x3);
+            if (overflow) {
+              vp9_highbd_fdct16x16_c(input, output, stride);
+              return;
+            }
+#endif  // DCT_HIGH_BIT_DEPTH
+            // Interleave to do the multiply by constants which gets us
+            // into 32 bits.
+            {
+              const __m128i t0 = _mm_unpacklo_epi16(x0, x3);
+              const __m128i t1 = _mm_unpackhi_epi16(x0, x3);
+              const __m128i t2 = _mm_unpacklo_epi16(x1, x2);
+              const __m128i t3 = _mm_unpackhi_epi16(x1, x2);
+              res02 = mult_round_shift(&t0, &t1, &k__cospi_p28_p04,
+                                       &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+              res14 = mult_round_shift(&t0, &t1, &k__cospi_m04_p28,
+                                       &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+              res10 = mult_round_shift(&t2, &t3, &k__cospi_p12_p20,
+                                       &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+              res06 = mult_round_shift(&t2, &t3, &k__cospi_m20_p12,
+                                       &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+#if DCT_HIGH_BIT_DEPTH
+              overflow = check_epi16_overflow_x4(&res02, &res14,
+                                                 &res10, &res06);
+              if (overflow) {
+                vp9_highbd_fdct16x16_c(input, output, stride);
+                return;
+              }
+#endif  // DCT_HIGH_BIT_DEPTH
+            }
+          }
+        }
+      }
+      // Work on the next eight values; step1 -> odd_results
+      {
+        // step 2
+        {
+          const __m128i t0 = _mm_unpacklo_epi16(step1_5, step1_2);
+          const __m128i t1 = _mm_unpackhi_epi16(step1_5, step1_2);
+          const __m128i t2 = _mm_unpacklo_epi16(step1_4, step1_3);
+          const __m128i t3 = _mm_unpackhi_epi16(step1_4, step1_3);
+          step2_2 = mult_round_shift(&t0, &t1, &k__cospi_p16_m16,
+                                     &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+          step2_3 = mult_round_shift(&t2, &t3, &k__cospi_p16_m16,
+                                     &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+          step2_5 = mult_round_shift(&t0, &t1, &k__cospi_p16_p16,
+                                     &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+          step2_4 = mult_round_shift(&t2, &t3, &k__cospi_p16_p16,
+                                     &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x4(&step2_2, &step2_3, &step2_5,
+                                             &step2_4);
+          if (overflow) {
+            vp9_highbd_fdct16x16_c(input, output, stride);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        // step 3
+        {
+          step3_0 = ADD_EPI16(step1_0, step2_3);
+          step3_1 = ADD_EPI16(step1_1, step2_2);
+          step3_2 = SUB_EPI16(step1_1, step2_2);
+          step3_3 = SUB_EPI16(step1_0, step2_3);
+          step3_4 = SUB_EPI16(step1_7, step2_4);
+          step3_5 = SUB_EPI16(step1_6, step2_5);
+          step3_6 = ADD_EPI16(step1_6, step2_5);
+          step3_7 = ADD_EPI16(step1_7, step2_4);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x8(&step3_0, &step3_1,
+                                             &step3_2, &step3_3,
+                                             &step3_4, &step3_5,
+                                             &step3_6, &step3_7);
+          if (overflow) {
+            vp9_highbd_fdct16x16_c(input, output, stride);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        // step 4
+        {
+          const __m128i t0 = _mm_unpacklo_epi16(step3_1, step3_6);
+          const __m128i t1 = _mm_unpackhi_epi16(step3_1, step3_6);
+          const __m128i t2 = _mm_unpacklo_epi16(step3_2, step3_5);
+          const __m128i t3 = _mm_unpackhi_epi16(step3_2, step3_5);
+          step2_1 = mult_round_shift(&t0, &t1, &k__cospi_m08_p24,
+                                     &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+          step2_2 = mult_round_shift(&t2, &t3, &k__cospi_p24_p08,
+                                     &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+          step2_6 = mult_round_shift(&t0, &t1, &k__cospi_p24_p08,
+                                     &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+          step2_5 = mult_round_shift(&t2, &t3, &k__cospi_p08_m24,
+                                     &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x4(&step2_1, &step2_2, &step2_6,
+                                             &step2_5);
+          if (overflow) {
+            vp9_highbd_fdct16x16_c(input, output, stride);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        // step 5
+        {
+          step1_0 = ADD_EPI16(step3_0, step2_1);
+          step1_1 = SUB_EPI16(step3_0, step2_1);
+          step1_2 = ADD_EPI16(step3_3, step2_2);
+          step1_3 = SUB_EPI16(step3_3, step2_2);
+          step1_4 = SUB_EPI16(step3_4, step2_5);
+          step1_5 = ADD_EPI16(step3_4, step2_5);
+          step1_6 = SUB_EPI16(step3_7, step2_6);
+          step1_7 = ADD_EPI16(step3_7, step2_6);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x8(&step1_0, &step1_1,
+                                             &step1_2, &step1_3,
+                                             &step1_4, &step1_5,
+                                             &step1_6, &step1_7);
+          if (overflow) {
+            vp9_highbd_fdct16x16_c(input, output, stride);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        // step 6
+        {
+          const __m128i t0 = _mm_unpacklo_epi16(step1_0, step1_7);
+          const __m128i t1 = _mm_unpackhi_epi16(step1_0, step1_7);
+          const __m128i t2 = _mm_unpacklo_epi16(step1_1, step1_6);
+          const __m128i t3 = _mm_unpackhi_epi16(step1_1, step1_6);
+          res01 = mult_round_shift(&t0, &t1, &k__cospi_p30_p02,
+                                   &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+          res09 = mult_round_shift(&t2, &t3, &k__cospi_p14_p18,
+                                   &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+          res15 = mult_round_shift(&t0, &t1, &k__cospi_m02_p30,
+                                   &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+          res07 = mult_round_shift(&t2, &t3, &k__cospi_m18_p14,
+                                   &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x4(&res01, &res09, &res15, &res07);
+          if (overflow) {
+            vp9_highbd_fdct16x16_c(input, output, stride);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+        {
+          const __m128i t0 = _mm_unpacklo_epi16(step1_2, step1_5);
+          const __m128i t1 = _mm_unpackhi_epi16(step1_2, step1_5);
+          const __m128i t2 = _mm_unpacklo_epi16(step1_3, step1_4);
+          const __m128i t3 = _mm_unpackhi_epi16(step1_3, step1_4);
+          res05 = mult_round_shift(&t0, &t1, &k__cospi_p22_p10,
+                                   &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+          res13 = mult_round_shift(&t2, &t3, &k__cospi_p06_p26,
+                                   &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+          res11 = mult_round_shift(&t0, &t1, &k__cospi_m10_p22,
+                                   &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+          res03 = mult_round_shift(&t2, &t3, &k__cospi_m26_p06,
+                                   &k__DCT_CONST_ROUNDING, DCT_CONST_BITS);
+#if DCT_HIGH_BIT_DEPTH
+          overflow = check_epi16_overflow_x4(&res05, &res13, &res11, &res03);
+          if (overflow) {
+            vp9_highbd_fdct16x16_c(input, output, stride);
+            return;
+          }
+#endif  // DCT_HIGH_BIT_DEPTH
+        }
+      }
+      // Transpose the results, do it as two 8x8 transposes.
+      transpose_and_output8x8(&res00, &res01, &res02, &res03,
+                              &res04, &res05, &res06, &res07,
+                              pass, out0, out1);
+      transpose_and_output8x8(&res08, &res09, &res10, &res11,
+                              &res12, &res13, &res14, &res15,
+                              pass, out0 + 8, out1 + 8);
+      if (pass == 0) {
+        out0 += 8*16;
+      } else {
+        out1 += 8*16;
+      }
+    }
+    // Setup in/out for next pass.
+    in = intermediate;
+  }
+}
+
+#undef ADD_EPI16
+#undef SUB_EPI16
diff --git a/media/libvpx/vp9/encoder/x86/vp9_dct_ssse3.c b/media/libvpx/vp9/encoder/x86/vp9_dct_ssse3.c
new file mode 100644
index 000000000..96038fee1
--- /dev/null
+++ b/media/libvpx/vp9/encoder/x86/vp9_dct_ssse3.c
@@ -0,0 +1,471 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <assert.h>
+#if defined(_MSC_VER) && _MSC_VER <= 1500
+// Need to include math.h before calling tmmintrin.h/intrin.h
+// in certain versions of MSVS.
+#include <math.h>
+#endif
+#include <tmmintrin.h>  // SSSE3
+
+#include "./vp9_rtcd.h"
+#include "vp9/common/x86/vp9_idct_intrin_sse2.h"
+
+void vp9_fdct8x8_quant_ssse3(const int16_t *input, int stride,
+                             int16_t* coeff_ptr, intptr_t n_coeffs,
+                             int skip_block, const int16_t* zbin_ptr,
+                             const int16_t* round_ptr, const int16_t* quant_ptr,
+                             const int16_t* quant_shift_ptr,
+                             int16_t* qcoeff_ptr,
+                             int16_t* dqcoeff_ptr, const int16_t* dequant_ptr,
+                             uint16_t* eob_ptr,
+                             const int16_t* scan_ptr,
+                             const int16_t* iscan_ptr) {
+  __m128i zero;
+  int pass;
+  // Constants
+  //    When we use them, in one case, they are all the same. In all others
+  //    it's a pair of them that we need to repeat four times. This is done
+  //    by constructing the 32 bit constant corresponding to that pair.
+  const __m128i k__dual_p16_p16 = dual_set_epi16(23170, 23170);
+  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
+  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
+  const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
+  const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
+  const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64);
+  const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_64);
+  const __m128i k__cospi_p12_p20 = pair_set_epi16(cospi_12_64, cospi_20_64);
+  const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64, cospi_12_64);
+  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
+  // Load input
+  __m128i in0  = _mm_load_si128((const __m128i *)(input + 0 * stride));
+  __m128i in1  = _mm_load_si128((const __m128i *)(input + 1 * stride));
+  __m128i in2  = _mm_load_si128((const __m128i *)(input + 2 * stride));
+  __m128i in3  = _mm_load_si128((const __m128i *)(input + 3 * stride));
+  __m128i in4  = _mm_load_si128((const __m128i *)(input + 4 * stride));
+  __m128i in5  = _mm_load_si128((const __m128i *)(input + 5 * stride));
+  __m128i in6  = _mm_load_si128((const __m128i *)(input + 6 * stride));
+  __m128i in7  = _mm_load_si128((const __m128i *)(input + 7 * stride));
+  __m128i *in[8];
+  int index = 0;
+
+  (void)scan_ptr;
+  (void)zbin_ptr;
+  (void)quant_shift_ptr;
+  (void)coeff_ptr;
+
+  // Pre-condition input (shift by two)
+  in0 = _mm_slli_epi16(in0, 2);
+  in1 = _mm_slli_epi16(in1, 2);
+  in2 = _mm_slli_epi16(in2, 2);
+  in3 = _mm_slli_epi16(in3, 2);
+  in4 = _mm_slli_epi16(in4, 2);
+  in5 = _mm_slli_epi16(in5, 2);
+  in6 = _mm_slli_epi16(in6, 2);
+  in7 = _mm_slli_epi16(in7, 2);
+
+  in[0] = &in0;
+  in[1] = &in1;
+  in[2] = &in2;
+  in[3] = &in3;
+  in[4] = &in4;
+  in[5] = &in5;
+  in[6] = &in6;
+  in[7] = &in7;
+
+  // We do two passes, first the columns, then the rows. The results of the
+  // first pass are transposed so that the same column code can be reused. The
+  // results of the second pass are also transposed so that the rows (processed
+  // as columns) are put back in row positions.
+  for (pass = 0; pass < 2; pass++) {
+    // To store results of each pass before the transpose.
+    __m128i res0, res1, res2, res3, res4, res5, res6, res7;
+    // Add/subtract
+    const __m128i q0 = _mm_add_epi16(in0, in7);
+    const __m128i q1 = _mm_add_epi16(in1, in6);
+    const __m128i q2 = _mm_add_epi16(in2, in5);
+    const __m128i q3 = _mm_add_epi16(in3, in4);
+    const __m128i q4 = _mm_sub_epi16(in3, in4);
+    const __m128i q5 = _mm_sub_epi16(in2, in5);
+    const __m128i q6 = _mm_sub_epi16(in1, in6);
+    const __m128i q7 = _mm_sub_epi16(in0, in7);
+    // Work on first four results
+    {
+      // Add/subtract
+      const __m128i r0 = _mm_add_epi16(q0, q3);
+      const __m128i r1 = _mm_add_epi16(q1, q2);
+      const __m128i r2 = _mm_sub_epi16(q1, q2);
+      const __m128i r3 = _mm_sub_epi16(q0, q3);
+      // Interleave to do the multiply by constants which gets us into 32bits
+      const __m128i t0 = _mm_unpacklo_epi16(r0, r1);
+      const __m128i t1 = _mm_unpackhi_epi16(r0, r1);
+      const __m128i t2 = _mm_unpacklo_epi16(r2, r3);
+      const __m128i t3 = _mm_unpackhi_epi16(r2, r3);
+
+      const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16);
+      const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_p16);
+      const __m128i u2 = _mm_madd_epi16(t0, k__cospi_p16_m16);
+      const __m128i u3 = _mm_madd_epi16(t1, k__cospi_p16_m16);
+
+      const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p24_p08);
+      const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p24_p08);
+      const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m08_p24);
+      const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m08_p24);
+      // dct_const_round_shift
+
+      const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
+      const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
+      const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
+      const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
+
+      const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
+      const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
+      const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
+      const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
+
+      const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+      const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+      const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+      const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+
+      const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
+      const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
+      const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
+      const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
+      // Combine
+
+      res0 = _mm_packs_epi32(w0, w1);
+      res4 = _mm_packs_epi32(w2, w3);
+      res2 = _mm_packs_epi32(w4, w5);
+      res6 = _mm_packs_epi32(w6, w7);
+    }
+    // Work on next four results
+    {
+      // Interleave to do the multiply by constants which gets us into 32bits
+      const __m128i d0 = _mm_sub_epi16(q6, q5);
+      const __m128i d1 = _mm_add_epi16(q6, q5);
+      const __m128i r0 = _mm_mulhrs_epi16(d0, k__dual_p16_p16);
+      const __m128i r1 = _mm_mulhrs_epi16(d1, k__dual_p16_p16);
+
+      // Add/subtract
+      const __m128i x0 = _mm_add_epi16(q4, r0);
+      const __m128i x1 = _mm_sub_epi16(q4, r0);
+      const __m128i x2 = _mm_sub_epi16(q7, r1);
+      const __m128i x3 = _mm_add_epi16(q7, r1);
+      // Interleave to do the multiply by constants which gets us into 32bits
+      const __m128i t0 = _mm_unpacklo_epi16(x0, x3);
+      const __m128i t1 = _mm_unpackhi_epi16(x0, x3);
+      const __m128i t2 = _mm_unpacklo_epi16(x1, x2);
+      const __m128i t3 = _mm_unpackhi_epi16(x1, x2);
+      const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p28_p04);
+      const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p28_p04);
+      const __m128i u2 = _mm_madd_epi16(t0, k__cospi_m04_p28);
+      const __m128i u3 = _mm_madd_epi16(t1, k__cospi_m04_p28);
+      const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p12_p20);
+      const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p12_p20);
+      const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m20_p12);
+      const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m20_p12);
+      // dct_const_round_shift
+      const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
+      const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
+      const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
+      const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
+      const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
+      const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
+      const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
+      const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
+      const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
+      const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
+      const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
+      const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
+      const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
+      const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
+      const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
+      const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
+      // Combine
+      res1 = _mm_packs_epi32(w0, w1);
+      res7 = _mm_packs_epi32(w2, w3);
+      res5 = _mm_packs_epi32(w4, w5);
+      res3 = _mm_packs_epi32(w6, w7);
+    }
+    // Transpose the 8x8.
+    {
+      // 00 01 02 03 04 05 06 07
+      // 10 11 12 13 14 15 16 17
+      // 20 21 22 23 24 25 26 27
+      // 30 31 32 33 34 35 36 37
+      // 40 41 42 43 44 45 46 47
+      // 50 51 52 53 54 55 56 57
+      // 60 61 62 63 64 65 66 67
+      // 70 71 72 73 74 75 76 77
+      const __m128i tr0_0 = _mm_unpacklo_epi16(res0, res1);
+      const __m128i tr0_1 = _mm_unpacklo_epi16(res2, res3);
+      const __m128i tr0_2 = _mm_unpackhi_epi16(res0, res1);
+      const __m128i tr0_3 = _mm_unpackhi_epi16(res2, res3);
+      const __m128i tr0_4 = _mm_unpacklo_epi16(res4, res5);
+      const __m128i tr0_5 = _mm_unpacklo_epi16(res6, res7);
+      const __m128i tr0_6 = _mm_unpackhi_epi16(res4, res5);
+      const __m128i tr0_7 = _mm_unpackhi_epi16(res6, res7);
+      // 00 10 01 11 02 12 03 13
+      // 20 30 21 31 22 32 23 33
+      // 04 14 05 15 06 16 07 17
+      // 24 34 25 35 26 36 27 37
+      // 40 50 41 51 42 52 43 53
+      // 60 70 61 71 62 72 63 73
+      // 54 54 55 55 56 56 57 57
+      // 64 74 65 75 66 76 67 77
+      const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
+      const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3);
+      const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
+      const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3);
+      const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5);
+      const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
+      const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5);
+      const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
+      // 00 10 20 30 01 11 21 31
+      // 40 50 60 70 41 51 61 71
+      // 02 12 22 32 03 13 23 33
+      // 42 52 62 72 43 53 63 73
+      // 04 14 24 34 05 15 21 36
+      // 44 54 64 74 45 55 61 76
+      // 06 16 26 36 07 17 27 37
+      // 46 56 66 76 47 57 67 77
+      in0 = _mm_unpacklo_epi64(tr1_0, tr1_4);
+      in1 = _mm_unpackhi_epi64(tr1_0, tr1_4);
+      in2 = _mm_unpacklo_epi64(tr1_2, tr1_6);
+      in3 = _mm_unpackhi_epi64(tr1_2, tr1_6);
+      in4 = _mm_unpacklo_epi64(tr1_1, tr1_5);
+      in5 = _mm_unpackhi_epi64(tr1_1, tr1_5);
+      in6 = _mm_unpacklo_epi64(tr1_3, tr1_7);
+      in7 = _mm_unpackhi_epi64(tr1_3, tr1_7);
+      // 00 10 20 30 40 50 60 70
+      // 01 11 21 31 41 51 61 71
+      // 02 12 22 32 42 52 62 72
+      // 03 13 23 33 43 53 63 73
+      // 04 14 24 34 44 54 64 74
+      // 05 15 25 35 45 55 65 75
+      // 06 16 26 36 46 56 66 76
+      // 07 17 27 37 47 57 67 77
+    }
+  }
+  // Post-condition output and store it
+  {
+    // Post-condition (division by two)
+    //    division of two 16 bits signed numbers using shifts
+    //    n / 2 = (n - (n >> 15)) >> 1
+    const __m128i sign_in0 = _mm_srai_epi16(in0, 15);
+    const __m128i sign_in1 = _mm_srai_epi16(in1, 15);
+    const __m128i sign_in2 = _mm_srai_epi16(in2, 15);
+    const __m128i sign_in3 = _mm_srai_epi16(in3, 15);
+    const __m128i sign_in4 = _mm_srai_epi16(in4, 15);
+    const __m128i sign_in5 = _mm_srai_epi16(in5, 15);
+    const __m128i sign_in6 = _mm_srai_epi16(in6, 15);
+    const __m128i sign_in7 = _mm_srai_epi16(in7, 15);
+    in0 = _mm_sub_epi16(in0, sign_in0);
+    in1 = _mm_sub_epi16(in1, sign_in1);
+    in2 = _mm_sub_epi16(in2, sign_in2);
+    in3 = _mm_sub_epi16(in3, sign_in3);
+    in4 = _mm_sub_epi16(in4, sign_in4);
+    in5 = _mm_sub_epi16(in5, sign_in5);
+    in6 = _mm_sub_epi16(in6, sign_in6);
+    in7 = _mm_sub_epi16(in7, sign_in7);
+    in0 = _mm_srai_epi16(in0, 1);
+    in1 = _mm_srai_epi16(in1, 1);
+    in2 = _mm_srai_epi16(in2, 1);
+    in3 = _mm_srai_epi16(in3, 1);
+    in4 = _mm_srai_epi16(in4, 1);
+    in5 = _mm_srai_epi16(in5, 1);
+    in6 = _mm_srai_epi16(in6, 1);
+    in7 = _mm_srai_epi16(in7, 1);
+  }
+
+  iscan_ptr += n_coeffs;
+  qcoeff_ptr += n_coeffs;
+  dqcoeff_ptr += n_coeffs;
+  n_coeffs = -n_coeffs;
+  zero = _mm_setzero_si128();
+
+  if (!skip_block) {
+    __m128i eob;
+    __m128i round, quant, dequant, thr;
+    int16_t nzflag;
+    {
+      __m128i coeff0, coeff1;
+
+      // Setup global values
+      {
+        round = _mm_load_si128((const __m128i*)round_ptr);
+        quant = _mm_load_si128((const __m128i*)quant_ptr);
+        dequant = _mm_load_si128((const __m128i*)dequant_ptr);
+      }
+
+      {
+        __m128i coeff0_sign, coeff1_sign;
+        __m128i qcoeff0, qcoeff1;
+        __m128i qtmp0, qtmp1;
+        // Do DC and first 15 AC
+        coeff0 = *in[0];
+        coeff1 = *in[1];
+
+        // Poor man's sign extract
+        coeff0_sign = _mm_srai_epi16(coeff0, 15);
+        coeff1_sign = _mm_srai_epi16(coeff1, 15);
+        qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        qcoeff0 = _mm_adds_epi16(qcoeff0, round);
+        round = _mm_unpackhi_epi64(round, round);
+        qcoeff1 = _mm_adds_epi16(qcoeff1, round);
+        qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
+        quant = _mm_unpackhi_epi64(quant, quant);
+        qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
+
+        // Reinsert signs
+        qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), qcoeff0);
+        _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
+
+        coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
+        dequant = _mm_unpackhi_epi64(dequant, dequant);
+        coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+
+        _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), coeff0);
+        _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, coeff1);
+      }
+
+      {
+        // Scan for eob
+        __m128i zero_coeff0, zero_coeff1;
+        __m128i nzero_coeff0, nzero_coeff1;
+        __m128i iscan0, iscan1;
+        __m128i eob1;
+        zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
+        zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
+        nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
+        nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
+        iscan0 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs));
+        iscan1 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs) + 1);
+        // Add one to convert from indices to counts
+        iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
+        iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
+        eob = _mm_and_si128(iscan0, nzero_coeff0);
+        eob1 = _mm_and_si128(iscan1, nzero_coeff1);
+        eob = _mm_max_epi16(eob, eob1);
+      }
+      n_coeffs += 8 * 2;
+    }
+
+    // AC only loop
+    index = 2;
+    thr = _mm_srai_epi16(dequant, 1);
+    while (n_coeffs < 0) {
+      __m128i coeff0, coeff1;
+      {
+        __m128i coeff0_sign, coeff1_sign;
+        __m128i qcoeff0, qcoeff1;
+        __m128i qtmp0, qtmp1;
+
+        assert(index < (int)(sizeof(in) / sizeof(in[0])) - 1);
+        coeff0 = *in[index];
+        coeff1 = *in[index + 1];
+
+        // Poor man's sign extract
+        coeff0_sign = _mm_srai_epi16(coeff0, 15);
+        coeff1_sign = _mm_srai_epi16(coeff1, 15);
+        qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        nzflag = _mm_movemask_epi8(_mm_cmpgt_epi16(qcoeff0, thr)) |
+            _mm_movemask_epi8(_mm_cmpgt_epi16(qcoeff1, thr));
+
+        if (nzflag) {
+          qcoeff0 = _mm_adds_epi16(qcoeff0, round);
+          qcoeff1 = _mm_adds_epi16(qcoeff1, round);
+          qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
+          qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
+
+          // Reinsert signs
+          qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign);
+          qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign);
+          qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+          qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+          _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), qcoeff0);
+          _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
+
+          coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
+          coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+
+          _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), coeff0);
+          _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, coeff1);
+        } else {
+          _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), zero);
+          _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, zero);
+
+          _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), zero);
+          _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, zero);
+        }
+      }
+
+      if (nzflag) {
+        // Scan for eob
+        __m128i zero_coeff0, zero_coeff1;
+        __m128i nzero_coeff0, nzero_coeff1;
+        __m128i iscan0, iscan1;
+        __m128i eob0, eob1;
+        zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
+        zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
+        nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
+        nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
+        iscan0 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs));
+        iscan1 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs) + 1);
+        // Add one to convert from indices to counts
+        iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
+        iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
+        eob0 = _mm_and_si128(iscan0, nzero_coeff0);
+        eob1 = _mm_and_si128(iscan1, nzero_coeff1);
+        eob0 = _mm_max_epi16(eob0, eob1);
+        eob = _mm_max_epi16(eob, eob0);
+      }
+      n_coeffs += 8 * 2;
+      index += 2;
+    }
+
+    // Accumulate EOB
+    {
+      __m128i eob_shuffled;
+      eob_shuffled = _mm_shuffle_epi32(eob, 0xe);
+      eob = _mm_max_epi16(eob, eob_shuffled);
+      eob_shuffled = _mm_shufflelo_epi16(eob, 0xe);
+      eob = _mm_max_epi16(eob, eob_shuffled);
+      eob_shuffled = _mm_shufflelo_epi16(eob, 0x1);
+      eob = _mm_max_epi16(eob, eob_shuffled);
+      *eob_ptr = _mm_extract_epi16(eob, 1);
+    }
+  } else {
+    do {
+      _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), zero);
+      _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, zero);
+      _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), zero);
+      _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, zero);
+      n_coeffs += 8 * 2;
+    } while (n_coeffs < 0);
+    *eob_ptr = 0;
+  }
+}
diff --git a/media/libvpx/vp9/encoder/x86/vp9_dct_ssse3_x86_64.asm b/media/libvpx/vp9/encoder/x86/vp9_dct_ssse3_x86_64.asm
new file mode 100644
index 000000000..3a29aba6f
--- /dev/null
+++ b/media/libvpx/vp9/encoder/x86/vp9_dct_ssse3_x86_64.asm
@@ -0,0 +1,255 @@
+;
+;  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+%include "third_party/x86inc/x86inc.asm"
+
+; This file provides SSSE3 version of the forward transformation. Part
+; of the macro definitions are originally derived from the ffmpeg project.
+; The current version applies to x86 64-bit only.
+
+SECTION_RODATA
+
+pw_11585x2: times 8 dw 23170
+pd_8192:    times 4 dd 8192
+
+%macro TRANSFORM_COEFFS 2
+pw_%1_%2:   dw  %1,  %2,  %1,  %2,  %1,  %2,  %1,  %2
+pw_%2_m%1:  dw  %2, -%1,  %2, -%1,  %2, -%1,  %2, -%1
+%endmacro
+
+TRANSFORM_COEFFS 11585,  11585
+TRANSFORM_COEFFS 15137,   6270
+TRANSFORM_COEFFS 16069,   3196
+TRANSFORM_COEFFS  9102,  13623
+
+SECTION .text
+
+%if ARCH_X86_64
+%macro SUM_SUB 3
+  psubw  m%3, m%1, m%2
+  paddw  m%1, m%2
+  SWAP    %2, %3
+%endmacro
+
+; butterfly operation
+%macro MUL_ADD_2X 6 ; dst1, dst2, src, round, coefs1, coefs2
+  pmaddwd            m%1, m%3, %5
+  pmaddwd            m%2, m%3, %6
+  paddd              m%1,  %4
+  paddd              m%2,  %4
+  psrad              m%1,  14
+  psrad              m%2,  14
+%endmacro
+
+%macro BUTTERFLY_4X 7 ; dst1, dst2, coef1, coef2, round, tmp1, tmp2
+  punpckhwd          m%6, m%2, m%1
+  MUL_ADD_2X         %7,  %6,  %6,  %5, [pw_%4_%3], [pw_%3_m%4]
+  punpcklwd          m%2, m%1
+  MUL_ADD_2X         %1,  %2,  %2,  %5, [pw_%4_%3], [pw_%3_m%4]
+  packssdw           m%1, m%7
+  packssdw           m%2, m%6
+%endmacro
+
+; matrix transpose
+%macro INTERLEAVE_2X 4
+  punpckh%1          m%4, m%2, m%3
+  punpckl%1          m%2, m%3
+  SWAP               %3,  %4
+%endmacro
+
+%macro TRANSPOSE8X8 9
+  INTERLEAVE_2X  wd, %1, %2, %9
+  INTERLEAVE_2X  wd, %3, %4, %9
+  INTERLEAVE_2X  wd, %5, %6, %9
+  INTERLEAVE_2X  wd, %7, %8, %9
+
+  INTERLEAVE_2X  dq, %1, %3, %9
+  INTERLEAVE_2X  dq, %2, %4, %9
+  INTERLEAVE_2X  dq, %5, %7, %9
+  INTERLEAVE_2X  dq, %6, %8, %9
+
+  INTERLEAVE_2X  qdq, %1, %5, %9
+  INTERLEAVE_2X  qdq, %3, %7, %9
+  INTERLEAVE_2X  qdq, %2, %6, %9
+  INTERLEAVE_2X  qdq, %4, %8, %9
+
+  SWAP  %2, %5
+  SWAP  %4, %7
+%endmacro
+
+; 1D forward 8x8 DCT transform
+%macro FDCT8_1D 1
+  SUM_SUB            0,  7,  9
+  SUM_SUB            1,  6,  9
+  SUM_SUB            2,  5,  9
+  SUM_SUB            3,  4,  9
+
+  SUM_SUB            0,  3,  9
+  SUM_SUB            1,  2,  9
+  SUM_SUB            6,  5,  9
+%if %1 == 0
+  SUM_SUB            0,  1,  9
+%endif
+
+  BUTTERFLY_4X       2,  3,  6270,  15137,  m8,  9,  10
+
+  pmulhrsw           m6, m12
+  pmulhrsw           m5, m12
+%if %1 == 0
+  pmulhrsw           m0, m12
+  pmulhrsw           m1, m12
+%else
+  BUTTERFLY_4X       1,  0,  11585, 11585,  m8,  9,  10
+  SWAP               0,  1
+%endif
+
+  SUM_SUB            4,  5,  9
+  SUM_SUB            7,  6,  9
+  BUTTERFLY_4X       4,  7,  3196,  16069,  m8,  9,  10
+  BUTTERFLY_4X       5,  6,  13623,  9102,  m8,  9,  10
+  SWAP               1,  4
+  SWAP               3,  6
+%endmacro
+
+%macro DIVIDE_ROUND_2X 4 ; dst1, dst2, tmp1, tmp2
+  psraw              m%3, m%1, 15
+  psraw              m%4, m%2, 15
+  psubw              m%1, m%3
+  psubw              m%2, m%4
+  psraw              m%1, 1
+  psraw              m%2, 1
+%endmacro
+
+INIT_XMM ssse3
+cglobal fdct8x8, 3, 5, 13, input, output, stride
+
+  mova               m8, [pd_8192]
+  mova              m12, [pw_11585x2]
+  pxor              m11, m11
+
+  lea                r3, [2 * strideq]
+  lea                r4, [4 * strideq]
+  mova               m0, [inputq]
+  mova               m1, [inputq + r3]
+  lea                inputq, [inputq + r4]
+  mova               m2, [inputq]
+  mova               m3, [inputq + r3]
+  lea                inputq, [inputq + r4]
+  mova               m4, [inputq]
+  mova               m5, [inputq + r3]
+  lea                inputq, [inputq + r4]
+  mova               m6, [inputq]
+  mova               m7, [inputq + r3]
+
+  ; left shift by 2 to increase forward transformation precision
+  psllw              m0, 2
+  psllw              m1, 2
+  psllw              m2, 2
+  psllw              m3, 2
+  psllw              m4, 2
+  psllw              m5, 2
+  psllw              m6, 2
+  psllw              m7, 2
+
+  ; column transform
+  FDCT8_1D  0
+  TRANSPOSE8X8 0, 1, 2, 3, 4, 5, 6, 7, 9
+
+  FDCT8_1D  1
+  TRANSPOSE8X8 0, 1, 2, 3, 4, 5, 6, 7, 9
+
+  DIVIDE_ROUND_2X   0, 1, 9, 10
+  DIVIDE_ROUND_2X   2, 3, 9, 10
+  DIVIDE_ROUND_2X   4, 5, 9, 10
+  DIVIDE_ROUND_2X   6, 7, 9, 10
+
+  mova              [outputq +   0], m0
+  mova              [outputq +  16], m1
+  mova              [outputq +  32], m2
+  mova              [outputq +  48], m3
+  mova              [outputq +  64], m4
+  mova              [outputq +  80], m5
+  mova              [outputq +  96], m6
+  mova              [outputq + 112], m7
+
+  RET
+
+%macro HMD8_1D 0
+  psubw              m8, m0, m1
+  psubw              m9, m2, m3
+  paddw              m0, m1
+  paddw              m2, m3
+  SWAP               1, 8
+  SWAP               3, 9
+  psubw              m8, m4, m5
+  psubw              m9, m6, m7
+  paddw              m4, m5
+  paddw              m6, m7
+  SWAP               5, 8
+  SWAP               7, 9
+
+  psubw              m8, m0, m2
+  psubw              m9, m1, m3
+  paddw              m0, m2
+  paddw              m1, m3
+  SWAP               2, 8
+  SWAP               3, 9
+  psubw              m8, m4, m6
+  psubw              m9, m5, m7
+  paddw              m4, m6
+  paddw              m5, m7
+  SWAP               6, 8
+  SWAP               7, 9
+
+  psubw              m8, m0, m4
+  psubw              m9, m1, m5
+  paddw              m0, m4
+  paddw              m1, m5
+  SWAP               4, 8
+  SWAP               5, 9
+  psubw              m8, m2, m6
+  psubw              m9, m3, m7
+  paddw              m2, m6
+  paddw              m3, m7
+  SWAP               6, 8
+  SWAP               7, 9
+%endmacro
+
+INIT_XMM ssse3
+cglobal hadamard_8x8, 3, 5, 10, input, stride, output
+  lea                r3, [2 * strideq]
+  lea                r4, [4 * strideq]
+
+  mova               m0, [inputq]
+  mova               m1, [inputq + r3]
+  lea                inputq, [inputq + r4]
+  mova               m2, [inputq]
+  mova               m3, [inputq + r3]
+  lea                inputq, [inputq + r4]
+  mova               m4, [inputq]
+  mova               m5, [inputq + r3]
+  lea                inputq, [inputq + r4]
+  mova               m6, [inputq]
+  mova               m7, [inputq + r3]
+
+  HMD8_1D
+  TRANSPOSE8X8 0, 1, 2, 3, 4, 5, 6, 7, 9
+  HMD8_1D
+
+  mova              [outputq +   0], m0
+  mova              [outputq +  16], m1
+  mova              [outputq +  32], m2
+  mova              [outputq +  48], m3
+  mova              [outputq +  64], m4
+  mova              [outputq +  80], m5
+  mova              [outputq +  96], m6
+  mova              [outputq + 112], m7
+
+  RET
+%endif
diff --git a/media/libvpx/vp9/encoder/x86/vp9_denoiser_sse2.c b/media/libvpx/vp9/encoder/x86/vp9_denoiser_sse2.c
new file mode 100644
index 000000000..bf7c7af77
--- /dev/null
+++ b/media/libvpx/vp9/encoder/x86/vp9_denoiser_sse2.c
@@ -0,0 +1,375 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <emmintrin.h>
+
+#include "./vpx_config.h"
+#include "./vp9_rtcd.h"
+
+#include "vpx_ports/emmintrin_compat.h"
+#include "vpx/vpx_integer.h"
+#include "vp9/common/vp9_reconinter.h"
+#include "vp9/encoder/vp9_context_tree.h"
+#include "vp9/encoder/vp9_denoiser.h"
+#include "vpx_mem/vpx_mem.h"
+
+// Compute the sum of all pixel differences of this MB.
