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
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
|
/*
* Copyright (c) 2018, 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 <tmmintrin.h>
#include "config/aom_config.h"
#include "config/aom_dsp_rtcd.h"
#include "aom_dsp/blend.h"
#include "aom/aom_integer.h"
#include "aom_dsp/x86/synonyms.h"
#include "aom_dsp/x86/masked_sad_intrin_ssse3.h"
static INLINE unsigned int masked_sad32xh_avx2(
const uint8_t *src_ptr, int src_stride, const uint8_t *a_ptr, int a_stride,
const uint8_t *b_ptr, int b_stride, const uint8_t *m_ptr, int m_stride,
int width, int height) {
int x, y;
__m256i res = _mm256_setzero_si256();
const __m256i mask_max = _mm256_set1_epi8((1 << AOM_BLEND_A64_ROUND_BITS));
const __m256i round_scale =
_mm256_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS));
for (y = 0; y < height; y++) {
for (x = 0; x < width; x += 32) {
const __m256i src = _mm256_lddqu_si256((const __m256i *)&src_ptr[x]);
const __m256i a = _mm256_lddqu_si256((const __m256i *)&a_ptr[x]);
const __m256i b = _mm256_lddqu_si256((const __m256i *)&b_ptr[x]);
const __m256i m = _mm256_lddqu_si256((const __m256i *)&m_ptr[x]);
const __m256i m_inv = _mm256_sub_epi8(mask_max, m);
// Calculate 16 predicted pixels.
// Note that the maximum value of any entry of 'pred_l' or 'pred_r'
// is 64 * 255, so we have plenty of space to add rounding constants.
const __m256i data_l = _mm256_unpacklo_epi8(a, b);
const __m256i mask_l = _mm256_unpacklo_epi8(m, m_inv);
__m256i pred_l = _mm256_maddubs_epi16(data_l, mask_l);
pred_l = _mm256_mulhrs_epi16(pred_l, round_scale);
const __m256i data_r = _mm256_unpackhi_epi8(a, b);
const __m256i mask_r = _mm256_unpackhi_epi8(m, m_inv);
__m256i pred_r = _mm256_maddubs_epi16(data_r, mask_r);
pred_r = _mm256_mulhrs_epi16(pred_r, round_scale);
const __m256i pred = _mm256_packus_epi16(pred_l, pred_r);
res = _mm256_add_epi32(res, _mm256_sad_epu8(pred, src));
}
src_ptr += src_stride;
a_ptr += a_stride;
b_ptr += b_stride;
m_ptr += m_stride;
}
// At this point, we have two 32-bit partial SADs in lanes 0 and 2 of 'res'.
res = _mm256_shuffle_epi32(res, 0xd8);
res = _mm256_permute4x64_epi64(res, 0xd8);
res = _mm256_hadd_epi32(res, res);
res = _mm256_hadd_epi32(res, res);
int32_t sad = _mm256_extract_epi32(res, 0);
return sad;
}
static INLINE __m256i xx_loadu2_m128i(const void *hi, const void *lo) {
__m128i a0 = _mm_lddqu_si128((const __m128i *)(lo));
__m128i a1 = _mm_lddqu_si128((const __m128i *)(hi));
__m256i a = _mm256_castsi128_si256(a0);
return _mm256_inserti128_si256(a, a1, 1);
}
static INLINE unsigned int masked_sad16xh_avx2(
const uint8_t *src_ptr, int src_stride, const uint8_t *a_ptr, int a_stride,
const uint8_t *b_ptr, int b_stride, const uint8_t *m_ptr, int m_stride,
int height) {
int y;
__m256i res = _mm256_setzero_si256();
const __m256i mask_max = _mm256_set1_epi8((1 << AOM_BLEND_A64_ROUND_BITS));
const __m256i round_scale =
_mm256_set1_epi16(1 << (15 - AOM_BLEND_A64_ROUND_BITS));
for (y = 0; y < height; y += 2) {
const __m256i src = xx_loadu2_m128i(src_ptr + src_stride, src_ptr);
const __m256i a = xx_loadu2_m128i(a_ptr + a_stride, a_ptr);
const __m256i b = xx_loadu2_m128i(b_ptr + b_stride, b_ptr);
const __m256i m = xx_loadu2_m128i(m_ptr + m_stride, m_ptr);
const __m256i m_inv = _mm256_sub_epi8(mask_max, m);
// Calculate 16 predicted pixels.
