summaryrefslogtreecommitdiffstats
path: root/third_party/aom/aom_dsp/x86/obmc_sad_sse4.c
blob: ad77f974c7ffdfde52ca4f268aa38d2c46949f76 (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
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
/*
 * Copyright (c) 2016, Alliance for Open Media. All rights reserved
 *
 * This source code is subject to the terms of the BSD 2 Clause License and
 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
 * was not distributed with this source code in the LICENSE file, you can
 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
 * Media Patent License 1.0 was not distributed with this source code in the
 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
 */

#include <assert.h>
#include <immintrin.h>

#include "./aom_config.h"
#include "aom_ports/mem.h"
#include "aom/aom_integer.h"

#include "aom_dsp/aom_dsp_common.h"
#include "aom_dsp/x86/synonyms.h"

////////////////////////////////////////////////////////////////////////////////
// 8 bit
////////////////////////////////////////////////////////////////////////////////

static INLINE unsigned int obmc_sad_w4(const uint8_t *pre, const int pre_stride,
                                       const int32_t *wsrc, const int32_t *mask,
                                       const int height) {
  const int pre_step = pre_stride - 4;
  int n = 0;
  __m128i v_sad_d = _mm_setzero_si128();

  do {
    const __m128i v_p_b = xx_loadl_32(pre + n);
    const __m128i v_m_d = xx_load_128(mask + n);
    const __m128i v_w_d = xx_load_128(wsrc + n);

    const __m128i v_p_d = _mm_cvtepu8_epi32(v_p_b);

    // Values in both pre and mask fit in 15 bits, and are packed at 32 bit
    // boundaries. We use pmaddwd, as it has lower latency on Haswell
    // than pmulld but produces the same result with these inputs.
    const __m128i v_pm_d = _mm_madd_epi16(v_p_d, v_m_d);

    const __m128i v_diff_d = _mm_sub_epi32(v_w_d, v_pm_d);
    const __m128i v_absdiff_d = _mm_abs_epi32(v_diff_d);

    // Rounded absolute difference
    const __m128i v_rad_d = xx_roundn_epu32(v_absdiff_d, 12);

    v_sad_d = _mm_add_epi32(v_sad_d, v_rad_d);

    n += 4;

    if (n % 4 == 0) pre += pre_step;
  } while (n < 4 * height);

  return xx_hsum_epi32_si32(v_sad_d);
}

static INLINE unsigned int obmc_sad_w8n(const uint8_t *pre,
                                        const int pre_stride,
                                        const int32_t *wsrc,
                                        const int32_t *mask, const int width,
                                        const int height) {
  const int pre_step = pre_stride - width;
  int n = 0;
  __m128i v_sad_d = _mm_setzero_si128();

  assert(width >= 8);
  assert(IS_POWER_OF_TWO(width));

  do {
    const __m128i v_p1_b = xx_loadl_32(pre + n + 4);
    const __m128i v_m1_d = xx_load_128(mask + n + 4);
    const __m128i v_w1_d = xx_load_128(wsrc + n + 4);
    const __m128i v_p0_b = xx_loadl_32(pre + n);
    const __m128i v_m0_d = xx_load_128(mask + n);
    const __m128i v_w0_d = xx_load_128(wsrc + n);

    const __m128i v_p0_d = _mm_cvtepu8_epi32(v_p0_b);
    const __m128i v_p1_d = _mm_cvtepu8_epi32(v_p1_b);

    // Values in both pre and mask fit in 15 bits, and are packed at 32 bit
    // boundaries. We use pmaddwd, as it has lower latency on Haswell
    // than pmulld but produces the same result with these inputs.
    const __m128i v_pm0_d = _mm_madd_epi16(v_p0_d, v_m0_d);
    const __m128i v_pm1_d = _mm_madd_epi16(v_p1_d, v_m1_d);

    const __m128i v_diff0_d = _mm_sub_epi32(v_w0_d, v_pm0_d);
    const __m128i v_diff1_d = _mm_sub_epi32(v_w1_d, v_pm1_d);
    const __m128i v_absdiff0_d = _mm_abs_epi32(v_diff0_d);
    const __m128i v_absdiff1_d = _mm_abs_epi32(v_diff1_d);