+static INLINE int sum_diff_16x1(__m128i acc_diff) {
+  const __m128i k_1 = _mm_set1_epi16(1);
+  const __m128i acc_diff_lo =
+      _mm_srai_epi16(_mm_unpacklo_epi8(acc_diff, acc_diff), 8);
+  const __m128i acc_diff_hi =
+      _mm_srai_epi16(_mm_unpackhi_epi8(acc_diff, acc_diff), 8);
+  const __m128i acc_diff_16 = _mm_add_epi16(acc_diff_lo, acc_diff_hi);
+  const __m128i hg_fe_dc_ba = _mm_madd_epi16(acc_diff_16, k_1);
+  const __m128i hgfe_dcba =
+      _mm_add_epi32(hg_fe_dc_ba, _mm_srli_si128(hg_fe_dc_ba, 8));
+  const __m128i hgfedcba =
+      _mm_add_epi32(hgfe_dcba, _mm_srli_si128(hgfe_dcba, 4));
+  return _mm_cvtsi128_si32(hgfedcba);
+}
+
+// Denoise a 16x1 vector.
+static INLINE __m128i vp9_denoiser_16x1_sse2(const uint8_t *sig,
+                                             const uint8_t *mc_running_avg_y,
+                                             uint8_t *running_avg_y,
+                                             const __m128i *k_0,
+                                             const __m128i *k_4,
+                                             const __m128i *k_8,
+                                             const __m128i *k_16,
+                                             const __m128i *l3,
+                                             const __m128i *l32,
+                                             const __m128i *l21,
+                                             __m128i acc_diff) {
+  // Calculate differences
+  const __m128i v_sig = _mm_loadu_si128((const __m128i *)(&sig[0]));
+  const __m128i v_mc_running_avg_y =
+      _mm_loadu_si128((const __m128i *)(&mc_running_avg_y[0]));
+  __m128i v_running_avg_y;
+  const __m128i pdiff = _mm_subs_epu8(v_mc_running_avg_y, v_sig);
+  const __m128i ndiff = _mm_subs_epu8(v_sig, v_mc_running_avg_y);
+  // Obtain the sign. FF if diff is negative.
+  const __m128i diff_sign = _mm_cmpeq_epi8(pdiff, *k_0);
+  // Clamp absolute difference to 16 to be used to get mask. Doing this
+  // allows us to use _mm_cmpgt_epi8, which operates on signed byte.
+  const __m128i clamped_absdiff =
+      _mm_min_epu8(_mm_or_si128(pdiff, ndiff), *k_16);
+  // Get masks for l2 l1 and l0 adjustments.
+  const __m128i mask2 = _mm_cmpgt_epi8(*k_16, clamped_absdiff);
+  const __m128i mask1 = _mm_cmpgt_epi8(*k_8, clamped_absdiff);
+  const __m128i mask0 = _mm_cmpgt_epi8(*k_4, clamped_absdiff);
+  // Get adjustments for l2, l1, and l0.
+  __m128i adj2 = _mm_and_si128(mask2, *l32);
+  const __m128i adj1 = _mm_and_si128(mask1, *l21);
+  const __m128i adj0 = _mm_and_si128(mask0, clamped_absdiff);
+  __m128i adj,  padj, nadj;
+
+  // Combine the adjustments and get absolute adjustments.
+  adj2 = _mm_add_epi8(adj2, adj1);
+  adj = _mm_sub_epi8(*l3, adj2);
+  adj = _mm_andnot_si128(mask0, adj);
+  adj = _mm_or_si128(adj, adj0);
+
+  // Restore the sign and get positive and negative adjustments.
+  padj = _mm_andnot_si128(diff_sign, adj);
+  nadj = _mm_and_si128(diff_sign, adj);
+
+  // Calculate filtered value.
+  v_running_avg_y = _mm_adds_epu8(v_sig, padj);
+  v_running_avg_y = _mm_subs_epu8(v_running_avg_y, nadj);
+  _mm_storeu_si128((__m128i *)running_avg_y, v_running_avg_y);
+
+  // Adjustments <=7, and each element in acc_diff can fit in signed
+  // char.
+  acc_diff = _mm_adds_epi8(acc_diff, padj);
+  acc_diff = _mm_subs_epi8(acc_diff, nadj);
+  return acc_diff;
+}
+
+// Denoise a 16x1 vector with a weaker filter.
+static INLINE __m128i vp9_denoiser_adj_16x1_sse2(
+    const uint8_t *sig, const uint8_t *mc_running_avg_y,
+    uint8_t *running_avg_y, const __m128i k_0,
+    const __m128i k_delta, __m128i acc_diff) {
+  __m128i v_running_avg_y = _mm_loadu_si128((__m128i *)(&running_avg_y[0]));
+  // Calculate differences.
+  const __m128i v_sig = _mm_loadu_si128((const __m128i *)(&sig[0]));
+  const __m128i v_mc_running_avg_y =
+      _mm_loadu_si128((const __m128i *)(&mc_running_avg_y[0]));
+  const __m128i pdiff = _mm_subs_epu8(v_mc_running_avg_y, v_sig);
+  const __m128i ndiff = _mm_subs_epu8(v_sig, v_mc_running_avg_y);
+  // Obtain the sign. FF if diff is negative.
+  const __m128i diff_sign = _mm_cmpeq_epi8(pdiff, k_0);
+  // Clamp absolute difference to delta to get the adjustment.
+  const __m128i adj =
+      _mm_min_epu8(_mm_or_si128(pdiff, ndiff), k_delta);
+  // Restore the sign and get positive and negative adjustments.
+  __m128i padj, nadj;
+  padj = _mm_andnot_si128(diff_sign, adj);
+  nadj = _mm_and_si128(diff_sign, adj);
+  // Calculate filtered value.
+  v_running_avg_y = _mm_subs_epu8(v_running_avg_y, padj);
+  v_running_avg_y = _mm_adds_epu8(v_running_avg_y, nadj);
+  _mm_storeu_si128((__m128i *)running_avg_y, v_running_avg_y);
+
+  // Accumulate the adjustments.
+  acc_diff = _mm_subs_epi8(acc_diff, padj);
+  acc_diff = _mm_adds_epi8(acc_diff, nadj);
+  return acc_diff;
+}
+
+// Denoiser for 4xM and 8xM blocks.
+static int vp9_denoiser_NxM_sse2_small(
+    const uint8_t *sig, int sig_stride, const uint8_t *mc_running_avg_y,
+    int mc_avg_y_stride, uint8_t *running_avg_y, int avg_y_stride,
+    int increase_denoising, BLOCK_SIZE bs, int motion_magnitude, int width) {
+  int sum_diff_thresh, r, sum_diff = 0;
+  const int shift_inc  = (increase_denoising &&
+                          motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ?
+                         1 : 0;
+  uint8_t sig_buffer[8][16], mc_running_buffer[8][16], running_buffer[8][16];
+  __m128i acc_diff = _mm_setzero_si128();
+  const __m128i k_0 = _mm_setzero_si128();
+  const __m128i k_4 = _mm_set1_epi8(4 + shift_inc);
+  const __m128i k_8 = _mm_set1_epi8(8);
+  const __m128i k_16 = _mm_set1_epi8(16);
+  // Modify each level's adjustment according to motion_magnitude.
+  const __m128i l3 = _mm_set1_epi8(
+      (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ? 7 + shift_inc : 6);
+  // Difference between level 3 and level 2 is 2.
+  const __m128i l32 = _mm_set1_epi8(2);
+  // Difference between level 2 and level 1 is 1.
+  const __m128i l21 = _mm_set1_epi8(1);
+  const uint8_t shift = (width == 4) ? 2 : 1;
+
+  for (r = 0; r < ((4 << b_height_log2_lookup[bs]) >> shift); ++r) {
+    memcpy(sig_buffer[r], sig, width);
+    memcpy(sig_buffer[r] + width, sig + sig_stride, width);
+    memcpy(mc_running_buffer[r], mc_running_avg_y, width);
+    memcpy(mc_running_buffer[r] + width,
+           mc_running_avg_y + mc_avg_y_stride, width);
+    memcpy(running_buffer[r], running_avg_y, width);
+    memcpy(running_buffer[r] + width, running_avg_y + avg_y_stride, width);
+    if (width == 4) {
+      memcpy(sig_buffer[r] + width * 2, sig + sig_stride * 2, width);
+      memcpy(sig_buffer[r] + width * 3, sig + sig_stride * 3, width);
+      memcpy(mc_running_buffer[r] + width * 2,
+             mc_running_avg_y + mc_avg_y_stride * 2, width);
+      memcpy(mc_running_buffer[r] + width * 3,
+             mc_running_avg_y + mc_avg_y_stride * 3, width);
+      memcpy(running_buffer[r] + width * 2,
+             running_avg_y + avg_y_stride * 2, width);
+      memcpy(running_buffer[r] + width * 3,
+             running_avg_y + avg_y_stride * 3, width);
+    }
+    acc_diff = vp9_denoiser_16x1_sse2(sig_buffer[r],
+                                      mc_running_buffer[r],
+                                      running_buffer[r],
+                                      &k_0, &k_4, &k_8, &k_16,
+                                      &l3, &l32, &l21, acc_diff);
+    memcpy(running_avg_y, running_buffer[r], width);
+    memcpy(running_avg_y + avg_y_stride, running_buffer[r] + width, width);
+    if (width == 4) {
+      memcpy(running_avg_y + avg_y_stride * 2,
+             running_buffer[r] + width * 2, width);
+      memcpy(running_avg_y + avg_y_stride * 3,
+             running_buffer[r] + width * 3, width);
+    }
+    // Update pointers for next iteration.
+    sig += (sig_stride << shift);
+    mc_running_avg_y += (mc_avg_y_stride << shift);
+    running_avg_y += (avg_y_stride << shift);
+  }
+
+  {
+    sum_diff = sum_diff_16x1(acc_diff);
+    sum_diff_thresh = total_adj_strong_thresh(bs, increase_denoising);
+    if (abs(sum_diff) > sum_diff_thresh) {
+      // Before returning to copy the block (i.e., apply no denoising),
+      // check if we can still apply some (weaker) temporal filtering to
+      // this block, that would otherwise not be denoised at all. Simplest
+      // is to apply an additional adjustment to running_avg_y to bring it
+      // closer to sig. The adjustment is capped by a maximum delta, and
+      // chosen such that in most cases the resulting sum_diff will be
+      // within the acceptable range given by sum_diff_thresh.
+
+      // The delta is set by the excess of absolute pixel diff over the
+      // threshold.
+      const int delta = ((abs(sum_diff) - sum_diff_thresh) >>
+                         num_pels_log2_lookup[bs]) + 1;
+      // Only apply the adjustment for max delta up to 3.
+      if (delta < 4) {
+        const __m128i k_delta = _mm_set1_epi8(delta);
+        running_avg_y -= avg_y_stride * (4 << b_height_log2_lookup[bs]);
+        for (r = 0; r < ((4 << b_height_log2_lookup[bs]) >> shift); ++r) {
+          acc_diff = vp9_denoiser_adj_16x1_sse2(
+              sig_buffer[r], mc_running_buffer[r], running_buffer[r],
+              k_0, k_delta, acc_diff);
+          memcpy(running_avg_y, running_buffer[r], width);
+          memcpy(running_avg_y + avg_y_stride,
+                 running_buffer[r] + width, width);
+          if (width == 4) {
+            memcpy(running_avg_y + avg_y_stride * 2,
+                   running_buffer[r] + width * 2, width);
+            memcpy(running_avg_y + avg_y_stride * 3,
+                   running_buffer[r] + width * 3, width);
+          }
+          // Update pointers for next iteration.
+          running_avg_y += (avg_y_stride << shift);
+        }
+        sum_diff = sum_diff_16x1(acc_diff);
+        if (abs(sum_diff) > sum_diff_thresh) {
+          return COPY_BLOCK;
+        }
+      } else {
+        return COPY_BLOCK;
+      }
+    }
+  }
+  return FILTER_BLOCK;
+}
+
+// Denoiser for 16xM, 32xM and 64xM blocks
+static int vp9_denoiser_NxM_sse2_big(const uint8_t *sig, int sig_stride,
+                                     const uint8_t *mc_running_avg_y,
+                                     int mc_avg_y_stride,
+                                     uint8_t *running_avg_y,
+                                     int avg_y_stride,
+                                     int increase_denoising, BLOCK_SIZE bs,
+                                     int motion_magnitude) {
+  int sum_diff_thresh, r, c, sum_diff = 0;
+  const int shift_inc  = (increase_denoising &&
+                          motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ?
+                         1 : 0;
+  __m128i acc_diff[4][4];
+  const __m128i k_0 = _mm_setzero_si128();
+  const __m128i k_4 = _mm_set1_epi8(4 + shift_inc);
+  const __m128i k_8 = _mm_set1_epi8(8);
+  const __m128i k_16 = _mm_set1_epi8(16);
+  // Modify each level's adjustment according to motion_magnitude.
+  const __m128i l3 = _mm_set1_epi8(
+      (motion_magnitude <= MOTION_MAGNITUDE_THRESHOLD) ? 7 + shift_inc : 6);
+  // Difference between level 3 and level 2 is 2.
+  const __m128i l32 = _mm_set1_epi8(2);
+  // Difference between level 2 and level 1 is 1.
+  const __m128i l21 = _mm_set1_epi8(1);
+
+  for (c = 0; c < 4; ++c) {
+    for (r = 0; r < 4; ++r) {
+      acc_diff[c][r] = _mm_setzero_si128();
+    }
+  }
+
+  for (r = 0; r < (4 << b_height_log2_lookup[bs]); ++r) {
+    for (c = 0; c < (4 << b_width_log2_lookup[bs]); c += 16) {
+      acc_diff[c>>4][r>>4] = vp9_denoiser_16x1_sse2(
+          sig, mc_running_avg_y, running_avg_y, &k_0, &k_4,
+          &k_8, &k_16, &l3, &l32, &l21, acc_diff[c>>4][r>>4]);
+      // Update pointers for next iteration.
+      sig += 16;
+      mc_running_avg_y += 16;
+      running_avg_y += 16;
+    }
+
+    if ((r + 1) % 16 == 0 || (bs == BLOCK_16X8 && r == 7)) {
+      for (c = 0; c < (4 << b_width_log2_lookup[bs]); c += 16) {
+        sum_diff += sum_diff_16x1(acc_diff[c>>4][r>>4]);
+      }
+    }
+
+    // Update pointers for next iteration.
+    sig = sig - 16 * ((4 << b_width_log2_lookup[bs]) >> 4) + sig_stride;
+    mc_running_avg_y = mc_running_avg_y -
+                       16 * ((4 << b_width_log2_lookup[bs]) >> 4) +
+                       mc_avg_y_stride;
+    running_avg_y = running_avg_y -
+                    16 * ((4 << b_width_log2_lookup[bs]) >> 4) +
+                    avg_y_stride;
+  }
+
+  {
+    sum_diff_thresh = total_adj_strong_thresh(bs, increase_denoising);
+    if (abs(sum_diff) > sum_diff_thresh) {
+      const int delta = ((abs(sum_diff) - sum_diff_thresh) >>
+                         num_pels_log2_lookup[bs]) + 1;
+
+      // Only apply the adjustment for max delta up to 3.
+      if (delta < 4) {
+        const __m128i k_delta = _mm_set1_epi8(delta);
+        sig -= sig_stride * (4 << b_height_log2_lookup[bs]);
+        mc_running_avg_y -= mc_avg_y_stride * (4 << b_height_log2_lookup[bs]);
+        running_avg_y -= avg_y_stride * (4 << b_height_log2_lookup[bs]);
+        sum_diff = 0;
+        for (r = 0; r < (4 << b_height_log2_lookup[bs]); ++r) {
+          for (c = 0; c < (4 << b_width_log2_lookup[bs]); c += 16) {
+            acc_diff[c>>4][r>>4] = vp9_denoiser_adj_16x1_sse2(
+                sig, mc_running_avg_y, running_avg_y, k_0,
+                k_delta, acc_diff[c>>4][r>>4]);
+            // Update pointers for next iteration.
+            sig += 16;
+            mc_running_avg_y += 16;
+            running_avg_y += 16;
+          }
+
+          if ((r + 1) % 16 == 0 || (bs == BLOCK_16X8 && r == 7)) {
+            for (c = 0; c < (4 << b_width_log2_lookup[bs]); c += 16) {
+              sum_diff += sum_diff_16x1(acc_diff[c>>4][r>>4]);
+            }
+          }
+          sig = sig - 16 * ((4 << b_width_log2_lookup[bs]) >> 4) + sig_stride;
+          mc_running_avg_y = mc_running_avg_y -
+                             16 * ((4 << b_width_log2_lookup[bs]) >> 4) +
+                             mc_avg_y_stride;
+          running_avg_y = running_avg_y -
+                          16 * ((4 << b_width_log2_lookup[bs]) >> 4) +
+                          avg_y_stride;
+        }
+        if (abs(sum_diff) > sum_diff_thresh) {
+          return COPY_BLOCK;
+        }
+      } else {
+        return COPY_BLOCK;
+      }
+    }
+  }
+  return FILTER_BLOCK;
+}
+
+int vp9_denoiser_filter_sse2(const uint8_t *sig, int sig_stride,
+                             const uint8_t *mc_avg,
+                             int mc_avg_stride,
+                             uint8_t *avg, int avg_stride,
+                             int increase_denoising,
+                             BLOCK_SIZE bs,
+                             int motion_magnitude) {
+  if (bs == BLOCK_4X4 || bs == BLOCK_4X8) {
+    return vp9_denoiser_NxM_sse2_small(sig, sig_stride,
+                                       mc_avg, mc_avg_stride,
+                                       avg, avg_stride,
+                                       increase_denoising,
+                                       bs, motion_magnitude, 4);
+  } else if (bs == BLOCK_8X4 || bs == BLOCK_8X8 || bs == BLOCK_8X16) {
+    return vp9_denoiser_NxM_sse2_small(sig, sig_stride,
+                                       mc_avg, mc_avg_stride,
+                                       avg, avg_stride,
+                                       increase_denoising,
+                                       bs, motion_magnitude, 8);
+  } else if (bs == BLOCK_16X8 || bs == BLOCK_16X16 || bs == BLOCK_16X32 ||
+             bs == BLOCK_32X16|| bs == BLOCK_32X32 || bs == BLOCK_32X64 ||
+             bs == BLOCK_64X32 || bs == BLOCK_64X64) {
+    return vp9_denoiser_NxM_sse2_big(sig, sig_stride,
+                                     mc_avg, mc_avg_stride,
+                                     avg, avg_stride,
+                                     increase_denoising,
+                                     bs, motion_magnitude);
+  } else {
+    return COPY_BLOCK;
+  }
+}
diff --git a/media/libvpx/vp9/encoder/x86/vp9_error_intrin_avx2.c b/media/libvpx/vp9/encoder/x86/vp9_error_intrin_avx2.c
new file mode 100644
index 000000000..dfebaab0a
--- /dev/null
+++ b/media/libvpx/vp9/encoder/x86/vp9_error_intrin_avx2.c
@@ -0,0 +1,73 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Usee of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <immintrin.h>  // AVX2
+
+#include "./vp9_rtcd.h"
+#include "vpx/vpx_integer.h"
+
+int64_t vp9_block_error_avx2(const int16_t *coeff,
+                             const int16_t *dqcoeff,
+                             intptr_t block_size,
+                             int64_t *ssz) {
+  __m256i sse_reg, ssz_reg, coeff_reg, dqcoeff_reg;
+  __m256i exp_dqcoeff_lo, exp_dqcoeff_hi, exp_coeff_lo, exp_coeff_hi;
+  __m256i sse_reg_64hi, ssz_reg_64hi;
+  __m128i sse_reg128, ssz_reg128;
+  int64_t sse;
+  int i;
+  const __m256i zero_reg = _mm256_set1_epi16(0);
+
+  // init sse and ssz registerd to zero
+  sse_reg = _mm256_set1_epi16(0);
+  ssz_reg = _mm256_set1_epi16(0);
+
+  for (i = 0 ; i < block_size ; i+= 16) {
+    // load 32 bytes from coeff and dqcoeff
+    coeff_reg = _mm256_loadu_si256((const __m256i *)(coeff + i));
+    dqcoeff_reg = _mm256_loadu_si256((const __m256i *)(dqcoeff + i));
+    // dqcoeff - coeff
+    dqcoeff_reg = _mm256_sub_epi16(dqcoeff_reg, coeff_reg);
+    // madd (dqcoeff - coeff)
+    dqcoeff_reg = _mm256_madd_epi16(dqcoeff_reg, dqcoeff_reg);
+    // madd coeff
+    coeff_reg = _mm256_madd_epi16(coeff_reg, coeff_reg);
+    // expand each double word of madd (dqcoeff - coeff) to quad word
+    exp_dqcoeff_lo = _mm256_unpacklo_epi32(dqcoeff_reg, zero_reg);
+    exp_dqcoeff_hi = _mm256_unpackhi_epi32(dqcoeff_reg, zero_reg);
+    // expand each double word of madd (coeff) to quad word
+    exp_coeff_lo = _mm256_unpacklo_epi32(coeff_reg, zero_reg);
+    exp_coeff_hi = _mm256_unpackhi_epi32(coeff_reg, zero_reg);
+    // add each quad word of madd (dqcoeff - coeff) and madd (coeff)
+    sse_reg = _mm256_add_epi64(sse_reg, exp_dqcoeff_lo);
+    ssz_reg = _mm256_add_epi64(ssz_reg, exp_coeff_lo);
+    sse_reg = _mm256_add_epi64(sse_reg, exp_dqcoeff_hi);
+    ssz_reg = _mm256_add_epi64(ssz_reg, exp_coeff_hi);
+  }
+  // save the higher 64 bit of each 128 bit lane
+  sse_reg_64hi = _mm256_srli_si256(sse_reg, 8);
+  ssz_reg_64hi = _mm256_srli_si256(ssz_reg, 8);
+  // add the higher 64 bit to the low 64 bit
+  sse_reg = _mm256_add_epi64(sse_reg, sse_reg_64hi);
+  ssz_reg = _mm256_add_epi64(ssz_reg, ssz_reg_64hi);
+
+  // add each 64 bit from each of the 128 bit lane of the 256 bit
+  sse_reg128 = _mm_add_epi64(_mm256_castsi256_si128(sse_reg),
+                             _mm256_extractf128_si256(sse_reg, 1));
+
+  ssz_reg128 = _mm_add_epi64(_mm256_castsi256_si128(ssz_reg),
+                             _mm256_extractf128_si256(ssz_reg, 1));
+
+  // store the results
+  _mm_storel_epi64((__m128i*)(&sse), sse_reg128);
+
+  _mm_storel_epi64((__m128i*)(ssz), ssz_reg128);
+  return sse;
+}
diff --git a/media/libvpx/vp9/encoder/x86/vp9_error_sse2.asm b/media/libvpx/vp9/encoder/x86/vp9_error_sse2.asm
new file mode 100644
index 000000000..56373e897
--- /dev/null
+++ b/media/libvpx/vp9/encoder/x86/vp9_error_sse2.asm
@@ -0,0 +1,120 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION .text
+
+; int64_t vp9_block_error(int16_t *coeff, int16_t *dqcoeff, intptr_t block_size,
+;                         int64_t *ssz)
+
+INIT_XMM sse2
+cglobal block_error, 3, 3, 8, uqc, dqc, size, ssz
+  pxor      m4, m4                 ; sse accumulator
+  pxor      m6, m6                 ; ssz accumulator
+  pxor      m5, m5                 ; dedicated zero register
+  lea     uqcq, [uqcq+sizeq*2]
+  lea     dqcq, [dqcq+sizeq*2]
+  neg    sizeq
+.loop:
+  mova      m2, [uqcq+sizeq*2]
+  mova      m0, [dqcq+sizeq*2]
+  mova      m3, [uqcq+sizeq*2+mmsize]
+  mova      m1, [dqcq+sizeq*2+mmsize]
+  psubw     m0, m2
+  psubw     m1, m3
+  ; individual errors are max. 15bit+sign, so squares are 30bit, and
+  ; thus the sum of 2 should fit in a 31bit integer (+ unused sign bit)
+  pmaddwd   m0, m0
+  pmaddwd   m1, m1
+  pmaddwd   m2, m2
+  pmaddwd   m3, m3
+  ; accumulate in 64bit
+  punpckldq m7, m0, m5
+  punpckhdq m0, m5
+  paddq     m4, m7
+  punpckldq m7, m1, m5
+  paddq     m4, m0
+  punpckhdq m1, m5
+  paddq     m4, m7
+  punpckldq m7, m2, m5
+  paddq     m4, m1
+  punpckhdq m2, m5
+  paddq     m6, m7
+  punpckldq m7, m3, m5
+  paddq     m6, m2
+  punpckhdq m3, m5
+  paddq     m6, m7
+  paddq     m6, m3
+  add    sizeq, mmsize
+  jl .loop
+
+  ; accumulate horizontally and store in return value
+  movhlps   m5, m4
+  movhlps   m7, m6
+  paddq     m4, m5
+  paddq     m6, m7
+%if ARCH_X86_64
+  movq    rax, m4
+  movq [sszq], m6
+%else
+  mov     eax, sszm
+  pshufd   m5, m4, 0x1
+  movq  [eax], m6
+  movd    eax, m4
+  movd    edx, m5
+%endif
+  RET
+
+; Compute the sum of squared difference between two int16_t vectors.