// Note that the maximum value of any entry of 'pred_l' or 'pred_r'
// is 64 * 255, so we have plenty of space to add rounding constants.
const __m256i data_l = _mm256_unpacklo_epi8(a, b);
const __m256i mask_l = _mm256_unpacklo_epi8(m, m_inv);
__m256i pred_l = _mm256_maddubs_epi16(data_l, mask_l);
pred_l = _mm256_mulhrs_epi16(pred_l, round_scale);
const __m256i data_r = _mm256_unpackhi_epi8(a, b);
const __m256i mask_r = _mm256_unpackhi_epi8(m, m_inv);
__m256i pred_r = _mm256_maddubs_epi16(data_r, mask_r);
pred_r = _mm256_mulhrs_epi16(pred_r, round_scale);
const __m256i pred = _mm256_packus_epi16(pred_l, pred_r);
res = _mm256_add_epi32(res, _mm256_sad_epu8(pred, src));
src_ptr += src_stride << 1;
a_ptr += a_stride << 1;
b_ptr += b_stride << 1;
m_ptr += m_stride << 1;
}
// At this point, we have two 32-bit partial SADs in lanes 0 and 2 of 'res'.
res = _mm256_shuffle_epi32(res, 0xd8);
res = _mm256_permute4x64_epi64(res, 0xd8);
res = _mm256_hadd_epi32(res, res);
res = _mm256_hadd_epi32(res, res);
int32_t sad = _mm256_extract_epi32(res, 0);
return sad;
}
static INLINE unsigned int aom_masked_sad_avx2(
const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride,
const uint8_t *second_pred, const uint8_t *msk, int msk_stride,
int invert_mask, int m, int n) {
unsigned int sad;
if (!invert_mask) {
switch (m) {
case 4:
sad = aom_masked_sad4xh_ssse3(src, src_stride, ref, ref_stride,
second_pred, m, msk, msk_stride, n);
break;
case 8:
sad = aom_masked_sad8xh_ssse3(src, src_stride, ref, ref_stride,
second_pred, m, msk, msk_stride, n);
break;
case 16:
sad = masked_sad16xh_avx2(src, src_stride, ref, ref_stride, second_pred,
m, msk, msk_stride, n);
break;
default:
sad = masked_sad32xh_avx2(src, src_stride, ref, ref_stride, second_pred,
m, msk, msk_stride, m, n);
break;
}
} else {
switch (m) {
case 4:
sad = aom_masked_sad4xh_ssse3(src, src_stride, second_pred, m, ref,
ref_stride, msk, msk_stride, n);
break;
case 8:
sad = aom_masked_sad8xh_ssse3(src, src_stride, second_pred, m, ref,
ref_stride, msk, msk_stride, n);
break;
case 16:
sad = masked_sad16xh_avx2(src, src_stride, second_pred, m, ref,
ref_stride, msk, msk_stride, n);
break;
default:
sad = masked_sad32xh_avx2(src, src_stride, second_pred, m, ref,
ref_stride, msk, msk_stride, m, n);
break;
}
}
return sad;
}
#define MASKSADMXN_AVX2(m, n) \
unsigned int aom_masked_sad##m##x##n##_avx2( \
const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride, \
const uint8_t *second_pred, const uint8_t *msk, int msk_stride, \
int invert_mask) { \
return aom_masked_sad_avx2(src, src_stride, ref, ref_stride, second_pred, \
msk, msk_stride, invert_mask, m, n); \
}
MASKSADMXN_AVX2(4, 4)
MASKSADMXN_AVX2(4, 8)
MASKSADMXN_AVX2(8, 4)
MASKSADMXN_AVX2(8, 8)
MASKSADMXN_AVX2(8, 16)
MASKSADMXN_AVX2(16, 8)
MASKSADMXN_AVX2(16, 16)
MASKSADMXN_AVX2(16, 32)
MASKSADMXN_AVX2(32, 16)
MASKSADMXN_AVX2(32, 32)
MASKSADMXN_AVX2(32, 64)
MASKSADMXN_AVX2(64, 32)
MASKSADMXN_AVX2(64, 64)
MASKSADMXN_AVX2(64, 128)
MASKSADMXN_AVX2(128, 64)
MASKSADMXN_AVX2(128, 128)
MASKSADMXN_AVX2(4, 16)
MASKSADMXN_AVX2(16, 4)
MASKSADMXN_AVX2(8, 32)
MASKSADMXN_AVX2(32, 8)
MASKSADMXN_AVX2(16, 64)
MASKSADMXN_AVX2(64, 16)
static INLINE unsigned int highbd_masked_sad8xh_avx2(
const uint8_t *src8, int src_stride, const uint8_t *a8, int a_stride,
const uint8_t *b8, int b_stride, const uint8_t *m_ptr, int m_stride,
int height) {
const uint16_t *src_ptr = CONVERT_TO_SHORTPTR(src8);
const uint16_t *a_ptr = CONVERT_TO_SHORTPTR(a8);
const uint16_t *b_ptr = CONVERT_TO_SHORTPTR(b8);
int y;
__m256i res = _mm256_setzero_si256();
const __m256i mask_max = _mm256_set1_epi16((1 << AOM_BLEND_A64_ROUND_BITS));
