summaryrefslogtreecommitdiffstats
path: root/media/libaom/src/aom_dsp/x86/sum_squares_sse2.c
blob: 22d7739ec4c5c29e56cc35d1d0f61c38a1b32f51 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
/*
 * Copyright (c) 2016, Alliance for Open Media. All rights reserved
 *
 * This source code is subject to the terms of the BSD 2 Clause License and
 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
 * was not distributed with this source code in the LICENSE file, you can
 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
 * Media Patent License 1.0 was not distributed with this source code in the
 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
 */

#include <assert.h>
#include <emmintrin.h>
#include <stdio.h>

#include "aom_dsp/x86/synonyms.h"
#include "aom_dsp/x86/sum_squares_sse2.h"
#include "config/aom_dsp_rtcd.h"

static INLINE __m128i xx_loadh_64(__m128i a, const void *b) {
  const __m128d ad = _mm_castsi128_pd(a);
  return _mm_castpd_si128(_mm_loadh_pd(ad, (double *)b));
}

static INLINE uint64_t xx_cvtsi128_si64(__m128i a) {
#if ARCH_X86_64
  return (uint64_t)_mm_cvtsi128_si64(a);
#else
  {
    uint64_t tmp;
    _mm_storel_epi64((__m128i *)&tmp, a);
    return tmp;
  }
#endif
}

static INLINE __m128i sum_squares_i16_4x4_sse2(const int16_t *src, int stride) {
  const __m128i v_val_0_w = xx_loadl_64(src + 0 * stride);
  const __m128i v_val_2_w = xx_loadl_64(src + 2 * stride);
  const __m128i v_val_01_w = xx_loadh_64(v_val_0_w, src + 1 * stride);
  const __m128i v_val_23_w = xx_loadh_64(v_val_2_w, src + 3 * stride);
  const __m128i v_sq_01_d = _mm_madd_epi16(v_val_01_w, v_val_01_w);
  const __m128i v_sq_23_d = _mm_madd_epi16(v_val_23_w, v_val_23_w);

  return _mm_add_epi32(v_sq_01_d, v_sq_23_d);
}

uint64_t aom_sum_squares_2d_i16_4x4_sse2(const int16_t *src, int stride) {
  const __m128i v_sum_0123_d = sum_squares_i16_4x4_sse2(src, stride);
  __m128i v_sum_d =
      _mm_add_epi32(v_sum_0123_d, _mm_srli_epi64(v_sum_0123_d, 32));
  v_sum_d = _mm_add_epi32(v_sum_d, _mm_srli_si128(v_sum_d, 8));
  return (uint64_t)_mm_cvtsi128_si32(v_sum_d);
}

uint64_t aom_sum_squares_2d_i16_4xn_sse2(const int16_t *src, int stride,
                                         int height) {
  int r = 0;
  __m128i v_acc_q = _mm_setzero_si128();
  do {
    const __m128i v_acc_d = sum_squares_i16_4x4_sse2(src, stride);
    v_acc_q = _mm_add_epi32(v_acc_q, v_acc_d);
    src += stride << 2;
    r += 4;
  } while (r < height);
  const __m128i v_zext_mask_q = xx_set1_64_from_32i(0xffffffff);
  __m128i v_acc_64 = _mm_add_epi64(_mm_srli_epi64(v_acc_q, 32),
                                   _mm_and_si128(v_acc_q, v_zext_mask_q));
  v_acc_64 = _mm_add_epi64(v_acc_64, _mm_srli_si128(v_acc_64, 8));
  return xx_cvtsi128_si64(v_acc_64);
}

#ifdef __GNUC__
// This prevents GCC/Clang from inlining this function into
// aom_sum_squares_2d_i16_sse2, which in turn saves some stack
// maintenance instructions in the common case of 4x4.
__attribute__((noinline))
#endif
uint64_t
aom_sum_squares_2d_i16_nxn_sse2(const int16_t *src, int stride, int width,
                                int height) {
  int r = 0;

  const __m128i v_zext_mask_q = xx_set1_64_from_32i(0xffffffff);
  __m128i v_acc_q = _mm_setzero_si128();

  do {
    __m128i v_acc_d = _mm_setzero_si128();
    int c = 0;
    do {
      const int16_t *b = src + c;

      const __m128i v_val_0_w = xx_load_128(b + 0 * stride);
      const __m128i v_val_1_w = xx_load_128(b + 1 * stride);
      const __m128i v_val_2_w = xx_load_128(b + 2 * stride);
      const __m128i v_val_3_w = xx_load_128(b + 3 * stride);

      const __m128i v_sq_0_d = _mm_madd_epi16(v_val_0_w, v_val_0_w);
      const __m128i v_sq_1_d = _mm_madd_epi16(v_val_1_w, v_val_1_w);
      const __m128i v_sq_2_d = _mm_madd_epi16(v_val_2_w, v_val_2_w);
      const __m128i v_sq_3_d = _mm_madd_epi16(v_val_3_w, v_val_3_w);

      const __m128i v_sum_01_d = _mm_add_epi32(v_sq_0_d, v_sq_1_d);
      const __m128i v_sum_23_d = _mm_add_epi32(v_sq_2_d, v_sq_3_d);

      const __m128i v_sum_0123_d = _mm_add_epi32(v_sum_01_d, v_sum_23_d);