+; int64_t vp9_block_error_fp(int16_t *coeff, int16_t *dqcoeff,
+;                            intptr_t block_size)
+
+INIT_XMM sse2
+cglobal block_error_fp, 3, 3, 6, uqc, dqc, size
+  pxor      m4, m4                 ; sse accumulator
+  pxor      m5, m5                 ; dedicated zero register
+  lea     uqcq, [uqcq+sizeq*2]
+  lea     dqcq, [dqcq+sizeq*2]
+  neg    sizeq
+.loop:
+  mova      m2, [uqcq+sizeq*2]
+  mova      m0, [dqcq+sizeq*2]
+  mova      m3, [uqcq+sizeq*2+mmsize]
+  mova      m1, [dqcq+sizeq*2+mmsize]
+  psubw     m0, m2
+  psubw     m1, m3
+  ; individual errors are max. 15bit+sign, so squares are 30bit, and
+  ; thus the sum of 2 should fit in a 31bit integer (+ unused sign bit)
+  pmaddwd   m0, m0
+  pmaddwd   m1, m1
+  ; accumulate in 64bit
+  punpckldq m3, m0, m5
+  punpckhdq m0, m5
+  paddq     m4, m3
+  punpckldq m3, m1, m5
+  paddq     m4, m0
+  punpckhdq m1, m5
+  paddq     m4, m3
+  paddq     m4, m1
+  add    sizeq, mmsize
+  jl .loop
+
+  ; accumulate horizontally and store in return value
+  movhlps   m5, m4
+  paddq     m4, m5
+%if ARCH_X86_64
+  movq    rax, m4
+%else
+  pshufd   m5, m4, 0x1
+  movd    eax, m4
+  movd    edx, m5
+%endif
+  RET
diff --git a/media/libvpx/vp9/encoder/x86/vp9_highbd_block_error_intrin_sse2.c b/media/libvpx/vp9/encoder/x86/vp9_highbd_block_error_intrin_sse2.c
new file mode 100644
index 000000000..c245ccafa
--- /dev/null
+++ b/media/libvpx/vp9/encoder/x86/vp9_highbd_block_error_intrin_sse2.c
@@ -0,0 +1,71 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <emmintrin.h>
+#include <stdio.h>
+
+#include "vp9/common/vp9_common.h"
+
+int64_t vp9_highbd_block_error_sse2(tran_low_t *coeff, tran_low_t *dqcoeff,
+                                    intptr_t block_size, int64_t *ssz,
+                                    int bps) {
+  int i, j, test;
+  uint32_t temp[4];
+  __m128i max, min, cmp0, cmp1, cmp2, cmp3;
+  int64_t error = 0, sqcoeff = 0;
+  const int shift = 2 * (bps - 8);
+  const int rounding = shift > 0 ? 1 << (shift - 1) : 0;
+
+  for (i = 0; i < block_size; i+=8) {
+    // Load the data into xmm registers
+    __m128i mm_coeff = _mm_load_si128((__m128i*) (coeff + i));
+    __m128i mm_coeff2 = _mm_load_si128((__m128i*) (coeff + i + 4));
+    __m128i mm_dqcoeff = _mm_load_si128((__m128i*) (dqcoeff + i));
+    __m128i mm_dqcoeff2 = _mm_load_si128((__m128i*) (dqcoeff + i + 4));
+    // Check if any values require more than 15 bit
+    max = _mm_set1_epi32(0x3fff);
+    min = _mm_set1_epi32(0xffffc000);
+    cmp0 = _mm_xor_si128(_mm_cmpgt_epi32(mm_coeff, max),
+            _mm_cmplt_epi32(mm_coeff, min));
+    cmp1 = _mm_xor_si128(_mm_cmpgt_epi32(mm_coeff2, max),
+            _mm_cmplt_epi32(mm_coeff2, min));
+    cmp2 = _mm_xor_si128(_mm_cmpgt_epi32(mm_dqcoeff, max),
+            _mm_cmplt_epi32(mm_dqcoeff, min));
+    cmp3 = _mm_xor_si128(_mm_cmpgt_epi32(mm_dqcoeff2, max),
+            _mm_cmplt_epi32(mm_dqcoeff2, min));
+    test = _mm_movemask_epi8(_mm_or_si128(_mm_or_si128(cmp0, cmp1),
+            _mm_or_si128(cmp2, cmp3)));
+
+    if (!test) {
+      __m128i mm_diff, error_sse2, sqcoeff_sse2;;
+      mm_coeff = _mm_packs_epi32(mm_coeff, mm_coeff2);
+      mm_dqcoeff = _mm_packs_epi32(mm_dqcoeff, mm_dqcoeff2);
+      mm_diff = _mm_sub_epi16(mm_coeff, mm_dqcoeff);
+      error_sse2 = _mm_madd_epi16(mm_diff, mm_diff);
+      sqcoeff_sse2 = _mm_madd_epi16(mm_coeff, mm_coeff);
+      _mm_storeu_si128((__m128i*)temp, error_sse2);
+      error = error + temp[0] + temp[1] + temp[2] + temp[3];
+      _mm_storeu_si128((__m128i*)temp, sqcoeff_sse2);
+      sqcoeff += temp[0] + temp[1] + temp[2] + temp[3];
+    } else {
+      for (j = 0; j < 8; j++) {
+        const int64_t diff = coeff[i + j] - dqcoeff[i + j];
+        error +=  diff * diff;
+        sqcoeff += (int64_t)coeff[i + j] * (int64_t)coeff[i + j];
+      }
+    }
+  }
+  assert(error >= 0 && sqcoeff >= 0);
+  error = (error + rounding) >> shift;
+  sqcoeff = (sqcoeff + rounding) >> shift;
+
+  *ssz = sqcoeff;
+  return error;
+}
diff --git a/media/libvpx/vp9/encoder/x86/vp9_highbd_quantize_intrin_sse2.c b/media/libvpx/vp9/encoder/x86/vp9_highbd_quantize_intrin_sse2.c
new file mode 100644
index 000000000..cbdd1c93e
--- /dev/null
+++ b/media/libvpx/vp9/encoder/x86/vp9_highbd_quantize_intrin_sse2.c
@@ -0,0 +1,181 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <emmintrin.h>
+
+#include "vpx_ports/mem.h"
+#include "vp9/common/vp9_common.h"
+
+#if CONFIG_VP9_HIGHBITDEPTH
+// from vp9_idct.h: typedef int32_t tran_low_t;
+void vp9_highbd_quantize_b_sse2(const tran_low_t *coeff_ptr,
+                                intptr_t count,
+                                int skip_block,
+                                const int16_t *zbin_ptr,
+                                const int16_t *round_ptr,
+                                const int16_t *quant_ptr,
+                                const int16_t *quant_shift_ptr,
+                                tran_low_t *qcoeff_ptr,
+                                tran_low_t *dqcoeff_ptr,
+                                const int16_t *dequant_ptr,
+                                uint16_t *eob_ptr,
+                                const int16_t *scan,
+                                const int16_t *iscan) {
+  int i, j, non_zero_regs = (int)count / 4, eob_i = -1;
+  __m128i zbins[2];
+  __m128i nzbins[2];
+
+  zbins[0] = _mm_set_epi32((int)zbin_ptr[1],
+                           (int)zbin_ptr[1],
+                           (int)zbin_ptr[1],
+                           (int)zbin_ptr[0]);
+  zbins[1] = _mm_set1_epi32((int)zbin_ptr[1]);
+
+  nzbins[0] = _mm_setzero_si128();
+  nzbins[1] = _mm_setzero_si128();
+  nzbins[0] = _mm_sub_epi32(nzbins[0], zbins[0]);
+  nzbins[1] = _mm_sub_epi32(nzbins[1], zbins[1]);
+
+  (void)scan;
+
+  memset(qcoeff_ptr, 0, count * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, count * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    // Pre-scan pass
+    for (i = ((int)count / 4) - 1; i >= 0; i--) {
+      __m128i coeffs, cmp1, cmp2;
+      int test;
+      coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4));
+      cmp1 = _mm_cmplt_epi32(coeffs, zbins[i != 0]);
+      cmp2 = _mm_cmpgt_epi32(coeffs, nzbins[i != 0]);
+      cmp1 = _mm_and_si128(cmp1, cmp2);
+      test = _mm_movemask_epi8(cmp1);
+      if (test == 0xffff)
+        non_zero_regs--;
+      else
+        break;
+    }
+
+    // Quantization pass:
+    for (i = 0; i < non_zero_regs; i++) {
+      __m128i coeffs, coeffs_sign, tmp1, tmp2;
+      int test;
+      int abs_coeff[4];
+      int coeff_sign[4];
+
+      coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4));
+      coeffs_sign = _mm_srai_epi32(coeffs, 31);
+      coeffs = _mm_sub_epi32(
+            _mm_xor_si128(coeffs, coeffs_sign), coeffs_sign);
+      tmp1 = _mm_cmpgt_epi32(coeffs, zbins[i != 0]);
+      tmp2 = _mm_cmpeq_epi32(coeffs, zbins[i != 0]);
+      tmp1 = _mm_or_si128(tmp1, tmp2);
+      test = _mm_movemask_epi8(tmp1);
+      _mm_storeu_si128((__m128i*)abs_coeff, coeffs);
+      _mm_storeu_si128((__m128i*)coeff_sign, coeffs_sign);
+
+      for (j = 0; j < 4; j++) {
+        if (test & (1 << (4 * j))) {
+          int k = 4 * i + j;
+          int64_t tmp = clamp(abs_coeff[j] + round_ptr[k != 0],
+                              INT32_MIN, INT32_MAX);
+          tmp = ((((tmp * quant_ptr[k != 0]) >> 16) + tmp) *
+                    quant_shift_ptr[k != 0]) >> 16;  // quantization
+          qcoeff_ptr[k] = (tmp ^ coeff_sign[j]) - coeff_sign[j];
+          dqcoeff_ptr[k] = qcoeff_ptr[k] * dequant_ptr[k != 0];
+          if (tmp)
+            eob_i = iscan[k] > eob_i ? iscan[k] : eob_i;
+        }
+      }
+    }
+  }
+  *eob_ptr = eob_i + 1;
+}
+
+
+void vp9_highbd_quantize_b_32x32_sse2(const tran_low_t *coeff_ptr,
+                                      intptr_t n_coeffs,
+                                      int skip_block,
+                                      const int16_t *zbin_ptr,
+                                      const int16_t *round_ptr,
+                                      const int16_t *quant_ptr,
+                                      const int16_t *quant_shift_ptr,
+                                      tran_low_t *qcoeff_ptr,
+                                      tran_low_t *dqcoeff_ptr,
+                                      const int16_t *dequant_ptr,
+                                      uint16_t *eob_ptr,
+                                      const int16_t *scan,
+                                      const int16_t *iscan) {
+  __m128i zbins[2];
+  __m128i nzbins[2];
+  int idx = 0;
+  int idx_arr[1024];
+  int i, eob = -1;
+  const int zbin0_tmp = ROUND_POWER_OF_TWO(zbin_ptr[0], 1);
+  const int zbin1_tmp = ROUND_POWER_OF_TWO(zbin_ptr[1], 1);
+  (void)scan;
+  zbins[0] = _mm_set_epi32(zbin1_tmp,
+                           zbin1_tmp,
+                           zbin1_tmp,
+                           zbin0_tmp);
+  zbins[1] = _mm_set1_epi32(zbin1_tmp);
+
+  nzbins[0] = _mm_setzero_si128();
+  nzbins[1] = _mm_setzero_si128();
+  nzbins[0] = _mm_sub_epi32(nzbins[0], zbins[0]);
+  nzbins[1] = _mm_sub_epi32(nzbins[1], zbins[1]);
+
+  memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
+  memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
+
+  if (!skip_block) {
+    // Pre-scan pass
+    for (i = 0; i < n_coeffs / 4; i++) {
+      __m128i coeffs, cmp1, cmp2;
+      int test;
+      coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4));
+      cmp1 = _mm_cmplt_epi32(coeffs, zbins[i != 0]);
+      cmp2 = _mm_cmpgt_epi32(coeffs, nzbins[i != 0]);
+      cmp1 = _mm_and_si128(cmp1, cmp2);
+      test = _mm_movemask_epi8(cmp1);
+      if (!(test & 0xf))
+        idx_arr[idx++] = i * 4;
+      if (!(test & 0xf0))
+        idx_arr[idx++] = i * 4 + 1;
+      if (!(test & 0xf00))
+        idx_arr[idx++] = i * 4 + 2;
+      if (!(test & 0xf000))
+        idx_arr[idx++] = i * 4 + 3;
+    }
+
+    // Quantization pass: only process the coefficients selected in
+    // pre-scan pass. Note: idx can be zero.
+    for (i = 0; i < idx; i++) {
+      const int rc = idx_arr[i];
+      const int coeff = coeff_ptr[rc];
+      const int coeff_sign = (coeff >> 31);
+      int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
+      int64_t tmp = clamp(abs_coeff +
+                          ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1),
+                          INT32_MIN, INT32_MAX);
+      tmp = ((((tmp * quant_ptr[rc != 0]) >> 16) + tmp) *
+               quant_shift_ptr[rc != 0]) >> 15;
+
+      qcoeff_ptr[rc] = (tmp ^ coeff_sign) - coeff_sign;
+      dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0] / 2;
+
+      if (tmp)
+        eob = iscan[idx_arr[i]] > eob ? iscan[idx_arr[i]] : eob;
+    }
+  }
+  *eob_ptr = eob + 1;
+}
+#endif
diff --git a/media/libvpx/vp9/encoder/x86/vp9_highbd_subpel_variance.asm b/media/libvpx/vp9/encoder/x86/vp9_highbd_subpel_variance.asm
new file mode 100644
index 000000000..4594bb1aa
--- /dev/null
+++ b/media/libvpx/vp9/encoder/x86/vp9_highbd_subpel_variance.asm
@@ -0,0 +1,1039 @@
+;
+;  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION_RODATA
+pw_8: times  8 dw  8
+bilin_filter_m_sse2: times  8 dw 16
+                     times  8 dw  0
+                     times  8 dw 14
+                     times  8 dw  2
+                     times  8 dw 12
+                     times  8 dw  4
+                     times  8 dw 10
+                     times  8 dw  6
+                     times 16 dw  8
+                     times  8 dw  6
+                     times  8 dw 10
+                     times  8 dw  4
+                     times  8 dw 12
+                     times  8 dw  2
+                     times  8 dw 14
+
+SECTION .text
+
+; int vp9_sub_pixel_varianceNxh(const uint8_t *src, ptrdiff_t src_stride,
+;                               int x_offset, int y_offset,
+;                               const uint8_t *dst, ptrdiff_t dst_stride,
+;                               int height, unsigned int *sse);
+;
+; This function returns the SE and stores SSE in the given pointer.
+
+%macro SUM_SSE 6 ; src1, dst1, src2, dst2, sum, sse
+  psubw                %3, %4
+  psubw                %1, %2
+  mova                 %4, %3       ; make copies to manipulate to calc sum
+  mova                 %2, %1       ; use originals for calc sse
+  pmaddwd              %3, %3
+  paddw                %4, %2
+  pmaddwd              %1, %1
+  movhlps              %2, %4
+  paddd                %6, %3
+  paddw                %4, %2
+  pxor                 %2, %2
+  pcmpgtw              %2, %4       ; mask for 0 > %4 (sum)
+  punpcklwd            %4, %2       ; sign-extend word to dword
+  paddd                %6, %1
+  paddd                %5, %4
+
+%endmacro
+
+%macro STORE_AND_RET 0
+%if mmsize == 16
+  ; if H=64 and W=16, we have 8 words of each 2(1bit)x64(6bit)x9bit=16bit
+  ; in m6, i.e. it _exactly_ fits in a signed word per word in the xmm reg.
+  ; We have to sign-extend it before adding the words within the register
+  ; and outputing to a dword.
+  movhlps              m3, m7
+  movhlps              m4, m6
+  paddd                m7, m3
+  paddd                m6, m4
+  pshufd               m3, m7, 0x1
+  pshufd               m4, m6, 0x1
+  paddd                m7, m3
+  paddd                m6, m4
+  mov                  r1, ssem         ; r1 = unsigned int *sse
+  movd               [r1], m7           ; store sse
+  movd                rax, m6           ; store sum as return value
+%endif
+  RET
+%endmacro
+
+%macro INC_SRC_BY_SRC_STRIDE  0
+%if ARCH_X86=1 && CONFIG_PIC=1
+  lea                srcq, [srcq + src_stridemp*2]
+%else
+  lea                srcq, [srcq + src_strideq*2]
+%endif
+%endmacro
+
+%macro INC_SRC_BY_SRC_2STRIDE  0
+%if ARCH_X86=1 && CONFIG_PIC=1
+  lea                srcq, [srcq + src_stridemp*4]
+%else
+  lea                srcq, [srcq + src_strideq*4]
+%endif
+%endmacro
+
+%macro SUBPEL_VARIANCE 1-2 0 ; W
+%define bilin_filter_m bilin_filter_m_sse2
+%define filter_idx_shift 5
+
+
+%ifdef PIC    ; 64bit PIC
+  %if %2 == 1 ; avg
+    cglobal highbd_sub_pixel_avg_variance%1xh, 9, 10, 13, src, src_stride, \
+                                      x_offset, y_offset, \
+                                      dst, dst_stride, \
+                                      sec, sec_stride, height, sse
+    %define sec_str sec_strideq
+  %else
+    cglobal highbd_sub_pixel_variance%1xh, 7, 8, 13, src, src_stride, x_offset, \
+                                  y_offset, dst, dst_stride, height, sse
+  %endif
+  %define h heightd
+  %define bilin_filter sseq
+%else
+  %if ARCH_X86=1 && CONFIG_PIC=1
+    %if %2 == 1 ; avg
+      cglobal highbd_sub_pixel_avg_variance%1xh, 7, 7, 13, src, src_stride, \
+                                  x_offset, y_offset, \
+                                  dst, dst_stride, \
+                                  sec, sec_stride, \
+                                  height, sse, g_bilin_filter, g_pw_8
+      %define h dword heightm
+      %define sec_str sec_stridemp
+
+      ; Store bilin_filter and pw_8 location in stack
+      GET_GOT eax
+      add esp, 4                ; restore esp
+
+      lea ecx, [GLOBAL(bilin_filter_m)]
+      mov g_bilin_filterm, ecx
+
+      lea ecx, [GLOBAL(pw_8)]
+      mov g_pw_8m, ecx
+
+      LOAD_IF_USED 0, 1         ; load eax, ecx back
+    %else
+      cglobal highbd_sub_pixel_variance%1xh, 7, 7, 13, src, src_stride, \
+                                x_offset, y_offset, dst, dst_stride, height, \
+                                sse, g_bilin_filter, g_pw_8
+      %define h heightd
+
+      ; Store bilin_filter and pw_8 location in stack
+      GET_GOT eax
+      add esp, 4                ; restore esp
+
+      lea ecx, [GLOBAL(bilin_filter_m)]
+      mov g_bilin_filterm, ecx
+
+      lea ecx, [GLOBAL(pw_8)]
+      mov g_pw_8m, ecx
+
+      LOAD_IF_USED 0, 1         ; load eax, ecx back
+    %endif
+  %else
+    %if %2 == 1 ; avg
+      cglobal highbd_sub_pixel_avg_variance%1xh, 7 + 2 * ARCH_X86_64, \
+                        7 + 2 * ARCH_X86_64, 13, src, src_stride, \
+                                             x_offset, y_offset, \
+                                             dst, dst_stride, \
+                                             sec, sec_stride, \
+                                             height, sse
+      %if ARCH_X86_64
+      %define h heightd
+      %define sec_str sec_strideq
+      %else
+      %define h dword heightm
+      %define sec_str sec_stridemp
+      %endif
+    %else
+      cglobal highbd_sub_pixel_variance%1xh, 7, 7, 13, src, src_stride, \
+                              x_offset, y_offset, dst, dst_stride, height, sse
+      %define h heightd
+    %endif
+
+    %define bilin_filter bilin_filter_m
+  %endif
+%endif
+
+  ASSERT               %1 <= 16         ; m6 overflows if w > 16
+  pxor                 m6, m6           ; sum
+  pxor                 m7, m7           ; sse
+
+%if %1 < 16
+  sar                   h, 1
+%endif
+%if %2 == 1 ; avg
+  shl             sec_str, 1
+%endif
+
+  ; FIXME(rbultje) replace by jumptable?
+  test          x_offsetd, x_offsetd
+  jnz .x_nonzero
+  ; x_offset == 0
+  test          y_offsetd, y_offsetd
+  jnz .x_zero_y_nonzero
+
+  ; x_offset == 0 && y_offset == 0
+.x_zero_y_zero_loop:
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m2, [srcq + 16]
+  mova                 m1, [dstq]
+  mova                 m3, [dstq + 16]
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  pavgw                m2, [secq+16]
+%endif
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  lea                srcq, [srcq + src_strideq*2]
+  lea                dstq, [dstq + dst_strideq*2]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%else ; %1 < 16
+  movu                 m0, [srcq]
+  movu                 m2, [srcq + src_strideq*2]
+  mova                 m1, [dstq]
+  mova                 m3, [dstq + dst_strideq*2]
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  add                secq, sec_str
+  pavgw                m2, [secq]
+%endif
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  lea                srcq, [srcq + src_strideq*4]
+  lea                dstq, [dstq + dst_strideq*4]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%endif
+  dec                   h
+  jg .x_zero_y_zero_loop
+  STORE_AND_RET
+
+.x_zero_y_nonzero:
+  cmp           y_offsetd, 8
+  jne .x_zero_y_nonhalf
+
+  ; x_offset == 0 && y_offset == 0.5
+.x_zero_y_half_loop:
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq+16]
+  movu                 m4, [srcq+src_strideq*2]
+  movu                 m5, [srcq+src_strideq*2+16]
+  mova                 m2, [dstq]
+  mova                 m3, [dstq+16]
+  pavgw                m0, m4
+  pavgw                m1, m5
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  pavgw                m1, [secq+16]
+%endif
+  SUM_SSE              m0, m2, m1, m3, m6, m7
+
+  lea                srcq, [srcq + src_strideq*2]
+  lea                dstq, [dstq + dst_strideq*2]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%else ; %1 < 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq+src_strideq*2]
+  movu                 m5, [srcq+src_strideq*4]
+  mova                 m2, [dstq]
+  mova                 m3, [dstq+dst_strideq*2]
+  pavgw                m0, m1
+  pavgw                m1, m5
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  add                secq, sec_str
+  pavgw                m1, [secq]
+%endif
+  SUM_SSE              m0, m2, m1, m3, m6, m7
+
+  lea                srcq, [srcq + src_strideq*4]
+  lea                dstq, [dstq + dst_strideq*4]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%endif
+  dec                   h
+  jg .x_zero_y_half_loop
+  STORE_AND_RET
+
+.x_zero_y_nonhalf:
+  ; x_offset == 0 && y_offset == bilin interpolation
+%ifdef PIC
+  lea        bilin_filter, [bilin_filter_m]
+%endif
+  shl           y_offsetd, filter_idx_shift
+%if ARCH_X86_64 && mmsize == 16
+  mova                 m8, [bilin_filter+y_offsetq]
+  mova                 m9, [bilin_filter+y_offsetq+16]
+  mova                m10, [pw_8]
+%define filter_y_a m8
+%define filter_y_b m9
+%define filter_rnd m10
+%else ; x86-32 or mmx
+%if ARCH_X86=1 && CONFIG_PIC=1
+; x_offset == 0, reuse x_offset reg
+%define tempq x_offsetq
+  add y_offsetq, g_bilin_filterm
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+  mov tempq, g_pw_8m
+%define filter_rnd [tempq]
+%else
+  add           y_offsetq, bilin_filter
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+%define filter_rnd [pw_8]
+%endif
+%endif
+
+.x_zero_y_other_loop:
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq + 16]
+  movu                 m4, [srcq+src_strideq*2]
+  movu                 m5, [srcq+src_strideq*2+16]
+  mova                 m2, [dstq]
+  mova                 m3, [dstq+16]
+  ; FIXME(rbultje) instead of out=((num-x)*in1+x*in2+rnd)>>log2(num), we can
+  ; also do out=in1+(((num-x)*(in2-in1)+rnd)>>log2(num)). Total number of
+  ; instructions is the same (5), but it is 1 mul instead of 2, so might be
+  ; slightly faster because of pmullw latency. It would also cut our rodata
+  ; tables in half for this function, and save 1-2 registers on x86-64.
+  pmullw               m1, filter_y_a
+  pmullw               m5, filter_y_b
+  paddw                m1, filter_rnd
+  pmullw               m0, filter_y_a
+  pmullw               m4, filter_y_b
+  paddw                m0, filter_rnd
+  paddw                m1, m5
+  paddw                m0, m4
+  psrlw                m1, 4
+  psrlw                m0, 4
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  pavgw                m1, [secq+16]
+%endif
+  SUM_SSE              m0, m2, m1, m3, m6, m7
+
+  lea                srcq, [srcq + src_strideq*2]
+  lea                dstq, [dstq + dst_strideq*2]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%else ; %1 < 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq+src_strideq*2]
+  movu                 m5, [srcq+src_strideq*4]
+  mova                 m4, m1
+  mova                 m2, [dstq]
+  mova                 m3, [dstq+dst_strideq*2]
+  pmullw               m1, filter_y_a
+  pmullw               m5, filter_y_b
+  paddw                m1, filter_rnd
+  pmullw               m0, filter_y_a
+  pmullw               m4, filter_y_b
+  paddw                m0, filter_rnd
+  paddw                m1, m5
+  paddw                m0, m4
+  psrlw                m1, 4
+  psrlw                m0, 4
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  add                secq, sec_str
+  pavgw                m1, [secq]
+%endif
+  SUM_SSE              m0, m2, m1, m3, m6, m7
+
+  lea                srcq, [srcq + src_strideq*4]
+  lea                dstq, [dstq + dst_strideq*4]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%endif
+  dec                   h
+  jg .x_zero_y_other_loop
+%undef filter_y_a
+%undef filter_y_b
+%undef filter_rnd
+  STORE_AND_RET
+
+.x_nonzero:
+  cmp           x_offsetd, 8
+  jne .x_nonhalf
+  ; x_offset == 0.5
+  test          y_offsetd, y_offsetd
+  jnz .x_half_y_nonzero
+
+  ; x_offset == 0.5 && y_offset == 0
+.x_half_y_zero_loop:
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq + 16]
+  movu                 m4, [srcq + 2]
+  movu                 m5, [srcq + 18]
+  mova                 m2, [dstq]
+  mova                 m3, [dstq + 16]
+  pavgw                m0, m4
+  pavgw                m1, m5
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  pavgw                m1, [secq+16]
+%endif
+  SUM_SSE              m0, m2, m1, m3, m6, m7
+
+  lea                srcq, [srcq + src_strideq*2]
+  lea                dstq, [dstq + dst_strideq*2]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%else ; %1 < 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq + src_strideq*2]
+  movu                 m4, [srcq + 2]
+  movu                 m5, [srcq + src_strideq*2 + 2]
+  mova                 m2, [dstq]
+  mova                 m3, [dstq + dst_strideq*2]
+  pavgw                m0, m4
+  pavgw                m1, m5
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  add                secq, sec_str
+  pavgw                m1, [secq]
+%endif
+  SUM_SSE              m0, m2, m1, m3, m6, m7
+
+  lea                srcq, [srcq + src_strideq*4]
+  lea                dstq, [dstq + dst_strideq*4]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%endif
+  dec                   h
+  jg .x_half_y_zero_loop
+  STORE_AND_RET
+
+.x_half_y_nonzero:
+  cmp           y_offsetd, 8
+  jne .x_half_y_nonhalf
+
+  ; x_offset == 0.5 && y_offset == 0.5
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq+16]
+  movu                 m2, [srcq+2]
+  movu                 m3, [srcq+18]
+  lea                srcq, [srcq + src_strideq*2]
+  pavgw                m0, m2
+  pavgw                m1, m3
+.x_half_y_half_loop:
+  movu                 m2, [srcq]
+  movu                 m3, [srcq + 16]
+  movu                 m4, [srcq + 2]
+  movu                 m5, [srcq + 18]
+  pavgw                m2, m4
+  pavgw                m3, m5
+  pavgw                m0, m2
+  pavgw                m1, m3
+  mova                 m4, [dstq]
+  mova                 m5, [dstq + 16]
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  pavgw                m1, [secq+16]
+%endif
+  SUM_SSE              m0, m4, m1, m5, m6, m7
+  mova                 m0, m2
+  mova                 m1, m3
+
+  lea                srcq, [srcq + src_strideq*2]
+  lea                dstq, [dstq + dst_strideq*2]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%else ; %1 < 16
+  movu                 m0, [srcq]
+  movu                 m2, [srcq+2]
+  lea                srcq, [srcq + src_strideq*2]
+  pavgw                m0, m2
+.x_half_y_half_loop:
+  movu                 m2, [srcq]
+  movu                 m3, [srcq + src_strideq*2]
+  movu                 m4, [srcq + 2]
+  movu                 m5, [srcq + src_strideq*2 + 2]
+  pavgw                m2, m4
+  pavgw                m3, m5
+  pavgw                m0, m2
+  pavgw                m2, m3
+  mova                 m4, [dstq]
+  mova                 m5, [dstq + dst_strideq*2]
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  add                secq, sec_str
+  pavgw                m2, [secq]
+%endif
+  SUM_SSE              m0, m4, m2, m5, m6, m7
+  mova                 m0, m3
+
+  lea                srcq, [srcq + src_strideq*4]
+  lea                dstq, [dstq + dst_strideq*4]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%endif
+  dec                   h
+  jg .x_half_y_half_loop
+  STORE_AND_RET
+
+.x_half_y_nonhalf:
+  ; x_offset == 0.5 && y_offset == bilin interpolation
+%ifdef PIC
+  lea        bilin_filter, [bilin_filter_m]
+%endif
+  shl           y_offsetd, filter_idx_shift
+%if ARCH_X86_64 && mmsize == 16
+  mova                 m8, [bilin_filter+y_offsetq]
+  mova                 m9, [bilin_filter+y_offsetq+16]
+  mova                m10, [pw_8]
+%define filter_y_a m8
+%define filter_y_b m9
+%define filter_rnd m10
+%else  ; x86_32
+%if ARCH_X86=1 && CONFIG_PIC=1
+; x_offset == 0.5. We can reuse x_offset reg
+%define tempq x_offsetq
+  add y_offsetq, g_bilin_filterm
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+  mov tempq, g_pw_8m
+%define filter_rnd [tempq]
+%else
+  add           y_offsetq, bilin_filter
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+%define filter_rnd [pw_8]
+%endif
+%endif
+
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq+16]
+  movu                 m2, [srcq+2]
+  movu                 m3, [srcq+18]
+  lea                srcq, [srcq + src_strideq*2]
+  pavgw                m0, m2
+  pavgw                m1, m3
+.x_half_y_other_loop:
+  movu                 m2, [srcq]
+  movu                 m3, [srcq+16]
+  movu                 m4, [srcq+2]
+  movu                 m5, [srcq+18]
+  pavgw                m2, m4
+  pavgw                m3, m5
+  mova                 m4, m2
+  mova                 m5, m3
+  pmullw               m1, filter_y_a
+  pmullw               m3, filter_y_b
+  paddw                m1, filter_rnd
+  paddw                m1, m3
+  pmullw               m0, filter_y_a
+  pmullw               m2, filter_y_b
+  paddw                m0, filter_rnd
+  psrlw                m1, 4
+  paddw                m0, m2
+  mova                 m2, [dstq]
+  psrlw                m0, 4
+  mova                 m3, [dstq+16]
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  pavgw                m1, [secq+16]
+%endif
+  SUM_SSE              m0, m2, m1, m3, m6, m7
+  mova                 m0, m4
+  mova                 m1, m5
+
+  lea                srcq, [srcq + src_strideq*2]
+  lea                dstq, [dstq + dst_strideq*2]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%else ; %1 < 16
+  movu                 m0, [srcq]
+  movu                 m2, [srcq+2]
+  lea                srcq, [srcq + src_strideq*2]
+  pavgw                m0, m2
+.x_half_y_other_loop:
+  movu                 m2, [srcq]
+  movu                 m3, [srcq+src_strideq*2]
+  movu                 m4, [srcq+2]
+  movu                 m5, [srcq+src_strideq*2+2]
+  pavgw                m2, m4
+  pavgw                m3, m5
+  mova                 m4, m2
+  mova                 m5, m3
+  pmullw               m4, filter_y_a
+  pmullw               m3, filter_y_b
+  paddw                m4, filter_rnd
+  paddw                m4, m3
+  pmullw               m0, filter_y_a
+  pmullw               m2, filter_y_b
+  paddw                m0, filter_rnd
+  psrlw                m4, 4
+  paddw                m0, m2
+  mova                 m2, [dstq]
+  psrlw                m0, 4
+  mova                 m3, [dstq+dst_strideq*2]
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  add                secq, sec_str
+  pavgw                m4, [secq]
+%endif
+  SUM_SSE              m0, m2, m4, m3, m6, m7
+  mova                 m0, m5
+
+  lea                srcq, [srcq + src_strideq*4]
+  lea                dstq, [dstq + dst_strideq*4]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%endif
+  dec                   h
+  jg .x_half_y_other_loop
+%undef filter_y_a
+%undef filter_y_b
+%undef filter_rnd
+  STORE_AND_RET
+
+.x_nonhalf:
+  test          y_offsetd, y_offsetd
+  jnz .x_nonhalf_y_nonzero
+
+  ; x_offset == bilin interpolation && y_offset == 0
+%ifdef PIC
+  lea        bilin_filter, [bilin_filter_m]
+%endif
+  shl           x_offsetd, filter_idx_shift
+%if ARCH_X86_64 && mmsize == 16
+  mova                 m8, [bilin_filter+x_offsetq]
+  mova                 m9, [bilin_filter+x_offsetq+16]
+  mova                m10, [pw_8]
+%define filter_x_a m8
+%define filter_x_b m9
+%define filter_rnd m10
+%else    ; x86-32
+%if ARCH_X86=1 && CONFIG_PIC=1
+; y_offset == 0. We can reuse y_offset reg.