const __m256i round_const =
_mm256_set1_epi32((1 << AOM_BLEND_A64_ROUND_BITS) >> 1);
const __m256i one = _mm256_set1_epi16(1);
for (y = 0; y < height; y += 2) {
const __m256i src = xx_loadu2_m128i(src_ptr + src_stride, src_ptr);
const __m256i a = xx_loadu2_m128i(a_ptr + a_stride, a_ptr);
const __m256i b = xx_loadu2_m128i(b_ptr + b_stride, b_ptr);
// Zero-extend mask to 16 bits
const __m256i m = _mm256_cvtepu8_epi16(_mm_unpacklo_epi64(
_mm_loadl_epi64((const __m128i *)(m_ptr)),
_mm_loadl_epi64((const __m128i *)(m_ptr + m_stride))));
const __m256i m_inv = _mm256_sub_epi16(mask_max, m);
const __m256i data_l = _mm256_unpacklo_epi16(a, b);
const __m256i mask_l = _mm256_unpacklo_epi16(m, m_inv);
__m256i pred_l = _mm256_madd_epi16(data_l, mask_l);
pred_l = _mm256_srai_epi32(_mm256_add_epi32(pred_l, round_const),
AOM_BLEND_A64_ROUND_BITS);
const __m256i data_r = _mm256_unpackhi_epi16(a, b);
const __m256i mask_r = _mm256_unpackhi_epi16(m, m_inv);
__m256i pred_r = _mm256_madd_epi16(data_r, mask_r);
pred_r = _mm256_srai_epi32(_mm256_add_epi32(pred_r, round_const),
AOM_BLEND_A64_ROUND_BITS);
// Note: the maximum value in pred_l/r is (2^bd)-1 < 2^15,
// so it is safe to do signed saturation here.
const __m256i pred = _mm256_packs_epi32(pred_l, pred_r);
// There is no 16-bit SAD instruction, so we have to synthesize
// an 8-element SAD. We do this by storing 4 32-bit partial SADs,
// and accumulating them at the end
const __m256i diff = _mm256_abs_epi16(_mm256_sub_epi16(pred, src));
res = _mm256_add_epi32(res, _mm256_madd_epi16(diff, one));
src_ptr += src_stride << 1;
a_ptr += a_stride << 1;
b_ptr += b_stride << 1;
m_ptr += m_stride << 1;
}
// At this point, we have four 32-bit partial SADs stored in 'res'.
res = _mm256_hadd_epi32(res, res);
res = _mm256_hadd_epi32(res, res);
int sad = _mm256_extract_epi32(res, 0) + _mm256_extract_epi32(res, 4);
return sad;
}
static INLINE unsigned int highbd_masked_sad16xh_avx2(
const uint8_t *src8, int src_stride, const uint8_t *a8, int a_stride,
const uint8_t *b8, int b_stride, const uint8_t *m_ptr, int m_stride,
int width, int height) {
const uint16_t *src_ptr = CONVERT_TO_SHORTPTR(src8);
const uint16_t *a_ptr = CONVERT_TO_SHORTPTR(a8);
const uint16_t *b_ptr = CONVERT_TO_SHORTPTR(b8);
int x, y;
__m256i res = _mm256_setzero_si256();
const __m256i mask_max = _mm256_set1_epi16((1 << AOM_BLEND_A64_ROUND_BITS));
const __m256i round_const =
_mm256_set1_epi32((1 << AOM_BLEND_A64_ROUND_BITS) >> 1);
const __m256i one = _mm256_set1_epi16(1);
for (y = 0; y < height; y++) {
for (x = 0; x < width; x += 16) {
const __m256i src = _mm256_lddqu_si256((const __m256i *)&src_ptr[x]);
const __m256i a = _mm256_lddqu_si256((const __m256i *)&a_ptr[x]);
const __m256i b = _mm256_lddqu_si256((const __m256i *)&b_ptr[x]);
// Zero-extend mask to 16 bits
const __m256i m =
_mm256_cvtepu8_epi16(_mm_lddqu_si128((const __m128i *)&m_ptr[x]));
const __m256i m_inv = _mm256_sub_epi16(mask_max, m);
const __m256i data_l = _mm256_unpacklo_epi16(a, b);
const __m256i mask_l = _mm256_unpacklo_epi16(m, m_inv);
__m256i pred_l = _mm256_madd_epi16(data_l, mask_l);
pred_l = _mm256_srai_epi32(_mm256_add_epi32(pred_l, round_const),
AOM_BLEND_A64_ROUND_BITS);
const __m256i data_r = _mm256_unpackhi_epi16(a, b);
const __m256i mask_r = _mm256_unpackhi_epi16(m, m_inv);
__m256i pred_r = _mm256_madd_epi16(data_r, mask_r);
pred_r = _mm256_srai_epi32(_mm256_add_epi32(pred_r, round_const),
AOM_BLEND_A64_ROUND_BITS);
// Note: the maximum value in pred_l/r is (2^bd)-1 < 2^15,
// so it is safe to do signed saturation here.