    // Rounded absolute difference
    const __m128i v_rad0_d = xx_roundn_epu32(v_absdiff0_d, 12);
    const __m128i v_rad1_d = xx_roundn_epu32(v_absdiff1_d, 12);

    v_sad_d = _mm_add_epi32(v_sad_d, v_rad0_d);
    v_sad_d = _mm_add_epi32(v_sad_d, v_rad1_d);

    n += 8;

    if (n % width == 0) pre += pre_step;
  } while (n < width * height);

  return xx_hsum_epi32_si32(v_sad_d);
}

#define OBMCSADWXH(w, h)                                       \
  unsigned int aom_obmc_sad##w##x##h##_sse4_1(                 \
      const uint8_t *pre, int pre_stride, const int32_t *wsrc, \
      const int32_t *msk) {                                    \
    if (w == 4) {                                              \
      return obmc_sad_w4(pre, pre_stride, wsrc, msk, h);       \
    } else {                                                   \
      return obmc_sad_w8n(pre, pre_stride, wsrc, msk, w, h);   \
    }                                                          \
  }

#if CONFIG_EXT_PARTITION
OBMCSADWXH(128, 128)
OBMCSADWXH(128, 64)
OBMCSADWXH(64, 128)
#endif  // CONFIG_EXT_PARTITION
OBMCSADWXH(64, 64)
OBMCSADWXH(64, 32)
OBMCSADWXH(32, 64)
OBMCSADWXH(32, 32)
OBMCSADWXH(32, 16)
OBMCSADWXH(16, 32)
OBMCSADWXH(16, 16)
OBMCSADWXH(16, 8)
OBMCSADWXH(8, 16)
OBMCSADWXH(8, 8)
OBMCSADWXH(8, 4)
OBMCSADWXH(4, 8)
OBMCSADWXH(4, 4)

////////////////////////////////////////////////////////////////////////////////
// High bit-depth
////////////////////////////////////////////////////////////////////////////////

#if CONFIG_HIGHBITDEPTH
static INLINE unsigned int hbd_obmc_sad_w4(const uint8_t *pre8,
                                           const int pre_stride,
                                           const int32_t *wsrc,
                                           const int32_t *mask,
                                           const int height) {
  const uint16_t *pre = CONVERT_TO_SHORTPTR(pre8);
  const int pre_step = pre_stride - 4;
  int n = 0;
  __m128i v_sad_d = _mm_setzero_si128();

  do {
    const __m128i v_p_w = xx_loadl_64(pre + n);
    const __m128i v_m_d = xx_load_128(mask + n);
    const __m128i v_w_d = xx_load_128(wsrc + n);

    const __m128i v_p_d = _mm_cvtepu16_epi32(v_p_w);

    // Values in both pre and mask fit in 15 bits, and are packed at 32 bit
    // boundaries. We use pmaddwd, as it has lower latency on Haswell
    // than pmulld but produces the same result with these inputs.
    const __m128i v_pm_d = _mm_madd_epi16(v_p_d, v_m_d);

    const __m128i v_diff_d = _mm_sub_epi32(v_w_d, v_pm_d);
    const __m128i v_absdiff_d = _mm_abs_epi32(v_diff_d);

