move vp9_avg to vpx_dsp

Change-Id: I7bc991abea383db1f86c1bb0f2e849837b54d90f

2 files changed, 544 insertions, 0 deletions

diff --git a/vpx_dsp/x86/avg_intrin_sse2.c b/vpx_dsp/x86/avg_intrin_sse2.c
new file mode 100644
index 000000000..f9af6cf97
--- /dev/null
+++ b/vpx_dsp/x86/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 "./vpx_dsp_rtcd.h"
+#include "vpx_ports/mem.h"
+
+void vpx_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 vpx_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 vpx_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 vpx_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 vpx_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;
+    vpx_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);
+
+    b0 = _mm_srai_epi16(b0, 1);
+    b1 = _mm_srai_epi16(b1, 1);
+    b2 = _mm_srai_epi16(b2, 1);
+    b3 = _mm_srai_epi16(b3, 1);
+
+    coeff0 = _mm_add_epi16(b0, b2);
+    coeff1 = _mm_add_epi16(b1, b3);
+    _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);
+    _mm_store_si128((__m128i *)(coeff + 128), coeff2);
+    _mm_store_si128((__m128i *)(coeff + 192), coeff3);
+
+    coeff += 8;
+  }
+}
+
+int vpx_satd_sse2(const int16_t *coeff, int length) {
+  int i;
+  const __m128i zero = _mm_setzero_si128();
+  __m128i accum = zero;
+
+  for (i = 0; i < length; i += 8) {
+    const __m128i src_line = _mm_load_si128((const __m128i *)coeff);
+    const __m128i inv = _mm_sub_epi16(zero, src_line);
+    const __m128i abs = _mm_max_epi16(src_line, inv);  // abs(src_line)
+    const __m128i abs_lo = _mm_unpacklo_epi16(abs, zero);
+    const __m128i abs_hi = _mm_unpackhi_epi16(abs, zero);
+    const __m128i sum = _mm_add_epi32(abs_lo, abs_hi);
+    accum = _mm_add_epi32(accum, sum);
+    coeff += 8;
+  }
+
+  {  // cascading summation of accum
+    __m128i hi = _mm_srli_si128(accum, 8);
+    accum = _mm_add_epi32(accum, hi);
+    hi = _mm_srli_epi64(accum, 32);
+    accum = _mm_add_epi32(accum, hi);
+  }
+
+  return _mm_cvtsi128_si32(accum);
+}
+
+void vpx_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 vpx_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 vpx_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/vpx_dsp/x86/avg_ssse3_x86_64.asm b/vpx_dsp/x86/avg_ssse3_x86_64.asm
new file mode 100644
index 000000000..26412e8e4
--- /dev/null
+++ b/vpx_dsp/x86/avg_ssse3_x86_64.asm
@@ -0,0 +1,121 @@
+;
+;  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.
+;
+
+%define private_prefix vpx
+
+%include "third_party/x86inc/x86inc.asm"
+
+; This file provides SSSE3 version of the hadamard transformation. Part
+; of the macro definitions are originally derived from the ffmpeg project.
+; The current version applies to x86 64-bit only.
+
+SECTION .text
+
+%if ARCH_X86_64
+; 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
+
+%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