obmc_variance_sse4.c 13.8 KB
Newer Older
1
/*
Yaowu Xu's avatar
Yaowu Xu committed
2
 * Copyright (c) 2016, Alliance for Open Media. All rights reserved
3
 *
Yaowu Xu's avatar
Yaowu Xu committed
4 5 6 7 8 9
 * 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.
10 11 12 13 14
 */

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

Yaowu Xu's avatar
Yaowu Xu committed
15
#include "./aom_config.h"
16
#include "aom_ports/mem.h"
Yaowu Xu's avatar
Yaowu Xu committed
17
#include "aom/aom_integer.h"
18

Yaowu Xu's avatar
Yaowu Xu committed
19 20
#include "aom_dsp/aom_dsp_common.h"
#include "aom_dsp/aom_filter.h"
21 22
#include "aom_dsp/x86/obmc_intrinsic_ssse3.h"
#include "aom_dsp/x86/synonyms.h"
23 24 25 26 27

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

clang-format's avatar
clang-format committed
28 29 30
static INLINE void obmc_variance_w4(const uint8_t *pre, const int pre_stride,
                                    const int32_t *wsrc, const int32_t *mask,
                                    unsigned int *const sse, int *const sum,
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
                                    const int h) {
  const int pre_step = pre_stride - 4;
  int n = 0;
  __m128i v_sum_d = _mm_setzero_si128();
  __m128i v_sse_d = _mm_setzero_si128();

  assert(IS_POWER_OF_TWO(h));

  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_rdiff_d = xx_roundn_epi32(v_diff_d, 12);
    const __m128i v_sqrdiff_d = _mm_mullo_epi32(v_rdiff_d, v_rdiff_d);

    v_sum_d = _mm_add_epi32(v_sum_d, v_rdiff_d);
    v_sse_d = _mm_add_epi32(v_sse_d, v_sqrdiff_d);

    n += 4;

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

  *sum = xx_hsum_epi32_si32(v_sum_d);
  *sse = xx_hsum_epi32_si32(v_sse_d);
}

clang-format's avatar
clang-format committed
67 68 69 70
static INLINE void obmc_variance_w8n(const uint8_t *pre, const int pre_stride,
                                     const int32_t *wsrc, const int32_t *mask,
                                     unsigned int *const sse, int *const sum,
                                     const int w, const int h) {
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
  const int pre_step = pre_stride - w;
  int n = 0;
  __m128i v_sum_d = _mm_setzero_si128();
  __m128i v_sse_d = _mm_setzero_si128();

  assert(w >= 8);
  assert(IS_POWER_OF_TWO(w));
  assert(IS_POWER_OF_TWO(h));

  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_rdiff0_d = xx_roundn_epi32(v_diff0_d, 12);
    const __m128i v_rdiff1_d = xx_roundn_epi32(v_diff1_d, 12);
    const __m128i v_rdiff01_w = _mm_packs_epi32(v_rdiff0_d, v_rdiff1_d);
    const __m128i v_sqrdiff_d = _mm_madd_epi16(v_rdiff01_w, v_rdiff01_w);

    v_sum_d = _mm_add_epi32(v_sum_d, v_rdiff0_d);
    v_sum_d = _mm_add_epi32(v_sum_d, v_rdiff1_d);
    v_sse_d = _mm_add_epi32(v_sse_d, v_sqrdiff_d);

    n += 8;

    if (n % w == 0) pre += pre_step;
  } while (n < w * h);