    // Rounded absolute difference
    const __m128i v_rad_d = xx_roundn_epu32(v_absdiff_d, 12);

    v_sad_d = _mm_add_epi32(v_sad_d, v_rad_d);

    n += 4;

    if (n % 4 == 0) pre += pre_step;
  } while (n < 4 * height);

  return xx_hsum_epi32_si32(v_sad_d);
}

static INLINE unsigned int hbd_obmc_sad_w8n(const uint8_t *pre8,
                                            const int pre_stride,
                                            const int32_t *wsrc,
                                            const int32_t *mask,
                                            const int width, const int height) {
  const uint16_t *pre = CONVERT_TO_SHORTPTR(pre8);
  const int pre_step = pre_stride - width;
  int n = 0;
  __m128i v_sad_d = _mm_setzero_si128();

  assert(width >= 8);
  assert(IS_POWER_OF_TWO(width));

  do {
    const __m128i v_p1_w = xx_loadl_64(pre + n + 4);
    const __m128i v_m1_d = xx_load_128(mask + n + 4);
    const __m128i v_w1_d = xx_load_128(wsrc + n + 4);
    const __m128i v_p0_w = xx_loadl_64(pre + n);
    const __m128i v_m0_d = xx_load_128(mask + n);
    const __m128i v_w0_d = xx_load_128(wsrc + n);

    const __m128i v_p0_d = _mm_cvtepu16_epi32(v_p0_w);
    const __m128i v_p1_d = _mm_cvtepu16_epi32(v_p1_w);

    // Values in both pre and mask fit in 15 bits, and are packed at 32 bit
    // boundaries. We use pmaddwd, as it has lower latency on Haswell
    // than pmulld but produces the same result with these inputs.
    const __m128i v_pm0_d = _mm_madd_epi16(v_p0_d, v_m0_d);
    const __m128i v_pm1_d = _mm_madd_epi16(v_p1_d, v_m1_d);

    const __m128i v_diff0_d = _mm_sub_epi32(v_w0_d, v_pm0_d);
    const __m128i v_diff1_d = _mm_sub_epi32(v_w1_d, v_pm1_d);
    const __m128i v_absdiff0_d = _mm_abs_epi32(v_diff0_d);
    const __m128i v_absdiff1_d = _mm_abs_epi32(v_diff1_d);

    // Rounded absolute difference
    const __m128i v_rad0_d = xx_roundn_epu32(v_absdiff0_d, 12);
    const __m128i v_rad1_d = xx_roundn_epu32(v_absdiff1_d, 12);

    v_sad_d = _mm_add_epi32(v_sad_d, v_rad0_d);
    v_sad_d = _mm_add_epi32(v_sad_d, v_rad1_d);

    n += 8;

    if (n % width == 0) pre += pre_step;
  } while (n < width * height);

  return xx_hsum_epi32_si32(v_sad_d);
}

#define HBD_OBMCSADWXH(w, h)                                      \
  unsigned int aom_highbd_obmc_sad##w##x##h##_sse4_1(             \
      const uint8_t *pre, int pre_stride, const int32_t *wsrc,    \
      const int32_t *mask) {                                      \
    if (w == 4) {                                                 \
      return hbd_obmc_sad_w4(pre, pre_stride, wsrc, mask, h);     \
    } else {                                                      \
      return hbd_obmc_sad_w8n(pre, pre_stride, wsrc, mask, w, h); \
    }                                                             \
  }

#if CONFIG_EXT_PARTITION
HBD_OBMCSADWXH(128, 128)
HBD_OBMCSADWXH(128, 64)
HBD_OBMCSADWXH(64, 128)
#endif  // CONFIG_EXT_PARTITION
HBD_OBMCSADWXH(64, 64)
HBD_OBMCSADWXH(64, 32)
HBD_OBMCSADWXH(32, 64)
HBD_OBMCSADWXH(32, 32)
HBD_OBMCSADWXH(32, 16)
HBD_OBMCSADWXH(16, 32)
HBD_OBMCSADWXH(16, 16)
HBD_OBMCSADWXH(16, 8)
HBD_OBMCSADWXH(8, 16)
HBD_OBMCSADWXH(8, 8)
HBD_OBMCSADWXH(8, 4)
HBD_OBMCSADWXH(4, 8)
HBD_OBMCSADWXH(4, 4)
#endif  // CONFIG_HIGHBITDEPTH