convolve_avx2.c 10.7 KB
Newer Older
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
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
/*
 * Copyright (c) 2017, 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 <immintrin.h>

#include "aom_dsp/aom_dsp_common.h"
#include "./av1_rtcd.h"

#if CONFIG_CONVOLVE_ROUND
static const uint32_t sindex[8] = { 0, 4, 1, 5, 2, 6, 3, 7 };

// 16 epi16 pixels
static INLINE void pixel_clamp_avx2(__m256i *u, int bd) {
  const __m256i one = _mm256_set1_epi16(1);
  const __m256i max = _mm256_sub_epi16(_mm256_slli_epi16(one, bd), one);
  __m256i clamped, mask;

  mask = _mm256_cmpgt_epi16(*u, max);
  clamped = _mm256_andnot_si256(mask, *u);
  mask = _mm256_and_si256(mask, max);
  clamped = _mm256_or_si256(mask, clamped);

  const __m256i zero = _mm256_setzero_si256();
  mask = _mm256_cmpgt_epi16(clamped, zero);
  *u = _mm256_and_si256(clamped, mask);
}

// 8 epi16 pixels
static INLINE void pixel_clamp_sse2(__m128i *u, int bd) {
  const __m128i one = _mm_set1_epi16(1);
  const __m128i max = _mm_sub_epi16(_mm_slli_epi16(one, bd), one);
  __m128i clamped, mask;

  mask = _mm_cmpgt_epi16(*u, max);
  clamped = _mm_andnot_si128(mask, *u);
  mask = _mm_and_si128(mask, max);
  clamped = _mm_or_si128(mask, clamped);

  const __m128i zero = _mm_setzero_si128();
  mask = _mm_cmpgt_epi16(clamped, zero);
  *u = _mm_and_si128(clamped, mask);
}

// Work on multiple of 32 pixels
static INLINE void cal_rounding_32xn_avx2(const int32_t *src, uint8_t *dst,
                                          const __m256i *rnd, int shift,
                                          int num) {
  do {
    __m256i x0 = _mm256_loadu_si256((const __m256i *)src);
    __m256i x1 = _mm256_loadu_si256((const __m256i *)src + 1);
    __m256i x2 = _mm256_loadu_si256((const __m256i *)src + 2);
    __m256i x3 = _mm256_loadu_si256((const __m256i *)src + 3);

    x0 = _mm256_add_epi32(x0, *rnd);
    x1 = _mm256_add_epi32(x1, *rnd);
    x2 = _mm256_add_epi32(x2, *rnd);
    x3 = _mm256_add_epi32(x3, *rnd);

    x0 = _mm256_srai_epi32(x0, shift);
    x1 = _mm256_srai_epi32(x1, shift);
    x2 = _mm256_srai_epi32(x2, shift);
    x3 = _mm256_srai_epi32(x3, shift);

    x0 = _mm256_packs_epi32(x0, x1);
    x2 = _mm256_packs_epi32(x2, x3);

    pixel_clamp_avx2(&x0, 8);
    pixel_clamp_avx2(&x2, 8);

    x0 = _mm256_packus_epi16(x0, x2);
    x1 = _mm256_loadu_si256((const __m256i *)sindex);
    x2 = _mm256_permutevar8x32_epi32(x0, x1);

    _mm256_storeu_si256((__m256i *)dst, x2);
    src += 32;
    dst += 32;
    num--;
  } while (num > 0);
}

static INLINE void cal_rounding_16_avx2(const int32_t *src, uint8_t *dst,
                                        const __m256i *rnd, int shift) {
  __m256i x0 = _mm256_loadu_si256((const __m256i *)src);
  __m256i x1 = _mm256_loadu_si256((const __m256i *)src + 1);

  x0 = _mm256_add_epi32(x0, *rnd);
  x1 = _mm256_add_epi32(x1, *rnd);

  x0 = _mm256_srai_epi32(x0, shift);
  x1 = _mm256_srai_epi32(x1, shift);

  x0 = _mm256_packs_epi32(x0, x1);
  pixel_clamp_avx2(&x0, 8);

  const __m256i x2 = _mm256_packus_epi16(x0, x0);
  x1 = _mm256_loadu_si256((const __m256i *)sindex);
  x0 = _mm256_permutevar8x32_epi32(x2, x1);