      v_acc_d = _mm_add_epi32(v_acc_d, v_sum_0123_d);
      c += 8;
    } while (c < width);

    v_acc_q = _mm_add_epi64(v_acc_q, _mm_and_si128(v_acc_d, v_zext_mask_q));
    v_acc_q = _mm_add_epi64(v_acc_q, _mm_srli_epi64(v_acc_d, 32));

    src += 4 * stride;
    r += 4;
  } while (r < height);

  v_acc_q = _mm_add_epi64(v_acc_q, _mm_srli_si128(v_acc_q, 8));
  return xx_cvtsi128_si64(v_acc_q);
}

uint64_t aom_sum_squares_2d_i16_sse2(const int16_t *src, int stride, int width,
                                     int height) {
  // 4 elements per row only requires half an XMM register, so this
  // must be a special case, but also note that over 75% of all calls
  // are with size == 4, so it is also the common case.
  if (LIKELY(width == 4 && height == 4)) {
    return aom_sum_squares_2d_i16_4x4_sse2(src, stride);
  } else if (LIKELY(width == 4 && (height & 3) == 0)) {
    return aom_sum_squares_2d_i16_4xn_sse2(src, stride, height);
  } else if (LIKELY((width & 7) == 0 && (height & 3) == 0)) {
    // Generic case
    return aom_sum_squares_2d_i16_nxn_sse2(src, stride, width, height);
  } else {
    return aom_sum_squares_2d_i16_c(src, stride, width, height);
  }
}

//////////////////////////////////////////////////////////////////////////////
// 1D version
//////////////////////////////////////////////////////////////////////////////

static uint64_t aom_sum_squares_i16_64n_sse2(const int16_t *src, uint32_t n) {
  const __m128i v_zext_mask_q = xx_set1_64_from_32i(0xffffffff);
  __m128i v_acc0_q = _mm_setzero_si128();
  __m128i v_acc1_q = _mm_setzero_si128();

  const int16_t *const end = src + n;

  assert(n % 64 == 0);

  while (src < end) {
    const __m128i v_val_0_w = xx_load_128(src);
    const __m128i v_val_1_w = xx_load_128(src + 8);
    const __m128i v_val_2_w = xx_load_128(src + 16);
    const __m128i v_val_3_w = xx_load_128(src + 24);
    const __m128i v_val_4_w = xx_load_128(src + 32);
    const __m128i v_val_5_w = xx_load_128(src + 40);
    const __m128i v_val_6_w = xx_load_128(src + 48);
    const __m128i v_val_7_w = xx_load_128(src + 56);

    const __m128i v_sq_0_d = _mm_madd_epi16(v_val_0_w, v_val_0_w);
    const __m128i v_sq_1_d = _mm_madd_epi16(v_val_1_w, v_val_1_w);
    const __m128i v_sq_2_d = _mm_madd_epi16(v_val_2_w, v_val_2_w);
    const __m128i v_sq_3_d = _mm_madd_epi16(v_val_3_w, v_val_3_w);
    const __m128i v_sq_4_d = _mm_madd_epi16(v_val_4_w, v_val_4_w);
    const __m128i v_sq_5_d = _mm_madd_epi16(v_val_5_w, v_val_5_w);
    const __m128i v_sq_6_d = _mm_madd_epi16(v_val_6_w, v_val_6_w);
    const __m128i v_sq_7_d = _mm_madd_epi16(v_val_7_w, v_val_7_w);

    const __m128i v_sum_01_d = _mm_add_epi32(v_sq_0_d, v_sq_1_d);
    const __m128i v_sum_23_d = _mm_add_epi32(v_sq_2_d, v_sq_3_d);
    const __m128i v_sum_45_d = _mm_add_epi32(v_sq_4_d, v_sq_5_d);
    const __m128i v_sum_67_d = _mm_add_epi32(v_sq_6_d, v_sq_7_d);

    const __m128i v_sum_0123_d = _mm_add_epi32(v_sum_01_d, v_sum_23_d);
    const __m128i v_sum_4567_d = _mm_add_epi32(v_sum_45_d, v_sum_67_d);

    const __m128i v_sum_d = _mm_add_epi32(v_sum_0123_d, v_sum_4567_d);

    v_acc0_q = _mm_add_epi64(v_acc0_q, _mm_and_si128(v_sum_d, v_zext_mask_q));
    v_acc1_q = _mm_add_epi64(v_acc1_q, _mm_srli_epi64(v_sum_d, 32));

    src += 64;
  }

  v_acc0_q = _mm_add_epi64(v_acc0_q, v_acc1_q);
  v_acc0_q = _mm_add_epi64(v_acc0_q, _mm_srli_si128(v_acc0_q, 8));
  return xx_cvtsi128_si64(v_acc0_q);
}

uint64_t aom_sum_squares_i16_sse2(const int16_t *src, uint32_t n) {
  if (n % 64 == 0) {
    return aom_sum_squares_i16_64n_sse2(src, n);
  } else if (n > 64) {
    int k = n & ~(64 - 1);
    return aom_sum_squares_i16_64n_sse2(src, k) +
           aom_sum_squares_i16_c(src + k, n - k);
  } else {
    return aom_sum_squares_i16_c(src, n);
  }
}