+%define tempq y_offsetq
+  add x_offsetq, g_bilin_filterm
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+  mov tempq, g_pw_8m
+%define filter_rnd [tempq]
+%else
+  add           x_offsetq, bilin_filter
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+%define filter_rnd [pw_8]
+%endif
+%endif
+
+.x_other_y_zero_loop:
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq+16]
+  movu                 m2, [srcq+2]
+  movu                 m3, [srcq+18]
+  mova                 m4, [dstq]
+  mova                 m5, [dstq+16]
+  pmullw               m1, filter_x_a
+  pmullw               m3, filter_x_b
+  paddw                m1, filter_rnd
+  pmullw               m0, filter_x_a
+  pmullw               m2, filter_x_b
+  paddw                m0, filter_rnd
+  paddw                m1, m3
+  paddw                m0, m2
+  psrlw                m1, 4
+  psrlw                m0, 4
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  pavgw                m1, [secq+16]
+%endif
+  SUM_SSE              m0, m4, m1, m5, m6, m7
+
+  lea                srcq, [srcq+src_strideq*2]
+  lea                dstq, [dstq+dst_strideq*2]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%else ; %1 < 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq+src_strideq*2]
+  movu                 m2, [srcq+2]
+  movu                 m3, [srcq+src_strideq*2+2]
+  mova                 m4, [dstq]
+  mova                 m5, [dstq+dst_strideq*2]
+  pmullw               m1, filter_x_a
+  pmullw               m3, filter_x_b
+  paddw                m1, filter_rnd
+  pmullw               m0, filter_x_a
+  pmullw               m2, filter_x_b
+  paddw                m0, filter_rnd
+  paddw                m1, m3
+  paddw                m0, m2
+  psrlw                m1, 4
+  psrlw                m0, 4
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  add                secq, sec_str
+  pavgw                m1, [secq]
+%endif
+  SUM_SSE              m0, m4, m1, m5, m6, m7
+
+  lea                srcq, [srcq+src_strideq*4]
+  lea                dstq, [dstq+dst_strideq*4]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%endif
+  dec                   h
+  jg .x_other_y_zero_loop
+%undef filter_x_a
+%undef filter_x_b
+%undef filter_rnd
+  STORE_AND_RET
+
+.x_nonhalf_y_nonzero:
+  cmp           y_offsetd, 8
+  jne .x_nonhalf_y_nonhalf
+
+  ; x_offset == bilin interpolation && y_offset == 0.5
+%ifdef PIC
+  lea        bilin_filter, [bilin_filter_m]
+%endif
+  shl           x_offsetd, filter_idx_shift
+%if ARCH_X86_64 && mmsize == 16
+  mova                 m8, [bilin_filter+x_offsetq]
+  mova                 m9, [bilin_filter+x_offsetq+16]
+  mova                m10, [pw_8]
+%define filter_x_a m8
+%define filter_x_b m9
+%define filter_rnd m10
+%else    ; x86-32
+%if ARCH_X86=1 && CONFIG_PIC=1
+; y_offset == 0.5. We can reuse y_offset reg.
+%define tempq y_offsetq
+  add x_offsetq, g_bilin_filterm
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+  mov tempq, g_pw_8m
+%define filter_rnd [tempq]
+%else
+  add           x_offsetq, bilin_filter
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+%define filter_rnd [pw_8]
+%endif
+%endif
+
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq+16]
+  movu                 m2, [srcq+2]
+  movu                 m3, [srcq+18]
+  pmullw               m0, filter_x_a
+  pmullw               m2, filter_x_b
+  paddw                m0, filter_rnd
+  pmullw               m1, filter_x_a
+  pmullw               m3, filter_x_b
+  paddw                m1, filter_rnd
+  paddw                m0, m2
+  paddw                m1, m3
+  psrlw                m0, 4
+  psrlw                m1, 4
+  lea                srcq, [srcq+src_strideq*2]
+.x_other_y_half_loop:
+  movu                 m2, [srcq]
+  movu                 m3, [srcq+16]
+  movu                 m4, [srcq+2]
+  movu                 m5, [srcq+18]
+  pmullw               m2, filter_x_a
+  pmullw               m4, filter_x_b
+  paddw                m2, filter_rnd
+  pmullw               m3, filter_x_a
+  pmullw               m5, filter_x_b
+  paddw                m3, filter_rnd
+  paddw                m2, m4
+  paddw                m3, m5
+  mova                 m4, [dstq]
+  mova                 m5, [dstq+16]
+  psrlw                m2, 4
+  psrlw                m3, 4
+  pavgw                m0, m2
+  pavgw                m1, m3
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  pavgw                m1, [secq+16]
+%endif
+  SUM_SSE              m0, m4, m1, m5, m6, m7
+  mova                 m0, m2
+  mova                 m1, m3
+
+  lea                srcq, [srcq+src_strideq*2]
+  lea                dstq, [dstq+dst_strideq*2]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%else ; %1 < 16
+  movu                 m0, [srcq]
+  movu                 m2, [srcq+2]
+  pmullw               m0, filter_x_a
+  pmullw               m2, filter_x_b
+  paddw                m0, filter_rnd
+  paddw                m0, m2
+  psrlw                m0, 4
+  lea                srcq, [srcq+src_strideq*2]
+.x_other_y_half_loop:
+  movu                 m2, [srcq]
+  movu                 m3, [srcq+src_strideq*2]
+  movu                 m4, [srcq+2]
+  movu                 m5, [srcq+src_strideq*2+2]
+  pmullw               m2, filter_x_a
+  pmullw               m4, filter_x_b
+  paddw                m2, filter_rnd
+  pmullw               m3, filter_x_a
+  pmullw               m5, filter_x_b
+  paddw                m3, filter_rnd
+  paddw                m2, m4
+  paddw                m3, m5
+  mova                 m4, [dstq]
+  mova                 m5, [dstq+dst_strideq*2]
+  psrlw                m2, 4
+  psrlw                m3, 4
+  pavgw                m0, m2
+  pavgw                m2, m3
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  add                secq, sec_str
+  pavgw                m2, [secq]
+%endif
+  SUM_SSE              m0, m4, m2, m5, m6, m7
+  mova                 m0, m3
+
+  lea                srcq, [srcq+src_strideq*4]
+  lea                dstq, [dstq+dst_strideq*4]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%endif
+  dec                   h
+  jg .x_other_y_half_loop
+%undef filter_x_a
+%undef filter_x_b
+%undef filter_rnd
+  STORE_AND_RET
+
+.x_nonhalf_y_nonhalf:
+; loading filter - this is same as in 8-bit depth
+%ifdef PIC
+  lea        bilin_filter, [bilin_filter_m]
+%endif
+  shl           x_offsetd, filter_idx_shift ; filter_idx_shift = 5
+  shl           y_offsetd, filter_idx_shift
+%if ARCH_X86_64 && mmsize == 16
+  mova                 m8, [bilin_filter+x_offsetq]
+  mova                 m9, [bilin_filter+x_offsetq+16]
+  mova                m10, [bilin_filter+y_offsetq]
+  mova                m11, [bilin_filter+y_offsetq+16]
+  mova                m12, [pw_8]
+%define filter_x_a m8
+%define filter_x_b m9
+%define filter_y_a m10
+%define filter_y_b m11
+%define filter_rnd m12
+%else   ; x86-32
+%if ARCH_X86=1 && CONFIG_PIC=1
+; In this case, there is NO unused register. Used src_stride register. Later,
+; src_stride has to be loaded from stack when it is needed.
+%define tempq src_strideq
+  mov tempq, g_bilin_filterm
+  add           x_offsetq, tempq
+  add           y_offsetq, tempq
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+
+  mov tempq, g_pw_8m
+%define filter_rnd [tempq]
+%else
+  add           x_offsetq, bilin_filter
+  add           y_offsetq, bilin_filter
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+%define filter_rnd [pw_8]
+%endif
+%endif
+; end of load filter
+
+  ; x_offset == bilin interpolation && y_offset == bilin interpolation
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m2, [srcq+2]
+  movu                 m1, [srcq+16]
+  movu                 m3, [srcq+18]
+  pmullw               m0, filter_x_a
+  pmullw               m2, filter_x_b
+  paddw                m0, filter_rnd
+  pmullw               m1, filter_x_a
+  pmullw               m3, filter_x_b
+  paddw                m1, filter_rnd
+  paddw                m0, m2
+  paddw                m1, m3
+  psrlw                m0, 4
+  psrlw                m1, 4
+
+  INC_SRC_BY_SRC_STRIDE
+
+.x_other_y_other_loop:
+  movu                 m2, [srcq]
+  movu                 m4, [srcq+2]
+  movu                 m3, [srcq+16]
+  movu                 m5, [srcq+18]
+  pmullw               m2, filter_x_a
+  pmullw               m4, filter_x_b
+  paddw                m2, filter_rnd
+  pmullw               m3, filter_x_a
+  pmullw               m5, filter_x_b
+  paddw                m3, filter_rnd
+  paddw                m2, m4
+  paddw                m3, m5
+  psrlw                m2, 4
+  psrlw                m3, 4
+  mova                 m4, m2
+  mova                 m5, m3
+  pmullw               m0, filter_y_a
+  pmullw               m2, filter_y_b
+  paddw                m0, filter_rnd
+  pmullw               m1, filter_y_a
+  pmullw               m3, filter_y_b
+  paddw                m0, m2
+  paddw                m1, filter_rnd
+  mova                 m2, [dstq]
+  paddw                m1, m3
+  psrlw                m0, 4
+  psrlw                m1, 4
+  mova                 m3, [dstq+16]
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  pavgw                m1, [secq+16]
+%endif
+  SUM_SSE              m0, m2, m1, m3, m6, m7
+  mova                 m0, m4
+  mova                 m1, m5
+
+  INC_SRC_BY_SRC_STRIDE
+  lea                dstq, [dstq + dst_strideq * 2]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%else ; %1 < 16
+  movu                 m0, [srcq]
+  movu                 m2, [srcq+2]
+  pmullw               m0, filter_x_a
+  pmullw               m2, filter_x_b
+  paddw                m0, filter_rnd
+  paddw                m0, m2
+  psrlw                m0, 4
+
+  INC_SRC_BY_SRC_STRIDE
+
+.x_other_y_other_loop:
+  movu                 m2, [srcq]
+  movu                 m4, [srcq+2]
+  movu                 m3, [srcq+src_strideq*2]
+  movu                 m5, [srcq+src_strideq*2+2]
+  pmullw               m2, filter_x_a
+  pmullw               m4, filter_x_b
+  paddw                m2, filter_rnd
+  pmullw               m3, filter_x_a
+  pmullw               m5, filter_x_b
+  paddw                m3, filter_rnd
+  paddw                m2, m4
+  paddw                m3, m5
+  psrlw                m2, 4
+  psrlw                m3, 4
+  mova                 m4, m2
+  mova                 m5, m3
+  pmullw               m0, filter_y_a
+  pmullw               m2, filter_y_b
+  paddw                m0, filter_rnd
+  pmullw               m4, filter_y_a
+  pmullw               m3, filter_y_b
+  paddw                m0, m2
+  paddw                m4, filter_rnd
+  mova                 m2, [dstq]
+  paddw                m4, m3
+  psrlw                m0, 4
+  psrlw                m4, 4
+  mova                 m3, [dstq+dst_strideq*2]
+%if %2 == 1 ; avg
+  pavgw                m0, [secq]
+  add                secq, sec_str
+  pavgw                m4, [secq]
+%endif
+  SUM_SSE              m0, m2, m4, m3, m6, m7
+  mova                 m0, m5
+
+  INC_SRC_BY_SRC_2STRIDE
+  lea                dstq, [dstq + dst_strideq * 4]
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+%endif
+  dec                   h
+  jg .x_other_y_other_loop
+%undef filter_x_a
+%undef filter_x_b
+%undef filter_y_a
+%undef filter_y_b
+%undef filter_rnd
+  STORE_AND_RET
+%endmacro
+
+INIT_XMM sse2
+SUBPEL_VARIANCE  8
+SUBPEL_VARIANCE 16
+
+INIT_XMM sse2
+SUBPEL_VARIANCE  8, 1
+SUBPEL_VARIANCE 16, 1
diff --git a/media/libvpx/vp9/encoder/x86/vp9_highbd_variance_sse2.c b/media/libvpx/vp9/encoder/x86/vp9_highbd_variance_sse2.c
new file mode 100644
index 000000000..29b7b2782
--- /dev/null
+++ b/media/libvpx/vp9/encoder/x86/vp9_highbd_variance_sse2.c
@@ -0,0 +1,349 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include "./vpx_config.h"
+#include "vp9/common/vp9_common.h"
+
+#include "vp9/encoder/vp9_variance.h"
+#include "vpx_ports/mem.h"
+
+#define DECL(w, opt) \
+int vp9_highbd_sub_pixel_variance##w##xh_##opt(const uint16_t *src, \
+                                               ptrdiff_t src_stride, \
+                                               int x_offset, int y_offset, \
+                                               const uint16_t *dst, \
+                                               ptrdiff_t dst_stride, \
+                                               int height, unsigned int *sse);
+#define DECLS(opt1, opt2) \
+DECL(8, opt1); \
+DECL(16, opt1)
+
+DECLS(sse2, sse);
+// DECLS(ssse3, ssse3);
+#undef DECLS
+#undef DECL
+
+#define FN(w, h, wf, wlog2, hlog2, opt, cast) \
+uint32_t vp9_highbd_sub_pixel_variance##w##x##h##_##opt(const uint8_t *src8, \
+                                                        int src_stride, \
+                                                        int x_offset, \
+                                                        int y_offset, \
+                                                        const uint8_t *dst8, \
+                                                        int dst_stride, \
+                                                        uint32_t *sse_ptr) { \
+  uint32_t sse; \
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8); \
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); \
+  int se = vp9_highbd_sub_pixel_variance##wf##xh_##opt(src, src_stride, \
+                                                       x_offset, y_offset, \
+                                                       dst, dst_stride, h, \
+                                                       &sse); \
+  if (w > wf) { \
+    unsigned int sse2; \
+    int se2 = vp9_highbd_sub_pixel_variance##wf##xh_##opt(src + 16, \
+                                                          src_stride, \
+                                                          x_offset, y_offset, \
+                                                          dst + 16, \
+                                                          dst_stride, \
+                                                          h, &sse2); \
+    se += se2; \
+    sse += sse2; \
+    if (w > wf * 2) { \
+      se2 = vp9_highbd_sub_pixel_variance##wf##xh_##opt(src + 32, src_stride, \
+                                                        x_offset, y_offset, \
+                                                        dst + 32, dst_stride, \
+                                                        h, &sse2); \
+      se += se2; \
+      sse += sse2; \
+      se2 = vp9_highbd_sub_pixel_variance##wf##xh_##opt( \
+          src + 48, src_stride, x_offset, y_offset, \
+          dst + 48, dst_stride, h, &sse2); \
+      se += se2; \
+      sse += sse2; \
+    } \
+  } \
+  *sse_ptr = sse; \
+  return sse - ((cast se * se) >> (wlog2 + hlog2)); \
+} \
+\
+uint32_t vp9_highbd_10_sub_pixel_variance##w##x##h##_##opt( \
+    const uint8_t *src8, int src_stride, int x_offset, int y_offset, \
+    const uint8_t *dst8, int dst_stride, uint32_t *sse_ptr) { \
+  uint32_t sse; \
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8); \
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); \
+  int se = vp9_highbd_sub_pixel_variance##wf##xh_##opt(src, src_stride, \
+                                                       x_offset, y_offset, \
+                                                       dst, dst_stride, \
+                                                       h, &sse); \
+  if (w > wf) { \
+    uint32_t sse2; \
+    int se2 = vp9_highbd_sub_pixel_variance##wf##xh_##opt(src + 16, \
+                                                          src_stride, \
+                                                          x_offset, y_offset, \
+                                                          dst + 16, \
+                                                          dst_stride, \
+                                                          h, &sse2); \
+    se += se2; \
+    sse += sse2; \
+    if (w > wf * 2) { \
+      se2 = vp9_highbd_sub_pixel_variance##wf##xh_##opt(src + 32, src_stride, \
+                                                        x_offset, y_offset, \
+                                                        dst + 32, dst_stride, \
+                                                        h, &sse2); \
+      se += se2; \
+      sse += sse2; \
+      se2 = vp9_highbd_sub_pixel_variance##wf##xh_##opt(src + 48, src_stride, \
+                                                        x_offset, y_offset, \
+                                                        dst + 48, dst_stride, \
+                                                        h, &sse2); \
+      se += se2; \
+      sse += sse2; \
+    } \
+  } \
+  se = ROUND_POWER_OF_TWO(se, 2); \
+  sse = ROUND_POWER_OF_TWO(sse, 4); \
+  *sse_ptr = sse; \
+  return sse - ((cast se * se) >> (wlog2 + hlog2)); \
+} \
+\
+uint32_t vp9_highbd_12_sub_pixel_variance##w##x##h##_##opt( \
+    const uint8_t *src8, int src_stride, int x_offset, int y_offset, \
+    const uint8_t *dst8, int dst_stride, uint32_t *sse_ptr) { \
+  int start_row; \
+  uint32_t sse; \
+  int se = 0; \
+  uint64_t long_sse = 0; \
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8); \
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); \
+  for (start_row = 0; start_row < h; start_row +=16) { \
+    uint32_t sse2; \
+    int height = h - start_row < 16 ? h - start_row : 16; \
+    int se2 = vp9_highbd_sub_pixel_variance##wf##xh_##opt( \
+        src + (start_row * src_stride), src_stride, \
+        x_offset, y_offset, dst + (start_row * dst_stride), \
+        dst_stride, height, &sse2); \
+    se += se2; \
+    long_sse += sse2; \
+    if (w > wf) { \
+      se2 = vp9_highbd_sub_pixel_variance##wf##xh_##opt( \
+          src + 16 + (start_row * src_stride), src_stride, \
+          x_offset, y_offset, dst + 16 + (start_row * dst_stride), \
+          dst_stride, height, &sse2); \
+      se += se2; \
+      long_sse += sse2; \
+      if (w > wf * 2) { \
+        se2 = vp9_highbd_sub_pixel_variance##wf##xh_##opt( \
+            src + 32 + (start_row * src_stride), src_stride, \
+            x_offset, y_offset, dst + 32 + (start_row * dst_stride), \
+            dst_stride, height, &sse2); \
+        se += se2; \
+        long_sse += sse2; \
+        se2 = vp9_highbd_sub_pixel_variance##wf##xh_##opt( \
+            src + 48 + (start_row * src_stride), src_stride, \
+            x_offset, y_offset, dst + 48 + (start_row * dst_stride), \
+            dst_stride, height, &sse2); \
+        se += se2; \
+        long_sse += sse2; \
+      }\
+    } \
+  } \
+  se = ROUND_POWER_OF_TWO(se, 4); \
+  sse = ROUND_POWER_OF_TWO(long_sse, 8); \
+  *sse_ptr = sse; \
+  return sse - ((cast se * se) >> (wlog2 + hlog2)); \
+}
+
+#define FNS(opt1, opt2) \
+FN(64, 64, 16, 6, 6, opt1, (int64_t)); \
+FN(64, 32, 16, 6, 5, opt1, (int64_t)); \
+FN(32, 64, 16, 5, 6, opt1, (int64_t)); \
+FN(32, 32, 16, 5, 5, opt1, (int64_t)); \
+FN(32, 16, 16, 5, 4, opt1, (int64_t)); \
+FN(16, 32, 16, 4, 5, opt1, (int64_t)); \
+FN(16, 16, 16, 4, 4, opt1, (int64_t)); \
+FN(16, 8, 16, 4, 3, opt1, (int64_t)); \
+FN(8, 16, 8, 3, 4, opt1, (int64_t)); \
+FN(8, 8, 8, 3, 3, opt1, (int64_t)); \
+FN(8, 4, 8, 3, 2, opt1, (int64_t));
+
+
+FNS(sse2, sse);
+
+#undef FNS
+#undef FN
+
+#define DECL(w, opt) \
+int vp9_highbd_sub_pixel_avg_variance##w##xh_##opt(const uint16_t *src, \
+                                                   ptrdiff_t src_stride, \
+                                                   int x_offset, int y_offset, \
+                                                   const uint16_t *dst, \
+                                                   ptrdiff_t dst_stride, \
+                                                   const uint16_t *sec, \
+                                                   ptrdiff_t sec_stride, \
+                                                   int height, \
+                                                   unsigned int *sse);
+#define DECLS(opt1) \
+DECL(16, opt1) \
+DECL(8, opt1)
+
+DECLS(sse2);
+#undef DECL
+#undef DECLS
+
+#define FN(w, h, wf, wlog2, hlog2, opt, cast) \
+uint32_t vp9_highbd_sub_pixel_avg_variance##w##x##h##_##opt( \
+    const uint8_t *src8, int src_stride, int x_offset, int y_offset, \
+    const uint8_t *dst8, int dst_stride, uint32_t *sse_ptr, \
+    const uint8_t *sec8) { \
+  uint32_t sse; \
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8); \
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); \
+  uint16_t *sec = CONVERT_TO_SHORTPTR(sec8); \
+  int se = vp9_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
+               src, src_stride, x_offset, \
+               y_offset, dst, dst_stride, sec, w, h, &sse); \
+  if (w > wf) { \
+    uint32_t sse2; \
+    int se2 = vp9_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
+                  src + 16, src_stride, x_offset, y_offset, \
+                  dst + 16, dst_stride, sec + 16, w, h, &sse2); \
+    se += se2; \
+    sse += sse2; \
+    if (w > wf * 2) { \
+      se2 = vp9_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
+                src + 32, src_stride, x_offset, y_offset, \
+                dst + 32, dst_stride, sec + 32, w, h, &sse2); \
+      se += se2; \
+      sse += sse2; \
+      se2 = vp9_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
+                src + 48, src_stride, x_offset, y_offset, \
+                dst + 48, dst_stride, sec + 48, w, h, &sse2); \
+      se += se2; \
+      sse += sse2; \
+    } \
+  } \
+  *sse_ptr = sse; \
+  return sse - ((cast se * se) >> (wlog2 + hlog2)); \
+} \
+\
+uint32_t vp9_highbd_10_sub_pixel_avg_variance##w##x##h##_##opt( \
+    const uint8_t *src8, int src_stride, int x_offset, int y_offset, \
+    const uint8_t *dst8, int dst_stride, uint32_t *sse_ptr, \
+    const uint8_t *sec8) { \
+  uint32_t sse; \
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8); \
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); \
+  uint16_t *sec = CONVERT_TO_SHORTPTR(sec8); \
+  int se = vp9_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
+                                            src, src_stride, x_offset, \
+                                            y_offset, dst, dst_stride, \
+                                            sec, w, h, &sse); \
+  if (w > wf) { \
+    uint32_t sse2; \
+    int se2 = vp9_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
+                                            src + 16, src_stride, \
+                                            x_offset, y_offset, \
+                                            dst + 16, dst_stride, \
+                                            sec + 16, w, h, &sse2); \
+    se += se2; \
+    sse += sse2; \
+    if (w > wf * 2) { \
+      se2 = vp9_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
+                                            src + 32, src_stride, \
+                                            x_offset, y_offset, \
+                                            dst + 32, dst_stride, \
+                                            sec + 32, w, h, &sse2); \
+      se += se2; \
+      sse += sse2; \
+      se2 = vp9_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
+                                            src + 48, src_stride, \
+                                            x_offset, y_offset, \
+                                            dst + 48, dst_stride, \
+                                            sec + 48, w, h, &sse2); \
+      se += se2; \
+      sse += sse2; \
+    } \
+  } \
+  se = ROUND_POWER_OF_TWO(se, 2); \
+  sse = ROUND_POWER_OF_TWO(sse, 4); \
+  *sse_ptr = sse; \
+  return sse - ((cast se * se) >> (wlog2 + hlog2)); \
+} \
+\
+uint32_t vp9_highbd_12_sub_pixel_avg_variance##w##x##h##_##opt( \
+    const uint8_t *src8, int src_stride, int x_offset, int y_offset, \
+    const uint8_t *dst8, int dst_stride, uint32_t *sse_ptr, \
+    const uint8_t *sec8) { \
+  int start_row; \
+  uint32_t sse; \
+  int se = 0; \
+  uint64_t long_sse = 0; \
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8); \
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); \
+  uint16_t *sec = CONVERT_TO_SHORTPTR(sec8); \
+  for (start_row = 0; start_row < h; start_row +=16) { \
+    uint32_t sse2; \
+    int height = h - start_row < 16 ? h - start_row : 16; \
+    int se2 = vp9_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
+                src + (start_row * src_stride), src_stride, x_offset, \
+                y_offset, dst + (start_row * dst_stride), dst_stride, \
+                sec + (start_row * w), w, height, &sse2); \
+    se += se2; \
+    long_sse += sse2; \
+    if (w > wf) { \
+      se2 = vp9_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
+                src + 16 + (start_row * src_stride), src_stride, \
+                x_offset, y_offset, \
+                dst + 16 + (start_row * dst_stride), dst_stride, \
+                sec + 16 + (start_row * w), w, height, &sse2); \
+      se += se2; \
+      long_sse += sse2; \
+      if (w > wf * 2) { \
+        se2 = vp9_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
+                src + 32 + (start_row * src_stride), src_stride, \
+                x_offset, y_offset, \
+                dst + 32 + (start_row * dst_stride), dst_stride, \
+                sec + 32 + (start_row * w), w, height, &sse2); \
+        se += se2; \
+        long_sse += sse2; \
+        se2 = vp9_highbd_sub_pixel_avg_variance##wf##xh_##opt( \
+                src + 48 + (start_row * src_stride), src_stride, \
+                x_offset, y_offset, \
+                dst + 48 + (start_row * dst_stride), dst_stride, \
+                sec + 48 + (start_row * w), w, height, &sse2); \
+        se += se2; \
+        long_sse += sse2; \
+      } \
+    } \
+  } \
+  se = ROUND_POWER_OF_TWO(se, 4); \
+  sse = ROUND_POWER_OF_TWO(long_sse, 8); \
+  *sse_ptr = sse; \
+  return sse - ((cast se * se) >> (wlog2 + hlog2)); \
+}
+
+
+#define FNS(opt1) \
+FN(64, 64, 16, 6, 6, opt1, (int64_t)); \
+FN(64, 32, 16, 6, 5, opt1, (int64_t)); \
+FN(32, 64, 16, 5, 6, opt1, (int64_t)); \
+FN(32, 32, 16, 5, 5, opt1, (int64_t)); \
+FN(32, 16, 16, 5, 4, opt1, (int64_t)); \
+FN(16, 32, 16, 4, 5, opt1, (int64_t)); \
+FN(16, 16, 16, 4, 4, opt1, (int64_t)); \
+FN(16, 8, 16, 4, 3, opt1, (int64_t)); \
+FN(8, 16, 8, 4, 3, opt1, (int64_t)); \
+FN(8, 8, 8, 3, 3, opt1, (int64_t)); \
+FN(8, 4, 8, 3, 2, opt1, (int64_t));
+
+FNS(sse2);
+
+#undef FNS
+#undef FN
diff --git a/media/libvpx/vp9/encoder/x86/vp9_quantize_sse2.c b/media/libvpx/vp9/encoder/x86/vp9_quantize_sse2.c
new file mode 100644
index 000000000..71fdfd716
--- /dev/null
+++ b/media/libvpx/vp9/encoder/x86/vp9_quantize_sse2.c
@@ -0,0 +1,419 @@
+/*
+ *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <emmintrin.h>
+#include <xmmintrin.h>
+
+#include "./vp9_rtcd.h"
+#include "vpx/vpx_integer.h"
+
+void vp9_quantize_b_sse2(const int16_t* coeff_ptr, intptr_t n_coeffs,
+                         int skip_block, const int16_t* zbin_ptr,
+                         const int16_t* round_ptr, const int16_t* quant_ptr,
+                         const int16_t* quant_shift_ptr, int16_t* qcoeff_ptr,
+                         int16_t* dqcoeff_ptr, const int16_t* dequant_ptr,
+                         uint16_t* eob_ptr,
+                         const int16_t* scan_ptr,
+                         const int16_t* iscan_ptr) {
+  __m128i zero;
+  (void)scan_ptr;
+
+  coeff_ptr += n_coeffs;
+  iscan_ptr += n_coeffs;
+  qcoeff_ptr += n_coeffs;
+  dqcoeff_ptr += n_coeffs;
+  n_coeffs = -n_coeffs;
+  zero = _mm_setzero_si128();
+  if (!skip_block) {
+    __m128i eob;
+    __m128i zbin;
+    __m128i round, quant, dequant, shift;
+    {
+      __m128i coeff0, coeff1;
+
+      // Setup global values
+      {
+        __m128i pw_1;
+        zbin = _mm_load_si128((const __m128i*)zbin_ptr);
+        round = _mm_load_si128((const __m128i*)round_ptr);
+        quant = _mm_load_si128((const __m128i*)quant_ptr);
+        pw_1 = _mm_set1_epi16(1);
+        zbin = _mm_sub_epi16(zbin, pw_1);
+        dequant = _mm_load_si128((const __m128i*)dequant_ptr);
+        shift = _mm_load_si128((const __m128i*)quant_shift_ptr);
+      }
+
+      {
+        __m128i coeff0_sign, coeff1_sign;
+        __m128i qcoeff0, qcoeff1;
+        __m128i qtmp0, qtmp1;
+        __m128i cmp_mask0, cmp_mask1;
+        // Do DC and first 15 AC
+        coeff0 = _mm_load_si128((const __m128i*)(coeff_ptr + n_coeffs));
+        coeff1 = _mm_load_si128((const __m128i*)(coeff_ptr + n_coeffs) + 1);
+
+        // Poor man's sign extract
+        coeff0_sign = _mm_srai_epi16(coeff0, 15);
+        coeff1_sign = _mm_srai_epi16(coeff1, 15);
+        qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin);
+        zbin = _mm_unpackhi_epi64(zbin, zbin);  // Switch DC to AC
+        cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);
+        qcoeff0 = _mm_adds_epi16(qcoeff0, round);
+        round = _mm_unpackhi_epi64(round, round);
+        qcoeff1 = _mm_adds_epi16(qcoeff1, round);
+        qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
+        quant = _mm_unpackhi_epi64(quant, quant);
+        qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
+        qtmp0 = _mm_add_epi16(qtmp0, qcoeff0);
+        qtmp1 = _mm_add_epi16(qtmp1, qcoeff1);
+        qcoeff0 = _mm_mulhi_epi16(qtmp0, shift);
+        shift = _mm_unpackhi_epi64(shift, shift);
+        qcoeff1 = _mm_mulhi_epi16(qtmp1, shift);
+
+        // Reinsert signs
+        qcoeff0 = _mm_xor_si128(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(qcoeff1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        // Mask out zbin threshold coeffs
+        qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
+        qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
+
+        _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), qcoeff0);
+        _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
+
+        coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
+        dequant = _mm_unpackhi_epi64(dequant, dequant);
+        coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+
+        _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), coeff0);
+        _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, coeff1);
+      }
+
+      {
+        // Scan for eob
+        __m128i zero_coeff0, zero_coeff1;
+        __m128i nzero_coeff0, nzero_coeff1;
+        __m128i iscan0, iscan1;
+        __m128i eob1;
+        zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
+        zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
+        nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
+        nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
+        iscan0 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs));
+        iscan1 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs) + 1);
+        // Add one to convert from indices to counts
+        iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
+        iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
+        eob = _mm_and_si128(iscan0, nzero_coeff0);
+        eob1 = _mm_and_si128(iscan1, nzero_coeff1);
+        eob = _mm_max_epi16(eob, eob1);
+      }
+      n_coeffs += 8 * 2;
+    }
+
+    // AC only loop
+    while (n_coeffs < 0) {
+      __m128i coeff0, coeff1;
+      {
+        __m128i coeff0_sign, coeff1_sign;
+        __m128i qcoeff0, qcoeff1;
+        __m128i qtmp0, qtmp1;
+        __m128i cmp_mask0, cmp_mask1;
+
+        coeff0 = _mm_load_si128((const __m128i*)(coeff_ptr + n_coeffs));
+        coeff1 = _mm_load_si128((const __m128i*)(coeff_ptr + n_coeffs) + 1);
+
+        // Poor man's sign extract
+        coeff0_sign = _mm_srai_epi16(coeff0, 15);
+        coeff1_sign = _mm_srai_epi16(coeff1, 15);
+        qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin);
+        cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);
+        qcoeff0 = _mm_adds_epi16(qcoeff0, round);
+        qcoeff1 = _mm_adds_epi16(qcoeff1, round);
+        qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
+        qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
+        qtmp0 = _mm_add_epi16(qtmp0, qcoeff0);
+        qtmp1 = _mm_add_epi16(qtmp1, qcoeff1);
+        qcoeff0 = _mm_mulhi_epi16(qtmp0, shift);
+        qcoeff1 = _mm_mulhi_epi16(qtmp1, shift);
+
+        // Reinsert signs
+        qcoeff0 = _mm_xor_si128(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(qcoeff1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        // Mask out zbin threshold coeffs
+        qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
+        qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
+
+        _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), qcoeff0);
+        _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
+
+        coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
+        coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+
+        _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), coeff0);
+        _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, coeff1);
+      }
+
+      {
+        // Scan for eob
+        __m128i zero_coeff0, zero_coeff1;
+        __m128i nzero_coeff0, nzero_coeff1;
+        __m128i iscan0, iscan1;
+        __m128i eob0, eob1;
+        zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
+        zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
+        nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
+        nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
+        iscan0 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs));
+        iscan1 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs) + 1);
+        // Add one to convert from indices to counts
+        iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
+        iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
+        eob0 = _mm_and_si128(iscan0, nzero_coeff0);
+        eob1 = _mm_and_si128(iscan1, nzero_coeff1);
+        eob0 = _mm_max_epi16(eob0, eob1);
+        eob = _mm_max_epi16(eob, eob0);
+      }
+      n_coeffs += 8 * 2;
+    }
+
+    // Accumulate EOB
+    {
+      __m128i eob_shuffled;
+      eob_shuffled = _mm_shuffle_epi32(eob, 0xe);
+      eob = _mm_max_epi16(eob, eob_shuffled);
+      eob_shuffled = _mm_shufflelo_epi16(eob, 0xe);
+      eob = _mm_max_epi16(eob, eob_shuffled);
+      eob_shuffled = _mm_shufflelo_epi16(eob, 0x1);
+      eob = _mm_max_epi16(eob, eob_shuffled);
+      *eob_ptr = _mm_extract_epi16(eob, 1);
+    }
+  } else {
+    do {
+      _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), zero);
+      _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, zero);
+      _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), zero);
+      _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, zero);
+      n_coeffs += 8 * 2;
+    } while (n_coeffs < 0);
+    *eob_ptr = 0;
+  }
+}
+
+void vp9_quantize_fp_sse2(const int16_t* coeff_ptr, intptr_t n_coeffs,
+                          int skip_block, const int16_t* zbin_ptr,
+                          const int16_t* round_ptr, const int16_t* quant_ptr,
+                          const int16_t* quant_shift_ptr, int16_t* qcoeff_ptr,
+                          int16_t* dqcoeff_ptr, const int16_t* dequant_ptr,
+                          uint16_t* eob_ptr,
+                          const int16_t* scan_ptr,
+                          const int16_t* iscan_ptr) {
+  __m128i zero;
+  __m128i thr;
+  int16_t nzflag;
+  (void)scan_ptr;
+  (void)zbin_ptr;
+  (void)quant_shift_ptr;
+
+  coeff_ptr += n_coeffs;
+  iscan_ptr += n_coeffs;
+  qcoeff_ptr += n_coeffs;
+  dqcoeff_ptr += n_coeffs;
+  n_coeffs = -n_coeffs;
+  zero = _mm_setzero_si128();
+
+  if (!skip_block) {
+    __m128i eob;
+    __m128i round, quant, dequant;
+    {
+      __m128i coeff0, coeff1;
+
+      // Setup global values
+      {
+        round = _mm_load_si128((const __m128i*)round_ptr);
+        quant = _mm_load_si128((const __m128i*)quant_ptr);
+        dequant = _mm_load_si128((const __m128i*)dequant_ptr);
+      }
+
+      {
+        __m128i coeff0_sign, coeff1_sign;
+        __m128i qcoeff0, qcoeff1;
+        __m128i qtmp0, qtmp1;
+        // Do DC and first 15 AC
+        coeff0 = _mm_load_si128((const __m128i*)(coeff_ptr + n_coeffs));
+        coeff1 = _mm_load_si128((const __m128i*)(coeff_ptr + n_coeffs) + 1);
+
+        // Poor man's sign extract
+        coeff0_sign = _mm_srai_epi16(coeff0, 15);
+        coeff1_sign = _mm_srai_epi16(coeff1, 15);
+        qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        qcoeff0 = _mm_adds_epi16(qcoeff0, round);
+        round = _mm_unpackhi_epi64(round, round);
+        qcoeff1 = _mm_adds_epi16(qcoeff1, round);
+        qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
+        quant = _mm_unpackhi_epi64(quant, quant);
+        qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
+
+        // Reinsert signs
+        qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), qcoeff0);
+        _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
+
+        coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
+        dequant = _mm_unpackhi_epi64(dequant, dequant);
+        coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+
+        _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), coeff0);
+        _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, coeff1);
+      }
+
+      {
+        // Scan for eob
+        __m128i zero_coeff0, zero_coeff1;
+        __m128i nzero_coeff0, nzero_coeff1;
+        __m128i iscan0, iscan1;
+        __m128i eob1;
+        zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
+        zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
+        nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
+        nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
+        iscan0 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs));
+        iscan1 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs) + 1);
+        // Add one to convert from indices to counts
+        iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
+        iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
+        eob = _mm_and_si128(iscan0, nzero_coeff0);
+        eob1 = _mm_and_si128(iscan1, nzero_coeff1);
+        eob = _mm_max_epi16(eob, eob1);
+      }
+      n_coeffs += 8 * 2;
+    }
+
+    thr = _mm_srai_epi16(dequant, 1);
+
+    // AC only loop
+    while (n_coeffs < 0) {
+      __m128i coeff0, coeff1;
+      {
+        __m128i coeff0_sign, coeff1_sign;
+        __m128i qcoeff0, qcoeff1;
+        __m128i qtmp0, qtmp1;
+
+        coeff0 = _mm_load_si128((const __m128i*)(coeff_ptr + n_coeffs));
+        coeff1 = _mm_load_si128((const __m128i*)(coeff_ptr + n_coeffs) + 1);
+
+        // Poor man's sign extract
+        coeff0_sign = _mm_srai_epi16(coeff0, 15);
+        coeff1_sign = _mm_srai_epi16(coeff1, 15);
+        qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
+        qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
+        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+        nzflag = _mm_movemask_epi8(_mm_cmpgt_epi16(qcoeff0, thr)) |
+            _mm_movemask_epi8(_mm_cmpgt_epi16(qcoeff1, thr));
+
+        if (nzflag) {
+          qcoeff0 = _mm_adds_epi16(qcoeff0, round);
+          qcoeff1 = _mm_adds_epi16(qcoeff1, round);
+          qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
+          qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
+
+          // Reinsert signs
+          qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign);
+          qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign);
+          qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
+          qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
+
+          _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), qcoeff0);
+          _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
+
+          coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
+          coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
+
+          _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), coeff0);
+          _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, coeff1);
+        } else {
+          _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), zero);
+          _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, zero);
+
+          _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), zero);
+          _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, zero);
+        }
+      }
+
+      if (nzflag) {
+        // Scan for eob
+        __m128i zero_coeff0, zero_coeff1;
+        __m128i nzero_coeff0, nzero_coeff1;
+        __m128i iscan0, iscan1;
+        __m128i eob0, eob1;
+        zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
+        zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
+        nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
+        nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
+        iscan0 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs));
+        iscan1 = _mm_load_si128((const __m128i*)(iscan_ptr + n_coeffs) + 1);
+        // Add one to convert from indices to counts
+        iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
+        iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
+        eob0 = _mm_and_si128(iscan0, nzero_coeff0);
+        eob1 = _mm_and_si128(iscan1, nzero_coeff1);
+        eob0 = _mm_max_epi16(eob0, eob1);
+        eob = _mm_max_epi16(eob, eob0);
+      }
+      n_coeffs += 8 * 2;
+    }
+
+    // Accumulate EOB
+    {
+      __m128i eob_shuffled;
+      eob_shuffled = _mm_shuffle_epi32(eob, 0xe);
+      eob = _mm_max_epi16(eob, eob_shuffled);
+      eob_shuffled = _mm_shufflelo_epi16(eob, 0xe);
+      eob = _mm_max_epi16(eob, eob_shuffled);
+      eob_shuffled = _mm_shufflelo_epi16(eob, 0x1);
+      eob = _mm_max_epi16(eob, eob_shuffled);
+      *eob_ptr = _mm_extract_epi16(eob, 1);
+    }
+  } else {
+    do {
+      _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs), zero);
+      _mm_store_si128((__m128i*)(dqcoeff_ptr + n_coeffs) + 1, zero);
+      _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs), zero);
+      _mm_store_si128((__m128i*)(qcoeff_ptr + n_coeffs) + 1, zero);
+      n_coeffs += 8 * 2;
+    } while (n_coeffs < 0);
+    *eob_ptr = 0;
+  }
+}
diff --git a/media/libvpx/vp9/encoder/x86/vp9_quantize_ssse3_x86_64.asm b/media/libvpx/vp9/encoder/x86/vp9_quantize_ssse3_x86_64.asm
new file mode 100644
index 000000000..449d52b22
--- /dev/null
+++ b/media/libvpx/vp9/encoder/x86/vp9_quantize_ssse3_x86_64.asm
@@ -0,0 +1,399 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION_RODATA
+pw_1: times 8 dw 1
+
+SECTION .text
+
+; TODO(yunqingwang)fix quantize_b code for skip=1 case.