const __m256i pred = _mm256_packs_epi32(pred_l, pred_r);
// There is no 16-bit SAD instruction, so we have to synthesize
// an 8-element SAD. We do this by storing 4 32-bit partial SADs,
// and accumulating them at the end
const __m256i diff = _mm256_abs_epi16(_mm256_sub_epi16(pred, src));
res = _mm256_add_epi32(res, _mm256_madd_epi16(diff, one));
}
src_ptr += src_stride;
a_ptr += a_stride;
b_ptr += b_stride;
m_ptr += m_stride;
}
// At this point, we have four 32-bit partial SADs stored in 'res'.
res = _mm256_hadd_epi32(res, res);
res = _mm256_hadd_epi32(res, res);
int sad = _mm256_extract_epi32(res, 0) + _mm256_extract_epi32(res, 4);
return sad;
}
static INLINE unsigned int aom_highbd_masked_sad_avx2(
const uint8_t *src, int src_stride, const uint8_t *ref, int ref_stride,
const uint8_t *second_pred, const uint8_t *msk, int msk_stride,
int invert_mask, int m, int n) {
unsigned int sad;
if (!invert_mask) {
switch (m) {
case 4:
sad =
aom_highbd_masked_sad4xh_ssse3(src, src_stride, ref, ref_stride,
second_pred, m, msk, msk_stride, n);
break;
case 8:
sad = highbd_masked_sad8xh_avx2(src, src_stride, ref, ref_stride,
second_pred, m, msk, msk_stride, n);
break;
default:
sad = highbd_masked_sad16xh_avx2(src, src_stride, ref, ref_stride,
second_pred, m, msk, msk_stride, m, n);
break;
}
} else {
switch (m) {
case 4:
sad =
aom_highbd_masked_sad4xh_ssse3(src, src_stride, second_pred, m, ref,
ref_stride, msk, msk_stride, n);
break;
case 8:
sad = highbd_masked_sad8xh_avx2(src, src_stride, second_pred, m, ref,
ref_stride, msk, msk_stride, n);
break;
default:
sad = highbd_masked_sad16xh_avx2(src, src_stride, second_pred, m, ref,
ref_stride, msk, msk_stride, m, n);
break;
}
}
return sad;
}
#define HIGHBD_MASKSADMXN_AVX2(m, n) \
unsigned int aom_highbd_masked_sad##m##x##n##_avx2( \
const uint8_t *src8, int src_stride, const uint8_t *ref8, \
int ref_stride, const uint8_t *second_pred8, const uint8_t *msk, \
int msk_stride, int invert_mask) { \
return aom_highbd_masked_sad_avx2(src8, src_stride, ref8, ref_stride, \
second_pred8, msk, msk_stride, \
invert_mask, m, n); \
}
HIGHBD_MASKSADMXN_AVX2(4, 4)
HIGHBD_MASKSADMXN_AVX2(4, 8)
HIGHBD_MASKSADMXN_AVX2(8, 4)
HIGHBD_MASKSADMXN_AVX2(8, 8)
HIGHBD_MASKSADMXN_AVX2(8, 16)
HIGHBD_MASKSADMXN_AVX2(16, 8)
HIGHBD_MASKSADMXN_AVX2(16, 16)
HIGHBD_MASKSADMXN_AVX2(16, 32)
HIGHBD_MASKSADMXN_AVX2(32, 16)
HIGHBD_MASKSADMXN_AVX2(32, 32)
HIGHBD_MASKSADMXN_AVX2(32, 64)
HIGHBD_MASKSADMXN_AVX2(64, 32)
HIGHBD_MASKSADMXN_AVX2(64, 64)
HIGHBD_MASKSADMXN_AVX2(64, 128)
HIGHBD_MASKSADMXN_AVX2(128, 64)
HIGHBD_MASKSADMXN_AVX2(128, 128)
HIGHBD_MASKSADMXN_AVX2(4, 16)
HIGHBD_MASKSADMXN_AVX2(16, 4)
HIGHBD_MASKSADMXN_AVX2(8, 32)
HIGHBD_MASKSADMXN_AVX2(32, 8)
HIGHBD_MASKSADMXN_AVX2(16, 64)
HIGHBD_MASKSADMXN_AVX2(64, 16)
|