  _mm_storeu_si128((__m128i *)dst, _mm256_castsi256_si128(x0));
}

static INLINE void cal_rounding_8_avx2(const int32_t *src, uint8_t *dst,
                                       const __m256i *rnd, int shift) {
  __m256i x0 = _mm256_loadu_si256((const __m256i *)src);
  x0 = _mm256_add_epi32(x0, *rnd);
  x0 = _mm256_srai_epi32(x0, shift);

  x0 = _mm256_packs_epi32(x0, x0);
  pixel_clamp_avx2(&x0, 8);

  x0 = _mm256_packus_epi16(x0, x0);
  const __m256i x1 = _mm256_loadu_si256((const __m256i *)sindex);
  x0 = _mm256_permutevar8x32_epi32(x0, x1);

  _mm_storel_epi64((__m128i *)dst, _mm256_castsi256_si128(x0));
}

static INLINE void cal_rounding_4_sse2(const int32_t *src, uint8_t *dst,
                                       const __m128i *rnd, int shift) {
  __m128i x = _mm_loadu_si128((const __m128i *)src);
  x = _mm_add_epi32(x, *rnd);
  x = _mm_srai_epi32(x, shift);

  x = _mm_packs_epi32(x, x);
  pixel_clamp_sse2(&x, 8);

  x = _mm_packus_epi16(x, x);
  *(uint32_t *)dst = _mm_cvtsi128_si32(x);
}

void av1_convolve_rounding_avx2(const int32_t *src, int src_stride,
                                uint8_t *dst, int dst_stride, int w, int h,
                                int bits) {
  const __m256i rnd_num = _mm256_set1_epi32((int32_t)(1 << (bits - 1)));
  const __m128i rnd_num_sse2 = _mm256_castsi256_si128(rnd_num);

  if (w > 64) {  // width = 128
    do {
      cal_rounding_32xn_avx2(src, dst, &rnd_num, bits, 4);
      src += src_stride;
      dst += dst_stride;
      h--;
    } while (h > 0);
  } else if (w > 32) {  // width = 64
    do {
      cal_rounding_32xn_avx2(src, dst, &rnd_num, bits, 2);
      src += src_stride;
      dst += dst_stride;
      h--;
    } while (h > 0);
  } else if (w > 16) {  // width = 32
    do {
      cal_rounding_32xn_avx2(src, dst, &rnd_num, bits, 1);
      src += src_stride;
      dst += dst_stride;
      h--;
    } while (h > 0);
  } else if (w > 8) {  // width = 16
    do {
      cal_rounding_16_avx2(src, dst, &rnd_num, bits);
      src += src_stride;
      dst += dst_stride;
      h--;
    } while (h > 0);
  } else if (w > 4) {  // width = 8
    do {
      cal_rounding_8_avx2(src, dst, &rnd_num, bits);
      src += src_stride;
      dst += dst_stride;
      h--;
    } while (h > 0);
  } else if (w > 2) {  // width = 4
    do {
      cal_rounding_4_sse2(src, dst, &rnd_num_sse2, bits);
      src += src_stride;
      dst += dst_stride;
      h--;
    } while (h > 0);
  } else {  // width = 2
    do {
      dst[0] = clip_pixel(ROUND_POWER_OF_TWO(src[0], bits));
      dst[1] = clip_pixel(ROUND_POWER_OF_TWO(src[1], bits));
      src += src_stride;
      dst += dst_stride;
      h--;
    } while (h > 0);
  }
}

#if CONFIG_HIGHBITDEPTH
static INLINE void cal_highbd_rounding_32xn_avx2(const int32_t *src,
                                                 uint16_t *dst,
                                                 const __m256i *rnd, int shift,
                                                 int num, int bd) {
  do {
    __m256i x0 = _mm256_loadu_si256((const __m256i *)src);
    __m256i x1 = _mm256_loadu_si256((const __m256i *)src + 1);
    __m256i x2 = _mm256_loadu_si256((const __m256i *)src + 2);
    __m256i x3 = _mm256_loadu_si256((const __m256i *)src + 3);

    x0 = _mm256_add_epi32(x0, *rnd);
    x1 = _mm256_add_epi32(x1, *rnd);
    x2 = _mm256_add_epi32(x2, *rnd);
    x3 = _mm256_add_epi32(x3, *rnd);

    x0 = _mm256_srai_epi32(x0, shift);
    x1 = _mm256_srai_epi32(x1, shift);
    x2 = _mm256_srai_epi32(x2, shift);
    x3 = _mm256_srai_epi32(x3, shift);

    x0 = _mm256_packs_epi32(x0, x1);
    x2 = _mm256_packs_epi32(x2, x3);

    pixel_clamp_avx2(&x0, bd);
    pixel_clamp_avx2(&x2, bd);

    x0 = _mm256_permute4x64_epi64(x0, 0xD8);
    x2 = _mm256_permute4x64_epi64(x2, 0xD8);