+%macro QUANTIZE_FN 2
+cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \
+                                shift, qcoeff, dqcoeff, dequant, \
+                                eob, scan, iscan
+  cmp                    dword skipm, 0
+  jne .blank
+
+  ; actual quantize loop - setup pointers, rounders, etc.
+  movifnidn                   coeffq, coeffmp
+  movifnidn                  ncoeffq, ncoeffmp
+  mov                             r2, dequantmp
+  movifnidn                    zbinq, zbinmp
+  movifnidn                   roundq, roundmp
+  movifnidn                   quantq, quantmp
+  mova                            m0, [zbinq]              ; m0 = zbin
+  mova                            m1, [roundq]             ; m1 = round
+  mova                            m2, [quantq]             ; m2 = quant
+%ifidn %1, b_32x32
+  pcmpeqw                         m5, m5
+  psrlw                           m5, 15
+  paddw                           m0, m5
+  paddw                           m1, m5
+  psrlw                           m0, 1                    ; m0 = (m0 + 1) / 2
+  psrlw                           m1, 1                    ; m1 = (m1 + 1) / 2
+%endif
+  mova                            m3, [r2q]                ; m3 = dequant
+  psubw                           m0, [pw_1]
+  mov                             r2, shiftmp
+  mov                             r3, qcoeffmp
+  mova                            m4, [r2]                 ; m4 = shift
+  mov                             r4, dqcoeffmp
+  mov                             r5, iscanmp
+%ifidn %1, b_32x32
+  psllw                           m4, 1
+%endif
+  pxor                            m5, m5                   ; m5 = dedicated zero
+  DEFINE_ARGS coeff, ncoeff, d1, qcoeff, dqcoeff, iscan, d2, d3, d4, d5, eob
+  lea                         coeffq, [  coeffq+ncoeffq*2]
+  lea                         iscanq, [  iscanq+ncoeffq*2]
+  lea                        qcoeffq, [ qcoeffq+ncoeffq*2]
+  lea                       dqcoeffq, [dqcoeffq+ncoeffq*2]
+  neg                        ncoeffq
+
+  ; get DC and first 15 AC coeffs
+  mova                            m9, [  coeffq+ncoeffq*2+ 0] ; m9 = c[i]
+  mova                           m10, [  coeffq+ncoeffq*2+16] ; m10 = c[i]
+  pabsw                           m6, m9                   ; m6 = abs(m9)
+  pabsw                          m11, m10                  ; m11 = abs(m10)
+  pcmpgtw                         m7, m6, m0               ; m7 = c[i] >= zbin
+  punpckhqdq                      m0, m0
+  pcmpgtw                        m12, m11, m0              ; m12 = c[i] >= zbin
+  paddsw                          m6, m1                   ; m6 += round
+  punpckhqdq                      m1, m1
+  paddsw                         m11, m1                   ; m11 += round
+  pmulhw                          m8, m6, m2               ; m8 = m6*q>>16
+  punpckhqdq                      m2, m2
+  pmulhw                         m13, m11, m2              ; m13 = m11*q>>16
+  paddw                           m8, m6                   ; m8 += m6
+  paddw                          m13, m11                  ; m13 += m11
+  pmulhw                          m8, m4                   ; m8 = m8*qsh>>16
+  punpckhqdq                      m4, m4
+  pmulhw                         m13, m4                   ; m13 = m13*qsh>>16
+  psignw                          m8, m9                   ; m8 = reinsert sign
+  psignw                         m13, m10                  ; m13 = reinsert sign
+  pand                            m8, m7
+  pand                           m13, m12
+  mova        [qcoeffq+ncoeffq*2+ 0], m8
+  mova        [qcoeffq+ncoeffq*2+16], m13
+%ifidn %1, b_32x32
+  pabsw                           m8, m8
+  pabsw                          m13, m13
+%endif
+  pmullw                          m8, m3                   ; dqc[i] = qc[i] * q
+  punpckhqdq                      m3, m3
+  pmullw                         m13, m3                   ; dqc[i] = qc[i] * q
+%ifidn %1, b_32x32
+  psrlw                           m8, 1
+  psrlw                          m13, 1
+  psignw                          m8, m9
+  psignw                         m13, m10
+%endif
+  mova       [dqcoeffq+ncoeffq*2+ 0], m8
+  mova       [dqcoeffq+ncoeffq*2+16], m13
+  pcmpeqw                         m8, m5                   ; m8 = c[i] == 0
+  pcmpeqw                        m13, m5                   ; m13 = c[i] == 0
+  mova                            m6, [  iscanq+ncoeffq*2+ 0] ; m6 = scan[i]
+  mova                           m11, [  iscanq+ncoeffq*2+16] ; m11 = scan[i]
+  psubw                           m6, m7                   ; m6 = scan[i] + 1
+  psubw                          m11, m12                  ; m11 = scan[i] + 1
+  pandn                           m8, m6                   ; m8 = max(eob)
+  pandn                          m13, m11                  ; m13 = max(eob)
+  pmaxsw                          m8, m13
+  add                        ncoeffq, mmsize
+  jz .accumulate_eob
+
+.ac_only_loop:
+  mova                            m9, [  coeffq+ncoeffq*2+ 0] ; m9 = c[i]
+  mova                           m10, [  coeffq+ncoeffq*2+16] ; m10 = c[i]
+  pabsw                           m6, m9                   ; m6 = abs(m9)
+  pabsw                          m11, m10                  ; m11 = abs(m10)
+  pcmpgtw                         m7, m6, m0               ; m7 = c[i] >= zbin
+  pcmpgtw                        m12, m11, m0              ; m12 = c[i] >= zbin
+%ifidn %1, b_32x32
+  pmovmskb                       r6d, m7
+  pmovmskb                       r2d, m12
+  or                              r6, r2
+  jz .skip_iter
+%endif
+  paddsw                          m6, m1                   ; m6 += round
+  paddsw                         m11, m1                   ; m11 += round
+  pmulhw                         m14, m6, m2               ; m14 = m6*q>>16
+  pmulhw                         m13, m11, m2              ; m13 = m11*q>>16
+  paddw                          m14, m6                   ; m14 += m6
+  paddw                          m13, m11                  ; m13 += m11
+  pmulhw                         m14, m4                   ; m14 = m14*qsh>>16
+  pmulhw                         m13, m4                   ; m13 = m13*qsh>>16
+  psignw                         m14, m9                   ; m14 = reinsert sign
+  psignw                         m13, m10                  ; m13 = reinsert sign
+  pand                           m14, m7
+  pand                           m13, m12
+  mova        [qcoeffq+ncoeffq*2+ 0], m14
+  mova        [qcoeffq+ncoeffq*2+16], m13
+%ifidn %1, b_32x32
+  pabsw                          m14, m14
+  pabsw                          m13, m13
+%endif
+  pmullw                         m14, m3                   ; dqc[i] = qc[i] * q
+  pmullw                         m13, m3                   ; dqc[i] = qc[i] * q
+%ifidn %1, b_32x32
+  psrlw                          m14, 1
+  psrlw                          m13, 1
+  psignw                         m14, m9
+  psignw                         m13, m10
+%endif
+  mova       [dqcoeffq+ncoeffq*2+ 0], m14
+  mova       [dqcoeffq+ncoeffq*2+16], m13
+  pcmpeqw                        m14, m5                   ; m14 = c[i] == 0
+  pcmpeqw                        m13, m5                   ; m13 = c[i] == 0
+  mova                            m6, [  iscanq+ncoeffq*2+ 0] ; m6 = scan[i]
+  mova                           m11, [  iscanq+ncoeffq*2+16] ; m11 = scan[i]
+  psubw                           m6, m7                   ; m6 = scan[i] + 1
+  psubw                          m11, m12                  ; m11 = scan[i] + 1
+  pandn                          m14, m6                   ; m14 = max(eob)
+  pandn                          m13, m11                  ; m13 = max(eob)
+  pmaxsw                          m8, m14
+  pmaxsw                          m8, m13
+  add                        ncoeffq, mmsize
+  jl .ac_only_loop
+
+%ifidn %1, b_32x32
+  jmp .accumulate_eob
+.skip_iter:
+  mova        [qcoeffq+ncoeffq*2+ 0], m5
+  mova        [qcoeffq+ncoeffq*2+16], m5
+  mova       [dqcoeffq+ncoeffq*2+ 0], m5
+  mova       [dqcoeffq+ncoeffq*2+16], m5
+  add                        ncoeffq, mmsize
+  jl .ac_only_loop
+%endif
+
+.accumulate_eob:
+  ; horizontally accumulate/max eobs and write into [eob] memory pointer
+  mov                             r2, eobmp
+  pshufd                          m7, m8, 0xe
+  pmaxsw                          m8, m7
+  pshuflw                         m7, m8, 0xe
+  pmaxsw                          m8, m7
+  pshuflw                         m7, m8, 0x1
+  pmaxsw                          m8, m7
+  pextrw                          r6, m8, 0
+  mov                             [r2], r6
+  RET
+
+  ; skip-block, i.e. just write all zeroes
+.blank:
+  mov                             r0, dqcoeffmp
+  movifnidn                  ncoeffq, ncoeffmp
+  mov                             r2, qcoeffmp
+  mov                             r3, eobmp
+  DEFINE_ARGS dqcoeff, ncoeff, qcoeff, eob
+  lea                       dqcoeffq, [dqcoeffq+ncoeffq*2]
+  lea                        qcoeffq, [ qcoeffq+ncoeffq*2]
+  neg                        ncoeffq
+  pxor                            m7, m7
+.blank_loop:
+  mova       [dqcoeffq+ncoeffq*2+ 0], m7
+  mova       [dqcoeffq+ncoeffq*2+16], m7
+  mova        [qcoeffq+ncoeffq*2+ 0], m7
+  mova        [qcoeffq+ncoeffq*2+16], m7
+  add                        ncoeffq, mmsize
+  jl .blank_loop
+  mov                    word [eobq], 0
+  RET
+%endmacro
+
+INIT_XMM ssse3
+QUANTIZE_FN b, 7
+QUANTIZE_FN b_32x32, 7
+
+%macro QUANTIZE_FP 2
+cglobal quantize_%1, 0, %2, 15, coeff, ncoeff, skip, zbin, round, quant, \
+                                shift, qcoeff, dqcoeff, dequant, \
+                                eob, scan, iscan
+  cmp                    dword skipm, 0
+  jne .blank
+
+  ; actual quantize loop - setup pointers, rounders, etc.
+  movifnidn                   coeffq, coeffmp
+  movifnidn                  ncoeffq, ncoeffmp
+  mov                             r2, dequantmp
+  movifnidn                    zbinq, zbinmp
+  movifnidn                   roundq, roundmp
+  movifnidn                   quantq, quantmp
+  mova                            m1, [roundq]             ; m1 = round
+  mova                            m2, [quantq]             ; m2 = quant
+%ifidn %1, fp_32x32
+  pcmpeqw                         m5, m5
+  psrlw                           m5, 15
+  paddw                           m1, m5
+  psrlw                           m1, 1                    ; m1 = (m1 + 1) / 2
+%endif
+  mova                            m3, [r2q]                ; m3 = dequant
+  mov                             r3, qcoeffmp
+  mov                             r4, dqcoeffmp
+  mov                             r5, iscanmp
+%ifidn %1, fp_32x32
+  psllw                           m2, 1
+%endif
+  pxor                            m5, m5                   ; m5 = dedicated zero
+
+  lea                         coeffq, [  coeffq+ncoeffq*2]
+  lea                            r5q, [  r5q+ncoeffq*2]
+  lea                            r3q, [ r3q+ncoeffq*2]
+  lea                            r4q, [r4q+ncoeffq*2]
+  neg                        ncoeffq
+
+  ; get DC and first 15 AC coeffs
+  mova                            m9, [  coeffq+ncoeffq*2+ 0] ; m9 = c[i]
+  mova                           m10, [  coeffq+ncoeffq*2+16] ; m10 = c[i]
+  pabsw                           m6, m9                   ; m6 = abs(m9)
+  pabsw                          m11, m10                  ; m11 = abs(m10)
+  pcmpeqw                         m7, m7
+
+  paddsw                          m6, m1                   ; m6 += round
+  punpckhqdq                      m1, m1
+  paddsw                         m11, m1                   ; m11 += round
+  pmulhw                          m8, m6, m2               ; m8 = m6*q>>16
+  punpckhqdq                      m2, m2
+  pmulhw                         m13, m11, m2              ; m13 = m11*q>>16
+  psignw                          m8, m9                   ; m8 = reinsert sign
+  psignw                         m13, m10                  ; m13 = reinsert sign
+  mova            [r3q+ncoeffq*2+ 0], m8
+  mova            [r3q+ncoeffq*2+16], m13
+%ifidn %1, fp_32x32
+  pabsw                           m8, m8
+  pabsw                          m13, m13
+%endif
+  pmullw                          m8, m3                   ; r4[i] = r3[i] * q
+  punpckhqdq                      m3, m3
+  pmullw                         m13, m3                   ; r4[i] = r3[i] * q
+%ifidn %1, fp_32x32
+  psrlw                           m8, 1
+  psrlw                          m13, 1
+  psignw                          m8, m9
+  psignw                         m13, m10
+  psrlw                           m0, m3, 2
+%else
+  psrlw                           m0, m3, 1
+%endif
+  mova            [r4q+ncoeffq*2+ 0], m8
+  mova            [r4q+ncoeffq*2+16], m13
+  pcmpeqw                         m8, m5                   ; m8 = c[i] == 0
+  pcmpeqw                        m13, m5                   ; m13 = c[i] == 0
+  mova                            m6, [  r5q+ncoeffq*2+ 0] ; m6 = scan[i]
+  mova                           m11, [  r5q+ncoeffq*2+16] ; m11 = scan[i]
+  psubw                           m6, m7                   ; m6 = scan[i] + 1
+  psubw                          m11, m7                   ; m11 = scan[i] + 1
+  pandn                           m8, m6                   ; m8 = max(eob)
+  pandn                          m13, m11                  ; m13 = max(eob)
+  pmaxsw                          m8, m13
+  add                        ncoeffq, mmsize
+  jz .accumulate_eob
+
+.ac_only_loop:
+  mova                            m9, [  coeffq+ncoeffq*2+ 0] ; m9 = c[i]
+  mova                           m10, [  coeffq+ncoeffq*2+16] ; m10 = c[i]
+  pabsw                           m6, m9                   ; m6 = abs(m9)
+  pabsw                          m11, m10                  ; m11 = abs(m10)
+
+  pcmpgtw                         m7, m6,  m0
+  pcmpgtw                        m12, m11, m0
+  pmovmskb                       r6d, m7
+  pmovmskb                       r2d, m12
+
+  or                              r6, r2
+  jz .skip_iter
+
+  pcmpeqw                         m7, m7
+
+  paddsw                          m6, m1                   ; m6 += round
+  paddsw                         m11, m1                   ; m11 += round
+  pmulhw                         m14, m6, m2               ; m14 = m6*q>>16
+  pmulhw                         m13, m11, m2              ; m13 = m11*q>>16
+  psignw                         m14, m9                   ; m14 = reinsert sign
+  psignw                         m13, m10                  ; m13 = reinsert sign
+  mova            [r3q+ncoeffq*2+ 0], m14
+  mova            [r3q+ncoeffq*2+16], m13
+%ifidn %1, fp_32x32
+  pabsw                          m14, m14
+  pabsw                          m13, m13
+%endif
+  pmullw                         m14, m3                   ; r4[i] = r3[i] * q
+  pmullw                         m13, m3                   ; r4[i] = r3[i] * q
+%ifidn %1, fp_32x32
+  psrlw                          m14, 1
+  psrlw                          m13, 1
+  psignw                         m14, m9
+  psignw                         m13, m10
+%endif
+  mova            [r4q+ncoeffq*2+ 0], m14
+  mova            [r4q+ncoeffq*2+16], m13
+  pcmpeqw                        m14, m5                   ; m14 = c[i] == 0
+  pcmpeqw                        m13, m5                   ; m13 = c[i] == 0
+  mova                            m6, [  r5q+ncoeffq*2+ 0] ; m6 = scan[i]
+  mova                           m11, [  r5q+ncoeffq*2+16] ; m11 = scan[i]
+  psubw                           m6, m7                   ; m6 = scan[i] + 1
+  psubw                          m11, m7                   ; m11 = scan[i] + 1
+  pandn                          m14, m6                   ; m14 = max(eob)
+  pandn                          m13, m11                  ; m13 = max(eob)
+  pmaxsw                          m8, m14
+  pmaxsw                          m8, m13
+  add                        ncoeffq, mmsize
+  jl .ac_only_loop
+
+  jmp .accumulate_eob
+.skip_iter:
+  mova            [r3q+ncoeffq*2+ 0], m5
+  mova            [r3q+ncoeffq*2+16], m5
+  mova            [r4q+ncoeffq*2+ 0], m5
+  mova            [r4q+ncoeffq*2+16], m5
+  add                        ncoeffq, mmsize
+  jl .ac_only_loop
+
+.accumulate_eob:
+  ; horizontally accumulate/max eobs and write into [eob] memory pointer
+  mov                             r2, eobmp
+  pshufd                          m7, m8, 0xe
+  pmaxsw                          m8, m7
+  pshuflw                         m7, m8, 0xe
+  pmaxsw                          m8, m7
+  pshuflw                         m7, m8, 0x1
+  pmaxsw                          m8, m7
+  pextrw                          r6, m8, 0
+  mov                           [r2], r6
+  RET
+
+  ; skip-block, i.e. just write all zeroes
+.blank:
+  mov                             r0, dqcoeffmp
+  movifnidn                  ncoeffq, ncoeffmp
+  mov                             r2, qcoeffmp
+  mov                             r3, eobmp
+
+  lea                            r0q, [r0q+ncoeffq*2]
+  lea                            r2q, [r2q+ncoeffq*2]
+  neg                        ncoeffq
+  pxor                            m7, m7
+.blank_loop:
+  mova            [r0q+ncoeffq*2+ 0], m7
+  mova            [r0q+ncoeffq*2+16], m7
+  mova            [r2q+ncoeffq*2+ 0], m7
+  mova            [r2q+ncoeffq*2+16], m7
+  add                        ncoeffq, mmsize
+  jl .blank_loop
+  mov                     word [r3q], 0
+  RET
+%endmacro
+
+INIT_XMM ssse3
+QUANTIZE_FP fp, 7
+QUANTIZE_FP fp_32x32, 7
diff --git a/media/libvpx/vp9/encoder/x86/vp9_ssim_opt_x86_64.asm b/media/libvpx/vp9/encoder/x86/vp9_ssim_opt_x86_64.asm
new file mode 100644
index 000000000..455d10d2c
--- /dev/null
+++ b/media/libvpx/vp9/encoder/x86/vp9_ssim_opt_x86_64.asm
@@ -0,0 +1,216 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%include "vpx_ports/x86_abi_support.asm"
+
+; tabulate_ssim - sums sum_s,sum_r,sum_sq_s,sum_sq_r, sum_sxr
+%macro TABULATE_SSIM 0
+        paddusw         xmm15, xmm3  ; sum_s
+        paddusw         xmm14, xmm4  ; sum_r
+        movdqa          xmm1, xmm3
+        pmaddwd         xmm1, xmm1
+        paddd           xmm13, xmm1 ; sum_sq_s
+        movdqa          xmm2, xmm4
+        pmaddwd         xmm2, xmm2
+        paddd           xmm12, xmm2 ; sum_sq_r
+        pmaddwd         xmm3, xmm4
+        paddd           xmm11, xmm3  ; sum_sxr
+%endmacro
+
+; Sum across the register %1 starting with q words
+%macro SUM_ACROSS_Q 1
+        movdqa          xmm2,%1
+        punpckldq       %1,xmm0
+        punpckhdq       xmm2,xmm0
+        paddq           %1,xmm2
+        movdqa          xmm2,%1
+        punpcklqdq      %1,xmm0
+        punpckhqdq      xmm2,xmm0
+        paddq           %1,xmm2
+%endmacro
+
+; Sum across the register %1 starting with q words
+%macro SUM_ACROSS_W 1
+        movdqa          xmm1, %1
+        punpcklwd       %1,xmm0
+        punpckhwd       xmm1,xmm0
+        paddd           %1, xmm1
+        SUM_ACROSS_Q    %1
+%endmacro
+;void ssim_parms_sse2(
+;    unsigned char *s,
+;    int sp,
+;    unsigned char *r,
+;    int rp
+;    unsigned long *sum_s,
+;    unsigned long *sum_r,
+;    unsigned long *sum_sq_s,
+;    unsigned long *sum_sq_r,
+;    unsigned long *sum_sxr);
+;
+; TODO: Use parm passing through structure, probably don't need the pxors
+; ( calling app will initialize to 0 ) could easily fit everything in sse2
+; without too much hastle, and can probably do better estimates with psadw
+; or pavgb At this point this is just meant to be first pass for calculating
+; all the parms needed for 16x16 ssim so we can play with dssim as distortion
+; in mode selection code.
+global sym(vp9_ssim_parms_16x16_sse2) PRIVATE
+sym(vp9_ssim_parms_16x16_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 9
+    SAVE_XMM 15
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    mov             rsi,        arg(0) ;s
+    mov             rcx,        arg(1) ;sp
+    mov             rdi,        arg(2) ;r
+    mov             rax,        arg(3) ;rp
+
+    pxor            xmm0, xmm0
+    pxor            xmm15,xmm15  ;sum_s
+    pxor            xmm14,xmm14  ;sum_r
+    pxor            xmm13,xmm13  ;sum_sq_s
+    pxor            xmm12,xmm12  ;sum_sq_r
+    pxor            xmm11,xmm11  ;sum_sxr
+
+    mov             rdx, 16      ;row counter
+.NextRow:
+
+    ;grab source and reference pixels
+    movdqu          xmm5, [rsi]
+    movdqu          xmm6, [rdi]
+    movdqa          xmm3, xmm5
+    movdqa          xmm4, xmm6
+    punpckhbw       xmm3, xmm0 ; high_s
+    punpckhbw       xmm4, xmm0 ; high_r
+
+    TABULATE_SSIM
+
+    movdqa          xmm3, xmm5
+    movdqa          xmm4, xmm6
+    punpcklbw       xmm3, xmm0 ; low_s
+    punpcklbw       xmm4, xmm0 ; low_r
+
+    TABULATE_SSIM
+
+    add             rsi, rcx   ; next s row
+    add             rdi, rax   ; next r row
+
+    dec             rdx        ; counter
+    jnz .NextRow
+
+    SUM_ACROSS_W    xmm15
+    SUM_ACROSS_W    xmm14
+    SUM_ACROSS_Q    xmm13
+    SUM_ACROSS_Q    xmm12
+    SUM_ACROSS_Q    xmm11
+
+    mov             rdi,arg(4)
+    movd            [rdi], xmm15;
+    mov             rdi,arg(5)
+    movd            [rdi], xmm14;
+    mov             rdi,arg(6)
+    movd            [rdi], xmm13;
+    mov             rdi,arg(7)
+    movd            [rdi], xmm12;
+    mov             rdi,arg(8)
+    movd            [rdi], xmm11;
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+;void ssim_parms_sse2(
+;    unsigned char *s,
+;    int sp,
+;    unsigned char *r,
+;    int rp
+;    unsigned long *sum_s,
+;    unsigned long *sum_r,
+;    unsigned long *sum_sq_s,
+;    unsigned long *sum_sq_r,
+;    unsigned long *sum_sxr);
+;
+; TODO: Use parm passing through structure, probably don't need the pxors
+; ( calling app will initialize to 0 ) could easily fit everything in sse2
+; without too much hastle, and can probably do better estimates with psadw
+; or pavgb At this point this is just meant to be first pass for calculating
+; all the parms needed for 16x16 ssim so we can play with dssim as distortion
+; in mode selection code.