  *sum = xx_hsum_epi32_si32(v_sum_d);
  *sse = xx_hsum_epi32_si32(v_sse_d);
}

clang-format's avatar
clang-format committed
118
#define OBMCVARWXH(W, H)                                               \
Yaowu Xu's avatar
Yaowu Xu committed
119
  unsigned int aom_obmc_variance##W##x##H##_sse4_1(                    \
clang-format's avatar
clang-format committed
120 121 122 123 124 125 126 127
      const uint8_t *pre, int pre_stride, const int32_t *wsrc,         \
      const int32_t *mask, unsigned int *sse) {                        \
    int sum;                                                           \
    if (W == 4) {                                                      \
      obmc_variance_w4(pre, pre_stride, wsrc, mask, sse, &sum, H);     \
    } else {                                                           \
      obmc_variance_w8n(pre, pre_stride, wsrc, mask, sse, &sum, W, H); \
    }                                                                  \
128
    return *sse - (unsigned int)(((int64_t)sum * sum) / (W * H));      \
clang-format's avatar
clang-format committed
129
  }
130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148

#if CONFIG_EXT_PARTITION
OBMCVARWXH(128, 128)
OBMCVARWXH(128, 64)
OBMCVARWXH(64, 128)
#endif  // CONFIG_EXT_PARTITION
OBMCVARWXH(64, 64)
OBMCVARWXH(64, 32)
OBMCVARWXH(32, 64)
OBMCVARWXH(32, 32)
OBMCVARWXH(32, 16)
OBMCVARWXH(16, 32)
OBMCVARWXH(16, 16)
OBMCVARWXH(16, 8)
OBMCVARWXH(8, 16)
OBMCVARWXH(8, 8)
OBMCVARWXH(8, 4)
OBMCVARWXH(4, 8)
OBMCVARWXH(4, 4)
149 150 151 152 153 154
#if CONFIG_EXT_PARTITION_TYPES
OBMCVARWXH(4, 16)
OBMCVARWXH(16, 4)
OBMCVARWXH(8, 32)
OBMCVARWXH(32, 8)
#endif
155 156 157 158 159

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

160
#if CONFIG_HIGHBITDEPTH
clang-format's avatar
clang-format committed
161 162 163
static INLINE void hbd_obmc_variance_w4(
    const uint8_t *pre8, const int pre_stride, const int32_t *wsrc,
    const int32_t *mask, uint64_t *const sse, int64_t *const sum, const int h) {
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
  const uint16_t *pre = CONVERT_TO_SHORTPTR(pre8);
  const int pre_step = pre_stride - 4;
  int n = 0;
  __m128i v_sum_d = _mm_setzero_si128();
  __m128i v_sse_d = _mm_setzero_si128();

  assert(IS_POWER_OF_TWO(h));

  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_rdiff_d = xx_roundn_epi32(v_diff_d, 12);
    const __m128i v_sqrdiff_d = _mm_mullo_epi32(v_rdiff_d, v_rdiff_d);

    v_sum_d = _mm_add_epi32(v_sum_d, v_rdiff_d);
    v_sse_d = _mm_add_epi32(v_sse_d, v_sqrdiff_d);

    n += 4;

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

  *sum = xx_hsum_epi32_si32(v_sum_d);
  *sse = xx_hsum_epi32_si32(v_sse_d);
}

clang-format's avatar
clang-format committed
200 201 202 203
static INLINE void hbd_obmc_variance_w8n(
    const uint8_t *pre8, const int pre_stride, const int32_t *wsrc,
    const int32_t *mask, uint64_t *const sse, int64_t *const sum, const int w,
    const int h) {
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
  const uint16_t *pre = CONVERT_TO_SHORTPTR(pre8);
  const int pre_step = pre_stride - w;
  int n = 0;
  __m128i v_sum_d = _mm_setzero_si128();
  __m128i v_sse_d = _mm_setzero_si128();

  assert(w >= 8);
  assert(IS_POWER_OF_TWO(w));
  assert(IS_POWER_OF_TWO(h));