    _mm256_storeu_si256((__m256i *)dst, x0);
    _mm256_storeu_si256((__m256i *)(dst + 16), x2);
    src += 32;
    dst += 32;
    num--;
  } while (num > 0);
}

static INLINE void cal_highbd_rounding_16_avx2(const int32_t *src,
                                               uint16_t *dst,
                                               const __m256i *rnd, int shift,
                                               int bd) {
  __m256i x0 = _mm256_loadu_si256((const __m256i *)src);
  __m256i x1 = _mm256_loadu_si256((const __m256i *)src + 1);

  x0 = _mm256_add_epi32(x0, *rnd);
  x1 = _mm256_add_epi32(x1, *rnd);

  x0 = _mm256_srai_epi32(x0, shift);
  x1 = _mm256_srai_epi32(x1, shift);

  x0 = _mm256_packs_epi32(x0, x1);
  pixel_clamp_avx2(&x0, bd);

  x0 = _mm256_permute4x64_epi64(x0, 0xD8);
  _mm256_storeu_si256((__m256i *)dst, x0);
}

static INLINE void cal_highbd_rounding_8_avx2(const int32_t *src, uint16_t *dst,
                                              const __m256i *rnd, int shift,
                                              int bd) {
  __m256i x = _mm256_loadu_si256((const __m256i *)src);
  x = _mm256_add_epi32(x, *rnd);
  x = _mm256_srai_epi32(x, shift);

  x = _mm256_packs_epi32(x, x);
  pixel_clamp_avx2(&x, bd);

  x = _mm256_permute4x64_epi64(x, 0xD8);
  _mm_storeu_si128((__m128i *)dst, _mm256_castsi256_si128(x));
}

static INLINE void cal_highbd_rounding_4_sse2(const int32_t *src, uint16_t *dst,
                                              const __m128i *rnd, int shift,
                                              int bd) {
  __m128i x = _mm_loadu_si128((const __m128i *)src);
  x = _mm_add_epi32(x, *rnd);
  x = _mm_srai_epi32(x, shift);

  x = _mm_packs_epi32(x, x);
  pixel_clamp_sse2(&x, bd);
  _mm_storel_epi64((__m128i *)dst, x);
}

void av1_highbd_convolve_rounding_avx2(const int32_t *src, int src_stride,
                                       uint8_t *dst8, int dst_stride, int w,
                                       int h, int bits, int bd) {
  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
  const __m256i rnd_num = _mm256_set1_epi32((int32_t)(1 << (bits - 1)));
  const __m128i rnd_num_sse2 = _mm256_castsi256_si128(rnd_num);

  if (w > 64) {  // width = 128
    do {
      cal_highbd_rounding_32xn_avx2(src, dst, &rnd_num, bits, 4, bd);
      src += src_stride;
      dst += dst_stride;
      h--;
    } while (h > 0);
  } else if (w > 32) {  // width = 64
    do {
      cal_highbd_rounding_32xn_avx2(src, dst, &rnd_num, bits, 2, bd);
      src += src_stride;
      dst += dst_stride;
      h--;
    } while (h > 0);
  } else if (w > 16) {  // width = 32
    do {
      cal_highbd_rounding_32xn_avx2(src, dst, &rnd_num, bits, 1, bd);
      src += src_stride;
      dst += dst_stride;
      h--;
    } while (h > 0);
  } else if (w > 8) {  // width = 16
    do {
      cal_highbd_rounding_16_avx2(src, dst, &rnd_num, bits, bd);
      src += src_stride;
      dst += dst_stride;
      h--;
    } while (h > 0);
  } else if (w > 4) {  // width = 8
    do {
      cal_highbd_rounding_8_avx2(src, dst, &rnd_num, bits, bd);
      src += src_stride;
      dst += dst_stride;
      h--;
    } while (h > 0);
  } else if (w > 2) {  // width = 4
    do {
      cal_highbd_rounding_4_sse2(src, dst, &rnd_num_sse2, bits, bd);
      src += src_stride;
      dst += dst_stride;
      h--;
    } while (h > 0);
  } else {  // width = 2
    do {
      dst[0] = clip_pixel_highbd(ROUND_POWER_OF_TWO(src[0], bits), bd);
      dst[1] = clip_pixel_highbd(ROUND_POWER_OF_TWO(src[1], bits), bd);
      src += src_stride;
      dst += dst_stride;
      h--;
    } while (h > 0);
  }
}
#endif  // CONFIG_HIGHBITDEPTH
#endif  // CONFIG_CONVOLVE_ROUND