+global sym(vp9_ssim_parms_8x8_sse2) PRIVATE
+sym(vp9_ssim_parms_8x8_sse2):
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 9
+    SAVE_XMM 15
+    push        rsi
+    push        rdi
+    ; end prolog
+
+    mov             rsi,        arg(0) ;s
+    mov             rcx,        arg(1) ;sp
+    mov             rdi,        arg(2) ;r
+    mov             rax,        arg(3) ;rp
+
+    pxor            xmm0, xmm0
+    pxor            xmm15,xmm15  ;sum_s
+    pxor            xmm14,xmm14  ;sum_r
+    pxor            xmm13,xmm13  ;sum_sq_s
+    pxor            xmm12,xmm12  ;sum_sq_r
+    pxor            xmm11,xmm11  ;sum_sxr
+
+    mov             rdx, 8      ;row counter
+.NextRow:
+
+    ;grab source and reference pixels
+    movq            xmm3, [rsi]
+    movq            xmm4, [rdi]
+    punpcklbw       xmm3, xmm0 ; low_s
+    punpcklbw       xmm4, xmm0 ; low_r
+
+    TABULATE_SSIM
+
+    add             rsi, rcx   ; next s row
+    add             rdi, rax   ; next r row
+
+    dec             rdx        ; counter
+    jnz .NextRow
+
+    SUM_ACROSS_W    xmm15
+    SUM_ACROSS_W    xmm14
+    SUM_ACROSS_Q    xmm13
+    SUM_ACROSS_Q    xmm12
+    SUM_ACROSS_Q    xmm11
+
+    mov             rdi,arg(4)
+    movd            [rdi], xmm15;
+    mov             rdi,arg(5)
+    movd            [rdi], xmm14;
+    mov             rdi,arg(6)
+    movd            [rdi], xmm13;
+    mov             rdi,arg(7)
+    movd            [rdi], xmm12;
+    mov             rdi,arg(8)
+    movd            [rdi], xmm11;
+
+    ; begin epilog
+    pop         rdi
+    pop         rsi
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
diff --git a/media/libvpx/vp9/encoder/x86/vp9_subpel_variance.asm b/media/libvpx/vp9/encoder/x86/vp9_subpel_variance.asm
new file mode 100644
index 000000000..292cf34d1
--- /dev/null
+++ b/media/libvpx/vp9/encoder/x86/vp9_subpel_variance.asm
@@ -0,0 +1,1396 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION_RODATA
+pw_8: times  8 dw  8
+bilin_filter_m_sse2: times  8 dw 16
+                     times  8 dw  0
+                     times  8 dw 14
+                     times  8 dw  2
+                     times  8 dw 12
+                     times  8 dw  4
+                     times  8 dw 10
+                     times  8 dw  6
+                     times 16 dw  8
+                     times  8 dw  6
+                     times  8 dw 10
+                     times  8 dw  4
+                     times  8 dw 12
+                     times  8 dw  2
+                     times  8 dw 14
+
+bilin_filter_m_ssse3: times  8 db 16,  0
+                      times  8 db 14,  2
+                      times  8 db 12,  4
+                      times  8 db 10,  6
+                      times 16 db  8
+                      times  8 db  6, 10
+                      times  8 db  4, 12
+                      times  8 db  2, 14
+
+SECTION .text
+
+; int vp9_sub_pixel_varianceNxh(const uint8_t *src, ptrdiff_t src_stride,
+;                               int x_offset, int y_offset,
+;                               const uint8_t *dst, ptrdiff_t dst_stride,
+;                               int height, unsigned int *sse);
+;
+; This function returns the SE and stores SSE in the given pointer.
+
+%macro SUM_SSE 6 ; src1, dst1, src2, dst2, sum, sse
+  psubw                %3, %4
+  psubw                %1, %2
+  paddw                %5, %3
+  pmaddwd              %3, %3
+  paddw                %5, %1
+  pmaddwd              %1, %1
+  paddd                %6, %3
+  paddd                %6, %1
+%endmacro
+
+%macro STORE_AND_RET 0
+%if mmsize == 16
+  ; if H=64 and W=16, we have 8 words of each 2(1bit)x64(6bit)x9bit=16bit
+  ; in m6, i.e. it _exactly_ fits in a signed word per word in the xmm reg.
+  ; We have to sign-extend it before adding the words within the register
+  ; and outputing to a dword.
+  pcmpgtw              m5, m6           ; mask for 0 > x
+  movhlps              m3, m7
+  punpcklwd            m4, m6, m5
+  punpckhwd            m6, m5           ; sign-extend m6 word->dword
+  paddd                m7, m3
+  paddd                m6, m4
+  pshufd               m3, m7, 0x1
+  movhlps              m4, m6
+  paddd                m7, m3
+  paddd                m6, m4
+  mov                  r1, ssem         ; r1 = unsigned int *sse
+  pshufd               m4, m6, 0x1
+  movd               [r1], m7           ; store sse
+  paddd                m6, m4
+  movd               raxd, m6           ; store sum as return value
+%else ; mmsize == 8
+  pshufw               m4, m6, 0xe
+  pshufw               m3, m7, 0xe
+  paddw                m6, m4
+  paddd                m7, m3
+  pcmpgtw              m5, m6           ; mask for 0 > x
+  mov                  r1, ssem         ; r1 = unsigned int *sse
+  punpcklwd            m6, m5           ; sign-extend m6 word->dword
+  movd               [r1], m7           ; store sse
+  pshufw               m4, m6, 0xe
+  paddd                m6, m4
+  movd               raxd, m6           ; store sum as return value
+%endif
+  RET
+%endmacro
+
+%macro INC_SRC_BY_SRC_STRIDE  0
+%if ARCH_X86=1 && CONFIG_PIC=1
+  add                srcq, src_stridemp
+%else
+  add                srcq, src_strideq
+%endif
+%endmacro
+
+%macro SUBPEL_VARIANCE 1-2 0 ; W
+%if cpuflag(ssse3)
+%define bilin_filter_m bilin_filter_m_ssse3
+%define filter_idx_shift 4
+%else
+%define bilin_filter_m bilin_filter_m_sse2
+%define filter_idx_shift 5
+%endif
+; FIXME(rbultje) only bilinear filters use >8 registers, and ssse3 only uses
+; 11, not 13, if the registers are ordered correctly. May make a minor speed
+; difference on Win64
+
+%ifdef PIC    ; 64bit PIC
+  %if %2 == 1 ; avg
+    cglobal sub_pixel_avg_variance%1xh, 9, 10, 13, src, src_stride, \
+                                      x_offset, y_offset, \
+                                      dst, dst_stride, \
+                                      sec, sec_stride, height, sse
+    %define sec_str sec_strideq
+  %else
+    cglobal sub_pixel_variance%1xh, 7, 8, 13, src, src_stride, x_offset, \
+                                  y_offset, dst, dst_stride, height, sse
+  %endif
+  %define h heightd
+  %define bilin_filter sseq
+%else
+  %if ARCH_X86=1 && CONFIG_PIC=1
+    %if %2 == 1 ; avg
+      cglobal sub_pixel_avg_variance%1xh, 7, 7, 13, src, src_stride, \
+                                  x_offset, y_offset, \
+                                  dst, dst_stride, \
+                                  sec, sec_stride, \
+                                  height, sse, g_bilin_filter, g_pw_8
+      %define h dword heightm
+      %define sec_str sec_stridemp
+
+      ;Store bilin_filter and pw_8 location in stack
+      GET_GOT eax
+      add esp, 4                ; restore esp
+
+      lea ecx, [GLOBAL(bilin_filter_m)]
+      mov g_bilin_filterm, ecx
+
+      lea ecx, [GLOBAL(pw_8)]
+      mov g_pw_8m, ecx
+
+      LOAD_IF_USED 0, 1         ; load eax, ecx back
+    %else
+      cglobal sub_pixel_variance%1xh, 7, 7, 13, src, src_stride, x_offset, \
+                                y_offset, dst, dst_stride, height, sse, \
+                                g_bilin_filter, g_pw_8
+      %define h heightd
+
+      ;Store bilin_filter and pw_8 location in stack
+      GET_GOT eax
+      add esp, 4                ; restore esp
+
+      lea ecx, [GLOBAL(bilin_filter_m)]
+      mov g_bilin_filterm, ecx
+
+      lea ecx, [GLOBAL(pw_8)]
+      mov g_pw_8m, ecx
+
+      LOAD_IF_USED 0, 1         ; load eax, ecx back
+    %endif
+  %else
+    %if %2 == 1 ; avg
+      cglobal sub_pixel_avg_variance%1xh, 7 + 2 * ARCH_X86_64, \
+                        7 + 2 * ARCH_X86_64, 13, src, src_stride, \
+                                             x_offset, y_offset, \
+                                             dst, dst_stride, \
+                                             sec, sec_stride, \
+                                             height, sse
+      %if ARCH_X86_64
+      %define h heightd
+      %define sec_str sec_strideq
+      %else
+      %define h dword heightm
+      %define sec_str sec_stridemp
+      %endif
+    %else
+      cglobal sub_pixel_variance%1xh, 7, 7, 13, src, src_stride, x_offset, \
+                              y_offset, dst, dst_stride, height, sse
+      %define h heightd
+    %endif
+
+    %define bilin_filter bilin_filter_m
+  %endif
+%endif
+
+  ASSERT               %1 <= 16         ; m6 overflows if w > 16
+  pxor                 m6, m6           ; sum
+  pxor                 m7, m7           ; sse
+  ; FIXME(rbultje) if both filters are bilinear, we don't actually use m5; we
+  ; could perhaps use it for something more productive then
+  pxor                 m5, m5           ; dedicated zero register
+%if %1 < 16
+  sar                   h, 1
+%if %2 == 1 ; avg
+  shl             sec_str, 1
+%endif
+%endif
+
+  ; FIXME(rbultje) replace by jumptable?
+  test          x_offsetd, x_offsetd
+  jnz .x_nonzero
+  ; x_offset == 0
+  test          y_offsetd, y_offsetd
+  jnz .x_zero_y_nonzero
+
+  ; x_offset == 0 && y_offset == 0
+.x_zero_y_zero_loop:
+%if %1 == 16
+  movu                 m0, [srcq]
+  mova                 m1, [dstq]
+%if %2 == 1 ; avg
+  pavgb                m0, [secq]
+  punpckhbw            m3, m1, m5
+  punpcklbw            m1, m5
+%endif
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%if %2 == 0 ; !avg
+  punpckhbw            m3, m1, m5
+  punpcklbw            m1, m5
+%endif
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  add                srcq, src_strideq
+  add                dstq, dst_strideq
+%else ; %1 < 16
+  movh                 m0, [srcq]
+%if %2 == 1 ; avg
+%if mmsize == 16
+  movhps               m0, [srcq+src_strideq]
+%else ; mmsize == 8
+  punpckldq            m0, [srcq+src_strideq]
+%endif
+%else ; !avg
+  movh                 m2, [srcq+src_strideq]
+%endif
+  movh                 m1, [dstq]
+  movh                 m3, [dstq+dst_strideq]
+%if %2 == 1 ; avg
+  pavgb                m0, [secq]
+  punpcklbw            m3, m5
+  punpcklbw            m1, m5
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%else ; !avg
+  punpcklbw            m0, m5
+  punpcklbw            m2, m5
+  punpcklbw            m3, m5
+  punpcklbw            m1, m5
+%endif
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  lea                srcq, [srcq+src_strideq*2]
+  lea                dstq, [dstq+dst_strideq*2]
+%endif
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+  dec                   h
+  jg .x_zero_y_zero_loop
+  STORE_AND_RET
+
+.x_zero_y_nonzero:
+  cmp           y_offsetd, 8
+  jne .x_zero_y_nonhalf
+
+  ; x_offset == 0 && y_offset == 0.5
+.x_zero_y_half_loop:
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m4, [srcq+src_strideq]
+  mova                 m1, [dstq]
+  pavgb                m0, m4
+  punpckhbw            m3, m1, m5
+%if %2 == 1 ; avg
+  pavgb                m0, [secq]
+%endif
+  punpcklbw            m1, m5
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  add                srcq, src_strideq
+  add                dstq, dst_strideq
+%else ; %1 < 16
+  movh                 m0, [srcq]
+  movh                 m2, [srcq+src_strideq]
+%if %2 == 1 ; avg
+%if mmsize == 16
+  movhps               m2, [srcq+src_strideq*2]
+%else ; mmsize == 8
+%if %1 == 4
+  movh                 m1, [srcq+src_strideq*2]
+  punpckldq            m2, m1
+%else
+  punpckldq            m2, [srcq+src_strideq*2]
+%endif
+%endif
+  movh                 m1, [dstq]
+%if mmsize == 16
+  movlhps              m0, m2
+%else ; mmsize == 8
+  punpckldq            m0, m2
+%endif
+  movh                 m3, [dstq+dst_strideq]
+  pavgb                m0, m2
+  punpcklbw            m1, m5
+  pavgb                m0, [secq]
+  punpcklbw            m3, m5
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%else ; !avg
+  movh                 m4, [srcq+src_strideq*2]
+  movh                 m1, [dstq]
+  pavgb                m0, m2
+  movh                 m3, [dstq+dst_strideq]
+  pavgb                m2, m4
+  punpcklbw            m0, m5
+  punpcklbw            m2, m5
+  punpcklbw            m3, m5
+  punpcklbw            m1, m5
+%endif
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  lea                srcq, [srcq+src_strideq*2]
+  lea                dstq, [dstq+dst_strideq*2]
+%endif
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+  dec                   h
+  jg .x_zero_y_half_loop
+  STORE_AND_RET
+
+.x_zero_y_nonhalf:
+  ; x_offset == 0 && y_offset == bilin interpolation
+%ifdef PIC
+  lea        bilin_filter, [bilin_filter_m]
+%endif
+  shl           y_offsetd, filter_idx_shift
+%if ARCH_X86_64 && mmsize == 16
+  mova                 m8, [bilin_filter+y_offsetq]
+%if notcpuflag(ssse3) ; FIXME(rbultje) don't scatter registers on x86-64
+  mova                 m9, [bilin_filter+y_offsetq+16]
+%endif
+  mova                m10, [pw_8]
+%define filter_y_a m8
+%define filter_y_b m9
+%define filter_rnd m10
+%else ; x86-32 or mmx
+%if ARCH_X86=1 && CONFIG_PIC=1
+; x_offset == 0, reuse x_offset reg
+%define tempq x_offsetq
+  add y_offsetq, g_bilin_filterm
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+  mov tempq, g_pw_8m
+%define filter_rnd [tempq]
+%else
+  add           y_offsetq, bilin_filter
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+%define filter_rnd [pw_8]
+%endif
+%endif
+
+.x_zero_y_other_loop:
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m4, [srcq+src_strideq]
+  mova                 m1, [dstq]
+%if cpuflag(ssse3)
+  punpckhbw            m2, m0, m4
+  punpcklbw            m0, m4
+  pmaddubsw            m2, filter_y_a
+  pmaddubsw            m0, filter_y_a
+  paddw                m2, filter_rnd
+  paddw                m0, filter_rnd
+%else
+  punpckhbw            m2, m0, m5
+  punpckhbw            m3, m4, m5
+  punpcklbw            m0, m5
+  punpcklbw            m4, m5
+  ; FIXME(rbultje) instead of out=((num-x)*in1+x*in2+rnd)>>log2(num), we can
+  ; also do out=in1+(((num-x)*(in2-in1)+rnd)>>log2(num)). Total number of
+  ; instructions is the same (5), but it is 1 mul instead of 2, so might be
+  ; slightly faster because of pmullw latency. It would also cut our rodata
+  ; tables in half for this function, and save 1-2 registers on x86-64.
+  pmullw               m2, filter_y_a
+  pmullw               m3, filter_y_b
+  paddw                m2, filter_rnd
+  pmullw               m0, filter_y_a
+  pmullw               m4, filter_y_b
+  paddw                m0, filter_rnd
+  paddw                m2, m3
+  paddw                m0, m4
+%endif
+  psraw                m2, 4
+  psraw                m0, 4
+%if %2 == 1 ; avg
+  ; FIXME(rbultje) pipeline
+  packuswb             m0, m2
+  pavgb                m0, [secq]
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%endif
+  punpckhbw            m3, m1, m5
+  punpcklbw            m1, m5
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  add                srcq, src_strideq
+  add                dstq, dst_strideq
+%else ; %1 < 16
+  movh                 m0, [srcq]
+  movh                 m2, [srcq+src_strideq]
+  movh                 m4, [srcq+src_strideq*2]
+  movh                 m3, [dstq+dst_strideq]
+%if cpuflag(ssse3)
+  movh                 m1, [dstq]
+  punpcklbw            m0, m2
+  punpcklbw            m2, m4
+  pmaddubsw            m0, filter_y_a
+  pmaddubsw            m2, filter_y_a
+  punpcklbw            m3, m5
+  paddw                m2, filter_rnd
+  paddw                m0, filter_rnd
+%else
+  punpcklbw            m0, m5
+  punpcklbw            m2, m5
+  punpcklbw            m4, m5
+  pmullw               m0, filter_y_a
+  pmullw               m1, m2, filter_y_b
+  punpcklbw            m3, m5
+  paddw                m0, filter_rnd
+  pmullw               m2, filter_y_a
+  pmullw               m4, filter_y_b
+  paddw                m0, m1
+  paddw                m2, filter_rnd
+  movh                 m1, [dstq]
+  paddw                m2, m4
+%endif
+  psraw                m0, 4
+  psraw                m2, 4
+%if %2 == 1 ; avg
+  ; FIXME(rbultje) pipeline
+  packuswb             m0, m2
+  pavgb                m0, [secq]
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%endif
+  punpcklbw            m1, m5
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  lea                srcq, [srcq+src_strideq*2]
+  lea                dstq, [dstq+dst_strideq*2]
+%endif
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+  dec                   h
+  jg .x_zero_y_other_loop
+%undef filter_y_a
+%undef filter_y_b
+%undef filter_rnd
+  STORE_AND_RET
+
+.x_nonzero:
+  cmp           x_offsetd, 8
+  jne .x_nonhalf
+  ; x_offset == 0.5
+  test          y_offsetd, y_offsetd
+  jnz .x_half_y_nonzero
+
+  ; x_offset == 0.5 && y_offset == 0
+.x_half_y_zero_loop:
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m4, [srcq+1]
+  mova                 m1, [dstq]
+  pavgb                m0, m4
+  punpckhbw            m3, m1, m5
+%if %2 == 1 ; avg
+  pavgb                m0, [secq]
+%endif
+  punpcklbw            m1, m5
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  add                srcq, src_strideq
+  add                dstq, dst_strideq
+%else ; %1 < 16
+  movh                 m0, [srcq]
+  movh                 m4, [srcq+1]
+%if %2 == 1 ; avg
+%if mmsize == 16
+  movhps               m0, [srcq+src_strideq]
+  movhps               m4, [srcq+src_strideq+1]
+%else ; mmsize == 8
+  punpckldq            m0, [srcq+src_strideq]
+  punpckldq            m4, [srcq+src_strideq+1]
+%endif
+  movh                 m1, [dstq]
+  movh                 m3, [dstq+dst_strideq]
+  pavgb                m0, m4
+  punpcklbw            m3, m5
+  pavgb                m0, [secq]
+  punpcklbw            m1, m5
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%else ; !avg
+  movh                 m2, [srcq+src_strideq]
+  movh                 m1, [dstq]
+  pavgb                m0, m4
+  movh                 m4, [srcq+src_strideq+1]
+  movh                 m3, [dstq+dst_strideq]
+  pavgb                m2, m4
+  punpcklbw            m0, m5
+  punpcklbw            m2, m5
+  punpcklbw            m3, m5
+  punpcklbw            m1, m5
+%endif
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  lea                srcq, [srcq+src_strideq*2]
+  lea                dstq, [dstq+dst_strideq*2]
+%endif
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+  dec                   h
+  jg .x_half_y_zero_loop
+  STORE_AND_RET
+
+.x_half_y_nonzero:
+  cmp           y_offsetd, 8
+  jne .x_half_y_nonhalf
+
+  ; x_offset == 0.5 && y_offset == 0.5
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m3, [srcq+1]
+  add                srcq, src_strideq
+  pavgb                m0, m3
+.x_half_y_half_loop:
+  movu                 m4, [srcq]
+  movu                 m3, [srcq+1]
+  mova                 m1, [dstq]
+  pavgb                m4, m3
+  punpckhbw            m3, m1, m5
+  pavgb                m0, m4
+%if %2 == 1 ; avg
+  punpcklbw            m1, m5
+  pavgb                m0, [secq]
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%else
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+  punpcklbw            m1, m5
+%endif
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+  mova                 m0, m4
+
+  add                srcq, src_strideq
+  add                dstq, dst_strideq
+%else ; %1 < 16
+  movh                 m0, [srcq]
+  movh                 m3, [srcq+1]
+  add                srcq, src_strideq
+  pavgb                m0, m3
+.x_half_y_half_loop:
+  movh                 m2, [srcq]
+  movh                 m3, [srcq+1]
+%if %2 == 1 ; avg
+%if mmsize == 16
+  movhps               m2, [srcq+src_strideq]
+  movhps               m3, [srcq+src_strideq+1]
+%else
+%if %1 == 4
+  movh                 m1, [srcq+src_strideq]
+  punpckldq            m2, m1
+  movh                 m1, [srcq+src_strideq+1]
+  punpckldq            m3, m1
+%else
+  punpckldq            m2, [srcq+src_strideq]
+  punpckldq            m3, [srcq+src_strideq+1]
+%endif
+%endif
+  pavgb                m2, m3
+%if mmsize == 16
+  movlhps              m0, m2
+  movhlps              m4, m2
+%else ; mmsize == 8
+  punpckldq            m0, m2
+  pshufw               m4, m2, 0xe
+%endif
+  movh                 m1, [dstq]
+  pavgb                m0, m2
+  movh                 m3, [dstq+dst_strideq]
+  pavgb                m0, [secq]
+  punpcklbw            m3, m5
+  punpcklbw            m1, m5
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%else ; !avg
+  movh                 m4, [srcq+src_strideq]
+  movh                 m1, [srcq+src_strideq+1]
+  pavgb                m2, m3
+  pavgb                m4, m1
+  pavgb                m0, m2
+  pavgb                m2, m4
+  movh                 m1, [dstq]
+  movh                 m3, [dstq+dst_strideq]
+  punpcklbw            m0, m5
+  punpcklbw            m2, m5
+  punpcklbw            m3, m5
+  punpcklbw            m1, m5
+%endif
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+  mova                 m0, m4
+
+  lea                srcq, [srcq+src_strideq*2]
+  lea                dstq, [dstq+dst_strideq*2]
+%endif
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+  dec                   h
+  jg .x_half_y_half_loop
+  STORE_AND_RET
+
+.x_half_y_nonhalf:
+  ; x_offset == 0.5 && y_offset == bilin interpolation
+%ifdef PIC
+  lea        bilin_filter, [bilin_filter_m]
+%endif
+  shl           y_offsetd, filter_idx_shift
+%if ARCH_X86_64 && mmsize == 16
+  mova                 m8, [bilin_filter+y_offsetq]
+%if notcpuflag(ssse3) ; FIXME(rbultje) don't scatter registers on x86-64
+  mova                 m9, [bilin_filter+y_offsetq+16]
+%endif
+  mova                m10, [pw_8]
+%define filter_y_a m8
+%define filter_y_b m9
+%define filter_rnd m10
+%else  ;x86_32
+%if ARCH_X86=1 && CONFIG_PIC=1
+; x_offset == 0.5. We can reuse x_offset reg
+%define tempq x_offsetq
+  add y_offsetq, g_bilin_filterm
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+  mov tempq, g_pw_8m
+%define filter_rnd [tempq]
+%else
+  add           y_offsetq, bilin_filter
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+%define filter_rnd [pw_8]
+%endif
+%endif
+
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m3, [srcq+1]
+  add                srcq, src_strideq
+  pavgb                m0, m3
+.x_half_y_other_loop:
+  movu                 m4, [srcq]
+  movu                 m2, [srcq+1]
+  mova                 m1, [dstq]
+  pavgb                m4, m2
+%if cpuflag(ssse3)
+  punpckhbw            m2, m0, m4
+  punpcklbw            m0, m4
+  pmaddubsw            m2, filter_y_a
+  pmaddubsw            m0, filter_y_a
+  paddw                m2, filter_rnd
+  paddw                m0, filter_rnd
+  psraw                m2, 4
+%else
+  punpckhbw            m2, m0, m5
+  punpckhbw            m3, m4, m5
+  pmullw               m2, filter_y_a
+  pmullw               m3, filter_y_b
+  paddw                m2, filter_rnd
+  punpcklbw            m0, m5
+  paddw                m2, m3
+  punpcklbw            m3, m4, m5
+  pmullw               m0, filter_y_a
+  pmullw               m3, filter_y_b
+  paddw                m0, filter_rnd
+  psraw                m2, 4
+  paddw                m0, m3
+%endif
+  punpckhbw            m3, m1, m5
+  psraw                m0, 4
+%if %2 == 1 ; avg
+  ; FIXME(rbultje) pipeline
+  packuswb             m0, m2
+  pavgb                m0, [secq]
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%endif
+  punpcklbw            m1, m5
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+  mova                 m0, m4
+
+  add                srcq, src_strideq
+  add                dstq, dst_strideq
+%else ; %1 < 16
+  movh                 m0, [srcq]
+  movh                 m3, [srcq+1]
+  add                srcq, src_strideq
+  pavgb                m0, m3
+%if notcpuflag(ssse3)
+  punpcklbw            m0, m5
+%endif
+.x_half_y_other_loop:
+  movh                 m2, [srcq]
+  movh                 m1, [srcq+1]
+  movh                 m4, [srcq+src_strideq]
+  movh                 m3, [srcq+src_strideq+1]
+  pavgb                m2, m1
+  pavgb                m4, m3
+  movh                 m3, [dstq+dst_strideq]
+%if cpuflag(ssse3)
+  movh                 m1, [dstq]
+  punpcklbw            m0, m2
+  punpcklbw            m2, m4
+  pmaddubsw            m0, filter_y_a
+  pmaddubsw            m2, filter_y_a
+  punpcklbw            m3, m5
+  paddw                m0, filter_rnd
+  paddw                m2, filter_rnd
+%else
+  punpcklbw            m2, m5
+  punpcklbw            m4, m5
+  pmullw               m0, filter_y_a
+  pmullw               m1, m2, filter_y_b
+  punpcklbw            m3, m5
+  paddw                m0, filter_rnd
+  pmullw               m2, filter_y_a
+  paddw                m0, m1
+  pmullw               m1, m4, filter_y_b
+  paddw                m2, filter_rnd
+  paddw                m2, m1
+  movh                 m1, [dstq]
+%endif
+  psraw                m0, 4
+  psraw                m2, 4
+%if %2 == 1 ; avg
+  ; FIXME(rbultje) pipeline
+  packuswb             m0, m2
+  pavgb                m0, [secq]
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%endif
+  punpcklbw            m1, m5
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+  mova                 m0, m4
+
+  lea                srcq, [srcq+src_strideq*2]
+  lea                dstq, [dstq+dst_strideq*2]
+%endif
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+  dec                   h
+  jg .x_half_y_other_loop
+%undef filter_y_a
+%undef filter_y_b
+%undef filter_rnd
+  STORE_AND_RET
+
+.x_nonhalf:
+  test          y_offsetd, y_offsetd
+  jnz .x_nonhalf_y_nonzero
+
+  ; x_offset == bilin interpolation && y_offset == 0
+%ifdef PIC
+  lea        bilin_filter, [bilin_filter_m]
+%endif
+  shl           x_offsetd, filter_idx_shift
+%if ARCH_X86_64 && mmsize == 16
+  mova                 m8, [bilin_filter+x_offsetq]
+%if notcpuflag(ssse3) ; FIXME(rbultje) don't scatter registers on x86-64
+  mova                 m9, [bilin_filter+x_offsetq+16]
+%endif
+  mova                m10, [pw_8]
+%define filter_x_a m8
+%define filter_x_b m9
+%define filter_rnd m10
+%else    ; x86-32
+%if ARCH_X86=1 && CONFIG_PIC=1
+;y_offset == 0. We can reuse y_offset reg.
+%define tempq y_offsetq
+  add x_offsetq, g_bilin_filterm
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+  mov tempq, g_pw_8m
+%define filter_rnd [tempq]
+%else
+  add           x_offsetq, bilin_filter
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+%define filter_rnd [pw_8]
+%endif
+%endif
+
+.x_other_y_zero_loop:
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m4, [srcq+1]
+  mova                 m1, [dstq]
+%if cpuflag(ssse3)
+  punpckhbw            m2, m0, m4
+  punpcklbw            m0, m4
+  pmaddubsw            m2, filter_x_a
+  pmaddubsw            m0, filter_x_a
+  paddw                m2, filter_rnd
+  paddw                m0, filter_rnd
+%else
+  punpckhbw            m2, m0, m5
+  punpckhbw            m3, m4, m5
+  punpcklbw            m0, m5
+  punpcklbw            m4, m5
+  pmullw               m2, filter_x_a
+  pmullw               m3, filter_x_b
+  paddw                m2, filter_rnd
+  pmullw               m0, filter_x_a
+  pmullw               m4, filter_x_b
+  paddw                m0, filter_rnd
+  paddw                m2, m3
+  paddw                m0, m4
+%endif
+  psraw                m2, 4
+  psraw                m0, 4
+%if %2 == 1 ; avg
+  ; FIXME(rbultje) pipeline
+  packuswb             m0, m2
+  pavgb                m0, [secq]
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%endif
+  punpckhbw            m3, m1, m5
+  punpcklbw            m1, m5
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  add                srcq, src_strideq
+  add                dstq, dst_strideq
+%else ; %1 < 16
+  movh                 m0, [srcq]
+  movh                 m1, [srcq+1]
+  movh                 m2, [srcq+src_strideq]
+  movh                 m4, [srcq+src_strideq+1]
+  movh                 m3, [dstq+dst_strideq]
+%if cpuflag(ssse3)
+  punpcklbw            m0, m1
+  movh                 m1, [dstq]
+  punpcklbw            m2, m4
+  pmaddubsw            m0, filter_x_a
+  pmaddubsw            m2, filter_x_a
+  punpcklbw            m3, m5
+  paddw                m0, filter_rnd
+  paddw                m2, filter_rnd
+%else
+  punpcklbw            m0, m5
+  punpcklbw            m1, m5
+  punpcklbw            m2, m5
+  punpcklbw            m4, m5
+  pmullw               m0, filter_x_a
+  pmullw               m1, filter_x_b
+  punpcklbw            m3, m5
+  paddw                m0, filter_rnd
+  pmullw               m2, filter_x_a
+  pmullw               m4, filter_x_b
+  paddw                m0, m1
+  paddw                m2, filter_rnd
+  movh                 m1, [dstq]
+  paddw                m2, m4
+%endif
+  psraw                m0, 4
+  psraw                m2, 4
+%if %2 == 1 ; avg
+  ; FIXME(rbultje) pipeline
+  packuswb             m0, m2
+  pavgb                m0, [secq]
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%endif
+  punpcklbw            m1, m5
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+
+  lea                srcq, [srcq+src_strideq*2]
+  lea                dstq, [dstq+dst_strideq*2]
+%endif
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+  dec                   h
+  jg .x_other_y_zero_loop
+%undef filter_x_a
+%undef filter_x_b
+%undef filter_rnd
+  STORE_AND_RET
+
+.x_nonhalf_y_nonzero:
+  cmp           y_offsetd, 8
+  jne .x_nonhalf_y_nonhalf
+
+  ; x_offset == bilin interpolation && y_offset == 0.5
+%ifdef PIC
+  lea        bilin_filter, [bilin_filter_m]
+%endif
+  shl           x_offsetd, filter_idx_shift
+%if ARCH_X86_64 && mmsize == 16
+  mova                 m8, [bilin_filter+x_offsetq]
+%if notcpuflag(ssse3) ; FIXME(rbultje) don't scatter registers on x86-64
+  mova                 m9, [bilin_filter+x_offsetq+16]
+%endif
+  mova                m10, [pw_8]
+%define filter_x_a m8
+%define filter_x_b m9
+%define filter_rnd m10
+%else    ; x86-32
+%if ARCH_X86=1 && CONFIG_PIC=1
+; y_offset == 0.5. We can reuse y_offset reg.