  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_rdiff0_d = xx_roundn_epi32(v_diff0_d, 12);
    const __m128i v_rdiff1_d = xx_roundn_epi32(v_diff1_d, 12);
    const __m128i v_rdiff01_w = _mm_packs_epi32(v_rdiff0_d, v_rdiff1_d);
    const __m128i v_sqrdiff_d = _mm_madd_epi16(v_rdiff01_w, v_rdiff01_w);

    v_sum_d = _mm_add_epi32(v_sum_d, v_rdiff0_d);
    v_sum_d = _mm_add_epi32(v_sum_d, v_rdiff1_d);
    v_sse_d = _mm_add_epi32(v_sse_d, v_sqrdiff_d);

    n += 8;

    if (n % w == 0) pre += pre_step;
  } while (n < w * h);

  *sum += xx_hsum_epi32_si64(v_sum_d);
  *sse += xx_hsum_epi32_si64(v_sse_d);
}

static INLINE void highbd_obmc_variance(const uint8_t *pre8, int pre_stride,
                                        const int32_t *wsrc,
clang-format's avatar
clang-format committed
254
                                        const int32_t *mask, int w, int h,
255 256 257 258 259 260 261 262 263 264 265 266 267 268
                                        unsigned int *sse, int *sum) {
  int64_t sum64 = 0;
  uint64_t sse64 = 0;
  if (w == 4) {
    hbd_obmc_variance_w4(pre8, pre_stride, wsrc, mask, &sse64, &sum64, h);
  } else {
    hbd_obmc_variance_w8n(pre8, pre_stride, wsrc, mask, &sse64, &sum64, w, h);
  }
  *sum = (int)sum64;
  *sse = (unsigned int)sse64;
}

static INLINE void highbd_10_obmc_variance(const uint8_t *pre8, int pre_stride,
                                           const int32_t *wsrc,
clang-format's avatar
clang-format committed
269
                                           const int32_t *mask, int w, int h,
270 271 272 273 274 275 276 277 278 279 280 281 282 283
                                           unsigned int *sse, int *sum) {
  int64_t sum64 = 0;
  uint64_t sse64 = 0;
  if (w == 4) {
    hbd_obmc_variance_w4(pre8, pre_stride, wsrc, mask, &sse64, &sum64, h);
  } else {
    hbd_obmc_variance_w8n(pre8, pre_stride, wsrc, mask, &sse64, &sum64, w, h);
  }
  *sum = (int)ROUND_POWER_OF_TWO(sum64, 2);
  *sse = (unsigned int)ROUND_POWER_OF_TWO(sse64, 4);
}

static INLINE void highbd_12_obmc_variance(const uint8_t *pre8, int pre_stride,
                                           const int32_t *wsrc,
clang-format's avatar
clang-format committed
284
                                           const int32_t *mask, int w, int h,
285 286 287 288 289
                                           unsigned int *sse, int *sum) {
  int64_t sum64 = 0;
  uint64_t sse64 = 0;
  if (w == 128) {
    do {
clang-format's avatar
clang-format committed
290 291
      hbd_obmc_variance_w8n(pre8, pre_stride, wsrc, mask, &sse64, &sum64, 128,
                            32);
292 293 294 295 296 297 298
      pre8 += 32 * pre_stride;
      wsrc += 32 * 128;
      mask += 32 * 128;
      h -= 32;
    } while (h > 0);
  } else if (w == 64 && h >= 128) {
    do {
clang-format's avatar
clang-format committed
299 300
      hbd_obmc_variance_w8n(pre8, pre_stride, wsrc, mask, &sse64, &sum64, 64,
                            64);
301 302 303 304 305 306 307 308 309 310 311 312 313 314
      pre8 += 64 * pre_stride;
      wsrc += 64 * 64;
      mask += 64 * 64;
      h -= 64;
    } while (h > 0);
  } else if (w == 4) {
    hbd_obmc_variance_w4(pre8, pre_stride, wsrc, mask, &sse64, &sum64, h);
  } else {
    hbd_obmc_variance_w8n(pre8, pre_stride, wsrc, mask, &sse64, &sum64, w, h);
  }
  *sum = (int)ROUND_POWER_OF_TWO(sum64, 4);
  *sse = (unsigned int)ROUND_POWER_OF_TWO(sse64, 8);
}