+%define tempq y_offsetq
+  add x_offsetq, g_bilin_filterm
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+  mov tempq, g_pw_8m
+%define filter_rnd [tempq]
+%else
+  add           x_offsetq, bilin_filter
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+%define filter_rnd [pw_8]
+%endif
+%endif
+
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq+1]
+%if cpuflag(ssse3)
+  punpckhbw            m2, m0, m1
+  punpcklbw            m0, m1
+  pmaddubsw            m2, filter_x_a
+  pmaddubsw            m0, filter_x_a
+  paddw                m2, filter_rnd
+  paddw                m0, filter_rnd
+%else
+  punpckhbw            m2, m0, m5
+  punpckhbw            m3, m1, m5
+  punpcklbw            m0, m5
+  punpcklbw            m1, m5
+  pmullw               m0, filter_x_a
+  pmullw               m1, filter_x_b
+  paddw                m0, filter_rnd
+  pmullw               m2, filter_x_a
+  pmullw               m3, filter_x_b
+  paddw                m2, filter_rnd
+  paddw                m0, m1
+  paddw                m2, m3
+%endif
+  psraw                m0, 4
+  psraw                m2, 4
+  add                srcq, src_strideq
+  packuswb             m0, m2
+.x_other_y_half_loop:
+  movu                 m4, [srcq]
+  movu                 m3, [srcq+1]
+%if cpuflag(ssse3)
+  mova                 m1, [dstq]
+  punpckhbw            m2, m4, m3
+  punpcklbw            m4, m3
+  pmaddubsw            m2, filter_x_a
+  pmaddubsw            m4, filter_x_a
+  paddw                m2, filter_rnd
+  paddw                m4, filter_rnd
+  psraw                m2, 4
+  psraw                m4, 4
+  packuswb             m4, m2
+  pavgb                m0, m4
+  punpckhbw            m3, m1, m5
+  punpcklbw            m1, m5
+%else
+  punpckhbw            m2, m4, m5
+  punpckhbw            m1, m3, m5
+  punpcklbw            m4, m5
+  punpcklbw            m3, m5
+  pmullw               m4, filter_x_a
+  pmullw               m3, filter_x_b
+  paddw                m4, filter_rnd
+  pmullw               m2, filter_x_a
+  pmullw               m1, filter_x_b
+  paddw                m2, filter_rnd
+  paddw                m4, m3
+  paddw                m2, m1
+  mova                 m1, [dstq]
+  psraw                m4, 4
+  psraw                m2, 4
+  punpckhbw            m3, m1, m5
+  ; FIXME(rbultje) the repeated pack/unpack here around m0/m2 is because we
+  ; have a 1-register shortage to be able to store the backup of the bilin
+  ; filtered second line as words as cache for the next line. Packing into
+  ; a byte costs 1 pack and 2 unpacks, but saves a register.
+  packuswb             m4, m2
+  punpcklbw            m1, m5
+  pavgb                m0, m4
+%endif
+%if %2 == 1 ; avg
+  ; FIXME(rbultje) pipeline
+  pavgb                m0, [secq]
+%endif
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+  mova                 m0, m4
+
+  add                srcq, src_strideq
+  add                dstq, dst_strideq
+%else ; %1 < 16
+  movh                 m0, [srcq]
+  movh                 m1, [srcq+1]
+%if cpuflag(ssse3)
+  punpcklbw            m0, m1
+  pmaddubsw            m0, filter_x_a
+  paddw                m0, filter_rnd
+%else
+  punpcklbw            m0, m5
+  punpcklbw            m1, m5
+  pmullw               m0, filter_x_a
+  pmullw               m1, filter_x_b
+  paddw                m0, filter_rnd
+  paddw                m0, m1
+%endif
+  add                srcq, src_strideq
+  psraw                m0, 4
+.x_other_y_half_loop:
+  movh                 m2, [srcq]
+  movh                 m1, [srcq+1]
+  movh                 m4, [srcq+src_strideq]
+  movh                 m3, [srcq+src_strideq+1]
+%if cpuflag(ssse3)
+  punpcklbw            m2, m1
+  punpcklbw            m4, m3
+  pmaddubsw            m2, filter_x_a
+  pmaddubsw            m4, filter_x_a
+  movh                 m1, [dstq]
+  movh                 m3, [dstq+dst_strideq]
+  paddw                m2, filter_rnd
+  paddw                m4, filter_rnd
+%else
+  punpcklbw            m2, m5
+  punpcklbw            m1, m5
+  punpcklbw            m4, m5
+  punpcklbw            m3, m5
+  pmullw               m2, filter_x_a
+  pmullw               m1, filter_x_b
+  paddw                m2, filter_rnd
+  pmullw               m4, filter_x_a
+  pmullw               m3, filter_x_b
+  paddw                m4, filter_rnd
+  paddw                m2, m1
+  movh                 m1, [dstq]
+  paddw                m4, m3
+  movh                 m3, [dstq+dst_strideq]
+%endif
+  psraw                m2, 4
+  psraw                m4, 4
+  pavgw                m0, m2
+  pavgw                m2, m4
+%if %2 == 1 ; avg
+  ; FIXME(rbultje) pipeline - also consider going to bytes here
+  packuswb             m0, m2
+  pavgb                m0, [secq]
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%endif
+  punpcklbw            m3, m5
+  punpcklbw            m1, m5
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+  mova                 m0, m4
+
+  lea                srcq, [srcq+src_strideq*2]
+  lea                dstq, [dstq+dst_strideq*2]
+%endif
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+  dec                   h
+  jg .x_other_y_half_loop
+%undef filter_x_a
+%undef filter_x_b
+%undef filter_rnd
+  STORE_AND_RET
+
+.x_nonhalf_y_nonhalf:
+%ifdef PIC
+  lea        bilin_filter, [bilin_filter_m]
+%endif
+  shl           x_offsetd, filter_idx_shift
+  shl           y_offsetd, filter_idx_shift
+%if ARCH_X86_64 && mmsize == 16
+  mova                 m8, [bilin_filter+x_offsetq]
+%if notcpuflag(ssse3) ; FIXME(rbultje) don't scatter registers on x86-64
+  mova                 m9, [bilin_filter+x_offsetq+16]
+%endif
+  mova                m10, [bilin_filter+y_offsetq]
+%if notcpuflag(ssse3) ; FIXME(rbultje) don't scatter registers on x86-64
+  mova                m11, [bilin_filter+y_offsetq+16]
+%endif
+  mova                m12, [pw_8]
+%define filter_x_a m8
+%define filter_x_b m9
+%define filter_y_a m10
+%define filter_y_b m11
+%define filter_rnd m12
+%else   ; x86-32
+%if ARCH_X86=1 && CONFIG_PIC=1
+; In this case, there is NO unused register. Used src_stride register. Later,
+; src_stride has to be loaded from stack when it is needed.
+%define tempq src_strideq
+  mov tempq, g_bilin_filterm
+  add           x_offsetq, tempq
+  add           y_offsetq, tempq
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+
+  mov tempq, g_pw_8m
+%define filter_rnd [tempq]
+%else
+  add           x_offsetq, bilin_filter
+  add           y_offsetq, bilin_filter
+%define filter_x_a [x_offsetq]
+%define filter_x_b [x_offsetq+16]
+%define filter_y_a [y_offsetq]
+%define filter_y_b [y_offsetq+16]
+%define filter_rnd [pw_8]
+%endif
+%endif
+
+  ; x_offset == bilin interpolation && y_offset == bilin interpolation
+%if %1 == 16
+  movu                 m0, [srcq]
+  movu                 m1, [srcq+1]
+%if cpuflag(ssse3)
+  punpckhbw            m2, m0, m1
+  punpcklbw            m0, m1
+  pmaddubsw            m2, filter_x_a
+  pmaddubsw            m0, filter_x_a
+  paddw                m2, filter_rnd
+  paddw                m0, filter_rnd
+%else
+  punpckhbw            m2, m0, m5
+  punpckhbw            m3, m1, m5
+  punpcklbw            m0, m5
+  punpcklbw            m1, m5
+  pmullw               m0, filter_x_a
+  pmullw               m1, filter_x_b
+  paddw                m0, filter_rnd
+  pmullw               m2, filter_x_a
+  pmullw               m3, filter_x_b
+  paddw                m2, filter_rnd
+  paddw                m0, m1
+  paddw                m2, m3
+%endif
+  psraw                m0, 4
+  psraw                m2, 4
+
+  INC_SRC_BY_SRC_STRIDE
+
+  packuswb             m0, m2
+.x_other_y_other_loop:
+%if cpuflag(ssse3)
+  movu                 m4, [srcq]
+  movu                 m3, [srcq+1]
+  mova                 m1, [dstq]
+  punpckhbw            m2, m4, m3
+  punpcklbw            m4, m3
+  pmaddubsw            m2, filter_x_a
+  pmaddubsw            m4, filter_x_a
+  punpckhbw            m3, m1, m5
+  paddw                m2, filter_rnd
+  paddw                m4, filter_rnd
+  psraw                m2, 4
+  psraw                m4, 4
+  packuswb             m4, m2
+  punpckhbw            m2, m0, m4
+  punpcklbw            m0, m4
+  pmaddubsw            m2, filter_y_a
+  pmaddubsw            m0, filter_y_a
+  punpcklbw            m1, m5
+  paddw                m2, filter_rnd
+  paddw                m0, filter_rnd
+  psraw                m2, 4
+  psraw                m0, 4
+%else
+  movu                 m3, [srcq]
+  movu                 m4, [srcq+1]
+  punpckhbw            m1, m3, m5
+  punpckhbw            m2, m4, m5
+  punpcklbw            m3, m5
+  punpcklbw            m4, m5
+  pmullw               m3, filter_x_a
+  pmullw               m4, filter_x_b
+  paddw                m3, filter_rnd
+  pmullw               m1, filter_x_a
+  pmullw               m2, filter_x_b
+  paddw                m1, filter_rnd
+  paddw                m3, m4
+  paddw                m1, m2
+  psraw                m3, 4
+  psraw                m1, 4
+  packuswb             m4, m3, m1
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+  pmullw               m2, filter_y_a
+  pmullw               m1, filter_y_b
+  paddw                m2, filter_rnd
+  pmullw               m0, filter_y_a
+  pmullw               m3, filter_y_b
+  paddw                m2, m1
+  mova                 m1, [dstq]
+  paddw                m0, filter_rnd
+  psraw                m2, 4
+  paddw                m0, m3
+  punpckhbw            m3, m1, m5
+  psraw                m0, 4
+  punpcklbw            m1, m5
+%endif
+%if %2 == 1 ; avg
+  ; FIXME(rbultje) pipeline
+  packuswb             m0, m2
+  pavgb                m0, [secq]
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%endif
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+  mova                 m0, m4
+
+  INC_SRC_BY_SRC_STRIDE
+  add                dstq, dst_strideq
+%else ; %1 < 16
+  movh                 m0, [srcq]
+  movh                 m1, [srcq+1]
+%if cpuflag(ssse3)
+  punpcklbw            m0, m1
+  pmaddubsw            m0, filter_x_a
+  paddw                m0, filter_rnd
+%else
+  punpcklbw            m0, m5
+  punpcklbw            m1, m5
+  pmullw               m0, filter_x_a
+  pmullw               m1, filter_x_b
+  paddw                m0, filter_rnd
+  paddw                m0, m1
+%endif
+  psraw                m0, 4
+%if cpuflag(ssse3)
+  packuswb             m0, m0
+%endif
+
+  INC_SRC_BY_SRC_STRIDE
+
+.x_other_y_other_loop:
+  movh                 m2, [srcq]
+  movh                 m1, [srcq+1]
+
+  INC_SRC_BY_SRC_STRIDE
+  movh                 m4, [srcq]
+  movh                 m3, [srcq+1]
+
+%if cpuflag(ssse3)
+  punpcklbw            m2, m1
+  punpcklbw            m4, m3
+  pmaddubsw            m2, filter_x_a
+  pmaddubsw            m4, filter_x_a
+  movh                 m3, [dstq+dst_strideq]
+  movh                 m1, [dstq]
+  paddw                m2, filter_rnd
+  paddw                m4, filter_rnd
+  psraw                m2, 4
+  psraw                m4, 4
+  packuswb             m2, m2
+  packuswb             m4, m4
+  punpcklbw            m0, m2
+  punpcklbw            m2, m4
+  pmaddubsw            m0, filter_y_a
+  pmaddubsw            m2, filter_y_a
+  punpcklbw            m3, m5
+  paddw                m0, filter_rnd
+  paddw                m2, filter_rnd
+  psraw                m0, 4
+  psraw                m2, 4
+  punpcklbw            m1, m5
+%else
+  punpcklbw            m2, m5
+  punpcklbw            m1, m5
+  punpcklbw            m4, m5
+  punpcklbw            m3, m5
+  pmullw               m2, filter_x_a
+  pmullw               m1, filter_x_b
+  paddw                m2, filter_rnd
+  pmullw               m4, filter_x_a
+  pmullw               m3, filter_x_b
+  paddw                m4, filter_rnd
+  paddw                m2, m1
+  paddw                m4, m3
+  psraw                m2, 4
+  psraw                m4, 4
+  pmullw               m0, filter_y_a
+  pmullw               m3, m2, filter_y_b
+  paddw                m0, filter_rnd
+  pmullw               m2, filter_y_a
+  pmullw               m1, m4, filter_y_b
+  paddw                m2, filter_rnd
+  paddw                m0, m3
+  movh                 m3, [dstq+dst_strideq]
+  paddw                m2, m1
+  movh                 m1, [dstq]
+  psraw                m0, 4
+  psraw                m2, 4
+  punpcklbw            m3, m5
+  punpcklbw            m1, m5
+%endif
+%if %2 == 1 ; avg
+  ; FIXME(rbultje) pipeline
+  packuswb             m0, m2
+  pavgb                m0, [secq]
+  punpckhbw            m2, m0, m5
+  punpcklbw            m0, m5
+%endif
+  SUM_SSE              m0, m1, m2, m3, m6, m7
+  mova                 m0, m4
+
+  INC_SRC_BY_SRC_STRIDE
+  lea                dstq, [dstq+dst_strideq*2]
+%endif
+%if %2 == 1 ; avg
+  add                secq, sec_str
+%endif
+  dec                   h
+  jg .x_other_y_other_loop
+%undef filter_x_a
+%undef filter_x_b
+%undef filter_y_a
+%undef filter_y_b
+%undef filter_rnd
+  STORE_AND_RET
+%endmacro
+
+; FIXME(rbultje) the non-bilinear versions (i.e. x=0,8&&y=0,8) are identical
+; between the ssse3 and non-ssse3 version. It may make sense to merge their
+; code in the sense that the ssse3 version would jump to the appropriate
+; location in the sse/2 version, rather than duplicating that code in the
+; binary.
+
+INIT_MMX sse
+SUBPEL_VARIANCE  4
+INIT_XMM sse2
+SUBPEL_VARIANCE  8
+SUBPEL_VARIANCE 16
+
+INIT_MMX ssse3
+SUBPEL_VARIANCE  4
+INIT_XMM ssse3
+SUBPEL_VARIANCE  8
+SUBPEL_VARIANCE 16
+
+INIT_MMX sse
+SUBPEL_VARIANCE  4, 1
+INIT_XMM sse2
+SUBPEL_VARIANCE  8, 1
+SUBPEL_VARIANCE 16, 1
+
+INIT_MMX ssse3
+SUBPEL_VARIANCE  4, 1
+INIT_XMM ssse3
+SUBPEL_VARIANCE  8, 1
+SUBPEL_VARIANCE 16, 1
diff --git a/media/libvpx/vp9/encoder/x86/vp9_subpel_variance_impl_intrin_avx2.c b/media/libvpx/vp9/encoder/x86/vp9_subpel_variance_impl_intrin_avx2.c
new file mode 100644
index 000000000..b1c797520
--- /dev/null
+++ b/media/libvpx/vp9/encoder/x86/vp9_subpel_variance_impl_intrin_avx2.c
@@ -0,0 +1,525 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <immintrin.h>  // AVX2
+
+#include "./vp9_rtcd.h"
+#include "vpx_ports/mem.h"
+#include "vp9/encoder/vp9_variance.h"
+
+DECLARE_ALIGNED(32, static const uint8_t, bilinear_filters_avx2[512]) = {
+  16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0,
+  16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0, 16, 0,
+  14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2,
+  14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2,
+  12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4,
+  12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4,
+  10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6,
+  10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6,
+  8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+  8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8, 8,
+  6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10,
+  6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10, 6, 10,
+  4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12,
+  4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12, 4, 12,
+  2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14,
+  2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14, 2, 14,
+};
+
+#define FILTER_SRC(filter) \
+  /* filter the source */ \
+  exp_src_lo = _mm256_maddubs_epi16(exp_src_lo, filter); \
+  exp_src_hi = _mm256_maddubs_epi16(exp_src_hi, filter); \
+  \
+  /* add 8 to source */ \
+  exp_src_lo = _mm256_add_epi16(exp_src_lo, pw8); \
+  exp_src_hi = _mm256_add_epi16(exp_src_hi, pw8); \
+  \
+  /* divide source by 16 */ \
+  exp_src_lo = _mm256_srai_epi16(exp_src_lo, 4); \
+  exp_src_hi = _mm256_srai_epi16(exp_src_hi, 4);
+
+#define MERGE_WITH_SRC(src_reg, reg) \
+  exp_src_lo = _mm256_unpacklo_epi8(src_reg, reg); \
+  exp_src_hi = _mm256_unpackhi_epi8(src_reg, reg);
+
+#define LOAD_SRC_DST \
+  /* load source and destination */ \
+  src_reg = _mm256_loadu_si256((__m256i const *) (src)); \
+  dst_reg = _mm256_loadu_si256((__m256i const *) (dst));
+
+#define AVG_NEXT_SRC(src_reg, size_stride) \
+  src_next_reg = _mm256_loadu_si256((__m256i const *) \
+                                   (src + size_stride)); \
+  /* average between current and next stride source */ \
+  src_reg = _mm256_avg_epu8(src_reg, src_next_reg);
+
+#define MERGE_NEXT_SRC(src_reg, size_stride) \
+  src_next_reg = _mm256_loadu_si256((__m256i const *) \
+                                   (src + size_stride)); \
+  MERGE_WITH_SRC(src_reg, src_next_reg)
+
+#define CALC_SUM_SSE_INSIDE_LOOP \
+  /* expand each byte to 2 bytes */ \
+  exp_dst_lo = _mm256_unpacklo_epi8(dst_reg, zero_reg); \
+  exp_dst_hi = _mm256_unpackhi_epi8(dst_reg, zero_reg); \
+  /* source - dest */ \
+  exp_src_lo = _mm256_sub_epi16(exp_src_lo, exp_dst_lo); \
+  exp_src_hi = _mm256_sub_epi16(exp_src_hi, exp_dst_hi); \
+  /* caculate sum */ \
+  sum_reg = _mm256_add_epi16(sum_reg, exp_src_lo); \
+  exp_src_lo = _mm256_madd_epi16(exp_src_lo, exp_src_lo); \
+  sum_reg = _mm256_add_epi16(sum_reg, exp_src_hi); \
+  exp_src_hi = _mm256_madd_epi16(exp_src_hi, exp_src_hi); \
+  /* calculate sse */ \
+  sse_reg = _mm256_add_epi32(sse_reg, exp_src_lo); \
+  sse_reg = _mm256_add_epi32(sse_reg, exp_src_hi);
+
+// final calculation to sum and sse
+#define CALC_SUM_AND_SSE \
+  res_cmp = _mm256_cmpgt_epi16(zero_reg, sum_reg); \
+  sse_reg_hi = _mm256_srli_si256(sse_reg, 8); \
+  sum_reg_lo = _mm256_unpacklo_epi16(sum_reg, res_cmp); \
+  sum_reg_hi = _mm256_unpackhi_epi16(sum_reg, res_cmp); \
+  sse_reg = _mm256_add_epi32(sse_reg, sse_reg_hi); \
+  sum_reg = _mm256_add_epi32(sum_reg_lo, sum_reg_hi); \
+  \
+  sse_reg_hi = _mm256_srli_si256(sse_reg, 4); \
+  sum_reg_hi = _mm256_srli_si256(sum_reg, 8); \
+  \
+  sse_reg = _mm256_add_epi32(sse_reg, sse_reg_hi); \
+  sum_reg = _mm256_add_epi32(sum_reg, sum_reg_hi); \
+  *((int*)sse)= _mm_cvtsi128_si32(_mm256_castsi256_si128(sse_reg)) + \
+                _mm_cvtsi128_si32(_mm256_extractf128_si256(sse_reg, 1)); \
+  sum_reg_hi = _mm256_srli_si256(sum_reg, 4); \
+  sum_reg = _mm256_add_epi32(sum_reg, sum_reg_hi); \
+  sum = _mm_cvtsi128_si32(_mm256_castsi256_si128(sum_reg)) + \
+        _mm_cvtsi128_si32(_mm256_extractf128_si256(sum_reg, 1));
+
+
+unsigned int vp9_sub_pixel_variance32xh_avx2(const uint8_t *src,
+                                             int src_stride,
+                                             int x_offset,
+                                             int y_offset,
+                                             const uint8_t *dst,
+                                             int dst_stride,
+                                             int height,
+                                             unsigned int *sse) {
+  __m256i src_reg, dst_reg, exp_src_lo, exp_src_hi, exp_dst_lo, exp_dst_hi;
+  __m256i sse_reg, sum_reg, sse_reg_hi, res_cmp, sum_reg_lo, sum_reg_hi;
+  __m256i zero_reg;
+  int i, sum;
+  sum_reg = _mm256_set1_epi16(0);
+  sse_reg = _mm256_set1_epi16(0);
+  zero_reg = _mm256_set1_epi16(0);
+
+  // x_offset = 0 and y_offset = 0
+  if (x_offset == 0) {
+    if (y_offset == 0) {
+      for (i = 0; i < height ; i++) {
+        LOAD_SRC_DST
+        // expend each byte to 2 bytes
+        MERGE_WITH_SRC(src_reg, zero_reg)
+        CALC_SUM_SSE_INSIDE_LOOP
+        src+= src_stride;
+        dst+= dst_stride;
+      }
+    // x_offset = 0 and y_offset = 8
+    } else if (y_offset == 8) {
+      __m256i src_next_reg;
+      for (i = 0; i < height ; i++) {
+        LOAD_SRC_DST
+        AVG_NEXT_SRC(src_reg, src_stride)
+        // expend each byte to 2 bytes
+        MERGE_WITH_SRC(src_reg, zero_reg)
+        CALC_SUM_SSE_INSIDE_LOOP
+        src+= src_stride;
+        dst+= dst_stride;
+      }
+    // x_offset = 0 and y_offset = bilin interpolation
+    } else {
+      __m256i filter, pw8, src_next_reg;
+
+      y_offset <<= 5;
+      filter = _mm256_load_si256((__m256i const *)
+               (bilinear_filters_avx2 + y_offset));
+      pw8 = _mm256_set1_epi16(8);
+      for (i = 0; i < height ; i++) {
+        LOAD_SRC_DST
+        MERGE_NEXT_SRC(src_reg, src_stride)
+        FILTER_SRC(filter)
+        CALC_SUM_SSE_INSIDE_LOOP
+        src+= src_stride;
+        dst+= dst_stride;
+      }
+    }
+  // x_offset = 8  and y_offset = 0
+  } else if (x_offset == 8) {
+    if (y_offset == 0) {
+      __m256i src_next_reg;
+      for (i = 0; i < height ; i++) {
+        LOAD_SRC_DST
+        AVG_NEXT_SRC(src_reg, 1)
+        // expand each byte to 2 bytes
+        MERGE_WITH_SRC(src_reg, zero_reg)
+        CALC_SUM_SSE_INSIDE_LOOP
+        src+= src_stride;
+        dst+= dst_stride;
+      }
+    // x_offset = 8  and y_offset = 8
+    } else if (y_offset == 8) {
+      __m256i src_next_reg, src_avg;
+      // load source and another source starting from the next
+      // following byte
+      src_reg = _mm256_loadu_si256((__m256i const *) (src));
+      AVG_NEXT_SRC(src_reg, 1)
+      for (i = 0; i < height ; i++) {
+        src_avg = src_reg;
+        src+= src_stride;
+        LOAD_SRC_DST
+        AVG_NEXT_SRC(src_reg, 1)
+        // average between previous average to current average
+        src_avg = _mm256_avg_epu8(src_avg, src_reg);
+        // expand each byte to 2 bytes
+        MERGE_WITH_SRC(src_avg, zero_reg)
+        // save current source average
+        CALC_SUM_SSE_INSIDE_LOOP
+        dst+= dst_stride;
+      }
+    // x_offset = 8  and y_offset = bilin interpolation
+    } else {
+      __m256i filter, pw8, src_next_reg, src_avg;
+      y_offset <<= 5;
+      filter = _mm256_load_si256((__m256i const *)
+               (bilinear_filters_avx2 + y_offset));
+      pw8 = _mm256_set1_epi16(8);
+      // load source and another source starting from the next
+      // following byte
+      src_reg = _mm256_loadu_si256((__m256i const *) (src));
+      AVG_NEXT_SRC(src_reg, 1)
+      for (i = 0; i < height ; i++) {
+        // save current source average
+        src_avg = src_reg;
+        src+= src_stride;
+        LOAD_SRC_DST
+        AVG_NEXT_SRC(src_reg, 1)
+        MERGE_WITH_SRC(src_avg, src_reg)
+        FILTER_SRC(filter)
+        CALC_SUM_SSE_INSIDE_LOOP
+        dst+= dst_stride;
+      }
+    }
+  // x_offset = bilin interpolation and y_offset = 0
+  } else {
+    if (y_offset == 0) {
+      __m256i filter, pw8, src_next_reg;
+      x_offset <<= 5;
+      filter = _mm256_load_si256((__m256i const *)
+               (bilinear_filters_avx2 + x_offset));
+      pw8 = _mm256_set1_epi16(8);
+      for (i = 0; i < height ; i++) {
+        LOAD_SRC_DST
+        MERGE_NEXT_SRC(src_reg, 1)
+        FILTER_SRC(filter)
+        CALC_SUM_SSE_INSIDE_LOOP
+        src+= src_stride;
+        dst+= dst_stride;
+      }
+    // x_offset = bilin interpolation and y_offset = 8
+    } else if (y_offset == 8) {
+      __m256i filter, pw8, src_next_reg, src_pack;
+      x_offset <<= 5;
+      filter = _mm256_load_si256((__m256i const *)
+               (bilinear_filters_avx2 + x_offset));
+      pw8 = _mm256_set1_epi16(8);
+      src_reg = _mm256_loadu_si256((__m256i const *) (src));
+      MERGE_NEXT_SRC(src_reg, 1)
+      FILTER_SRC(filter)
+      // convert each 16 bit to 8 bit to each low and high lane source
+      src_pack =  _mm256_packus_epi16(exp_src_lo, exp_src_hi);
+      for (i = 0; i < height ; i++) {
+        src+= src_stride;
+        LOAD_SRC_DST
+        MERGE_NEXT_SRC(src_reg, 1)
+        FILTER_SRC(filter)
+        src_reg =  _mm256_packus_epi16(exp_src_lo, exp_src_hi);
+        // average between previous pack to the current
+        src_pack = _mm256_avg_epu8(src_pack, src_reg);
+        MERGE_WITH_SRC(src_pack, zero_reg)
+        CALC_SUM_SSE_INSIDE_LOOP
+        src_pack = src_reg;
+        dst+= dst_stride;
+      }
+    // x_offset = bilin interpolation and y_offset = bilin interpolation
+    } else {
+      __m256i xfilter, yfilter, pw8, src_next_reg, src_pack;
+      x_offset <<= 5;
+      xfilter = _mm256_load_si256((__m256i const *)
+                (bilinear_filters_avx2 + x_offset));
+      y_offset <<= 5;
+      yfilter = _mm256_load_si256((__m256i const *)
+                (bilinear_filters_avx2 + y_offset));
+      pw8 = _mm256_set1_epi16(8);
+      // load source and another source starting from the next
+      // following byte
+      src_reg = _mm256_loadu_si256((__m256i const *) (src));
+      MERGE_NEXT_SRC(src_reg, 1)
+
+      FILTER_SRC(xfilter)
+      // convert each 16 bit to 8 bit to each low and high lane source
+      src_pack = _mm256_packus_epi16(exp_src_lo, exp_src_hi);
+      for (i = 0; i < height ; i++) {
+        src+= src_stride;
+        LOAD_SRC_DST
+        MERGE_NEXT_SRC(src_reg, 1)
+        FILTER_SRC(xfilter)
+        src_reg = _mm256_packus_epi16(exp_src_lo, exp_src_hi);
+        // merge previous pack to current pack source
+        MERGE_WITH_SRC(src_pack, src_reg)
+        // filter the source
+        FILTER_SRC(yfilter)
+        src_pack = src_reg;
+        CALC_SUM_SSE_INSIDE_LOOP
+        dst+= dst_stride;
+      }
+    }
+  }
+  CALC_SUM_AND_SSE
+  return sum;
+}
+
+unsigned int vp9_sub_pixel_avg_variance32xh_avx2(const uint8_t *src,
+                                             int src_stride,
+                                             int x_offset,
+                                             int y_offset,
+                                             const uint8_t *dst,
+                                             int dst_stride,
+                                             const uint8_t *sec,
+                                             int sec_stride,
+                                             int height,
+                                             unsigned int *sse) {
+  __m256i sec_reg;
+  __m256i src_reg, dst_reg, exp_src_lo, exp_src_hi, exp_dst_lo, exp_dst_hi;
+  __m256i sse_reg, sum_reg, sse_reg_hi, res_cmp, sum_reg_lo, sum_reg_hi;
+  __m256i zero_reg;
+  int i, sum;
+  sum_reg = _mm256_set1_epi16(0);
+  sse_reg = _mm256_set1_epi16(0);
+  zero_reg = _mm256_set1_epi16(0);
+
+  // x_offset = 0 and y_offset = 0
+  if (x_offset == 0) {
+    if (y_offset == 0) {
+      for (i = 0; i < height ; i++) {
+        LOAD_SRC_DST
+        sec_reg = _mm256_loadu_si256((__m256i const *) (sec));
+        src_reg = _mm256_avg_epu8(src_reg, sec_reg);
+        sec+= sec_stride;
+        // expend each byte to 2 bytes
+        MERGE_WITH_SRC(src_reg, zero_reg)
+        CALC_SUM_SSE_INSIDE_LOOP
+        src+= src_stride;
+        dst+= dst_stride;
+      }
+    } else if (y_offset == 8) {
+      __m256i src_next_reg;
+      for (i = 0; i < height ; i++) {
+        LOAD_SRC_DST
+        AVG_NEXT_SRC(src_reg, src_stride)
+        sec_reg = _mm256_loadu_si256((__m256i const *) (sec));
+        src_reg = _mm256_avg_epu8(src_reg, sec_reg);
+        sec+= sec_stride;
+        // expend each byte to 2 bytes
+        MERGE_WITH_SRC(src_reg, zero_reg)
+        CALC_SUM_SSE_INSIDE_LOOP
+        src+= src_stride;
+        dst+= dst_stride;
+      }
+    // x_offset = 0 and y_offset = bilin interpolation
+    } else {
+      __m256i filter, pw8, src_next_reg;
+
+      y_offset <<= 5;
+      filter = _mm256_load_si256((__m256i const *)
+                 (bilinear_filters_avx2 + y_offset));
+      pw8 = _mm256_set1_epi16(8);
+      for (i = 0; i < height ; i++) {
+        LOAD_SRC_DST
+        MERGE_NEXT_SRC(src_reg, src_stride)
+        FILTER_SRC(filter)
+        src_reg = _mm256_packus_epi16(exp_src_lo, exp_src_hi);
+        sec_reg = _mm256_loadu_si256((__m256i const *) (sec));
+        src_reg = _mm256_avg_epu8(src_reg, sec_reg);
+        sec+= sec_stride;
+        MERGE_WITH_SRC(src_reg, zero_reg)
+        CALC_SUM_SSE_INSIDE_LOOP
+        src+= src_stride;
+        dst+= dst_stride;
+      }
+    }
+  // x_offset = 8  and y_offset = 0
+  } else if (x_offset == 8) {
+    if (y_offset == 0) {
+      __m256i src_next_reg;
+      for (i = 0; i < height ; i++) {
+        LOAD_SRC_DST
+        AVG_NEXT_SRC(src_reg, 1)
+        sec_reg = _mm256_loadu_si256((__m256i const *) (sec));
+        src_reg = _mm256_avg_epu8(src_reg, sec_reg);
+        sec+= sec_stride;
+        // expand each byte to 2 bytes
+        MERGE_WITH_SRC(src_reg, zero_reg)
+        CALC_SUM_SSE_INSIDE_LOOP
+        src+= src_stride;
+        dst+= dst_stride;
+      }
+    // x_offset = 8  and y_offset = 8
+    } else if (y_offset == 8) {
+      __m256i src_next_reg, src_avg;
+      // load source and another source starting from the next
+      // following byte
+      src_reg = _mm256_loadu_si256((__m256i const *) (src));
+      AVG_NEXT_SRC(src_reg, 1)
+      for (i = 0; i < height ; i++) {
+        // save current source average
+        src_avg = src_reg;
+        src+= src_stride;
+        LOAD_SRC_DST
+        AVG_NEXT_SRC(src_reg, 1)
+        // average between previous average to current average
+        src_avg = _mm256_avg_epu8(src_avg, src_reg);
+        sec_reg = _mm256_loadu_si256((__m256i const *) (sec));
+        src_avg = _mm256_avg_epu8(src_avg, sec_reg);
+        sec+= sec_stride;
+        // expand each byte to 2 bytes
+        MERGE_WITH_SRC(src_avg, zero_reg)
+        CALC_SUM_SSE_INSIDE_LOOP
+        dst+= dst_stride;
+      }
+    // x_offset = 8  and y_offset = bilin interpolation
+    } else {
+      __m256i filter, pw8, src_next_reg, src_avg;
+      y_offset <<= 5;
+      filter = _mm256_load_si256((__m256i const *)
+               (bilinear_filters_avx2 + y_offset));
+      pw8 = _mm256_set1_epi16(8);