clang-format's avatar
clang-format committed
315
#define HBD_OBMCVARWXH(W, H)                                               \
Yaowu Xu's avatar
Yaowu Xu committed
316
  unsigned int aom_highbd_obmc_variance##W##x##H##_sse4_1(                 \
clang-format's avatar
clang-format committed
317 318 319 320
      const uint8_t *pre, int pre_stride, const int32_t *wsrc,             \
      const int32_t *mask, unsigned int *sse) {                            \
    int sum;                                                               \
    highbd_obmc_variance(pre, pre_stride, wsrc, mask, W, H, sse, &sum);    \
321
    return *sse - (unsigned int)(((int64_t)sum * sum) / (W * H));          \
clang-format's avatar
clang-format committed
322 323
  }                                                                        \
                                                                           \
Yaowu Xu's avatar
Yaowu Xu committed
324
  unsigned int aom_highbd_10_obmc_variance##W##x##H##_sse4_1(              \
clang-format's avatar
clang-format committed
325 326 327
      const uint8_t *pre, int pre_stride, const int32_t *wsrc,             \
      const int32_t *mask, unsigned int *sse) {                            \
    int sum;                                                               \
328
    int64_t var;                                                           \
clang-format's avatar
clang-format committed
329
    highbd_10_obmc_variance(pre, pre_stride, wsrc, mask, W, H, sse, &sum); \
330 331
    var = (int64_t)(*sse) - (((int64_t)sum * sum) / (W * H));              \
    return (var >= 0) ? (uint32_t)var : 0;                                 \
clang-format's avatar
clang-format committed
332 333
  }                                                                        \
                                                                           \
Yaowu Xu's avatar
Yaowu Xu committed
334
  unsigned int aom_highbd_12_obmc_variance##W##x##H##_sse4_1(              \
clang-format's avatar
clang-format committed
335 336 337
      const uint8_t *pre, int pre_stride, const int32_t *wsrc,             \
      const int32_t *mask, unsigned int *sse) {                            \
    int sum;                                                               \
338
    int64_t var;                                                           \
clang-format's avatar
clang-format committed
339
    highbd_12_obmc_variance(pre, pre_stride, wsrc, mask, W, H, sse, &sum); \
340 341
    var = (int64_t)(*sse) - (((int64_t)sum * sum) / (W * H));              \
    return (var >= 0) ? (uint32_t)var : 0;                                 \
clang-format's avatar
clang-format committed
342
  }
343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361

#if CONFIG_EXT_PARTITION
HBD_OBMCVARWXH(128, 128)
HBD_OBMCVARWXH(128, 64)
HBD_OBMCVARWXH(64, 128)
#endif  // CONFIG_EXT_PARTITION
HBD_OBMCVARWXH(64, 64)
HBD_OBMCVARWXH(64, 32)
HBD_OBMCVARWXH(32, 64)
HBD_OBMCVARWXH(32, 32)
HBD_OBMCVARWXH(32, 16)
HBD_OBMCVARWXH(16, 32)
HBD_OBMCVARWXH(16, 16)
HBD_OBMCVARWXH(16, 8)
HBD_OBMCVARWXH(8, 16)
HBD_OBMCVARWXH(8, 8)
HBD_OBMCVARWXH(8, 4)
HBD_OBMCVARWXH(4, 8)
HBD_OBMCVARWXH(4, 4)
362 363 364 365 366 367
#if CONFIG_EXT_PARTITION_TYPES
HBD_OBMCVARWXH(4, 16)
HBD_OBMCVARWXH(16, 4)
HBD_OBMCVARWXH(8, 32)
HBD_OBMCVARWXH(32, 8)
#endif
368
#endif  // CONFIG_HIGHBITDEPTH