+      // load source and another source starting from the next
+      // following byte
+      src_reg = _mm256_loadu_si256((__m256i const *) (src));
+      AVG_NEXT_SRC(src_reg, 1)
+      for (i = 0; i < height ; i++) {
+        // save current source average
+        src_avg = src_reg;
+        src+= src_stride;
+        LOAD_SRC_DST
+        AVG_NEXT_SRC(src_reg, 1)
+        MERGE_WITH_SRC(src_avg, src_reg)
+        FILTER_SRC(filter)
+        src_avg = _mm256_packus_epi16(exp_src_lo, exp_src_hi);
+        sec_reg = _mm256_loadu_si256((__m256i const *) (sec));
+        src_avg = _mm256_avg_epu8(src_avg, sec_reg);
+        // expand each byte to 2 bytes
+        MERGE_WITH_SRC(src_avg, zero_reg)
+        sec+= sec_stride;
+        CALC_SUM_SSE_INSIDE_LOOP
+        dst+= dst_stride;
+      }
+    }
+  // x_offset = bilin interpolation and y_offset = 0
+  } else {
+    if (y_offset == 0) {
+      __m256i filter, pw8, src_next_reg;
+      x_offset <<= 5;
+      filter = _mm256_load_si256((__m256i const *)
+               (bilinear_filters_avx2 + x_offset));
+      pw8 = _mm256_set1_epi16(8);
+      for (i = 0; i < height ; i++) {
+        LOAD_SRC_DST
+        MERGE_NEXT_SRC(src_reg, 1)
+        FILTER_SRC(filter)
+        src_reg = _mm256_packus_epi16(exp_src_lo, exp_src_hi);
+        sec_reg = _mm256_loadu_si256((__m256i const *) (sec));
+        src_reg = _mm256_avg_epu8(src_reg, sec_reg);
+        MERGE_WITH_SRC(src_reg, zero_reg)
+        sec+= sec_stride;
+        CALC_SUM_SSE_INSIDE_LOOP
+        src+= src_stride;
+        dst+= dst_stride;
+      }
+    // x_offset = bilin interpolation and y_offset = 8
+    } else if (y_offset == 8) {
+      __m256i filter, pw8, src_next_reg, src_pack;
+      x_offset <<= 5;
+      filter = _mm256_load_si256((__m256i const *)
+               (bilinear_filters_avx2 + x_offset));
+      pw8 = _mm256_set1_epi16(8);
+      src_reg = _mm256_loadu_si256((__m256i const *) (src));
+      MERGE_NEXT_SRC(src_reg, 1)
+      FILTER_SRC(filter)
+      // convert each 16 bit to 8 bit to each low and high lane source
+      src_pack =  _mm256_packus_epi16(exp_src_lo, exp_src_hi);
+      for (i = 0; i < height ; i++) {
+        src+= src_stride;
+        LOAD_SRC_DST
+        MERGE_NEXT_SRC(src_reg, 1)
+        FILTER_SRC(filter)
+        src_reg =  _mm256_packus_epi16(exp_src_lo, exp_src_hi);
+        // average between previous pack to the current
+        src_pack = _mm256_avg_epu8(src_pack, src_reg);
+        sec_reg = _mm256_loadu_si256((__m256i const *) (sec));
+        src_pack = _mm256_avg_epu8(src_pack, sec_reg);
+        sec+= sec_stride;
+        MERGE_WITH_SRC(src_pack, zero_reg)
+        src_pack = src_reg;
+        CALC_SUM_SSE_INSIDE_LOOP
+        dst+= dst_stride;
+      }
+    // x_offset = bilin interpolation and y_offset = bilin interpolation
+    } else {
+      __m256i xfilter, yfilter, pw8, src_next_reg, src_pack;
+      x_offset <<= 5;
+      xfilter = _mm256_load_si256((__m256i const *)
+                (bilinear_filters_avx2 + x_offset));
+      y_offset <<= 5;
+      yfilter = _mm256_load_si256((__m256i const *)
+                (bilinear_filters_avx2 + y_offset));
+      pw8 = _mm256_set1_epi16(8);
+      // load source and another source starting from the next
+      // following byte
+      src_reg = _mm256_loadu_si256((__m256i const *) (src));
+      MERGE_NEXT_SRC(src_reg, 1)
+
+      FILTER_SRC(xfilter)
+      // convert each 16 bit to 8 bit to each low and high lane source
+      src_pack = _mm256_packus_epi16(exp_src_lo, exp_src_hi);
+      for (i = 0; i < height ; i++) {
+        src+= src_stride;
+        LOAD_SRC_DST
+        MERGE_NEXT_SRC(src_reg, 1)
+        FILTER_SRC(xfilter)
+        src_reg = _mm256_packus_epi16(exp_src_lo, exp_src_hi);
+        // merge previous pack to current pack source
+        MERGE_WITH_SRC(src_pack, src_reg)
+        // filter the source
+        FILTER_SRC(yfilter)
+        src_pack = _mm256_packus_epi16(exp_src_lo, exp_src_hi);
+        sec_reg = _mm256_loadu_si256((__m256i const *) (sec));
+        src_pack = _mm256_avg_epu8(src_pack, sec_reg);
+        MERGE_WITH_SRC(src_pack, zero_reg)
+        src_pack = src_reg;
+        sec+= sec_stride;
+        CALC_SUM_SSE_INSIDE_LOOP
+        dst+= dst_stride;
+      }
+    }
+  }
+  CALC_SUM_AND_SSE
+  return sum;
+}
diff --git a/media/libvpx/vp9/encoder/x86/vp9_subtract_sse2.asm b/media/libvpx/vp9/encoder/x86/vp9_subtract_sse2.asm
new file mode 100644
index 000000000..982408083
--- /dev/null
+++ b/media/libvpx/vp9/encoder/x86/vp9_subtract_sse2.asm
@@ -0,0 +1,127 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+%include "third_party/x86inc/x86inc.asm"
+
+SECTION .text
+
+; void vp9_subtract_block(int rows, int cols,
+;                         int16_t *diff, ptrdiff_t diff_stride,
+;                         const uint8_t *src, ptrdiff_t src_stride,
+;                         const uint8_t *pred, ptrdiff_t pred_stride)
+
+INIT_XMM sse2
+cglobal subtract_block, 7, 7, 8, \
+                        rows, cols, diff, diff_stride, src, src_stride, \
+                        pred, pred_stride
+%define pred_str colsq
+  pxor                  m7, m7         ; dedicated zero register
+  cmp                colsd, 4
+  je .case_4
+  cmp                colsd, 8
+  je .case_8
+  cmp                colsd, 16
+  je .case_16
+  cmp                colsd, 32
+  je .case_32
+
+%macro loop16 6
+  mova                  m0, [srcq+%1]
+  mova                  m4, [srcq+%2]
+  mova                  m1, [predq+%3]
+  mova                  m5, [predq+%4]
+  punpckhbw             m2, m0, m7
+  punpckhbw             m3, m1, m7
+  punpcklbw             m0, m7
+  punpcklbw             m1, m7
+  psubw                 m2, m3
+  psubw                 m0, m1
+  punpckhbw             m1, m4, m7
+  punpckhbw             m3, m5, m7
+  punpcklbw             m4, m7
+  punpcklbw             m5, m7
+  psubw                 m1, m3
+  psubw                 m4, m5
+  mova [diffq+mmsize*0+%5], m0
+  mova [diffq+mmsize*1+%5], m2
+  mova [diffq+mmsize*0+%6], m4
+  mova [diffq+mmsize*1+%6], m1
+%endmacro
+
+  mov             pred_str, pred_stridemp
+.loop_64:
+  loop16 0*mmsize, 1*mmsize, 0*mmsize, 1*mmsize, 0*mmsize, 2*mmsize
+  loop16 2*mmsize, 3*mmsize, 2*mmsize, 3*mmsize, 4*mmsize, 6*mmsize
+  lea                diffq, [diffq+diff_strideq*2]
+  add                predq, pred_str
+  add                 srcq, src_strideq
+  dec                rowsd
+  jg .loop_64
+  RET
+
+.case_32:
+  mov             pred_str, pred_stridemp
+.loop_32:
+  loop16 0, mmsize, 0, mmsize, 0, 2*mmsize
+  lea                diffq, [diffq+diff_strideq*2]
+  add                predq, pred_str
+  add                 srcq, src_strideq
+  dec                rowsd
+  jg .loop_32
+  RET
+
+.case_16:
+  mov             pred_str, pred_stridemp
+.loop_16:
+  loop16 0, src_strideq, 0, pred_str, 0, diff_strideq*2
+  lea                diffq, [diffq+diff_strideq*4]
+  lea                predq, [predq+pred_str*2]
+  lea                 srcq, [srcq+src_strideq*2]
+  sub                rowsd, 2
+  jg .loop_16
+  RET
+
+%macro loop_h 0
+  movh                  m0, [srcq]
+  movh                  m2, [srcq+src_strideq]
+  movh                  m1, [predq]
+  movh                  m3, [predq+pred_str]
+  punpcklbw             m0, m7
+  punpcklbw             m1, m7
+  punpcklbw             m2, m7
+  punpcklbw             m3, m7
+  psubw                 m0, m1
+  psubw                 m2, m3
+  mova             [diffq], m0
+  mova [diffq+diff_strideq*2], m2
+%endmacro
+
+.case_8:
+  mov             pred_str, pred_stridemp
+.loop_8:
+  loop_h
+  lea                diffq, [diffq+diff_strideq*4]
+  lea                 srcq, [srcq+src_strideq*2]
+  lea                predq, [predq+pred_str*2]
+  sub                rowsd, 2
+  jg .loop_8
+  RET
+
+INIT_MMX
+.case_4:
+  mov             pred_str, pred_stridemp
+.loop_4:
+  loop_h
+  lea                diffq, [diffq+diff_strideq*4]
+  lea                 srcq, [srcq+src_strideq*2]
+  lea                predq, [predq+pred_str*2]
+  sub                rowsd, 2
+  jg .loop_4
+  RET
diff --git a/media/libvpx/vp9/encoder/x86/vp9_temporal_filter_apply_sse2.asm b/media/libvpx/vp9/encoder/x86/vp9_temporal_filter_apply_sse2.asm
new file mode 100644
index 000000000..21aaa9383
--- /dev/null
+++ b/media/libvpx/vp9/encoder/x86/vp9_temporal_filter_apply_sse2.asm
@@ -0,0 +1,212 @@
+;
+;  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+;
+;  Use of this source code is governed by a BSD-style license
+;  that can be found in the LICENSE file in the root of the source
+;  tree. An additional intellectual property rights grant can be found
+;  in the file PATENTS.  All contributing project authors may
+;  be found in the AUTHORS file in the root of the source tree.
+;
+
+
+%include "vpx_ports/x86_abi_support.asm"
+
+; void vp9_temporal_filter_apply_sse2 | arg
+;  (unsigned char  *frame1,           |  0
+;   unsigned int    stride,           |  1
+;   unsigned char  *frame2,           |  2
+;   unsigned int    block_width,      |  3
+;   unsigned int    block_height,     |  4
+;   int             strength,         |  5
+;   int             filter_weight,    |  6
+;   unsigned int   *accumulator,      |  7
+;   unsigned short *count)            |  8
+global sym(vp9_temporal_filter_apply_sse2) PRIVATE
+sym(vp9_temporal_filter_apply_sse2):
+
+    push        rbp
+    mov         rbp, rsp
+    SHADOW_ARGS_TO_STACK 9
+    SAVE_XMM 7
+    GET_GOT     rbx
+    push        rsi
+    push        rdi
+    ALIGN_STACK 16, rax
+    %define block_width    0
+    %define block_height  16
+    %define strength      32
+    %define filter_weight 48
+    %define rounding_bit  64
+    %define rbp_backup    80
+    %define stack_size    96
+    sub         rsp,           stack_size
+    mov         [rsp + rbp_backup], rbp
+    ; end prolog
+
+        mov         edx,            arg(3)
+        mov         [rsp + block_width], rdx
+        mov         edx,            arg(4)
+        mov         [rsp + block_height], rdx
+        movd        xmm6,           arg(5)
+        movdqa      [rsp + strength], xmm6 ; where strength is used, all 16 bytes are read
+
+        ; calculate the rounding bit outside the loop
+        ; 0x8000 >> (16 - strength)
+        mov         rdx,            16
+        sub         rdx,            arg(5) ; 16 - strength
+        movq        xmm4,           rdx    ; can't use rdx w/ shift
+        movdqa      xmm5,           [GLOBAL(_const_top_bit)]
+        psrlw       xmm5,           xmm4
+        movdqa      [rsp + rounding_bit], xmm5
+
+        mov         rsi,            arg(0) ; src/frame1
+        mov         rdx,            arg(2) ; predictor frame
+        mov         rdi,            arg(7) ; accumulator
+        mov         rax,            arg(8) ; count
+
+        ; dup the filter weight and store for later
+        movd        xmm0,           arg(6) ; filter_weight
+        pshuflw     xmm0,           xmm0, 0
+        punpcklwd   xmm0,           xmm0
+        movdqa      [rsp + filter_weight], xmm0
+
+        mov         rbp,            arg(1) ; stride
+        pxor        xmm7,           xmm7   ; zero for extraction
+
+        mov         rcx,            [rsp + block_width]
+        imul        rcx,            [rsp + block_height]
+        add         rcx,            rdx
+        cmp         dword ptr [rsp + block_width], 8
+        jne         .temporal_filter_apply_load_16
+
+.temporal_filter_apply_load_8:
+        movq        xmm0,           [rsi]  ; first row
+        lea         rsi,            [rsi + rbp] ; += stride
+        punpcklbw   xmm0,           xmm7   ; src[ 0- 7]
+        movq        xmm1,           [rsi]  ; second row
+        lea         rsi,            [rsi + rbp] ; += stride
+        punpcklbw   xmm1,           xmm7   ; src[ 8-15]
+        jmp         .temporal_filter_apply_load_finished
+
+.temporal_filter_apply_load_16:
+        movdqa      xmm0,           [rsi]  ; src (frame1)
+        lea         rsi,            [rsi + rbp] ; += stride
+        movdqa      xmm1,           xmm0
+        punpcklbw   xmm0,           xmm7   ; src[ 0- 7]
+        punpckhbw   xmm1,           xmm7   ; src[ 8-15]
+
+.temporal_filter_apply_load_finished:
+        movdqa      xmm2,           [rdx]  ; predictor (frame2)
+        movdqa      xmm3,           xmm2
+        punpcklbw   xmm2,           xmm7   ; pred[ 0- 7]
+        punpckhbw   xmm3,           xmm7   ; pred[ 8-15]
+
+        ; modifier = src_byte - pixel_value
+        psubw       xmm0,           xmm2   ; src - pred[ 0- 7]
+        psubw       xmm1,           xmm3   ; src - pred[ 8-15]
+
+        ; modifier *= modifier
+        pmullw      xmm0,           xmm0   ; modifer[ 0- 7]^2
+        pmullw      xmm1,           xmm1   ; modifer[ 8-15]^2
+
+        ; modifier *= 3
+        pmullw      xmm0,           [GLOBAL(_const_3w)]
+        pmullw      xmm1,           [GLOBAL(_const_3w)]
+
+        ; modifer += 0x8000 >> (16 - strength)
+        paddw       xmm0,           [rsp + rounding_bit]
+        paddw       xmm1,           [rsp + rounding_bit]
+
+        ; modifier >>= strength
+        psrlw       xmm0,           [rsp + strength]
+        psrlw       xmm1,           [rsp + strength]
+
+        ; modifier = 16 - modifier
+        ; saturation takes care of modifier > 16
+        movdqa      xmm3,           [GLOBAL(_const_16w)]
+        movdqa      xmm2,           [GLOBAL(_const_16w)]
+        psubusw     xmm3,           xmm1
+        psubusw     xmm2,           xmm0
+
+        ; modifier *= filter_weight
+        pmullw      xmm2,           [rsp + filter_weight]
+        pmullw      xmm3,           [rsp + filter_weight]
+
+        ; count
+        movdqa      xmm4,           [rax]
+        movdqa      xmm5,           [rax+16]
+        ; += modifier
+        paddw       xmm4,           xmm2
+        paddw       xmm5,           xmm3
+        ; write back
+        movdqa      [rax],          xmm4
+        movdqa      [rax+16],       xmm5
+        lea         rax,            [rax + 16*2] ; count += 16*(sizeof(short))
+
+        ; load and extract the predictor up to shorts
+        pxor        xmm7,           xmm7
+        movdqa      xmm0,           [rdx]
+        lea         rdx,            [rdx + 16*1] ; pred += 16*(sizeof(char))
+        movdqa      xmm1,           xmm0
+        punpcklbw   xmm0,           xmm7   ; pred[ 0- 7]
+        punpckhbw   xmm1,           xmm7   ; pred[ 8-15]
+
+        ; modifier *= pixel_value
+        pmullw      xmm0,           xmm2
+        pmullw      xmm1,           xmm3
+
+        ; expand to double words
+        movdqa      xmm2,           xmm0
+        punpcklwd   xmm0,           xmm7   ; [ 0- 3]
+        punpckhwd   xmm2,           xmm7   ; [ 4- 7]
+        movdqa      xmm3,           xmm1
+        punpcklwd   xmm1,           xmm7   ; [ 8-11]
+        punpckhwd   xmm3,           xmm7   ; [12-15]
+
+        ; accumulator
+        movdqa      xmm4,           [rdi]
+        movdqa      xmm5,           [rdi+16]
+        movdqa      xmm6,           [rdi+32]
+        movdqa      xmm7,           [rdi+48]
+        ; += modifier
+        paddd       xmm4,           xmm0
+        paddd       xmm5,           xmm2
+        paddd       xmm6,           xmm1
+        paddd       xmm7,           xmm3
+        ; write back
+        movdqa      [rdi],          xmm4
+        movdqa      [rdi+16],       xmm5
+        movdqa      [rdi+32],       xmm6
+        movdqa      [rdi+48],       xmm7
+        lea         rdi,            [rdi + 16*4] ; accumulator += 16*(sizeof(int))
+
+        cmp         rdx,            rcx
+        je          .temporal_filter_apply_epilog
+        pxor        xmm7,           xmm7   ; zero for extraction
+        cmp         dword ptr [rsp + block_width], 16
+        je          .temporal_filter_apply_load_16
+        jmp         .temporal_filter_apply_load_8
+
+.temporal_filter_apply_epilog:
+    ; begin epilog
+    mov         rbp,            [rsp + rbp_backup]
+    add         rsp,            stack_size
+    pop         rsp
+    pop         rdi
+    pop         rsi
+    RESTORE_GOT
+    RESTORE_XMM
+    UNSHADOW_ARGS
+    pop         rbp
+    ret
+
+SECTION_RODATA
+align 16
+_const_3w:
+    times 8 dw 3
+align 16
+_const_top_bit:
+    times 8 dw 1<<15
+align 16
+_const_16w
+    times 8 dw 16
diff --git a/media/libvpx/vp9/encoder/x86/vp9_variance_avx2.c b/media/libvpx/vp9/encoder/x86/vp9_variance_avx2.c
new file mode 100644
index 000000000..8cd071de5
--- /dev/null
+++ b/media/libvpx/vp9/encoder/x86/vp9_variance_avx2.c
@@ -0,0 +1,104 @@
+/*
+ *  Copyright (c) 2012 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+#include "./vp9_rtcd.h"
+#include "./vpx_config.h"
+
+#include "vp9/encoder/vp9_variance.h"
+#include "vpx_ports/mem.h"
+
+unsigned int vp9_sub_pixel_variance32xh_avx2(const uint8_t *src, int src_stride,
+                                             int x_offset, int y_offset,
+                                             const uint8_t *dst, int dst_stride,
+                                             int height,
+                                             unsigned int *sse);
+
+unsigned int vp9_sub_pixel_avg_variance32xh_avx2(const uint8_t *src,
+                                                 int src_stride,
+                                                 int x_offset,
+                                                 int y_offset,
+                                                 const uint8_t *dst,
+                                                 int dst_stride,
+                                                 const uint8_t *sec,
+                                                 int sec_stride,
+                                                 int height,
+                                                 unsigned int *sseptr);
+
+unsigned int vp9_sub_pixel_variance64x64_avx2(const uint8_t *src,
+                                              int src_stride,
+                                              int x_offset,
+                                              int y_offset,
+                                              const uint8_t *dst,
+                                              int dst_stride,
+                                              unsigned int *sse) {
+  unsigned int sse1;
+  const int se1 = vp9_sub_pixel_variance32xh_avx2(src, src_stride, x_offset,
+                                                  y_offset, dst, dst_stride,
+                                                  64, &sse1);
+  unsigned int sse2;
+  const int se2 = vp9_sub_pixel_variance32xh_avx2(src + 32, src_stride,
+                                                  x_offset, y_offset,
+                                                  dst + 32, dst_stride,
+                                                  64, &sse2);
+  const int se = se1 + se2;
+  *sse = sse1 + sse2;
+  return *sse - (((int64_t)se * se) >> 12);
+}
+
+unsigned int vp9_sub_pixel_variance32x32_avx2(const uint8_t *src,
+                                              int src_stride,
+                                              int x_offset,
+                                              int y_offset,
+                                              const uint8_t *dst,
+                                              int dst_stride,
+                                              unsigned int *sse) {
+  const int se = vp9_sub_pixel_variance32xh_avx2(src, src_stride, x_offset,
+                                                 y_offset, dst, dst_stride,
+                                                 32, sse);
+  return *sse - (((int64_t)se * se) >> 10);
+}
+
+unsigned int vp9_sub_pixel_avg_variance64x64_avx2(const uint8_t *src,
+                                                  int src_stride,
+                                                  int x_offset,
+                                                  int y_offset,
+                                                  const uint8_t *dst,
+                                                  int dst_stride,
+                                                  unsigned int *sse,
+                                                  const uint8_t *sec) {
+  unsigned int sse1;
+  const int se1 = vp9_sub_pixel_avg_variance32xh_avx2(src, src_stride, x_offset,
+                                                      y_offset, dst, dst_stride,
+                                                      sec, 64, 64, &sse1);
+  unsigned int sse2;
+  const int se2 =
+      vp9_sub_pixel_avg_variance32xh_avx2(src + 32, src_stride, x_offset,
+                                          y_offset, dst + 32, dst_stride,
+                                          sec + 32, 64, 64, &sse2);
+  const int se = se1 + se2;
+
+  *sse = sse1 + sse2;
+
+  return *sse - (((int64_t)se * se) >> 12);
+}
+
+unsigned int vp9_sub_pixel_avg_variance32x32_avx2(const uint8_t *src,
+                                                  int src_stride,
+                                                  int x_offset,
+                                                  int y_offset,
+                                                  const uint8_t *dst,
+                                                  int dst_stride,
+                                                  unsigned int *sse,
+                                                  const uint8_t *sec) {
+  // processing 32 element in parallel
+  const int se = vp9_sub_pixel_avg_variance32xh_avx2(src, src_stride, x_offset,
+                                                     y_offset, dst, dst_stride,
+                                                     sec, 32, 32, sse);
+  return *sse - (((int64_t)se * se) >> 10);
+}
diff --git a/media/libvpx/vp9/encoder/x86/vp9_variance_sse2.c b/media/libvpx/vp9/encoder/x86/vp9_variance_sse2.c
new file mode 100644
index 000000000..961efe34e
--- /dev/null
+++ b/media/libvpx/vp9/encoder/x86/vp9_variance_sse2.c
@@ -0,0 +1,182 @@
+/*
+ *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include <emmintrin.h>  // SSE2
+
+#include "./vp9_rtcd.h"
+#include "./vpx_config.h"
+
+#include "vp9/encoder/vp9_variance.h"
+#include "vpx_ports/mem.h"
+
+// The 2 unused parameters are place holders for PIC enabled build.
+#define DECL(w, opt) \
+int vp9_sub_pixel_variance##w##xh_##opt(const uint8_t *src, \
+                                        ptrdiff_t src_stride, \
+                                        int x_offset, int y_offset, \
+                                        const uint8_t *dst, \
+                                        ptrdiff_t dst_stride, \
+                                        int height, unsigned int *sse, \
+                                        void *unused0, void *unused)
+#define DECLS(opt1, opt2) \
+DECL(4, opt2); \
+DECL(8, opt1); \
+DECL(16, opt1)
+
+DECLS(sse2, sse);
+DECLS(ssse3, ssse3);
+#undef DECLS
+#undef DECL
+
+#define FN(w, h, wf, wlog2, hlog2, opt, cast) \
+unsigned int vp9_sub_pixel_variance##w##x##h##_##opt(const uint8_t *src, \
+                                                     int src_stride, \
+                                                     int x_offset, \
+                                                     int y_offset, \
+                                                     const uint8_t *dst, \
+                                                     int dst_stride, \
+                                                     unsigned int *sse_ptr) { \
+  unsigned int sse; \
+  int se = vp9_sub_pixel_variance##wf##xh_##opt(src, src_stride, x_offset, \
+                                                y_offset, dst, dst_stride, \
+                                                h, &sse, NULL, NULL); \
+  if (w > wf) { \
+    unsigned int sse2; \
+    int se2 = vp9_sub_pixel_variance##wf##xh_##opt(src + 16, src_stride, \
+                                                   x_offset, y_offset, \
+                                                   dst + 16, dst_stride, \
+                                                   h, &sse2, NULL, NULL); \
+    se += se2; \
+    sse += sse2; \
+    if (w > wf * 2) { \
+      se2 = vp9_sub_pixel_variance##wf##xh_##opt(src + 32, src_stride, \
+                                                 x_offset, y_offset, \
+                                                 dst + 32, dst_stride, \
+                                                 h, &sse2, NULL, NULL); \
+      se += se2; \
+      sse += sse2; \
+      se2 = vp9_sub_pixel_variance##wf##xh_##opt(src + 48, src_stride, \
+                                                 x_offset, y_offset, \
+                                                 dst + 48, dst_stride, \
+                                                 h, &sse2, NULL, NULL); \
+      se += se2; \
+      sse += sse2; \
+    } \
+  } \
+  *sse_ptr = sse; \
+  return sse - ((cast se * se) >> (wlog2 + hlog2)); \
+}
+
+#define FNS(opt1, opt2) \
+FN(64, 64, 16, 6, 6, opt1, (int64_t)); \
+FN(64, 32, 16, 6, 5, opt1, (int64_t)); \
+FN(32, 64, 16, 5, 6, opt1, (int64_t)); \
+FN(32, 32, 16, 5, 5, opt1, (int64_t)); \
+FN(32, 16, 16, 5, 4, opt1, (int64_t)); \
+FN(16, 32, 16, 4, 5, opt1, (int64_t)); \
+FN(16, 16, 16, 4, 4, opt1, (unsigned int)); \
+FN(16,  8, 16, 4, 3, opt1, (unsigned int)); \
+FN(8,  16,  8, 3, 4, opt1, (unsigned int)); \
+FN(8,   8,  8, 3, 3, opt1, (unsigned int)); \
+FN(8,   4,  8, 3, 2, opt1, (unsigned int)); \
+FN(4,   8,  4, 2, 3, opt2, (unsigned int)); \
+FN(4,   4,  4, 2, 2, opt2, (unsigned int))
+
+FNS(sse2, sse);
+FNS(ssse3, ssse3);
+
+#undef FNS
+#undef FN
+
+// The 2 unused parameters are place holders for PIC enabled build.
+#define DECL(w, opt) \
+int vp9_sub_pixel_avg_variance##w##xh_##opt(const uint8_t *src, \
+                                            ptrdiff_t src_stride, \
+                                            int x_offset, int y_offset, \
+                                            const uint8_t *dst, \
+                                            ptrdiff_t dst_stride, \
+                                            const uint8_t *sec, \
+                                            ptrdiff_t sec_stride, \
+                                            int height, unsigned int *sse, \
+                                            void *unused0, void *unused)
+#define DECLS(opt1, opt2) \
+DECL(4, opt2); \
+DECL(8, opt1); \
+DECL(16, opt1)
+
+DECLS(sse2, sse);
+DECLS(ssse3, ssse3);
+#undef DECL
+#undef DECLS
+
+#define FN(w, h, wf, wlog2, hlog2, opt, cast) \
+unsigned int vp9_sub_pixel_avg_variance##w##x##h##_##opt(const uint8_t *src, \
+                                                         int src_stride, \
+                                                         int x_offset, \
+                                                         int y_offset, \
+                                                         const uint8_t *dst, \
+                                                         int dst_stride, \
+                                                         unsigned int *sseptr, \
+                                                         const uint8_t *sec) { \
+  unsigned int sse; \
+  int se = vp9_sub_pixel_avg_variance##wf##xh_##opt(src, src_stride, x_offset, \
+                                                    y_offset, dst, dst_stride, \
+                                                    sec, w, h, &sse, NULL, \
+                                                    NULL); \
+  if (w > wf) { \
+    unsigned int sse2; \
+    int se2 = vp9_sub_pixel_avg_variance##wf##xh_##opt(src + 16, src_stride, \
+                                                       x_offset, y_offset, \
+                                                       dst + 16, dst_stride, \
+                                                       sec + 16, w, h, &sse2, \
+                                                       NULL, NULL); \
+    se += se2; \
+    sse += sse2; \
+    if (w > wf * 2) { \
+      se2 = vp9_sub_pixel_avg_variance##wf##xh_##opt(src + 32, src_stride, \
+                                                     x_offset, y_offset, \
+                                                     dst + 32, dst_stride, \
+                                                     sec + 32, w, h, &sse2, \
+                                                     NULL, NULL); \
+      se += se2; \
+      sse += sse2; \
+      se2 = vp9_sub_pixel_avg_variance##wf##xh_##opt(src + 48, src_stride, \
+                                                     x_offset, y_offset, \
+                                                     dst + 48, dst_stride, \
+                                                     sec + 48, w, h, &sse2, \
+                                                     NULL, NULL); \
+      se += se2; \
+      sse += sse2; \
+    } \
+  } \
+  *sseptr = sse; \
+  return sse - ((cast se * se) >> (wlog2 + hlog2)); \
+}
+
+#define FNS(opt1, opt2) \
+FN(64, 64, 16, 6, 6, opt1, (int64_t)); \
+FN(64, 32, 16, 6, 5, opt1, (int64_t)); \
+FN(32, 64, 16, 5, 6, opt1, (int64_t)); \
+FN(32, 32, 16, 5, 5, opt1, (int64_t)); \
+FN(32, 16, 16, 5, 4, opt1, (int64_t)); \
+FN(16, 32, 16, 4, 5, opt1, (int64_t)); \
+FN(16, 16, 16, 4, 4, opt1, (unsigned int)); \
+FN(16,  8, 16, 4, 3, opt1, (unsigned int)); \
+FN(8,  16,  8, 3, 4, opt1, (unsigned int)); \
+FN(8,   8,  8, 3, 3, opt1, (unsigned int)); \
+FN(8,   4,  8, 3, 2, opt1, (unsigned int)); \
+FN(4,   8,  4, 2, 3, opt2, (unsigned int)); \
+FN(4,   4,  4, 2, 2, opt2, (unsigned int))
+
+FNS(sse2, sse);
+FNS(ssse3, ssse3);
+
+#undef FNS
+#undef FN